Loading...
1/*
2 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
3 * Internal non-public definitions that provide either classic
4 * or preemptible semantics.
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
10 *
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
15 *
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, you can access it online at
18 * http://www.gnu.org/licenses/gpl-2.0.html.
19 *
20 * Copyright Red Hat, 2009
21 * Copyright IBM Corporation, 2009
22 *
23 * Author: Ingo Molnar <mingo@elte.hu>
24 * Paul E. McKenney <paulmck@linux.vnet.ibm.com>
25 */
26
27#include <linux/delay.h>
28#include <linux/gfp.h>
29#include <linux/oom.h>
30#include <linux/smpboot.h>
31#include "../time/tick-internal.h"
32
33#ifdef CONFIG_RCU_BOOST
34
35#include "../locking/rtmutex_common.h"
36
37/*
38 * Control variables for per-CPU and per-rcu_node kthreads. These
39 * handle all flavors of RCU.
40 */
41static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
42DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
43DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
44DEFINE_PER_CPU(char, rcu_cpu_has_work);
45
46#else /* #ifdef CONFIG_RCU_BOOST */
47
48/*
49 * Some architectures do not define rt_mutexes, but if !CONFIG_RCU_BOOST,
50 * all uses are in dead code. Provide a definition to keep the compiler
51 * happy, but add WARN_ON_ONCE() to complain if used in the wrong place.
52 * This probably needs to be excluded from -rt builds.
53 */
54#define rt_mutex_owner(a) ({ WARN_ON_ONCE(1); NULL; })
55
56#endif /* #else #ifdef CONFIG_RCU_BOOST */
57
58#ifdef CONFIG_RCU_NOCB_CPU
59static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
60static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */
61static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
62#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
63
64/*
65 * Check the RCU kernel configuration parameters and print informative
66 * messages about anything out of the ordinary.
67 */
68static void __init rcu_bootup_announce_oddness(void)
69{
70 if (IS_ENABLED(CONFIG_RCU_TRACE))
71 pr_info("\tRCU debugfs-based tracing is enabled.\n");
72 if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
73 (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
74 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
75 RCU_FANOUT);
76 if (rcu_fanout_exact)
77 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
78 if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ))
79 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
80 if (IS_ENABLED(CONFIG_PROVE_RCU))
81 pr_info("\tRCU lockdep checking is enabled.\n");
82 if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST_RUNNABLE))
83 pr_info("\tRCU torture testing starts during boot.\n");
84 if (RCU_NUM_LVLS >= 4)
85 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
86 if (RCU_FANOUT_LEAF != 16)
87 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
88 RCU_FANOUT_LEAF);
89 if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
90 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
91 if (nr_cpu_ids != NR_CPUS)
92 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
93 if (IS_ENABLED(CONFIG_RCU_BOOST))
94 pr_info("\tRCU kthread priority: %d.\n", kthread_prio);
95}
96
97#ifdef CONFIG_PREEMPT_RCU
98
99RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
100static struct rcu_state *const rcu_state_p = &rcu_preempt_state;
101static struct rcu_data __percpu *const rcu_data_p = &rcu_preempt_data;
102
103static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
104 bool wake);
105
106/*
107 * Tell them what RCU they are running.
108 */
109static void __init rcu_bootup_announce(void)
110{
111 pr_info("Preemptible hierarchical RCU implementation.\n");
112 rcu_bootup_announce_oddness();
113}
114
115/* Flags for rcu_preempt_ctxt_queue() decision table. */
116#define RCU_GP_TASKS 0x8
117#define RCU_EXP_TASKS 0x4
118#define RCU_GP_BLKD 0x2
119#define RCU_EXP_BLKD 0x1
120
121/*
122 * Queues a task preempted within an RCU-preempt read-side critical
123 * section into the appropriate location within the ->blkd_tasks list,
124 * depending on the states of any ongoing normal and expedited grace
125 * periods. The ->gp_tasks pointer indicates which element the normal
126 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
127 * indicates which element the expedited grace period is waiting on (again,
128 * NULL if none). If a grace period is waiting on a given element in the
129 * ->blkd_tasks list, it also waits on all subsequent elements. Thus,
130 * adding a task to the tail of the list blocks any grace period that is
131 * already waiting on one of the elements. In contrast, adding a task
132 * to the head of the list won't block any grace period that is already
133 * waiting on one of the elements.
134 *
135 * This queuing is imprecise, and can sometimes make an ongoing grace
136 * period wait for a task that is not strictly speaking blocking it.
137 * Given the choice, we needlessly block a normal grace period rather than
138 * blocking an expedited grace period.
139 *
140 * Note that an endless sequence of expedited grace periods still cannot
141 * indefinitely postpone a normal grace period. Eventually, all of the
142 * fixed number of preempted tasks blocking the normal grace period that are
143 * not also blocking the expedited grace period will resume and complete
144 * their RCU read-side critical sections. At that point, the ->gp_tasks
145 * pointer will equal the ->exp_tasks pointer, at which point the end of
146 * the corresponding expedited grace period will also be the end of the
147 * normal grace period.
148 */
149static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
150 __releases(rnp->lock) /* But leaves rrupts disabled. */
151{
152 int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
153 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
154 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
155 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
156 struct task_struct *t = current;
157
158 /*
159 * Decide where to queue the newly blocked task. In theory,
160 * this could be an if-statement. In practice, when I tried
161 * that, it was quite messy.
162 */
163 switch (blkd_state) {
164 case 0:
165 case RCU_EXP_TASKS:
166 case RCU_EXP_TASKS + RCU_GP_BLKD:
167 case RCU_GP_TASKS:
168 case RCU_GP_TASKS + RCU_EXP_TASKS:
169
170 /*
171 * Blocking neither GP, or first task blocking the normal
172 * GP but not blocking the already-waiting expedited GP.
173 * Queue at the head of the list to avoid unnecessarily
174 * blocking the already-waiting GPs.
175 */
176 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
177 break;
178
179 case RCU_EXP_BLKD:
180 case RCU_GP_BLKD:
181 case RCU_GP_BLKD + RCU_EXP_BLKD:
182 case RCU_GP_TASKS + RCU_EXP_BLKD:
183 case RCU_GP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
184 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
185
186 /*
187 * First task arriving that blocks either GP, or first task
188 * arriving that blocks the expedited GP (with the normal
189 * GP already waiting), or a task arriving that blocks
190 * both GPs with both GPs already waiting. Queue at the
191 * tail of the list to avoid any GP waiting on any of the
192 * already queued tasks that are not blocking it.
193 */
194 list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
195 break;
196
197 case RCU_EXP_TASKS + RCU_EXP_BLKD:
198 case RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
199 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_EXP_BLKD:
200
201 /*
202 * Second or subsequent task blocking the expedited GP.
203 * The task either does not block the normal GP, or is the
204 * first task blocking the normal GP. Queue just after
205 * the first task blocking the expedited GP.
206 */
207 list_add(&t->rcu_node_entry, rnp->exp_tasks);
208 break;
209
210 case RCU_GP_TASKS + RCU_GP_BLKD:
211 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
212
213 /*
214 * Second or subsequent task blocking the normal GP.
215 * The task does not block the expedited GP. Queue just
216 * after the first task blocking the normal GP.
217 */
218 list_add(&t->rcu_node_entry, rnp->gp_tasks);
219 break;
220
221 default:
222
223 /* Yet another exercise in excessive paranoia. */
224 WARN_ON_ONCE(1);
225 break;
226 }
227
228 /*
229 * We have now queued the task. If it was the first one to
230 * block either grace period, update the ->gp_tasks and/or
231 * ->exp_tasks pointers, respectively, to reference the newly
232 * blocked tasks.
233 */
234 if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD))
235 rnp->gp_tasks = &t->rcu_node_entry;
236 if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
237 rnp->exp_tasks = &t->rcu_node_entry;
238 raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
239
240 /*
241 * Report the quiescent state for the expedited GP. This expedited
242 * GP should not be able to end until we report, so there should be
243 * no need to check for a subsequent expedited GP. (Though we are
244 * still in a quiescent state in any case.)
245 */
246 if (blkd_state & RCU_EXP_BLKD &&
247 t->rcu_read_unlock_special.b.exp_need_qs) {
248 t->rcu_read_unlock_special.b.exp_need_qs = false;
249 rcu_report_exp_rdp(rdp->rsp, rdp, true);
250 } else {
251 WARN_ON_ONCE(t->rcu_read_unlock_special.b.exp_need_qs);
252 }
253}
254
255/*
256 * Record a preemptible-RCU quiescent state for the specified CPU. Note
257 * that this just means that the task currently running on the CPU is
258 * not in a quiescent state. There might be any number of tasks blocked
259 * while in an RCU read-side critical section.
260 *
261 * As with the other rcu_*_qs() functions, callers to this function
262 * must disable preemption.
263 */
264static void rcu_preempt_qs(void)
265{
266 if (__this_cpu_read(rcu_data_p->cpu_no_qs.s)) {
267 trace_rcu_grace_period(TPS("rcu_preempt"),
268 __this_cpu_read(rcu_data_p->gpnum),
269 TPS("cpuqs"));
270 __this_cpu_write(rcu_data_p->cpu_no_qs.b.norm, false);
271 barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */
272 current->rcu_read_unlock_special.b.need_qs = false;
273 }
274}
275
276/*
277 * We have entered the scheduler, and the current task might soon be
278 * context-switched away from. If this task is in an RCU read-side
279 * critical section, we will no longer be able to rely on the CPU to
280 * record that fact, so we enqueue the task on the blkd_tasks list.
281 * The task will dequeue itself when it exits the outermost enclosing
282 * RCU read-side critical section. Therefore, the current grace period
283 * cannot be permitted to complete until the blkd_tasks list entries
284 * predating the current grace period drain, in other words, until
285 * rnp->gp_tasks becomes NULL.
286 *
287 * Caller must disable interrupts.
288 */
289static void rcu_preempt_note_context_switch(void)
290{
291 struct task_struct *t = current;
292 struct rcu_data *rdp;
293 struct rcu_node *rnp;
294
295 if (t->rcu_read_lock_nesting > 0 &&
296 !t->rcu_read_unlock_special.b.blocked) {
297
298 /* Possibly blocking in an RCU read-side critical section. */
299 rdp = this_cpu_ptr(rcu_state_p->rda);
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(rdp->rsp->name,
313 t->pid,
314 (rnp->qsmask & rdp->grpmask)
315 ? rnp->gpnum
316 : rnp->gpnum + 1);
317 rcu_preempt_ctxt_queue(rnp, rdp);
318 } else if (t->rcu_read_lock_nesting < 0 &&
319 t->rcu_read_unlock_special.s) {
320
321 /*
322 * Complete exit from RCU read-side critical section on
323 * behalf of preempted instance of __rcu_read_unlock().
324 */
325 rcu_read_unlock_special(t);
326 }
327
328 /*
329 * Either we were not in an RCU read-side critical section to
330 * begin with, or we have now recorded that critical section
331 * globally. Either way, we can now note a quiescent state
332 * for this CPU. Again, if we were in an RCU read-side critical
333 * section, and if that critical section was blocking the current
334 * grace period, then the fact that the task has been enqueued
335 * means that we continue to block the current grace period.
336 */
337 rcu_preempt_qs();
338}
339
340/*
341 * Check for preempted RCU readers blocking the current grace period
342 * for the specified rcu_node structure. If the caller needs a reliable
343 * answer, it must hold the rcu_node's ->lock.
344 */
345static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
346{
347 return rnp->gp_tasks != NULL;
348}
349
350/*
351 * Advance a ->blkd_tasks-list pointer to the next entry, instead
352 * returning NULL if at the end of the list.
353 */
354static struct list_head *rcu_next_node_entry(struct task_struct *t,
355 struct rcu_node *rnp)
356{
357 struct list_head *np;
358
359 np = t->rcu_node_entry.next;
360 if (np == &rnp->blkd_tasks)
361 np = NULL;
362 return np;
363}
364
365/*
366 * Return true if the specified rcu_node structure has tasks that were
367 * preempted within an RCU read-side critical section.
368 */
369static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
370{
371 return !list_empty(&rnp->blkd_tasks);
372}
373
374/*
375 * Handle special cases during rcu_read_unlock(), such as needing to
376 * notify RCU core processing or task having blocked during the RCU
377 * read-side critical section.
378 */
379void rcu_read_unlock_special(struct task_struct *t)
380{
381 bool empty_exp;
382 bool empty_norm;
383 bool empty_exp_now;
384 unsigned long flags;
385 struct list_head *np;
386 bool drop_boost_mutex = false;
387 struct rcu_data *rdp;
388 struct rcu_node *rnp;
389 union rcu_special special;
390
391 /* NMI handlers cannot block and cannot safely manipulate state. */
392 if (in_nmi())
393 return;
394
395 local_irq_save(flags);
396
397 /*
398 * If RCU core is waiting for this CPU to exit its critical section,
399 * report the fact that it has exited. Because irqs are disabled,
400 * t->rcu_read_unlock_special cannot change.
401 */
402 special = t->rcu_read_unlock_special;
403 if (special.b.need_qs) {
404 rcu_preempt_qs();
405 t->rcu_read_unlock_special.b.need_qs = false;
406 if (!t->rcu_read_unlock_special.s) {
407 local_irq_restore(flags);
408 return;
409 }
410 }
411
412 /*
413 * Respond to a request for an expedited grace period, but only if
414 * we were not preempted, meaning that we were running on the same
415 * CPU throughout. If we were preempted, the exp_need_qs flag
416 * would have been cleared at the time of the first preemption,
417 * and the quiescent state would be reported when we were dequeued.
418 */
419 if (special.b.exp_need_qs) {
420 WARN_ON_ONCE(special.b.blocked);
421 t->rcu_read_unlock_special.b.exp_need_qs = false;
422 rdp = this_cpu_ptr(rcu_state_p->rda);
423 rcu_report_exp_rdp(rcu_state_p, rdp, true);
424 if (!t->rcu_read_unlock_special.s) {
425 local_irq_restore(flags);
426 return;
427 }
428 }
429
430 /* Hardware IRQ handlers cannot block, complain if they get here. */
431 if (in_irq() || in_serving_softirq()) {
432 lockdep_rcu_suspicious(__FILE__, __LINE__,
433 "rcu_read_unlock() from irq or softirq with blocking in critical section!!!\n");
434 pr_alert("->rcu_read_unlock_special: %#x (b: %d, enq: %d nq: %d)\n",
435 t->rcu_read_unlock_special.s,
436 t->rcu_read_unlock_special.b.blocked,
437 t->rcu_read_unlock_special.b.exp_need_qs,
438 t->rcu_read_unlock_special.b.need_qs);
439 local_irq_restore(flags);
440 return;
441 }
442
443 /* Clean up if blocked during RCU read-side critical section. */
444 if (special.b.blocked) {
445 t->rcu_read_unlock_special.b.blocked = false;
446
447 /*
448 * Remove this task from the list it blocked on. The task
449 * now remains queued on the rcu_node corresponding to the
450 * CPU it first blocked on, so there is no longer any need
451 * to loop. Retain a WARN_ON_ONCE() out of sheer paranoia.
452 */
453 rnp = t->rcu_blocked_node;
454 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
455 WARN_ON_ONCE(rnp != t->rcu_blocked_node);
456 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
457 empty_exp = sync_rcu_preempt_exp_done(rnp);
458 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
459 np = rcu_next_node_entry(t, rnp);
460 list_del_init(&t->rcu_node_entry);
461 t->rcu_blocked_node = NULL;
462 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
463 rnp->gpnum, t->pid);
464 if (&t->rcu_node_entry == rnp->gp_tasks)
465 rnp->gp_tasks = np;
466 if (&t->rcu_node_entry == rnp->exp_tasks)
467 rnp->exp_tasks = np;
468 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
469 if (&t->rcu_node_entry == rnp->boost_tasks)
470 rnp->boost_tasks = np;
471 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
472 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
473 }
474
475 /*
476 * If this was the last task on the current list, and if
477 * we aren't waiting on any CPUs, report the quiescent state.
478 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
479 * so we must take a snapshot of the expedited state.
480 */
481 empty_exp_now = sync_rcu_preempt_exp_done(rnp);
482 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
483 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
484 rnp->gpnum,
485 0, rnp->qsmask,
486 rnp->level,
487 rnp->grplo,
488 rnp->grphi,
489 !!rnp->gp_tasks);
490 rcu_report_unblock_qs_rnp(rcu_state_p, rnp, flags);
491 } else {
492 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
493 }
494
495 /* Unboost if we were boosted. */
496 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
497 rt_mutex_unlock(&rnp->boost_mtx);
498
499 /*
500 * If this was the last task on the expedited lists,
501 * then we need to report up the rcu_node hierarchy.
502 */
503 if (!empty_exp && empty_exp_now)
504 rcu_report_exp_rnp(rcu_state_p, rnp, true);
505 } else {
506 local_irq_restore(flags);
507 }
508}
509
510/*
511 * Dump detailed information for all tasks blocking the current RCU
512 * grace period on the specified rcu_node structure.
513 */
514static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
515{
516 unsigned long flags;
517 struct task_struct *t;
518
519 raw_spin_lock_irqsave_rcu_node(rnp, flags);
520 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
521 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
522 return;
523 }
524 t = list_entry(rnp->gp_tasks->prev,
525 struct task_struct, rcu_node_entry);
526 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
527 sched_show_task(t);
528 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
529}
530
531/*
532 * Dump detailed information for all tasks blocking the current RCU
533 * grace period.
534 */
535static void rcu_print_detail_task_stall(struct rcu_state *rsp)
536{
537 struct rcu_node *rnp = rcu_get_root(rsp);
538
539 rcu_print_detail_task_stall_rnp(rnp);
540 rcu_for_each_leaf_node(rsp, rnp)
541 rcu_print_detail_task_stall_rnp(rnp);
542}
543
544static void rcu_print_task_stall_begin(struct rcu_node *rnp)
545{
546 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
547 rnp->level, rnp->grplo, rnp->grphi);
548}
549
550static void rcu_print_task_stall_end(void)
551{
552 pr_cont("\n");
553}
554
555/*
556 * Scan the current list of tasks blocked within RCU read-side critical
557 * sections, printing out the tid of each.
558 */
559static int rcu_print_task_stall(struct rcu_node *rnp)
560{
561 struct task_struct *t;
562 int ndetected = 0;
563
564 if (!rcu_preempt_blocked_readers_cgp(rnp))
565 return 0;
566 rcu_print_task_stall_begin(rnp);
567 t = list_entry(rnp->gp_tasks->prev,
568 struct task_struct, rcu_node_entry);
569 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
570 pr_cont(" P%d", t->pid);
571 ndetected++;
572 }
573 rcu_print_task_stall_end();
574 return ndetected;
575}
576
577/*
578 * Scan the current list of tasks blocked within RCU read-side critical
579 * sections, printing out the tid of each that is blocking the current
580 * expedited grace period.
581 */
582static int rcu_print_task_exp_stall(struct rcu_node *rnp)
583{
584 struct task_struct *t;
585 int ndetected = 0;
586
587 if (!rnp->exp_tasks)
588 return 0;
589 t = list_entry(rnp->exp_tasks->prev,
590 struct task_struct, rcu_node_entry);
591 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
592 pr_cont(" P%d", t->pid);
593 ndetected++;
594 }
595 return ndetected;
596}
597
598/*
599 * Check that the list of blocked tasks for the newly completed grace
600 * period is in fact empty. It is a serious bug to complete a grace
601 * period that still has RCU readers blocked! This function must be
602 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
603 * must be held by the caller.
604 *
605 * Also, if there are blocked tasks on the list, they automatically
606 * block the newly created grace period, so set up ->gp_tasks accordingly.
607 */
608static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
609{
610 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
611 if (rcu_preempt_has_tasks(rnp))
612 rnp->gp_tasks = rnp->blkd_tasks.next;
613 WARN_ON_ONCE(rnp->qsmask);
614}
615
616/*
617 * Check for a quiescent state from the current CPU. When a task blocks,
618 * the task is recorded in the corresponding CPU's rcu_node structure,
619 * which is checked elsewhere.
620 *
621 * Caller must disable hard irqs.
622 */
623static void rcu_preempt_check_callbacks(void)
624{
625 struct task_struct *t = current;
626
627 if (t->rcu_read_lock_nesting == 0) {
628 rcu_preempt_qs();
629 return;
630 }
631 if (t->rcu_read_lock_nesting > 0 &&
632 __this_cpu_read(rcu_data_p->core_needs_qs) &&
633 __this_cpu_read(rcu_data_p->cpu_no_qs.b.norm))
634 t->rcu_read_unlock_special.b.need_qs = true;
635}
636
637#ifdef CONFIG_RCU_BOOST
638
639static void rcu_preempt_do_callbacks(void)
640{
641 rcu_do_batch(rcu_state_p, this_cpu_ptr(rcu_data_p));
642}
643
644#endif /* #ifdef CONFIG_RCU_BOOST */
645
646/*
647 * Queue a preemptible-RCU callback for invocation after a grace period.
648 */
649void call_rcu(struct rcu_head *head, rcu_callback_t func)
650{
651 __call_rcu(head, func, rcu_state_p, -1, 0);
652}
653EXPORT_SYMBOL_GPL(call_rcu);
654
655/**
656 * synchronize_rcu - wait until a grace period has elapsed.
657 *
658 * Control will return to the caller some time after a full grace
659 * period has elapsed, in other words after all currently executing RCU
660 * read-side critical sections have completed. Note, however, that
661 * upon return from synchronize_rcu(), the caller might well be executing
662 * concurrently with new RCU read-side critical sections that began while
663 * synchronize_rcu() was waiting. RCU read-side critical sections are
664 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
665 *
666 * See the description of synchronize_sched() for more detailed information
667 * on memory ordering guarantees.
668 */
669void synchronize_rcu(void)
670{
671 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
672 lock_is_held(&rcu_lock_map) ||
673 lock_is_held(&rcu_sched_lock_map),
674 "Illegal synchronize_rcu() in RCU read-side critical section");
675 if (!rcu_scheduler_active)
676 return;
677 if (rcu_gp_is_expedited())
678 synchronize_rcu_expedited();
679 else
680 wait_rcu_gp(call_rcu);
681}
682EXPORT_SYMBOL_GPL(synchronize_rcu);
683
684/*
685 * Remote handler for smp_call_function_single(). If there is an
686 * RCU read-side critical section in effect, request that the
687 * next rcu_read_unlock() record the quiescent state up the
688 * ->expmask fields in the rcu_node tree. Otherwise, immediately
689 * report the quiescent state.
690 */
691static void sync_rcu_exp_handler(void *info)
692{
693 struct rcu_data *rdp;
694 struct rcu_state *rsp = info;
695 struct task_struct *t = current;
696
697 /*
698 * Within an RCU read-side critical section, request that the next
699 * rcu_read_unlock() report. Unless this RCU read-side critical
700 * section has already blocked, in which case it is already set
701 * up for the expedited grace period to wait on it.
702 */
703 if (t->rcu_read_lock_nesting > 0 &&
704 !t->rcu_read_unlock_special.b.blocked) {
705 t->rcu_read_unlock_special.b.exp_need_qs = true;
706 return;
707 }
708
709 /*
710 * We are either exiting an RCU read-side critical section (negative
711 * values of t->rcu_read_lock_nesting) or are not in one at all
712 * (zero value of t->rcu_read_lock_nesting). Or we are in an RCU
713 * read-side critical section that blocked before this expedited
714 * grace period started. Either way, we can immediately report
715 * the quiescent state.
716 */
717 rdp = this_cpu_ptr(rsp->rda);
718 rcu_report_exp_rdp(rsp, rdp, true);
719}
720
721/**
722 * synchronize_rcu_expedited - Brute-force RCU grace period
723 *
724 * Wait for an RCU-preempt grace period, but expedite it. The basic
725 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
726 * the ->blkd_tasks lists and wait for this list to drain. This consumes
727 * significant time on all CPUs and is unfriendly to real-time workloads,
728 * so is thus not recommended for any sort of common-case code.
729 * In fact, if you are using synchronize_rcu_expedited() in a loop,
730 * please restructure your code to batch your updates, and then Use a
731 * single synchronize_rcu() instead.
732 */
733void synchronize_rcu_expedited(void)
734{
735 struct rcu_node *rnp;
736 struct rcu_node *rnp_unlock;
737 struct rcu_state *rsp = rcu_state_p;
738 unsigned long s;
739
740 /* If expedited grace periods are prohibited, fall back to normal. */
741 if (rcu_gp_is_normal()) {
742 wait_rcu_gp(call_rcu);
743 return;
744 }
745
746 s = rcu_exp_gp_seq_snap(rsp);
747
748 rnp_unlock = exp_funnel_lock(rsp, s);
749 if (rnp_unlock == NULL)
750 return; /* Someone else did our work for us. */
751
752 rcu_exp_gp_seq_start(rsp);
753
754 /* Initialize the rcu_node tree in preparation for the wait. */
755 sync_rcu_exp_select_cpus(rsp, sync_rcu_exp_handler);
756
757 /* Wait for snapshotted ->blkd_tasks lists to drain. */
758 rnp = rcu_get_root(rsp);
759 synchronize_sched_expedited_wait(rsp);
760
761 /* Clean up and exit. */
762 rcu_exp_gp_seq_end(rsp);
763 mutex_unlock(&rnp_unlock->exp_funnel_mutex);
764}
765EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
766
767/**
768 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
769 *
770 * Note that this primitive does not necessarily wait for an RCU grace period
771 * to complete. For example, if there are no RCU callbacks queued anywhere
772 * in the system, then rcu_barrier() is within its rights to return
773 * immediately, without waiting for anything, much less an RCU grace period.
774 */
775void rcu_barrier(void)
776{
777 _rcu_barrier(rcu_state_p);
778}
779EXPORT_SYMBOL_GPL(rcu_barrier);
780
781/*
782 * Initialize preemptible RCU's state structures.
783 */
784static void __init __rcu_init_preempt(void)
785{
786 rcu_init_one(rcu_state_p);
787}
788
789/*
790 * Check for a task exiting while in a preemptible-RCU read-side
791 * critical section, clean up if so. No need to issue warnings,
792 * as debug_check_no_locks_held() already does this if lockdep
793 * is enabled.
794 */
795void exit_rcu(void)
796{
797 struct task_struct *t = current;
798
799 if (likely(list_empty(¤t->rcu_node_entry)))
800 return;
801 t->rcu_read_lock_nesting = 1;
802 barrier();
803 t->rcu_read_unlock_special.b.blocked = true;
804 __rcu_read_unlock();
805}
806
807#else /* #ifdef CONFIG_PREEMPT_RCU */
808
809static struct rcu_state *const rcu_state_p = &rcu_sched_state;
810
811/*
812 * Tell them what RCU they are running.
813 */
814static void __init rcu_bootup_announce(void)
815{
816 pr_info("Hierarchical RCU implementation.\n");
817 rcu_bootup_announce_oddness();
818}
819
820/*
821 * Because preemptible RCU does not exist, we never have to check for
822 * CPUs being in quiescent states.
823 */
824static void rcu_preempt_note_context_switch(void)
825{
826}
827
828/*
829 * Because preemptible RCU does not exist, there are never any preempted
830 * RCU readers.
831 */
832static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
833{
834 return 0;
835}
836
837/*
838 * Because there is no preemptible RCU, there can be no readers blocked.
839 */
840static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
841{
842 return false;
843}
844
845/*
846 * Because preemptible RCU does not exist, we never have to check for
847 * tasks blocked within RCU read-side critical sections.
848 */
849static void rcu_print_detail_task_stall(struct rcu_state *rsp)
850{
851}
852
853/*
854 * Because preemptible RCU does not exist, we never have to check for
855 * tasks blocked within RCU read-side critical sections.
856 */
857static int rcu_print_task_stall(struct rcu_node *rnp)
858{
859 return 0;
860}
861
862/*
863 * Because preemptible RCU does not exist, we never have to check for
864 * tasks blocked within RCU read-side critical sections that are
865 * blocking the current expedited grace period.
866 */
867static int rcu_print_task_exp_stall(struct rcu_node *rnp)
868{
869 return 0;
870}
871
872/*
873 * Because there is no preemptible RCU, there can be no readers blocked,
874 * so there is no need to check for blocked tasks. So check only for
875 * bogus qsmask values.
876 */
877static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
878{
879 WARN_ON_ONCE(rnp->qsmask);
880}
881
882/*
883 * Because preemptible RCU does not exist, it never has any callbacks
884 * to check.
885 */
886static void rcu_preempt_check_callbacks(void)
887{
888}
889
890/*
891 * Wait for an rcu-preempt grace period, but make it happen quickly.
892 * But because preemptible RCU does not exist, map to rcu-sched.
893 */
894void synchronize_rcu_expedited(void)
895{
896 synchronize_sched_expedited();
897}
898EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
899
900/*
901 * Because preemptible RCU does not exist, rcu_barrier() is just
902 * another name for rcu_barrier_sched().
903 */
904void rcu_barrier(void)
905{
906 rcu_barrier_sched();
907}
908EXPORT_SYMBOL_GPL(rcu_barrier);
909
910/*
911 * Because preemptible RCU does not exist, it need not be initialized.
912 */
913static void __init __rcu_init_preempt(void)
914{
915}
916
917/*
918 * Because preemptible RCU does not exist, tasks cannot possibly exit
919 * while in preemptible RCU read-side critical sections.
920 */
921void exit_rcu(void)
922{
923}
924
925#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
926
927#ifdef CONFIG_RCU_BOOST
928
929#include "../locking/rtmutex_common.h"
930
931#ifdef CONFIG_RCU_TRACE
932
933static void rcu_initiate_boost_trace(struct rcu_node *rnp)
934{
935 if (!rcu_preempt_has_tasks(rnp))
936 rnp->n_balk_blkd_tasks++;
937 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
938 rnp->n_balk_exp_gp_tasks++;
939 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
940 rnp->n_balk_boost_tasks++;
941 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
942 rnp->n_balk_notblocked++;
943 else if (rnp->gp_tasks != NULL &&
944 ULONG_CMP_LT(jiffies, rnp->boost_time))
945 rnp->n_balk_notyet++;
946 else
947 rnp->n_balk_nos++;
948}
949
950#else /* #ifdef CONFIG_RCU_TRACE */
951
952static void rcu_initiate_boost_trace(struct rcu_node *rnp)
953{
954}
955
956#endif /* #else #ifdef CONFIG_RCU_TRACE */
957
958static void rcu_wake_cond(struct task_struct *t, int status)
959{
960 /*
961 * If the thread is yielding, only wake it when this
962 * is invoked from idle
963 */
964 if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
965 wake_up_process(t);
966}
967
968/*
969 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
970 * or ->boost_tasks, advancing the pointer to the next task in the
971 * ->blkd_tasks list.
972 *
973 * Note that irqs must be enabled: boosting the task can block.
974 * Returns 1 if there are more tasks needing to be boosted.
975 */
976static int rcu_boost(struct rcu_node *rnp)
977{
978 unsigned long flags;
979 struct task_struct *t;
980 struct list_head *tb;
981
982 if (READ_ONCE(rnp->exp_tasks) == NULL &&
983 READ_ONCE(rnp->boost_tasks) == NULL)
984 return 0; /* Nothing left to boost. */
985
986 raw_spin_lock_irqsave_rcu_node(rnp, flags);
987
988 /*
989 * Recheck under the lock: all tasks in need of boosting
990 * might exit their RCU read-side critical sections on their own.
991 */
992 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
993 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
994 return 0;
995 }
996
997 /*
998 * Preferentially boost tasks blocking expedited grace periods.
999 * This cannot starve the normal grace periods because a second
1000 * expedited grace period must boost all blocked tasks, including
1001 * those blocking the pre-existing normal grace period.
1002 */
1003 if (rnp->exp_tasks != NULL) {
1004 tb = rnp->exp_tasks;
1005 rnp->n_exp_boosts++;
1006 } else {
1007 tb = rnp->boost_tasks;
1008 rnp->n_normal_boosts++;
1009 }
1010 rnp->n_tasks_boosted++;
1011
1012 /*
1013 * We boost task t by manufacturing an rt_mutex that appears to
1014 * be held by task t. We leave a pointer to that rt_mutex where
1015 * task t can find it, and task t will release the mutex when it
1016 * exits its outermost RCU read-side critical section. Then
1017 * simply acquiring this artificial rt_mutex will boost task
1018 * t's priority. (Thanks to tglx for suggesting this approach!)
1019 *
1020 * Note that task t must acquire rnp->lock to remove itself from
1021 * the ->blkd_tasks list, which it will do from exit() if from
1022 * nowhere else. We therefore are guaranteed that task t will
1023 * stay around at least until we drop rnp->lock. Note that
1024 * rnp->lock also resolves races between our priority boosting
1025 * and task t's exiting its outermost RCU read-side critical
1026 * section.
1027 */
1028 t = container_of(tb, struct task_struct, rcu_node_entry);
1029 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1030 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1031 /* Lock only for side effect: boosts task t's priority. */
1032 rt_mutex_lock(&rnp->boost_mtx);
1033 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1034
1035 return READ_ONCE(rnp->exp_tasks) != NULL ||
1036 READ_ONCE(rnp->boost_tasks) != NULL;
1037}
1038
1039/*
1040 * Priority-boosting kthread, one per leaf rcu_node.
1041 */
1042static int rcu_boost_kthread(void *arg)
1043{
1044 struct rcu_node *rnp = (struct rcu_node *)arg;
1045 int spincnt = 0;
1046 int more2boost;
1047
1048 trace_rcu_utilization(TPS("Start boost kthread@init"));
1049 for (;;) {
1050 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1051 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1052 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1053 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1054 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1055 more2boost = rcu_boost(rnp);
1056 if (more2boost)
1057 spincnt++;
1058 else
1059 spincnt = 0;
1060 if (spincnt > 10) {
1061 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1062 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1063 schedule_timeout_interruptible(2);
1064 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1065 spincnt = 0;
1066 }
1067 }
1068 /* NOTREACHED */
1069 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1070 return 0;
1071}
1072
1073/*
1074 * Check to see if it is time to start boosting RCU readers that are
1075 * blocking the current grace period, and, if so, tell the per-rcu_node
1076 * kthread to start boosting them. If there is an expedited grace
1077 * period in progress, it is always time to boost.
1078 *
1079 * The caller must hold rnp->lock, which this function releases.
1080 * The ->boost_kthread_task is immortal, so we don't need to worry
1081 * about it going away.
1082 */
1083static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1084 __releases(rnp->lock)
1085{
1086 struct task_struct *t;
1087
1088 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1089 rnp->n_balk_exp_gp_tasks++;
1090 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1091 return;
1092 }
1093 if (rnp->exp_tasks != NULL ||
1094 (rnp->gp_tasks != NULL &&
1095 rnp->boost_tasks == NULL &&
1096 rnp->qsmask == 0 &&
1097 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1098 if (rnp->exp_tasks == NULL)
1099 rnp->boost_tasks = rnp->gp_tasks;
1100 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1101 t = rnp->boost_kthread_task;
1102 if (t)
1103 rcu_wake_cond(t, rnp->boost_kthread_status);
1104 } else {
1105 rcu_initiate_boost_trace(rnp);
1106 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1107 }
1108}
1109
1110/*
1111 * Wake up the per-CPU kthread to invoke RCU callbacks.
1112 */
1113static void invoke_rcu_callbacks_kthread(void)
1114{
1115 unsigned long flags;
1116
1117 local_irq_save(flags);
1118 __this_cpu_write(rcu_cpu_has_work, 1);
1119 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1120 current != __this_cpu_read(rcu_cpu_kthread_task)) {
1121 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1122 __this_cpu_read(rcu_cpu_kthread_status));
1123 }
1124 local_irq_restore(flags);
1125}
1126
1127/*
1128 * Is the current CPU running the RCU-callbacks kthread?
1129 * Caller must have preemption disabled.
1130 */
1131static bool rcu_is_callbacks_kthread(void)
1132{
1133 return __this_cpu_read(rcu_cpu_kthread_task) == current;
1134}
1135
1136#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1137
1138/*
1139 * Do priority-boost accounting for the start of a new grace period.
1140 */
1141static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1142{
1143 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1144}
1145
1146/*
1147 * Create an RCU-boost kthread for the specified node if one does not
1148 * already exist. We only create this kthread for preemptible RCU.
1149 * Returns zero if all is well, a negated errno otherwise.
1150 */
1151static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1152 struct rcu_node *rnp)
1153{
1154 int rnp_index = rnp - &rsp->node[0];
1155 unsigned long flags;
1156 struct sched_param sp;
1157 struct task_struct *t;
1158
1159 if (rcu_state_p != rsp)
1160 return 0;
1161
1162 if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1163 return 0;
1164
1165 rsp->boost = 1;
1166 if (rnp->boost_kthread_task != NULL)
1167 return 0;
1168 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1169 "rcub/%d", rnp_index);
1170 if (IS_ERR(t))
1171 return PTR_ERR(t);
1172 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1173 rnp->boost_kthread_task = t;
1174 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1175 sp.sched_priority = kthread_prio;
1176 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1177 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1178 return 0;
1179}
1180
1181static void rcu_kthread_do_work(void)
1182{
1183 rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1184 rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1185 rcu_preempt_do_callbacks();
1186}
1187
1188static void rcu_cpu_kthread_setup(unsigned int cpu)
1189{
1190 struct sched_param sp;
1191
1192 sp.sched_priority = kthread_prio;
1193 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1194}
1195
1196static void rcu_cpu_kthread_park(unsigned int cpu)
1197{
1198 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1199}
1200
1201static int rcu_cpu_kthread_should_run(unsigned int cpu)
1202{
1203 return __this_cpu_read(rcu_cpu_has_work);
1204}
1205
1206/*
1207 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1208 * RCU softirq used in flavors and configurations of RCU that do not
1209 * support RCU priority boosting.
1210 */
1211static void rcu_cpu_kthread(unsigned int cpu)
1212{
1213 unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1214 char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1215 int spincnt;
1216
1217 for (spincnt = 0; spincnt < 10; spincnt++) {
1218 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1219 local_bh_disable();
1220 *statusp = RCU_KTHREAD_RUNNING;
1221 this_cpu_inc(rcu_cpu_kthread_loops);
1222 local_irq_disable();
1223 work = *workp;
1224 *workp = 0;
1225 local_irq_enable();
1226 if (work)
1227 rcu_kthread_do_work();
1228 local_bh_enable();
1229 if (*workp == 0) {
1230 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1231 *statusp = RCU_KTHREAD_WAITING;
1232 return;
1233 }
1234 }
1235 *statusp = RCU_KTHREAD_YIELDING;
1236 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1237 schedule_timeout_interruptible(2);
1238 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1239 *statusp = RCU_KTHREAD_WAITING;
1240}
1241
1242/*
1243 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1244 * served by the rcu_node in question. The CPU hotplug lock is still
1245 * held, so the value of rnp->qsmaskinit will be stable.
1246 *
1247 * We don't include outgoingcpu in the affinity set, use -1 if there is
1248 * no outgoing CPU. If there are no CPUs left in the affinity set,
1249 * this function allows the kthread to execute on any CPU.
1250 */
1251static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1252{
1253 struct task_struct *t = rnp->boost_kthread_task;
1254 unsigned long mask = rcu_rnp_online_cpus(rnp);
1255 cpumask_var_t cm;
1256 int cpu;
1257
1258 if (!t)
1259 return;
1260 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1261 return;
1262 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1263 if ((mask & 0x1) && cpu != outgoingcpu)
1264 cpumask_set_cpu(cpu, cm);
1265 if (cpumask_weight(cm) == 0)
1266 cpumask_setall(cm);
1267 set_cpus_allowed_ptr(t, cm);
1268 free_cpumask_var(cm);
1269}
1270
1271static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1272 .store = &rcu_cpu_kthread_task,
1273 .thread_should_run = rcu_cpu_kthread_should_run,
1274 .thread_fn = rcu_cpu_kthread,
1275 .thread_comm = "rcuc/%u",
1276 .setup = rcu_cpu_kthread_setup,
1277 .park = rcu_cpu_kthread_park,
1278};
1279
1280/*
1281 * Spawn boost kthreads -- called as soon as the scheduler is running.
1282 */
1283static void __init rcu_spawn_boost_kthreads(void)
1284{
1285 struct rcu_node *rnp;
1286 int cpu;
1287
1288 for_each_possible_cpu(cpu)
1289 per_cpu(rcu_cpu_has_work, cpu) = 0;
1290 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1291 rcu_for_each_leaf_node(rcu_state_p, rnp)
1292 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1293}
1294
1295static void rcu_prepare_kthreads(int cpu)
1296{
1297 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
1298 struct rcu_node *rnp = rdp->mynode;
1299
1300 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1301 if (rcu_scheduler_fully_active)
1302 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1303}
1304
1305#else /* #ifdef CONFIG_RCU_BOOST */
1306
1307static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1308 __releases(rnp->lock)
1309{
1310 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1311}
1312
1313static void invoke_rcu_callbacks_kthread(void)
1314{
1315 WARN_ON_ONCE(1);
1316}
1317
1318static bool rcu_is_callbacks_kthread(void)
1319{
1320 return false;
1321}
1322
1323static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1324{
1325}
1326
1327static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1328{
1329}
1330
1331static void __init rcu_spawn_boost_kthreads(void)
1332{
1333}
1334
1335static void rcu_prepare_kthreads(int cpu)
1336{
1337}
1338
1339#endif /* #else #ifdef CONFIG_RCU_BOOST */
1340
1341#if !defined(CONFIG_RCU_FAST_NO_HZ)
1342
1343/*
1344 * Check to see if any future RCU-related work will need to be done
1345 * by the current CPU, even if none need be done immediately, returning
1346 * 1 if so. This function is part of the RCU implementation; it is -not-
1347 * an exported member of the RCU API.
1348 *
1349 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1350 * any flavor of RCU.
1351 */
1352int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1353{
1354 *nextevt = KTIME_MAX;
1355 return IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL)
1356 ? 0 : rcu_cpu_has_callbacks(NULL);
1357}
1358
1359/*
1360 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1361 * after it.
1362 */
1363static void rcu_cleanup_after_idle(void)
1364{
1365}
1366
1367/*
1368 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1369 * is nothing.
1370 */
1371static void rcu_prepare_for_idle(void)
1372{
1373}
1374
1375/*
1376 * Don't bother keeping a running count of the number of RCU callbacks
1377 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1378 */
1379static void rcu_idle_count_callbacks_posted(void)
1380{
1381}
1382
1383#else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1384
1385/*
1386 * This code is invoked when a CPU goes idle, at which point we want
1387 * to have the CPU do everything required for RCU so that it can enter
1388 * the energy-efficient dyntick-idle mode. This is handled by a
1389 * state machine implemented by rcu_prepare_for_idle() below.
1390 *
1391 * The following three proprocessor symbols control this state machine:
1392 *
1393 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1394 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1395 * is sized to be roughly one RCU grace period. Those energy-efficiency
1396 * benchmarkers who might otherwise be tempted to set this to a large
1397 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1398 * system. And if you are -that- concerned about energy efficiency,
1399 * just power the system down and be done with it!
1400 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1401 * permitted to sleep in dyntick-idle mode with only lazy RCU
1402 * callbacks pending. Setting this too high can OOM your system.
1403 *
1404 * The values below work well in practice. If future workloads require
1405 * adjustment, they can be converted into kernel config parameters, though
1406 * making the state machine smarter might be a better option.
1407 */
1408#define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1409#define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1410
1411static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1412module_param(rcu_idle_gp_delay, int, 0644);
1413static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1414module_param(rcu_idle_lazy_gp_delay, int, 0644);
1415
1416/*
1417 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1418 * only if it has been awhile since the last time we did so. Afterwards,
1419 * if there are any callbacks ready for immediate invocation, return true.
1420 */
1421static bool __maybe_unused rcu_try_advance_all_cbs(void)
1422{
1423 bool cbs_ready = false;
1424 struct rcu_data *rdp;
1425 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1426 struct rcu_node *rnp;
1427 struct rcu_state *rsp;
1428
1429 /* Exit early if we advanced recently. */
1430 if (jiffies == rdtp->last_advance_all)
1431 return false;
1432 rdtp->last_advance_all = jiffies;
1433
1434 for_each_rcu_flavor(rsp) {
1435 rdp = this_cpu_ptr(rsp->rda);
1436 rnp = rdp->mynode;
1437
1438 /*
1439 * Don't bother checking unless a grace period has
1440 * completed since we last checked and there are
1441 * callbacks not yet ready to invoke.
1442 */
1443 if ((rdp->completed != rnp->completed ||
1444 unlikely(READ_ONCE(rdp->gpwrap))) &&
1445 rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1446 note_gp_changes(rsp, rdp);
1447
1448 if (cpu_has_callbacks_ready_to_invoke(rdp))
1449 cbs_ready = true;
1450 }
1451 return cbs_ready;
1452}
1453
1454/*
1455 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1456 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1457 * caller to set the timeout based on whether or not there are non-lazy
1458 * callbacks.
1459 *
1460 * The caller must have disabled interrupts.
1461 */
1462int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1463{
1464 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1465 unsigned long dj;
1466
1467 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL)) {
1468 *nextevt = KTIME_MAX;
1469 return 0;
1470 }
1471
1472 /* Snapshot to detect later posting of non-lazy callback. */
1473 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1474
1475 /* If no callbacks, RCU doesn't need the CPU. */
1476 if (!rcu_cpu_has_callbacks(&rdtp->all_lazy)) {
1477 *nextevt = KTIME_MAX;
1478 return 0;
1479 }
1480
1481 /* Attempt to advance callbacks. */
1482 if (rcu_try_advance_all_cbs()) {
1483 /* Some ready to invoke, so initiate later invocation. */
1484 invoke_rcu_core();
1485 return 1;
1486 }
1487 rdtp->last_accelerate = jiffies;
1488
1489 /* Request timer delay depending on laziness, and round. */
1490 if (!rdtp->all_lazy) {
1491 dj = round_up(rcu_idle_gp_delay + jiffies,
1492 rcu_idle_gp_delay) - jiffies;
1493 } else {
1494 dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1495 }
1496 *nextevt = basemono + dj * TICK_NSEC;
1497 return 0;
1498}
1499
1500/*
1501 * Prepare a CPU for idle from an RCU perspective. The first major task
1502 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1503 * The second major task is to check to see if a non-lazy callback has
1504 * arrived at a CPU that previously had only lazy callbacks. The third
1505 * major task is to accelerate (that is, assign grace-period numbers to)
1506 * any recently arrived callbacks.
1507 *
1508 * The caller must have disabled interrupts.
1509 */
1510static void rcu_prepare_for_idle(void)
1511{
1512 bool needwake;
1513 struct rcu_data *rdp;
1514 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1515 struct rcu_node *rnp;
1516 struct rcu_state *rsp;
1517 int tne;
1518
1519 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) ||
1520 rcu_is_nocb_cpu(smp_processor_id()))
1521 return;
1522
1523 /* Handle nohz enablement switches conservatively. */
1524 tne = READ_ONCE(tick_nohz_active);
1525 if (tne != rdtp->tick_nohz_enabled_snap) {
1526 if (rcu_cpu_has_callbacks(NULL))
1527 invoke_rcu_core(); /* force nohz to see update. */
1528 rdtp->tick_nohz_enabled_snap = tne;
1529 return;
1530 }
1531 if (!tne)
1532 return;
1533
1534 /*
1535 * If a non-lazy callback arrived at a CPU having only lazy
1536 * callbacks, invoke RCU core for the side-effect of recalculating
1537 * idle duration on re-entry to idle.
1538 */
1539 if (rdtp->all_lazy &&
1540 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1541 rdtp->all_lazy = false;
1542 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1543 invoke_rcu_core();
1544 return;
1545 }
1546
1547 /*
1548 * If we have not yet accelerated this jiffy, accelerate all
1549 * callbacks on this CPU.
1550 */
1551 if (rdtp->last_accelerate == jiffies)
1552 return;
1553 rdtp->last_accelerate = jiffies;
1554 for_each_rcu_flavor(rsp) {
1555 rdp = this_cpu_ptr(rsp->rda);
1556 if (!*rdp->nxttail[RCU_DONE_TAIL])
1557 continue;
1558 rnp = rdp->mynode;
1559 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1560 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1561 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1562 if (needwake)
1563 rcu_gp_kthread_wake(rsp);
1564 }
1565}
1566
1567/*
1568 * Clean up for exit from idle. Attempt to advance callbacks based on
1569 * any grace periods that elapsed while the CPU was idle, and if any
1570 * callbacks are now ready to invoke, initiate invocation.
1571 */
1572static void rcu_cleanup_after_idle(void)
1573{
1574 if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) ||
1575 rcu_is_nocb_cpu(smp_processor_id()))
1576 return;
1577 if (rcu_try_advance_all_cbs())
1578 invoke_rcu_core();
1579}
1580
1581/*
1582 * Keep a running count of the number of non-lazy callbacks posted
1583 * on this CPU. This running counter (which is never decremented) allows
1584 * rcu_prepare_for_idle() to detect when something out of the idle loop
1585 * posts a callback, even if an equal number of callbacks are invoked.
1586 * Of course, callbacks should only be posted from within a trace event
1587 * designed to be called from idle or from within RCU_NONIDLE().
1588 */
1589static void rcu_idle_count_callbacks_posted(void)
1590{
1591 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1592}
1593
1594/*
1595 * Data for flushing lazy RCU callbacks at OOM time.
1596 */
1597static atomic_t oom_callback_count;
1598static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1599
1600/*
1601 * RCU OOM callback -- decrement the outstanding count and deliver the
1602 * wake-up if we are the last one.
1603 */
1604static void rcu_oom_callback(struct rcu_head *rhp)
1605{
1606 if (atomic_dec_and_test(&oom_callback_count))
1607 wake_up(&oom_callback_wq);
1608}
1609
1610/*
1611 * Post an rcu_oom_notify callback on the current CPU if it has at
1612 * least one lazy callback. This will unnecessarily post callbacks
1613 * to CPUs that already have a non-lazy callback at the end of their
1614 * callback list, but this is an infrequent operation, so accept some
1615 * extra overhead to keep things simple.
1616 */
1617static void rcu_oom_notify_cpu(void *unused)
1618{
1619 struct rcu_state *rsp;
1620 struct rcu_data *rdp;
1621
1622 for_each_rcu_flavor(rsp) {
1623 rdp = raw_cpu_ptr(rsp->rda);
1624 if (rdp->qlen_lazy != 0) {
1625 atomic_inc(&oom_callback_count);
1626 rsp->call(&rdp->oom_head, rcu_oom_callback);
1627 }
1628 }
1629}
1630
1631/*
1632 * If low on memory, ensure that each CPU has a non-lazy callback.
1633 * This will wake up CPUs that have only lazy callbacks, in turn
1634 * ensuring that they free up the corresponding memory in a timely manner.
1635 * Because an uncertain amount of memory will be freed in some uncertain
1636 * timeframe, we do not claim to have freed anything.
1637 */
1638static int rcu_oom_notify(struct notifier_block *self,
1639 unsigned long notused, void *nfreed)
1640{
1641 int cpu;
1642
1643 /* Wait for callbacks from earlier instance to complete. */
1644 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1645 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1646
1647 /*
1648 * Prevent premature wakeup: ensure that all increments happen
1649 * before there is a chance of the counter reaching zero.
1650 */
1651 atomic_set(&oom_callback_count, 1);
1652
1653 for_each_online_cpu(cpu) {
1654 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1655 cond_resched_rcu_qs();
1656 }
1657
1658 /* Unconditionally decrement: no need to wake ourselves up. */
1659 atomic_dec(&oom_callback_count);
1660
1661 return NOTIFY_OK;
1662}
1663
1664static struct notifier_block rcu_oom_nb = {
1665 .notifier_call = rcu_oom_notify
1666};
1667
1668static int __init rcu_register_oom_notifier(void)
1669{
1670 register_oom_notifier(&rcu_oom_nb);
1671 return 0;
1672}
1673early_initcall(rcu_register_oom_notifier);
1674
1675#endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1676
1677#ifdef CONFIG_RCU_FAST_NO_HZ
1678
1679static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1680{
1681 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1682 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1683
1684 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1685 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1686 ulong2long(nlpd),
1687 rdtp->all_lazy ? 'L' : '.',
1688 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1689}
1690
1691#else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1692
1693static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1694{
1695 *cp = '\0';
1696}
1697
1698#endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1699
1700/* Initiate the stall-info list. */
1701static void print_cpu_stall_info_begin(void)
1702{
1703 pr_cont("\n");
1704}
1705
1706/*
1707 * Print out diagnostic information for the specified stalled CPU.
1708 *
1709 * If the specified CPU is aware of the current RCU grace period
1710 * (flavor specified by rsp), then print the number of scheduling
1711 * clock interrupts the CPU has taken during the time that it has
1712 * been aware. Otherwise, print the number of RCU grace periods
1713 * that this CPU is ignorant of, for example, "1" if the CPU was
1714 * aware of the previous grace period.
1715 *
1716 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1717 */
1718static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1719{
1720 char fast_no_hz[72];
1721 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1722 struct rcu_dynticks *rdtp = rdp->dynticks;
1723 char *ticks_title;
1724 unsigned long ticks_value;
1725
1726 if (rsp->gpnum == rdp->gpnum) {
1727 ticks_title = "ticks this GP";
1728 ticks_value = rdp->ticks_this_gp;
1729 } else {
1730 ticks_title = "GPs behind";
1731 ticks_value = rsp->gpnum - rdp->gpnum;
1732 }
1733 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1734 pr_err("\t%d-%c%c%c: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u fqs=%ld %s\n",
1735 cpu,
1736 "O."[!!cpu_online(cpu)],
1737 "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
1738 "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
1739 ticks_value, ticks_title,
1740 atomic_read(&rdtp->dynticks) & 0xfff,
1741 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1742 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1743 READ_ONCE(rsp->n_force_qs) - rsp->n_force_qs_gpstart,
1744 fast_no_hz);
1745}
1746
1747/* Terminate the stall-info list. */
1748static void print_cpu_stall_info_end(void)
1749{
1750 pr_err("\t");
1751}
1752
1753/* Zero ->ticks_this_gp for all flavors of RCU. */
1754static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1755{
1756 rdp->ticks_this_gp = 0;
1757 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1758}
1759
1760/* Increment ->ticks_this_gp for all flavors of RCU. */
1761static void increment_cpu_stall_ticks(void)
1762{
1763 struct rcu_state *rsp;
1764
1765 for_each_rcu_flavor(rsp)
1766 raw_cpu_inc(rsp->rda->ticks_this_gp);
1767}
1768
1769#ifdef CONFIG_RCU_NOCB_CPU
1770
1771/*
1772 * Offload callback processing from the boot-time-specified set of CPUs
1773 * specified by rcu_nocb_mask. For each CPU in the set, there is a
1774 * kthread created that pulls the callbacks from the corresponding CPU,
1775 * waits for a grace period to elapse, and invokes the callbacks.
1776 * The no-CBs CPUs do a wake_up() on their kthread when they insert
1777 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1778 * has been specified, in which case each kthread actively polls its
1779 * CPU. (Which isn't so great for energy efficiency, but which does
1780 * reduce RCU's overhead on that CPU.)
1781 *
1782 * This is intended to be used in conjunction with Frederic Weisbecker's
1783 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1784 * running CPU-bound user-mode computations.
1785 *
1786 * Offloading of callback processing could also in theory be used as
1787 * an energy-efficiency measure because CPUs with no RCU callbacks
1788 * queued are more aggressive about entering dyntick-idle mode.
1789 */
1790
1791
1792/* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1793static int __init rcu_nocb_setup(char *str)
1794{
1795 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1796 have_rcu_nocb_mask = true;
1797 cpulist_parse(str, rcu_nocb_mask);
1798 return 1;
1799}
1800__setup("rcu_nocbs=", rcu_nocb_setup);
1801
1802static int __init parse_rcu_nocb_poll(char *arg)
1803{
1804 rcu_nocb_poll = 1;
1805 return 0;
1806}
1807early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1808
1809/*
1810 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1811 * grace period.
1812 */
1813static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
1814{
1815 swake_up_all(sq);
1816}
1817
1818/*
1819 * Set the root rcu_node structure's ->need_future_gp field
1820 * based on the sum of those of all rcu_node structures. This does
1821 * double-count the root rcu_node structure's requests, but this
1822 * is necessary to handle the possibility of a rcu_nocb_kthread()
1823 * having awakened during the time that the rcu_node structures
1824 * were being updated for the end of the previous grace period.
1825 */
1826static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
1827{
1828 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
1829}
1830
1831static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
1832{
1833 return &rnp->nocb_gp_wq[rnp->completed & 0x1];
1834}
1835
1836static void rcu_init_one_nocb(struct rcu_node *rnp)
1837{
1838 init_swait_queue_head(&rnp->nocb_gp_wq[0]);
1839 init_swait_queue_head(&rnp->nocb_gp_wq[1]);
1840}
1841
1842#ifndef CONFIG_RCU_NOCB_CPU_ALL
1843/* Is the specified CPU a no-CBs CPU? */
1844bool rcu_is_nocb_cpu(int cpu)
1845{
1846 if (have_rcu_nocb_mask)
1847 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1848 return false;
1849}
1850#endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1851
1852/*
1853 * Kick the leader kthread for this NOCB group.
1854 */
1855static void wake_nocb_leader(struct rcu_data *rdp, bool force)
1856{
1857 struct rcu_data *rdp_leader = rdp->nocb_leader;
1858
1859 if (!READ_ONCE(rdp_leader->nocb_kthread))
1860 return;
1861 if (READ_ONCE(rdp_leader->nocb_leader_sleep) || force) {
1862 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
1863 WRITE_ONCE(rdp_leader->nocb_leader_sleep, false);
1864 swake_up(&rdp_leader->nocb_wq);
1865 }
1866}
1867
1868/*
1869 * Does the specified CPU need an RCU callback for the specified flavor
1870 * of rcu_barrier()?
1871 */
1872static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
1873{
1874 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1875 unsigned long ret;
1876#ifdef CONFIG_PROVE_RCU
1877 struct rcu_head *rhp;
1878#endif /* #ifdef CONFIG_PROVE_RCU */
1879
1880 /*
1881 * Check count of all no-CBs callbacks awaiting invocation.
1882 * There needs to be a barrier before this function is called,
1883 * but associated with a prior determination that no more
1884 * callbacks would be posted. In the worst case, the first
1885 * barrier in _rcu_barrier() suffices (but the caller cannot
1886 * necessarily rely on this, not a substitute for the caller
1887 * getting the concurrency design right!). There must also be
1888 * a barrier between the following load an posting of a callback
1889 * (if a callback is in fact needed). This is associated with an
1890 * atomic_inc() in the caller.
1891 */
1892 ret = atomic_long_read(&rdp->nocb_q_count);
1893
1894#ifdef CONFIG_PROVE_RCU
1895 rhp = READ_ONCE(rdp->nocb_head);
1896 if (!rhp)
1897 rhp = READ_ONCE(rdp->nocb_gp_head);
1898 if (!rhp)
1899 rhp = READ_ONCE(rdp->nocb_follower_head);
1900
1901 /* Having no rcuo kthread but CBs after scheduler starts is bad! */
1902 if (!READ_ONCE(rdp->nocb_kthread) && rhp &&
1903 rcu_scheduler_fully_active) {
1904 /* RCU callback enqueued before CPU first came online??? */
1905 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
1906 cpu, rhp->func);
1907 WARN_ON_ONCE(1);
1908 }
1909#endif /* #ifdef CONFIG_PROVE_RCU */
1910
1911 return !!ret;
1912}
1913
1914/*
1915 * Enqueue the specified string of rcu_head structures onto the specified
1916 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
1917 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
1918 * counts are supplied by rhcount and rhcount_lazy.
1919 *
1920 * If warranted, also wake up the kthread servicing this CPUs queues.
1921 */
1922static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
1923 struct rcu_head *rhp,
1924 struct rcu_head **rhtp,
1925 int rhcount, int rhcount_lazy,
1926 unsigned long flags)
1927{
1928 int len;
1929 struct rcu_head **old_rhpp;
1930 struct task_struct *t;
1931
1932 /* Enqueue the callback on the nocb list and update counts. */
1933 atomic_long_add(rhcount, &rdp->nocb_q_count);
1934 /* rcu_barrier() relies on ->nocb_q_count add before xchg. */
1935 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
1936 WRITE_ONCE(*old_rhpp, rhp);
1937 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
1938 smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
1939
1940 /* If we are not being polled and there is a kthread, awaken it ... */
1941 t = READ_ONCE(rdp->nocb_kthread);
1942 if (rcu_nocb_poll || !t) {
1943 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1944 TPS("WakeNotPoll"));
1945 return;
1946 }
1947 len = atomic_long_read(&rdp->nocb_q_count);
1948 if (old_rhpp == &rdp->nocb_head) {
1949 if (!irqs_disabled_flags(flags)) {
1950 /* ... if queue was empty ... */
1951 wake_nocb_leader(rdp, false);
1952 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1953 TPS("WakeEmpty"));
1954 } else {
1955 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE;
1956 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1957 TPS("WakeEmptyIsDeferred"));
1958 }
1959 rdp->qlen_last_fqs_check = 0;
1960 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
1961 /* ... or if many callbacks queued. */
1962 if (!irqs_disabled_flags(flags)) {
1963 wake_nocb_leader(rdp, true);
1964 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1965 TPS("WakeOvf"));
1966 } else {
1967 rdp->nocb_defer_wakeup = RCU_NOGP_WAKE_FORCE;
1968 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1969 TPS("WakeOvfIsDeferred"));
1970 }
1971 rdp->qlen_last_fqs_check = LONG_MAX / 2;
1972 } else {
1973 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
1974 }
1975 return;
1976}
1977
1978/*
1979 * This is a helper for __call_rcu(), which invokes this when the normal
1980 * callback queue is inoperable. If this is not a no-CBs CPU, this
1981 * function returns failure back to __call_rcu(), which can complain
1982 * appropriately.
1983 *
1984 * Otherwise, this function queues the callback where the corresponding
1985 * "rcuo" kthread can find it.
1986 */
1987static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
1988 bool lazy, unsigned long flags)
1989{
1990
1991 if (!rcu_is_nocb_cpu(rdp->cpu))
1992 return false;
1993 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
1994 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
1995 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
1996 (unsigned long)rhp->func,
1997 -atomic_long_read(&rdp->nocb_q_count_lazy),
1998 -atomic_long_read(&rdp->nocb_q_count));
1999 else
2000 trace_rcu_callback(rdp->rsp->name, rhp,
2001 -atomic_long_read(&rdp->nocb_q_count_lazy),
2002 -atomic_long_read(&rdp->nocb_q_count));
2003
2004 /*
2005 * If called from an extended quiescent state with interrupts
2006 * disabled, invoke the RCU core in order to allow the idle-entry
2007 * deferred-wakeup check to function.
2008 */
2009 if (irqs_disabled_flags(flags) &&
2010 !rcu_is_watching() &&
2011 cpu_online(smp_processor_id()))
2012 invoke_rcu_core();
2013
2014 return true;
2015}
2016
2017/*
2018 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2019 * not a no-CBs CPU.
2020 */
2021static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2022 struct rcu_data *rdp,
2023 unsigned long flags)
2024{
2025 long ql = rsp->qlen;
2026 long qll = rsp->qlen_lazy;
2027
2028 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2029 if (!rcu_is_nocb_cpu(smp_processor_id()))
2030 return false;
2031 rsp->qlen = 0;
2032 rsp->qlen_lazy = 0;
2033
2034 /* First, enqueue the donelist, if any. This preserves CB ordering. */
2035 if (rsp->orphan_donelist != NULL) {
2036 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
2037 rsp->orphan_donetail, ql, qll, flags);
2038 ql = qll = 0;
2039 rsp->orphan_donelist = NULL;
2040 rsp->orphan_donetail = &rsp->orphan_donelist;
2041 }
2042 if (rsp->orphan_nxtlist != NULL) {
2043 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
2044 rsp->orphan_nxttail, ql, qll, flags);
2045 ql = qll = 0;
2046 rsp->orphan_nxtlist = NULL;
2047 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2048 }
2049 return true;
2050}
2051
2052/*
2053 * If necessary, kick off a new grace period, and either way wait
2054 * for a subsequent grace period to complete.
2055 */
2056static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2057{
2058 unsigned long c;
2059 bool d;
2060 unsigned long flags;
2061 bool needwake;
2062 struct rcu_node *rnp = rdp->mynode;
2063
2064 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2065 needwake = rcu_start_future_gp(rnp, rdp, &c);
2066 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2067 if (needwake)
2068 rcu_gp_kthread_wake(rdp->rsp);
2069
2070 /*
2071 * Wait for the grace period. Do so interruptibly to avoid messing
2072 * up the load average.
2073 */
2074 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
2075 for (;;) {
2076 swait_event_interruptible(
2077 rnp->nocb_gp_wq[c & 0x1],
2078 (d = ULONG_CMP_GE(READ_ONCE(rnp->completed), c)));
2079 if (likely(d))
2080 break;
2081 WARN_ON(signal_pending(current));
2082 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
2083 }
2084 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
2085 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2086}
2087
2088/*
2089 * Leaders come here to wait for additional callbacks to show up.
2090 * This function does not return until callbacks appear.
2091 */
2092static void nocb_leader_wait(struct rcu_data *my_rdp)
2093{
2094 bool firsttime = true;
2095 bool gotcbs;
2096 struct rcu_data *rdp;
2097 struct rcu_head **tail;
2098
2099wait_again:
2100
2101 /* Wait for callbacks to appear. */
2102 if (!rcu_nocb_poll) {
2103 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep");
2104 swait_event_interruptible(my_rdp->nocb_wq,
2105 !READ_ONCE(my_rdp->nocb_leader_sleep));
2106 /* Memory barrier handled by smp_mb() calls below and repoll. */
2107 } else if (firsttime) {
2108 firsttime = false; /* Don't drown trace log with "Poll"! */
2109 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll");
2110 }
2111
2112 /*
2113 * Each pass through the following loop checks a follower for CBs.
2114 * We are our own first follower. Any CBs found are moved to
2115 * nocb_gp_head, where they await a grace period.
2116 */
2117 gotcbs = false;
2118 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2119 rdp->nocb_gp_head = READ_ONCE(rdp->nocb_head);
2120 if (!rdp->nocb_gp_head)
2121 continue; /* No CBs here, try next follower. */
2122
2123 /* Move callbacks to wait-for-GP list, which is empty. */
2124 WRITE_ONCE(rdp->nocb_head, NULL);
2125 rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2126 gotcbs = true;
2127 }
2128
2129 /*
2130 * If there were no callbacks, sleep a bit, rescan after a
2131 * memory barrier, and go retry.
2132 */
2133 if (unlikely(!gotcbs)) {
2134 if (!rcu_nocb_poll)
2135 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu,
2136 "WokeEmpty");
2137 WARN_ON(signal_pending(current));
2138 schedule_timeout_interruptible(1);
2139
2140 /* Rescan in case we were a victim of memory ordering. */
2141 my_rdp->nocb_leader_sleep = true;
2142 smp_mb(); /* Ensure _sleep true before scan. */
2143 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower)
2144 if (READ_ONCE(rdp->nocb_head)) {
2145 /* Found CB, so short-circuit next wait. */
2146 my_rdp->nocb_leader_sleep = false;
2147 break;
2148 }
2149 goto wait_again;
2150 }
2151
2152 /* Wait for one grace period. */
2153 rcu_nocb_wait_gp(my_rdp);
2154
2155 /*
2156 * We left ->nocb_leader_sleep unset to reduce cache thrashing.
2157 * We set it now, but recheck for new callbacks while
2158 * traversing our follower list.
2159 */
2160 my_rdp->nocb_leader_sleep = true;
2161 smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */
2162
2163 /* Each pass through the following loop wakes a follower, if needed. */
2164 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2165 if (READ_ONCE(rdp->nocb_head))
2166 my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/
2167 if (!rdp->nocb_gp_head)
2168 continue; /* No CBs, so no need to wake follower. */
2169
2170 /* Append callbacks to follower's "done" list. */
2171 tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail);
2172 *tail = rdp->nocb_gp_head;
2173 smp_mb__after_atomic(); /* Store *tail before wakeup. */
2174 if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
2175 /*
2176 * List was empty, wake up the follower.
2177 * Memory barriers supplied by atomic_long_add().
2178 */
2179 swake_up(&rdp->nocb_wq);
2180 }
2181 }
2182
2183 /* If we (the leader) don't have CBs, go wait some more. */
2184 if (!my_rdp->nocb_follower_head)
2185 goto wait_again;
2186}
2187
2188/*
2189 * Followers come here to wait for additional callbacks to show up.
2190 * This function does not return until callbacks appear.
2191 */
2192static void nocb_follower_wait(struct rcu_data *rdp)
2193{
2194 bool firsttime = true;
2195
2196 for (;;) {
2197 if (!rcu_nocb_poll) {
2198 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2199 "FollowerSleep");
2200 swait_event_interruptible(rdp->nocb_wq,
2201 READ_ONCE(rdp->nocb_follower_head));
2202 } else if (firsttime) {
2203 /* Don't drown trace log with "Poll"! */
2204 firsttime = false;
2205 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll");
2206 }
2207 if (smp_load_acquire(&rdp->nocb_follower_head)) {
2208 /* ^^^ Ensure CB invocation follows _head test. */
2209 return;
2210 }
2211 if (!rcu_nocb_poll)
2212 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2213 "WokeEmpty");
2214 WARN_ON(signal_pending(current));
2215 schedule_timeout_interruptible(1);
2216 }
2217}
2218
2219/*
2220 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2221 * callbacks queued by the corresponding no-CBs CPU, however, there is
2222 * an optional leader-follower relationship so that the grace-period
2223 * kthreads don't have to do quite so many wakeups.
2224 */
2225static int rcu_nocb_kthread(void *arg)
2226{
2227 int c, cl;
2228 struct rcu_head *list;
2229 struct rcu_head *next;
2230 struct rcu_head **tail;
2231 struct rcu_data *rdp = arg;
2232
2233 /* Each pass through this loop invokes one batch of callbacks */
2234 for (;;) {
2235 /* Wait for callbacks. */
2236 if (rdp->nocb_leader == rdp)
2237 nocb_leader_wait(rdp);
2238 else
2239 nocb_follower_wait(rdp);
2240
2241 /* Pull the ready-to-invoke callbacks onto local list. */
2242 list = READ_ONCE(rdp->nocb_follower_head);
2243 BUG_ON(!list);
2244 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty");
2245 WRITE_ONCE(rdp->nocb_follower_head, NULL);
2246 tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head);
2247
2248 /* Each pass through the following loop invokes a callback. */
2249 trace_rcu_batch_start(rdp->rsp->name,
2250 atomic_long_read(&rdp->nocb_q_count_lazy),
2251 atomic_long_read(&rdp->nocb_q_count), -1);
2252 c = cl = 0;
2253 while (list) {
2254 next = list->next;
2255 /* Wait for enqueuing to complete, if needed. */
2256 while (next == NULL && &list->next != tail) {
2257 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2258 TPS("WaitQueue"));
2259 schedule_timeout_interruptible(1);
2260 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2261 TPS("WokeQueue"));
2262 next = list->next;
2263 }
2264 debug_rcu_head_unqueue(list);
2265 local_bh_disable();
2266 if (__rcu_reclaim(rdp->rsp->name, list))
2267 cl++;
2268 c++;
2269 local_bh_enable();
2270 list = next;
2271 }
2272 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2273 smp_mb__before_atomic(); /* _add after CB invocation. */
2274 atomic_long_add(-c, &rdp->nocb_q_count);
2275 atomic_long_add(-cl, &rdp->nocb_q_count_lazy);
2276 rdp->n_nocbs_invoked += c;
2277 }
2278 return 0;
2279}
2280
2281/* Is a deferred wakeup of rcu_nocb_kthread() required? */
2282static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2283{
2284 return READ_ONCE(rdp->nocb_defer_wakeup);
2285}
2286
2287/* Do a deferred wakeup of rcu_nocb_kthread(). */
2288static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2289{
2290 int ndw;
2291
2292 if (!rcu_nocb_need_deferred_wakeup(rdp))
2293 return;
2294 ndw = READ_ONCE(rdp->nocb_defer_wakeup);
2295 WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOGP_WAKE_NOT);
2296 wake_nocb_leader(rdp, ndw == RCU_NOGP_WAKE_FORCE);
2297 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake"));
2298}
2299
2300void __init rcu_init_nohz(void)
2301{
2302 int cpu;
2303 bool need_rcu_nocb_mask = true;
2304 struct rcu_state *rsp;
2305
2306#ifdef CONFIG_RCU_NOCB_CPU_NONE
2307 need_rcu_nocb_mask = false;
2308#endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
2309
2310#if defined(CONFIG_NO_HZ_FULL)
2311 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2312 need_rcu_nocb_mask = true;
2313#endif /* #if defined(CONFIG_NO_HZ_FULL) */
2314
2315 if (!have_rcu_nocb_mask && need_rcu_nocb_mask) {
2316 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2317 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2318 return;
2319 }
2320 have_rcu_nocb_mask = true;
2321 }
2322 if (!have_rcu_nocb_mask)
2323 return;
2324
2325#ifdef CONFIG_RCU_NOCB_CPU_ZERO
2326 pr_info("\tOffload RCU callbacks from CPU 0\n");
2327 cpumask_set_cpu(0, rcu_nocb_mask);
2328#endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
2329#ifdef CONFIG_RCU_NOCB_CPU_ALL
2330 pr_info("\tOffload RCU callbacks from all CPUs\n");
2331 cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
2332#endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
2333#if defined(CONFIG_NO_HZ_FULL)
2334 if (tick_nohz_full_running)
2335 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2336#endif /* #if defined(CONFIG_NO_HZ_FULL) */
2337
2338 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2339 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
2340 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2341 rcu_nocb_mask);
2342 }
2343 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2344 cpumask_pr_args(rcu_nocb_mask));
2345 if (rcu_nocb_poll)
2346 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2347
2348 for_each_rcu_flavor(rsp) {
2349 for_each_cpu(cpu, rcu_nocb_mask)
2350 init_nocb_callback_list(per_cpu_ptr(rsp->rda, cpu));
2351 rcu_organize_nocb_kthreads(rsp);
2352 }
2353}
2354
2355/* Initialize per-rcu_data variables for no-CBs CPUs. */
2356static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2357{
2358 rdp->nocb_tail = &rdp->nocb_head;
2359 init_swait_queue_head(&rdp->nocb_wq);
2360 rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2361}
2362
2363/*
2364 * If the specified CPU is a no-CBs CPU that does not already have its
2365 * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are
2366 * brought online out of order, this can require re-organizing the
2367 * leader-follower relationships.
2368 */
2369static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu)
2370{
2371 struct rcu_data *rdp;
2372 struct rcu_data *rdp_last;
2373 struct rcu_data *rdp_old_leader;
2374 struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu);
2375 struct task_struct *t;
2376
2377 /*
2378 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2379 * then nothing to do.
2380 */
2381 if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread)
2382 return;
2383
2384 /* If we didn't spawn the leader first, reorganize! */
2385 rdp_old_leader = rdp_spawn->nocb_leader;
2386 if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) {
2387 rdp_last = NULL;
2388 rdp = rdp_old_leader;
2389 do {
2390 rdp->nocb_leader = rdp_spawn;
2391 if (rdp_last && rdp != rdp_spawn)
2392 rdp_last->nocb_next_follower = rdp;
2393 if (rdp == rdp_spawn) {
2394 rdp = rdp->nocb_next_follower;
2395 } else {
2396 rdp_last = rdp;
2397 rdp = rdp->nocb_next_follower;
2398 rdp_last->nocb_next_follower = NULL;
2399 }
2400 } while (rdp);
2401 rdp_spawn->nocb_next_follower = rdp_old_leader;
2402 }
2403
2404 /* Spawn the kthread for this CPU and RCU flavor. */
2405 t = kthread_run(rcu_nocb_kthread, rdp_spawn,
2406 "rcuo%c/%d", rsp->abbr, cpu);
2407 BUG_ON(IS_ERR(t));
2408 WRITE_ONCE(rdp_spawn->nocb_kthread, t);
2409}
2410
2411/*
2412 * If the specified CPU is a no-CBs CPU that does not already have its
2413 * rcuo kthreads, spawn them.
2414 */
2415static void rcu_spawn_all_nocb_kthreads(int cpu)
2416{
2417 struct rcu_state *rsp;
2418
2419 if (rcu_scheduler_fully_active)
2420 for_each_rcu_flavor(rsp)
2421 rcu_spawn_one_nocb_kthread(rsp, cpu);
2422}
2423
2424/*
2425 * Once the scheduler is running, spawn rcuo kthreads for all online
2426 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2427 * non-boot CPUs come online -- if this changes, we will need to add
2428 * some mutual exclusion.
2429 */
2430static void __init rcu_spawn_nocb_kthreads(void)
2431{
2432 int cpu;
2433
2434 for_each_online_cpu(cpu)
2435 rcu_spawn_all_nocb_kthreads(cpu);
2436}
2437
2438/* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */
2439static int rcu_nocb_leader_stride = -1;
2440module_param(rcu_nocb_leader_stride, int, 0444);
2441
2442/*
2443 * Initialize leader-follower relationships for all no-CBs CPU.
2444 */
2445static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp)
2446{
2447 int cpu;
2448 int ls = rcu_nocb_leader_stride;
2449 int nl = 0; /* Next leader. */
2450 struct rcu_data *rdp;
2451 struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */
2452 struct rcu_data *rdp_prev = NULL;
2453
2454 if (!have_rcu_nocb_mask)
2455 return;
2456 if (ls == -1) {
2457 ls = int_sqrt(nr_cpu_ids);
2458 rcu_nocb_leader_stride = ls;
2459 }
2460
2461 /*
2462 * Each pass through this loop sets up one rcu_data structure and
2463 * spawns one rcu_nocb_kthread().
2464 */
2465 for_each_cpu(cpu, rcu_nocb_mask) {
2466 rdp = per_cpu_ptr(rsp->rda, cpu);
2467 if (rdp->cpu >= nl) {
2468 /* New leader, set up for followers & next leader. */
2469 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2470 rdp->nocb_leader = rdp;
2471 rdp_leader = rdp;
2472 } else {
2473 /* Another follower, link to previous leader. */
2474 rdp->nocb_leader = rdp_leader;
2475 rdp_prev->nocb_next_follower = rdp;
2476 }
2477 rdp_prev = rdp;
2478 }
2479}
2480
2481/* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2482static bool init_nocb_callback_list(struct rcu_data *rdp)
2483{
2484 if (!rcu_is_nocb_cpu(rdp->cpu))
2485 return false;
2486
2487 /* If there are early-boot callbacks, move them to nocb lists. */
2488 if (rdp->nxtlist) {
2489 rdp->nocb_head = rdp->nxtlist;
2490 rdp->nocb_tail = rdp->nxttail[RCU_NEXT_TAIL];
2491 atomic_long_set(&rdp->nocb_q_count, rdp->qlen);
2492 atomic_long_set(&rdp->nocb_q_count_lazy, rdp->qlen_lazy);
2493 rdp->nxtlist = NULL;
2494 rdp->qlen = 0;
2495 rdp->qlen_lazy = 0;
2496 }
2497 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2498 return true;
2499}
2500
2501#else /* #ifdef CONFIG_RCU_NOCB_CPU */
2502
2503static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
2504{
2505 WARN_ON_ONCE(1); /* Should be dead code. */
2506 return false;
2507}
2508
2509static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
2510{
2511}
2512
2513static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2514{
2515}
2516
2517static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
2518{
2519 return NULL;
2520}
2521
2522static void rcu_init_one_nocb(struct rcu_node *rnp)
2523{
2524}
2525
2526static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2527 bool lazy, unsigned long flags)
2528{
2529 return false;
2530}
2531
2532static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2533 struct rcu_data *rdp,
2534 unsigned long flags)
2535{
2536 return false;
2537}
2538
2539static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2540{
2541}
2542
2543static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2544{
2545 return false;
2546}
2547
2548static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2549{
2550}
2551
2552static void rcu_spawn_all_nocb_kthreads(int cpu)
2553{
2554}
2555
2556static void __init rcu_spawn_nocb_kthreads(void)
2557{
2558}
2559
2560static bool init_nocb_callback_list(struct rcu_data *rdp)
2561{
2562 return false;
2563}
2564
2565#endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2566
2567/*
2568 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2569 * arbitrarily long period of time with the scheduling-clock tick turned
2570 * off. RCU will be paying attention to this CPU because it is in the
2571 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2572 * machine because the scheduling-clock tick has been disabled. Therefore,
2573 * if an adaptive-ticks CPU is failing to respond to the current grace
2574 * period and has not be idle from an RCU perspective, kick it.
2575 */
2576static void __maybe_unused rcu_kick_nohz_cpu(int cpu)
2577{
2578#ifdef CONFIG_NO_HZ_FULL
2579 if (tick_nohz_full_cpu(cpu))
2580 smp_send_reschedule(cpu);
2581#endif /* #ifdef CONFIG_NO_HZ_FULL */
2582}
2583
2584
2585#ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2586
2587static int full_sysidle_state; /* Current system-idle state. */
2588#define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2589#define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2590#define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2591#define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2592#define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2593
2594/*
2595 * Invoked to note exit from irq or task transition to idle. Note that
2596 * usermode execution does -not- count as idle here! After all, we want
2597 * to detect full-system idle states, not RCU quiescent states and grace
2598 * periods. The caller must have disabled interrupts.
2599 */
2600static void rcu_sysidle_enter(int irq)
2601{
2602 unsigned long j;
2603 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2604
2605 /* If there are no nohz_full= CPUs, no need to track this. */
2606 if (!tick_nohz_full_enabled())
2607 return;
2608
2609 /* Adjust nesting, check for fully idle. */
2610 if (irq) {
2611 rdtp->dynticks_idle_nesting--;
2612 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2613 if (rdtp->dynticks_idle_nesting != 0)
2614 return; /* Still not fully idle. */
2615 } else {
2616 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2617 DYNTICK_TASK_NEST_VALUE) {
2618 rdtp->dynticks_idle_nesting = 0;
2619 } else {
2620 rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2621 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2622 return; /* Still not fully idle. */
2623 }
2624 }
2625
2626 /* Record start of fully idle period. */
2627 j = jiffies;
2628 WRITE_ONCE(rdtp->dynticks_idle_jiffies, j);
2629 smp_mb__before_atomic();
2630 atomic_inc(&rdtp->dynticks_idle);
2631 smp_mb__after_atomic();
2632 WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2633}
2634
2635/*
2636 * Unconditionally force exit from full system-idle state. This is
2637 * invoked when a normal CPU exits idle, but must be called separately
2638 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2639 * is that the timekeeping CPU is permitted to take scheduling-clock
2640 * interrupts while the system is in system-idle state, and of course
2641 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2642 * interrupt from any other type of interrupt.
2643 */
2644void rcu_sysidle_force_exit(void)
2645{
2646 int oldstate = READ_ONCE(full_sysidle_state);
2647 int newoldstate;
2648
2649 /*
2650 * Each pass through the following loop attempts to exit full
2651 * system-idle state. If contention proves to be a problem,
2652 * a trylock-based contention tree could be used here.
2653 */
2654 while (oldstate > RCU_SYSIDLE_SHORT) {
2655 newoldstate = cmpxchg(&full_sysidle_state,
2656 oldstate, RCU_SYSIDLE_NOT);
2657 if (oldstate == newoldstate &&
2658 oldstate == RCU_SYSIDLE_FULL_NOTED) {
2659 rcu_kick_nohz_cpu(tick_do_timer_cpu);
2660 return; /* We cleared it, done! */
2661 }
2662 oldstate = newoldstate;
2663 }
2664 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2665}
2666
2667/*
2668 * Invoked to note entry to irq or task transition from idle. Note that
2669 * usermode execution does -not- count as idle here! The caller must
2670 * have disabled interrupts.
2671 */
2672static void rcu_sysidle_exit(int irq)
2673{
2674 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
2675
2676 /* If there are no nohz_full= CPUs, no need to track this. */
2677 if (!tick_nohz_full_enabled())
2678 return;
2679
2680 /* Adjust nesting, check for already non-idle. */
2681 if (irq) {
2682 rdtp->dynticks_idle_nesting++;
2683 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2684 if (rdtp->dynticks_idle_nesting != 1)
2685 return; /* Already non-idle. */
2686 } else {
2687 /*
2688 * Allow for irq misnesting. Yes, it really is possible
2689 * to enter an irq handler then never leave it, and maybe
2690 * also vice versa. Handle both possibilities.
2691 */
2692 if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2693 rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2694 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2695 return; /* Already non-idle. */
2696 } else {
2697 rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2698 }
2699 }
2700
2701 /* Record end of idle period. */
2702 smp_mb__before_atomic();
2703 atomic_inc(&rdtp->dynticks_idle);
2704 smp_mb__after_atomic();
2705 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2706
2707 /*
2708 * If we are the timekeeping CPU, we are permitted to be non-idle
2709 * during a system-idle state. This must be the case, because
2710 * the timekeeping CPU has to take scheduling-clock interrupts
2711 * during the time that the system is transitioning to full
2712 * system-idle state. This means that the timekeeping CPU must
2713 * invoke rcu_sysidle_force_exit() directly if it does anything
2714 * more than take a scheduling-clock interrupt.
2715 */
2716 if (smp_processor_id() == tick_do_timer_cpu)
2717 return;
2718
2719 /* Update system-idle state: We are clearly no longer fully idle! */
2720 rcu_sysidle_force_exit();
2721}
2722
2723/*
2724 * Check to see if the current CPU is idle. Note that usermode execution
2725 * does not count as idle. The caller must have disabled interrupts,
2726 * and must be running on tick_do_timer_cpu.
2727 */
2728static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2729 unsigned long *maxj)
2730{
2731 int cur;
2732 unsigned long j;
2733 struct rcu_dynticks *rdtp = rdp->dynticks;
2734
2735 /* If there are no nohz_full= CPUs, don't check system-wide idleness. */
2736 if (!tick_nohz_full_enabled())
2737 return;
2738
2739 /*
2740 * If some other CPU has already reported non-idle, if this is
2741 * not the flavor of RCU that tracks sysidle state, or if this
2742 * is an offline or the timekeeping CPU, nothing to do.
2743 */
2744 if (!*isidle || rdp->rsp != rcu_state_p ||
2745 cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2746 return;
2747 /* Verify affinity of current kthread. */
2748 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2749
2750 /* Pick up current idle and NMI-nesting counter and check. */
2751 cur = atomic_read(&rdtp->dynticks_idle);
2752 if (cur & 0x1) {
2753 *isidle = false; /* We are not idle! */
2754 return;
2755 }
2756 smp_mb(); /* Read counters before timestamps. */
2757
2758 /* Pick up timestamps. */
2759 j = READ_ONCE(rdtp->dynticks_idle_jiffies);
2760 /* If this CPU entered idle more recently, update maxj timestamp. */
2761 if (ULONG_CMP_LT(*maxj, j))
2762 *maxj = j;
2763}
2764
2765/*
2766 * Is this the flavor of RCU that is handling full-system idle?
2767 */
2768static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2769{
2770 return rsp == rcu_state_p;
2771}
2772
2773/*
2774 * Return a delay in jiffies based on the number of CPUs, rcu_node
2775 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2776 * systems more time to transition to full-idle state in order to
2777 * avoid the cache thrashing that otherwise occur on the state variable.
2778 * Really small systems (less than a couple of tens of CPUs) should
2779 * instead use a single global atomically incremented counter, and later
2780 * versions of this will automatically reconfigure themselves accordingly.
2781 */
2782static unsigned long rcu_sysidle_delay(void)
2783{
2784 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2785 return 0;
2786 return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2787}
2788
2789/*
2790 * Advance the full-system-idle state. This is invoked when all of
2791 * the non-timekeeping CPUs are idle.
2792 */
2793static void rcu_sysidle(unsigned long j)
2794{
2795 /* Check the current state. */
2796 switch (READ_ONCE(full_sysidle_state)) {
2797 case RCU_SYSIDLE_NOT:
2798
2799 /* First time all are idle, so note a short idle period. */
2800 WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_SHORT);
2801 break;
2802
2803 case RCU_SYSIDLE_SHORT:
2804
2805 /*
2806 * Idle for a bit, time to advance to next state?
2807 * cmpxchg failure means race with non-idle, let them win.
2808 */
2809 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2810 (void)cmpxchg(&full_sysidle_state,
2811 RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2812 break;
2813
2814 case RCU_SYSIDLE_LONG:
2815
2816 /*
2817 * Do an additional check pass before advancing to full.
2818 * cmpxchg failure means race with non-idle, let them win.
2819 */
2820 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2821 (void)cmpxchg(&full_sysidle_state,
2822 RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2823 break;
2824
2825 default:
2826 break;
2827 }
2828}
2829
2830/*
2831 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2832 * back to the beginning.
2833 */
2834static void rcu_sysidle_cancel(void)
2835{
2836 smp_mb();
2837 if (full_sysidle_state > RCU_SYSIDLE_SHORT)
2838 WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_NOT);
2839}
2840
2841/*
2842 * Update the sysidle state based on the results of a force-quiescent-state
2843 * scan of the CPUs' dyntick-idle state.
2844 */
2845static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
2846 unsigned long maxj, bool gpkt)
2847{
2848 if (rsp != rcu_state_p)
2849 return; /* Wrong flavor, ignore. */
2850 if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2851 return; /* Running state machine from timekeeping CPU. */
2852 if (isidle)
2853 rcu_sysidle(maxj); /* More idle! */
2854 else
2855 rcu_sysidle_cancel(); /* Idle is over. */
2856}
2857
2858/*
2859 * Wrapper for rcu_sysidle_report() when called from the grace-period
2860 * kthread's context.
2861 */
2862static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2863 unsigned long maxj)
2864{
2865 /* If there are no nohz_full= CPUs, no need to track this. */
2866 if (!tick_nohz_full_enabled())
2867 return;
2868
2869 rcu_sysidle_report(rsp, isidle, maxj, true);
2870}
2871
2872/* Callback and function for forcing an RCU grace period. */
2873struct rcu_sysidle_head {
2874 struct rcu_head rh;
2875 int inuse;
2876};
2877
2878static void rcu_sysidle_cb(struct rcu_head *rhp)
2879{
2880 struct rcu_sysidle_head *rshp;
2881
2882 /*
2883 * The following memory barrier is needed to replace the
2884 * memory barriers that would normally be in the memory
2885 * allocator.
2886 */
2887 smp_mb(); /* grace period precedes setting inuse. */
2888
2889 rshp = container_of(rhp, struct rcu_sysidle_head, rh);
2890 WRITE_ONCE(rshp->inuse, 0);
2891}
2892
2893/*
2894 * Check to see if the system is fully idle, other than the timekeeping CPU.
2895 * The caller must have disabled interrupts. This is not intended to be
2896 * called unless tick_nohz_full_enabled().
2897 */
2898bool rcu_sys_is_idle(void)
2899{
2900 static struct rcu_sysidle_head rsh;
2901 int rss = READ_ONCE(full_sysidle_state);
2902
2903 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
2904 return false;
2905
2906 /* Handle small-system case by doing a full scan of CPUs. */
2907 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
2908 int oldrss = rss - 1;
2909
2910 /*
2911 * One pass to advance to each state up to _FULL.
2912 * Give up if any pass fails to advance the state.
2913 */
2914 while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
2915 int cpu;
2916 bool isidle = true;
2917 unsigned long maxj = jiffies - ULONG_MAX / 4;
2918 struct rcu_data *rdp;
2919
2920 /* Scan all the CPUs looking for nonidle CPUs. */
2921 for_each_possible_cpu(cpu) {
2922 rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
2923 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
2924 if (!isidle)
2925 break;
2926 }
2927 rcu_sysidle_report(rcu_state_p, isidle, maxj, false);
2928 oldrss = rss;
2929 rss = READ_ONCE(full_sysidle_state);
2930 }
2931 }
2932
2933 /* If this is the first observation of an idle period, record it. */
2934 if (rss == RCU_SYSIDLE_FULL) {
2935 rss = cmpxchg(&full_sysidle_state,
2936 RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
2937 return rss == RCU_SYSIDLE_FULL;
2938 }
2939
2940 smp_mb(); /* ensure rss load happens before later caller actions. */
2941
2942 /* If already fully idle, tell the caller (in case of races). */
2943 if (rss == RCU_SYSIDLE_FULL_NOTED)
2944 return true;
2945
2946 /*
2947 * If we aren't there yet, and a grace period is not in flight,
2948 * initiate a grace period. Either way, tell the caller that
2949 * we are not there yet. We use an xchg() rather than an assignment
2950 * to make up for the memory barriers that would otherwise be
2951 * provided by the memory allocator.
2952 */
2953 if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
2954 !rcu_gp_in_progress(rcu_state_p) &&
2955 !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
2956 call_rcu(&rsh.rh, rcu_sysidle_cb);
2957 return false;
2958}
2959
2960/*
2961 * Initialize dynticks sysidle state for CPUs coming online.
2962 */
2963static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2964{
2965 rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
2966}
2967
2968#else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2969
2970static void rcu_sysidle_enter(int irq)
2971{
2972}
2973
2974static void rcu_sysidle_exit(int irq)
2975{
2976}
2977
2978static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2979 unsigned long *maxj)
2980{
2981}
2982
2983static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2984{
2985 return false;
2986}
2987
2988static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2989 unsigned long maxj)
2990{
2991}
2992
2993static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2994{
2995}
2996
2997#endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2998
2999/*
3000 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
3001 * grace-period kthread will do force_quiescent_state() processing?
3002 * The idea is to avoid waking up RCU core processing on such a
3003 * CPU unless the grace period has extended for too long.
3004 *
3005 * This code relies on the fact that all NO_HZ_FULL CPUs are also
3006 * CONFIG_RCU_NOCB_CPU CPUs.
3007 */
3008static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
3009{
3010#ifdef CONFIG_NO_HZ_FULL
3011 if (tick_nohz_full_cpu(smp_processor_id()) &&
3012 (!rcu_gp_in_progress(rsp) ||
3013 ULONG_CMP_LT(jiffies, READ_ONCE(rsp->gp_start) + HZ)))
3014 return true;
3015#endif /* #ifdef CONFIG_NO_HZ_FULL */
3016 return false;
3017}
3018
3019/*
3020 * Bind the grace-period kthread for the sysidle flavor of RCU to the
3021 * timekeeping CPU.
3022 */
3023static void rcu_bind_gp_kthread(void)
3024{
3025 int __maybe_unused cpu;
3026
3027 if (!tick_nohz_full_enabled())
3028 return;
3029#ifdef CONFIG_NO_HZ_FULL_SYSIDLE
3030 cpu = tick_do_timer_cpu;
3031 if (cpu >= 0 && cpu < nr_cpu_ids)
3032 set_cpus_allowed_ptr(current, cpumask_of(cpu));
3033#else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3034 housekeeping_affine(current);
3035#endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3036}
3037
3038/* Record the current task on dyntick-idle entry. */
3039static void rcu_dynticks_task_enter(void)
3040{
3041#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3042 WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
3043#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
3044}
3045
3046/* Record no current task on dyntick-idle exit. */
3047static void rcu_dynticks_task_exit(void)
3048{
3049#if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3050 WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
3051#endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
3052}
1/* SPDX-License-Identifier: GPL-2.0+ */
2/*
3 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
4 * Internal non-public definitions that provide either classic
5 * or preemptible semantics.
6 *
7 * Copyright Red Hat, 2009
8 * Copyright IBM Corporation, 2009
9 *
10 * Author: Ingo Molnar <mingo@elte.hu>
11 * Paul E. McKenney <paulmck@linux.ibm.com>
12 */
13
14#include "../locking/rtmutex_common.h"
15
16static bool rcu_rdp_is_offloaded(struct rcu_data *rdp)
17{
18 /*
19 * In order to read the offloaded state of an rdp in a safe
20 * and stable way and prevent from its value to be changed
21 * under us, we must either hold the barrier mutex, the cpu
22 * hotplug lock (read or write) or the nocb lock. Local
23 * non-preemptible reads are also safe. NOCB kthreads and
24 * timers have their own means of synchronization against the
25 * offloaded state updaters.
26 */
27 RCU_LOCKDEP_WARN(
28 !(lockdep_is_held(&rcu_state.barrier_mutex) ||
29 (IS_ENABLED(CONFIG_HOTPLUG_CPU) && lockdep_is_cpus_held()) ||
30 rcu_lockdep_is_held_nocb(rdp) ||
31 (rdp == this_cpu_ptr(&rcu_data) &&
32 !(IS_ENABLED(CONFIG_PREEMPT_COUNT) && preemptible())) ||
33 rcu_current_is_nocb_kthread(rdp)),
34 "Unsafe read of RCU_NOCB offloaded state"
35 );
36
37 return rcu_segcblist_is_offloaded(&rdp->cblist);
38}
39
40/*
41 * Check the RCU kernel configuration parameters and print informative
42 * messages about anything out of the ordinary.
43 */
44static void __init rcu_bootup_announce_oddness(void)
45{
46 if (IS_ENABLED(CONFIG_RCU_TRACE))
47 pr_info("\tRCU event tracing is enabled.\n");
48 if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
49 (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
50 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d.\n",
51 RCU_FANOUT);
52 if (rcu_fanout_exact)
53 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
54 if (IS_ENABLED(CONFIG_PROVE_RCU))
55 pr_info("\tRCU lockdep checking is enabled.\n");
56 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
57 pr_info("\tRCU strict (and thus non-scalable) grace periods are enabled.\n");
58 if (RCU_NUM_LVLS >= 4)
59 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
60 if (RCU_FANOUT_LEAF != 16)
61 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
62 RCU_FANOUT_LEAF);
63 if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
64 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n",
65 rcu_fanout_leaf);
66 if (nr_cpu_ids != NR_CPUS)
67 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%u.\n", NR_CPUS, nr_cpu_ids);
68#ifdef CONFIG_RCU_BOOST
69 pr_info("\tRCU priority boosting: priority %d delay %d ms.\n",
70 kthread_prio, CONFIG_RCU_BOOST_DELAY);
71#endif
72 if (blimit != DEFAULT_RCU_BLIMIT)
73 pr_info("\tBoot-time adjustment of callback invocation limit to %ld.\n", blimit);
74 if (qhimark != DEFAULT_RCU_QHIMARK)
75 pr_info("\tBoot-time adjustment of callback high-water mark to %ld.\n", qhimark);
76 if (qlowmark != DEFAULT_RCU_QLOMARK)
77 pr_info("\tBoot-time adjustment of callback low-water mark to %ld.\n", qlowmark);
78 if (qovld != DEFAULT_RCU_QOVLD)
79 pr_info("\tBoot-time adjustment of callback overload level to %ld.\n", qovld);
80 if (jiffies_till_first_fqs != ULONG_MAX)
81 pr_info("\tBoot-time adjustment of first FQS scan delay to %ld jiffies.\n", jiffies_till_first_fqs);
82 if (jiffies_till_next_fqs != ULONG_MAX)
83 pr_info("\tBoot-time adjustment of subsequent FQS scan delay to %ld jiffies.\n", jiffies_till_next_fqs);
84 if (jiffies_till_sched_qs != ULONG_MAX)
85 pr_info("\tBoot-time adjustment of scheduler-enlistment delay to %ld jiffies.\n", jiffies_till_sched_qs);
86 if (rcu_kick_kthreads)
87 pr_info("\tKick kthreads if too-long grace period.\n");
88 if (IS_ENABLED(CONFIG_DEBUG_OBJECTS_RCU_HEAD))
89 pr_info("\tRCU callback double-/use-after-free debug is enabled.\n");
90 if (gp_preinit_delay)
91 pr_info("\tRCU debug GP pre-init slowdown %d jiffies.\n", gp_preinit_delay);
92 if (gp_init_delay)
93 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_init_delay);
94 if (gp_cleanup_delay)
95 pr_info("\tRCU debug GP cleanup slowdown %d jiffies.\n", gp_cleanup_delay);
96 if (!use_softirq)
97 pr_info("\tRCU_SOFTIRQ processing moved to rcuc kthreads.\n");
98 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG))
99 pr_info("\tRCU debug extended QS entry/exit.\n");
100 rcupdate_announce_bootup_oddness();
101}
102
103#ifdef CONFIG_PREEMPT_RCU
104
105static void rcu_report_exp_rnp(struct rcu_node *rnp, bool wake);
106static void rcu_read_unlock_special(struct task_struct *t);
107
108/*
109 * Tell them what RCU they are running.
110 */
111static void __init rcu_bootup_announce(void)
112{
113 pr_info("Preemptible hierarchical RCU implementation.\n");
114 rcu_bootup_announce_oddness();
115}
116
117/* Flags for rcu_preempt_ctxt_queue() decision table. */
118#define RCU_GP_TASKS 0x8
119#define RCU_EXP_TASKS 0x4
120#define RCU_GP_BLKD 0x2
121#define RCU_EXP_BLKD 0x1
122
123/*
124 * Queues a task preempted within an RCU-preempt read-side critical
125 * section into the appropriate location within the ->blkd_tasks list,
126 * depending on the states of any ongoing normal and expedited grace
127 * periods. The ->gp_tasks pointer indicates which element the normal
128 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
129 * indicates which element the expedited grace period is waiting on (again,
130 * NULL if none). If a grace period is waiting on a given element in the
131 * ->blkd_tasks list, it also waits on all subsequent elements. Thus,
132 * adding a task to the tail of the list blocks any grace period that is
133 * already waiting on one of the elements. In contrast, adding a task
134 * to the head of the list won't block any grace period that is already
135 * waiting on one of the elements.
136 *
137 * This queuing is imprecise, and can sometimes make an ongoing grace
138 * period wait for a task that is not strictly speaking blocking it.
139 * Given the choice, we needlessly block a normal grace period rather than
140 * blocking an expedited grace period.
141 *
142 * Note that an endless sequence of expedited grace periods still cannot
143 * indefinitely postpone a normal grace period. Eventually, all of the
144 * fixed number of preempted tasks blocking the normal grace period that are
145 * not also blocking the expedited grace period will resume and complete
146 * their RCU read-side critical sections. At that point, the ->gp_tasks
147 * pointer will equal the ->exp_tasks pointer, at which point the end of
148 * the corresponding expedited grace period will also be the end of the
149 * normal grace period.
150 */
151static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
152 __releases(rnp->lock) /* But leaves rrupts disabled. */
153{
154 int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
155 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
156 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
157 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
158 struct task_struct *t = current;
159
160 raw_lockdep_assert_held_rcu_node(rnp);
161 WARN_ON_ONCE(rdp->mynode != rnp);
162 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
163 /* RCU better not be waiting on newly onlined CPUs! */
164 WARN_ON_ONCE(rnp->qsmaskinitnext & ~rnp->qsmaskinit & rnp->qsmask &
165 rdp->grpmask);
166
167 /*
168 * Decide where to queue the newly blocked task. In theory,
169 * this could be an if-statement. In practice, when I tried
170 * that, it was quite messy.
171 */
172 switch (blkd_state) {
173 case 0:
174 case RCU_EXP_TASKS:
175 case RCU_EXP_TASKS + RCU_GP_BLKD:
176 case RCU_GP_TASKS:
177 case RCU_GP_TASKS + RCU_EXP_TASKS:
178
179 /*
180 * Blocking neither GP, or first task blocking the normal
181 * GP but not blocking the already-waiting expedited GP.
182 * Queue at the head of the list to avoid unnecessarily
183 * blocking the already-waiting GPs.
184 */
185 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
186 break;
187
188 case RCU_EXP_BLKD:
189 case RCU_GP_BLKD:
190 case RCU_GP_BLKD + RCU_EXP_BLKD:
191 case RCU_GP_TASKS + RCU_EXP_BLKD:
192 case RCU_GP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
193 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
194
195 /*
196 * First task arriving that blocks either GP, or first task
197 * arriving that blocks the expedited GP (with the normal
198 * GP already waiting), or a task arriving that blocks
199 * both GPs with both GPs already waiting. Queue at the
200 * tail of the list to avoid any GP waiting on any of the
201 * already queued tasks that are not blocking it.
202 */
203 list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
204 break;
205
206 case RCU_EXP_TASKS + RCU_EXP_BLKD:
207 case RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
208 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_EXP_BLKD:
209
210 /*
211 * Second or subsequent task blocking the expedited GP.
212 * The task either does not block the normal GP, or is the
213 * first task blocking the normal GP. Queue just after
214 * the first task blocking the expedited GP.
215 */
216 list_add(&t->rcu_node_entry, rnp->exp_tasks);
217 break;
218
219 case RCU_GP_TASKS + RCU_GP_BLKD:
220 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
221
222 /*
223 * Second or subsequent task blocking the normal GP.
224 * The task does not block the expedited GP. Queue just
225 * after the first task blocking the normal GP.
226 */
227 list_add(&t->rcu_node_entry, rnp->gp_tasks);
228 break;
229
230 default:
231
232 /* Yet another exercise in excessive paranoia. */
233 WARN_ON_ONCE(1);
234 break;
235 }
236
237 /*
238 * We have now queued the task. If it was the first one to
239 * block either grace period, update the ->gp_tasks and/or
240 * ->exp_tasks pointers, respectively, to reference the newly
241 * blocked tasks.
242 */
243 if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD)) {
244 WRITE_ONCE(rnp->gp_tasks, &t->rcu_node_entry);
245 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq);
246 }
247 if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
248 WRITE_ONCE(rnp->exp_tasks, &t->rcu_node_entry);
249 WARN_ON_ONCE(!(blkd_state & RCU_GP_BLKD) !=
250 !(rnp->qsmask & rdp->grpmask));
251 WARN_ON_ONCE(!(blkd_state & RCU_EXP_BLKD) !=
252 !(rnp->expmask & rdp->grpmask));
253 raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
254
255 /*
256 * Report the quiescent state for the expedited GP. This expedited
257 * GP should not be able to end until we report, so there should be
258 * no need to check for a subsequent expedited GP. (Though we are
259 * still in a quiescent state in any case.)
260 */
261 if (blkd_state & RCU_EXP_BLKD && rdp->cpu_no_qs.b.exp)
262 rcu_report_exp_rdp(rdp);
263 else
264 WARN_ON_ONCE(rdp->cpu_no_qs.b.exp);
265}
266
267/*
268 * Record a preemptible-RCU quiescent state for the specified CPU.
269 * Note that this does not necessarily mean that the task currently running
270 * on the CPU is in a quiescent state: Instead, it means that the current
271 * grace period need not wait on any RCU read-side critical section that
272 * starts later on this CPU. It also means that if the current task is
273 * in an RCU read-side critical section, it has already added itself to
274 * some leaf rcu_node structure's ->blkd_tasks list. In addition to the
275 * current task, there might be any number of other tasks blocked while
276 * in an RCU read-side critical section.
277 *
278 * Unlike non-preemptible-RCU, quiescent state reports for expedited
279 * grace periods are handled separately via deferred quiescent states
280 * and context switch events.
281 *
282 * Callers to this function must disable preemption.
283 */
284static void rcu_qs(void)
285{
286 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!\n");
287 if (__this_cpu_read(rcu_data.cpu_no_qs.b.norm)) {
288 trace_rcu_grace_period(TPS("rcu_preempt"),
289 __this_cpu_read(rcu_data.gp_seq),
290 TPS("cpuqs"));
291 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
292 barrier(); /* Coordinate with rcu_flavor_sched_clock_irq(). */
293 WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, false);
294 }
295}
296
297/*
298 * We have entered the scheduler, and the current task might soon be
299 * context-switched away from. If this task is in an RCU read-side
300 * critical section, we will no longer be able to rely on the CPU to
301 * record that fact, so we enqueue the task on the blkd_tasks list.
302 * The task will dequeue itself when it exits the outermost enclosing
303 * RCU read-side critical section. Therefore, the current grace period
304 * cannot be permitted to complete until the blkd_tasks list entries
305 * predating the current grace period drain, in other words, until
306 * rnp->gp_tasks becomes NULL.
307 *
308 * Caller must disable interrupts.
309 */
310void rcu_note_context_switch(bool preempt)
311{
312 struct task_struct *t = current;
313 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
314 struct rcu_node *rnp;
315
316 trace_rcu_utilization(TPS("Start context switch"));
317 lockdep_assert_irqs_disabled();
318 WARN_ONCE(!preempt && rcu_preempt_depth() > 0, "Voluntary context switch within RCU read-side critical section!");
319 if (rcu_preempt_depth() > 0 &&
320 !t->rcu_read_unlock_special.b.blocked) {
321
322 /* Possibly blocking in an RCU read-side critical section. */
323 rnp = rdp->mynode;
324 raw_spin_lock_rcu_node(rnp);
325 t->rcu_read_unlock_special.b.blocked = true;
326 t->rcu_blocked_node = rnp;
327
328 /*
329 * Verify the CPU's sanity, trace the preemption, and
330 * then queue the task as required based on the states
331 * of any ongoing and expedited grace periods.
332 */
333 WARN_ON_ONCE(!rcu_rdp_cpu_online(rdp));
334 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
335 trace_rcu_preempt_task(rcu_state.name,
336 t->pid,
337 (rnp->qsmask & rdp->grpmask)
338 ? rnp->gp_seq
339 : rcu_seq_snap(&rnp->gp_seq));
340 rcu_preempt_ctxt_queue(rnp, rdp);
341 } else {
342 rcu_preempt_deferred_qs(t);
343 }
344
345 /*
346 * Either we were not in an RCU read-side critical section to
347 * begin with, or we have now recorded that critical section
348 * globally. Either way, we can now note a quiescent state
349 * for this CPU. Again, if we were in an RCU read-side critical
350 * section, and if that critical section was blocking the current
351 * grace period, then the fact that the task has been enqueued
352 * means that we continue to block the current grace period.
353 */
354 rcu_qs();
355 if (rdp->cpu_no_qs.b.exp)
356 rcu_report_exp_rdp(rdp);
357 rcu_tasks_qs(current, preempt);
358 trace_rcu_utilization(TPS("End context switch"));
359}
360EXPORT_SYMBOL_GPL(rcu_note_context_switch);
361
362/*
363 * Check for preempted RCU readers blocking the current grace period
364 * for the specified rcu_node structure. If the caller needs a reliable
365 * answer, it must hold the rcu_node's ->lock.
366 */
367static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
368{
369 return READ_ONCE(rnp->gp_tasks) != NULL;
370}
371
372/* limit value for ->rcu_read_lock_nesting. */
373#define RCU_NEST_PMAX (INT_MAX / 2)
374
375static void rcu_preempt_read_enter(void)
376{
377 WRITE_ONCE(current->rcu_read_lock_nesting, READ_ONCE(current->rcu_read_lock_nesting) + 1);
378}
379
380static int rcu_preempt_read_exit(void)
381{
382 int ret = READ_ONCE(current->rcu_read_lock_nesting) - 1;
383
384 WRITE_ONCE(current->rcu_read_lock_nesting, ret);
385 return ret;
386}
387
388static void rcu_preempt_depth_set(int val)
389{
390 WRITE_ONCE(current->rcu_read_lock_nesting, val);
391}
392
393/*
394 * Preemptible RCU implementation for rcu_read_lock().
395 * Just increment ->rcu_read_lock_nesting, shared state will be updated
396 * if we block.
397 */
398void __rcu_read_lock(void)
399{
400 rcu_preempt_read_enter();
401 if (IS_ENABLED(CONFIG_PROVE_LOCKING))
402 WARN_ON_ONCE(rcu_preempt_depth() > RCU_NEST_PMAX);
403 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) && rcu_state.gp_kthread)
404 WRITE_ONCE(current->rcu_read_unlock_special.b.need_qs, true);
405 barrier(); /* critical section after entry code. */
406}
407EXPORT_SYMBOL_GPL(__rcu_read_lock);
408
409/*
410 * Preemptible RCU implementation for rcu_read_unlock().
411 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
412 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
413 * invoke rcu_read_unlock_special() to clean up after a context switch
414 * in an RCU read-side critical section and other special cases.
415 */
416void __rcu_read_unlock(void)
417{
418 struct task_struct *t = current;
419
420 barrier(); // critical section before exit code.
421 if (rcu_preempt_read_exit() == 0) {
422 barrier(); // critical-section exit before .s check.
423 if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
424 rcu_read_unlock_special(t);
425 }
426 if (IS_ENABLED(CONFIG_PROVE_LOCKING)) {
427 int rrln = rcu_preempt_depth();
428
429 WARN_ON_ONCE(rrln < 0 || rrln > RCU_NEST_PMAX);
430 }
431}
432EXPORT_SYMBOL_GPL(__rcu_read_unlock);
433
434/*
435 * Advance a ->blkd_tasks-list pointer to the next entry, instead
436 * returning NULL if at the end of the list.
437 */
438static struct list_head *rcu_next_node_entry(struct task_struct *t,
439 struct rcu_node *rnp)
440{
441 struct list_head *np;
442
443 np = t->rcu_node_entry.next;
444 if (np == &rnp->blkd_tasks)
445 np = NULL;
446 return np;
447}
448
449/*
450 * Return true if the specified rcu_node structure has tasks that were
451 * preempted within an RCU read-side critical section.
452 */
453static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
454{
455 return !list_empty(&rnp->blkd_tasks);
456}
457
458/*
459 * Report deferred quiescent states. The deferral time can
460 * be quite short, for example, in the case of the call from
461 * rcu_read_unlock_special().
462 */
463static notrace void
464rcu_preempt_deferred_qs_irqrestore(struct task_struct *t, unsigned long flags)
465{
466 bool empty_exp;
467 bool empty_norm;
468 bool empty_exp_now;
469 struct list_head *np;
470 bool drop_boost_mutex = false;
471 struct rcu_data *rdp;
472 struct rcu_node *rnp;
473 union rcu_special special;
474
475 /*
476 * If RCU core is waiting for this CPU to exit its critical section,
477 * report the fact that it has exited. Because irqs are disabled,
478 * t->rcu_read_unlock_special cannot change.
479 */
480 special = t->rcu_read_unlock_special;
481 rdp = this_cpu_ptr(&rcu_data);
482 if (!special.s && !rdp->cpu_no_qs.b.exp) {
483 local_irq_restore(flags);
484 return;
485 }
486 t->rcu_read_unlock_special.s = 0;
487 if (special.b.need_qs) {
488 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
489 rdp->cpu_no_qs.b.norm = false;
490 rcu_report_qs_rdp(rdp);
491 udelay(rcu_unlock_delay);
492 } else {
493 rcu_qs();
494 }
495 }
496
497 /*
498 * Respond to a request by an expedited grace period for a
499 * quiescent state from this CPU. Note that requests from
500 * tasks are handled when removing the task from the
501 * blocked-tasks list below.
502 */
503 if (rdp->cpu_no_qs.b.exp)
504 rcu_report_exp_rdp(rdp);
505
506 /* Clean up if blocked during RCU read-side critical section. */
507 if (special.b.blocked) {
508
509 /*
510 * Remove this task from the list it blocked on. The task
511 * now remains queued on the rcu_node corresponding to the
512 * CPU it first blocked on, so there is no longer any need
513 * to loop. Retain a WARN_ON_ONCE() out of sheer paranoia.
514 */
515 rnp = t->rcu_blocked_node;
516 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
517 WARN_ON_ONCE(rnp != t->rcu_blocked_node);
518 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
519 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
520 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq &&
521 (!empty_norm || rnp->qsmask));
522 empty_exp = sync_rcu_exp_done(rnp);
523 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
524 np = rcu_next_node_entry(t, rnp);
525 list_del_init(&t->rcu_node_entry);
526 t->rcu_blocked_node = NULL;
527 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
528 rnp->gp_seq, t->pid);
529 if (&t->rcu_node_entry == rnp->gp_tasks)
530 WRITE_ONCE(rnp->gp_tasks, np);
531 if (&t->rcu_node_entry == rnp->exp_tasks)
532 WRITE_ONCE(rnp->exp_tasks, np);
533 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
534 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
535 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx.rtmutex) == t;
536 if (&t->rcu_node_entry == rnp->boost_tasks)
537 WRITE_ONCE(rnp->boost_tasks, np);
538 }
539
540 /*
541 * If this was the last task on the current list, and if
542 * we aren't waiting on any CPUs, report the quiescent state.
543 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
544 * so we must take a snapshot of the expedited state.
545 */
546 empty_exp_now = sync_rcu_exp_done(rnp);
547 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
548 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
549 rnp->gp_seq,
550 0, rnp->qsmask,
551 rnp->level,
552 rnp->grplo,
553 rnp->grphi,
554 !!rnp->gp_tasks);
555 rcu_report_unblock_qs_rnp(rnp, flags);
556 } else {
557 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
558 }
559
560 /*
561 * If this was the last task on the expedited lists,
562 * then we need to report up the rcu_node hierarchy.
563 */
564 if (!empty_exp && empty_exp_now)
565 rcu_report_exp_rnp(rnp, true);
566
567 /* Unboost if we were boosted. */
568 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
569 rt_mutex_futex_unlock(&rnp->boost_mtx.rtmutex);
570 } else {
571 local_irq_restore(flags);
572 }
573}
574
575/*
576 * Is a deferred quiescent-state pending, and are we also not in
577 * an RCU read-side critical section? It is the caller's responsibility
578 * to ensure it is otherwise safe to report any deferred quiescent
579 * states. The reason for this is that it is safe to report a
580 * quiescent state during context switch even though preemption
581 * is disabled. This function cannot be expected to understand these
582 * nuances, so the caller must handle them.
583 */
584static notrace bool rcu_preempt_need_deferred_qs(struct task_struct *t)
585{
586 return (__this_cpu_read(rcu_data.cpu_no_qs.b.exp) ||
587 READ_ONCE(t->rcu_read_unlock_special.s)) &&
588 rcu_preempt_depth() == 0;
589}
590
591/*
592 * Report a deferred quiescent state if needed and safe to do so.
593 * As with rcu_preempt_need_deferred_qs(), "safe" involves only
594 * not being in an RCU read-side critical section. The caller must
595 * evaluate safety in terms of interrupt, softirq, and preemption
596 * disabling.
597 */
598notrace void rcu_preempt_deferred_qs(struct task_struct *t)
599{
600 unsigned long flags;
601
602 if (!rcu_preempt_need_deferred_qs(t))
603 return;
604 local_irq_save(flags);
605 rcu_preempt_deferred_qs_irqrestore(t, flags);
606}
607
608/*
609 * Minimal handler to give the scheduler a chance to re-evaluate.
610 */
611static void rcu_preempt_deferred_qs_handler(struct irq_work *iwp)
612{
613 struct rcu_data *rdp;
614
615 rdp = container_of(iwp, struct rcu_data, defer_qs_iw);
616 rdp->defer_qs_iw_pending = false;
617}
618
619/*
620 * Handle special cases during rcu_read_unlock(), such as needing to
621 * notify RCU core processing or task having blocked during the RCU
622 * read-side critical section.
623 */
624static void rcu_read_unlock_special(struct task_struct *t)
625{
626 unsigned long flags;
627 bool irqs_were_disabled;
628 bool preempt_bh_were_disabled =
629 !!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
630
631 /* NMI handlers cannot block and cannot safely manipulate state. */
632 if (in_nmi())
633 return;
634
635 local_irq_save(flags);
636 irqs_were_disabled = irqs_disabled_flags(flags);
637 if (preempt_bh_were_disabled || irqs_were_disabled) {
638 bool expboost; // Expedited GP in flight or possible boosting.
639 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
640 struct rcu_node *rnp = rdp->mynode;
641
642 expboost = (t->rcu_blocked_node && READ_ONCE(t->rcu_blocked_node->exp_tasks)) ||
643 (rdp->grpmask & READ_ONCE(rnp->expmask)) ||
644 (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) &&
645 ((rdp->grpmask & READ_ONCE(rnp->qsmask)) || t->rcu_blocked_node)) ||
646 (IS_ENABLED(CONFIG_RCU_BOOST) && irqs_were_disabled &&
647 t->rcu_blocked_node);
648 // Need to defer quiescent state until everything is enabled.
649 if (use_softirq && (in_hardirq() || (expboost && !irqs_were_disabled))) {
650 // Using softirq, safe to awaken, and either the
651 // wakeup is free or there is either an expedited
652 // GP in flight or a potential need to deboost.
653 raise_softirq_irqoff(RCU_SOFTIRQ);
654 } else {
655 // Enabling BH or preempt does reschedule, so...
656 // Also if no expediting and no possible deboosting,
657 // slow is OK. Plus nohz_full CPUs eventually get
658 // tick enabled.
659 set_tsk_need_resched(current);
660 set_preempt_need_resched();
661 if (IS_ENABLED(CONFIG_IRQ_WORK) && irqs_were_disabled &&
662 expboost && !rdp->defer_qs_iw_pending && cpu_online(rdp->cpu)) {
663 // Get scheduler to re-evaluate and call hooks.
664 // If !IRQ_WORK, FQS scan will eventually IPI.
665 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) &&
666 IS_ENABLED(CONFIG_PREEMPT_RT))
667 rdp->defer_qs_iw = IRQ_WORK_INIT_HARD(
668 rcu_preempt_deferred_qs_handler);
669 else
670 init_irq_work(&rdp->defer_qs_iw,
671 rcu_preempt_deferred_qs_handler);
672 rdp->defer_qs_iw_pending = true;
673 irq_work_queue_on(&rdp->defer_qs_iw, rdp->cpu);
674 }
675 }
676 local_irq_restore(flags);
677 return;
678 }
679 rcu_preempt_deferred_qs_irqrestore(t, flags);
680}
681
682/*
683 * Check that the list of blocked tasks for the newly completed grace
684 * period is in fact empty. It is a serious bug to complete a grace
685 * period that still has RCU readers blocked! This function must be
686 * invoked -before- updating this rnp's ->gp_seq.
687 *
688 * Also, if there are blocked tasks on the list, they automatically
689 * block the newly created grace period, so set up ->gp_tasks accordingly.
690 */
691static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
692{
693 struct task_struct *t;
694
695 RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
696 raw_lockdep_assert_held_rcu_node(rnp);
697 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
698 dump_blkd_tasks(rnp, 10);
699 if (rcu_preempt_has_tasks(rnp) &&
700 (rnp->qsmaskinit || rnp->wait_blkd_tasks)) {
701 WRITE_ONCE(rnp->gp_tasks, rnp->blkd_tasks.next);
702 t = container_of(rnp->gp_tasks, struct task_struct,
703 rcu_node_entry);
704 trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
705 rnp->gp_seq, t->pid);
706 }
707 WARN_ON_ONCE(rnp->qsmask);
708}
709
710/*
711 * Check for a quiescent state from the current CPU, including voluntary
712 * context switches for Tasks RCU. When a task blocks, the task is
713 * recorded in the corresponding CPU's rcu_node structure, which is checked
714 * elsewhere, hence this function need only check for quiescent states
715 * related to the current CPU, not to those related to tasks.
716 */
717static void rcu_flavor_sched_clock_irq(int user)
718{
719 struct task_struct *t = current;
720
721 lockdep_assert_irqs_disabled();
722 if (rcu_preempt_depth() > 0 ||
723 (preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK))) {
724 /* No QS, force context switch if deferred. */
725 if (rcu_preempt_need_deferred_qs(t)) {
726 set_tsk_need_resched(t);
727 set_preempt_need_resched();
728 }
729 } else if (rcu_preempt_need_deferred_qs(t)) {
730 rcu_preempt_deferred_qs(t); /* Report deferred QS. */
731 return;
732 } else if (!WARN_ON_ONCE(rcu_preempt_depth())) {
733 rcu_qs(); /* Report immediate QS. */
734 return;
735 }
736
737 /* If GP is oldish, ask for help from rcu_read_unlock_special(). */
738 if (rcu_preempt_depth() > 0 &&
739 __this_cpu_read(rcu_data.core_needs_qs) &&
740 __this_cpu_read(rcu_data.cpu_no_qs.b.norm) &&
741 !t->rcu_read_unlock_special.b.need_qs &&
742 time_after(jiffies, rcu_state.gp_start + HZ))
743 t->rcu_read_unlock_special.b.need_qs = true;
744}
745
746/*
747 * Check for a task exiting while in a preemptible-RCU read-side
748 * critical section, clean up if so. No need to issue warnings, as
749 * debug_check_no_locks_held() already does this if lockdep is enabled.
750 * Besides, if this function does anything other than just immediately
751 * return, there was a bug of some sort. Spewing warnings from this
752 * function is like as not to simply obscure important prior warnings.
753 */
754void exit_rcu(void)
755{
756 struct task_struct *t = current;
757
758 if (unlikely(!list_empty(¤t->rcu_node_entry))) {
759 rcu_preempt_depth_set(1);
760 barrier();
761 WRITE_ONCE(t->rcu_read_unlock_special.b.blocked, true);
762 } else if (unlikely(rcu_preempt_depth())) {
763 rcu_preempt_depth_set(1);
764 } else {
765 return;
766 }
767 __rcu_read_unlock();
768 rcu_preempt_deferred_qs(current);
769}
770
771/*
772 * Dump the blocked-tasks state, but limit the list dump to the
773 * specified number of elements.
774 */
775static void
776dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
777{
778 int cpu;
779 int i;
780 struct list_head *lhp;
781 struct rcu_data *rdp;
782 struct rcu_node *rnp1;
783
784 raw_lockdep_assert_held_rcu_node(rnp);
785 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
786 __func__, rnp->grplo, rnp->grphi, rnp->level,
787 (long)READ_ONCE(rnp->gp_seq), (long)rnp->completedqs);
788 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
789 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx\n",
790 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext);
791 pr_info("%s: ->gp_tasks %p ->boost_tasks %p ->exp_tasks %p\n",
792 __func__, READ_ONCE(rnp->gp_tasks), data_race(rnp->boost_tasks),
793 READ_ONCE(rnp->exp_tasks));
794 pr_info("%s: ->blkd_tasks", __func__);
795 i = 0;
796 list_for_each(lhp, &rnp->blkd_tasks) {
797 pr_cont(" %p", lhp);
798 if (++i >= ncheck)
799 break;
800 }
801 pr_cont("\n");
802 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
803 rdp = per_cpu_ptr(&rcu_data, cpu);
804 pr_info("\t%d: %c online: %ld(%d) offline: %ld(%d)\n",
805 cpu, ".o"[rcu_rdp_cpu_online(rdp)],
806 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
807 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
808 }
809}
810
811#else /* #ifdef CONFIG_PREEMPT_RCU */
812
813/*
814 * If strict grace periods are enabled, and if the calling
815 * __rcu_read_unlock() marks the beginning of a quiescent state, immediately
816 * report that quiescent state and, if requested, spin for a bit.
817 */
818void rcu_read_unlock_strict(void)
819{
820 struct rcu_data *rdp;
821
822 if (irqs_disabled() || preempt_count() || !rcu_state.gp_kthread)
823 return;
824 rdp = this_cpu_ptr(&rcu_data);
825 rdp->cpu_no_qs.b.norm = false;
826 rcu_report_qs_rdp(rdp);
827 udelay(rcu_unlock_delay);
828}
829EXPORT_SYMBOL_GPL(rcu_read_unlock_strict);
830
831/*
832 * Tell them what RCU they are running.
833 */
834static void __init rcu_bootup_announce(void)
835{
836 pr_info("Hierarchical RCU implementation.\n");
837 rcu_bootup_announce_oddness();
838}
839
840/*
841 * Note a quiescent state for PREEMPTION=n. Because we do not need to know
842 * how many quiescent states passed, just if there was at least one since
843 * the start of the grace period, this just sets a flag. The caller must
844 * have disabled preemption.
845 */
846static void rcu_qs(void)
847{
848 RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!");
849 if (!__this_cpu_read(rcu_data.cpu_no_qs.s))
850 return;
851 trace_rcu_grace_period(TPS("rcu_sched"),
852 __this_cpu_read(rcu_data.gp_seq), TPS("cpuqs"));
853 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
854 if (__this_cpu_read(rcu_data.cpu_no_qs.b.exp))
855 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
856}
857
858/*
859 * Register an urgently needed quiescent state. If there is an
860 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
861 * dyntick-idle quiescent state visible to other CPUs, which will in
862 * some cases serve for expedited as well as normal grace periods.
863 * Either way, register a lightweight quiescent state.
864 */
865void rcu_all_qs(void)
866{
867 unsigned long flags;
868
869 if (!raw_cpu_read(rcu_data.rcu_urgent_qs))
870 return;
871 preempt_disable(); // For CONFIG_PREEMPT_COUNT=y kernels
872 /* Load rcu_urgent_qs before other flags. */
873 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
874 preempt_enable();
875 return;
876 }
877 this_cpu_write(rcu_data.rcu_urgent_qs, false);
878 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs))) {
879 local_irq_save(flags);
880 rcu_momentary_dyntick_idle();
881 local_irq_restore(flags);
882 }
883 rcu_qs();
884 preempt_enable();
885}
886EXPORT_SYMBOL_GPL(rcu_all_qs);
887
888/*
889 * Note a PREEMPTION=n context switch. The caller must have disabled interrupts.
890 */
891void rcu_note_context_switch(bool preempt)
892{
893 trace_rcu_utilization(TPS("Start context switch"));
894 rcu_qs();
895 /* Load rcu_urgent_qs before other flags. */
896 if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs)))
897 goto out;
898 this_cpu_write(rcu_data.rcu_urgent_qs, false);
899 if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs)))
900 rcu_momentary_dyntick_idle();
901out:
902 rcu_tasks_qs(current, preempt);
903 trace_rcu_utilization(TPS("End context switch"));
904}
905EXPORT_SYMBOL_GPL(rcu_note_context_switch);
906
907/*
908 * Because preemptible RCU does not exist, there are never any preempted
909 * RCU readers.
910 */
911static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
912{
913 return 0;
914}
915
916/*
917 * Because there is no preemptible RCU, there can be no readers blocked.
918 */
919static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
920{
921 return false;
922}
923
924/*
925 * Because there is no preemptible RCU, there can be no deferred quiescent
926 * states.
927 */
928static notrace bool rcu_preempt_need_deferred_qs(struct task_struct *t)
929{
930 return false;
931}
932
933// Except that we do need to respond to a request by an expedited
934// grace period for a quiescent state from this CPU. Note that in
935// non-preemptible kernels, there can be no context switches within RCU
936// read-side critical sections, which in turn means that the leaf rcu_node
937// structure's blocked-tasks list is always empty. is therefore no need to
938// actually check it. Instead, a quiescent state from this CPU suffices,
939// and this function is only called from such a quiescent state.
940notrace void rcu_preempt_deferred_qs(struct task_struct *t)
941{
942 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
943
944 if (rdp->cpu_no_qs.b.exp)
945 rcu_report_exp_rdp(rdp);
946}
947
948/*
949 * Because there is no preemptible RCU, there can be no readers blocked,
950 * so there is no need to check for blocked tasks. So check only for
951 * bogus qsmask values.
952 */
953static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
954{
955 WARN_ON_ONCE(rnp->qsmask);
956}
957
958/*
959 * Check to see if this CPU is in a non-context-switch quiescent state,
960 * namely user mode and idle loop.
961 */
962static void rcu_flavor_sched_clock_irq(int user)
963{
964 if (user || rcu_is_cpu_rrupt_from_idle()) {
965
966 /*
967 * Get here if this CPU took its interrupt from user
968 * mode or from the idle loop, and if this is not a
969 * nested interrupt. In this case, the CPU is in
970 * a quiescent state, so note it.
971 *
972 * No memory barrier is required here because rcu_qs()
973 * references only CPU-local variables that other CPUs
974 * neither access nor modify, at least not while the
975 * corresponding CPU is online.
976 */
977 rcu_qs();
978 }
979}
980
981/*
982 * Because preemptible RCU does not exist, tasks cannot possibly exit
983 * while in preemptible RCU read-side critical sections.
984 */
985void exit_rcu(void)
986{
987}
988
989/*
990 * Dump the guaranteed-empty blocked-tasks state. Trust but verify.
991 */
992static void
993dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
994{
995 WARN_ON_ONCE(!list_empty(&rnp->blkd_tasks));
996}
997
998#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
999
1000/*
1001 * If boosting, set rcuc kthreads to realtime priority.
1002 */
1003static void rcu_cpu_kthread_setup(unsigned int cpu)
1004{
1005 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1006#ifdef CONFIG_RCU_BOOST
1007 struct sched_param sp;
1008
1009 sp.sched_priority = kthread_prio;
1010 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1011#endif /* #ifdef CONFIG_RCU_BOOST */
1012
1013 WRITE_ONCE(rdp->rcuc_activity, jiffies);
1014}
1015
1016static bool rcu_is_callbacks_nocb_kthread(struct rcu_data *rdp)
1017{
1018#ifdef CONFIG_RCU_NOCB_CPU
1019 return rdp->nocb_cb_kthread == current;
1020#else
1021 return false;
1022#endif
1023}
1024
1025/*
1026 * Is the current CPU running the RCU-callbacks kthread?
1027 * Caller must have preemption disabled.
1028 */
1029static bool rcu_is_callbacks_kthread(struct rcu_data *rdp)
1030{
1031 return rdp->rcu_cpu_kthread_task == current ||
1032 rcu_is_callbacks_nocb_kthread(rdp);
1033}
1034
1035#ifdef CONFIG_RCU_BOOST
1036
1037/*
1038 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1039 * or ->boost_tasks, advancing the pointer to the next task in the
1040 * ->blkd_tasks list.
1041 *
1042 * Note that irqs must be enabled: boosting the task can block.
1043 * Returns 1 if there are more tasks needing to be boosted.
1044 */
1045static int rcu_boost(struct rcu_node *rnp)
1046{
1047 unsigned long flags;
1048 struct task_struct *t;
1049 struct list_head *tb;
1050
1051 if (READ_ONCE(rnp->exp_tasks) == NULL &&
1052 READ_ONCE(rnp->boost_tasks) == NULL)
1053 return 0; /* Nothing left to boost. */
1054
1055 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1056
1057 /*
1058 * Recheck under the lock: all tasks in need of boosting
1059 * might exit their RCU read-side critical sections on their own.
1060 */
1061 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1062 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1063 return 0;
1064 }
1065
1066 /*
1067 * Preferentially boost tasks blocking expedited grace periods.
1068 * This cannot starve the normal grace periods because a second
1069 * expedited grace period must boost all blocked tasks, including
1070 * those blocking the pre-existing normal grace period.
1071 */
1072 if (rnp->exp_tasks != NULL)
1073 tb = rnp->exp_tasks;
1074 else
1075 tb = rnp->boost_tasks;
1076
1077 /*
1078 * We boost task t by manufacturing an rt_mutex that appears to
1079 * be held by task t. We leave a pointer to that rt_mutex where
1080 * task t can find it, and task t will release the mutex when it
1081 * exits its outermost RCU read-side critical section. Then
1082 * simply acquiring this artificial rt_mutex will boost task
1083 * t's priority. (Thanks to tglx for suggesting this approach!)
1084 *
1085 * Note that task t must acquire rnp->lock to remove itself from
1086 * the ->blkd_tasks list, which it will do from exit() if from
1087 * nowhere else. We therefore are guaranteed that task t will
1088 * stay around at least until we drop rnp->lock. Note that
1089 * rnp->lock also resolves races between our priority boosting
1090 * and task t's exiting its outermost RCU read-side critical
1091 * section.
1092 */
1093 t = container_of(tb, struct task_struct, rcu_node_entry);
1094 rt_mutex_init_proxy_locked(&rnp->boost_mtx.rtmutex, t);
1095 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1096 /* Lock only for side effect: boosts task t's priority. */
1097 rt_mutex_lock(&rnp->boost_mtx);
1098 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
1099 rnp->n_boosts++;
1100
1101 return READ_ONCE(rnp->exp_tasks) != NULL ||
1102 READ_ONCE(rnp->boost_tasks) != NULL;
1103}
1104
1105/*
1106 * Priority-boosting kthread, one per leaf rcu_node.
1107 */
1108static int rcu_boost_kthread(void *arg)
1109{
1110 struct rcu_node *rnp = (struct rcu_node *)arg;
1111 int spincnt = 0;
1112 int more2boost;
1113
1114 trace_rcu_utilization(TPS("Start boost kthread@init"));
1115 for (;;) {
1116 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_WAITING);
1117 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1118 rcu_wait(READ_ONCE(rnp->boost_tasks) ||
1119 READ_ONCE(rnp->exp_tasks));
1120 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1121 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_RUNNING);
1122 more2boost = rcu_boost(rnp);
1123 if (more2boost)
1124 spincnt++;
1125 else
1126 spincnt = 0;
1127 if (spincnt > 10) {
1128 WRITE_ONCE(rnp->boost_kthread_status, RCU_KTHREAD_YIELDING);
1129 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1130 schedule_timeout_idle(2);
1131 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1132 spincnt = 0;
1133 }
1134 }
1135 /* NOTREACHED */
1136 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1137 return 0;
1138}
1139
1140/*
1141 * Check to see if it is time to start boosting RCU readers that are
1142 * blocking the current grace period, and, if so, tell the per-rcu_node
1143 * kthread to start boosting them. If there is an expedited grace
1144 * period in progress, it is always time to boost.
1145 *
1146 * The caller must hold rnp->lock, which this function releases.
1147 * The ->boost_kthread_task is immortal, so we don't need to worry
1148 * about it going away.
1149 */
1150static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1151 __releases(rnp->lock)
1152{
1153 raw_lockdep_assert_held_rcu_node(rnp);
1154 if (!rnp->boost_kthread_task ||
1155 (!rcu_preempt_blocked_readers_cgp(rnp) && !rnp->exp_tasks)) {
1156 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1157 return;
1158 }
1159 if (rnp->exp_tasks != NULL ||
1160 (rnp->gp_tasks != NULL &&
1161 rnp->boost_tasks == NULL &&
1162 rnp->qsmask == 0 &&
1163 (!time_after(rnp->boost_time, jiffies) || rcu_state.cbovld ||
1164 IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)))) {
1165 if (rnp->exp_tasks == NULL)
1166 WRITE_ONCE(rnp->boost_tasks, rnp->gp_tasks);
1167 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1168 rcu_wake_cond(rnp->boost_kthread_task,
1169 READ_ONCE(rnp->boost_kthread_status));
1170 } else {
1171 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1172 }
1173}
1174
1175#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1176
1177/*
1178 * Do priority-boost accounting for the start of a new grace period.
1179 */
1180static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1181{
1182 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1183}
1184
1185/*
1186 * Create an RCU-boost kthread for the specified node if one does not
1187 * already exist. We only create this kthread for preemptible RCU.
1188 */
1189static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1190{
1191 unsigned long flags;
1192 int rnp_index = rnp - rcu_get_root();
1193 struct sched_param sp;
1194 struct task_struct *t;
1195
1196 mutex_lock(&rnp->boost_kthread_mutex);
1197 if (rnp->boost_kthread_task || !rcu_scheduler_fully_active)
1198 goto out;
1199
1200 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1201 "rcub/%d", rnp_index);
1202 if (WARN_ON_ONCE(IS_ERR(t)))
1203 goto out;
1204
1205 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1206 rnp->boost_kthread_task = t;
1207 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1208 sp.sched_priority = kthread_prio;
1209 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1210 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1211
1212 out:
1213 mutex_unlock(&rnp->boost_kthread_mutex);
1214}
1215
1216/*
1217 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1218 * served by the rcu_node in question. The CPU hotplug lock is still
1219 * held, so the value of rnp->qsmaskinit will be stable.
1220 *
1221 * We don't include outgoingcpu in the affinity set, use -1 if there is
1222 * no outgoing CPU. If there are no CPUs left in the affinity set,
1223 * this function allows the kthread to execute on any CPU.
1224 *
1225 * Any future concurrent calls are serialized via ->boost_kthread_mutex.
1226 */
1227static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1228{
1229 struct task_struct *t = rnp->boost_kthread_task;
1230 unsigned long mask;
1231 cpumask_var_t cm;
1232 int cpu;
1233
1234 if (!t)
1235 return;
1236 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1237 return;
1238 mutex_lock(&rnp->boost_kthread_mutex);
1239 mask = rcu_rnp_online_cpus(rnp);
1240 for_each_leaf_node_possible_cpu(rnp, cpu)
1241 if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
1242 cpu != outgoingcpu)
1243 cpumask_set_cpu(cpu, cm);
1244 cpumask_and(cm, cm, housekeeping_cpumask(HK_TYPE_RCU));
1245 if (cpumask_empty(cm)) {
1246 cpumask_copy(cm, housekeeping_cpumask(HK_TYPE_RCU));
1247 if (outgoingcpu >= 0)
1248 cpumask_clear_cpu(outgoingcpu, cm);
1249 }
1250 set_cpus_allowed_ptr(t, cm);
1251 mutex_unlock(&rnp->boost_kthread_mutex);
1252 free_cpumask_var(cm);
1253}
1254
1255#else /* #ifdef CONFIG_RCU_BOOST */
1256
1257static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1258 __releases(rnp->lock)
1259{
1260 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1261}
1262
1263static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1264{
1265}
1266
1267static void rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1268{
1269}
1270
1271static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1272{
1273}
1274
1275#endif /* #else #ifdef CONFIG_RCU_BOOST */
1276
1277/*
1278 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
1279 * grace-period kthread will do force_quiescent_state() processing?
1280 * The idea is to avoid waking up RCU core processing on such a
1281 * CPU unless the grace period has extended for too long.
1282 *
1283 * This code relies on the fact that all NO_HZ_FULL CPUs are also
1284 * RCU_NOCB_CPU CPUs.
1285 */
1286static bool rcu_nohz_full_cpu(void)
1287{
1288#ifdef CONFIG_NO_HZ_FULL
1289 if (tick_nohz_full_cpu(smp_processor_id()) &&
1290 (!rcu_gp_in_progress() ||
1291 time_before(jiffies, READ_ONCE(rcu_state.gp_start) + HZ)))
1292 return true;
1293#endif /* #ifdef CONFIG_NO_HZ_FULL */
1294 return false;
1295}
1296
1297/*
1298 * Bind the RCU grace-period kthreads to the housekeeping CPU.
1299 */
1300static void rcu_bind_gp_kthread(void)
1301{
1302 if (!tick_nohz_full_enabled())
1303 return;
1304 housekeeping_affine(current, HK_TYPE_RCU);
1305}