2 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
3 * Internal non-public definitions that provide either classic
4 * or preemptible semantics.
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.
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.
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.
20 * Copyright Red Hat, 2009
21 * Copyright IBM Corporation, 2009
23 * Author: Ingo Molnar <mingo@elte.hu>
24 * Paul E. McKenney <paulmck@linux.vnet.ibm.com>
27 #include <linux/delay.h>
28 #include <linux/gfp.h>
29 #include <linux/oom.h>
30 #include <linux/sched/debug.h>
31 #include <linux/smpboot.h>
32 #include <uapi/linux/sched/types.h>
33 #include "../time/tick-internal.h"
35 #ifdef CONFIG_RCU_BOOST
37 #include "../locking/rtmutex_common.h"
40 * Control variables for per-CPU and per-rcu_node kthreads. These
41 * handle all flavors of RCU.
43 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
44 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
45 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
46 DEFINE_PER_CPU(char, rcu_cpu_has_work);
48 #else /* #ifdef CONFIG_RCU_BOOST */
51 * Some architectures do not define rt_mutexes, but if !CONFIG_RCU_BOOST,
52 * all uses are in dead code. Provide a definition to keep the compiler
53 * happy, but add WARN_ON_ONCE() to complain if used in the wrong place.
54 * This probably needs to be excluded from -rt builds.
56 #define rt_mutex_owner(a) ({ WARN_ON_ONCE(1); NULL; })
58 #endif /* #else #ifdef CONFIG_RCU_BOOST */
60 #ifdef CONFIG_RCU_NOCB_CPU
61 static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
62 static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */
63 static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
64 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
67 * Check the RCU kernel configuration parameters and print informative
68 * messages about anything out of the ordinary.
70 static void __init rcu_bootup_announce_oddness(void)
72 if (IS_ENABLED(CONFIG_RCU_TRACE))
73 pr_info("\tRCU event tracing is enabled.\n");
74 if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
75 (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
76 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
79 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
80 if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ))
81 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
82 if (IS_ENABLED(CONFIG_PROVE_RCU))
83 pr_info("\tRCU lockdep checking is enabled.\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",
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=%u.\n", NR_CPUS, nr_cpu_ids);
93 #ifdef CONFIG_RCU_BOOST
94 pr_info("\tRCU priority boosting: priority %d delay %d ms.\n", kthread_prio, CONFIG_RCU_BOOST_DELAY);
96 if (blimit != DEFAULT_RCU_BLIMIT)
97 pr_info("\tBoot-time adjustment of callback invocation limit to %ld.\n", blimit);
98 if (qhimark != DEFAULT_RCU_QHIMARK)
99 pr_info("\tBoot-time adjustment of callback high-water mark to %ld.\n", qhimark);
100 if (qlowmark != DEFAULT_RCU_QLOMARK)
101 pr_info("\tBoot-time adjustment of callback low-water mark to %ld.\n", qlowmark);
102 if (jiffies_till_first_fqs != ULONG_MAX)
103 pr_info("\tBoot-time adjustment of first FQS scan delay to %ld jiffies.\n", jiffies_till_first_fqs);
104 if (jiffies_till_next_fqs != ULONG_MAX)
105 pr_info("\tBoot-time adjustment of subsequent FQS scan delay to %ld jiffies.\n", jiffies_till_next_fqs);
106 if (rcu_kick_kthreads)
107 pr_info("\tKick kthreads if too-long grace period.\n");
108 if (IS_ENABLED(CONFIG_DEBUG_OBJECTS_RCU_HEAD))
109 pr_info("\tRCU callback double-/use-after-free debug enabled.\n");
110 if (gp_preinit_delay)
111 pr_info("\tRCU debug GP pre-init slowdown %d jiffies.\n", gp_preinit_delay);
113 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_init_delay);
114 if (gp_cleanup_delay)
115 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_cleanup_delay);
116 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG))
117 pr_info("\tRCU debug extended QS entry/exit.\n");
118 rcupdate_announce_bootup_oddness();
121 #ifdef CONFIG_PREEMPT_RCU
123 RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
124 static struct rcu_state *const rcu_state_p = &rcu_preempt_state;
125 static struct rcu_data __percpu *const rcu_data_p = &rcu_preempt_data;
127 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
131 * Tell them what RCU they are running.
133 static void __init rcu_bootup_announce(void)
135 pr_info("Preemptible hierarchical RCU implementation.\n");
136 rcu_bootup_announce_oddness();
139 /* Flags for rcu_preempt_ctxt_queue() decision table. */
140 #define RCU_GP_TASKS 0x8
141 #define RCU_EXP_TASKS 0x4
142 #define RCU_GP_BLKD 0x2
143 #define RCU_EXP_BLKD 0x1
146 * Queues a task preempted within an RCU-preempt read-side critical
147 * section into the appropriate location within the ->blkd_tasks list,
148 * depending on the states of any ongoing normal and expedited grace
149 * periods. The ->gp_tasks pointer indicates which element the normal
150 * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
151 * indicates which element the expedited grace period is waiting on (again,
152 * NULL if none). If a grace period is waiting on a given element in the
153 * ->blkd_tasks list, it also waits on all subsequent elements. Thus,
154 * adding a task to the tail of the list blocks any grace period that is
155 * already waiting on one of the elements. In contrast, adding a task
156 * to the head of the list won't block any grace period that is already
157 * waiting on one of the elements.
159 * This queuing is imprecise, and can sometimes make an ongoing grace
160 * period wait for a task that is not strictly speaking blocking it.
161 * Given the choice, we needlessly block a normal grace period rather than
162 * blocking an expedited grace period.
164 * Note that an endless sequence of expedited grace periods still cannot
165 * indefinitely postpone a normal grace period. Eventually, all of the
166 * fixed number of preempted tasks blocking the normal grace period that are
167 * not also blocking the expedited grace period will resume and complete
168 * their RCU read-side critical sections. At that point, the ->gp_tasks
169 * pointer will equal the ->exp_tasks pointer, at which point the end of
170 * the corresponding expedited grace period will also be the end of the
171 * normal grace period.
173 static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
174 __releases(rnp->lock) /* But leaves rrupts disabled. */
176 int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
177 (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
178 (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
179 (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
180 struct task_struct *t = current;
182 lockdep_assert_held(&rnp->lock);
183 WARN_ON_ONCE(rdp->mynode != rnp);
184 WARN_ON_ONCE(rnp->level != rcu_num_lvls - 1);
187 * Decide where to queue the newly blocked task. In theory,
188 * this could be an if-statement. In practice, when I tried
189 * that, it was quite messy.
191 switch (blkd_state) {
194 case RCU_EXP_TASKS + RCU_GP_BLKD:
196 case RCU_GP_TASKS + RCU_EXP_TASKS:
199 * Blocking neither GP, or first task blocking the normal
200 * GP but not blocking the already-waiting expedited GP.
201 * Queue at the head of the list to avoid unnecessarily
202 * blocking the already-waiting GPs.
204 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
209 case RCU_GP_BLKD + RCU_EXP_BLKD:
210 case RCU_GP_TASKS + RCU_EXP_BLKD:
211 case RCU_GP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
212 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
215 * First task arriving that blocks either GP, or first task
216 * arriving that blocks the expedited GP (with the normal
217 * GP already waiting), or a task arriving that blocks
218 * both GPs with both GPs already waiting. Queue at the
219 * tail of the list to avoid any GP waiting on any of the
220 * already queued tasks that are not blocking it.
222 list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
225 case RCU_EXP_TASKS + RCU_EXP_BLKD:
226 case RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
227 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_EXP_BLKD:
230 * Second or subsequent task blocking the expedited GP.
231 * The task either does not block the normal GP, or is the
232 * first task blocking the normal GP. Queue just after
233 * the first task blocking the expedited GP.
235 list_add(&t->rcu_node_entry, rnp->exp_tasks);
238 case RCU_GP_TASKS + RCU_GP_BLKD:
239 case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
242 * Second or subsequent task blocking the normal GP.
243 * The task does not block the expedited GP. Queue just
244 * after the first task blocking the normal GP.
246 list_add(&t->rcu_node_entry, rnp->gp_tasks);
251 /* Yet another exercise in excessive paranoia. */
257 * We have now queued the task. If it was the first one to
258 * block either grace period, update the ->gp_tasks and/or
259 * ->exp_tasks pointers, respectively, to reference the newly
262 if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD))
263 rnp->gp_tasks = &t->rcu_node_entry;
264 if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
265 rnp->exp_tasks = &t->rcu_node_entry;
266 WARN_ON_ONCE(!(blkd_state & RCU_GP_BLKD) !=
267 !(rnp->qsmask & rdp->grpmask));
268 WARN_ON_ONCE(!(blkd_state & RCU_EXP_BLKD) !=
269 !(rnp->expmask & rdp->grpmask));
270 raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
273 * Report the quiescent state for the expedited GP. This expedited
274 * GP should not be able to end until we report, so there should be
275 * no need to check for a subsequent expedited GP. (Though we are
276 * still in a quiescent state in any case.)
278 if (blkd_state & RCU_EXP_BLKD &&
279 t->rcu_read_unlock_special.b.exp_need_qs) {
280 t->rcu_read_unlock_special.b.exp_need_qs = false;
281 rcu_report_exp_rdp(rdp->rsp, rdp, true);
283 WARN_ON_ONCE(t->rcu_read_unlock_special.b.exp_need_qs);
288 * Record a preemptible-RCU quiescent state for the specified CPU. Note
289 * that this just means that the task currently running on the CPU is
290 * not in a quiescent state. There might be any number of tasks blocked
291 * while in an RCU read-side critical section.
293 * As with the other rcu_*_qs() functions, callers to this function
294 * must disable preemption.
296 static void rcu_preempt_qs(void)
298 RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_qs() invoked with preemption enabled!!!\n");
299 if (__this_cpu_read(rcu_data_p->cpu_no_qs.s)) {
300 trace_rcu_grace_period(TPS("rcu_preempt"),
301 __this_cpu_read(rcu_data_p->gpnum),
303 __this_cpu_write(rcu_data_p->cpu_no_qs.b.norm, false);
304 barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */
305 current->rcu_read_unlock_special.b.need_qs = false;
310 * We have entered the scheduler, and the current task might soon be
311 * context-switched away from. If this task is in an RCU read-side
312 * critical section, we will no longer be able to rely on the CPU to
313 * record that fact, so we enqueue the task on the blkd_tasks list.
314 * The task will dequeue itself when it exits the outermost enclosing
315 * RCU read-side critical section. Therefore, the current grace period
316 * cannot be permitted to complete until the blkd_tasks list entries
317 * predating the current grace period drain, in other words, until
318 * rnp->gp_tasks becomes NULL.
320 * Caller must disable interrupts.
322 static void rcu_preempt_note_context_switch(bool preempt)
324 struct task_struct *t = current;
325 struct rcu_data *rdp;
326 struct rcu_node *rnp;
328 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_preempt_note_context_switch() invoked with interrupts enabled!!!\n");
329 WARN_ON_ONCE(!preempt && t->rcu_read_lock_nesting > 0);
330 if (t->rcu_read_lock_nesting > 0 &&
331 !t->rcu_read_unlock_special.b.blocked) {
333 /* Possibly blocking in an RCU read-side critical section. */
334 rdp = this_cpu_ptr(rcu_state_p->rda);
336 raw_spin_lock_rcu_node(rnp);
337 t->rcu_read_unlock_special.b.blocked = true;
338 t->rcu_blocked_node = rnp;
341 * Verify the CPU's sanity, trace the preemption, and
342 * then queue the task as required based on the states
343 * of any ongoing and expedited grace periods.
345 WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
346 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
347 trace_rcu_preempt_task(rdp->rsp->name,
349 (rnp->qsmask & rdp->grpmask)
352 rcu_preempt_ctxt_queue(rnp, rdp);
353 } else if (t->rcu_read_lock_nesting < 0 &&
354 t->rcu_read_unlock_special.s) {
357 * Complete exit from RCU read-side critical section on
358 * behalf of preempted instance of __rcu_read_unlock().
360 rcu_read_unlock_special(t);
364 * Either we were not in an RCU read-side critical section to
365 * begin with, or we have now recorded that critical section
366 * globally. Either way, we can now note a quiescent state
367 * for this CPU. Again, if we were in an RCU read-side critical
368 * section, and if that critical section was blocking the current
369 * grace period, then the fact that the task has been enqueued
370 * means that we continue to block the current grace period.
376 * Check for preempted RCU readers blocking the current grace period
377 * for the specified rcu_node structure. If the caller needs a reliable
378 * answer, it must hold the rcu_node's ->lock.
380 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
382 return rnp->gp_tasks != NULL;
386 * Advance a ->blkd_tasks-list pointer to the next entry, instead
387 * returning NULL if at the end of the list.
389 static struct list_head *rcu_next_node_entry(struct task_struct *t,
390 struct rcu_node *rnp)
392 struct list_head *np;
394 np = t->rcu_node_entry.next;
395 if (np == &rnp->blkd_tasks)
401 * Return true if the specified rcu_node structure has tasks that were
402 * preempted within an RCU read-side critical section.
404 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
406 return !list_empty(&rnp->blkd_tasks);
410 * Handle special cases during rcu_read_unlock(), such as needing to
411 * notify RCU core processing or task having blocked during the RCU
412 * read-side critical section.
414 void rcu_read_unlock_special(struct task_struct *t)
420 struct list_head *np;
421 bool drop_boost_mutex = false;
422 struct rcu_data *rdp;
423 struct rcu_node *rnp;
424 union rcu_special special;
426 /* NMI handlers cannot block and cannot safely manipulate state. */
430 local_irq_save(flags);
433 * If RCU core is waiting for this CPU to exit its critical section,
434 * report the fact that it has exited. Because irqs are disabled,
435 * t->rcu_read_unlock_special cannot change.
437 special = t->rcu_read_unlock_special;
438 if (special.b.need_qs) {
440 t->rcu_read_unlock_special.b.need_qs = false;
441 if (!t->rcu_read_unlock_special.s) {
442 local_irq_restore(flags);
448 * Respond to a request for an expedited grace period, but only if
449 * we were not preempted, meaning that we were running on the same
450 * CPU throughout. If we were preempted, the exp_need_qs flag
451 * would have been cleared at the time of the first preemption,
452 * and the quiescent state would be reported when we were dequeued.
454 if (special.b.exp_need_qs) {
455 WARN_ON_ONCE(special.b.blocked);
456 t->rcu_read_unlock_special.b.exp_need_qs = false;
457 rdp = this_cpu_ptr(rcu_state_p->rda);
458 rcu_report_exp_rdp(rcu_state_p, rdp, true);
459 if (!t->rcu_read_unlock_special.s) {
460 local_irq_restore(flags);
465 /* Hardware IRQ handlers cannot block, complain if they get here. */
466 if (in_irq() || in_serving_softirq()) {
467 lockdep_rcu_suspicious(__FILE__, __LINE__,
468 "rcu_read_unlock() from irq or softirq with blocking in critical section!!!\n");
469 pr_alert("->rcu_read_unlock_special: %#x (b: %d, enq: %d nq: %d)\n",
470 t->rcu_read_unlock_special.s,
471 t->rcu_read_unlock_special.b.blocked,
472 t->rcu_read_unlock_special.b.exp_need_qs,
473 t->rcu_read_unlock_special.b.need_qs);
474 local_irq_restore(flags);
478 /* Clean up if blocked during RCU read-side critical section. */
479 if (special.b.blocked) {
480 t->rcu_read_unlock_special.b.blocked = false;
483 * Remove this task from the list it blocked on. The task
484 * now remains queued on the rcu_node corresponding to the
485 * CPU it first blocked on, so there is no longer any need
486 * to loop. Retain a WARN_ON_ONCE() out of sheer paranoia.
488 rnp = t->rcu_blocked_node;
489 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
490 WARN_ON_ONCE(rnp != t->rcu_blocked_node);
491 WARN_ON_ONCE(rnp->level != rcu_num_lvls - 1);
492 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
493 empty_exp = sync_rcu_preempt_exp_done(rnp);
494 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
495 np = rcu_next_node_entry(t, rnp);
496 list_del_init(&t->rcu_node_entry);
497 t->rcu_blocked_node = NULL;
498 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
500 if (&t->rcu_node_entry == rnp->gp_tasks)
502 if (&t->rcu_node_entry == rnp->exp_tasks)
504 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
505 /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
506 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
507 if (&t->rcu_node_entry == rnp->boost_tasks)
508 rnp->boost_tasks = np;
512 * If this was the last task on the current list, and if
513 * we aren't waiting on any CPUs, report the quiescent state.
514 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
515 * so we must take a snapshot of the expedited state.
517 empty_exp_now = sync_rcu_preempt_exp_done(rnp);
518 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
519 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
526 rcu_report_unblock_qs_rnp(rcu_state_p, rnp, flags);
528 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
531 /* Unboost if we were boosted. */
532 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
533 rt_mutex_unlock(&rnp->boost_mtx);
536 * If this was the last task on the expedited lists,
537 * then we need to report up the rcu_node hierarchy.
539 if (!empty_exp && empty_exp_now)
540 rcu_report_exp_rnp(rcu_state_p, rnp, true);
542 local_irq_restore(flags);
547 * Dump detailed information for all tasks blocking the current RCU
548 * grace period on the specified rcu_node structure.
550 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
553 struct task_struct *t;
555 raw_spin_lock_irqsave_rcu_node(rnp, flags);
556 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
557 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
560 t = list_entry(rnp->gp_tasks->prev,
561 struct task_struct, rcu_node_entry);
562 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
564 * We could be printing a lot while holding a spinlock.
565 * Avoid triggering hard lockup.
567 touch_nmi_watchdog();
570 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
574 * Dump detailed information for all tasks blocking the current RCU
577 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
579 struct rcu_node *rnp = rcu_get_root(rsp);
581 rcu_print_detail_task_stall_rnp(rnp);
582 rcu_for_each_leaf_node(rsp, rnp)
583 rcu_print_detail_task_stall_rnp(rnp);
586 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
588 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
589 rnp->level, rnp->grplo, rnp->grphi);
592 static void rcu_print_task_stall_end(void)
598 * Scan the current list of tasks blocked within RCU read-side critical
599 * sections, printing out the tid of each.
601 static int rcu_print_task_stall(struct rcu_node *rnp)
603 struct task_struct *t;
606 if (!rcu_preempt_blocked_readers_cgp(rnp))
608 rcu_print_task_stall_begin(rnp);
609 t = list_entry(rnp->gp_tasks->prev,
610 struct task_struct, rcu_node_entry);
611 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
612 pr_cont(" P%d", t->pid);
615 rcu_print_task_stall_end();
620 * Scan the current list of tasks blocked within RCU read-side critical
621 * sections, printing out the tid of each that is blocking the current
622 * expedited grace period.
624 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
626 struct task_struct *t;
631 t = list_entry(rnp->exp_tasks->prev,
632 struct task_struct, rcu_node_entry);
633 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
634 pr_cont(" P%d", t->pid);
641 * Check that the list of blocked tasks for the newly completed grace
642 * period is in fact empty. It is a serious bug to complete a grace
643 * period that still has RCU readers blocked! This function must be
644 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
645 * must be held by the caller.
647 * Also, if there are blocked tasks on the list, they automatically
648 * block the newly created grace period, so set up ->gp_tasks accordingly.
650 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
652 struct task_struct *t;
654 RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
655 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
656 if (rcu_preempt_has_tasks(rnp)) {
657 rnp->gp_tasks = rnp->blkd_tasks.next;
658 t = container_of(rnp->gp_tasks, struct task_struct,
660 trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
663 WARN_ON_ONCE(rnp->qsmask);
667 * Check for a quiescent state from the current CPU. When a task blocks,
668 * the task is recorded in the corresponding CPU's rcu_node structure,
669 * which is checked elsewhere.
671 * Caller must disable hard irqs.
673 static void rcu_preempt_check_callbacks(void)
675 struct task_struct *t = current;
677 if (t->rcu_read_lock_nesting == 0) {
681 if (t->rcu_read_lock_nesting > 0 &&
682 __this_cpu_read(rcu_data_p->core_needs_qs) &&
683 __this_cpu_read(rcu_data_p->cpu_no_qs.b.norm))
684 t->rcu_read_unlock_special.b.need_qs = true;
687 #ifdef CONFIG_RCU_BOOST
689 static void rcu_preempt_do_callbacks(void)
691 rcu_do_batch(rcu_state_p, this_cpu_ptr(rcu_data_p));
694 #endif /* #ifdef CONFIG_RCU_BOOST */
697 * call_rcu() - Queue an RCU callback for invocation after a grace period.
698 * @head: structure to be used for queueing the RCU updates.
699 * @func: actual callback function to be invoked after the grace period
701 * The callback function will be invoked some time after a full grace
702 * period elapses, in other words after all pre-existing RCU read-side
703 * critical sections have completed. However, the callback function
704 * might well execute concurrently with RCU read-side critical sections
705 * that started after call_rcu() was invoked. RCU read-side critical
706 * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
709 * Note that all CPUs must agree that the grace period extended beyond
710 * all pre-existing RCU read-side critical section. On systems with more
711 * than one CPU, this means that when "func()" is invoked, each CPU is
712 * guaranteed to have executed a full memory barrier since the end of its
713 * last RCU read-side critical section whose beginning preceded the call
714 * to call_rcu(). It also means that each CPU executing an RCU read-side
715 * critical section that continues beyond the start of "func()" must have
716 * executed a memory barrier after the call_rcu() but before the beginning
717 * of that RCU read-side critical section. Note that these guarantees
718 * include CPUs that are offline, idle, or executing in user mode, as
719 * well as CPUs that are executing in the kernel.
721 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
722 * resulting RCU callback function "func()", then both CPU A and CPU B are
723 * guaranteed to execute a full memory barrier during the time interval
724 * between the call to call_rcu() and the invocation of "func()" -- even
725 * if CPU A and CPU B are the same CPU (but again only if the system has
726 * more than one CPU).
728 void call_rcu(struct rcu_head *head, rcu_callback_t func)
730 __call_rcu(head, func, rcu_state_p, -1, 0);
732 EXPORT_SYMBOL_GPL(call_rcu);
735 * synchronize_rcu - wait until a grace period has elapsed.
737 * Control will return to the caller some time after a full grace
738 * period has elapsed, in other words after all currently executing RCU
739 * read-side critical sections have completed. Note, however, that
740 * upon return from synchronize_rcu(), the caller might well be executing
741 * concurrently with new RCU read-side critical sections that began while
742 * synchronize_rcu() was waiting. RCU read-side critical sections are
743 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
745 * See the description of synchronize_sched() for more detailed
746 * information on memory-ordering guarantees. However, please note
747 * that -only- the memory-ordering guarantees apply. For example,
748 * synchronize_rcu() is -not- guaranteed to wait on things like code
749 * protected by preempt_disable(), instead, synchronize_rcu() is -only-
750 * guaranteed to wait on RCU read-side critical sections, that is, sections
751 * of code protected by rcu_read_lock().
753 void synchronize_rcu(void)
755 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
756 lock_is_held(&rcu_lock_map) ||
757 lock_is_held(&rcu_sched_lock_map),
758 "Illegal synchronize_rcu() in RCU read-side critical section");
759 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
761 if (rcu_gp_is_expedited())
762 synchronize_rcu_expedited();
764 wait_rcu_gp(call_rcu);
766 EXPORT_SYMBOL_GPL(synchronize_rcu);
769 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
771 * Note that this primitive does not necessarily wait for an RCU grace period
772 * to complete. For example, if there are no RCU callbacks queued anywhere
773 * in the system, then rcu_barrier() is within its rights to return
774 * immediately, without waiting for anything, much less an RCU grace period.
776 void rcu_barrier(void)
778 _rcu_barrier(rcu_state_p);
780 EXPORT_SYMBOL_GPL(rcu_barrier);
783 * Initialize preemptible RCU's state structures.
785 static void __init __rcu_init_preempt(void)
787 rcu_init_one(rcu_state_p);
791 * Check for a task exiting while in a preemptible-RCU read-side
792 * critical section, clean up if so. No need to issue warnings,
793 * as debug_check_no_locks_held() already does this if lockdep
798 struct task_struct *t = current;
800 if (likely(list_empty(¤t->rcu_node_entry)))
802 t->rcu_read_lock_nesting = 1;
804 t->rcu_read_unlock_special.b.blocked = true;
808 #else /* #ifdef CONFIG_PREEMPT_RCU */
810 static struct rcu_state *const rcu_state_p = &rcu_sched_state;
813 * Tell them what RCU they are running.
815 static void __init rcu_bootup_announce(void)
817 pr_info("Hierarchical RCU implementation.\n");
818 rcu_bootup_announce_oddness();
822 * Because preemptible RCU does not exist, we never have to check for
823 * CPUs being in quiescent states.
825 static void rcu_preempt_note_context_switch(bool preempt)
830 * Because preemptible RCU does not exist, there are never any preempted
833 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
839 * Because there is no preemptible RCU, there can be no readers blocked.
841 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
847 * Because preemptible RCU does not exist, we never have to check for
848 * tasks blocked within RCU read-side critical sections.
850 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
855 * Because preemptible RCU does not exist, we never have to check for
856 * tasks blocked within RCU read-side critical sections.
858 static int rcu_print_task_stall(struct rcu_node *rnp)
864 * Because preemptible RCU does not exist, we never have to check for
865 * tasks blocked within RCU read-side critical sections that are
866 * blocking the current expedited grace period.
868 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
874 * Because there is no preemptible RCU, there can be no readers blocked,
875 * so there is no need to check for blocked tasks. So check only for
876 * bogus qsmask values.
878 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
880 WARN_ON_ONCE(rnp->qsmask);
884 * Because preemptible RCU does not exist, it never has any callbacks
887 static void rcu_preempt_check_callbacks(void)
892 * Because preemptible RCU does not exist, rcu_barrier() is just
893 * another name for rcu_barrier_sched().
895 void rcu_barrier(void)
899 EXPORT_SYMBOL_GPL(rcu_barrier);
902 * Because preemptible RCU does not exist, it need not be initialized.
904 static void __init __rcu_init_preempt(void)
909 * Because preemptible RCU does not exist, tasks cannot possibly exit
910 * while in preemptible RCU read-side critical sections.
916 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
918 #ifdef CONFIG_RCU_BOOST
920 #include "../locking/rtmutex_common.h"
922 static void rcu_wake_cond(struct task_struct *t, int status)
925 * If the thread is yielding, only wake it when this
926 * is invoked from idle
928 if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
933 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
934 * or ->boost_tasks, advancing the pointer to the next task in the
937 * Note that irqs must be enabled: boosting the task can block.
938 * Returns 1 if there are more tasks needing to be boosted.
940 static int rcu_boost(struct rcu_node *rnp)
943 struct task_struct *t;
944 struct list_head *tb;
946 if (READ_ONCE(rnp->exp_tasks) == NULL &&
947 READ_ONCE(rnp->boost_tasks) == NULL)
948 return 0; /* Nothing left to boost. */
950 raw_spin_lock_irqsave_rcu_node(rnp, flags);
953 * Recheck under the lock: all tasks in need of boosting
954 * might exit their RCU read-side critical sections on their own.
956 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
957 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
962 * Preferentially boost tasks blocking expedited grace periods.
963 * This cannot starve the normal grace periods because a second
964 * expedited grace period must boost all blocked tasks, including
965 * those blocking the pre-existing normal grace period.
967 if (rnp->exp_tasks != NULL) {
971 tb = rnp->boost_tasks;
972 rnp->n_normal_boosts++;
974 rnp->n_tasks_boosted++;
977 * We boost task t by manufacturing an rt_mutex that appears to
978 * be held by task t. We leave a pointer to that rt_mutex where
979 * task t can find it, and task t will release the mutex when it
980 * exits its outermost RCU read-side critical section. Then
981 * simply acquiring this artificial rt_mutex will boost task
982 * t's priority. (Thanks to tglx for suggesting this approach!)
984 * Note that task t must acquire rnp->lock to remove itself from
985 * the ->blkd_tasks list, which it will do from exit() if from
986 * nowhere else. We therefore are guaranteed that task t will
987 * stay around at least until we drop rnp->lock. Note that
988 * rnp->lock also resolves races between our priority boosting
989 * and task t's exiting its outermost RCU read-side critical
992 t = container_of(tb, struct task_struct, rcu_node_entry);
993 rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
994 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
995 /* Lock only for side effect: boosts task t's priority. */
996 rt_mutex_lock(&rnp->boost_mtx);
997 rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */
999 return READ_ONCE(rnp->exp_tasks) != NULL ||
1000 READ_ONCE(rnp->boost_tasks) != NULL;
1004 * Priority-boosting kthread, one per leaf rcu_node.
1006 static int rcu_boost_kthread(void *arg)
1008 struct rcu_node *rnp = (struct rcu_node *)arg;
1012 trace_rcu_utilization(TPS("Start boost kthread@init"));
1014 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1015 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1016 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1017 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1018 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1019 more2boost = rcu_boost(rnp);
1025 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1026 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1027 schedule_timeout_interruptible(2);
1028 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1033 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1038 * Check to see if it is time to start boosting RCU readers that are
1039 * blocking the current grace period, and, if so, tell the per-rcu_node
1040 * kthread to start boosting them. If there is an expedited grace
1041 * period in progress, it is always time to boost.
1043 * The caller must hold rnp->lock, which this function releases.
1044 * The ->boost_kthread_task is immortal, so we don't need to worry
1045 * about it going away.
1047 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1048 __releases(rnp->lock)
1050 struct task_struct *t;
1052 lockdep_assert_held(&rnp->lock);
1053 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1054 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1057 if (rnp->exp_tasks != NULL ||
1058 (rnp->gp_tasks != NULL &&
1059 rnp->boost_tasks == NULL &&
1061 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1062 if (rnp->exp_tasks == NULL)
1063 rnp->boost_tasks = rnp->gp_tasks;
1064 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1065 t = rnp->boost_kthread_task;
1067 rcu_wake_cond(t, rnp->boost_kthread_status);
1069 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1074 * Wake up the per-CPU kthread to invoke RCU callbacks.
1076 static void invoke_rcu_callbacks_kthread(void)
1078 unsigned long flags;
1080 local_irq_save(flags);
1081 __this_cpu_write(rcu_cpu_has_work, 1);
1082 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1083 current != __this_cpu_read(rcu_cpu_kthread_task)) {
1084 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1085 __this_cpu_read(rcu_cpu_kthread_status));
1087 local_irq_restore(flags);
1091 * Is the current CPU running the RCU-callbacks kthread?
1092 * Caller must have preemption disabled.
1094 static bool rcu_is_callbacks_kthread(void)
1096 return __this_cpu_read(rcu_cpu_kthread_task) == current;
1099 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1102 * Do priority-boost accounting for the start of a new grace period.
1104 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1106 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1110 * Create an RCU-boost kthread for the specified node if one does not
1111 * already exist. We only create this kthread for preemptible RCU.
1112 * Returns zero if all is well, a negated errno otherwise.
1114 static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1115 struct rcu_node *rnp)
1117 int rnp_index = rnp - &rsp->node[0];
1118 unsigned long flags;
1119 struct sched_param sp;
1120 struct task_struct *t;
1122 if (rcu_state_p != rsp)
1125 if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1129 if (rnp->boost_kthread_task != NULL)
1131 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1132 "rcub/%d", rnp_index);
1135 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1136 rnp->boost_kthread_task = t;
1137 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1138 sp.sched_priority = kthread_prio;
1139 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1140 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1144 static void rcu_kthread_do_work(void)
1146 rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1147 rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1148 rcu_preempt_do_callbacks();
1151 static void rcu_cpu_kthread_setup(unsigned int cpu)
1153 struct sched_param sp;
1155 sp.sched_priority = kthread_prio;
1156 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1159 static void rcu_cpu_kthread_park(unsigned int cpu)
1161 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1164 static int rcu_cpu_kthread_should_run(unsigned int cpu)
1166 return __this_cpu_read(rcu_cpu_has_work);
1170 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1171 * RCU softirq used in flavors and configurations of RCU that do not
1172 * support RCU priority boosting.
1174 static void rcu_cpu_kthread(unsigned int cpu)
1176 unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1177 char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1180 for (spincnt = 0; spincnt < 10; spincnt++) {
1181 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1183 *statusp = RCU_KTHREAD_RUNNING;
1184 this_cpu_inc(rcu_cpu_kthread_loops);
1185 local_irq_disable();
1190 rcu_kthread_do_work();
1193 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1194 *statusp = RCU_KTHREAD_WAITING;
1198 *statusp = RCU_KTHREAD_YIELDING;
1199 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1200 schedule_timeout_interruptible(2);
1201 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1202 *statusp = RCU_KTHREAD_WAITING;
1206 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1207 * served by the rcu_node in question. The CPU hotplug lock is still
1208 * held, so the value of rnp->qsmaskinit will be stable.
1210 * We don't include outgoingcpu in the affinity set, use -1 if there is
1211 * no outgoing CPU. If there are no CPUs left in the affinity set,
1212 * this function allows the kthread to execute on any CPU.
1214 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1216 struct task_struct *t = rnp->boost_kthread_task;
1217 unsigned long mask = rcu_rnp_online_cpus(rnp);
1223 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1225 for_each_leaf_node_possible_cpu(rnp, cpu)
1226 if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
1228 cpumask_set_cpu(cpu, cm);
1229 if (cpumask_weight(cm) == 0)
1231 set_cpus_allowed_ptr(t, cm);
1232 free_cpumask_var(cm);
1235 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1236 .store = &rcu_cpu_kthread_task,
1237 .thread_should_run = rcu_cpu_kthread_should_run,
1238 .thread_fn = rcu_cpu_kthread,
1239 .thread_comm = "rcuc/%u",
1240 .setup = rcu_cpu_kthread_setup,
1241 .park = rcu_cpu_kthread_park,
1245 * Spawn boost kthreads -- called as soon as the scheduler is running.
1247 static void __init rcu_spawn_boost_kthreads(void)
1249 struct rcu_node *rnp;
1252 for_each_possible_cpu(cpu)
1253 per_cpu(rcu_cpu_has_work, cpu) = 0;
1254 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1255 rcu_for_each_leaf_node(rcu_state_p, rnp)
1256 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1259 static void rcu_prepare_kthreads(int cpu)
1261 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
1262 struct rcu_node *rnp = rdp->mynode;
1264 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1265 if (rcu_scheduler_fully_active)
1266 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1269 #else /* #ifdef CONFIG_RCU_BOOST */
1271 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1272 __releases(rnp->lock)
1274 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1277 static void invoke_rcu_callbacks_kthread(void)
1282 static bool rcu_is_callbacks_kthread(void)
1287 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1291 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1295 static void __init rcu_spawn_boost_kthreads(void)
1299 static void rcu_prepare_kthreads(int cpu)
1303 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1305 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1308 * Check to see if any future RCU-related work will need to be done
1309 * by the current CPU, even if none need be done immediately, returning
1310 * 1 if so. This function is part of the RCU implementation; it is -not-
1311 * an exported member of the RCU API.
1313 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1314 * any flavor of RCU.
1316 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1318 *nextevt = KTIME_MAX;
1319 return rcu_cpu_has_callbacks(NULL);
1323 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1326 static void rcu_cleanup_after_idle(void)
1331 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1334 static void rcu_prepare_for_idle(void)
1339 * Don't bother keeping a running count of the number of RCU callbacks
1340 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1342 static void rcu_idle_count_callbacks_posted(void)
1346 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1349 * This code is invoked when a CPU goes idle, at which point we want
1350 * to have the CPU do everything required for RCU so that it can enter
1351 * the energy-efficient dyntick-idle mode. This is handled by a
1352 * state machine implemented by rcu_prepare_for_idle() below.
1354 * The following three proprocessor symbols control this state machine:
1356 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1357 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1358 * is sized to be roughly one RCU grace period. Those energy-efficiency
1359 * benchmarkers who might otherwise be tempted to set this to a large
1360 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1361 * system. And if you are -that- concerned about energy efficiency,
1362 * just power the system down and be done with it!
1363 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1364 * permitted to sleep in dyntick-idle mode with only lazy RCU
1365 * callbacks pending. Setting this too high can OOM your system.
1367 * The values below work well in practice. If future workloads require
1368 * adjustment, they can be converted into kernel config parameters, though
1369 * making the state machine smarter might be a better option.
1371 #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1372 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1374 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1375 module_param(rcu_idle_gp_delay, int, 0644);
1376 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1377 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1380 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1381 * only if it has been awhile since the last time we did so. Afterwards,
1382 * if there are any callbacks ready for immediate invocation, return true.
1384 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1386 bool cbs_ready = false;
1387 struct rcu_data *rdp;
1388 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1389 struct rcu_node *rnp;
1390 struct rcu_state *rsp;
1392 /* Exit early if we advanced recently. */
1393 if (jiffies == rdtp->last_advance_all)
1395 rdtp->last_advance_all = jiffies;
1397 for_each_rcu_flavor(rsp) {
1398 rdp = this_cpu_ptr(rsp->rda);
1402 * Don't bother checking unless a grace period has
1403 * completed since we last checked and there are
1404 * callbacks not yet ready to invoke.
1406 if ((rdp->completed != rnp->completed ||
1407 unlikely(READ_ONCE(rdp->gpwrap))) &&
1408 rcu_segcblist_pend_cbs(&rdp->cblist))
1409 note_gp_changes(rsp, rdp);
1411 if (rcu_segcblist_ready_cbs(&rdp->cblist))
1418 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1419 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1420 * caller to set the timeout based on whether or not there are non-lazy
1423 * The caller must have disabled interrupts.
1425 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1427 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1430 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_needs_cpu() invoked with irqs enabled!!!");
1432 /* Snapshot to detect later posting of non-lazy callback. */
1433 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1435 /* If no callbacks, RCU doesn't need the CPU. */
1436 if (!rcu_cpu_has_callbacks(&rdtp->all_lazy)) {
1437 *nextevt = KTIME_MAX;
1441 /* Attempt to advance callbacks. */
1442 if (rcu_try_advance_all_cbs()) {
1443 /* Some ready to invoke, so initiate later invocation. */
1447 rdtp->last_accelerate = jiffies;
1449 /* Request timer delay depending on laziness, and round. */
1450 if (!rdtp->all_lazy) {
1451 dj = round_up(rcu_idle_gp_delay + jiffies,
1452 rcu_idle_gp_delay) - jiffies;
1454 dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1456 *nextevt = basemono + dj * TICK_NSEC;
1461 * Prepare a CPU for idle from an RCU perspective. The first major task
1462 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1463 * The second major task is to check to see if a non-lazy callback has
1464 * arrived at a CPU that previously had only lazy callbacks. The third
1465 * major task is to accelerate (that is, assign grace-period numbers to)
1466 * any recently arrived callbacks.
1468 * The caller must have disabled interrupts.
1470 static void rcu_prepare_for_idle(void)
1473 struct rcu_data *rdp;
1474 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1475 struct rcu_node *rnp;
1476 struct rcu_state *rsp;
1479 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_prepare_for_idle() invoked with irqs enabled!!!");
1480 if (rcu_is_nocb_cpu(smp_processor_id()))
1483 /* Handle nohz enablement switches conservatively. */
1484 tne = READ_ONCE(tick_nohz_active);
1485 if (tne != rdtp->tick_nohz_enabled_snap) {
1486 if (rcu_cpu_has_callbacks(NULL))
1487 invoke_rcu_core(); /* force nohz to see update. */
1488 rdtp->tick_nohz_enabled_snap = tne;
1495 * If a non-lazy callback arrived at a CPU having only lazy
1496 * callbacks, invoke RCU core for the side-effect of recalculating
1497 * idle duration on re-entry to idle.
1499 if (rdtp->all_lazy &&
1500 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1501 rdtp->all_lazy = false;
1502 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1508 * If we have not yet accelerated this jiffy, accelerate all
1509 * callbacks on this CPU.
1511 if (rdtp->last_accelerate == jiffies)
1513 rdtp->last_accelerate = jiffies;
1514 for_each_rcu_flavor(rsp) {
1515 rdp = this_cpu_ptr(rsp->rda);
1516 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1519 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1520 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1521 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1523 rcu_gp_kthread_wake(rsp);
1528 * Clean up for exit from idle. Attempt to advance callbacks based on
1529 * any grace periods that elapsed while the CPU was idle, and if any
1530 * callbacks are now ready to invoke, initiate invocation.
1532 static void rcu_cleanup_after_idle(void)
1534 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_cleanup_after_idle() invoked with irqs enabled!!!");
1535 if (rcu_is_nocb_cpu(smp_processor_id()))
1537 if (rcu_try_advance_all_cbs())
1542 * Keep a running count of the number of non-lazy callbacks posted
1543 * on this CPU. This running counter (which is never decremented) allows
1544 * rcu_prepare_for_idle() to detect when something out of the idle loop
1545 * posts a callback, even if an equal number of callbacks are invoked.
1546 * Of course, callbacks should only be posted from within a trace event
1547 * designed to be called from idle or from within RCU_NONIDLE().
1549 static void rcu_idle_count_callbacks_posted(void)
1551 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1555 * Data for flushing lazy RCU callbacks at OOM time.
1557 static atomic_t oom_callback_count;
1558 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1561 * RCU OOM callback -- decrement the outstanding count and deliver the
1562 * wake-up if we are the last one.
1564 static void rcu_oom_callback(struct rcu_head *rhp)
1566 if (atomic_dec_and_test(&oom_callback_count))
1567 wake_up(&oom_callback_wq);
1571 * Post an rcu_oom_notify callback on the current CPU if it has at
1572 * least one lazy callback. This will unnecessarily post callbacks
1573 * to CPUs that already have a non-lazy callback at the end of their
1574 * callback list, but this is an infrequent operation, so accept some
1575 * extra overhead to keep things simple.
1577 static void rcu_oom_notify_cpu(void *unused)
1579 struct rcu_state *rsp;
1580 struct rcu_data *rdp;
1582 for_each_rcu_flavor(rsp) {
1583 rdp = raw_cpu_ptr(rsp->rda);
1584 if (rcu_segcblist_n_lazy_cbs(&rdp->cblist)) {
1585 atomic_inc(&oom_callback_count);
1586 rsp->call(&rdp->oom_head, rcu_oom_callback);
1592 * If low on memory, ensure that each CPU has a non-lazy callback.
1593 * This will wake up CPUs that have only lazy callbacks, in turn
1594 * ensuring that they free up the corresponding memory in a timely manner.
1595 * Because an uncertain amount of memory will be freed in some uncertain
1596 * timeframe, we do not claim to have freed anything.
1598 static int rcu_oom_notify(struct notifier_block *self,
1599 unsigned long notused, void *nfreed)
1603 /* Wait for callbacks from earlier instance to complete. */
1604 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1605 smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1608 * Prevent premature wakeup: ensure that all increments happen
1609 * before there is a chance of the counter reaching zero.
1611 atomic_set(&oom_callback_count, 1);
1613 for_each_online_cpu(cpu) {
1614 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1615 cond_resched_rcu_qs();
1618 /* Unconditionally decrement: no need to wake ourselves up. */
1619 atomic_dec(&oom_callback_count);
1624 static struct notifier_block rcu_oom_nb = {
1625 .notifier_call = rcu_oom_notify
1628 static int __init rcu_register_oom_notifier(void)
1630 register_oom_notifier(&rcu_oom_nb);
1633 early_initcall(rcu_register_oom_notifier);
1635 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1637 #ifdef CONFIG_RCU_FAST_NO_HZ
1639 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1641 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1642 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1644 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1645 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1647 rdtp->all_lazy ? 'L' : '.',
1648 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1651 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1653 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1658 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1660 /* Initiate the stall-info list. */
1661 static void print_cpu_stall_info_begin(void)
1667 * Print out diagnostic information for the specified stalled CPU.
1669 * If the specified CPU is aware of the current RCU grace period
1670 * (flavor specified by rsp), then print the number of scheduling
1671 * clock interrupts the CPU has taken during the time that it has
1672 * been aware. Otherwise, print the number of RCU grace periods
1673 * that this CPU is ignorant of, for example, "1" if the CPU was
1674 * aware of the previous grace period.
1676 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1678 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1680 char fast_no_hz[72];
1681 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1682 struct rcu_dynticks *rdtp = rdp->dynticks;
1684 unsigned long ticks_value;
1687 * We could be printing a lot while holding a spinlock. Avoid
1688 * triggering hard lockup.
1690 touch_nmi_watchdog();
1692 if (rsp->gpnum == rdp->gpnum) {
1693 ticks_title = "ticks this GP";
1694 ticks_value = rdp->ticks_this_gp;
1696 ticks_title = "GPs behind";
1697 ticks_value = rsp->gpnum - rdp->gpnum;
1699 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1700 pr_err("\t%d-%c%c%c: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u fqs=%ld %s\n",
1702 "O."[!!cpu_online(cpu)],
1703 "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
1704 "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
1705 ticks_value, ticks_title,
1706 rcu_dynticks_snap(rdtp) & 0xfff,
1707 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1708 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1709 READ_ONCE(rsp->n_force_qs) - rsp->n_force_qs_gpstart,
1713 /* Terminate the stall-info list. */
1714 static void print_cpu_stall_info_end(void)
1719 /* Zero ->ticks_this_gp for all flavors of RCU. */
1720 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1722 rdp->ticks_this_gp = 0;
1723 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1726 /* Increment ->ticks_this_gp for all flavors of RCU. */
1727 static void increment_cpu_stall_ticks(void)
1729 struct rcu_state *rsp;
1731 for_each_rcu_flavor(rsp)
1732 raw_cpu_inc(rsp->rda->ticks_this_gp);
1735 #ifdef CONFIG_RCU_NOCB_CPU
1738 * Offload callback processing from the boot-time-specified set of CPUs
1739 * specified by rcu_nocb_mask. For each CPU in the set, there is a
1740 * kthread created that pulls the callbacks from the corresponding CPU,
1741 * waits for a grace period to elapse, and invokes the callbacks.
1742 * The no-CBs CPUs do a wake_up() on their kthread when they insert
1743 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1744 * has been specified, in which case each kthread actively polls its
1745 * CPU. (Which isn't so great for energy efficiency, but which does
1746 * reduce RCU's overhead on that CPU.)
1748 * This is intended to be used in conjunction with Frederic Weisbecker's
1749 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1750 * running CPU-bound user-mode computations.
1752 * Offloading of callback processing could also in theory be used as
1753 * an energy-efficiency measure because CPUs with no RCU callbacks
1754 * queued are more aggressive about entering dyntick-idle mode.
1758 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1759 static int __init rcu_nocb_setup(char *str)
1761 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1762 have_rcu_nocb_mask = true;
1763 cpulist_parse(str, rcu_nocb_mask);
1766 __setup("rcu_nocbs=", rcu_nocb_setup);
1768 static int __init parse_rcu_nocb_poll(char *arg)
1770 rcu_nocb_poll = true;
1773 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1776 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1779 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
1785 * Set the root rcu_node structure's ->need_future_gp field
1786 * based on the sum of those of all rcu_node structures. This does
1787 * double-count the root rcu_node structure's requests, but this
1788 * is necessary to handle the possibility of a rcu_nocb_kthread()
1789 * having awakened during the time that the rcu_node structures
1790 * were being updated for the end of the previous grace period.
1792 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
1794 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
1797 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
1799 return &rnp->nocb_gp_wq[rnp->completed & 0x1];
1802 static void rcu_init_one_nocb(struct rcu_node *rnp)
1804 init_swait_queue_head(&rnp->nocb_gp_wq[0]);
1805 init_swait_queue_head(&rnp->nocb_gp_wq[1]);
1808 /* Is the specified CPU a no-CBs CPU? */
1809 bool rcu_is_nocb_cpu(int cpu)
1811 if (have_rcu_nocb_mask)
1812 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1817 * Kick the leader kthread for this NOCB group. Caller holds ->nocb_lock
1818 * and this function releases it.
1820 static void __wake_nocb_leader(struct rcu_data *rdp, bool force,
1821 unsigned long flags)
1822 __releases(rdp->nocb_lock)
1824 struct rcu_data *rdp_leader = rdp->nocb_leader;
1826 lockdep_assert_held(&rdp->nocb_lock);
1827 if (!READ_ONCE(rdp_leader->nocb_kthread)) {
1828 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1831 if (rdp_leader->nocb_leader_sleep || force) {
1832 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
1833 WRITE_ONCE(rdp_leader->nocb_leader_sleep, false);
1834 del_timer(&rdp->nocb_timer);
1835 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1836 smp_mb(); /* ->nocb_leader_sleep before swake_up(). */
1837 swake_up(&rdp_leader->nocb_wq);
1839 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1844 * Kick the leader kthread for this NOCB group, but caller has not
1847 static void wake_nocb_leader(struct rcu_data *rdp, bool force)
1849 unsigned long flags;
1851 raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
1852 __wake_nocb_leader(rdp, force, flags);
1856 * Arrange to wake the leader kthread for this NOCB group at some
1857 * future time when it is safe to do so.
1859 static void wake_nocb_leader_defer(struct rcu_data *rdp, int waketype,
1862 unsigned long flags;
1864 raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
1865 if (rdp->nocb_defer_wakeup == RCU_NOCB_WAKE_NOT)
1866 mod_timer(&rdp->nocb_timer, jiffies + 1);
1867 WRITE_ONCE(rdp->nocb_defer_wakeup, waketype);
1868 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, reason);
1869 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1873 * Does the specified CPU need an RCU callback for the specified flavor
1876 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
1878 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1880 #ifdef CONFIG_PROVE_RCU
1881 struct rcu_head *rhp;
1882 #endif /* #ifdef CONFIG_PROVE_RCU */
1885 * Check count of all no-CBs callbacks awaiting invocation.
1886 * There needs to be a barrier before this function is called,
1887 * but associated with a prior determination that no more
1888 * callbacks would be posted. In the worst case, the first
1889 * barrier in _rcu_barrier() suffices (but the caller cannot
1890 * necessarily rely on this, not a substitute for the caller
1891 * getting the concurrency design right!). There must also be
1892 * a barrier between the following load an posting of a callback
1893 * (if a callback is in fact needed). This is associated with an
1894 * atomic_inc() in the caller.
1896 ret = atomic_long_read(&rdp->nocb_q_count);
1898 #ifdef CONFIG_PROVE_RCU
1899 rhp = READ_ONCE(rdp->nocb_head);
1901 rhp = READ_ONCE(rdp->nocb_gp_head);
1903 rhp = READ_ONCE(rdp->nocb_follower_head);
1905 /* Having no rcuo kthread but CBs after scheduler starts is bad! */
1906 if (!READ_ONCE(rdp->nocb_kthread) && rhp &&
1907 rcu_scheduler_fully_active) {
1908 /* RCU callback enqueued before CPU first came online??? */
1909 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
1913 #endif /* #ifdef CONFIG_PROVE_RCU */
1919 * Enqueue the specified string of rcu_head structures onto the specified
1920 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
1921 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
1922 * counts are supplied by rhcount and rhcount_lazy.
1924 * If warranted, also wake up the kthread servicing this CPUs queues.
1926 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
1927 struct rcu_head *rhp,
1928 struct rcu_head **rhtp,
1929 int rhcount, int rhcount_lazy,
1930 unsigned long flags)
1933 struct rcu_head **old_rhpp;
1934 struct task_struct *t;
1936 /* Enqueue the callback on the nocb list and update counts. */
1937 atomic_long_add(rhcount, &rdp->nocb_q_count);
1938 /* rcu_barrier() relies on ->nocb_q_count add before xchg. */
1939 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
1940 WRITE_ONCE(*old_rhpp, rhp);
1941 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
1942 smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
1944 /* If we are not being polled and there is a kthread, awaken it ... */
1945 t = READ_ONCE(rdp->nocb_kthread);
1946 if (rcu_nocb_poll || !t) {
1947 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1948 TPS("WakeNotPoll"));
1951 len = atomic_long_read(&rdp->nocb_q_count);
1952 if (old_rhpp == &rdp->nocb_head) {
1953 if (!irqs_disabled_flags(flags)) {
1954 /* ... if queue was empty ... */
1955 wake_nocb_leader(rdp, false);
1956 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1959 wake_nocb_leader_defer(rdp, RCU_NOCB_WAKE,
1960 TPS("WakeEmptyIsDeferred"));
1962 rdp->qlen_last_fqs_check = 0;
1963 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
1964 /* ... or if many callbacks queued. */
1965 if (!irqs_disabled_flags(flags)) {
1966 wake_nocb_leader(rdp, true);
1967 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
1970 wake_nocb_leader_defer(rdp, RCU_NOCB_WAKE,
1971 TPS("WakeOvfIsDeferred"));
1973 rdp->qlen_last_fqs_check = LONG_MAX / 2;
1975 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
1981 * This is a helper for __call_rcu(), which invokes this when the normal
1982 * callback queue is inoperable. If this is not a no-CBs CPU, this
1983 * function returns failure back to __call_rcu(), which can complain
1986 * Otherwise, this function queues the callback where the corresponding
1987 * "rcuo" kthread can find it.
1989 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
1990 bool lazy, unsigned long flags)
1993 if (!rcu_is_nocb_cpu(rdp->cpu))
1995 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
1996 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
1997 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
1998 (unsigned long)rhp->func,
1999 -atomic_long_read(&rdp->nocb_q_count_lazy),
2000 -atomic_long_read(&rdp->nocb_q_count));
2002 trace_rcu_callback(rdp->rsp->name, rhp,
2003 -atomic_long_read(&rdp->nocb_q_count_lazy),
2004 -atomic_long_read(&rdp->nocb_q_count));
2007 * If called from an extended quiescent state with interrupts
2008 * disabled, invoke the RCU core in order to allow the idle-entry
2009 * deferred-wakeup check to function.
2011 if (irqs_disabled_flags(flags) &&
2012 !rcu_is_watching() &&
2013 cpu_online(smp_processor_id()))
2020 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2023 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_data *my_rdp,
2024 struct rcu_data *rdp,
2025 unsigned long flags)
2027 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_nocb_adopt_orphan_cbs() invoked with irqs enabled!!!");
2028 if (!rcu_is_nocb_cpu(smp_processor_id()))
2029 return false; /* Not NOCBs CPU, caller must migrate CBs. */
2030 __call_rcu_nocb_enqueue(my_rdp, rcu_segcblist_head(&rdp->cblist),
2031 rcu_segcblist_tail(&rdp->cblist),
2032 rcu_segcblist_n_cbs(&rdp->cblist),
2033 rcu_segcblist_n_lazy_cbs(&rdp->cblist), flags);
2034 rcu_segcblist_init(&rdp->cblist);
2035 rcu_segcblist_disable(&rdp->cblist);
2040 * If necessary, kick off a new grace period, and either way wait
2041 * for a subsequent grace period to complete.
2043 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2047 unsigned long flags;
2049 struct rcu_node *rnp = rdp->mynode;
2051 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2052 needwake = rcu_start_future_gp(rnp, rdp, &c);
2053 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2055 rcu_gp_kthread_wake(rdp->rsp);
2058 * Wait for the grace period. Do so interruptibly to avoid messing
2059 * up the load average.
2061 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
2063 swait_event_interruptible(
2064 rnp->nocb_gp_wq[c & 0x1],
2065 (d = ULONG_CMP_GE(READ_ONCE(rnp->completed), c)));
2068 WARN_ON(signal_pending(current));
2069 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
2071 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
2072 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2076 * Leaders come here to wait for additional callbacks to show up.
2077 * This function does not return until callbacks appear.
2079 static void nocb_leader_wait(struct rcu_data *my_rdp)
2081 bool firsttime = true;
2082 unsigned long flags;
2084 struct rcu_data *rdp;
2085 struct rcu_head **tail;
2089 /* Wait for callbacks to appear. */
2090 if (!rcu_nocb_poll) {
2091 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, TPS("Sleep"));
2092 swait_event_interruptible(my_rdp->nocb_wq,
2093 !READ_ONCE(my_rdp->nocb_leader_sleep));
2094 raw_spin_lock_irqsave(&my_rdp->nocb_lock, flags);
2095 my_rdp->nocb_leader_sleep = true;
2096 WRITE_ONCE(my_rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
2097 del_timer(&my_rdp->nocb_timer);
2098 raw_spin_unlock_irqrestore(&my_rdp->nocb_lock, flags);
2099 } else if (firsttime) {
2100 firsttime = false; /* Don't drown trace log with "Poll"! */
2101 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, TPS("Poll"));
2105 * Each pass through the following loop checks a follower for CBs.
2106 * We are our own first follower. Any CBs found are moved to
2107 * nocb_gp_head, where they await a grace period.
2110 smp_mb(); /* wakeup and _sleep before ->nocb_head reads. */
2111 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2112 rdp->nocb_gp_head = READ_ONCE(rdp->nocb_head);
2113 if (!rdp->nocb_gp_head)
2114 continue; /* No CBs here, try next follower. */
2116 /* Move callbacks to wait-for-GP list, which is empty. */
2117 WRITE_ONCE(rdp->nocb_head, NULL);
2118 rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2122 /* No callbacks? Sleep a bit if polling, and go retry. */
2123 if (unlikely(!gotcbs)) {
2124 WARN_ON(signal_pending(current));
2125 if (rcu_nocb_poll) {
2126 schedule_timeout_interruptible(1);
2128 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu,
2134 /* Wait for one grace period. */
2135 rcu_nocb_wait_gp(my_rdp);
2137 /* Each pass through the following loop wakes a follower, if needed. */
2138 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2139 if (!rcu_nocb_poll &&
2140 READ_ONCE(rdp->nocb_head) &&
2141 READ_ONCE(my_rdp->nocb_leader_sleep)) {
2142 raw_spin_lock_irqsave(&my_rdp->nocb_lock, flags);
2143 my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/
2144 raw_spin_unlock_irqrestore(&my_rdp->nocb_lock, flags);
2146 if (!rdp->nocb_gp_head)
2147 continue; /* No CBs, so no need to wake follower. */
2149 /* Append callbacks to follower's "done" list. */
2150 raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
2151 tail = rdp->nocb_follower_tail;
2152 rdp->nocb_follower_tail = rdp->nocb_gp_tail;
2153 *tail = rdp->nocb_gp_head;
2154 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
2155 if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
2156 /* List was empty, so wake up the follower. */
2157 swake_up(&rdp->nocb_wq);
2161 /* If we (the leader) don't have CBs, go wait some more. */
2162 if (!my_rdp->nocb_follower_head)
2167 * Followers come here to wait for additional callbacks to show up.
2168 * This function does not return until callbacks appear.
2170 static void nocb_follower_wait(struct rcu_data *rdp)
2173 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("FollowerSleep"));
2174 swait_event_interruptible(rdp->nocb_wq,
2175 READ_ONCE(rdp->nocb_follower_head));
2176 if (smp_load_acquire(&rdp->nocb_follower_head)) {
2177 /* ^^^ Ensure CB invocation follows _head test. */
2180 WARN_ON(signal_pending(current));
2181 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WokeEmpty"));
2186 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2187 * callbacks queued by the corresponding no-CBs CPU, however, there is
2188 * an optional leader-follower relationship so that the grace-period
2189 * kthreads don't have to do quite so many wakeups.
2191 static int rcu_nocb_kthread(void *arg)
2194 unsigned long flags;
2195 struct rcu_head *list;
2196 struct rcu_head *next;
2197 struct rcu_head **tail;
2198 struct rcu_data *rdp = arg;
2200 /* Each pass through this loop invokes one batch of callbacks */
2202 /* Wait for callbacks. */
2203 if (rdp->nocb_leader == rdp)
2204 nocb_leader_wait(rdp);
2206 nocb_follower_wait(rdp);
2208 /* Pull the ready-to-invoke callbacks onto local list. */
2209 raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
2210 list = rdp->nocb_follower_head;
2211 rdp->nocb_follower_head = NULL;
2212 tail = rdp->nocb_follower_tail;
2213 rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2214 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
2216 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WokeNonEmpty"));
2218 /* Each pass through the following loop invokes a callback. */
2219 trace_rcu_batch_start(rdp->rsp->name,
2220 atomic_long_read(&rdp->nocb_q_count_lazy),
2221 atomic_long_read(&rdp->nocb_q_count), -1);
2225 /* Wait for enqueuing to complete, if needed. */
2226 while (next == NULL && &list->next != tail) {
2227 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2229 schedule_timeout_interruptible(1);
2230 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2234 debug_rcu_head_unqueue(list);
2236 if (__rcu_reclaim(rdp->rsp->name, list))
2240 cond_resched_rcu_qs();
2243 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2244 smp_mb__before_atomic(); /* _add after CB invocation. */
2245 atomic_long_add(-c, &rdp->nocb_q_count);
2246 atomic_long_add(-cl, &rdp->nocb_q_count_lazy);
2247 rdp->n_nocbs_invoked += c;
2252 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2253 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2255 return READ_ONCE(rdp->nocb_defer_wakeup);
2258 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2259 static void do_nocb_deferred_wakeup_common(struct rcu_data *rdp)
2261 unsigned long flags;
2264 raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
2265 if (!rcu_nocb_need_deferred_wakeup(rdp)) {
2266 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
2269 ndw = READ_ONCE(rdp->nocb_defer_wakeup);
2270 WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
2271 __wake_nocb_leader(rdp, ndw == RCU_NOCB_WAKE_FORCE, flags);
2272 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake"));
2275 /* Do a deferred wakeup of rcu_nocb_kthread() from a timer handler. */
2276 static void do_nocb_deferred_wakeup_timer(unsigned long x)
2278 do_nocb_deferred_wakeup_common((struct rcu_data *)x);
2282 * Do a deferred wakeup of rcu_nocb_kthread() from fastpath.
2283 * This means we do an inexact common-case check. Note that if
2284 * we miss, ->nocb_timer will eventually clean things up.
2286 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2288 if (rcu_nocb_need_deferred_wakeup(rdp))
2289 do_nocb_deferred_wakeup_common(rdp);
2292 void __init rcu_init_nohz(void)
2295 bool need_rcu_nocb_mask = true;
2296 struct rcu_state *rsp;
2298 #if defined(CONFIG_NO_HZ_FULL)
2299 if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2300 need_rcu_nocb_mask = true;
2301 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2303 if (!have_rcu_nocb_mask && need_rcu_nocb_mask) {
2304 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2305 pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2308 have_rcu_nocb_mask = true;
2310 if (!have_rcu_nocb_mask)
2313 #if defined(CONFIG_NO_HZ_FULL)
2314 if (tick_nohz_full_running)
2315 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2316 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2318 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2319 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
2320 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2323 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2324 cpumask_pr_args(rcu_nocb_mask));
2326 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2328 for_each_rcu_flavor(rsp) {
2329 for_each_cpu(cpu, rcu_nocb_mask)
2330 init_nocb_callback_list(per_cpu_ptr(rsp->rda, cpu));
2331 rcu_organize_nocb_kthreads(rsp);
2335 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2336 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2338 rdp->nocb_tail = &rdp->nocb_head;
2339 init_swait_queue_head(&rdp->nocb_wq);
2340 rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2341 raw_spin_lock_init(&rdp->nocb_lock);
2342 setup_timer(&rdp->nocb_timer, do_nocb_deferred_wakeup_timer,
2343 (unsigned long)rdp);
2347 * If the specified CPU is a no-CBs CPU that does not already have its
2348 * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are
2349 * brought online out of order, this can require re-organizing the
2350 * leader-follower relationships.
2352 static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu)
2354 struct rcu_data *rdp;
2355 struct rcu_data *rdp_last;
2356 struct rcu_data *rdp_old_leader;
2357 struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu);
2358 struct task_struct *t;
2361 * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2362 * then nothing to do.
2364 if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread)
2367 /* If we didn't spawn the leader first, reorganize! */
2368 rdp_old_leader = rdp_spawn->nocb_leader;
2369 if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) {
2371 rdp = rdp_old_leader;
2373 rdp->nocb_leader = rdp_spawn;
2374 if (rdp_last && rdp != rdp_spawn)
2375 rdp_last->nocb_next_follower = rdp;
2376 if (rdp == rdp_spawn) {
2377 rdp = rdp->nocb_next_follower;
2380 rdp = rdp->nocb_next_follower;
2381 rdp_last->nocb_next_follower = NULL;
2384 rdp_spawn->nocb_next_follower = rdp_old_leader;
2387 /* Spawn the kthread for this CPU and RCU flavor. */
2388 t = kthread_run(rcu_nocb_kthread, rdp_spawn,
2389 "rcuo%c/%d", rsp->abbr, cpu);
2391 WRITE_ONCE(rdp_spawn->nocb_kthread, t);
2395 * If the specified CPU is a no-CBs CPU that does not already have its
2396 * rcuo kthreads, spawn them.
2398 static void rcu_spawn_all_nocb_kthreads(int cpu)
2400 struct rcu_state *rsp;
2402 if (rcu_scheduler_fully_active)
2403 for_each_rcu_flavor(rsp)
2404 rcu_spawn_one_nocb_kthread(rsp, cpu);
2408 * Once the scheduler is running, spawn rcuo kthreads for all online
2409 * no-CBs CPUs. This assumes that the early_initcall()s happen before
2410 * non-boot CPUs come online -- if this changes, we will need to add
2411 * some mutual exclusion.
2413 static void __init rcu_spawn_nocb_kthreads(void)
2417 for_each_online_cpu(cpu)
2418 rcu_spawn_all_nocb_kthreads(cpu);
2421 /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */
2422 static int rcu_nocb_leader_stride = -1;
2423 module_param(rcu_nocb_leader_stride, int, 0444);
2426 * Initialize leader-follower relationships for all no-CBs CPU.
2428 static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp)
2431 int ls = rcu_nocb_leader_stride;
2432 int nl = 0; /* Next leader. */
2433 struct rcu_data *rdp;
2434 struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */
2435 struct rcu_data *rdp_prev = NULL;
2437 if (!have_rcu_nocb_mask)
2440 ls = int_sqrt(nr_cpu_ids);
2441 rcu_nocb_leader_stride = ls;
2445 * Each pass through this loop sets up one rcu_data structure.
2446 * Should the corresponding CPU come online in the future, then
2447 * we will spawn the needed set of rcu_nocb_kthread() kthreads.
2449 for_each_cpu(cpu, rcu_nocb_mask) {
2450 rdp = per_cpu_ptr(rsp->rda, cpu);
2451 if (rdp->cpu >= nl) {
2452 /* New leader, set up for followers & next leader. */
2453 nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2454 rdp->nocb_leader = rdp;
2457 /* Another follower, link to previous leader. */
2458 rdp->nocb_leader = rdp_leader;
2459 rdp_prev->nocb_next_follower = rdp;
2465 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2466 static bool init_nocb_callback_list(struct rcu_data *rdp)
2468 if (!rcu_is_nocb_cpu(rdp->cpu))
2471 /* If there are early-boot callbacks, move them to nocb lists. */
2472 if (!rcu_segcblist_empty(&rdp->cblist)) {
2473 rdp->nocb_head = rcu_segcblist_head(&rdp->cblist);
2474 rdp->nocb_tail = rcu_segcblist_tail(&rdp->cblist);
2475 atomic_long_set(&rdp->nocb_q_count,
2476 rcu_segcblist_n_cbs(&rdp->cblist));
2477 atomic_long_set(&rdp->nocb_q_count_lazy,
2478 rcu_segcblist_n_lazy_cbs(&rdp->cblist));
2479 rcu_segcblist_init(&rdp->cblist);
2481 rcu_segcblist_disable(&rdp->cblist);
2485 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2487 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
2489 WARN_ON_ONCE(1); /* Should be dead code. */
2493 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
2497 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2501 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
2506 static void rcu_init_one_nocb(struct rcu_node *rnp)
2510 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2511 bool lazy, unsigned long flags)
2516 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_data *my_rdp,
2517 struct rcu_data *rdp,
2518 unsigned long flags)
2523 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2527 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2532 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2536 static void rcu_spawn_all_nocb_kthreads(int cpu)
2540 static void __init rcu_spawn_nocb_kthreads(void)
2544 static bool init_nocb_callback_list(struct rcu_data *rdp)
2549 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2552 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2553 * arbitrarily long period of time with the scheduling-clock tick turned
2554 * off. RCU will be paying attention to this CPU because it is in the
2555 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2556 * machine because the scheduling-clock tick has been disabled. Therefore,
2557 * if an adaptive-ticks CPU is failing to respond to the current grace
2558 * period and has not be idle from an RCU perspective, kick it.
2560 static void __maybe_unused rcu_kick_nohz_cpu(int cpu)
2562 #ifdef CONFIG_NO_HZ_FULL
2563 if (tick_nohz_full_cpu(cpu))
2564 smp_send_reschedule(cpu);
2565 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2569 * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2570 * grace-period kthread will do force_quiescent_state() processing?
2571 * The idea is to avoid waking up RCU core processing on such a
2572 * CPU unless the grace period has extended for too long.
2574 * This code relies on the fact that all NO_HZ_FULL CPUs are also
2575 * CONFIG_RCU_NOCB_CPU CPUs.
2577 static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
2579 #ifdef CONFIG_NO_HZ_FULL
2580 if (tick_nohz_full_cpu(smp_processor_id()) &&
2581 (!rcu_gp_in_progress(rsp) ||
2582 ULONG_CMP_LT(jiffies, READ_ONCE(rsp->gp_start) + HZ)))
2584 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2589 * Bind the grace-period kthread for the sysidle flavor of RCU to the
2592 static void rcu_bind_gp_kthread(void)
2594 int __maybe_unused cpu;
2596 if (!tick_nohz_full_enabled())
2598 housekeeping_affine(current);
2601 /* Record the current task on dyntick-idle entry. */
2602 static void rcu_dynticks_task_enter(void)
2604 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2605 WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
2606 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2609 /* Record no current task on dyntick-idle exit. */
2610 static void rcu_dynticks_task_exit(void)
2612 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2613 WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
2614 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */