Loading...
1/* SPDX-License-Identifier: GPL-2.0+ */
2/*
3 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
4 * Internal non-public definitions that provide either classic
5 * or preemptible semantics.
6 *
7 * Copyright Red Hat, 2009
8 * Copyright IBM Corporation, 2009
9 *
10 * Author: Ingo Molnar <mingo@elte.hu>
11 * Paul E. McKenney <paulmck@linux.ibm.com>
12 */
13
14#include "../locking/rtmutex_common.h"
15
16#ifdef CONFIG_RCU_NOCB_CPU
17static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
18static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
19#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
20
21/*
22 * Check the RCU kernel configuration parameters and print informative
23 * messages about anything out of the ordinary.
24 */
25static void __init rcu_bootup_announce_oddness(void)
26{
27 if (IS_ENABLED(CONFIG_RCU_TRACE))
28 pr_info("\tRCU event tracing is enabled.\n");
29 if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
30 (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
31 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d.\n",
32 RCU_FANOUT);
33 if (rcu_fanout_exact)
34 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
35 if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ))
36 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
37 if (IS_ENABLED(CONFIG_PROVE_RCU))
38 pr_info("\tRCU lockdep checking is enabled.\n");
39 if (RCU_NUM_LVLS >= 4)
40 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
41 if (RCU_FANOUT_LEAF != 16)
42 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
43 RCU_FANOUT_LEAF);
44 if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
45 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n",
46 rcu_fanout_leaf);
47 if (nr_cpu_ids != NR_CPUS)
48 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%u.\n", NR_CPUS, nr_cpu_ids);
49#ifdef CONFIG_RCU_BOOST
50 pr_info("\tRCU priority boosting: priority %d delay %d ms.\n",
51 kthread_prio, CONFIG_RCU_BOOST_DELAY);
52#endif
53 if (blimit != DEFAULT_RCU_BLIMIT)
54 pr_info("\tBoot-time adjustment of callback invocation limit to %ld.\n", blimit);
55 if (qhimark != DEFAULT_RCU_QHIMARK)
56 pr_info("\tBoot-time adjustment of callback high-water mark to %ld.\n", qhimark);
57 if (qlowmark != DEFAULT_RCU_QLOMARK)
58 pr_info("\tBoot-time adjustment of callback low-water mark to %ld.\n", qlowmark);
59 if (qovld != DEFAULT_RCU_QOVLD)
60 pr_info("\tBoot-time adjustment of callback overload level to %ld.\n", qovld);
61 if (jiffies_till_first_fqs != ULONG_MAX)
62 pr_info("\tBoot-time adjustment of first FQS scan delay to %ld jiffies.\n", jiffies_till_first_fqs);
63 if (jiffies_till_next_fqs != ULONG_MAX)
64 pr_info("\tBoot-time adjustment of subsequent FQS scan delay to %ld jiffies.\n", jiffies_till_next_fqs);
65 if (jiffies_till_sched_qs != ULONG_MAX)
66 pr_info("\tBoot-time adjustment of scheduler-enlistment delay to %ld jiffies.\n", jiffies_till_sched_qs);
67 if (rcu_kick_kthreads)
68 pr_info("\tKick kthreads if too-long grace period.\n");
69 if (IS_ENABLED(CONFIG_DEBUG_OBJECTS_RCU_HEAD))
70 pr_info("\tRCU callback double-/use-after-free debug enabled.\n");
71 if (gp_preinit_delay)
72 pr_info("\tRCU debug GP pre-init slowdown %d jiffies.\n", gp_preinit_delay);
73 if (gp_init_delay)
74 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_init_delay);
75 if (gp_cleanup_delay)
76 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_cleanup_delay);
77 if (!use_softirq)
78 pr_info("\tRCU_SOFTIRQ processing moved to rcuc kthreads.\n");
79 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG))
80 pr_info("\tRCU debug extended QS entry/exit.\n");
81 rcupdate_announce_bootup_oddness();
82}
83
84#ifdef CONFIG_PREEMPT_RCU
85
86static void rcu_report_exp_rnp(struct rcu_node *rnp, bool wake);
87static void rcu_read_unlock_special(struct task_struct *t);
88
89/*
90 * Tell them what RCU they are running.
91 */
92static void __init rcu_bootup_announce(void)
93{
94 pr_info("Preemptible hierarchical RCU implementation.\n");
95 rcu_bootup_announce_oddness();
96}
97
98/* Flags for rcu_preempt_ctxt_queue() decision table. */
99#define RCU_GP_TASKS 0x8
100#define RCU_EXP_TASKS 0x4
101#define RCU_GP_BLKD 0x2
102#define RCU_EXP_BLKD 0x1
103
104/*
105 * Queues a task preempted within an RCU-preempt read-side critical
106 * section into the appropriate location within the ->blkd_tasks list,
107 * depending on the states of any ongoing normal and expedited grace
108 * periods. The ->gp_tasks pointer indicates which element the normal
109 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
110 * indicates which element the expedited grace period is waiting on (again,
111 * NULL if none). If a grace period is waiting on a given element in the
112 * ->blkd_tasks list, it also waits on all subsequent elements. Thus,
113 * adding a task to the tail of the list blocks any grace period that is
114 * already waiting on one of the elements. In contrast, adding a task
115 * to the head of the list won't block any grace period that is already
116 * waiting on one of the elements.
117 *
118 * This queuing is imprecise, and can sometimes make an ongoing grace
119 * period wait for a task that is not strictly speaking blocking it.
120 * Given the choice, we needlessly block a normal grace period rather than
121 * blocking an expedited grace period.
122 *
123 * Note that an endless sequence of expedited grace periods still cannot
124 * indefinitely postpone a normal grace period. Eventually, all of the
125 * fixed number of preempted tasks blocking the normal grace period that are
126 * not also blocking the expedited grace period will resume and complete
127 * their RCU read-side critical sections. At that point, the ->gp_tasks
128 * pointer will equal the ->exp_tasks pointer, at which point the end of
129 * the corresponding expedited grace period will also be the end of the
130 * normal grace period.
131 */
132static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
133 __releases(rnp->lock) /* But leaves rrupts disabled. */
134{
135 int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
136 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
137 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
138 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
139 struct task_struct *t = current;
140
141 raw_lockdep_assert_held_rcu_node(rnp);
142 WARN_ON_ONCE(rdp->mynode != rnp);
143 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
144 /* RCU better not be waiting on newly onlined CPUs! */
145 WARN_ON_ONCE(rnp->qsmaskinitnext & ~rnp->qsmaskinit & rnp->qsmask &
146 rdp->grpmask);
147
148 /*
149 * Decide where to queue the newly blocked task. In theory,
150 * this could be an if-statement. In practice, when I tried
151 * that, it was quite messy.
152 */
153 switch (blkd_state) {
154 case 0:
155 case RCU_EXP_TASKS:
156 case RCU_EXP_TASKS + RCU_GP_BLKD:
157 case RCU_GP_TASKS:
158 case RCU_GP_TASKS + RCU_EXP_TASKS:
159
160 /*
161 * Blocking neither GP, or first task blocking the normal
162 * GP but not blocking the already-waiting expedited GP.
163 * Queue at the head of the list to avoid unnecessarily
164 * blocking the already-waiting GPs.
165 */
166 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
167 break;
168
169 case RCU_EXP_BLKD:
170 case RCU_GP_BLKD:
171 case RCU_GP_BLKD + RCU_EXP_BLKD:
172 case RCU_GP_TASKS + RCU_EXP_BLKD:
173 case RCU_GP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
174 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
175
176 /*
177 * First task arriving that blocks either GP, or first task
178 * arriving that blocks the expedited GP (with the normal
179 * GP already waiting), or a task arriving that blocks
180 * both GPs with both GPs already waiting. Queue at the
181 * tail of the list to avoid any GP waiting on any of the
182 * already queued tasks that are not blocking it.
183 */
184 list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
185 break;
186
187 case RCU_EXP_TASKS + RCU_EXP_BLKD:
188 case RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
189 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_EXP_BLKD:
190
191 /*
192 * Second or subsequent task blocking the expedited GP.
193 * The task either does not block the normal GP, or is the
194 * first task blocking the normal GP. Queue just after
195 * the first task blocking the expedited GP.
196 */
197 list_add(&t->rcu_node_entry, rnp->exp_tasks);
198 break;
199
200 case RCU_GP_TASKS + RCU_GP_BLKD:
201 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
202
203 /*
204 * Second or subsequent task blocking the normal GP.
205 * The task does not block the expedited GP. Queue just
206 * after the first task blocking the normal GP.
207 */
208 list_add(&t->rcu_node_entry, rnp->gp_tasks);
209 break;
210
211 default:
212
213 /* Yet another exercise in excessive paranoia. */
214 WARN_ON_ONCE(1);
215 break;
216 }
217
218 /*
219 * We have now queued the task. If it was the first one to
220 * block either grace period, update the ->gp_tasks and/or
221 * ->exp_tasks pointers, respectively, to reference the newly
222 * blocked tasks.
223 */
224 if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD)) {
225 WRITE_ONCE(rnp->gp_tasks, &t->rcu_node_entry);
226 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq);
227 }
228 if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
229 WRITE_ONCE(rnp->exp_tasks, &t->rcu_node_entry);
230 WARN_ON_ONCE(!(blkd_state & RCU_GP_BLKD) !=
231 !(rnp->qsmask & rdp->grpmask));
232 WARN_ON_ONCE(!(blkd_state & RCU_EXP_BLKD) !=
233 !(rnp->expmask & rdp->grpmask));
234 raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
235
236 /*
237 * Report the quiescent state for the expedited GP. This expedited
238 * GP should not be able to end until we report, so there should be
239 * no need to check for a subsequent expedited GP. (Though we are
240 * still in a quiescent state in any case.)
241 */
242 if (blkd_state & RCU_EXP_BLKD && rdp->exp_deferred_qs)
243 rcu_report_exp_rdp(rdp);
244 else
245 WARN_ON_ONCE(rdp->exp_deferred_qs);
246}
247
248/*
249 * Record a preemptible-RCU quiescent state for the specified CPU.
250 * Note that this does not necessarily mean that the task currently running
251 * on the CPU is in a quiescent state: Instead, it means that the current
252 * grace period need not wait on any RCU read-side critical section that
253 * starts later on this CPU. It also means that if the current task is
254 * in an RCU read-side critical section, it has already added itself to
255 * some leaf rcu_node structure's ->blkd_tasks list. In addition to the
256 * current task, there might be any number of other tasks blocked while
257 * in an RCU read-side critical section.
258 *
259 * Callers to this function must disable preemption.
260 */
261static void rcu_qs(void)
262{
263 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!\n");
264 if (__this_cpu_read(rcu_data.cpu_no_qs.s)) {
265 trace_rcu_grace_period(TPS("rcu_preempt"),
266 __this_cpu_read(rcu_data.gp_seq),
267 TPS("cpuqs"));
268 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
269 barrier(); /* Coordinate with rcu_flavor_sched_clock_irq(). */
270 WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, false);
271 }
272}
273
274/*
275 * We have entered the scheduler, and the current task might soon be
276 * context-switched away from. If this task is in an RCU read-side
277 * critical section, we will no longer be able to rely on the CPU to
278 * record that fact, so we enqueue the task on the blkd_tasks list.
279 * The task will dequeue itself when it exits the outermost enclosing
280 * RCU read-side critical section. Therefore, the current grace period
281 * cannot be permitted to complete until the blkd_tasks list entries
282 * predating the current grace period drain, in other words, until
283 * rnp->gp_tasks becomes NULL.
284 *
285 * Caller must disable interrupts.
286 */
287void rcu_note_context_switch(bool preempt)
288{
289 struct task_struct *t = current;
290 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
291 struct rcu_node *rnp;
292
293 trace_rcu_utilization(TPS("Start context switch"));
294 lockdep_assert_irqs_disabled();
295 WARN_ON_ONCE(!preempt && rcu_preempt_depth() > 0);
296 if (rcu_preempt_depth() > 0 &&
297 !t->rcu_read_unlock_special.b.blocked) {
298
299 /* Possibly blocking in an RCU read-side critical section. */
300 rnp = rdp->mynode;
301 raw_spin_lock_rcu_node(rnp);
302 t->rcu_read_unlock_special.b.blocked = true;
303 t->rcu_blocked_node = rnp;
304
305 /*
306 * Verify the CPU's sanity, trace the preemption, and
307 * then queue the task as required based on the states
308 * of any ongoing and expedited grace periods.
309 */
310 WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
311 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
312 trace_rcu_preempt_task(rcu_state.name,
313 t->pid,
314 (rnp->qsmask & rdp->grpmask)
315 ? rnp->gp_seq
316 : rcu_seq_snap(&rnp->gp_seq));
317 rcu_preempt_ctxt_queue(rnp, rdp);
318 } else {
319 rcu_preempt_deferred_qs(t);
320 }
321
322 /*
323 * Either we were not in an RCU read-side critical section to
324 * begin with, or we have now recorded that critical section
325 * globally. Either way, we can now note a quiescent state
326 * for this CPU. Again, if we were in an RCU read-side critical
327 * section, and if that critical section was blocking the current
328 * grace period, then the fact that the task has been enqueued
329 * means that we continue to block the current grace period.
330 */
331 rcu_qs();
332 if (rdp->exp_deferred_qs)
333 rcu_report_exp_rdp(rdp);
334 rcu_tasks_qs(current, preempt);
335 trace_rcu_utilization(TPS("End context switch"));
336}
337EXPORT_SYMBOL_GPL(rcu_note_context_switch);
338
339/*
340 * Check for preempted RCU readers blocking the current grace period
341 * for the specified rcu_node structure. If the caller needs a reliable
342 * answer, it must hold the rcu_node's ->lock.
343 */
344static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
345{
346 return READ_ONCE(rnp->gp_tasks) != NULL;
347}
348
349/* limit value for ->rcu_read_lock_nesting. */
350#define RCU_NEST_PMAX (INT_MAX / 2)
351
352static void rcu_preempt_read_enter(void)
353{
354 current->rcu_read_lock_nesting++;
355}
356
357static int rcu_preempt_read_exit(void)
358{
359 return --current->rcu_read_lock_nesting;
360}
361
362static void rcu_preempt_depth_set(int val)
363{
364 current->rcu_read_lock_nesting = val;
365}
366
367/*
368 * Preemptible RCU implementation for rcu_read_lock().
369 * Just increment ->rcu_read_lock_nesting, shared state will be updated
370 * if we block.
371 */
372void __rcu_read_lock(void)
373{
374 rcu_preempt_read_enter();
375 if (IS_ENABLED(CONFIG_PROVE_LOCKING))
376 WARN_ON_ONCE(rcu_preempt_depth() > RCU_NEST_PMAX);
377 barrier(); /* critical section after entry code. */
378}
379EXPORT_SYMBOL_GPL(__rcu_read_lock);
380
381/*
382 * Preemptible RCU implementation for rcu_read_unlock().
383 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
384 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
385 * invoke rcu_read_unlock_special() to clean up after a context switch
386 * in an RCU read-side critical section and other special cases.
387 */
388void __rcu_read_unlock(void)
389{
390 struct task_struct *t = current;
391
392 if (rcu_preempt_read_exit() == 0) {
393 barrier(); /* critical section before exit code. */
394 if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
395 rcu_read_unlock_special(t);
396 }
397 if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
398 int rrln = rcu_preempt_depth();
399
400 WARN_ON_ONCE(rrln < 0 || rrln > RCU_NEST_PMAX);
401 }
402}
403EXPORT_SYMBOL_GPL(__rcu_read_unlock);
404
405/*
406 * Advance a ->blkd_tasks-list pointer to the next entry, instead
407 * returning NULL if at the end of the list.
408 */
409static struct list_head *rcu_next_node_entry(struct task_struct *t,
410 struct rcu_node *rnp)
411{
412 struct list_head *np;
413
414 np = t->rcu_node_entry.next;
415 if (np == &rnp->blkd_tasks)
416 np = NULL;
417 return np;
418}
419
420/*
421 * Return true if the specified rcu_node structure has tasks that were
422 * preempted within an RCU read-side critical section.
423 */
424static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
425{
426 return !list_empty(&rnp->blkd_tasks);
427}
428
429/*
430 * Report deferred quiescent states. The deferral time can
431 * be quite short, for example, in the case of the call from
432 * rcu_read_unlock_special().
433 */
434static void
435rcu_preempt_deferred_qs_irqrestore(struct task_struct *t, unsigned long flags)
436{
437 bool empty_exp;
438 bool empty_norm;
439 bool empty_exp_now;
440 struct list_head *np;
441 bool drop_boost_mutex = false;
442 struct rcu_data *rdp;
443 struct rcu_node *rnp;
444 union rcu_special special;
445
446 /*
447 * If RCU core is waiting for this CPU to exit its critical section,
448 * report the fact that it has exited. Because irqs are disabled,
449 * t->rcu_read_unlock_special cannot change.
450 */
451 special = t->rcu_read_unlock_special;
452 rdp = this_cpu_ptr(&rcu_data);
453 if (!special.s && !rdp->exp_deferred_qs) {
454 local_irq_restore(flags);
455 return;
456 }
457 t->rcu_read_unlock_special.s = 0;
458 if (special.b.need_qs)
459 rcu_qs();
460
461 /*
462 * Respond to a request by an expedited grace period for a
463 * quiescent state from this CPU. Note that requests from
464 * tasks are handled when removing the task from the
465 * blocked-tasks list below.
466 */
467 if (rdp->exp_deferred_qs)
468 rcu_report_exp_rdp(rdp);
469
470 /* Clean up if blocked during RCU read-side critical section. */
471 if (special.b.blocked) {
472
473 /*
474 * Remove this task from the list it blocked on. The task
475 * now remains queued on the rcu_node corresponding to the
476 * CPU it first blocked on, so there is no longer any need
477 * to loop. Retain a WARN_ON_ONCE() out of sheer paranoia.
478 */
479 rnp = t->rcu_blocked_node;
480 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
481 WARN_ON_ONCE(rnp != t->rcu_blocked_node);
482 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
483 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
484 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq &&
485 (!empty_norm || rnp->qsmask));
486 empty_exp = sync_rcu_exp_done(rnp);
487 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
488 np = rcu_next_node_entry(t, rnp);
489 list_del_init(&t->rcu_node_entry);
490 t->rcu_blocked_node = NULL;
491 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
492 rnp->gp_seq, t->pid);
493 if (&t->rcu_node_entry == rnp->gp_tasks)
494 WRITE_ONCE(rnp->gp_tasks, np);
495 if (&t->rcu_node_entry == rnp->exp_tasks)
496 WRITE_ONCE(rnp->exp_tasks, np);
497 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
498 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
499 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
500 if (&t->rcu_node_entry == rnp->boost_tasks)
501 WRITE_ONCE(rnp->boost_tasks, np);
502 }
503
504 /*
505 * If this was the last task on the current list, and if
506 * we aren't waiting on any CPUs, report the quiescent state.
507 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
508 * so we must take a snapshot of the expedited state.
509 */
510 empty_exp_now = sync_rcu_exp_done(rnp);
511 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
512 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
513 rnp->gp_seq,
514 0, rnp->qsmask,
515 rnp->level,
516 rnp->grplo,
517 rnp->grphi,
518 !!rnp->gp_tasks);
519 rcu_report_unblock_qs_rnp(rnp, flags);
520 } else {
521 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
522 }
523
524 /* Unboost if we were boosted. */
525 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
526 rt_mutex_futex_unlock(&rnp->boost_mtx);
527
528 /*
529 * If this was the last task on the expedited lists,
530 * then we need to report up the rcu_node hierarchy.
531 */
532 if (!empty_exp && empty_exp_now)
533 rcu_report_exp_rnp(rnp, true);
534 } else {
535 local_irq_restore(flags);
536 }
537}
538
539/*
540 * Is a deferred quiescent-state pending, and are we also not in
541 * an RCU read-side critical section? It is the caller's responsibility
542 * to ensure it is otherwise safe to report any deferred quiescent
543 * states. The reason for this is that it is safe to report a
544 * quiescent state during context switch even though preemption
545 * is disabled. This function cannot be expected to understand these
546 * nuances, so the caller must handle them.
547 */
548static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
549{
550 return (__this_cpu_read(rcu_data.exp_deferred_qs) ||
551 READ_ONCE(t->rcu_read_unlock_special.s)) &&
552 rcu_preempt_depth() == 0;
553}
554
555/*
556 * Report a deferred quiescent state if needed and safe to do so.
557 * As with rcu_preempt_need_deferred_qs(), "safe" involves only
558 * not being in an RCU read-side critical section. The caller must
559 * evaluate safety in terms of interrupt, softirq, and preemption
560 * disabling.
561 */
562static void rcu_preempt_deferred_qs(struct task_struct *t)
563{
564 unsigned long flags;
565
566 if (!rcu_preempt_need_deferred_qs(t))
567 return;
568 local_irq_save(flags);
569 rcu_preempt_deferred_qs_irqrestore(t, flags);
570}
571
572/*
573 * Minimal handler to give the scheduler a chance to re-evaluate.
574 */
575static void rcu_preempt_deferred_qs_handler(struct irq_work *iwp)
576{
577 struct rcu_data *rdp;
578
579 rdp = container_of(iwp, struct rcu_data, defer_qs_iw);
580 rdp->defer_qs_iw_pending = false;
581}
582
583/*
584 * Handle special cases during rcu_read_unlock(), such as needing to
585 * notify RCU core processing or task having blocked during the RCU
586 * read-side critical section.
587 */
588static void rcu_read_unlock_special(struct task_struct *t)
589{
590 unsigned long flags;
591 bool preempt_bh_were_disabled =
592 !!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
593 bool irqs_were_disabled;
594
595 /* NMI handlers cannot block and cannot safely manipulate state. */
596 if (in_nmi())
597 return;
598
599 local_irq_save(flags);
600 irqs_were_disabled = irqs_disabled_flags(flags);
601 if (preempt_bh_were_disabled || irqs_were_disabled) {
602 bool exp;
603 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
604 struct rcu_node *rnp = rdp->mynode;
605
606 exp = (t->rcu_blocked_node &&
607 READ_ONCE(t->rcu_blocked_node->exp_tasks)) ||
608 (rdp->grpmask & READ_ONCE(rnp->expmask));
609 // Need to defer quiescent state until everything is enabled.
610 if (use_softirq && (in_irq() || (exp && !irqs_were_disabled))) {
611 // Using softirq, safe to awaken, and either the
612 // wakeup is free or there is an expedited GP.
613 raise_softirq_irqoff(RCU_SOFTIRQ);
614 } else {
615 // Enabling BH or preempt does reschedule, so...
616 // Also if no expediting, slow is OK.
617 // Plus nohz_full CPUs eventually get tick enabled.
618 set_tsk_need_resched(current);
619 set_preempt_need_resched();
620 if (IS_ENABLED(CONFIG_IRQ_WORK) && irqs_were_disabled &&
621 !rdp->defer_qs_iw_pending && exp) {
622 // Get scheduler to re-evaluate and call hooks.
623 // If !IRQ_WORK, FQS scan will eventually IPI.
624 init_irq_work(&rdp->defer_qs_iw,
625 rcu_preempt_deferred_qs_handler);
626 rdp->defer_qs_iw_pending = true;
627 irq_work_queue_on(&rdp->defer_qs_iw, rdp->cpu);
628 }
629 }
630 local_irq_restore(flags);
631 return;
632 }
633 rcu_preempt_deferred_qs_irqrestore(t, flags);
634}
635
636/*
637 * Check that the list of blocked tasks for the newly completed grace
638 * period is in fact empty. It is a serious bug to complete a grace
639 * period that still has RCU readers blocked! This function must be
640 * invoked -before- updating this rnp's ->gp_seq.
641 *
642 * Also, if there are blocked tasks on the list, they automatically
643 * block the newly created grace period, so set up ->gp_tasks accordingly.
644 */
645static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
646{
647 struct task_struct *t;
648
649 RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
650 raw_lockdep_assert_held_rcu_node(rnp);
651 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
652 dump_blkd_tasks(rnp, 10);
653 if (rcu_preempt_has_tasks(rnp) &&
654 (rnp->qsmaskinit || rnp->wait_blkd_tasks)) {
655 WRITE_ONCE(rnp->gp_tasks, rnp->blkd_tasks.next);
656 t = container_of(rnp->gp_tasks, struct task_struct,
657 rcu_node_entry);
658 trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
659 rnp->gp_seq, t->pid);
660 }
661 WARN_ON_ONCE(rnp->qsmask);
662}
663
664/*
665 * Check for a quiescent state from the current CPU, including voluntary
666 * context switches for Tasks RCU. When a task blocks, the task is
667 * recorded in the corresponding CPU's rcu_node structure, which is checked
668 * elsewhere, hence this function need only check for quiescent states
669 * related to the current CPU, not to those related to tasks.
670 */
671static void rcu_flavor_sched_clock_irq(int user)
672{
673 struct task_struct *t = current;
674
675 if (user || rcu_is_cpu_rrupt_from_idle()) {
676 rcu_note_voluntary_context_switch(current);
677 }
678 if (rcu_preempt_depth() > 0 ||
679 (preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK))) {
680 /* No QS, force context switch if deferred. */
681 if (rcu_preempt_need_deferred_qs(t)) {
682 set_tsk_need_resched(t);
683 set_preempt_need_resched();
684 }
685 } else if (rcu_preempt_need_deferred_qs(t)) {
686 rcu_preempt_deferred_qs(t); /* Report deferred QS. */
687 return;
688 } else if (!WARN_ON_ONCE(rcu_preempt_depth())) {
689 rcu_qs(); /* Report immediate QS. */
690 return;
691 }
692
693 /* If GP is oldish, ask for help from rcu_read_unlock_special(). */
694 if (rcu_preempt_depth() > 0 &&
695 __this_cpu_read(rcu_data.core_needs_qs) &&
696 __this_cpu_read(rcu_data.cpu_no_qs.b.norm) &&
697 !t->rcu_read_unlock_special.b.need_qs &&
698 time_after(jiffies, rcu_state.gp_start + HZ))
699 t->rcu_read_unlock_special.b.need_qs = true;
700}
701
702/*
703 * Check for a task exiting while in a preemptible-RCU read-side
704 * critical section, clean up if so. No need to issue warnings, as
705 * debug_check_no_locks_held() already does this if lockdep is enabled.
706 * Besides, if this function does anything other than just immediately
707 * return, there was a bug of some sort. Spewing warnings from this
708 * function is like as not to simply obscure important prior warnings.
709 */
710void exit_rcu(void)
711{
712 struct task_struct *t = current;
713
714 if (unlikely(!list_empty(¤t->rcu_node_entry))) {
715 rcu_preempt_depth_set(1);
716 barrier();
717 WRITE_ONCE(t->rcu_read_unlock_special.b.blocked, true);
718 } else if (unlikely(rcu_preempt_depth())) {
719 rcu_preempt_depth_set(1);
720 } else {
721 return;
722 }
723 __rcu_read_unlock();
724 rcu_preempt_deferred_qs(current);
725}
726
727/*
728 * Dump the blocked-tasks state, but limit the list dump to the
729 * specified number of elements.
730 */
731static void
732dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
733{
734 int cpu;
735 int i;
736 struct list_head *lhp;
737 bool onl;
738 struct rcu_data *rdp;
739 struct rcu_node *rnp1;
740
741 raw_lockdep_assert_held_rcu_node(rnp);
742 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
743 __func__, rnp->grplo, rnp->grphi, rnp->level,
744 (long)READ_ONCE(rnp->gp_seq), (long)rnp->completedqs);
745 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
746 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx\n",
747 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext);
748 pr_info("%s: ->gp_tasks %p ->boost_tasks %p ->exp_tasks %p\n",
749 __func__, READ_ONCE(rnp->gp_tasks), data_race(rnp->boost_tasks),
750 READ_ONCE(rnp->exp_tasks));
751 pr_info("%s: ->blkd_tasks", __func__);
752 i = 0;
753 list_for_each(lhp, &rnp->blkd_tasks) {
754 pr_cont(" %p", lhp);
755 if (++i >= ncheck)
756 break;
757 }
758 pr_cont("\n");
759 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
760 rdp = per_cpu_ptr(&rcu_data, cpu);
761 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
762 pr_info("\t%d: %c online: %ld(%d) offline: %ld(%d)\n",
763 cpu, ".o"[onl],
764 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
765 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
766 }
767}
768
769#else /* #ifdef CONFIG_PREEMPT_RCU */
770
771/*
772 * Tell them what RCU they are running.
773 */
774static void __init rcu_bootup_announce(void)
775{
776 pr_info("Hierarchical RCU implementation.\n");
777 rcu_bootup_announce_oddness();
778}
779
780/*
781 * Note a quiescent state for PREEMPTION=n. Because we do not need to know
782 * how many quiescent states passed, just if there was at least one since
783 * the start of the grace period, this just sets a flag. The caller must
784 * have disabled preemption.
785 */
786static void rcu_qs(void)
787{
788 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!");
789 if (!__this_cpu_read(rcu_data.cpu_no_qs.s))
790 return;
791 trace_rcu_grace_period(TPS("rcu_sched"),
792 __this_cpu_read(rcu_data.gp_seq), TPS("cpuqs"));
793 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
794 if (!__this_cpu_read(rcu_data.cpu_no_qs.b.exp))
795 return;
796 __this_cpu_write(rcu_data.cpu_no_qs.b.exp, false);
797 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
798}
799
800/*
801 * Register an urgently needed quiescent state. If there is an
802 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
803 * dyntick-idle quiescent state visible to other CPUs, which will in
804 * some cases serve for expedited as well as normal grace periods.
805 * Either way, register a lightweight quiescent state.
806 */
807void rcu_all_qs(void)
808{
809 unsigned long flags;
810
811 if (!raw_cpu_read(rcu_data.rcu_urgent_qs))
812 return;
813 preempt_disable();
814 /* Load rcu_urgent_qs before other flags. */
815 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
816 preempt_enable();
817 return;
818 }
819 this_cpu_write(rcu_data.rcu_urgent_qs, false);
820 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs))) {
821 local_irq_save(flags);
822 rcu_momentary_dyntick_idle();
823 local_irq_restore(flags);
824 }
825 rcu_qs();
826 preempt_enable();
827}
828EXPORT_SYMBOL_GPL(rcu_all_qs);
829
830/*
831 * Note a PREEMPTION=n context switch. The caller must have disabled interrupts.
832 */
833void rcu_note_context_switch(bool preempt)
834{
835 trace_rcu_utilization(TPS("Start context switch"));
836 rcu_qs();
837 /* Load rcu_urgent_qs before other flags. */
838 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs)))
839 goto out;
840 this_cpu_write(rcu_data.rcu_urgent_qs, false);
841 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs)))
842 rcu_momentary_dyntick_idle();
843 rcu_tasks_qs(current, preempt);
844out:
845 trace_rcu_utilization(TPS("End context switch"));
846}
847EXPORT_SYMBOL_GPL(rcu_note_context_switch);
848
849/*
850 * Because preemptible RCU does not exist, there are never any preempted
851 * RCU readers.
852 */
853static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
854{
855 return 0;
856}
857
858/*
859 * Because there is no preemptible RCU, there can be no readers blocked.
860 */
861static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
862{
863 return false;
864}
865
866/*
867 * Because there is no preemptible RCU, there can be no deferred quiescent
868 * states.
869 */
870static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
871{
872 return false;
873}
874static void rcu_preempt_deferred_qs(struct task_struct *t) { }
875
876/*
877 * Because there is no preemptible RCU, there can be no readers blocked,
878 * so there is no need to check for blocked tasks. So check only for
879 * bogus qsmask values.
880 */
881static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
882{
883 WARN_ON_ONCE(rnp->qsmask);
884}
885
886/*
887 * Check to see if this CPU is in a non-context-switch quiescent state,
888 * namely user mode and idle loop.
889 */
890static void rcu_flavor_sched_clock_irq(int user)
891{
892 if (user || rcu_is_cpu_rrupt_from_idle()) {
893
894 /*
895 * Get here if this CPU took its interrupt from user
896 * mode or from the idle loop, and if this is not a
897 * nested interrupt. In this case, the CPU is in
898 * a quiescent state, so note it.
899 *
900 * No memory barrier is required here because rcu_qs()
901 * references only CPU-local variables that other CPUs
902 * neither access nor modify, at least not while the
903 * corresponding CPU is online.
904 */
905
906 rcu_qs();
907 }
908}
909
910/*
911 * Because preemptible RCU does not exist, tasks cannot possibly exit
912 * while in preemptible RCU read-side critical sections.
913 */
914void exit_rcu(void)
915{
916}
917
918/*
919 * Dump the guaranteed-empty blocked-tasks state. Trust but verify.
920 */
921static void
922dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
923{
924 WARN_ON_ONCE(!list_empty(&rnp->blkd_tasks));
925}
926
927#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
928
929/*
930 * If boosting, set rcuc kthreads to realtime priority.
931 */
932static void rcu_cpu_kthread_setup(unsigned int cpu)
933{
934#ifdef CONFIG_RCU_BOOST
935 struct sched_param sp;
936
937 sp.sched_priority = kthread_prio;
938 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
939#endif /* #ifdef CONFIG_RCU_BOOST */
940}
941
942#ifdef CONFIG_RCU_BOOST
943
944/*
945 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
946 * or ->boost_tasks, advancing the pointer to the next task in the
947 * ->blkd_tasks list.
948 *
949 * Note that irqs must be enabled: boosting the task can block.
950 * Returns 1 if there are more tasks needing to be boosted.
951 */
952static int rcu_boost(struct rcu_node *rnp)
953{
954 unsigned long flags;
955 struct task_struct *t;
956 struct list_head *tb;
957
958 if (READ_ONCE(rnp->exp_tasks) == NULL &&
959 READ_ONCE(rnp->boost_tasks) == NULL)
960 return 0; /* Nothing left to boost. */
961
962 raw_spin_lock_irqsave_rcu_node(rnp, flags);
963
964 /*
965 * Recheck under the lock: all tasks in need of boosting
966 * might exit their RCU read-side critical sections on their own.
967 */
968 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
969 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
970 return 0;
971 }
972
973 /*
974 * Preferentially boost tasks blocking expedited grace periods.
975 * This cannot starve the normal grace periods because a second
976 * expedited grace period must boost all blocked tasks, including
977 * those blocking the pre-existing normal grace period.
978 */
979 if (rnp->exp_tasks != NULL)
980 tb = rnp->exp_tasks;
981 else
982 tb = rnp->boost_tasks;
983
984 /*
985 * We boost task t by manufacturing an rt_mutex that appears to
986 * be held by task t. We leave a pointer to that rt_mutex where
987 * task t can find it, and task t will release the mutex when it
988 * exits its outermost RCU read-side critical section. Then
989 * simply acquiring this artificial rt_mutex will boost task
990 * t's priority. (Thanks to tglx for suggesting this approach!)
991 *
992 * Note that task t must acquire rnp->lock to remove itself from
993 * the ->blkd_tasks list, which it will do from exit() if from
994 * nowhere else. We therefore are guaranteed that task t will
995 * stay around at least until we drop rnp->lock. Note that
996 * rnp->lock also resolves races between our priority boosting
997 * and task t's exiting its outermost RCU read-side critical
998 * section.
999 */
1000 t = container_of(tb, struct task_struct, rcu_node_entry);
1001 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1002 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1003 /* Lock only for side effect: boosts task t's priority. */
1004 rt_mutex_lock(&rnp->boost_mtx);
1005 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1006
1007 return READ_ONCE(rnp->exp_tasks) != NULL ||
1008 READ_ONCE(rnp->boost_tasks) != NULL;
1009}
1010
1011/*
1012 * Priority-boosting kthread, one per leaf rcu_node.
1013 */
1014static int rcu_boost_kthread(void *arg)
1015{
1016 struct rcu_node *rnp = (struct rcu_node *)arg;
1017 int spincnt = 0;
1018 int more2boost;
1019
1020 trace_rcu_utilization(TPS("Start boost kthread@init"));
1021 for (;;) {
1022 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_WAITING);
1023 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1024 rcu_wait(READ_ONCE(rnp->boost_tasks) ||
1025 READ_ONCE(rnp->exp_tasks));
1026 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1027 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_RUNNING);
1028 more2boost = rcu_boost(rnp);
1029 if (more2boost)
1030 spincnt++;
1031 else
1032 spincnt = 0;
1033 if (spincnt > 10) {
1034 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_YIELDING);
1035 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1036 schedule_timeout_idle(2);
1037 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1038 spincnt = 0;
1039 }
1040 }
1041 /* NOTREACHED */
1042 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1043 return 0;
1044}
1045
1046/*
1047 * Check to see if it is time to start boosting RCU readers that are
1048 * blocking the current grace period, and, if so, tell the per-rcu_node
1049 * kthread to start boosting them. If there is an expedited grace
1050 * period in progress, it is always time to boost.
1051 *
1052 * The caller must hold rnp->lock, which this function releases.
1053 * The ->boost_kthread_task is immortal, so we don't need to worry
1054 * about it going away.
1055 */
1056static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1057 __releases(rnp->lock)
1058{
1059 raw_lockdep_assert_held_rcu_node(rnp);
1060 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1061 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1062 return;
1063 }
1064 if (rnp->exp_tasks != NULL ||
1065 (rnp->gp_tasks != NULL &&
1066 rnp->boost_tasks == NULL &&
1067 rnp->qsmask == 0 &&
1068 (!time_after(rnp->boost_time, jiffies) || rcu_state.cbovld))) {
1069 if (rnp->exp_tasks == NULL)
1070 WRITE_ONCE(rnp->boost_tasks, rnp->gp_tasks);
1071 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1072 rcu_wake_cond(rnp->boost_kthread_task,
1073 READ_ONCE(rnp->boost_kthread_status));
1074 } else {
1075 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1076 }
1077}
1078
1079/*
1080 * Is the current CPU running the RCU-callbacks kthread?
1081 * Caller must have preemption disabled.
1082 */
1083static bool rcu_is_callbacks_kthread(void)
1084{
1085 return __this_cpu_read(rcu_data.rcu_cpu_kthread_task) == current;
1086}
1087
1088#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1089
1090/*
1091 * Do priority-boost accounting for the start of a new grace period.
1092 */
1093static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1094{
1095 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1096}
1097
1098/*
1099 * Create an RCU-boost kthread for the specified node if one does not
1100 * already exist. We only create this kthread for preemptible RCU.
1101 * Returns zero if all is well, a negated errno otherwise.
1102 */
1103static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1104{
1105 int rnp_index = rnp - rcu_get_root();
1106 unsigned long flags;
1107 struct sched_param sp;
1108 struct task_struct *t;
1109
1110 if (!IS_ENABLED(CONFIG_PREEMPT_RCU))
1111 return;
1112
1113 if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1114 return;
1115
1116 rcu_state.boost = 1;
1117
1118 if (rnp->boost_kthread_task != NULL)
1119 return;
1120
1121 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1122 "rcub/%d", rnp_index);
1123 if (WARN_ON_ONCE(IS_ERR(t)))
1124 return;
1125
1126 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1127 rnp->boost_kthread_task = t;
1128 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1129 sp.sched_priority = kthread_prio;
1130 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1131 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1132}
1133
1134/*
1135 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1136 * served by the rcu_node in question. The CPU hotplug lock is still
1137 * held, so the value of rnp->qsmaskinit will be stable.
1138 *
1139 * We don't include outgoingcpu in the affinity set, use -1 if there is
1140 * no outgoing CPU. If there are no CPUs left in the affinity set,
1141 * this function allows the kthread to execute on any CPU.
1142 */
1143static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1144{
1145 struct task_struct *t = rnp->boost_kthread_task;
1146 unsigned long mask = rcu_rnp_online_cpus(rnp);
1147 cpumask_var_t cm;
1148 int cpu;
1149
1150 if (!t)
1151 return;
1152 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1153 return;
1154 for_each_leaf_node_possible_cpu(rnp, cpu)
1155 if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
1156 cpu != outgoingcpu)
1157 cpumask_set_cpu(cpu, cm);
1158 if (cpumask_weight(cm) == 0)
1159 cpumask_setall(cm);
1160 set_cpus_allowed_ptr(t, cm);
1161 free_cpumask_var(cm);
1162}
1163
1164/*
1165 * Spawn boost kthreads -- called as soon as the scheduler is running.
1166 */
1167static void __init rcu_spawn_boost_kthreads(void)
1168{
1169 struct rcu_node *rnp;
1170
1171 rcu_for_each_leaf_node(rnp)
1172 rcu_spawn_one_boost_kthread(rnp);
1173}
1174
1175static void rcu_prepare_kthreads(int cpu)
1176{
1177 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1178 struct rcu_node *rnp = rdp->mynode;
1179
1180 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1181 if (rcu_scheduler_fully_active)
1182 rcu_spawn_one_boost_kthread(rnp);
1183}
1184
1185#else /* #ifdef CONFIG_RCU_BOOST */
1186
1187static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1188 __releases(rnp->lock)
1189{
1190 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1191}
1192
1193static bool rcu_is_callbacks_kthread(void)
1194{
1195 return false;
1196}
1197
1198static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1199{
1200}
1201
1202static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1203{
1204}
1205
1206static void __init rcu_spawn_boost_kthreads(void)
1207{
1208}
1209
1210static void rcu_prepare_kthreads(int cpu)
1211{
1212}
1213
1214#endif /* #else #ifdef CONFIG_RCU_BOOST */
1215
1216#if !defined(CONFIG_RCU_FAST_NO_HZ)
1217
1218/*
1219 * Check to see if any future non-offloaded RCU-related work will need
1220 * to be done by the current CPU, even if none need be done immediately,
1221 * returning 1 if so. This function is part of the RCU implementation;
1222 * it is -not- an exported member of the RCU API.
1223 *
1224 * Because we not have RCU_FAST_NO_HZ, just check whether or not this
1225 * CPU has RCU callbacks queued.
1226 */
1227int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1228{
1229 *nextevt = KTIME_MAX;
1230 return !rcu_segcblist_empty(&this_cpu_ptr(&rcu_data)->cblist) &&
1231 !rcu_segcblist_is_offloaded(&this_cpu_ptr(&rcu_data)->cblist);
1232}
1233
1234/*
1235 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1236 * after it.
1237 */
1238static void rcu_cleanup_after_idle(void)
1239{
1240}
1241
1242/*
1243 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1244 * is nothing.
1245 */
1246static void rcu_prepare_for_idle(void)
1247{
1248}
1249
1250#else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1251
1252/*
1253 * This code is invoked when a CPU goes idle, at which point we want
1254 * to have the CPU do everything required for RCU so that it can enter
1255 * the energy-efficient dyntick-idle mode.
1256 *
1257 * The following preprocessor symbol controls this:
1258 *
1259 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1260 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1261 * is sized to be roughly one RCU grace period. Those energy-efficiency
1262 * benchmarkers who might otherwise be tempted to set this to a large
1263 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1264 * system. And if you are -that- concerned about energy efficiency,
1265 * just power the system down and be done with it!
1266 *
1267 * The value below works well in practice. If future workloads require
1268 * adjustment, they can be converted into kernel config parameters, though
1269 * making the state machine smarter might be a better option.
1270 */
1271#define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1272
1273static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1274module_param(rcu_idle_gp_delay, int, 0644);
1275
1276/*
1277 * Try to advance callbacks on the current CPU, but only if it has been
1278 * awhile since the last time we did so. Afterwards, if there are any
1279 * callbacks ready for immediate invocation, return true.
1280 */
1281static bool __maybe_unused rcu_try_advance_all_cbs(void)
1282{
1283 bool cbs_ready = false;
1284 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1285 struct rcu_node *rnp;
1286
1287 /* Exit early if we advanced recently. */
1288 if (jiffies == rdp->last_advance_all)
1289 return false;
1290 rdp->last_advance_all = jiffies;
1291
1292 rnp = rdp->mynode;
1293
1294 /*
1295 * Don't bother checking unless a grace period has
1296 * completed since we last checked and there are
1297 * callbacks not yet ready to invoke.
1298 */
1299 if ((rcu_seq_completed_gp(rdp->gp_seq,
1300 rcu_seq_current(&rnp->gp_seq)) ||
1301 unlikely(READ_ONCE(rdp->gpwrap))) &&
1302 rcu_segcblist_pend_cbs(&rdp->cblist))
1303 note_gp_changes(rdp);
1304
1305 if (rcu_segcblist_ready_cbs(&rdp->cblist))
1306 cbs_ready = true;
1307 return cbs_ready;
1308}
1309
1310/*
1311 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1312 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1313 * caller about what to set the timeout.
1314 *
1315 * The caller must have disabled interrupts.
1316 */
1317int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1318{
1319 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1320 unsigned long dj;
1321
1322 lockdep_assert_irqs_disabled();
1323
1324 /* If no non-offloaded callbacks, RCU doesn't need the CPU. */
1325 if (rcu_segcblist_empty(&rdp->cblist) ||
1326 rcu_segcblist_is_offloaded(&this_cpu_ptr(&rcu_data)->cblist)) {
1327 *nextevt = KTIME_MAX;
1328 return 0;
1329 }
1330
1331 /* Attempt to advance callbacks. */
1332 if (rcu_try_advance_all_cbs()) {
1333 /* Some ready to invoke, so initiate later invocation. */
1334 invoke_rcu_core();
1335 return 1;
1336 }
1337 rdp->last_accelerate = jiffies;
1338
1339 /* Request timer and round. */
1340 dj = round_up(rcu_idle_gp_delay + jiffies, rcu_idle_gp_delay) - jiffies;
1341
1342 *nextevt = basemono + dj * TICK_NSEC;
1343 return 0;
1344}
1345
1346/*
1347 * Prepare a CPU for idle from an RCU perspective. The first major task is to
1348 * sense whether nohz mode has been enabled or disabled via sysfs. The second
1349 * major task is to accelerate (that is, assign grace-period numbers to) any
1350 * recently arrived callbacks.
1351 *
1352 * The caller must have disabled interrupts.
1353 */
1354static void rcu_prepare_for_idle(void)
1355{
1356 bool needwake;
1357 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1358 struct rcu_node *rnp;
1359 int tne;
1360
1361 lockdep_assert_irqs_disabled();
1362 if (rcu_segcblist_is_offloaded(&rdp->cblist))
1363 return;
1364
1365 /* Handle nohz enablement switches conservatively. */
1366 tne = READ_ONCE(tick_nohz_active);
1367 if (tne != rdp->tick_nohz_enabled_snap) {
1368 if (!rcu_segcblist_empty(&rdp->cblist))
1369 invoke_rcu_core(); /* force nohz to see update. */
1370 rdp->tick_nohz_enabled_snap = tne;
1371 return;
1372 }
1373 if (!tne)
1374 return;
1375
1376 /*
1377 * If we have not yet accelerated this jiffy, accelerate all
1378 * callbacks on this CPU.
1379 */
1380 if (rdp->last_accelerate == jiffies)
1381 return;
1382 rdp->last_accelerate = jiffies;
1383 if (rcu_segcblist_pend_cbs(&rdp->cblist)) {
1384 rnp = rdp->mynode;
1385 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1386 needwake = rcu_accelerate_cbs(rnp, rdp);
1387 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1388 if (needwake)
1389 rcu_gp_kthread_wake();
1390 }
1391}
1392
1393/*
1394 * Clean up for exit from idle. Attempt to advance callbacks based on
1395 * any grace periods that elapsed while the CPU was idle, and if any
1396 * callbacks are now ready to invoke, initiate invocation.
1397 */
1398static void rcu_cleanup_after_idle(void)
1399{
1400 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1401
1402 lockdep_assert_irqs_disabled();
1403 if (rcu_segcblist_is_offloaded(&rdp->cblist))
1404 return;
1405 if (rcu_try_advance_all_cbs())
1406 invoke_rcu_core();
1407}
1408
1409#endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1410
1411#ifdef CONFIG_RCU_NOCB_CPU
1412
1413/*
1414 * Offload callback processing from the boot-time-specified set of CPUs
1415 * specified by rcu_nocb_mask. For the CPUs in the set, there are kthreads
1416 * created that pull the callbacks from the corresponding CPU, wait for
1417 * a grace period to elapse, and invoke the callbacks. These kthreads
1418 * are organized into GP kthreads, which manage incoming callbacks, wait for
1419 * grace periods, and awaken CB kthreads, and the CB kthreads, which only
1420 * invoke callbacks. Each GP kthread invokes its own CBs. The no-CBs CPUs
1421 * do a wake_up() on their GP kthread when they insert a callback into any
1422 * empty list, unless the rcu_nocb_poll boot parameter has been specified,
1423 * in which case each kthread actively polls its CPU. (Which isn't so great
1424 * for energy efficiency, but which does reduce RCU's overhead on that CPU.)
1425 *
1426 * This is intended to be used in conjunction with Frederic Weisbecker's
1427 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1428 * running CPU-bound user-mode computations.
1429 *
1430 * Offloading of callbacks can also be used as an energy-efficiency
1431 * measure because CPUs with no RCU callbacks queued are more aggressive
1432 * about entering dyntick-idle mode.
1433 */
1434
1435
1436/*
1437 * Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters.
1438 * The string after the "rcu_nocbs=" is either "all" for all CPUs, or a
1439 * comma-separated list of CPUs and/or CPU ranges. If an invalid list is
1440 * given, a warning is emitted and all CPUs are offloaded.
1441 */
1442static int __init rcu_nocb_setup(char *str)
1443{
1444 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1445 if (!strcasecmp(str, "all"))
1446 cpumask_setall(rcu_nocb_mask);
1447 else
1448 if (cpulist_parse(str, rcu_nocb_mask)) {
1449 pr_warn("rcu_nocbs= bad CPU range, all CPUs set\n");
1450 cpumask_setall(rcu_nocb_mask);
1451 }
1452 return 1;
1453}
1454__setup("rcu_nocbs=", rcu_nocb_setup);
1455
1456static int __init parse_rcu_nocb_poll(char *arg)
1457{
1458 rcu_nocb_poll = true;
1459 return 0;
1460}
1461early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1462
1463/*
1464 * Don't bother bypassing ->cblist if the call_rcu() rate is low.
1465 * After all, the main point of bypassing is to avoid lock contention
1466 * on ->nocb_lock, which only can happen at high call_rcu() rates.
1467 */
1468int nocb_nobypass_lim_per_jiffy = 16 * 1000 / HZ;
1469module_param(nocb_nobypass_lim_per_jiffy, int, 0);
1470
1471/*
1472 * Acquire the specified rcu_data structure's ->nocb_bypass_lock. If the
1473 * lock isn't immediately available, increment ->nocb_lock_contended to
1474 * flag the contention.
1475 */
1476static void rcu_nocb_bypass_lock(struct rcu_data *rdp)
1477 __acquires(&rdp->nocb_bypass_lock)
1478{
1479 lockdep_assert_irqs_disabled();
1480 if (raw_spin_trylock(&rdp->nocb_bypass_lock))
1481 return;
1482 atomic_inc(&rdp->nocb_lock_contended);
1483 WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
1484 smp_mb__after_atomic(); /* atomic_inc() before lock. */
1485 raw_spin_lock(&rdp->nocb_bypass_lock);
1486 smp_mb__before_atomic(); /* atomic_dec() after lock. */
1487 atomic_dec(&rdp->nocb_lock_contended);
1488}
1489
1490/*
1491 * Spinwait until the specified rcu_data structure's ->nocb_lock is
1492 * not contended. Please note that this is extremely special-purpose,
1493 * relying on the fact that at most two kthreads and one CPU contend for
1494 * this lock, and also that the two kthreads are guaranteed to have frequent
1495 * grace-period-duration time intervals between successive acquisitions
1496 * of the lock. This allows us to use an extremely simple throttling
1497 * mechanism, and further to apply it only to the CPU doing floods of
1498 * call_rcu() invocations. Don't try this at home!
1499 */
1500static void rcu_nocb_wait_contended(struct rcu_data *rdp)
1501{
1502 WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
1503 while (WARN_ON_ONCE(atomic_read(&rdp->nocb_lock_contended)))
1504 cpu_relax();
1505}
1506
1507/*
1508 * Conditionally acquire the specified rcu_data structure's
1509 * ->nocb_bypass_lock.
1510 */
1511static bool rcu_nocb_bypass_trylock(struct rcu_data *rdp)
1512{
1513 lockdep_assert_irqs_disabled();
1514 return raw_spin_trylock(&rdp->nocb_bypass_lock);
1515}
1516
1517/*
1518 * Release the specified rcu_data structure's ->nocb_bypass_lock.
1519 */
1520static void rcu_nocb_bypass_unlock(struct rcu_data *rdp)
1521 __releases(&rdp->nocb_bypass_lock)
1522{
1523 lockdep_assert_irqs_disabled();
1524 raw_spin_unlock(&rdp->nocb_bypass_lock);
1525}
1526
1527/*
1528 * Acquire the specified rcu_data structure's ->nocb_lock, but only
1529 * if it corresponds to a no-CBs CPU.
1530 */
1531static void rcu_nocb_lock(struct rcu_data *rdp)
1532{
1533 lockdep_assert_irqs_disabled();
1534 if (!rcu_segcblist_is_offloaded(&rdp->cblist))
1535 return;
1536 raw_spin_lock(&rdp->nocb_lock);
1537}
1538
1539/*
1540 * Release the specified rcu_data structure's ->nocb_lock, but only
1541 * if it corresponds to a no-CBs CPU.
1542 */
1543static void rcu_nocb_unlock(struct rcu_data *rdp)
1544{
1545 if (rcu_segcblist_is_offloaded(&rdp->cblist)) {
1546 lockdep_assert_irqs_disabled();
1547 raw_spin_unlock(&rdp->nocb_lock);
1548 }
1549}
1550
1551/*
1552 * Release the specified rcu_data structure's ->nocb_lock and restore
1553 * interrupts, but only if it corresponds to a no-CBs CPU.
1554 */
1555static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
1556 unsigned long flags)
1557{
1558 if (rcu_segcblist_is_offloaded(&rdp->cblist)) {
1559 lockdep_assert_irqs_disabled();
1560 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1561 } else {
1562 local_irq_restore(flags);
1563 }
1564}
1565
1566/* Lockdep check that ->cblist may be safely accessed. */
1567static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
1568{
1569 lockdep_assert_irqs_disabled();
1570 if (rcu_segcblist_is_offloaded(&rdp->cblist))
1571 lockdep_assert_held(&rdp->nocb_lock);
1572}
1573
1574/*
1575 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1576 * grace period.
1577 */
1578static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
1579{
1580 swake_up_all(sq);
1581}
1582
1583static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
1584{
1585 return &rnp->nocb_gp_wq[rcu_seq_ctr(rnp->gp_seq) & 0x1];
1586}
1587
1588static void rcu_init_one_nocb(struct rcu_node *rnp)
1589{
1590 init_swait_queue_head(&rnp->nocb_gp_wq[0]);
1591 init_swait_queue_head(&rnp->nocb_gp_wq[1]);
1592}
1593
1594/* Is the specified CPU a no-CBs CPU? */
1595bool rcu_is_nocb_cpu(int cpu)
1596{
1597 if (cpumask_available(rcu_nocb_mask))
1598 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1599 return false;
1600}
1601
1602/*
1603 * Kick the GP kthread for this NOCB group. Caller holds ->nocb_lock
1604 * and this function releases it.
1605 */
1606static void wake_nocb_gp(struct rcu_data *rdp, bool force,
1607 unsigned long flags)
1608 __releases(rdp->nocb_lock)
1609{
1610 bool needwake = false;
1611 struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
1612
1613 lockdep_assert_held(&rdp->nocb_lock);
1614 if (!READ_ONCE(rdp_gp->nocb_gp_kthread)) {
1615 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1616 TPS("AlreadyAwake"));
1617 rcu_nocb_unlock_irqrestore(rdp, flags);
1618 return;
1619 }
1620 del_timer(&rdp->nocb_timer);
1621 rcu_nocb_unlock_irqrestore(rdp, flags);
1622 raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
1623 if (force || READ_ONCE(rdp_gp->nocb_gp_sleep)) {
1624 WRITE_ONCE(rdp_gp->nocb_gp_sleep, false);
1625 needwake = true;
1626 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DoWake"));
1627 }
1628 raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
1629 if (needwake)
1630 wake_up_process(rdp_gp->nocb_gp_kthread);
1631}
1632
1633/*
1634 * Arrange to wake the GP kthread for this NOCB group at some future
1635 * time when it is safe to do so.
1636 */
1637static void wake_nocb_gp_defer(struct rcu_data *rdp, int waketype,
1638 const char *reason)
1639{
1640 if (rdp->nocb_defer_wakeup == RCU_NOCB_WAKE_NOT)
1641 mod_timer(&rdp->nocb_timer, jiffies + 1);
1642 if (rdp->nocb_defer_wakeup < waketype)
1643 WRITE_ONCE(rdp->nocb_defer_wakeup, waketype);
1644 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, reason);
1645}
1646
1647/*
1648 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1649 * However, if there is a callback to be enqueued and if ->nocb_bypass
1650 * proves to be initially empty, just return false because the no-CB GP
1651 * kthread may need to be awakened in this case.
1652 *
1653 * Note that this function always returns true if rhp is NULL.
1654 */
1655static bool rcu_nocb_do_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1656 unsigned long j)
1657{
1658 struct rcu_cblist rcl;
1659
1660 WARN_ON_ONCE(!rcu_segcblist_is_offloaded(&rdp->cblist));
1661 rcu_lockdep_assert_cblist_protected(rdp);
1662 lockdep_assert_held(&rdp->nocb_bypass_lock);
1663 if (rhp && !rcu_cblist_n_cbs(&rdp->nocb_bypass)) {
1664 raw_spin_unlock(&rdp->nocb_bypass_lock);
1665 return false;
1666 }
1667 /* Note: ->cblist.len already accounts for ->nocb_bypass contents. */
1668 if (rhp)
1669 rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
1670 rcu_cblist_flush_enqueue(&rcl, &rdp->nocb_bypass, rhp);
1671 rcu_segcblist_insert_pend_cbs(&rdp->cblist, &rcl);
1672 WRITE_ONCE(rdp->nocb_bypass_first, j);
1673 rcu_nocb_bypass_unlock(rdp);
1674 return true;
1675}
1676
1677/*
1678 * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
1679 * However, if there is a callback to be enqueued and if ->nocb_bypass
1680 * proves to be initially empty, just return false because the no-CB GP
1681 * kthread may need to be awakened in this case.
1682 *
1683 * Note that this function always returns true if rhp is NULL.
1684 */
1685static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1686 unsigned long j)
1687{
1688 if (!rcu_segcblist_is_offloaded(&rdp->cblist))
1689 return true;
1690 rcu_lockdep_assert_cblist_protected(rdp);
1691 rcu_nocb_bypass_lock(rdp);
1692 return rcu_nocb_do_flush_bypass(rdp, rhp, j);
1693}
1694
1695/*
1696 * If the ->nocb_bypass_lock is immediately available, flush the
1697 * ->nocb_bypass queue into ->cblist.
1698 */
1699static void rcu_nocb_try_flush_bypass(struct rcu_data *rdp, unsigned long j)
1700{
1701 rcu_lockdep_assert_cblist_protected(rdp);
1702 if (!rcu_segcblist_is_offloaded(&rdp->cblist) ||
1703 !rcu_nocb_bypass_trylock(rdp))
1704 return;
1705 WARN_ON_ONCE(!rcu_nocb_do_flush_bypass(rdp, NULL, j));
1706}
1707
1708/*
1709 * See whether it is appropriate to use the ->nocb_bypass list in order
1710 * to control contention on ->nocb_lock. A limited number of direct
1711 * enqueues are permitted into ->cblist per jiffy. If ->nocb_bypass
1712 * is non-empty, further callbacks must be placed into ->nocb_bypass,
1713 * otherwise rcu_barrier() breaks. Use rcu_nocb_flush_bypass() to switch
1714 * back to direct use of ->cblist. However, ->nocb_bypass should not be
1715 * used if ->cblist is empty, because otherwise callbacks can be stranded
1716 * on ->nocb_bypass because we cannot count on the current CPU ever again
1717 * invoking call_rcu(). The general rule is that if ->nocb_bypass is
1718 * non-empty, the corresponding no-CBs grace-period kthread must not be
1719 * in an indefinite sleep state.
1720 *
1721 * Finally, it is not permitted to use the bypass during early boot,
1722 * as doing so would confuse the auto-initialization code. Besides
1723 * which, there is no point in worrying about lock contention while
1724 * there is only one CPU in operation.
1725 */
1726static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1727 bool *was_alldone, unsigned long flags)
1728{
1729 unsigned long c;
1730 unsigned long cur_gp_seq;
1731 unsigned long j = jiffies;
1732 long ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1733
1734 if (!rcu_segcblist_is_offloaded(&rdp->cblist)) {
1735 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1736 return false; /* Not offloaded, no bypassing. */
1737 }
1738 lockdep_assert_irqs_disabled();
1739
1740 // Don't use ->nocb_bypass during early boot.
1741 if (rcu_scheduler_active != RCU_SCHEDULER_RUNNING) {
1742 rcu_nocb_lock(rdp);
1743 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1744 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1745 return false;
1746 }
1747
1748 // If we have advanced to a new jiffy, reset counts to allow
1749 // moving back from ->nocb_bypass to ->cblist.
1750 if (j == rdp->nocb_nobypass_last) {
1751 c = rdp->nocb_nobypass_count + 1;
1752 } else {
1753 WRITE_ONCE(rdp->nocb_nobypass_last, j);
1754 c = rdp->nocb_nobypass_count - nocb_nobypass_lim_per_jiffy;
1755 if (ULONG_CMP_LT(rdp->nocb_nobypass_count,
1756 nocb_nobypass_lim_per_jiffy))
1757 c = 0;
1758 else if (c > nocb_nobypass_lim_per_jiffy)
1759 c = nocb_nobypass_lim_per_jiffy;
1760 }
1761 WRITE_ONCE(rdp->nocb_nobypass_count, c);
1762
1763 // If there hasn't yet been all that many ->cblist enqueues
1764 // this jiffy, tell the caller to enqueue onto ->cblist. But flush
1765 // ->nocb_bypass first.
1766 if (rdp->nocb_nobypass_count < nocb_nobypass_lim_per_jiffy) {
1767 rcu_nocb_lock(rdp);
1768 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1769 if (*was_alldone)
1770 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1771 TPS("FirstQ"));
1772 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, j));
1773 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1774 return false; // Caller must enqueue the callback.
1775 }
1776
1777 // If ->nocb_bypass has been used too long or is too full,
1778 // flush ->nocb_bypass to ->cblist.
1779 if ((ncbs && j != READ_ONCE(rdp->nocb_bypass_first)) ||
1780 ncbs >= qhimark) {
1781 rcu_nocb_lock(rdp);
1782 if (!rcu_nocb_flush_bypass(rdp, rhp, j)) {
1783 *was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
1784 if (*was_alldone)
1785 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1786 TPS("FirstQ"));
1787 WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1788 return false; // Caller must enqueue the callback.
1789 }
1790 if (j != rdp->nocb_gp_adv_time &&
1791 rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1792 rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
1793 rcu_advance_cbs_nowake(rdp->mynode, rdp);
1794 rdp->nocb_gp_adv_time = j;
1795 }
1796 rcu_nocb_unlock_irqrestore(rdp, flags);
1797 return true; // Callback already enqueued.
1798 }
1799
1800 // We need to use the bypass.
1801 rcu_nocb_wait_contended(rdp);
1802 rcu_nocb_bypass_lock(rdp);
1803 ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1804 rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
1805 rcu_cblist_enqueue(&rdp->nocb_bypass, rhp);
1806 if (!ncbs) {
1807 WRITE_ONCE(rdp->nocb_bypass_first, j);
1808 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("FirstBQ"));
1809 }
1810 rcu_nocb_bypass_unlock(rdp);
1811 smp_mb(); /* Order enqueue before wake. */
1812 if (ncbs) {
1813 local_irq_restore(flags);
1814 } else {
1815 // No-CBs GP kthread might be indefinitely asleep, if so, wake.
1816 rcu_nocb_lock(rdp); // Rare during call_rcu() flood.
1817 if (!rcu_segcblist_pend_cbs(&rdp->cblist)) {
1818 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1819 TPS("FirstBQwake"));
1820 __call_rcu_nocb_wake(rdp, true, flags);
1821 } else {
1822 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1823 TPS("FirstBQnoWake"));
1824 rcu_nocb_unlock_irqrestore(rdp, flags);
1825 }
1826 }
1827 return true; // Callback already enqueued.
1828}
1829
1830/*
1831 * Awaken the no-CBs grace-period kthead if needed, either due to it
1832 * legitimately being asleep or due to overload conditions.
1833 *
1834 * If warranted, also wake up the kthread servicing this CPUs queues.
1835 */
1836static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_alldone,
1837 unsigned long flags)
1838 __releases(rdp->nocb_lock)
1839{
1840 unsigned long cur_gp_seq;
1841 unsigned long j;
1842 long len;
1843 struct task_struct *t;
1844
1845 // If we are being polled or there is no kthread, just leave.
1846 t = READ_ONCE(rdp->nocb_gp_kthread);
1847 if (rcu_nocb_poll || !t) {
1848 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1849 TPS("WakeNotPoll"));
1850 rcu_nocb_unlock_irqrestore(rdp, flags);
1851 return;
1852 }
1853 // Need to actually to a wakeup.
1854 len = rcu_segcblist_n_cbs(&rdp->cblist);
1855 if (was_alldone) {
1856 rdp->qlen_last_fqs_check = len;
1857 if (!irqs_disabled_flags(flags)) {
1858 /* ... if queue was empty ... */
1859 wake_nocb_gp(rdp, false, flags);
1860 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1861 TPS("WakeEmpty"));
1862 } else {
1863 wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE,
1864 TPS("WakeEmptyIsDeferred"));
1865 rcu_nocb_unlock_irqrestore(rdp, flags);
1866 }
1867 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
1868 /* ... or if many callbacks queued. */
1869 rdp->qlen_last_fqs_check = len;
1870 j = jiffies;
1871 if (j != rdp->nocb_gp_adv_time &&
1872 rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1873 rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
1874 rcu_advance_cbs_nowake(rdp->mynode, rdp);
1875 rdp->nocb_gp_adv_time = j;
1876 }
1877 smp_mb(); /* Enqueue before timer_pending(). */
1878 if ((rdp->nocb_cb_sleep ||
1879 !rcu_segcblist_ready_cbs(&rdp->cblist)) &&
1880 !timer_pending(&rdp->nocb_bypass_timer))
1881 wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE_FORCE,
1882 TPS("WakeOvfIsDeferred"));
1883 rcu_nocb_unlock_irqrestore(rdp, flags);
1884 } else {
1885 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WakeNot"));
1886 rcu_nocb_unlock_irqrestore(rdp, flags);
1887 }
1888 return;
1889}
1890
1891/* Wake up the no-CBs GP kthread to flush ->nocb_bypass. */
1892static void do_nocb_bypass_wakeup_timer(struct timer_list *t)
1893{
1894 unsigned long flags;
1895 struct rcu_data *rdp = from_timer(rdp, t, nocb_bypass_timer);
1896
1897 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Timer"));
1898 rcu_nocb_lock_irqsave(rdp, flags);
1899 smp_mb__after_spinlock(); /* Timer expire before wakeup. */
1900 __call_rcu_nocb_wake(rdp, true, flags);
1901}
1902
1903/*
1904 * No-CBs GP kthreads come here to wait for additional callbacks to show up
1905 * or for grace periods to end.
1906 */
1907static void nocb_gp_wait(struct rcu_data *my_rdp)
1908{
1909 bool bypass = false;
1910 long bypass_ncbs;
1911 int __maybe_unused cpu = my_rdp->cpu;
1912 unsigned long cur_gp_seq;
1913 unsigned long flags;
1914 bool gotcbs = false;
1915 unsigned long j = jiffies;
1916 bool needwait_gp = false; // This prevents actual uninitialized use.
1917 bool needwake;
1918 bool needwake_gp;
1919 struct rcu_data *rdp;
1920 struct rcu_node *rnp;
1921 unsigned long wait_gp_seq = 0; // Suppress "use uninitialized" warning.
1922 bool wasempty = false;
1923
1924 /*
1925 * Each pass through the following loop checks for CBs and for the
1926 * nearest grace period (if any) to wait for next. The CB kthreads
1927 * and the global grace-period kthread are awakened if needed.
1928 */
1929 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_cb_rdp) {
1930 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Check"));
1931 rcu_nocb_lock_irqsave(rdp, flags);
1932 bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1933 if (bypass_ncbs &&
1934 (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + 1) ||
1935 bypass_ncbs > 2 * qhimark)) {
1936 // Bypass full or old, so flush it.
1937 (void)rcu_nocb_try_flush_bypass(rdp, j);
1938 bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
1939 } else if (!bypass_ncbs && rcu_segcblist_empty(&rdp->cblist)) {
1940 rcu_nocb_unlock_irqrestore(rdp, flags);
1941 continue; /* No callbacks here, try next. */
1942 }
1943 if (bypass_ncbs) {
1944 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1945 TPS("Bypass"));
1946 bypass = true;
1947 }
1948 rnp = rdp->mynode;
1949 if (bypass) { // Avoid race with first bypass CB.
1950 WRITE_ONCE(my_rdp->nocb_defer_wakeup,
1951 RCU_NOCB_WAKE_NOT);
1952 del_timer(&my_rdp->nocb_timer);
1953 }
1954 // Advance callbacks if helpful and low contention.
1955 needwake_gp = false;
1956 if (!rcu_segcblist_restempty(&rdp->cblist,
1957 RCU_NEXT_READY_TAIL) ||
1958 (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
1959 rcu_seq_done(&rnp->gp_seq, cur_gp_seq))) {
1960 raw_spin_lock_rcu_node(rnp); /* irqs disabled. */
1961 needwake_gp = rcu_advance_cbs(rnp, rdp);
1962 wasempty = rcu_segcblist_restempty(&rdp->cblist,
1963 RCU_NEXT_READY_TAIL);
1964 raw_spin_unlock_rcu_node(rnp); /* irqs disabled. */
1965 }
1966 // Need to wait on some grace period?
1967 WARN_ON_ONCE(wasempty &&
1968 !rcu_segcblist_restempty(&rdp->cblist,
1969 RCU_NEXT_READY_TAIL));
1970 if (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq)) {
1971 if (!needwait_gp ||
1972 ULONG_CMP_LT(cur_gp_seq, wait_gp_seq))
1973 wait_gp_seq = cur_gp_seq;
1974 needwait_gp = true;
1975 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
1976 TPS("NeedWaitGP"));
1977 }
1978 if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
1979 needwake = rdp->nocb_cb_sleep;
1980 WRITE_ONCE(rdp->nocb_cb_sleep, false);
1981 smp_mb(); /* CB invocation -after- GP end. */
1982 } else {
1983 needwake = false;
1984 }
1985 rcu_nocb_unlock_irqrestore(rdp, flags);
1986 if (needwake) {
1987 swake_up_one(&rdp->nocb_cb_wq);
1988 gotcbs = true;
1989 }
1990 if (needwake_gp)
1991 rcu_gp_kthread_wake();
1992 }
1993
1994 my_rdp->nocb_gp_bypass = bypass;
1995 my_rdp->nocb_gp_gp = needwait_gp;
1996 my_rdp->nocb_gp_seq = needwait_gp ? wait_gp_seq : 0;
1997 if (bypass && !rcu_nocb_poll) {
1998 // At least one child with non-empty ->nocb_bypass, so set
1999 // timer in order to avoid stranding its callbacks.
2000 raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
2001 mod_timer(&my_rdp->nocb_bypass_timer, j + 2);
2002 raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
2003 }
2004 if (rcu_nocb_poll) {
2005 /* Polling, so trace if first poll in the series. */
2006 if (gotcbs)
2007 trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Poll"));
2008 schedule_timeout_idle(1);
2009 } else if (!needwait_gp) {
2010 /* Wait for callbacks to appear. */
2011 trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Sleep"));
2012 swait_event_interruptible_exclusive(my_rdp->nocb_gp_wq,
2013 !READ_ONCE(my_rdp->nocb_gp_sleep));
2014 trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("EndSleep"));
2015 } else {
2016 rnp = my_rdp->mynode;
2017 trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("StartWait"));
2018 swait_event_interruptible_exclusive(
2019 rnp->nocb_gp_wq[rcu_seq_ctr(wait_gp_seq) & 0x1],
2020 rcu_seq_done(&rnp->gp_seq, wait_gp_seq) ||
2021 !READ_ONCE(my_rdp->nocb_gp_sleep));
2022 trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("EndWait"));
2023 }
2024 if (!rcu_nocb_poll) {
2025 raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
2026 if (bypass)
2027 del_timer(&my_rdp->nocb_bypass_timer);
2028 WRITE_ONCE(my_rdp->nocb_gp_sleep, true);
2029 raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
2030 }
2031 my_rdp->nocb_gp_seq = -1;
2032 WARN_ON(signal_pending(current));
2033}
2034
2035/*
2036 * No-CBs grace-period-wait kthread. There is one of these per group
2037 * of CPUs, but only once at least one CPU in that group has come online
2038 * at least once since boot. This kthread checks for newly posted
2039 * callbacks from any of the CPUs it is responsible for, waits for a
2040 * grace period, then awakens all of the rcu_nocb_cb_kthread() instances
2041 * that then have callback-invocation work to do.
2042 */
2043static int rcu_nocb_gp_kthread(void *arg)
2044{
2045 struct rcu_data *rdp = arg;
2046
2047 for (;;) {
2048 WRITE_ONCE(rdp->nocb_gp_loops, rdp->nocb_gp_loops + 1);
2049 nocb_gp_wait(rdp);
2050 cond_resched_tasks_rcu_qs();
2051 }
2052 return 0;
2053}
2054
2055/*
2056 * Invoke any ready callbacks from the corresponding no-CBs CPU,
2057 * then, if there are no more, wait for more to appear.
2058 */
2059static void nocb_cb_wait(struct rcu_data *rdp)
2060{
2061 unsigned long cur_gp_seq;
2062 unsigned long flags;
2063 bool needwake_gp = false;
2064 struct rcu_node *rnp = rdp->mynode;
2065
2066 local_irq_save(flags);
2067 rcu_momentary_dyntick_idle();
2068 local_irq_restore(flags);
2069 local_bh_disable();
2070 rcu_do_batch(rdp);
2071 local_bh_enable();
2072 lockdep_assert_irqs_enabled();
2073 rcu_nocb_lock_irqsave(rdp, flags);
2074 if (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
2075 rcu_seq_done(&rnp->gp_seq, cur_gp_seq) &&
2076 raw_spin_trylock_rcu_node(rnp)) { /* irqs already disabled. */
2077 needwake_gp = rcu_advance_cbs(rdp->mynode, rdp);
2078 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2079 }
2080 if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
2081 rcu_nocb_unlock_irqrestore(rdp, flags);
2082 if (needwake_gp)
2083 rcu_gp_kthread_wake();
2084 return;
2085 }
2086
2087 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("CBSleep"));
2088 WRITE_ONCE(rdp->nocb_cb_sleep, true);
2089 rcu_nocb_unlock_irqrestore(rdp, flags);
2090 if (needwake_gp)
2091 rcu_gp_kthread_wake();
2092 swait_event_interruptible_exclusive(rdp->nocb_cb_wq,
2093 !READ_ONCE(rdp->nocb_cb_sleep));
2094 if (!smp_load_acquire(&rdp->nocb_cb_sleep)) { /* VVV */
2095 /* ^^^ Ensure CB invocation follows _sleep test. */
2096 return;
2097 }
2098 WARN_ON(signal_pending(current));
2099 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WokeEmpty"));
2100}
2101
2102/*
2103 * Per-rcu_data kthread, but only for no-CBs CPUs. Repeatedly invoke
2104 * nocb_cb_wait() to do the dirty work.
2105 */
2106static int rcu_nocb_cb_kthread(void *arg)
2107{
2108 struct rcu_data *rdp = arg;
2109
2110 // Each pass through this loop does one callback batch, and,
2111 // if there are no more ready callbacks, waits for them.
2112 for (;;) {
2113 nocb_cb_wait(rdp);
2114 cond_resched_tasks_rcu_qs();
2115 }
2116 return 0;
2117}
2118
2119/* Is a deferred wakeup of rcu_nocb_kthread() required? */
2120static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2121{
2122 return READ_ONCE(rdp->nocb_defer_wakeup);
2123}
2124
2125/* Do a deferred wakeup of rcu_nocb_kthread(). */
2126static void do_nocb_deferred_wakeup_common(struct rcu_data *rdp)
2127{
2128 unsigned long flags;
2129 int ndw;
2130
2131 rcu_nocb_lock_irqsave(rdp, flags);
2132 if (!rcu_nocb_need_deferred_wakeup(rdp)) {
2133 rcu_nocb_unlock_irqrestore(rdp, flags);
2134 return;
2135 }
2136 ndw = READ_ONCE(rdp->nocb_defer_wakeup);
2137 WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
2138 wake_nocb_gp(rdp, ndw == RCU_NOCB_WAKE_FORCE, flags);
2139 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DeferredWake"));
2140}
2141
2142/* Do a deferred wakeup of rcu_nocb_kthread() from a timer handler. */
2143static void do_nocb_deferred_wakeup_timer(struct timer_list *t)
2144{
2145 struct rcu_data *rdp = from_timer(rdp, t, nocb_timer);
2146
2147 do_nocb_deferred_wakeup_common(rdp);
2148}
2149
2150/*
2151 * Do a deferred wakeup of rcu_nocb_kthread() from fastpath.
2152 * This means we do an inexact common-case check. Note that if
2153 * we miss, ->nocb_timer will eventually clean things up.
2154 */
2155static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2156{
2157 if (rcu_nocb_need_deferred_wakeup(rdp))
2158 do_nocb_deferred_wakeup_common(rdp);
2159}
2160
2161void __init rcu_init_nohz(void)
2162{
2163 int cpu;
2164 bool need_rcu_nocb_mask = false;
2165 struct rcu_data *rdp;
2166
2167#if defined(CONFIG_NO_HZ_FULL)
2168 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2169 need_rcu_nocb_mask = true;
2170#endif /* #if defined(CONFIG_NO_HZ_FULL) */
2171
2172 if (!cpumask_available(rcu_nocb_mask) && need_rcu_nocb_mask) {
2173 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2174 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2175 return;
2176 }
2177 }
2178 if (!cpumask_available(rcu_nocb_mask))
2179 return;
2180
2181#if defined(CONFIG_NO_HZ_FULL)
2182 if (tick_nohz_full_running)
2183 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2184#endif /* #if defined(CONFIG_NO_HZ_FULL) */
2185
2186 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2187 pr_info("\tNote: kernel parameter 'rcu_nocbs=', 'nohz_full', or 'isolcpus=' contains nonexistent CPUs.\n");
2188 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2189 rcu_nocb_mask);
2190 }
2191 if (cpumask_empty(rcu_nocb_mask))
2192 pr_info("\tOffload RCU callbacks from CPUs: (none).\n");
2193 else
2194 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2195 cpumask_pr_args(rcu_nocb_mask));
2196 if (rcu_nocb_poll)
2197 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2198
2199 for_each_cpu(cpu, rcu_nocb_mask) {
2200 rdp = per_cpu_ptr(&rcu_data, cpu);
2201 if (rcu_segcblist_empty(&rdp->cblist))
2202 rcu_segcblist_init(&rdp->cblist);
2203 rcu_segcblist_offload(&rdp->cblist);
2204 }
2205 rcu_organize_nocb_kthreads();
2206}
2207
2208/* Initialize per-rcu_data variables for no-CBs CPUs. */
2209static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2210{
2211 init_swait_queue_head(&rdp->nocb_cb_wq);
2212 init_swait_queue_head(&rdp->nocb_gp_wq);
2213 raw_spin_lock_init(&rdp->nocb_lock);
2214 raw_spin_lock_init(&rdp->nocb_bypass_lock);
2215 raw_spin_lock_init(&rdp->nocb_gp_lock);
2216 timer_setup(&rdp->nocb_timer, do_nocb_deferred_wakeup_timer, 0);
2217 timer_setup(&rdp->nocb_bypass_timer, do_nocb_bypass_wakeup_timer, 0);
2218 rcu_cblist_init(&rdp->nocb_bypass);
2219}
2220
2221/*
2222 * If the specified CPU is a no-CBs CPU that does not already have its
2223 * rcuo CB kthread, spawn it. Additionally, if the rcuo GP kthread
2224 * for this CPU's group has not yet been created, spawn it as well.
2225 */
2226static void rcu_spawn_one_nocb_kthread(int cpu)
2227{
2228 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2229 struct rcu_data *rdp_gp;
2230 struct task_struct *t;
2231
2232 /*
2233 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2234 * then nothing to do.
2235 */
2236 if (!rcu_is_nocb_cpu(cpu) || rdp->nocb_cb_kthread)
2237 return;
2238
2239 /* If we didn't spawn the GP kthread first, reorganize! */
2240 rdp_gp = rdp->nocb_gp_rdp;
2241 if (!rdp_gp->nocb_gp_kthread) {
2242 t = kthread_run(rcu_nocb_gp_kthread, rdp_gp,
2243 "rcuog/%d", rdp_gp->cpu);
2244 if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo GP kthread, OOM is now expected behavior\n", __func__))
2245 return;
2246 WRITE_ONCE(rdp_gp->nocb_gp_kthread, t);
2247 }
2248
2249 /* Spawn the kthread for this CPU. */
2250 t = kthread_run(rcu_nocb_cb_kthread, rdp,
2251 "rcuo%c/%d", rcu_state.abbr, cpu);
2252 if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo CB kthread, OOM is now expected behavior\n", __func__))
2253 return;
2254 WRITE_ONCE(rdp->nocb_cb_kthread, t);
2255 WRITE_ONCE(rdp->nocb_gp_kthread, rdp_gp->nocb_gp_kthread);
2256}
2257
2258/*
2259 * If the specified CPU is a no-CBs CPU that does not already have its
2260 * rcuo kthread, spawn it.
2261 */
2262static void rcu_spawn_cpu_nocb_kthread(int cpu)
2263{
2264 if (rcu_scheduler_fully_active)
2265 rcu_spawn_one_nocb_kthread(cpu);
2266}
2267
2268/*
2269 * Once the scheduler is running, spawn rcuo kthreads for all online
2270 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2271 * non-boot CPUs come online -- if this changes, we will need to add
2272 * some mutual exclusion.
2273 */
2274static void __init rcu_spawn_nocb_kthreads(void)
2275{
2276 int cpu;
2277
2278 for_each_online_cpu(cpu)
2279 rcu_spawn_cpu_nocb_kthread(cpu);
2280}
2281
2282/* How many CB CPU IDs per GP kthread? Default of -1 for sqrt(nr_cpu_ids). */
2283static int rcu_nocb_gp_stride = -1;
2284module_param(rcu_nocb_gp_stride, int, 0444);
2285
2286/*
2287 * Initialize GP-CB relationships for all no-CBs CPU.
2288 */
2289static void __init rcu_organize_nocb_kthreads(void)
2290{
2291 int cpu;
2292 bool firsttime = true;
2293 bool gotnocbs = false;
2294 bool gotnocbscbs = true;
2295 int ls = rcu_nocb_gp_stride;
2296 int nl = 0; /* Next GP kthread. */
2297 struct rcu_data *rdp;
2298 struct rcu_data *rdp_gp = NULL; /* Suppress misguided gcc warn. */
2299 struct rcu_data *rdp_prev = NULL;
2300
2301 if (!cpumask_available(rcu_nocb_mask))
2302 return;
2303 if (ls == -1) {
2304 ls = nr_cpu_ids / int_sqrt(nr_cpu_ids);
2305 rcu_nocb_gp_stride = ls;
2306 }
2307
2308 /*
2309 * Each pass through this loop sets up one rcu_data structure.
2310 * Should the corresponding CPU come online in the future, then
2311 * we will spawn the needed set of rcu_nocb_kthread() kthreads.
2312 */
2313 for_each_cpu(cpu, rcu_nocb_mask) {
2314 rdp = per_cpu_ptr(&rcu_data, cpu);
2315 if (rdp->cpu >= nl) {
2316 /* New GP kthread, set up for CBs & next GP. */
2317 gotnocbs = true;
2318 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2319 rdp->nocb_gp_rdp = rdp;
2320 rdp_gp = rdp;
2321 if (dump_tree) {
2322 if (!firsttime)
2323 pr_cont("%s\n", gotnocbscbs
2324 ? "" : " (self only)");
2325 gotnocbscbs = false;
2326 firsttime = false;
2327 pr_alert("%s: No-CB GP kthread CPU %d:",
2328 __func__, cpu);
2329 }
2330 } else {
2331 /* Another CB kthread, link to previous GP kthread. */
2332 gotnocbscbs = true;
2333 rdp->nocb_gp_rdp = rdp_gp;
2334 rdp_prev->nocb_next_cb_rdp = rdp;
2335 if (dump_tree)
2336 pr_cont(" %d", cpu);
2337 }
2338 rdp_prev = rdp;
2339 }
2340 if (gotnocbs && dump_tree)
2341 pr_cont("%s\n", gotnocbscbs ? "" : " (self only)");
2342}
2343
2344/*
2345 * Bind the current task to the offloaded CPUs. If there are no offloaded
2346 * CPUs, leave the task unbound. Splat if the bind attempt fails.
2347 */
2348void rcu_bind_current_to_nocb(void)
2349{
2350 if (cpumask_available(rcu_nocb_mask) && cpumask_weight(rcu_nocb_mask))
2351 WARN_ON(sched_setaffinity(current->pid, rcu_nocb_mask));
2352}
2353EXPORT_SYMBOL_GPL(rcu_bind_current_to_nocb);
2354
2355/*
2356 * Dump out nocb grace-period kthread state for the specified rcu_data
2357 * structure.
2358 */
2359static void show_rcu_nocb_gp_state(struct rcu_data *rdp)
2360{
2361 struct rcu_node *rnp = rdp->mynode;
2362
2363 pr_info("nocb GP %d %c%c%c%c%c%c %c[%c%c] %c%c:%ld rnp %d:%d %lu\n",
2364 rdp->cpu,
2365 "kK"[!!rdp->nocb_gp_kthread],
2366 "lL"[raw_spin_is_locked(&rdp->nocb_gp_lock)],
2367 "dD"[!!rdp->nocb_defer_wakeup],
2368 "tT"[timer_pending(&rdp->nocb_timer)],
2369 "bB"[timer_pending(&rdp->nocb_bypass_timer)],
2370 "sS"[!!rdp->nocb_gp_sleep],
2371 ".W"[swait_active(&rdp->nocb_gp_wq)],
2372 ".W"[swait_active(&rnp->nocb_gp_wq[0])],
2373 ".W"[swait_active(&rnp->nocb_gp_wq[1])],
2374 ".B"[!!rdp->nocb_gp_bypass],
2375 ".G"[!!rdp->nocb_gp_gp],
2376 (long)rdp->nocb_gp_seq,
2377 rnp->grplo, rnp->grphi, READ_ONCE(rdp->nocb_gp_loops));
2378}
2379
2380/* Dump out nocb kthread state for the specified rcu_data structure. */
2381static void show_rcu_nocb_state(struct rcu_data *rdp)
2382{
2383 struct rcu_segcblist *rsclp = &rdp->cblist;
2384 bool waslocked;
2385 bool wastimer;
2386 bool wassleep;
2387
2388 if (rdp->nocb_gp_rdp == rdp)
2389 show_rcu_nocb_gp_state(rdp);
2390
2391 pr_info(" CB %d->%d %c%c%c%c%c%c F%ld L%ld C%d %c%c%c%c%c q%ld\n",
2392 rdp->cpu, rdp->nocb_gp_rdp->cpu,
2393 "kK"[!!rdp->nocb_cb_kthread],
2394 "bB"[raw_spin_is_locked(&rdp->nocb_bypass_lock)],
2395 "cC"[!!atomic_read(&rdp->nocb_lock_contended)],
2396 "lL"[raw_spin_is_locked(&rdp->nocb_lock)],
2397 "sS"[!!rdp->nocb_cb_sleep],
2398 ".W"[swait_active(&rdp->nocb_cb_wq)],
2399 jiffies - rdp->nocb_bypass_first,
2400 jiffies - rdp->nocb_nobypass_last,
2401 rdp->nocb_nobypass_count,
2402 ".D"[rcu_segcblist_ready_cbs(rsclp)],
2403 ".W"[!rcu_segcblist_restempty(rsclp, RCU_DONE_TAIL)],
2404 ".R"[!rcu_segcblist_restempty(rsclp, RCU_WAIT_TAIL)],
2405 ".N"[!rcu_segcblist_restempty(rsclp, RCU_NEXT_READY_TAIL)],
2406 ".B"[!!rcu_cblist_n_cbs(&rdp->nocb_bypass)],
2407 rcu_segcblist_n_cbs(&rdp->cblist));
2408
2409 /* It is OK for GP kthreads to have GP state. */
2410 if (rdp->nocb_gp_rdp == rdp)
2411 return;
2412
2413 waslocked = raw_spin_is_locked(&rdp->nocb_gp_lock);
2414 wastimer = timer_pending(&rdp->nocb_timer);
2415 wassleep = swait_active(&rdp->nocb_gp_wq);
2416 if (!rdp->nocb_defer_wakeup && !rdp->nocb_gp_sleep &&
2417 !waslocked && !wastimer && !wassleep)
2418 return; /* Nothing untowards. */
2419
2420 pr_info(" !!! %c%c%c%c %c\n",
2421 "lL"[waslocked],
2422 "dD"[!!rdp->nocb_defer_wakeup],
2423 "tT"[wastimer],
2424 "sS"[!!rdp->nocb_gp_sleep],
2425 ".W"[wassleep]);
2426}
2427
2428#else /* #ifdef CONFIG_RCU_NOCB_CPU */
2429
2430/* No ->nocb_lock to acquire. */
2431static void rcu_nocb_lock(struct rcu_data *rdp)
2432{
2433}
2434
2435/* No ->nocb_lock to release. */
2436static void rcu_nocb_unlock(struct rcu_data *rdp)
2437{
2438}
2439
2440/* No ->nocb_lock to release. */
2441static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
2442 unsigned long flags)
2443{
2444 local_irq_restore(flags);
2445}
2446
2447/* Lockdep check that ->cblist may be safely accessed. */
2448static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
2449{
2450 lockdep_assert_irqs_disabled();
2451}
2452
2453static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
2454{
2455}
2456
2457static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
2458{
2459 return NULL;
2460}
2461
2462static void rcu_init_one_nocb(struct rcu_node *rnp)
2463{
2464}
2465
2466static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
2467 unsigned long j)
2468{
2469 return true;
2470}
2471
2472static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
2473 bool *was_alldone, unsigned long flags)
2474{
2475 return false;
2476}
2477
2478static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_empty,
2479 unsigned long flags)
2480{
2481 WARN_ON_ONCE(1); /* Should be dead code! */
2482}
2483
2484static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2485{
2486}
2487
2488static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2489{
2490 return false;
2491}
2492
2493static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2494{
2495}
2496
2497static void rcu_spawn_cpu_nocb_kthread(int cpu)
2498{
2499}
2500
2501static void __init rcu_spawn_nocb_kthreads(void)
2502{
2503}
2504
2505static void show_rcu_nocb_state(struct rcu_data *rdp)
2506{
2507}
2508
2509#endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2510
2511/*
2512 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2513 * grace-period kthread will do force_quiescent_state() processing?
2514 * The idea is to avoid waking up RCU core processing on such a
2515 * CPU unless the grace period has extended for too long.
2516 *
2517 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2518 * CONFIG_RCU_NOCB_CPU CPUs.
2519 */
2520static bool rcu_nohz_full_cpu(void)
2521{
2522#ifdef CONFIG_NO_HZ_FULL
2523 if (tick_nohz_full_cpu(smp_processor_id()) &&
2524 (!rcu_gp_in_progress() ||
2525 time_before(jiffies, READ_ONCE(rcu_state.gp_start) + HZ)))
2526 return true;
2527#endif /* #ifdef CONFIG_NO_HZ_FULL */
2528 return false;
2529}
2530
2531/*
2532 * Bind the RCU grace-period kthreads to the housekeeping CPU.
2533 */
2534static void rcu_bind_gp_kthread(void)
2535{
2536 if (!tick_nohz_full_enabled())
2537 return;
2538 housekeeping_affine(current, HK_FLAG_RCU);
2539}
2540
2541/* Record the current task on dyntick-idle entry. */
2542static void noinstr rcu_dynticks_task_enter(void)
2543{
2544#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2545 WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
2546#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2547}
2548
2549/* Record no current task on dyntick-idle exit. */
2550static void noinstr rcu_dynticks_task_exit(void)
2551{
2552#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2553 WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
2554#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2555}
2556
2557/* Turn on heavyweight RCU tasks trace readers on idle/user entry. */
2558static void rcu_dynticks_task_trace_enter(void)
2559{
2560#ifdef CONFIG_TASKS_RCU_TRACE
2561 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
2562 current->trc_reader_special.b.need_mb = true;
2563#endif /* #ifdef CONFIG_TASKS_RCU_TRACE */
2564}
2565
2566/* Turn off heavyweight RCU tasks trace readers on idle/user exit. */
2567static void rcu_dynticks_task_trace_exit(void)
2568{
2569#ifdef CONFIG_TASKS_RCU_TRACE
2570 if (IS_ENABLED(CONFIG_TASKS_TRACE_RCU_READ_MB))
2571 current->trc_reader_special.b.need_mb = false;
2572#endif /* #ifdef CONFIG_TASKS_RCU_TRACE */
2573}
1/*
2 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
3 * Internal non-public definitions that provide either classic
4 * or preemptible semantics.
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
10 *
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
15 *
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, you can access it online at
18 * http://www.gnu.org/licenses/gpl-2.0.html.
19 *
20 * Copyright Red Hat, 2009
21 * Copyright IBM Corporation, 2009
22 *
23 * Author: Ingo Molnar <mingo@elte.hu>
24 * Paul E. McKenney <paulmck@linux.vnet.ibm.com>
25 */
26
27#include <linux/delay.h>
28#include <linux/gfp.h>
29#include <linux/oom.h>
30#include <linux/smpboot.h>
31#include "../time/tick-internal.h"
32
33#define RCU_KTHREAD_PRIO 1
34
35#ifdef CONFIG_RCU_BOOST
36#define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO
37#else
38#define RCU_BOOST_PRIO RCU_KTHREAD_PRIO
39#endif
40
41#ifdef CONFIG_RCU_NOCB_CPU
42static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
43static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */
44static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
45static char __initdata nocb_buf[NR_CPUS * 5];
46#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
47
48/*
49 * Check the RCU kernel configuration parameters and print informative
50 * messages about anything out of the ordinary. If you like #ifdef, you
51 * will love this function.
52 */
53static void __init rcu_bootup_announce_oddness(void)
54{
55#ifdef CONFIG_RCU_TRACE
56 pr_info("\tRCU debugfs-based tracing is enabled.\n");
57#endif
58#if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
59 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
60 CONFIG_RCU_FANOUT);
61#endif
62#ifdef CONFIG_RCU_FANOUT_EXACT
63 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
64#endif
65#ifdef CONFIG_RCU_FAST_NO_HZ
66 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
67#endif
68#ifdef CONFIG_PROVE_RCU
69 pr_info("\tRCU lockdep checking is enabled.\n");
70#endif
71#ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
72 pr_info("\tRCU torture testing starts during boot.\n");
73#endif
74#if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE)
75 pr_info("\tDump stacks of tasks blocking RCU-preempt GP.\n");
76#endif
77#if defined(CONFIG_RCU_CPU_STALL_INFO)
78 pr_info("\tAdditional per-CPU info printed with stalls.\n");
79#endif
80#if NUM_RCU_LVL_4 != 0
81 pr_info("\tFour-level hierarchy is enabled.\n");
82#endif
83 if (rcu_fanout_leaf != CONFIG_RCU_FANOUT_LEAF)
84 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
85 if (nr_cpu_ids != NR_CPUS)
86 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
87#ifdef CONFIG_RCU_NOCB_CPU
88#ifndef CONFIG_RCU_NOCB_CPU_NONE
89 if (!have_rcu_nocb_mask) {
90 zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL);
91 have_rcu_nocb_mask = true;
92 }
93#ifdef CONFIG_RCU_NOCB_CPU_ZERO
94 pr_info("\tOffload RCU callbacks from CPU 0\n");
95 cpumask_set_cpu(0, rcu_nocb_mask);
96#endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
97#ifdef CONFIG_RCU_NOCB_CPU_ALL
98 pr_info("\tOffload RCU callbacks from all CPUs\n");
99 cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
100#endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
101#endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
102 if (have_rcu_nocb_mask) {
103 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
104 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
105 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
106 rcu_nocb_mask);
107 }
108 cpulist_scnprintf(nocb_buf, sizeof(nocb_buf), rcu_nocb_mask);
109 pr_info("\tOffload RCU callbacks from CPUs: %s.\n", nocb_buf);
110 if (rcu_nocb_poll)
111 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
112 }
113#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
114}
115
116#ifdef CONFIG_TREE_PREEMPT_RCU
117
118RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
119static struct rcu_state *rcu_state = &rcu_preempt_state;
120
121static int rcu_preempted_readers_exp(struct rcu_node *rnp);
122
123/*
124 * Tell them what RCU they are running.
125 */
126static void __init rcu_bootup_announce(void)
127{
128 pr_info("Preemptible hierarchical RCU implementation.\n");
129 rcu_bootup_announce_oddness();
130}
131
132/*
133 * Return the number of RCU-preempt batches processed thus far
134 * for debug and statistics.
135 */
136long rcu_batches_completed_preempt(void)
137{
138 return rcu_preempt_state.completed;
139}
140EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt);
141
142/*
143 * Return the number of RCU batches processed thus far for debug & stats.
144 */
145long rcu_batches_completed(void)
146{
147 return rcu_batches_completed_preempt();
148}
149EXPORT_SYMBOL_GPL(rcu_batches_completed);
150
151/*
152 * Force a quiescent state for preemptible RCU.
153 */
154void rcu_force_quiescent_state(void)
155{
156 force_quiescent_state(&rcu_preempt_state);
157}
158EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
159
160/*
161 * Record a preemptible-RCU quiescent state for the specified CPU. Note
162 * that this just means that the task currently running on the CPU is
163 * not in a quiescent state. There might be any number of tasks blocked
164 * while in an RCU read-side critical section.
165 *
166 * Unlike the other rcu_*_qs() functions, callers to this function
167 * must disable irqs in order to protect the assignment to
168 * ->rcu_read_unlock_special.
169 */
170static void rcu_preempt_qs(int cpu)
171{
172 struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu);
173
174 if (rdp->passed_quiesce == 0)
175 trace_rcu_grace_period(TPS("rcu_preempt"), rdp->gpnum, TPS("cpuqs"));
176 rdp->passed_quiesce = 1;
177 current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
178}
179
180/*
181 * We have entered the scheduler, and the current task might soon be
182 * context-switched away from. If this task is in an RCU read-side
183 * critical section, we will no longer be able to rely on the CPU to
184 * record that fact, so we enqueue the task on the blkd_tasks list.
185 * The task will dequeue itself when it exits the outermost enclosing
186 * RCU read-side critical section. Therefore, the current grace period
187 * cannot be permitted to complete until the blkd_tasks list entries
188 * predating the current grace period drain, in other words, until
189 * rnp->gp_tasks becomes NULL.
190 *
191 * Caller must disable preemption.
192 */
193static void rcu_preempt_note_context_switch(int cpu)
194{
195 struct task_struct *t = current;
196 unsigned long flags;
197 struct rcu_data *rdp;
198 struct rcu_node *rnp;
199
200 if (t->rcu_read_lock_nesting > 0 &&
201 (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) {
202
203 /* Possibly blocking in an RCU read-side critical section. */
204 rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu);
205 rnp = rdp->mynode;
206 raw_spin_lock_irqsave(&rnp->lock, flags);
207 smp_mb__after_unlock_lock();
208 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED;
209 t->rcu_blocked_node = rnp;
210
211 /*
212 * If this CPU has already checked in, then this task
213 * will hold up the next grace period rather than the
214 * current grace period. Queue the task accordingly.
215 * If the task is queued for the current grace period
216 * (i.e., this CPU has not yet passed through a quiescent
217 * state for the current grace period), then as long
218 * as that task remains queued, the current grace period
219 * cannot end. Note that there is some uncertainty as
220 * to exactly when the current grace period started.
221 * We take a conservative approach, which can result
222 * in unnecessarily waiting on tasks that started very
223 * slightly after the current grace period began. C'est
224 * la vie!!!
225 *
226 * But first, note that the current CPU must still be
227 * on line!
228 */
229 WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
230 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
231 if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
232 list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
233 rnp->gp_tasks = &t->rcu_node_entry;
234#ifdef CONFIG_RCU_BOOST
235 if (rnp->boost_tasks != NULL)
236 rnp->boost_tasks = rnp->gp_tasks;
237#endif /* #ifdef CONFIG_RCU_BOOST */
238 } else {
239 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
240 if (rnp->qsmask & rdp->grpmask)
241 rnp->gp_tasks = &t->rcu_node_entry;
242 }
243 trace_rcu_preempt_task(rdp->rsp->name,
244 t->pid,
245 (rnp->qsmask & rdp->grpmask)
246 ? rnp->gpnum
247 : rnp->gpnum + 1);
248 raw_spin_unlock_irqrestore(&rnp->lock, flags);
249 } else if (t->rcu_read_lock_nesting < 0 &&
250 t->rcu_read_unlock_special) {
251
252 /*
253 * Complete exit from RCU read-side critical section on
254 * behalf of preempted instance of __rcu_read_unlock().
255 */
256 rcu_read_unlock_special(t);
257 }
258
259 /*
260 * Either we were not in an RCU read-side critical section to
261 * begin with, or we have now recorded that critical section
262 * globally. Either way, we can now note a quiescent state
263 * for this CPU. Again, if we were in an RCU read-side critical
264 * section, and if that critical section was blocking the current
265 * grace period, then the fact that the task has been enqueued
266 * means that we continue to block the current grace period.
267 */
268 local_irq_save(flags);
269 rcu_preempt_qs(cpu);
270 local_irq_restore(flags);
271}
272
273/*
274 * Check for preempted RCU readers blocking the current grace period
275 * for the specified rcu_node structure. If the caller needs a reliable
276 * answer, it must hold the rcu_node's ->lock.
277 */
278static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
279{
280 return rnp->gp_tasks != NULL;
281}
282
283/*
284 * Record a quiescent state for all tasks that were previously queued
285 * on the specified rcu_node structure and that were blocking the current
286 * RCU grace period. The caller must hold the specified rnp->lock with
287 * irqs disabled, and this lock is released upon return, but irqs remain
288 * disabled.
289 */
290static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
291 __releases(rnp->lock)
292{
293 unsigned long mask;
294 struct rcu_node *rnp_p;
295
296 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
297 raw_spin_unlock_irqrestore(&rnp->lock, flags);
298 return; /* Still need more quiescent states! */
299 }
300
301 rnp_p = rnp->parent;
302 if (rnp_p == NULL) {
303 /*
304 * Either there is only one rcu_node in the tree,
305 * or tasks were kicked up to root rcu_node due to
306 * CPUs going offline.
307 */
308 rcu_report_qs_rsp(&rcu_preempt_state, flags);
309 return;
310 }
311
312 /* Report up the rest of the hierarchy. */
313 mask = rnp->grpmask;
314 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
315 raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
316 smp_mb__after_unlock_lock();
317 rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
318}
319
320/*
321 * Advance a ->blkd_tasks-list pointer to the next entry, instead
322 * returning NULL if at the end of the list.
323 */
324static struct list_head *rcu_next_node_entry(struct task_struct *t,
325 struct rcu_node *rnp)
326{
327 struct list_head *np;
328
329 np = t->rcu_node_entry.next;
330 if (np == &rnp->blkd_tasks)
331 np = NULL;
332 return np;
333}
334
335/*
336 * Handle special cases during rcu_read_unlock(), such as needing to
337 * notify RCU core processing or task having blocked during the RCU
338 * read-side critical section.
339 */
340void rcu_read_unlock_special(struct task_struct *t)
341{
342 int empty;
343 int empty_exp;
344 int empty_exp_now;
345 unsigned long flags;
346 struct list_head *np;
347#ifdef CONFIG_RCU_BOOST
348 struct rt_mutex *rbmp = NULL;
349#endif /* #ifdef CONFIG_RCU_BOOST */
350 struct rcu_node *rnp;
351 int special;
352
353 /* NMI handlers cannot block and cannot safely manipulate state. */
354 if (in_nmi())
355 return;
356
357 local_irq_save(flags);
358
359 /*
360 * If RCU core is waiting for this CPU to exit critical section,
361 * let it know that we have done so.
362 */
363 special = t->rcu_read_unlock_special;
364 if (special & RCU_READ_UNLOCK_NEED_QS) {
365 rcu_preempt_qs(smp_processor_id());
366 if (!t->rcu_read_unlock_special) {
367 local_irq_restore(flags);
368 return;
369 }
370 }
371
372 /* Hardware IRQ handlers cannot block, complain if they get here. */
373 if (WARN_ON_ONCE(in_irq() || in_serving_softirq())) {
374 local_irq_restore(flags);
375 return;
376 }
377
378 /* Clean up if blocked during RCU read-side critical section. */
379 if (special & RCU_READ_UNLOCK_BLOCKED) {
380 t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED;
381
382 /*
383 * Remove this task from the list it blocked on. The
384 * task can migrate while we acquire the lock, but at
385 * most one time. So at most two passes through loop.
386 */
387 for (;;) {
388 rnp = t->rcu_blocked_node;
389 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
390 smp_mb__after_unlock_lock();
391 if (rnp == t->rcu_blocked_node)
392 break;
393 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
394 }
395 empty = !rcu_preempt_blocked_readers_cgp(rnp);
396 empty_exp = !rcu_preempted_readers_exp(rnp);
397 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
398 np = rcu_next_node_entry(t, rnp);
399 list_del_init(&t->rcu_node_entry);
400 t->rcu_blocked_node = NULL;
401 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
402 rnp->gpnum, t->pid);
403 if (&t->rcu_node_entry == rnp->gp_tasks)
404 rnp->gp_tasks = np;
405 if (&t->rcu_node_entry == rnp->exp_tasks)
406 rnp->exp_tasks = np;
407#ifdef CONFIG_RCU_BOOST
408 if (&t->rcu_node_entry == rnp->boost_tasks)
409 rnp->boost_tasks = np;
410 /* Snapshot/clear ->rcu_boost_mutex with rcu_node lock held. */
411 if (t->rcu_boost_mutex) {
412 rbmp = t->rcu_boost_mutex;
413 t->rcu_boost_mutex = NULL;
414 }
415#endif /* #ifdef CONFIG_RCU_BOOST */
416
417 /*
418 * If this was the last task on the current list, and if
419 * we aren't waiting on any CPUs, report the quiescent state.
420 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
421 * so we must take a snapshot of the expedited state.
422 */
423 empty_exp_now = !rcu_preempted_readers_exp(rnp);
424 if (!empty && !rcu_preempt_blocked_readers_cgp(rnp)) {
425 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
426 rnp->gpnum,
427 0, rnp->qsmask,
428 rnp->level,
429 rnp->grplo,
430 rnp->grphi,
431 !!rnp->gp_tasks);
432 rcu_report_unblock_qs_rnp(rnp, flags);
433 } else {
434 raw_spin_unlock_irqrestore(&rnp->lock, flags);
435 }
436
437#ifdef CONFIG_RCU_BOOST
438 /* Unboost if we were boosted. */
439 if (rbmp)
440 rt_mutex_unlock(rbmp);
441#endif /* #ifdef CONFIG_RCU_BOOST */
442
443 /*
444 * If this was the last task on the expedited lists,
445 * then we need to report up the rcu_node hierarchy.
446 */
447 if (!empty_exp && empty_exp_now)
448 rcu_report_exp_rnp(&rcu_preempt_state, rnp, true);
449 } else {
450 local_irq_restore(flags);
451 }
452}
453
454#ifdef CONFIG_RCU_CPU_STALL_VERBOSE
455
456/*
457 * Dump detailed information for all tasks blocking the current RCU
458 * grace period on the specified rcu_node structure.
459 */
460static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
461{
462 unsigned long flags;
463 struct task_struct *t;
464
465 raw_spin_lock_irqsave(&rnp->lock, flags);
466 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
467 raw_spin_unlock_irqrestore(&rnp->lock, flags);
468 return;
469 }
470 t = list_entry(rnp->gp_tasks,
471 struct task_struct, rcu_node_entry);
472 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
473 sched_show_task(t);
474 raw_spin_unlock_irqrestore(&rnp->lock, flags);
475}
476
477/*
478 * Dump detailed information for all tasks blocking the current RCU
479 * grace period.
480 */
481static void rcu_print_detail_task_stall(struct rcu_state *rsp)
482{
483 struct rcu_node *rnp = rcu_get_root(rsp);
484
485 rcu_print_detail_task_stall_rnp(rnp);
486 rcu_for_each_leaf_node(rsp, rnp)
487 rcu_print_detail_task_stall_rnp(rnp);
488}
489
490#else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
491
492static void rcu_print_detail_task_stall(struct rcu_state *rsp)
493{
494}
495
496#endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
497
498#ifdef CONFIG_RCU_CPU_STALL_INFO
499
500static void rcu_print_task_stall_begin(struct rcu_node *rnp)
501{
502 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
503 rnp->level, rnp->grplo, rnp->grphi);
504}
505
506static void rcu_print_task_stall_end(void)
507{
508 pr_cont("\n");
509}
510
511#else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
512
513static void rcu_print_task_stall_begin(struct rcu_node *rnp)
514{
515}
516
517static void rcu_print_task_stall_end(void)
518{
519}
520
521#endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
522
523/*
524 * Scan the current list of tasks blocked within RCU read-side critical
525 * sections, printing out the tid of each.
526 */
527static int rcu_print_task_stall(struct rcu_node *rnp)
528{
529 struct task_struct *t;
530 int ndetected = 0;
531
532 if (!rcu_preempt_blocked_readers_cgp(rnp))
533 return 0;
534 rcu_print_task_stall_begin(rnp);
535 t = list_entry(rnp->gp_tasks,
536 struct task_struct, rcu_node_entry);
537 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
538 pr_cont(" P%d", t->pid);
539 ndetected++;
540 }
541 rcu_print_task_stall_end();
542 return ndetected;
543}
544
545/*
546 * Check that the list of blocked tasks for the newly completed grace
547 * period is in fact empty. It is a serious bug to complete a grace
548 * period that still has RCU readers blocked! This function must be
549 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
550 * must be held by the caller.
551 *
552 * Also, if there are blocked tasks on the list, they automatically
553 * block the newly created grace period, so set up ->gp_tasks accordingly.
554 */
555static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
556{
557 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
558 if (!list_empty(&rnp->blkd_tasks))
559 rnp->gp_tasks = rnp->blkd_tasks.next;
560 WARN_ON_ONCE(rnp->qsmask);
561}
562
563#ifdef CONFIG_HOTPLUG_CPU
564
565/*
566 * Handle tasklist migration for case in which all CPUs covered by the
567 * specified rcu_node have gone offline. Move them up to the root
568 * rcu_node. The reason for not just moving them to the immediate
569 * parent is to remove the need for rcu_read_unlock_special() to
570 * make more than two attempts to acquire the target rcu_node's lock.
571 * Returns true if there were tasks blocking the current RCU grace
572 * period.
573 *
574 * Returns 1 if there was previously a task blocking the current grace
575 * period on the specified rcu_node structure.
576 *
577 * The caller must hold rnp->lock with irqs disabled.
578 */
579static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
580 struct rcu_node *rnp,
581 struct rcu_data *rdp)
582{
583 struct list_head *lp;
584 struct list_head *lp_root;
585 int retval = 0;
586 struct rcu_node *rnp_root = rcu_get_root(rsp);
587 struct task_struct *t;
588
589 if (rnp == rnp_root) {
590 WARN_ONCE(1, "Last CPU thought to be offlined?");
591 return 0; /* Shouldn't happen: at least one CPU online. */
592 }
593
594 /* If we are on an internal node, complain bitterly. */
595 WARN_ON_ONCE(rnp != rdp->mynode);
596
597 /*
598 * Move tasks up to root rcu_node. Don't try to get fancy for
599 * this corner-case operation -- just put this node's tasks
600 * at the head of the root node's list, and update the root node's
601 * ->gp_tasks and ->exp_tasks pointers to those of this node's,
602 * if non-NULL. This might result in waiting for more tasks than
603 * absolutely necessary, but this is a good performance/complexity
604 * tradeoff.
605 */
606 if (rcu_preempt_blocked_readers_cgp(rnp) && rnp->qsmask == 0)
607 retval |= RCU_OFL_TASKS_NORM_GP;
608 if (rcu_preempted_readers_exp(rnp))
609 retval |= RCU_OFL_TASKS_EXP_GP;
610 lp = &rnp->blkd_tasks;
611 lp_root = &rnp_root->blkd_tasks;
612 while (!list_empty(lp)) {
613 t = list_entry(lp->next, typeof(*t), rcu_node_entry);
614 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
615 smp_mb__after_unlock_lock();
616 list_del(&t->rcu_node_entry);
617 t->rcu_blocked_node = rnp_root;
618 list_add(&t->rcu_node_entry, lp_root);
619 if (&t->rcu_node_entry == rnp->gp_tasks)
620 rnp_root->gp_tasks = rnp->gp_tasks;
621 if (&t->rcu_node_entry == rnp->exp_tasks)
622 rnp_root->exp_tasks = rnp->exp_tasks;
623#ifdef CONFIG_RCU_BOOST
624 if (&t->rcu_node_entry == rnp->boost_tasks)
625 rnp_root->boost_tasks = rnp->boost_tasks;
626#endif /* #ifdef CONFIG_RCU_BOOST */
627 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
628 }
629
630 rnp->gp_tasks = NULL;
631 rnp->exp_tasks = NULL;
632#ifdef CONFIG_RCU_BOOST
633 rnp->boost_tasks = NULL;
634 /*
635 * In case root is being boosted and leaf was not. Make sure
636 * that we boost the tasks blocking the current grace period
637 * in this case.
638 */
639 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
640 smp_mb__after_unlock_lock();
641 if (rnp_root->boost_tasks != NULL &&
642 rnp_root->boost_tasks != rnp_root->gp_tasks &&
643 rnp_root->boost_tasks != rnp_root->exp_tasks)
644 rnp_root->boost_tasks = rnp_root->gp_tasks;
645 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
646#endif /* #ifdef CONFIG_RCU_BOOST */
647
648 return retval;
649}
650
651#endif /* #ifdef CONFIG_HOTPLUG_CPU */
652
653/*
654 * Check for a quiescent state from the current CPU. When a task blocks,
655 * the task is recorded in the corresponding CPU's rcu_node structure,
656 * which is checked elsewhere.
657 *
658 * Caller must disable hard irqs.
659 */
660static void rcu_preempt_check_callbacks(int cpu)
661{
662 struct task_struct *t = current;
663
664 if (t->rcu_read_lock_nesting == 0) {
665 rcu_preempt_qs(cpu);
666 return;
667 }
668 if (t->rcu_read_lock_nesting > 0 &&
669 per_cpu(rcu_preempt_data, cpu).qs_pending)
670 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS;
671}
672
673#ifdef CONFIG_RCU_BOOST
674
675static void rcu_preempt_do_callbacks(void)
676{
677 rcu_do_batch(&rcu_preempt_state, this_cpu_ptr(&rcu_preempt_data));
678}
679
680#endif /* #ifdef CONFIG_RCU_BOOST */
681
682/*
683 * Queue a preemptible-RCU callback for invocation after a grace period.
684 */
685void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
686{
687 __call_rcu(head, func, &rcu_preempt_state, -1, 0);
688}
689EXPORT_SYMBOL_GPL(call_rcu);
690
691/*
692 * Queue an RCU callback for lazy invocation after a grace period.
693 * This will likely be later named something like "call_rcu_lazy()",
694 * but this change will require some way of tagging the lazy RCU
695 * callbacks in the list of pending callbacks. Until then, this
696 * function may only be called from __kfree_rcu().
697 */
698void kfree_call_rcu(struct rcu_head *head,
699 void (*func)(struct rcu_head *rcu))
700{
701 __call_rcu(head, func, &rcu_preempt_state, -1, 1);
702}
703EXPORT_SYMBOL_GPL(kfree_call_rcu);
704
705/**
706 * synchronize_rcu - wait until a grace period has elapsed.
707 *
708 * Control will return to the caller some time after a full grace
709 * period has elapsed, in other words after all currently executing RCU
710 * read-side critical sections have completed. Note, however, that
711 * upon return from synchronize_rcu(), the caller might well be executing
712 * concurrently with new RCU read-side critical sections that began while
713 * synchronize_rcu() was waiting. RCU read-side critical sections are
714 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
715 *
716 * See the description of synchronize_sched() for more detailed information
717 * on memory ordering guarantees.
718 */
719void synchronize_rcu(void)
720{
721 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
722 !lock_is_held(&rcu_lock_map) &&
723 !lock_is_held(&rcu_sched_lock_map),
724 "Illegal synchronize_rcu() in RCU read-side critical section");
725 if (!rcu_scheduler_active)
726 return;
727 if (rcu_expedited)
728 synchronize_rcu_expedited();
729 else
730 wait_rcu_gp(call_rcu);
731}
732EXPORT_SYMBOL_GPL(synchronize_rcu);
733
734static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
735static unsigned long sync_rcu_preempt_exp_count;
736static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
737
738/*
739 * Return non-zero if there are any tasks in RCU read-side critical
740 * sections blocking the current preemptible-RCU expedited grace period.
741 * If there is no preemptible-RCU expedited grace period currently in
742 * progress, returns zero unconditionally.
743 */
744static int rcu_preempted_readers_exp(struct rcu_node *rnp)
745{
746 return rnp->exp_tasks != NULL;
747}
748
749/*
750 * return non-zero if there is no RCU expedited grace period in progress
751 * for the specified rcu_node structure, in other words, if all CPUs and
752 * tasks covered by the specified rcu_node structure have done their bit
753 * for the current expedited grace period. Works only for preemptible
754 * RCU -- other RCU implementation use other means.
755 *
756 * Caller must hold sync_rcu_preempt_exp_mutex.
757 */
758static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
759{
760 return !rcu_preempted_readers_exp(rnp) &&
761 ACCESS_ONCE(rnp->expmask) == 0;
762}
763
764/*
765 * Report the exit from RCU read-side critical section for the last task
766 * that queued itself during or before the current expedited preemptible-RCU
767 * grace period. This event is reported either to the rcu_node structure on
768 * which the task was queued or to one of that rcu_node structure's ancestors,
769 * recursively up the tree. (Calm down, calm down, we do the recursion
770 * iteratively!)
771 *
772 * Most callers will set the "wake" flag, but the task initiating the
773 * expedited grace period need not wake itself.
774 *
775 * Caller must hold sync_rcu_preempt_exp_mutex.
776 */
777static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
778 bool wake)
779{
780 unsigned long flags;
781 unsigned long mask;
782
783 raw_spin_lock_irqsave(&rnp->lock, flags);
784 smp_mb__after_unlock_lock();
785 for (;;) {
786 if (!sync_rcu_preempt_exp_done(rnp)) {
787 raw_spin_unlock_irqrestore(&rnp->lock, flags);
788 break;
789 }
790 if (rnp->parent == NULL) {
791 raw_spin_unlock_irqrestore(&rnp->lock, flags);
792 if (wake) {
793 smp_mb(); /* EGP done before wake_up(). */
794 wake_up(&sync_rcu_preempt_exp_wq);
795 }
796 break;
797 }
798 mask = rnp->grpmask;
799 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
800 rnp = rnp->parent;
801 raw_spin_lock(&rnp->lock); /* irqs already disabled */
802 smp_mb__after_unlock_lock();
803 rnp->expmask &= ~mask;
804 }
805}
806
807/*
808 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
809 * grace period for the specified rcu_node structure. If there are no such
810 * tasks, report it up the rcu_node hierarchy.
811 *
812 * Caller must hold sync_rcu_preempt_exp_mutex and must exclude
813 * CPU hotplug operations.
814 */
815static void
816sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
817{
818 unsigned long flags;
819 int must_wait = 0;
820
821 raw_spin_lock_irqsave(&rnp->lock, flags);
822 smp_mb__after_unlock_lock();
823 if (list_empty(&rnp->blkd_tasks)) {
824 raw_spin_unlock_irqrestore(&rnp->lock, flags);
825 } else {
826 rnp->exp_tasks = rnp->blkd_tasks.next;
827 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
828 must_wait = 1;
829 }
830 if (!must_wait)
831 rcu_report_exp_rnp(rsp, rnp, false); /* Don't wake self. */
832}
833
834/**
835 * synchronize_rcu_expedited - Brute-force RCU grace period
836 *
837 * Wait for an RCU-preempt grace period, but expedite it. The basic
838 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
839 * the ->blkd_tasks lists and wait for this list to drain. This consumes
840 * significant time on all CPUs and is unfriendly to real-time workloads,
841 * so is thus not recommended for any sort of common-case code.
842 * In fact, if you are using synchronize_rcu_expedited() in a loop,
843 * please restructure your code to batch your updates, and then Use a
844 * single synchronize_rcu() instead.
845 *
846 * Note that it is illegal to call this function while holding any lock
847 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
848 * to call this function from a CPU-hotplug notifier. Failing to observe
849 * these restriction will result in deadlock.
850 */
851void synchronize_rcu_expedited(void)
852{
853 unsigned long flags;
854 struct rcu_node *rnp;
855 struct rcu_state *rsp = &rcu_preempt_state;
856 unsigned long snap;
857 int trycount = 0;
858
859 smp_mb(); /* Caller's modifications seen first by other CPUs. */
860 snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
861 smp_mb(); /* Above access cannot bleed into critical section. */
862
863 /*
864 * Block CPU-hotplug operations. This means that any CPU-hotplug
865 * operation that finds an rcu_node structure with tasks in the
866 * process of being boosted will know that all tasks blocking
867 * this expedited grace period will already be in the process of
868 * being boosted. This simplifies the process of moving tasks
869 * from leaf to root rcu_node structures.
870 */
871 get_online_cpus();
872
873 /*
874 * Acquire lock, falling back to synchronize_rcu() if too many
875 * lock-acquisition failures. Of course, if someone does the
876 * expedited grace period for us, just leave.
877 */
878 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
879 if (ULONG_CMP_LT(snap,
880 ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
881 put_online_cpus();
882 goto mb_ret; /* Others did our work for us. */
883 }
884 if (trycount++ < 10) {
885 udelay(trycount * num_online_cpus());
886 } else {
887 put_online_cpus();
888 wait_rcu_gp(call_rcu);
889 return;
890 }
891 }
892 if (ULONG_CMP_LT(snap, ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
893 put_online_cpus();
894 goto unlock_mb_ret; /* Others did our work for us. */
895 }
896
897 /* force all RCU readers onto ->blkd_tasks lists. */
898 synchronize_sched_expedited();
899
900 /* Initialize ->expmask for all non-leaf rcu_node structures. */
901 rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
902 raw_spin_lock_irqsave(&rnp->lock, flags);
903 smp_mb__after_unlock_lock();
904 rnp->expmask = rnp->qsmaskinit;
905 raw_spin_unlock_irqrestore(&rnp->lock, flags);
906 }
907
908 /* Snapshot current state of ->blkd_tasks lists. */
909 rcu_for_each_leaf_node(rsp, rnp)
910 sync_rcu_preempt_exp_init(rsp, rnp);
911 if (NUM_RCU_NODES > 1)
912 sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp));
913
914 put_online_cpus();
915
916 /* Wait for snapshotted ->blkd_tasks lists to drain. */
917 rnp = rcu_get_root(rsp);
918 wait_event(sync_rcu_preempt_exp_wq,
919 sync_rcu_preempt_exp_done(rnp));
920
921 /* Clean up and exit. */
922 smp_mb(); /* ensure expedited GP seen before counter increment. */
923 ACCESS_ONCE(sync_rcu_preempt_exp_count)++;
924unlock_mb_ret:
925 mutex_unlock(&sync_rcu_preempt_exp_mutex);
926mb_ret:
927 smp_mb(); /* ensure subsequent action seen after grace period. */
928}
929EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
930
931/**
932 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
933 *
934 * Note that this primitive does not necessarily wait for an RCU grace period
935 * to complete. For example, if there are no RCU callbacks queued anywhere
936 * in the system, then rcu_barrier() is within its rights to return
937 * immediately, without waiting for anything, much less an RCU grace period.
938 */
939void rcu_barrier(void)
940{
941 _rcu_barrier(&rcu_preempt_state);
942}
943EXPORT_SYMBOL_GPL(rcu_barrier);
944
945/*
946 * Initialize preemptible RCU's state structures.
947 */
948static void __init __rcu_init_preempt(void)
949{
950 rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
951}
952
953/*
954 * Check for a task exiting while in a preemptible-RCU read-side
955 * critical section, clean up if so. No need to issue warnings,
956 * as debug_check_no_locks_held() already does this if lockdep
957 * is enabled.
958 */
959void exit_rcu(void)
960{
961 struct task_struct *t = current;
962
963 if (likely(list_empty(¤t->rcu_node_entry)))
964 return;
965 t->rcu_read_lock_nesting = 1;
966 barrier();
967 t->rcu_read_unlock_special = RCU_READ_UNLOCK_BLOCKED;
968 __rcu_read_unlock();
969}
970
971#else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
972
973static struct rcu_state *rcu_state = &rcu_sched_state;
974
975/*
976 * Tell them what RCU they are running.
977 */
978static void __init rcu_bootup_announce(void)
979{
980 pr_info("Hierarchical RCU implementation.\n");
981 rcu_bootup_announce_oddness();
982}
983
984/*
985 * Return the number of RCU batches processed thus far for debug & stats.
986 */
987long rcu_batches_completed(void)
988{
989 return rcu_batches_completed_sched();
990}
991EXPORT_SYMBOL_GPL(rcu_batches_completed);
992
993/*
994 * Force a quiescent state for RCU, which, because there is no preemptible
995 * RCU, becomes the same as rcu-sched.
996 */
997void rcu_force_quiescent_state(void)
998{
999 rcu_sched_force_quiescent_state();
1000}
1001EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
1002
1003/*
1004 * Because preemptible RCU does not exist, we never have to check for
1005 * CPUs being in quiescent states.
1006 */
1007static void rcu_preempt_note_context_switch(int cpu)
1008{
1009}
1010
1011/*
1012 * Because preemptible RCU does not exist, there are never any preempted
1013 * RCU readers.
1014 */
1015static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
1016{
1017 return 0;
1018}
1019
1020#ifdef CONFIG_HOTPLUG_CPU
1021
1022/* Because preemptible RCU does not exist, no quieting of tasks. */
1023static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
1024{
1025 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1026}
1027
1028#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1029
1030/*
1031 * Because preemptible RCU does not exist, we never have to check for
1032 * tasks blocked within RCU read-side critical sections.
1033 */
1034static void rcu_print_detail_task_stall(struct rcu_state *rsp)
1035{
1036}
1037
1038/*
1039 * Because preemptible RCU does not exist, we never have to check for
1040 * tasks blocked within RCU read-side critical sections.
1041 */
1042static int rcu_print_task_stall(struct rcu_node *rnp)
1043{
1044 return 0;
1045}
1046
1047/*
1048 * Because there is no preemptible RCU, there can be no readers blocked,
1049 * so there is no need to check for blocked tasks. So check only for
1050 * bogus qsmask values.
1051 */
1052static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
1053{
1054 WARN_ON_ONCE(rnp->qsmask);
1055}
1056
1057#ifdef CONFIG_HOTPLUG_CPU
1058
1059/*
1060 * Because preemptible RCU does not exist, it never needs to migrate
1061 * tasks that were blocked within RCU read-side critical sections, and
1062 * such non-existent tasks cannot possibly have been blocking the current
1063 * grace period.
1064 */
1065static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
1066 struct rcu_node *rnp,
1067 struct rcu_data *rdp)
1068{
1069 return 0;
1070}
1071
1072#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1073
1074/*
1075 * Because preemptible RCU does not exist, it never has any callbacks
1076 * to check.
1077 */
1078static void rcu_preempt_check_callbacks(int cpu)
1079{
1080}
1081
1082/*
1083 * Queue an RCU callback for lazy invocation after a grace period.
1084 * This will likely be later named something like "call_rcu_lazy()",
1085 * but this change will require some way of tagging the lazy RCU
1086 * callbacks in the list of pending callbacks. Until then, this
1087 * function may only be called from __kfree_rcu().
1088 *
1089 * Because there is no preemptible RCU, we use RCU-sched instead.
1090 */
1091void kfree_call_rcu(struct rcu_head *head,
1092 void (*func)(struct rcu_head *rcu))
1093{
1094 __call_rcu(head, func, &rcu_sched_state, -1, 1);
1095}
1096EXPORT_SYMBOL_GPL(kfree_call_rcu);
1097
1098/*
1099 * Wait for an rcu-preempt grace period, but make it happen quickly.
1100 * But because preemptible RCU does not exist, map to rcu-sched.
1101 */
1102void synchronize_rcu_expedited(void)
1103{
1104 synchronize_sched_expedited();
1105}
1106EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
1107
1108#ifdef CONFIG_HOTPLUG_CPU
1109
1110/*
1111 * Because preemptible RCU does not exist, there is never any need to
1112 * report on tasks preempted in RCU read-side critical sections during
1113 * expedited RCU grace periods.
1114 */
1115static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
1116 bool wake)
1117{
1118}
1119
1120#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1121
1122/*
1123 * Because preemptible RCU does not exist, rcu_barrier() is just
1124 * another name for rcu_barrier_sched().
1125 */
1126void rcu_barrier(void)
1127{
1128 rcu_barrier_sched();
1129}
1130EXPORT_SYMBOL_GPL(rcu_barrier);
1131
1132/*
1133 * Because preemptible RCU does not exist, it need not be initialized.
1134 */
1135static void __init __rcu_init_preempt(void)
1136{
1137}
1138
1139/*
1140 * Because preemptible RCU does not exist, tasks cannot possibly exit
1141 * while in preemptible RCU read-side critical sections.
1142 */
1143void exit_rcu(void)
1144{
1145}
1146
1147#endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1148
1149#ifdef CONFIG_RCU_BOOST
1150
1151#include "../locking/rtmutex_common.h"
1152
1153#ifdef CONFIG_RCU_TRACE
1154
1155static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1156{
1157 if (list_empty(&rnp->blkd_tasks))
1158 rnp->n_balk_blkd_tasks++;
1159 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
1160 rnp->n_balk_exp_gp_tasks++;
1161 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
1162 rnp->n_balk_boost_tasks++;
1163 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
1164 rnp->n_balk_notblocked++;
1165 else if (rnp->gp_tasks != NULL &&
1166 ULONG_CMP_LT(jiffies, rnp->boost_time))
1167 rnp->n_balk_notyet++;
1168 else
1169 rnp->n_balk_nos++;
1170}
1171
1172#else /* #ifdef CONFIG_RCU_TRACE */
1173
1174static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1175{
1176}
1177
1178#endif /* #else #ifdef CONFIG_RCU_TRACE */
1179
1180static void rcu_wake_cond(struct task_struct *t, int status)
1181{
1182 /*
1183 * If the thread is yielding, only wake it when this
1184 * is invoked from idle
1185 */
1186 if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
1187 wake_up_process(t);
1188}
1189
1190/*
1191 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1192 * or ->boost_tasks, advancing the pointer to the next task in the
1193 * ->blkd_tasks list.
1194 *
1195 * Note that irqs must be enabled: boosting the task can block.
1196 * Returns 1 if there are more tasks needing to be boosted.
1197 */
1198static int rcu_boost(struct rcu_node *rnp)
1199{
1200 unsigned long flags;
1201 struct rt_mutex mtx;
1202 struct task_struct *t;
1203 struct list_head *tb;
1204
1205 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL)
1206 return 0; /* Nothing left to boost. */
1207
1208 raw_spin_lock_irqsave(&rnp->lock, flags);
1209 smp_mb__after_unlock_lock();
1210
1211 /*
1212 * Recheck under the lock: all tasks in need of boosting
1213 * might exit their RCU read-side critical sections on their own.
1214 */
1215 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1216 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1217 return 0;
1218 }
1219
1220 /*
1221 * Preferentially boost tasks blocking expedited grace periods.
1222 * This cannot starve the normal grace periods because a second
1223 * expedited grace period must boost all blocked tasks, including
1224 * those blocking the pre-existing normal grace period.
1225 */
1226 if (rnp->exp_tasks != NULL) {
1227 tb = rnp->exp_tasks;
1228 rnp->n_exp_boosts++;
1229 } else {
1230 tb = rnp->boost_tasks;
1231 rnp->n_normal_boosts++;
1232 }
1233 rnp->n_tasks_boosted++;
1234
1235 /*
1236 * We boost task t by manufacturing an rt_mutex that appears to
1237 * be held by task t. We leave a pointer to that rt_mutex where
1238 * task t can find it, and task t will release the mutex when it
1239 * exits its outermost RCU read-side critical section. Then
1240 * simply acquiring this artificial rt_mutex will boost task
1241 * t's priority. (Thanks to tglx for suggesting this approach!)
1242 *
1243 * Note that task t must acquire rnp->lock to remove itself from
1244 * the ->blkd_tasks list, which it will do from exit() if from
1245 * nowhere else. We therefore are guaranteed that task t will
1246 * stay around at least until we drop rnp->lock. Note that
1247 * rnp->lock also resolves races between our priority boosting
1248 * and task t's exiting its outermost RCU read-side critical
1249 * section.
1250 */
1251 t = container_of(tb, struct task_struct, rcu_node_entry);
1252 rt_mutex_init_proxy_locked(&mtx, t);
1253 t->rcu_boost_mutex = &mtx;
1254 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1255 rt_mutex_lock(&mtx); /* Side effect: boosts task t's priority. */
1256 rt_mutex_unlock(&mtx); /* Keep lockdep happy. */
1257
1258 return ACCESS_ONCE(rnp->exp_tasks) != NULL ||
1259 ACCESS_ONCE(rnp->boost_tasks) != NULL;
1260}
1261
1262/*
1263 * Priority-boosting kthread. One per leaf rcu_node and one for the
1264 * root rcu_node.
1265 */
1266static int rcu_boost_kthread(void *arg)
1267{
1268 struct rcu_node *rnp = (struct rcu_node *)arg;
1269 int spincnt = 0;
1270 int more2boost;
1271
1272 trace_rcu_utilization(TPS("Start boost kthread@init"));
1273 for (;;) {
1274 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1275 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1276 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1277 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1278 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1279 more2boost = rcu_boost(rnp);
1280 if (more2boost)
1281 spincnt++;
1282 else
1283 spincnt = 0;
1284 if (spincnt > 10) {
1285 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1286 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1287 schedule_timeout_interruptible(2);
1288 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1289 spincnt = 0;
1290 }
1291 }
1292 /* NOTREACHED */
1293 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1294 return 0;
1295}
1296
1297/*
1298 * Check to see if it is time to start boosting RCU readers that are
1299 * blocking the current grace period, and, if so, tell the per-rcu_node
1300 * kthread to start boosting them. If there is an expedited grace
1301 * period in progress, it is always time to boost.
1302 *
1303 * The caller must hold rnp->lock, which this function releases.
1304 * The ->boost_kthread_task is immortal, so we don't need to worry
1305 * about it going away.
1306 */
1307static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1308{
1309 struct task_struct *t;
1310
1311 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1312 rnp->n_balk_exp_gp_tasks++;
1313 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1314 return;
1315 }
1316 if (rnp->exp_tasks != NULL ||
1317 (rnp->gp_tasks != NULL &&
1318 rnp->boost_tasks == NULL &&
1319 rnp->qsmask == 0 &&
1320 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1321 if (rnp->exp_tasks == NULL)
1322 rnp->boost_tasks = rnp->gp_tasks;
1323 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1324 t = rnp->boost_kthread_task;
1325 if (t)
1326 rcu_wake_cond(t, rnp->boost_kthread_status);
1327 } else {
1328 rcu_initiate_boost_trace(rnp);
1329 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1330 }
1331}
1332
1333/*
1334 * Wake up the per-CPU kthread to invoke RCU callbacks.
1335 */
1336static void invoke_rcu_callbacks_kthread(void)
1337{
1338 unsigned long flags;
1339
1340 local_irq_save(flags);
1341 __this_cpu_write(rcu_cpu_has_work, 1);
1342 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1343 current != __this_cpu_read(rcu_cpu_kthread_task)) {
1344 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1345 __this_cpu_read(rcu_cpu_kthread_status));
1346 }
1347 local_irq_restore(flags);
1348}
1349
1350/*
1351 * Is the current CPU running the RCU-callbacks kthread?
1352 * Caller must have preemption disabled.
1353 */
1354static bool rcu_is_callbacks_kthread(void)
1355{
1356 return __this_cpu_read(rcu_cpu_kthread_task) == current;
1357}
1358
1359#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1360
1361/*
1362 * Do priority-boost accounting for the start of a new grace period.
1363 */
1364static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1365{
1366 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1367}
1368
1369/*
1370 * Create an RCU-boost kthread for the specified node if one does not
1371 * already exist. We only create this kthread for preemptible RCU.
1372 * Returns zero if all is well, a negated errno otherwise.
1373 */
1374static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1375 struct rcu_node *rnp)
1376{
1377 int rnp_index = rnp - &rsp->node[0];
1378 unsigned long flags;
1379 struct sched_param sp;
1380 struct task_struct *t;
1381
1382 if (&rcu_preempt_state != rsp)
1383 return 0;
1384
1385 if (!rcu_scheduler_fully_active || rnp->qsmaskinit == 0)
1386 return 0;
1387
1388 rsp->boost = 1;
1389 if (rnp->boost_kthread_task != NULL)
1390 return 0;
1391 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1392 "rcub/%d", rnp_index);
1393 if (IS_ERR(t))
1394 return PTR_ERR(t);
1395 raw_spin_lock_irqsave(&rnp->lock, flags);
1396 smp_mb__after_unlock_lock();
1397 rnp->boost_kthread_task = t;
1398 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1399 sp.sched_priority = RCU_BOOST_PRIO;
1400 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1401 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1402 return 0;
1403}
1404
1405static void rcu_kthread_do_work(void)
1406{
1407 rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1408 rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1409 rcu_preempt_do_callbacks();
1410}
1411
1412static void rcu_cpu_kthread_setup(unsigned int cpu)
1413{
1414 struct sched_param sp;
1415
1416 sp.sched_priority = RCU_KTHREAD_PRIO;
1417 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1418}
1419
1420static void rcu_cpu_kthread_park(unsigned int cpu)
1421{
1422 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1423}
1424
1425static int rcu_cpu_kthread_should_run(unsigned int cpu)
1426{
1427 return __this_cpu_read(rcu_cpu_has_work);
1428}
1429
1430/*
1431 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1432 * RCU softirq used in flavors and configurations of RCU that do not
1433 * support RCU priority boosting.
1434 */
1435static void rcu_cpu_kthread(unsigned int cpu)
1436{
1437 unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1438 char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1439 int spincnt;
1440
1441 for (spincnt = 0; spincnt < 10; spincnt++) {
1442 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1443 local_bh_disable();
1444 *statusp = RCU_KTHREAD_RUNNING;
1445 this_cpu_inc(rcu_cpu_kthread_loops);
1446 local_irq_disable();
1447 work = *workp;
1448 *workp = 0;
1449 local_irq_enable();
1450 if (work)
1451 rcu_kthread_do_work();
1452 local_bh_enable();
1453 if (*workp == 0) {
1454 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1455 *statusp = RCU_KTHREAD_WAITING;
1456 return;
1457 }
1458 }
1459 *statusp = RCU_KTHREAD_YIELDING;
1460 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1461 schedule_timeout_interruptible(2);
1462 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1463 *statusp = RCU_KTHREAD_WAITING;
1464}
1465
1466/*
1467 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1468 * served by the rcu_node in question. The CPU hotplug lock is still
1469 * held, so the value of rnp->qsmaskinit will be stable.
1470 *
1471 * We don't include outgoingcpu in the affinity set, use -1 if there is
1472 * no outgoing CPU. If there are no CPUs left in the affinity set,
1473 * this function allows the kthread to execute on any CPU.
1474 */
1475static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1476{
1477 struct task_struct *t = rnp->boost_kthread_task;
1478 unsigned long mask = rnp->qsmaskinit;
1479 cpumask_var_t cm;
1480 int cpu;
1481
1482 if (!t)
1483 return;
1484 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1485 return;
1486 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1487 if ((mask & 0x1) && cpu != outgoingcpu)
1488 cpumask_set_cpu(cpu, cm);
1489 if (cpumask_weight(cm) == 0) {
1490 cpumask_setall(cm);
1491 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++)
1492 cpumask_clear_cpu(cpu, cm);
1493 WARN_ON_ONCE(cpumask_weight(cm) == 0);
1494 }
1495 set_cpus_allowed_ptr(t, cm);
1496 free_cpumask_var(cm);
1497}
1498
1499static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1500 .store = &rcu_cpu_kthread_task,
1501 .thread_should_run = rcu_cpu_kthread_should_run,
1502 .thread_fn = rcu_cpu_kthread,
1503 .thread_comm = "rcuc/%u",
1504 .setup = rcu_cpu_kthread_setup,
1505 .park = rcu_cpu_kthread_park,
1506};
1507
1508/*
1509 * Spawn all kthreads -- called as soon as the scheduler is running.
1510 */
1511static int __init rcu_spawn_kthreads(void)
1512{
1513 struct rcu_node *rnp;
1514 int cpu;
1515
1516 rcu_scheduler_fully_active = 1;
1517 for_each_possible_cpu(cpu)
1518 per_cpu(rcu_cpu_has_work, cpu) = 0;
1519 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1520 rnp = rcu_get_root(rcu_state);
1521 (void)rcu_spawn_one_boost_kthread(rcu_state, rnp);
1522 if (NUM_RCU_NODES > 1) {
1523 rcu_for_each_leaf_node(rcu_state, rnp)
1524 (void)rcu_spawn_one_boost_kthread(rcu_state, rnp);
1525 }
1526 return 0;
1527}
1528early_initcall(rcu_spawn_kthreads);
1529
1530static void rcu_prepare_kthreads(int cpu)
1531{
1532 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
1533 struct rcu_node *rnp = rdp->mynode;
1534
1535 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1536 if (rcu_scheduler_fully_active)
1537 (void)rcu_spawn_one_boost_kthread(rcu_state, rnp);
1538}
1539
1540#else /* #ifdef CONFIG_RCU_BOOST */
1541
1542static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1543{
1544 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1545}
1546
1547static void invoke_rcu_callbacks_kthread(void)
1548{
1549 WARN_ON_ONCE(1);
1550}
1551
1552static bool rcu_is_callbacks_kthread(void)
1553{
1554 return false;
1555}
1556
1557static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1558{
1559}
1560
1561static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1562{
1563}
1564
1565static int __init rcu_scheduler_really_started(void)
1566{
1567 rcu_scheduler_fully_active = 1;
1568 return 0;
1569}
1570early_initcall(rcu_scheduler_really_started);
1571
1572static void rcu_prepare_kthreads(int cpu)
1573{
1574}
1575
1576#endif /* #else #ifdef CONFIG_RCU_BOOST */
1577
1578#if !defined(CONFIG_RCU_FAST_NO_HZ)
1579
1580/*
1581 * Check to see if any future RCU-related work will need to be done
1582 * by the current CPU, even if none need be done immediately, returning
1583 * 1 if so. This function is part of the RCU implementation; it is -not-
1584 * an exported member of the RCU API.
1585 *
1586 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1587 * any flavor of RCU.
1588 */
1589#ifndef CONFIG_RCU_NOCB_CPU_ALL
1590int rcu_needs_cpu(int cpu, unsigned long *delta_jiffies)
1591{
1592 *delta_jiffies = ULONG_MAX;
1593 return rcu_cpu_has_callbacks(cpu, NULL);
1594}
1595#endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1596
1597/*
1598 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1599 * after it.
1600 */
1601static void rcu_cleanup_after_idle(int cpu)
1602{
1603}
1604
1605/*
1606 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1607 * is nothing.
1608 */
1609static void rcu_prepare_for_idle(int cpu)
1610{
1611}
1612
1613/*
1614 * Don't bother keeping a running count of the number of RCU callbacks
1615 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1616 */
1617static void rcu_idle_count_callbacks_posted(void)
1618{
1619}
1620
1621#else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1622
1623/*
1624 * This code is invoked when a CPU goes idle, at which point we want
1625 * to have the CPU do everything required for RCU so that it can enter
1626 * the energy-efficient dyntick-idle mode. This is handled by a
1627 * state machine implemented by rcu_prepare_for_idle() below.
1628 *
1629 * The following three proprocessor symbols control this state machine:
1630 *
1631 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1632 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1633 * is sized to be roughly one RCU grace period. Those energy-efficiency
1634 * benchmarkers who might otherwise be tempted to set this to a large
1635 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1636 * system. And if you are -that- concerned about energy efficiency,
1637 * just power the system down and be done with it!
1638 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1639 * permitted to sleep in dyntick-idle mode with only lazy RCU
1640 * callbacks pending. Setting this too high can OOM your system.
1641 *
1642 * The values below work well in practice. If future workloads require
1643 * adjustment, they can be converted into kernel config parameters, though
1644 * making the state machine smarter might be a better option.
1645 */
1646#define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1647#define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1648
1649static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1650module_param(rcu_idle_gp_delay, int, 0644);
1651static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1652module_param(rcu_idle_lazy_gp_delay, int, 0644);
1653
1654extern int tick_nohz_active;
1655
1656/*
1657 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1658 * only if it has been awhile since the last time we did so. Afterwards,
1659 * if there are any callbacks ready for immediate invocation, return true.
1660 */
1661static bool __maybe_unused rcu_try_advance_all_cbs(void)
1662{
1663 bool cbs_ready = false;
1664 struct rcu_data *rdp;
1665 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1666 struct rcu_node *rnp;
1667 struct rcu_state *rsp;
1668
1669 /* Exit early if we advanced recently. */
1670 if (jiffies == rdtp->last_advance_all)
1671 return 0;
1672 rdtp->last_advance_all = jiffies;
1673
1674 for_each_rcu_flavor(rsp) {
1675 rdp = this_cpu_ptr(rsp->rda);
1676 rnp = rdp->mynode;
1677
1678 /*
1679 * Don't bother checking unless a grace period has
1680 * completed since we last checked and there are
1681 * callbacks not yet ready to invoke.
1682 */
1683 if (rdp->completed != rnp->completed &&
1684 rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1685 note_gp_changes(rsp, rdp);
1686
1687 if (cpu_has_callbacks_ready_to_invoke(rdp))
1688 cbs_ready = true;
1689 }
1690 return cbs_ready;
1691}
1692
1693/*
1694 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1695 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1696 * caller to set the timeout based on whether or not there are non-lazy
1697 * callbacks.
1698 *
1699 * The caller must have disabled interrupts.
1700 */
1701#ifndef CONFIG_RCU_NOCB_CPU_ALL
1702int rcu_needs_cpu(int cpu, unsigned long *dj)
1703{
1704 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1705
1706 /* Snapshot to detect later posting of non-lazy callback. */
1707 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1708
1709 /* If no callbacks, RCU doesn't need the CPU. */
1710 if (!rcu_cpu_has_callbacks(cpu, &rdtp->all_lazy)) {
1711 *dj = ULONG_MAX;
1712 return 0;
1713 }
1714
1715 /* Attempt to advance callbacks. */
1716 if (rcu_try_advance_all_cbs()) {
1717 /* Some ready to invoke, so initiate later invocation. */
1718 invoke_rcu_core();
1719 return 1;
1720 }
1721 rdtp->last_accelerate = jiffies;
1722
1723 /* Request timer delay depending on laziness, and round. */
1724 if (!rdtp->all_lazy) {
1725 *dj = round_up(rcu_idle_gp_delay + jiffies,
1726 rcu_idle_gp_delay) - jiffies;
1727 } else {
1728 *dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1729 }
1730 return 0;
1731}
1732#endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1733
1734/*
1735 * Prepare a CPU for idle from an RCU perspective. The first major task
1736 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1737 * The second major task is to check to see if a non-lazy callback has
1738 * arrived at a CPU that previously had only lazy callbacks. The third
1739 * major task is to accelerate (that is, assign grace-period numbers to)
1740 * any recently arrived callbacks.
1741 *
1742 * The caller must have disabled interrupts.
1743 */
1744static void rcu_prepare_for_idle(int cpu)
1745{
1746#ifndef CONFIG_RCU_NOCB_CPU_ALL
1747 struct rcu_data *rdp;
1748 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1749 struct rcu_node *rnp;
1750 struct rcu_state *rsp;
1751 int tne;
1752
1753 /* Handle nohz enablement switches conservatively. */
1754 tne = ACCESS_ONCE(tick_nohz_active);
1755 if (tne != rdtp->tick_nohz_enabled_snap) {
1756 if (rcu_cpu_has_callbacks(cpu, NULL))
1757 invoke_rcu_core(); /* force nohz to see update. */
1758 rdtp->tick_nohz_enabled_snap = tne;
1759 return;
1760 }
1761 if (!tne)
1762 return;
1763
1764 /* If this is a no-CBs CPU, no callbacks, just return. */
1765 if (rcu_is_nocb_cpu(cpu))
1766 return;
1767
1768 /*
1769 * If a non-lazy callback arrived at a CPU having only lazy
1770 * callbacks, invoke RCU core for the side-effect of recalculating
1771 * idle duration on re-entry to idle.
1772 */
1773 if (rdtp->all_lazy &&
1774 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1775 rdtp->all_lazy = false;
1776 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1777 invoke_rcu_core();
1778 return;
1779 }
1780
1781 /*
1782 * If we have not yet accelerated this jiffy, accelerate all
1783 * callbacks on this CPU.
1784 */
1785 if (rdtp->last_accelerate == jiffies)
1786 return;
1787 rdtp->last_accelerate = jiffies;
1788 for_each_rcu_flavor(rsp) {
1789 rdp = per_cpu_ptr(rsp->rda, cpu);
1790 if (!*rdp->nxttail[RCU_DONE_TAIL])
1791 continue;
1792 rnp = rdp->mynode;
1793 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1794 smp_mb__after_unlock_lock();
1795 rcu_accelerate_cbs(rsp, rnp, rdp);
1796 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1797 }
1798#endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1799}
1800
1801/*
1802 * Clean up for exit from idle. Attempt to advance callbacks based on
1803 * any grace periods that elapsed while the CPU was idle, and if any
1804 * callbacks are now ready to invoke, initiate invocation.
1805 */
1806static void rcu_cleanup_after_idle(int cpu)
1807{
1808#ifndef CONFIG_RCU_NOCB_CPU_ALL
1809 if (rcu_is_nocb_cpu(cpu))
1810 return;
1811 if (rcu_try_advance_all_cbs())
1812 invoke_rcu_core();
1813#endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1814}
1815
1816/*
1817 * Keep a running count of the number of non-lazy callbacks posted
1818 * on this CPU. This running counter (which is never decremented) allows
1819 * rcu_prepare_for_idle() to detect when something out of the idle loop
1820 * posts a callback, even if an equal number of callbacks are invoked.
1821 * Of course, callbacks should only be posted from within a trace event
1822 * designed to be called from idle or from within RCU_NONIDLE().
1823 */
1824static void rcu_idle_count_callbacks_posted(void)
1825{
1826 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1827}
1828
1829/*
1830 * Data for flushing lazy RCU callbacks at OOM time.
1831 */
1832static atomic_t oom_callback_count;
1833static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1834
1835/*
1836 * RCU OOM callback -- decrement the outstanding count and deliver the
1837 * wake-up if we are the last one.
1838 */
1839static void rcu_oom_callback(struct rcu_head *rhp)
1840{
1841 if (atomic_dec_and_test(&oom_callback_count))
1842 wake_up(&oom_callback_wq);
1843}
1844
1845/*
1846 * Post an rcu_oom_notify callback on the current CPU if it has at
1847 * least one lazy callback. This will unnecessarily post callbacks
1848 * to CPUs that already have a non-lazy callback at the end of their
1849 * callback list, but this is an infrequent operation, so accept some
1850 * extra overhead to keep things simple.
1851 */
1852static void rcu_oom_notify_cpu(void *unused)
1853{
1854 struct rcu_state *rsp;
1855 struct rcu_data *rdp;
1856
1857 for_each_rcu_flavor(rsp) {
1858 rdp = __this_cpu_ptr(rsp->rda);
1859 if (rdp->qlen_lazy != 0) {
1860 atomic_inc(&oom_callback_count);
1861 rsp->call(&rdp->oom_head, rcu_oom_callback);
1862 }
1863 }
1864}
1865
1866/*
1867 * If low on memory, ensure that each CPU has a non-lazy callback.
1868 * This will wake up CPUs that have only lazy callbacks, in turn
1869 * ensuring that they free up the corresponding memory in a timely manner.
1870 * Because an uncertain amount of memory will be freed in some uncertain
1871 * timeframe, we do not claim to have freed anything.
1872 */
1873static int rcu_oom_notify(struct notifier_block *self,
1874 unsigned long notused, void *nfreed)
1875{
1876 int cpu;
1877
1878 /* Wait for callbacks from earlier instance to complete. */
1879 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1880 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1881
1882 /*
1883 * Prevent premature wakeup: ensure that all increments happen
1884 * before there is a chance of the counter reaching zero.
1885 */
1886 atomic_set(&oom_callback_count, 1);
1887
1888 get_online_cpus();
1889 for_each_online_cpu(cpu) {
1890 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1891 cond_resched();
1892 }
1893 put_online_cpus();
1894
1895 /* Unconditionally decrement: no need to wake ourselves up. */
1896 atomic_dec(&oom_callback_count);
1897
1898 return NOTIFY_OK;
1899}
1900
1901static struct notifier_block rcu_oom_nb = {
1902 .notifier_call = rcu_oom_notify
1903};
1904
1905static int __init rcu_register_oom_notifier(void)
1906{
1907 register_oom_notifier(&rcu_oom_nb);
1908 return 0;
1909}
1910early_initcall(rcu_register_oom_notifier);
1911
1912#endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1913
1914#ifdef CONFIG_RCU_CPU_STALL_INFO
1915
1916#ifdef CONFIG_RCU_FAST_NO_HZ
1917
1918static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1919{
1920 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1921 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1922
1923 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1924 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1925 ulong2long(nlpd),
1926 rdtp->all_lazy ? 'L' : '.',
1927 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1928}
1929
1930#else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1931
1932static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1933{
1934 *cp = '\0';
1935}
1936
1937#endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1938
1939/* Initiate the stall-info list. */
1940static void print_cpu_stall_info_begin(void)
1941{
1942 pr_cont("\n");
1943}
1944
1945/*
1946 * Print out diagnostic information for the specified stalled CPU.
1947 *
1948 * If the specified CPU is aware of the current RCU grace period
1949 * (flavor specified by rsp), then print the number of scheduling
1950 * clock interrupts the CPU has taken during the time that it has
1951 * been aware. Otherwise, print the number of RCU grace periods
1952 * that this CPU is ignorant of, for example, "1" if the CPU was
1953 * aware of the previous grace period.
1954 *
1955 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1956 */
1957static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1958{
1959 char fast_no_hz[72];
1960 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1961 struct rcu_dynticks *rdtp = rdp->dynticks;
1962 char *ticks_title;
1963 unsigned long ticks_value;
1964
1965 if (rsp->gpnum == rdp->gpnum) {
1966 ticks_title = "ticks this GP";
1967 ticks_value = rdp->ticks_this_gp;
1968 } else {
1969 ticks_title = "GPs behind";
1970 ticks_value = rsp->gpnum - rdp->gpnum;
1971 }
1972 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1973 pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u %s\n",
1974 cpu, ticks_value, ticks_title,
1975 atomic_read(&rdtp->dynticks) & 0xfff,
1976 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1977 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1978 fast_no_hz);
1979}
1980
1981/* Terminate the stall-info list. */
1982static void print_cpu_stall_info_end(void)
1983{
1984 pr_err("\t");
1985}
1986
1987/* Zero ->ticks_this_gp for all flavors of RCU. */
1988static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1989{
1990 rdp->ticks_this_gp = 0;
1991 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1992}
1993
1994/* Increment ->ticks_this_gp for all flavors of RCU. */
1995static void increment_cpu_stall_ticks(void)
1996{
1997 struct rcu_state *rsp;
1998
1999 for_each_rcu_flavor(rsp)
2000 __this_cpu_ptr(rsp->rda)->ticks_this_gp++;
2001}
2002
2003#else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
2004
2005static void print_cpu_stall_info_begin(void)
2006{
2007 pr_cont(" {");
2008}
2009
2010static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
2011{
2012 pr_cont(" %d", cpu);
2013}
2014
2015static void print_cpu_stall_info_end(void)
2016{
2017 pr_cont("} ");
2018}
2019
2020static void zero_cpu_stall_ticks(struct rcu_data *rdp)
2021{
2022}
2023
2024static void increment_cpu_stall_ticks(void)
2025{
2026}
2027
2028#endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
2029
2030#ifdef CONFIG_RCU_NOCB_CPU
2031
2032/*
2033 * Offload callback processing from the boot-time-specified set of CPUs
2034 * specified by rcu_nocb_mask. For each CPU in the set, there is a
2035 * kthread created that pulls the callbacks from the corresponding CPU,
2036 * waits for a grace period to elapse, and invokes the callbacks.
2037 * The no-CBs CPUs do a wake_up() on their kthread when they insert
2038 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
2039 * has been specified, in which case each kthread actively polls its
2040 * CPU. (Which isn't so great for energy efficiency, but which does
2041 * reduce RCU's overhead on that CPU.)
2042 *
2043 * This is intended to be used in conjunction with Frederic Weisbecker's
2044 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
2045 * running CPU-bound user-mode computations.
2046 *
2047 * Offloading of callback processing could also in theory be used as
2048 * an energy-efficiency measure because CPUs with no RCU callbacks
2049 * queued are more aggressive about entering dyntick-idle mode.
2050 */
2051
2052
2053/* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
2054static int __init rcu_nocb_setup(char *str)
2055{
2056 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
2057 have_rcu_nocb_mask = true;
2058 cpulist_parse(str, rcu_nocb_mask);
2059 return 1;
2060}
2061__setup("rcu_nocbs=", rcu_nocb_setup);
2062
2063static int __init parse_rcu_nocb_poll(char *arg)
2064{
2065 rcu_nocb_poll = 1;
2066 return 0;
2067}
2068early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
2069
2070/*
2071 * Do any no-CBs CPUs need another grace period?
2072 *
2073 * Interrupts must be disabled. If the caller does not hold the root
2074 * rnp_node structure's ->lock, the results are advisory only.
2075 */
2076static int rcu_nocb_needs_gp(struct rcu_state *rsp)
2077{
2078 struct rcu_node *rnp = rcu_get_root(rsp);
2079
2080 return rnp->need_future_gp[(ACCESS_ONCE(rnp->completed) + 1) & 0x1];
2081}
2082
2083/*
2084 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
2085 * grace period.
2086 */
2087static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2088{
2089 wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]);
2090}
2091
2092/*
2093 * Set the root rcu_node structure's ->need_future_gp field
2094 * based on the sum of those of all rcu_node structures. This does
2095 * double-count the root rcu_node structure's requests, but this
2096 * is necessary to handle the possibility of a rcu_nocb_kthread()
2097 * having awakened during the time that the rcu_node structures
2098 * were being updated for the end of the previous grace period.
2099 */
2100static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2101{
2102 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
2103}
2104
2105static void rcu_init_one_nocb(struct rcu_node *rnp)
2106{
2107 init_waitqueue_head(&rnp->nocb_gp_wq[0]);
2108 init_waitqueue_head(&rnp->nocb_gp_wq[1]);
2109}
2110
2111#ifndef CONFIG_RCU_NOCB_CPU_ALL
2112/* Is the specified CPU a no-CPUs CPU? */
2113bool rcu_is_nocb_cpu(int cpu)
2114{
2115 if (have_rcu_nocb_mask)
2116 return cpumask_test_cpu(cpu, rcu_nocb_mask);
2117 return false;
2118}
2119#endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
2120
2121/*
2122 * Enqueue the specified string of rcu_head structures onto the specified
2123 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
2124 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
2125 * counts are supplied by rhcount and rhcount_lazy.
2126 *
2127 * If warranted, also wake up the kthread servicing this CPUs queues.
2128 */
2129static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
2130 struct rcu_head *rhp,
2131 struct rcu_head **rhtp,
2132 int rhcount, int rhcount_lazy,
2133 unsigned long flags)
2134{
2135 int len;
2136 struct rcu_head **old_rhpp;
2137 struct task_struct *t;
2138
2139 /* Enqueue the callback on the nocb list and update counts. */
2140 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
2141 ACCESS_ONCE(*old_rhpp) = rhp;
2142 atomic_long_add(rhcount, &rdp->nocb_q_count);
2143 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
2144
2145 /* If we are not being polled and there is a kthread, awaken it ... */
2146 t = ACCESS_ONCE(rdp->nocb_kthread);
2147 if (rcu_nocb_poll || !t) {
2148 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2149 TPS("WakeNotPoll"));
2150 return;
2151 }
2152 len = atomic_long_read(&rdp->nocb_q_count);
2153 if (old_rhpp == &rdp->nocb_head) {
2154 if (!irqs_disabled_flags(flags)) {
2155 wake_up(&rdp->nocb_wq); /* ... if queue was empty ... */
2156 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2157 TPS("WakeEmpty"));
2158 } else {
2159 rdp->nocb_defer_wakeup = true;
2160 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2161 TPS("WakeEmptyIsDeferred"));
2162 }
2163 rdp->qlen_last_fqs_check = 0;
2164 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
2165 wake_up_process(t); /* ... or if many callbacks queued. */
2166 rdp->qlen_last_fqs_check = LONG_MAX / 2;
2167 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeOvf"));
2168 } else {
2169 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
2170 }
2171 return;
2172}
2173
2174/*
2175 * This is a helper for __call_rcu(), which invokes this when the normal
2176 * callback queue is inoperable. If this is not a no-CBs CPU, this
2177 * function returns failure back to __call_rcu(), which can complain
2178 * appropriately.
2179 *
2180 * Otherwise, this function queues the callback where the corresponding
2181 * "rcuo" kthread can find it.
2182 */
2183static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2184 bool lazy, unsigned long flags)
2185{
2186
2187 if (!rcu_is_nocb_cpu(rdp->cpu))
2188 return 0;
2189 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
2190 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
2191 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
2192 (unsigned long)rhp->func,
2193 -atomic_long_read(&rdp->nocb_q_count_lazy),
2194 -atomic_long_read(&rdp->nocb_q_count));
2195 else
2196 trace_rcu_callback(rdp->rsp->name, rhp,
2197 -atomic_long_read(&rdp->nocb_q_count_lazy),
2198 -atomic_long_read(&rdp->nocb_q_count));
2199 return 1;
2200}
2201
2202/*
2203 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2204 * not a no-CBs CPU.
2205 */
2206static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2207 struct rcu_data *rdp,
2208 unsigned long flags)
2209{
2210 long ql = rsp->qlen;
2211 long qll = rsp->qlen_lazy;
2212
2213 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2214 if (!rcu_is_nocb_cpu(smp_processor_id()))
2215 return 0;
2216 rsp->qlen = 0;
2217 rsp->qlen_lazy = 0;
2218
2219 /* First, enqueue the donelist, if any. This preserves CB ordering. */
2220 if (rsp->orphan_donelist != NULL) {
2221 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
2222 rsp->orphan_donetail, ql, qll, flags);
2223 ql = qll = 0;
2224 rsp->orphan_donelist = NULL;
2225 rsp->orphan_donetail = &rsp->orphan_donelist;
2226 }
2227 if (rsp->orphan_nxtlist != NULL) {
2228 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
2229 rsp->orphan_nxttail, ql, qll, flags);
2230 ql = qll = 0;
2231 rsp->orphan_nxtlist = NULL;
2232 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2233 }
2234 return 1;
2235}
2236
2237/*
2238 * If necessary, kick off a new grace period, and either way wait
2239 * for a subsequent grace period to complete.
2240 */
2241static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2242{
2243 unsigned long c;
2244 bool d;
2245 unsigned long flags;
2246 struct rcu_node *rnp = rdp->mynode;
2247
2248 raw_spin_lock_irqsave(&rnp->lock, flags);
2249 smp_mb__after_unlock_lock();
2250 c = rcu_start_future_gp(rnp, rdp);
2251 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2252
2253 /*
2254 * Wait for the grace period. Do so interruptibly to avoid messing
2255 * up the load average.
2256 */
2257 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
2258 for (;;) {
2259 wait_event_interruptible(
2260 rnp->nocb_gp_wq[c & 0x1],
2261 (d = ULONG_CMP_GE(ACCESS_ONCE(rnp->completed), c)));
2262 if (likely(d))
2263 break;
2264 flush_signals(current);
2265 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
2266 }
2267 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
2268 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2269}
2270
2271/*
2272 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2273 * callbacks queued by the corresponding no-CBs CPU.
2274 */
2275static int rcu_nocb_kthread(void *arg)
2276{
2277 int c, cl;
2278 bool firsttime = 1;
2279 struct rcu_head *list;
2280 struct rcu_head *next;
2281 struct rcu_head **tail;
2282 struct rcu_data *rdp = arg;
2283
2284 /* Each pass through this loop invokes one batch of callbacks */
2285 for (;;) {
2286 /* If not polling, wait for next batch of callbacks. */
2287 if (!rcu_nocb_poll) {
2288 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2289 TPS("Sleep"));
2290 wait_event_interruptible(rdp->nocb_wq, rdp->nocb_head);
2291 /* Memory barrier provide by xchg() below. */
2292 } else if (firsttime) {
2293 firsttime = 0;
2294 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2295 TPS("Poll"));
2296 }
2297 list = ACCESS_ONCE(rdp->nocb_head);
2298 if (!list) {
2299 if (!rcu_nocb_poll)
2300 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2301 TPS("WokeEmpty"));
2302 schedule_timeout_interruptible(1);
2303 flush_signals(current);
2304 continue;
2305 }
2306 firsttime = 1;
2307 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2308 TPS("WokeNonEmpty"));
2309
2310 /*
2311 * Extract queued callbacks, update counts, and wait
2312 * for a grace period to elapse.
2313 */
2314 ACCESS_ONCE(rdp->nocb_head) = NULL;
2315 tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2316 c = atomic_long_xchg(&rdp->nocb_q_count, 0);
2317 cl = atomic_long_xchg(&rdp->nocb_q_count_lazy, 0);
2318 ACCESS_ONCE(rdp->nocb_p_count) += c;
2319 ACCESS_ONCE(rdp->nocb_p_count_lazy) += cl;
2320 rcu_nocb_wait_gp(rdp);
2321
2322 /* Each pass through the following loop invokes a callback. */
2323 trace_rcu_batch_start(rdp->rsp->name, cl, c, -1);
2324 c = cl = 0;
2325 while (list) {
2326 next = list->next;
2327 /* Wait for enqueuing to complete, if needed. */
2328 while (next == NULL && &list->next != tail) {
2329 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2330 TPS("WaitQueue"));
2331 schedule_timeout_interruptible(1);
2332 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2333 TPS("WokeQueue"));
2334 next = list->next;
2335 }
2336 debug_rcu_head_unqueue(list);
2337 local_bh_disable();
2338 if (__rcu_reclaim(rdp->rsp->name, list))
2339 cl++;
2340 c++;
2341 local_bh_enable();
2342 list = next;
2343 }
2344 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2345 ACCESS_ONCE(rdp->nocb_p_count) -= c;
2346 ACCESS_ONCE(rdp->nocb_p_count_lazy) -= cl;
2347 rdp->n_nocbs_invoked += c;
2348 }
2349 return 0;
2350}
2351
2352/* Is a deferred wakeup of rcu_nocb_kthread() required? */
2353static bool rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2354{
2355 return ACCESS_ONCE(rdp->nocb_defer_wakeup);
2356}
2357
2358/* Do a deferred wakeup of rcu_nocb_kthread(). */
2359static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2360{
2361 if (!rcu_nocb_need_deferred_wakeup(rdp))
2362 return;
2363 ACCESS_ONCE(rdp->nocb_defer_wakeup) = false;
2364 wake_up(&rdp->nocb_wq);
2365 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWakeEmpty"));
2366}
2367
2368/* Initialize per-rcu_data variables for no-CBs CPUs. */
2369static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2370{
2371 rdp->nocb_tail = &rdp->nocb_head;
2372 init_waitqueue_head(&rdp->nocb_wq);
2373}
2374
2375/* Create a kthread for each RCU flavor for each no-CBs CPU. */
2376static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp)
2377{
2378 int cpu;
2379 struct rcu_data *rdp;
2380 struct task_struct *t;
2381
2382 if (rcu_nocb_mask == NULL)
2383 return;
2384 for_each_cpu(cpu, rcu_nocb_mask) {
2385 rdp = per_cpu_ptr(rsp->rda, cpu);
2386 t = kthread_run(rcu_nocb_kthread, rdp,
2387 "rcuo%c/%d", rsp->abbr, cpu);
2388 BUG_ON(IS_ERR(t));
2389 ACCESS_ONCE(rdp->nocb_kthread) = t;
2390 }
2391}
2392
2393/* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2394static bool init_nocb_callback_list(struct rcu_data *rdp)
2395{
2396 if (rcu_nocb_mask == NULL ||
2397 !cpumask_test_cpu(rdp->cpu, rcu_nocb_mask))
2398 return false;
2399 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2400 return true;
2401}
2402
2403#else /* #ifdef CONFIG_RCU_NOCB_CPU */
2404
2405static int rcu_nocb_needs_gp(struct rcu_state *rsp)
2406{
2407 return 0;
2408}
2409
2410static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2411{
2412}
2413
2414static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2415{
2416}
2417
2418static void rcu_init_one_nocb(struct rcu_node *rnp)
2419{
2420}
2421
2422static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2423 bool lazy, unsigned long flags)
2424{
2425 return 0;
2426}
2427
2428static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2429 struct rcu_data *rdp,
2430 unsigned long flags)
2431{
2432 return 0;
2433}
2434
2435static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2436{
2437}
2438
2439static bool rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2440{
2441 return false;
2442}
2443
2444static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2445{
2446}
2447
2448static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp)
2449{
2450}
2451
2452static bool init_nocb_callback_list(struct rcu_data *rdp)
2453{
2454 return false;
2455}
2456
2457#endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2458
2459/*
2460 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2461 * arbitrarily long period of time with the scheduling-clock tick turned
2462 * off. RCU will be paying attention to this CPU because it is in the
2463 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2464 * machine because the scheduling-clock tick has been disabled. Therefore,
2465 * if an adaptive-ticks CPU is failing to respond to the current grace
2466 * period and has not be idle from an RCU perspective, kick it.
2467 */
2468static void rcu_kick_nohz_cpu(int cpu)
2469{
2470#ifdef CONFIG_NO_HZ_FULL
2471 if (tick_nohz_full_cpu(cpu))
2472 smp_send_reschedule(cpu);
2473#endif /* #ifdef CONFIG_NO_HZ_FULL */
2474}
2475
2476
2477#ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2478
2479/*
2480 * Define RCU flavor that holds sysidle state. This needs to be the
2481 * most active flavor of RCU.
2482 */
2483#ifdef CONFIG_PREEMPT_RCU
2484static struct rcu_state *rcu_sysidle_state = &rcu_preempt_state;
2485#else /* #ifdef CONFIG_PREEMPT_RCU */
2486static struct rcu_state *rcu_sysidle_state = &rcu_sched_state;
2487#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
2488
2489static int full_sysidle_state; /* Current system-idle state. */
2490#define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2491#define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2492#define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2493#define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2494#define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2495
2496/*
2497 * Invoked to note exit from irq or task transition to idle. Note that
2498 * usermode execution does -not- count as idle here! After all, we want
2499 * to detect full-system idle states, not RCU quiescent states and grace
2500 * periods. The caller must have disabled interrupts.
2501 */
2502static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq)
2503{
2504 unsigned long j;
2505
2506 /* Adjust nesting, check for fully idle. */
2507 if (irq) {
2508 rdtp->dynticks_idle_nesting--;
2509 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2510 if (rdtp->dynticks_idle_nesting != 0)
2511 return; /* Still not fully idle. */
2512 } else {
2513 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2514 DYNTICK_TASK_NEST_VALUE) {
2515 rdtp->dynticks_idle_nesting = 0;
2516 } else {
2517 rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2518 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2519 return; /* Still not fully idle. */
2520 }
2521 }
2522
2523 /* Record start of fully idle period. */
2524 j = jiffies;
2525 ACCESS_ONCE(rdtp->dynticks_idle_jiffies) = j;
2526 smp_mb__before_atomic_inc();
2527 atomic_inc(&rdtp->dynticks_idle);
2528 smp_mb__after_atomic_inc();
2529 WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2530}
2531
2532/*
2533 * Unconditionally force exit from full system-idle state. This is
2534 * invoked when a normal CPU exits idle, but must be called separately
2535 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2536 * is that the timekeeping CPU is permitted to take scheduling-clock
2537 * interrupts while the system is in system-idle state, and of course
2538 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2539 * interrupt from any other type of interrupt.
2540 */
2541void rcu_sysidle_force_exit(void)
2542{
2543 int oldstate = ACCESS_ONCE(full_sysidle_state);
2544 int newoldstate;
2545
2546 /*
2547 * Each pass through the following loop attempts to exit full
2548 * system-idle state. If contention proves to be a problem,
2549 * a trylock-based contention tree could be used here.
2550 */
2551 while (oldstate > RCU_SYSIDLE_SHORT) {
2552 newoldstate = cmpxchg(&full_sysidle_state,
2553 oldstate, RCU_SYSIDLE_NOT);
2554 if (oldstate == newoldstate &&
2555 oldstate == RCU_SYSIDLE_FULL_NOTED) {
2556 rcu_kick_nohz_cpu(tick_do_timer_cpu);
2557 return; /* We cleared it, done! */
2558 }
2559 oldstate = newoldstate;
2560 }
2561 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2562}
2563
2564/*
2565 * Invoked to note entry to irq or task transition from idle. Note that
2566 * usermode execution does -not- count as idle here! The caller must
2567 * have disabled interrupts.
2568 */
2569static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq)
2570{
2571 /* Adjust nesting, check for already non-idle. */
2572 if (irq) {
2573 rdtp->dynticks_idle_nesting++;
2574 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2575 if (rdtp->dynticks_idle_nesting != 1)
2576 return; /* Already non-idle. */
2577 } else {
2578 /*
2579 * Allow for irq misnesting. Yes, it really is possible
2580 * to enter an irq handler then never leave it, and maybe
2581 * also vice versa. Handle both possibilities.
2582 */
2583 if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2584 rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2585 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2586 return; /* Already non-idle. */
2587 } else {
2588 rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2589 }
2590 }
2591
2592 /* Record end of idle period. */
2593 smp_mb__before_atomic_inc();
2594 atomic_inc(&rdtp->dynticks_idle);
2595 smp_mb__after_atomic_inc();
2596 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2597
2598 /*
2599 * If we are the timekeeping CPU, we are permitted to be non-idle
2600 * during a system-idle state. This must be the case, because
2601 * the timekeeping CPU has to take scheduling-clock interrupts
2602 * during the time that the system is transitioning to full
2603 * system-idle state. This means that the timekeeping CPU must
2604 * invoke rcu_sysidle_force_exit() directly if it does anything
2605 * more than take a scheduling-clock interrupt.
2606 */
2607 if (smp_processor_id() == tick_do_timer_cpu)
2608 return;
2609
2610 /* Update system-idle state: We are clearly no longer fully idle! */
2611 rcu_sysidle_force_exit();
2612}
2613
2614/*
2615 * Check to see if the current CPU is idle. Note that usermode execution
2616 * does not count as idle. The caller must have disabled interrupts.
2617 */
2618static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2619 unsigned long *maxj)
2620{
2621 int cur;
2622 unsigned long j;
2623 struct rcu_dynticks *rdtp = rdp->dynticks;
2624
2625 /*
2626 * If some other CPU has already reported non-idle, if this is
2627 * not the flavor of RCU that tracks sysidle state, or if this
2628 * is an offline or the timekeeping CPU, nothing to do.
2629 */
2630 if (!*isidle || rdp->rsp != rcu_sysidle_state ||
2631 cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2632 return;
2633 if (rcu_gp_in_progress(rdp->rsp))
2634 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2635
2636 /* Pick up current idle and NMI-nesting counter and check. */
2637 cur = atomic_read(&rdtp->dynticks_idle);
2638 if (cur & 0x1) {
2639 *isidle = false; /* We are not idle! */
2640 return;
2641 }
2642 smp_mb(); /* Read counters before timestamps. */
2643
2644 /* Pick up timestamps. */
2645 j = ACCESS_ONCE(rdtp->dynticks_idle_jiffies);
2646 /* If this CPU entered idle more recently, update maxj timestamp. */
2647 if (ULONG_CMP_LT(*maxj, j))
2648 *maxj = j;
2649}
2650
2651/*
2652 * Is this the flavor of RCU that is handling full-system idle?
2653 */
2654static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2655{
2656 return rsp == rcu_sysidle_state;
2657}
2658
2659/*
2660 * Bind the grace-period kthread for the sysidle flavor of RCU to the
2661 * timekeeping CPU.
2662 */
2663static void rcu_bind_gp_kthread(void)
2664{
2665 int cpu = ACCESS_ONCE(tick_do_timer_cpu);
2666
2667 if (cpu < 0 || cpu >= nr_cpu_ids)
2668 return;
2669 if (raw_smp_processor_id() != cpu)
2670 set_cpus_allowed_ptr(current, cpumask_of(cpu));
2671}
2672
2673/*
2674 * Return a delay in jiffies based on the number of CPUs, rcu_node
2675 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2676 * systems more time to transition to full-idle state in order to
2677 * avoid the cache thrashing that otherwise occur on the state variable.
2678 * Really small systems (less than a couple of tens of CPUs) should
2679 * instead use a single global atomically incremented counter, and later
2680 * versions of this will automatically reconfigure themselves accordingly.
2681 */
2682static unsigned long rcu_sysidle_delay(void)
2683{
2684 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2685 return 0;
2686 return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2687}
2688
2689/*
2690 * Advance the full-system-idle state. This is invoked when all of
2691 * the non-timekeeping CPUs are idle.
2692 */
2693static void rcu_sysidle(unsigned long j)
2694{
2695 /* Check the current state. */
2696 switch (ACCESS_ONCE(full_sysidle_state)) {
2697 case RCU_SYSIDLE_NOT:
2698
2699 /* First time all are idle, so note a short idle period. */
2700 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_SHORT;
2701 break;
2702
2703 case RCU_SYSIDLE_SHORT:
2704
2705 /*
2706 * Idle for a bit, time to advance to next state?
2707 * cmpxchg failure means race with non-idle, let them win.
2708 */
2709 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2710 (void)cmpxchg(&full_sysidle_state,
2711 RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2712 break;
2713
2714 case RCU_SYSIDLE_LONG:
2715
2716 /*
2717 * Do an additional check pass before advancing to full.
2718 * cmpxchg failure means race with non-idle, let them win.
2719 */
2720 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2721 (void)cmpxchg(&full_sysidle_state,
2722 RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2723 break;
2724
2725 default:
2726 break;
2727 }
2728}
2729
2730/*
2731 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2732 * back to the beginning.
2733 */
2734static void rcu_sysidle_cancel(void)
2735{
2736 smp_mb();
2737 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_NOT;
2738}
2739
2740/*
2741 * Update the sysidle state based on the results of a force-quiescent-state
2742 * scan of the CPUs' dyntick-idle state.
2743 */
2744static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
2745 unsigned long maxj, bool gpkt)
2746{
2747 if (rsp != rcu_sysidle_state)
2748 return; /* Wrong flavor, ignore. */
2749 if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2750 return; /* Running state machine from timekeeping CPU. */
2751 if (isidle)
2752 rcu_sysidle(maxj); /* More idle! */
2753 else
2754 rcu_sysidle_cancel(); /* Idle is over. */
2755}
2756
2757/*
2758 * Wrapper for rcu_sysidle_report() when called from the grace-period
2759 * kthread's context.
2760 */
2761static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2762 unsigned long maxj)
2763{
2764 rcu_sysidle_report(rsp, isidle, maxj, true);
2765}
2766
2767/* Callback and function for forcing an RCU grace period. */
2768struct rcu_sysidle_head {
2769 struct rcu_head rh;
2770 int inuse;
2771};
2772
2773static void rcu_sysidle_cb(struct rcu_head *rhp)
2774{
2775 struct rcu_sysidle_head *rshp;
2776
2777 /*
2778 * The following memory barrier is needed to replace the
2779 * memory barriers that would normally be in the memory
2780 * allocator.
2781 */
2782 smp_mb(); /* grace period precedes setting inuse. */
2783
2784 rshp = container_of(rhp, struct rcu_sysidle_head, rh);
2785 ACCESS_ONCE(rshp->inuse) = 0;
2786}
2787
2788/*
2789 * Check to see if the system is fully idle, other than the timekeeping CPU.
2790 * The caller must have disabled interrupts.
2791 */
2792bool rcu_sys_is_idle(void)
2793{
2794 static struct rcu_sysidle_head rsh;
2795 int rss = ACCESS_ONCE(full_sysidle_state);
2796
2797 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
2798 return false;
2799
2800 /* Handle small-system case by doing a full scan of CPUs. */
2801 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
2802 int oldrss = rss - 1;
2803
2804 /*
2805 * One pass to advance to each state up to _FULL.
2806 * Give up if any pass fails to advance the state.
2807 */
2808 while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
2809 int cpu;
2810 bool isidle = true;
2811 unsigned long maxj = jiffies - ULONG_MAX / 4;
2812 struct rcu_data *rdp;
2813
2814 /* Scan all the CPUs looking for nonidle CPUs. */
2815 for_each_possible_cpu(cpu) {
2816 rdp = per_cpu_ptr(rcu_sysidle_state->rda, cpu);
2817 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
2818 if (!isidle)
2819 break;
2820 }
2821 rcu_sysidle_report(rcu_sysidle_state,
2822 isidle, maxj, false);
2823 oldrss = rss;
2824 rss = ACCESS_ONCE(full_sysidle_state);
2825 }
2826 }
2827
2828 /* If this is the first observation of an idle period, record it. */
2829 if (rss == RCU_SYSIDLE_FULL) {
2830 rss = cmpxchg(&full_sysidle_state,
2831 RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
2832 return rss == RCU_SYSIDLE_FULL;
2833 }
2834
2835 smp_mb(); /* ensure rss load happens before later caller actions. */
2836
2837 /* If already fully idle, tell the caller (in case of races). */
2838 if (rss == RCU_SYSIDLE_FULL_NOTED)
2839 return true;
2840
2841 /*
2842 * If we aren't there yet, and a grace period is not in flight,
2843 * initiate a grace period. Either way, tell the caller that
2844 * we are not there yet. We use an xchg() rather than an assignment
2845 * to make up for the memory barriers that would otherwise be
2846 * provided by the memory allocator.
2847 */
2848 if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
2849 !rcu_gp_in_progress(rcu_sysidle_state) &&
2850 !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
2851 call_rcu(&rsh.rh, rcu_sysidle_cb);
2852 return false;
2853}
2854
2855/*
2856 * Initialize dynticks sysidle state for CPUs coming online.
2857 */
2858static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2859{
2860 rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
2861}
2862
2863#else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2864
2865static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq)
2866{
2867}
2868
2869static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq)
2870{
2871}
2872
2873static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2874 unsigned long *maxj)
2875{
2876}
2877
2878static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2879{
2880 return false;
2881}
2882
2883static void rcu_bind_gp_kthread(void)
2884{
2885}
2886
2887static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2888 unsigned long maxj)
2889{
2890}
2891
2892static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2893{
2894}
2895
2896#endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2897
2898/*
2899 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2900 * grace-period kthread will do force_quiescent_state() processing?
2901 * The idea is to avoid waking up RCU core processing on such a
2902 * CPU unless the grace period has extended for too long.
2903 *
2904 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2905 * CONFIG_RCU_NOCB_CPU CPUs.
2906 */
2907static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
2908{
2909#ifdef CONFIG_NO_HZ_FULL
2910 if (tick_nohz_full_cpu(smp_processor_id()) &&
2911 (!rcu_gp_in_progress(rsp) ||
2912 ULONG_CMP_LT(jiffies, ACCESS_ONCE(rsp->gp_start) + HZ)))
2913 return 1;
2914#endif /* #ifdef CONFIG_NO_HZ_FULL */
2915 return 0;
2916}