357891104ea0797df85b8844a4a8b4cb79a1acab
[muen/linux.git] / kernel / rcu / tree_plugin.h
1 /*
2  * Read-Copy Update mechanism for mutual exclusion (tree-based version)
3  * Internal non-public definitions that provide either classic
4  * or preemptible semantics.
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License as published by
8  * the Free Software Foundation; either version 2 of the License, or
9  * (at your option) any later version.
10  *
11  * This program is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
14  * GNU General Public License for more details.
15  *
16  * You should have received a copy of the GNU General Public License
17  * along with this program; if not, you can access it online at
18  * http://www.gnu.org/licenses/gpl-2.0.html.
19  *
20  * Copyright Red Hat, 2009
21  * Copyright IBM Corporation, 2009
22  *
23  * Author: Ingo Molnar <mingo@elte.hu>
24  *         Paul E. McKenney <paulmck@linux.vnet.ibm.com>
25  */
26
27 #include <linux/delay.h>
28 #include <linux/gfp.h>
29 #include <linux/oom.h>
30 #include <linux/smpboot.h>
31 #include "../time/tick-internal.h"
32
33 #define RCU_KTHREAD_PRIO 1
34
35 #ifdef CONFIG_RCU_BOOST
36
37 #include "../locking/rtmutex_common.h"
38 #define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO
39
40 /*
41  * Control variables for per-CPU and per-rcu_node kthreads.  These
42  * handle all flavors of RCU.
43  */
44 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
45 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
46 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
47 DEFINE_PER_CPU(char, rcu_cpu_has_work);
48
49 #else /* #ifdef CONFIG_RCU_BOOST */
50
51 #define RCU_BOOST_PRIO RCU_KTHREAD_PRIO
52
53 #endif /* #else #ifdef CONFIG_RCU_BOOST */
54
55 #ifdef CONFIG_RCU_NOCB_CPU
56 static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
57 static bool have_rcu_nocb_mask;     /* Was rcu_nocb_mask allocated? */
58 static bool __read_mostly rcu_nocb_poll;    /* Offload kthread are to poll. */
59 static char __initdata nocb_buf[NR_CPUS * 5];
60 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
61
62 /*
63  * Check the RCU kernel configuration parameters and print informative
64  * messages about anything out of the ordinary.  If you like #ifdef, you
65  * will love this function.
66  */
67 static void __init rcu_bootup_announce_oddness(void)
68 {
69 #ifdef CONFIG_RCU_TRACE
70         pr_info("\tRCU debugfs-based tracing is enabled.\n");
71 #endif
72 #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
73         pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
74                CONFIG_RCU_FANOUT);
75 #endif
76 #ifdef CONFIG_RCU_FANOUT_EXACT
77         pr_info("\tHierarchical RCU autobalancing is disabled.\n");
78 #endif
79 #ifdef CONFIG_RCU_FAST_NO_HZ
80         pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
81 #endif
82 #ifdef CONFIG_PROVE_RCU
83         pr_info("\tRCU lockdep checking is enabled.\n");
84 #endif
85 #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
86         pr_info("\tRCU torture testing starts during boot.\n");
87 #endif
88 #if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE)
89         pr_info("\tDump stacks of tasks blocking RCU-preempt GP.\n");
90 #endif
91 #if defined(CONFIG_RCU_CPU_STALL_INFO)
92         pr_info("\tAdditional per-CPU info printed with stalls.\n");
93 #endif
94 #if NUM_RCU_LVL_4 != 0
95         pr_info("\tFour-level hierarchy is enabled.\n");
96 #endif
97         if (rcu_fanout_leaf != CONFIG_RCU_FANOUT_LEAF)
98                 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
99         if (nr_cpu_ids != NR_CPUS)
100                 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
101 }
102
103 #ifdef CONFIG_TREE_PREEMPT_RCU
104
105 RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
106 static struct rcu_state *rcu_state_p = &rcu_preempt_state;
107
108 static int rcu_preempted_readers_exp(struct rcu_node *rnp);
109
110 /*
111  * Tell them what RCU they are running.
112  */
113 static void __init rcu_bootup_announce(void)
114 {
115         pr_info("Preemptible hierarchical RCU implementation.\n");
116         rcu_bootup_announce_oddness();
117 }
118
119 /*
120  * Return the number of RCU-preempt batches processed thus far
121  * for debug and statistics.
122  */
123 static long rcu_batches_completed_preempt(void)
124 {
125         return rcu_preempt_state.completed;
126 }
127 EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt);
128
129 /*
130  * Return the number of RCU batches processed thus far for debug & stats.
131  */
132 long rcu_batches_completed(void)
133 {
134         return rcu_batches_completed_preempt();
135 }
136 EXPORT_SYMBOL_GPL(rcu_batches_completed);
137
138 /*
139  * Record a preemptible-RCU quiescent state for the specified CPU.  Note
140  * that this just means that the task currently running on the CPU is
141  * not in a quiescent state.  There might be any number of tasks blocked
142  * while in an RCU read-side critical section.
143  *
144  * As with the other rcu_*_qs() functions, callers to this function
145  * must disable preemption.
146  */
147 static void rcu_preempt_qs(void)
148 {
149         if (!__this_cpu_read(rcu_preempt_data.passed_quiesce)) {
150                 trace_rcu_grace_period(TPS("rcu_preempt"),
151                                        __this_cpu_read(rcu_preempt_data.gpnum),
152                                        TPS("cpuqs"));
153                 __this_cpu_write(rcu_preempt_data.passed_quiesce, 1);
154                 barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */
155                 current->rcu_read_unlock_special.b.need_qs = false;
156         }
157 }
158
159 /*
160  * We have entered the scheduler, and the current task might soon be
161  * context-switched away from.  If this task is in an RCU read-side
162  * critical section, we will no longer be able to rely on the CPU to
163  * record that fact, so we enqueue the task on the blkd_tasks list.
164  * The task will dequeue itself when it exits the outermost enclosing
165  * RCU read-side critical section.  Therefore, the current grace period
166  * cannot be permitted to complete until the blkd_tasks list entries
167  * predating the current grace period drain, in other words, until
168  * rnp->gp_tasks becomes NULL.
169  *
170  * Caller must disable preemption.
171  */
172 static void rcu_preempt_note_context_switch(int cpu)
173 {
174         struct task_struct *t = current;
175         unsigned long flags;
176         struct rcu_data *rdp;
177         struct rcu_node *rnp;
178
179         if (t->rcu_read_lock_nesting > 0 &&
180             !t->rcu_read_unlock_special.b.blocked) {
181
182                 /* Possibly blocking in an RCU read-side critical section. */
183                 rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu);
184                 rnp = rdp->mynode;
185                 raw_spin_lock_irqsave(&rnp->lock, flags);
186                 smp_mb__after_unlock_lock();
187                 t->rcu_read_unlock_special.b.blocked = true;
188                 t->rcu_blocked_node = rnp;
189
190                 /*
191                  * If this CPU has already checked in, then this task
192                  * will hold up the next grace period rather than the
193                  * current grace period.  Queue the task accordingly.
194                  * If the task is queued for the current grace period
195                  * (i.e., this CPU has not yet passed through a quiescent
196                  * state for the current grace period), then as long
197                  * as that task remains queued, the current grace period
198                  * cannot end.  Note that there is some uncertainty as
199                  * to exactly when the current grace period started.
200                  * We take a conservative approach, which can result
201                  * in unnecessarily waiting on tasks that started very
202                  * slightly after the current grace period began.  C'est
203                  * la vie!!!
204                  *
205                  * But first, note that the current CPU must still be
206                  * on line!
207                  */
208                 WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
209                 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
210                 if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
211                         list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
212                         rnp->gp_tasks = &t->rcu_node_entry;
213 #ifdef CONFIG_RCU_BOOST
214                         if (rnp->boost_tasks != NULL)
215                                 rnp->boost_tasks = rnp->gp_tasks;
216 #endif /* #ifdef CONFIG_RCU_BOOST */
217                 } else {
218                         list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
219                         if (rnp->qsmask & rdp->grpmask)
220                                 rnp->gp_tasks = &t->rcu_node_entry;
221                 }
222                 trace_rcu_preempt_task(rdp->rsp->name,
223                                        t->pid,
224                                        (rnp->qsmask & rdp->grpmask)
225                                        ? rnp->gpnum
226                                        : rnp->gpnum + 1);
227                 raw_spin_unlock_irqrestore(&rnp->lock, flags);
228         } else if (t->rcu_read_lock_nesting < 0 &&
229                    t->rcu_read_unlock_special.s) {
230
231                 /*
232                  * Complete exit from RCU read-side critical section on
233                  * behalf of preempted instance of __rcu_read_unlock().
234                  */
235                 rcu_read_unlock_special(t);
236         }
237
238         /*
239          * Either we were not in an RCU read-side critical section to
240          * begin with, or we have now recorded that critical section
241          * globally.  Either way, we can now note a quiescent state
242          * for this CPU.  Again, if we were in an RCU read-side critical
243          * section, and if that critical section was blocking the current
244          * grace period, then the fact that the task has been enqueued
245          * means that we continue to block the current grace period.
246          */
247         rcu_preempt_qs();
248 }
249
250 /*
251  * Check for preempted RCU readers blocking the current grace period
252  * for the specified rcu_node structure.  If the caller needs a reliable
253  * answer, it must hold the rcu_node's ->lock.
254  */
255 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
256 {
257         return rnp->gp_tasks != NULL;
258 }
259
260 /*
261  * Record a quiescent state for all tasks that were previously queued
262  * on the specified rcu_node structure and that were blocking the current
263  * RCU grace period.  The caller must hold the specified rnp->lock with
264  * irqs disabled, and this lock is released upon return, but irqs remain
265  * disabled.
266  */
267 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
268         __releases(rnp->lock)
269 {
270         unsigned long mask;
271         struct rcu_node *rnp_p;
272
273         if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
274                 raw_spin_unlock_irqrestore(&rnp->lock, flags);
275                 return;  /* Still need more quiescent states! */
276         }
277
278         rnp_p = rnp->parent;
279         if (rnp_p == NULL) {
280                 /*
281                  * Either there is only one rcu_node in the tree,
282                  * or tasks were kicked up to root rcu_node due to
283                  * CPUs going offline.
284                  */
285                 rcu_report_qs_rsp(&rcu_preempt_state, flags);
286                 return;
287         }
288
289         /* Report up the rest of the hierarchy. */
290         mask = rnp->grpmask;
291         raw_spin_unlock(&rnp->lock);    /* irqs remain disabled. */
292         raw_spin_lock(&rnp_p->lock);    /* irqs already disabled. */
293         smp_mb__after_unlock_lock();
294         rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
295 }
296
297 /*
298  * Advance a ->blkd_tasks-list pointer to the next entry, instead
299  * returning NULL if at the end of the list.
300  */
301 static struct list_head *rcu_next_node_entry(struct task_struct *t,
302                                              struct rcu_node *rnp)
303 {
304         struct list_head *np;
305
306         np = t->rcu_node_entry.next;
307         if (np == &rnp->blkd_tasks)
308                 np = NULL;
309         return np;
310 }
311
312 /*
313  * Handle special cases during rcu_read_unlock(), such as needing to
314  * notify RCU core processing or task having blocked during the RCU
315  * read-side critical section.
316  */
317 void rcu_read_unlock_special(struct task_struct *t)
318 {
319         int empty;
320         int empty_exp;
321         int empty_exp_now;
322         unsigned long flags;
323         struct list_head *np;
324 #ifdef CONFIG_RCU_BOOST
325         bool drop_boost_mutex = false;
326 #endif /* #ifdef CONFIG_RCU_BOOST */
327         struct rcu_node *rnp;
328         union rcu_special special;
329
330         /* NMI handlers cannot block and cannot safely manipulate state. */
331         if (in_nmi())
332                 return;
333
334         local_irq_save(flags);
335
336         /*
337          * If RCU core is waiting for this CPU to exit critical section,
338          * let it know that we have done so.  Because irqs are disabled,
339          * t->rcu_read_unlock_special cannot change.
340          */
341         special = t->rcu_read_unlock_special;
342         if (special.b.need_qs) {
343                 rcu_preempt_qs();
344                 if (!t->rcu_read_unlock_special.s) {
345                         local_irq_restore(flags);
346                         return;
347                 }
348         }
349
350         /* Hardware IRQ handlers cannot block, complain if they get here. */
351         if (WARN_ON_ONCE(in_irq() || in_serving_softirq())) {
352                 local_irq_restore(flags);
353                 return;
354         }
355
356         /* Clean up if blocked during RCU read-side critical section. */
357         if (special.b.blocked) {
358                 t->rcu_read_unlock_special.b.blocked = false;
359
360                 /*
361                  * Remove this task from the list it blocked on.  The
362                  * task can migrate while we acquire the lock, but at
363                  * most one time.  So at most two passes through loop.
364                  */
365                 for (;;) {
366                         rnp = t->rcu_blocked_node;
367                         raw_spin_lock(&rnp->lock);  /* irqs already disabled. */
368                         smp_mb__after_unlock_lock();
369                         if (rnp == t->rcu_blocked_node)
370                                 break;
371                         raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
372                 }
373                 empty = !rcu_preempt_blocked_readers_cgp(rnp);
374                 empty_exp = !rcu_preempted_readers_exp(rnp);
375                 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
376                 np = rcu_next_node_entry(t, rnp);
377                 list_del_init(&t->rcu_node_entry);
378                 t->rcu_blocked_node = NULL;
379                 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
380                                                 rnp->gpnum, t->pid);
381                 if (&t->rcu_node_entry == rnp->gp_tasks)
382                         rnp->gp_tasks = np;
383                 if (&t->rcu_node_entry == rnp->exp_tasks)
384                         rnp->exp_tasks = np;
385 #ifdef CONFIG_RCU_BOOST
386                 if (&t->rcu_node_entry == rnp->boost_tasks)
387                         rnp->boost_tasks = np;
388                 /* Snapshot ->boost_mtx ownership with rcu_node lock held. */
389                 drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
390 #endif /* #ifdef CONFIG_RCU_BOOST */
391
392                 /*
393                  * If this was the last task on the current list, and if
394                  * we aren't waiting on any CPUs, report the quiescent state.
395                  * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
396                  * so we must take a snapshot of the expedited state.
397                  */
398                 empty_exp_now = !rcu_preempted_readers_exp(rnp);
399                 if (!empty && !rcu_preempt_blocked_readers_cgp(rnp)) {
400                         trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
401                                                          rnp->gpnum,
402                                                          0, rnp->qsmask,
403                                                          rnp->level,
404                                                          rnp->grplo,
405                                                          rnp->grphi,
406                                                          !!rnp->gp_tasks);
407                         rcu_report_unblock_qs_rnp(rnp, flags);
408                 } else {
409                         raw_spin_unlock_irqrestore(&rnp->lock, flags);
410                 }
411
412 #ifdef CONFIG_RCU_BOOST
413                 /* Unboost if we were boosted. */
414                 if (drop_boost_mutex) {
415                         rt_mutex_unlock(&rnp->boost_mtx);
416                         complete(&rnp->boost_completion);
417                 }
418 #endif /* #ifdef CONFIG_RCU_BOOST */
419
420                 /*
421                  * If this was the last task on the expedited lists,
422                  * then we need to report up the rcu_node hierarchy.
423                  */
424                 if (!empty_exp && empty_exp_now)
425                         rcu_report_exp_rnp(&rcu_preempt_state, rnp, true);
426         } else {
427                 local_irq_restore(flags);
428         }
429 }
430
431 #ifdef CONFIG_RCU_CPU_STALL_VERBOSE
432
433 /*
434  * Dump detailed information for all tasks blocking the current RCU
435  * grace period on the specified rcu_node structure.
436  */
437 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
438 {
439         unsigned long flags;
440         struct task_struct *t;
441
442         raw_spin_lock_irqsave(&rnp->lock, flags);
443         if (!rcu_preempt_blocked_readers_cgp(rnp)) {
444                 raw_spin_unlock_irqrestore(&rnp->lock, flags);
445                 return;
446         }
447         t = list_entry(rnp->gp_tasks,
448                        struct task_struct, rcu_node_entry);
449         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
450                 sched_show_task(t);
451         raw_spin_unlock_irqrestore(&rnp->lock, flags);
452 }
453
454 /*
455  * Dump detailed information for all tasks blocking the current RCU
456  * grace period.
457  */
458 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
459 {
460         struct rcu_node *rnp = rcu_get_root(rsp);
461
462         rcu_print_detail_task_stall_rnp(rnp);
463         rcu_for_each_leaf_node(rsp, rnp)
464                 rcu_print_detail_task_stall_rnp(rnp);
465 }
466
467 #else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
468
469 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
470 {
471 }
472
473 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
474
475 #ifdef CONFIG_RCU_CPU_STALL_INFO
476
477 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
478 {
479         pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
480                rnp->level, rnp->grplo, rnp->grphi);
481 }
482
483 static void rcu_print_task_stall_end(void)
484 {
485         pr_cont("\n");
486 }
487
488 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
489
490 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
491 {
492 }
493
494 static void rcu_print_task_stall_end(void)
495 {
496 }
497
498 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
499
500 /*
501  * Scan the current list of tasks blocked within RCU read-side critical
502  * sections, printing out the tid of each.
503  */
504 static int rcu_print_task_stall(struct rcu_node *rnp)
505 {
506         struct task_struct *t;
507         int ndetected = 0;
508
509         if (!rcu_preempt_blocked_readers_cgp(rnp))
510                 return 0;
511         rcu_print_task_stall_begin(rnp);
512         t = list_entry(rnp->gp_tasks,
513                        struct task_struct, rcu_node_entry);
514         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
515                 pr_cont(" P%d", t->pid);
516                 ndetected++;
517         }
518         rcu_print_task_stall_end();
519         return ndetected;
520 }
521
522 /*
523  * Check that the list of blocked tasks for the newly completed grace
524  * period is in fact empty.  It is a serious bug to complete a grace
525  * period that still has RCU readers blocked!  This function must be
526  * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
527  * must be held by the caller.
528  *
529  * Also, if there are blocked tasks on the list, they automatically
530  * block the newly created grace period, so set up ->gp_tasks accordingly.
531  */
532 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
533 {
534         WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
535         if (!list_empty(&rnp->blkd_tasks))
536                 rnp->gp_tasks = rnp->blkd_tasks.next;
537         WARN_ON_ONCE(rnp->qsmask);
538 }
539
540 #ifdef CONFIG_HOTPLUG_CPU
541
542 /*
543  * Handle tasklist migration for case in which all CPUs covered by the
544  * specified rcu_node have gone offline.  Move them up to the root
545  * rcu_node.  The reason for not just moving them to the immediate
546  * parent is to remove the need for rcu_read_unlock_special() to
547  * make more than two attempts to acquire the target rcu_node's lock.
548  * Returns true if there were tasks blocking the current RCU grace
549  * period.
550  *
551  * Returns 1 if there was previously a task blocking the current grace
552  * period on the specified rcu_node structure.
553  *
554  * The caller must hold rnp->lock with irqs disabled.
555  */
556 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
557                                      struct rcu_node *rnp,
558                                      struct rcu_data *rdp)
559 {
560         struct list_head *lp;
561         struct list_head *lp_root;
562         int retval = 0;
563         struct rcu_node *rnp_root = rcu_get_root(rsp);
564         struct task_struct *t;
565
566         if (rnp == rnp_root) {
567                 WARN_ONCE(1, "Last CPU thought to be offlined?");
568                 return 0;  /* Shouldn't happen: at least one CPU online. */
569         }
570
571         /* If we are on an internal node, complain bitterly. */
572         WARN_ON_ONCE(rnp != rdp->mynode);
573
574         /*
575          * Move tasks up to root rcu_node.  Don't try to get fancy for
576          * this corner-case operation -- just put this node's tasks
577          * at the head of the root node's list, and update the root node's
578          * ->gp_tasks and ->exp_tasks pointers to those of this node's,
579          * if non-NULL.  This might result in waiting for more tasks than
580          * absolutely necessary, but this is a good performance/complexity
581          * tradeoff.
582          */
583         if (rcu_preempt_blocked_readers_cgp(rnp) && rnp->qsmask == 0)
584                 retval |= RCU_OFL_TASKS_NORM_GP;
585         if (rcu_preempted_readers_exp(rnp))
586                 retval |= RCU_OFL_TASKS_EXP_GP;
587         lp = &rnp->blkd_tasks;
588         lp_root = &rnp_root->blkd_tasks;
589         while (!list_empty(lp)) {
590                 t = list_entry(lp->next, typeof(*t), rcu_node_entry);
591                 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
592                 smp_mb__after_unlock_lock();
593                 list_del(&t->rcu_node_entry);
594                 t->rcu_blocked_node = rnp_root;
595                 list_add(&t->rcu_node_entry, lp_root);
596                 if (&t->rcu_node_entry == rnp->gp_tasks)
597                         rnp_root->gp_tasks = rnp->gp_tasks;
598                 if (&t->rcu_node_entry == rnp->exp_tasks)
599                         rnp_root->exp_tasks = rnp->exp_tasks;
600 #ifdef CONFIG_RCU_BOOST
601                 if (&t->rcu_node_entry == rnp->boost_tasks)
602                         rnp_root->boost_tasks = rnp->boost_tasks;
603 #endif /* #ifdef CONFIG_RCU_BOOST */
604                 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
605         }
606
607         rnp->gp_tasks = NULL;
608         rnp->exp_tasks = NULL;
609 #ifdef CONFIG_RCU_BOOST
610         rnp->boost_tasks = NULL;
611         /*
612          * In case root is being boosted and leaf was not.  Make sure
613          * that we boost the tasks blocking the current grace period
614          * in this case.
615          */
616         raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
617         smp_mb__after_unlock_lock();
618         if (rnp_root->boost_tasks != NULL &&
619             rnp_root->boost_tasks != rnp_root->gp_tasks &&
620             rnp_root->boost_tasks != rnp_root->exp_tasks)
621                 rnp_root->boost_tasks = rnp_root->gp_tasks;
622         raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
623 #endif /* #ifdef CONFIG_RCU_BOOST */
624
625         return retval;
626 }
627
628 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
629
630 /*
631  * Check for a quiescent state from the current CPU.  When a task blocks,
632  * the task is recorded in the corresponding CPU's rcu_node structure,
633  * which is checked elsewhere.
634  *
635  * Caller must disable hard irqs.
636  */
637 static void rcu_preempt_check_callbacks(int cpu)
638 {
639         struct task_struct *t = current;
640
641         if (t->rcu_read_lock_nesting == 0) {
642                 rcu_preempt_qs();
643                 return;
644         }
645         if (t->rcu_read_lock_nesting > 0 &&
646             per_cpu(rcu_preempt_data, cpu).qs_pending &&
647             !per_cpu(rcu_preempt_data, cpu).passed_quiesce)
648                 t->rcu_read_unlock_special.b.need_qs = true;
649 }
650
651 #ifdef CONFIG_RCU_BOOST
652
653 static void rcu_preempt_do_callbacks(void)
654 {
655         rcu_do_batch(&rcu_preempt_state, this_cpu_ptr(&rcu_preempt_data));
656 }
657
658 #endif /* #ifdef CONFIG_RCU_BOOST */
659
660 /*
661  * Queue a preemptible-RCU callback for invocation after a grace period.
662  */
663 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
664 {
665         __call_rcu(head, func, &rcu_preempt_state, -1, 0);
666 }
667 EXPORT_SYMBOL_GPL(call_rcu);
668
669 /**
670  * synchronize_rcu - wait until a grace period has elapsed.
671  *
672  * Control will return to the caller some time after a full grace
673  * period has elapsed, in other words after all currently executing RCU
674  * read-side critical sections have completed.  Note, however, that
675  * upon return from synchronize_rcu(), the caller might well be executing
676  * concurrently with new RCU read-side critical sections that began while
677  * synchronize_rcu() was waiting.  RCU read-side critical sections are
678  * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
679  *
680  * See the description of synchronize_sched() for more detailed information
681  * on memory ordering guarantees.
682  */
683 void synchronize_rcu(void)
684 {
685         rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
686                            !lock_is_held(&rcu_lock_map) &&
687                            !lock_is_held(&rcu_sched_lock_map),
688                            "Illegal synchronize_rcu() in RCU read-side critical section");
689         if (!rcu_scheduler_active)
690                 return;
691         if (rcu_expedited)
692                 synchronize_rcu_expedited();
693         else
694                 wait_rcu_gp(call_rcu);
695 }
696 EXPORT_SYMBOL_GPL(synchronize_rcu);
697
698 static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
699 static unsigned long sync_rcu_preempt_exp_count;
700 static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
701
702 /*
703  * Return non-zero if there are any tasks in RCU read-side critical
704  * sections blocking the current preemptible-RCU expedited grace period.
705  * If there is no preemptible-RCU expedited grace period currently in
706  * progress, returns zero unconditionally.
707  */
708 static int rcu_preempted_readers_exp(struct rcu_node *rnp)
709 {
710         return rnp->exp_tasks != NULL;
711 }
712
713 /*
714  * return non-zero if there is no RCU expedited grace period in progress
715  * for the specified rcu_node structure, in other words, if all CPUs and
716  * tasks covered by the specified rcu_node structure have done their bit
717  * for the current expedited grace period.  Works only for preemptible
718  * RCU -- other RCU implementation use other means.
719  *
720  * Caller must hold sync_rcu_preempt_exp_mutex.
721  */
722 static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
723 {
724         return !rcu_preempted_readers_exp(rnp) &&
725                ACCESS_ONCE(rnp->expmask) == 0;
726 }
727
728 /*
729  * Report the exit from RCU read-side critical section for the last task
730  * that queued itself during or before the current expedited preemptible-RCU
731  * grace period.  This event is reported either to the rcu_node structure on
732  * which the task was queued or to one of that rcu_node structure's ancestors,
733  * recursively up the tree.  (Calm down, calm down, we do the recursion
734  * iteratively!)
735  *
736  * Most callers will set the "wake" flag, but the task initiating the
737  * expedited grace period need not wake itself.
738  *
739  * Caller must hold sync_rcu_preempt_exp_mutex.
740  */
741 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
742                                bool wake)
743 {
744         unsigned long flags;
745         unsigned long mask;
746
747         raw_spin_lock_irqsave(&rnp->lock, flags);
748         smp_mb__after_unlock_lock();
749         for (;;) {
750                 if (!sync_rcu_preempt_exp_done(rnp)) {
751                         raw_spin_unlock_irqrestore(&rnp->lock, flags);
752                         break;
753                 }
754                 if (rnp->parent == NULL) {
755                         raw_spin_unlock_irqrestore(&rnp->lock, flags);
756                         if (wake) {
757                                 smp_mb(); /* EGP done before wake_up(). */
758                                 wake_up(&sync_rcu_preempt_exp_wq);
759                         }
760                         break;
761                 }
762                 mask = rnp->grpmask;
763                 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
764                 rnp = rnp->parent;
765                 raw_spin_lock(&rnp->lock); /* irqs already disabled */
766                 smp_mb__after_unlock_lock();
767                 rnp->expmask &= ~mask;
768         }
769 }
770
771 /*
772  * Snapshot the tasks blocking the newly started preemptible-RCU expedited
773  * grace period for the specified rcu_node structure.  If there are no such
774  * tasks, report it up the rcu_node hierarchy.
775  *
776  * Caller must hold sync_rcu_preempt_exp_mutex and must exclude
777  * CPU hotplug operations.
778  */
779 static void
780 sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
781 {
782         unsigned long flags;
783         int must_wait = 0;
784
785         raw_spin_lock_irqsave(&rnp->lock, flags);
786         smp_mb__after_unlock_lock();
787         if (list_empty(&rnp->blkd_tasks)) {
788                 raw_spin_unlock_irqrestore(&rnp->lock, flags);
789         } else {
790                 rnp->exp_tasks = rnp->blkd_tasks.next;
791                 rcu_initiate_boost(rnp, flags);  /* releases rnp->lock */
792                 must_wait = 1;
793         }
794         if (!must_wait)
795                 rcu_report_exp_rnp(rsp, rnp, false); /* Don't wake self. */
796 }
797
798 /**
799  * synchronize_rcu_expedited - Brute-force RCU grace period
800  *
801  * Wait for an RCU-preempt grace period, but expedite it.  The basic
802  * idea is to invoke synchronize_sched_expedited() to push all the tasks to
803  * the ->blkd_tasks lists and wait for this list to drain.  This consumes
804  * significant time on all CPUs and is unfriendly to real-time workloads,
805  * so is thus not recommended for any sort of common-case code.
806  * In fact, if you are using synchronize_rcu_expedited() in a loop,
807  * please restructure your code to batch your updates, and then Use a
808  * single synchronize_rcu() instead.
809  */
810 void synchronize_rcu_expedited(void)
811 {
812         unsigned long flags;
813         struct rcu_node *rnp;
814         struct rcu_state *rsp = &rcu_preempt_state;
815         unsigned long snap;
816         int trycount = 0;
817
818         smp_mb(); /* Caller's modifications seen first by other CPUs. */
819         snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
820         smp_mb(); /* Above access cannot bleed into critical section. */
821
822         /*
823          * Block CPU-hotplug operations.  This means that any CPU-hotplug
824          * operation that finds an rcu_node structure with tasks in the
825          * process of being boosted will know that all tasks blocking
826          * this expedited grace period will already be in the process of
827          * being boosted.  This simplifies the process of moving tasks
828          * from leaf to root rcu_node structures.
829          */
830         if (!try_get_online_cpus()) {
831                 /* CPU-hotplug operation in flight, fall back to normal GP. */
832                 wait_rcu_gp(call_rcu);
833                 return;
834         }
835
836         /*
837          * Acquire lock, falling back to synchronize_rcu() if too many
838          * lock-acquisition failures.  Of course, if someone does the
839          * expedited grace period for us, just leave.
840          */
841         while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
842                 if (ULONG_CMP_LT(snap,
843                     ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
844                         put_online_cpus();
845                         goto mb_ret; /* Others did our work for us. */
846                 }
847                 if (trycount++ < 10) {
848                         udelay(trycount * num_online_cpus());
849                 } else {
850                         put_online_cpus();
851                         wait_rcu_gp(call_rcu);
852                         return;
853                 }
854         }
855         if (ULONG_CMP_LT(snap, ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
856                 put_online_cpus();
857                 goto unlock_mb_ret; /* Others did our work for us. */
858         }
859
860         /* force all RCU readers onto ->blkd_tasks lists. */
861         synchronize_sched_expedited();
862
863         /* Initialize ->expmask for all non-leaf rcu_node structures. */
864         rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
865                 raw_spin_lock_irqsave(&rnp->lock, flags);
866                 smp_mb__after_unlock_lock();
867                 rnp->expmask = rnp->qsmaskinit;
868                 raw_spin_unlock_irqrestore(&rnp->lock, flags);
869         }
870
871         /* Snapshot current state of ->blkd_tasks lists. */
872         rcu_for_each_leaf_node(rsp, rnp)
873                 sync_rcu_preempt_exp_init(rsp, rnp);
874         if (NUM_RCU_NODES > 1)
875                 sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp));
876
877         put_online_cpus();
878
879         /* Wait for snapshotted ->blkd_tasks lists to drain. */
880         rnp = rcu_get_root(rsp);
881         wait_event(sync_rcu_preempt_exp_wq,
882                    sync_rcu_preempt_exp_done(rnp));
883
884         /* Clean up and exit. */
885         smp_mb(); /* ensure expedited GP seen before counter increment. */
886         ACCESS_ONCE(sync_rcu_preempt_exp_count) =
887                                         sync_rcu_preempt_exp_count + 1;
888 unlock_mb_ret:
889         mutex_unlock(&sync_rcu_preempt_exp_mutex);
890 mb_ret:
891         smp_mb(); /* ensure subsequent action seen after grace period. */
892 }
893 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
894
895 /**
896  * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
897  *
898  * Note that this primitive does not necessarily wait for an RCU grace period
899  * to complete.  For example, if there are no RCU callbacks queued anywhere
900  * in the system, then rcu_barrier() is within its rights to return
901  * immediately, without waiting for anything, much less an RCU grace period.
902  */
903 void rcu_barrier(void)
904 {
905         _rcu_barrier(&rcu_preempt_state);
906 }
907 EXPORT_SYMBOL_GPL(rcu_barrier);
908
909 /*
910  * Initialize preemptible RCU's state structures.
911  */
912 static void __init __rcu_init_preempt(void)
913 {
914         rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
915 }
916
917 /*
918  * Check for a task exiting while in a preemptible-RCU read-side
919  * critical section, clean up if so.  No need to issue warnings,
920  * as debug_check_no_locks_held() already does this if lockdep
921  * is enabled.
922  */
923 void exit_rcu(void)
924 {
925         struct task_struct *t = current;
926
927         if (likely(list_empty(&current->rcu_node_entry)))
928                 return;
929         t->rcu_read_lock_nesting = 1;
930         barrier();
931         t->rcu_read_unlock_special.b.blocked = true;
932         __rcu_read_unlock();
933 }
934
935 #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
936
937 static struct rcu_state *rcu_state_p = &rcu_sched_state;
938
939 /*
940  * Tell them what RCU they are running.
941  */
942 static void __init rcu_bootup_announce(void)
943 {
944         pr_info("Hierarchical RCU implementation.\n");
945         rcu_bootup_announce_oddness();
946 }
947
948 /*
949  * Return the number of RCU batches processed thus far for debug & stats.
950  */
951 long rcu_batches_completed(void)
952 {
953         return rcu_batches_completed_sched();
954 }
955 EXPORT_SYMBOL_GPL(rcu_batches_completed);
956
957 /*
958  * Because preemptible RCU does not exist, we never have to check for
959  * CPUs being in quiescent states.
960  */
961 static void rcu_preempt_note_context_switch(int cpu)
962 {
963 }
964
965 /*
966  * Because preemptible RCU does not exist, there are never any preempted
967  * RCU readers.
968  */
969 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
970 {
971         return 0;
972 }
973
974 #ifdef CONFIG_HOTPLUG_CPU
975
976 /* Because preemptible RCU does not exist, no quieting of tasks. */
977 static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
978         __releases(rnp->lock)
979 {
980         raw_spin_unlock_irqrestore(&rnp->lock, flags);
981 }
982
983 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
984
985 /*
986  * Because preemptible RCU does not exist, we never have to check for
987  * tasks blocked within RCU read-side critical sections.
988  */
989 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
990 {
991 }
992
993 /*
994  * Because preemptible RCU does not exist, we never have to check for
995  * tasks blocked within RCU read-side critical sections.
996  */
997 static int rcu_print_task_stall(struct rcu_node *rnp)
998 {
999         return 0;
1000 }
1001
1002 /*
1003  * Because there is no preemptible RCU, there can be no readers blocked,
1004  * so there is no need to check for blocked tasks.  So check only for
1005  * bogus qsmask values.
1006  */
1007 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
1008 {
1009         WARN_ON_ONCE(rnp->qsmask);
1010 }
1011
1012 #ifdef CONFIG_HOTPLUG_CPU
1013
1014 /*
1015  * Because preemptible RCU does not exist, it never needs to migrate
1016  * tasks that were blocked within RCU read-side critical sections, and
1017  * such non-existent tasks cannot possibly have been blocking the current
1018  * grace period.
1019  */
1020 static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
1021                                      struct rcu_node *rnp,
1022                                      struct rcu_data *rdp)
1023 {
1024         return 0;
1025 }
1026
1027 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1028
1029 /*
1030  * Because preemptible RCU does not exist, it never has any callbacks
1031  * to check.
1032  */
1033 static void rcu_preempt_check_callbacks(int cpu)
1034 {
1035 }
1036
1037 /*
1038  * Wait for an rcu-preempt grace period, but make it happen quickly.
1039  * But because preemptible RCU does not exist, map to rcu-sched.
1040  */
1041 void synchronize_rcu_expedited(void)
1042 {
1043         synchronize_sched_expedited();
1044 }
1045 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
1046
1047 #ifdef CONFIG_HOTPLUG_CPU
1048
1049 /*
1050  * Because preemptible RCU does not exist, there is never any need to
1051  * report on tasks preempted in RCU read-side critical sections during
1052  * expedited RCU grace periods.
1053  */
1054 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
1055                                bool wake)
1056 {
1057 }
1058
1059 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
1060
1061 /*
1062  * Because preemptible RCU does not exist, rcu_barrier() is just
1063  * another name for rcu_barrier_sched().
1064  */
1065 void rcu_barrier(void)
1066 {
1067         rcu_barrier_sched();
1068 }
1069 EXPORT_SYMBOL_GPL(rcu_barrier);
1070
1071 /*
1072  * Because preemptible RCU does not exist, it need not be initialized.
1073  */
1074 static void __init __rcu_init_preempt(void)
1075 {
1076 }
1077
1078 /*
1079  * Because preemptible RCU does not exist, tasks cannot possibly exit
1080  * while in preemptible RCU read-side critical sections.
1081  */
1082 void exit_rcu(void)
1083 {
1084 }
1085
1086 #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1087
1088 #ifdef CONFIG_RCU_BOOST
1089
1090 #include "../locking/rtmutex_common.h"
1091
1092 #ifdef CONFIG_RCU_TRACE
1093
1094 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1095 {
1096         if (list_empty(&rnp->blkd_tasks))
1097                 rnp->n_balk_blkd_tasks++;
1098         else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
1099                 rnp->n_balk_exp_gp_tasks++;
1100         else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
1101                 rnp->n_balk_boost_tasks++;
1102         else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
1103                 rnp->n_balk_notblocked++;
1104         else if (rnp->gp_tasks != NULL &&
1105                  ULONG_CMP_LT(jiffies, rnp->boost_time))
1106                 rnp->n_balk_notyet++;
1107         else
1108                 rnp->n_balk_nos++;
1109 }
1110
1111 #else /* #ifdef CONFIG_RCU_TRACE */
1112
1113 static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1114 {
1115 }
1116
1117 #endif /* #else #ifdef CONFIG_RCU_TRACE */
1118
1119 static void rcu_wake_cond(struct task_struct *t, int status)
1120 {
1121         /*
1122          * If the thread is yielding, only wake it when this
1123          * is invoked from idle
1124          */
1125         if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
1126                 wake_up_process(t);
1127 }
1128
1129 /*
1130  * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1131  * or ->boost_tasks, advancing the pointer to the next task in the
1132  * ->blkd_tasks list.
1133  *
1134  * Note that irqs must be enabled: boosting the task can block.
1135  * Returns 1 if there are more tasks needing to be boosted.
1136  */
1137 static int rcu_boost(struct rcu_node *rnp)
1138 {
1139         unsigned long flags;
1140         struct task_struct *t;
1141         struct list_head *tb;
1142
1143         if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL)
1144                 return 0;  /* Nothing left to boost. */
1145
1146         raw_spin_lock_irqsave(&rnp->lock, flags);
1147         smp_mb__after_unlock_lock();
1148
1149         /*
1150          * Recheck under the lock: all tasks in need of boosting
1151          * might exit their RCU read-side critical sections on their own.
1152          */
1153         if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1154                 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1155                 return 0;
1156         }
1157
1158         /*
1159          * Preferentially boost tasks blocking expedited grace periods.
1160          * This cannot starve the normal grace periods because a second
1161          * expedited grace period must boost all blocked tasks, including
1162          * those blocking the pre-existing normal grace period.
1163          */
1164         if (rnp->exp_tasks != NULL) {
1165                 tb = rnp->exp_tasks;
1166                 rnp->n_exp_boosts++;
1167         } else {
1168                 tb = rnp->boost_tasks;
1169                 rnp->n_normal_boosts++;
1170         }
1171         rnp->n_tasks_boosted++;
1172
1173         /*
1174          * We boost task t by manufacturing an rt_mutex that appears to
1175          * be held by task t.  We leave a pointer to that rt_mutex where
1176          * task t can find it, and task t will release the mutex when it
1177          * exits its outermost RCU read-side critical section.  Then
1178          * simply acquiring this artificial rt_mutex will boost task
1179          * t's priority.  (Thanks to tglx for suggesting this approach!)
1180          *
1181          * Note that task t must acquire rnp->lock to remove itself from
1182          * the ->blkd_tasks list, which it will do from exit() if from
1183          * nowhere else.  We therefore are guaranteed that task t will
1184          * stay around at least until we drop rnp->lock.  Note that
1185          * rnp->lock also resolves races between our priority boosting
1186          * and task t's exiting its outermost RCU read-side critical
1187          * section.
1188          */
1189         t = container_of(tb, struct task_struct, rcu_node_entry);
1190         rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1191         init_completion(&rnp->boost_completion);
1192         raw_spin_unlock_irqrestore(&rnp->lock, flags);
1193         /* Lock only for side effect: boosts task t's priority. */
1194         rt_mutex_lock(&rnp->boost_mtx);
1195         rt_mutex_unlock(&rnp->boost_mtx);  /* Then keep lockdep happy. */
1196
1197         /* Wait for boostee to be done w/boost_mtx before reinitializing. */
1198         wait_for_completion(&rnp->boost_completion);
1199
1200         return ACCESS_ONCE(rnp->exp_tasks) != NULL ||
1201                ACCESS_ONCE(rnp->boost_tasks) != NULL;
1202 }
1203
1204 /*
1205  * Priority-boosting kthread.  One per leaf rcu_node and one for the
1206  * root rcu_node.
1207  */
1208 static int rcu_boost_kthread(void *arg)
1209 {
1210         struct rcu_node *rnp = (struct rcu_node *)arg;
1211         int spincnt = 0;
1212         int more2boost;
1213
1214         trace_rcu_utilization(TPS("Start boost kthread@init"));
1215         for (;;) {
1216                 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1217                 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1218                 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1219                 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1220                 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1221                 more2boost = rcu_boost(rnp);
1222                 if (more2boost)
1223                         spincnt++;
1224                 else
1225                         spincnt = 0;
1226                 if (spincnt > 10) {
1227                         rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1228                         trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1229                         schedule_timeout_interruptible(2);
1230                         trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1231                         spincnt = 0;
1232                 }
1233         }
1234         /* NOTREACHED */
1235         trace_rcu_utilization(TPS("End boost kthread@notreached"));
1236         return 0;
1237 }
1238
1239 /*
1240  * Check to see if it is time to start boosting RCU readers that are
1241  * blocking the current grace period, and, if so, tell the per-rcu_node
1242  * kthread to start boosting them.  If there is an expedited grace
1243  * period in progress, it is always time to boost.
1244  *
1245  * The caller must hold rnp->lock, which this function releases.
1246  * The ->boost_kthread_task is immortal, so we don't need to worry
1247  * about it going away.
1248  */
1249 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1250         __releases(rnp->lock)
1251 {
1252         struct task_struct *t;
1253
1254         if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1255                 rnp->n_balk_exp_gp_tasks++;
1256                 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1257                 return;
1258         }
1259         if (rnp->exp_tasks != NULL ||
1260             (rnp->gp_tasks != NULL &&
1261              rnp->boost_tasks == NULL &&
1262              rnp->qsmask == 0 &&
1263              ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1264                 if (rnp->exp_tasks == NULL)
1265                         rnp->boost_tasks = rnp->gp_tasks;
1266                 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1267                 t = rnp->boost_kthread_task;
1268                 if (t)
1269                         rcu_wake_cond(t, rnp->boost_kthread_status);
1270         } else {
1271                 rcu_initiate_boost_trace(rnp);
1272                 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1273         }
1274 }
1275
1276 /*
1277  * Wake up the per-CPU kthread to invoke RCU callbacks.
1278  */
1279 static void invoke_rcu_callbacks_kthread(void)
1280 {
1281         unsigned long flags;
1282
1283         local_irq_save(flags);
1284         __this_cpu_write(rcu_cpu_has_work, 1);
1285         if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1286             current != __this_cpu_read(rcu_cpu_kthread_task)) {
1287                 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1288                               __this_cpu_read(rcu_cpu_kthread_status));
1289         }
1290         local_irq_restore(flags);
1291 }
1292
1293 /*
1294  * Is the current CPU running the RCU-callbacks kthread?
1295  * Caller must have preemption disabled.
1296  */
1297 static bool rcu_is_callbacks_kthread(void)
1298 {
1299         return __this_cpu_read(rcu_cpu_kthread_task) == current;
1300 }
1301
1302 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1303
1304 /*
1305  * Do priority-boost accounting for the start of a new grace period.
1306  */
1307 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1308 {
1309         rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1310 }
1311
1312 /*
1313  * Create an RCU-boost kthread for the specified node if one does not
1314  * already exist.  We only create this kthread for preemptible RCU.
1315  * Returns zero if all is well, a negated errno otherwise.
1316  */
1317 static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1318                                                  struct rcu_node *rnp)
1319 {
1320         int rnp_index = rnp - &rsp->node[0];
1321         unsigned long flags;
1322         struct sched_param sp;
1323         struct task_struct *t;
1324
1325         if (&rcu_preempt_state != rsp)
1326                 return 0;
1327
1328         if (!rcu_scheduler_fully_active || rnp->qsmaskinit == 0)
1329                 return 0;
1330
1331         rsp->boost = 1;
1332         if (rnp->boost_kthread_task != NULL)
1333                 return 0;
1334         t = kthread_create(rcu_boost_kthread, (void *)rnp,
1335                            "rcub/%d", rnp_index);
1336         if (IS_ERR(t))
1337                 return PTR_ERR(t);
1338         raw_spin_lock_irqsave(&rnp->lock, flags);
1339         smp_mb__after_unlock_lock();
1340         rnp->boost_kthread_task = t;
1341         raw_spin_unlock_irqrestore(&rnp->lock, flags);
1342         sp.sched_priority = RCU_BOOST_PRIO;
1343         sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1344         wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1345         return 0;
1346 }
1347
1348 static void rcu_kthread_do_work(void)
1349 {
1350         rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1351         rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1352         rcu_preempt_do_callbacks();
1353 }
1354
1355 static void rcu_cpu_kthread_setup(unsigned int cpu)
1356 {
1357         struct sched_param sp;
1358
1359         sp.sched_priority = RCU_KTHREAD_PRIO;
1360         sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1361 }
1362
1363 static void rcu_cpu_kthread_park(unsigned int cpu)
1364 {
1365         per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1366 }
1367
1368 static int rcu_cpu_kthread_should_run(unsigned int cpu)
1369 {
1370         return __this_cpu_read(rcu_cpu_has_work);
1371 }
1372
1373 /*
1374  * Per-CPU kernel thread that invokes RCU callbacks.  This replaces the
1375  * RCU softirq used in flavors and configurations of RCU that do not
1376  * support RCU priority boosting.
1377  */
1378 static void rcu_cpu_kthread(unsigned int cpu)
1379 {
1380         unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1381         char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1382         int spincnt;
1383
1384         for (spincnt = 0; spincnt < 10; spincnt++) {
1385                 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1386                 local_bh_disable();
1387                 *statusp = RCU_KTHREAD_RUNNING;
1388                 this_cpu_inc(rcu_cpu_kthread_loops);
1389                 local_irq_disable();
1390                 work = *workp;
1391                 *workp = 0;
1392                 local_irq_enable();
1393                 if (work)
1394                         rcu_kthread_do_work();
1395                 local_bh_enable();
1396                 if (*workp == 0) {
1397                         trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1398                         *statusp = RCU_KTHREAD_WAITING;
1399                         return;
1400                 }
1401         }
1402         *statusp = RCU_KTHREAD_YIELDING;
1403         trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1404         schedule_timeout_interruptible(2);
1405         trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1406         *statusp = RCU_KTHREAD_WAITING;
1407 }
1408
1409 /*
1410  * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1411  * served by the rcu_node in question.  The CPU hotplug lock is still
1412  * held, so the value of rnp->qsmaskinit will be stable.
1413  *
1414  * We don't include outgoingcpu in the affinity set, use -1 if there is
1415  * no outgoing CPU.  If there are no CPUs left in the affinity set,
1416  * this function allows the kthread to execute on any CPU.
1417  */
1418 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1419 {
1420         struct task_struct *t = rnp->boost_kthread_task;
1421         unsigned long mask = rnp->qsmaskinit;
1422         cpumask_var_t cm;
1423         int cpu;
1424
1425         if (!t)
1426                 return;
1427         if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1428                 return;
1429         for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1430                 if ((mask & 0x1) && cpu != outgoingcpu)
1431                         cpumask_set_cpu(cpu, cm);
1432         if (cpumask_weight(cm) == 0) {
1433                 cpumask_setall(cm);
1434                 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++)
1435                         cpumask_clear_cpu(cpu, cm);
1436                 WARN_ON_ONCE(cpumask_weight(cm) == 0);
1437         }
1438         set_cpus_allowed_ptr(t, cm);
1439         free_cpumask_var(cm);
1440 }
1441
1442 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1443         .store                  = &rcu_cpu_kthread_task,
1444         .thread_should_run      = rcu_cpu_kthread_should_run,
1445         .thread_fn              = rcu_cpu_kthread,
1446         .thread_comm            = "rcuc/%u",
1447         .setup                  = rcu_cpu_kthread_setup,
1448         .park                   = rcu_cpu_kthread_park,
1449 };
1450
1451 /*
1452  * Spawn boost kthreads -- called as soon as the scheduler is running.
1453  */
1454 static void __init rcu_spawn_boost_kthreads(void)
1455 {
1456         struct rcu_node *rnp;
1457         int cpu;
1458
1459         for_each_possible_cpu(cpu)
1460                 per_cpu(rcu_cpu_has_work, cpu) = 0;
1461         BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1462         rnp = rcu_get_root(rcu_state_p);
1463         (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1464         if (NUM_RCU_NODES > 1) {
1465                 rcu_for_each_leaf_node(rcu_state_p, rnp)
1466                         (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1467         }
1468 }
1469
1470 static void rcu_prepare_kthreads(int cpu)
1471 {
1472         struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
1473         struct rcu_node *rnp = rdp->mynode;
1474
1475         /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1476         if (rcu_scheduler_fully_active)
1477                 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1478 }
1479
1480 #else /* #ifdef CONFIG_RCU_BOOST */
1481
1482 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1483         __releases(rnp->lock)
1484 {
1485         raw_spin_unlock_irqrestore(&rnp->lock, flags);
1486 }
1487
1488 static void invoke_rcu_callbacks_kthread(void)
1489 {
1490         WARN_ON_ONCE(1);
1491 }
1492
1493 static bool rcu_is_callbacks_kthread(void)
1494 {
1495         return false;
1496 }
1497
1498 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1499 {
1500 }
1501
1502 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1503 {
1504 }
1505
1506 static void __init rcu_spawn_boost_kthreads(void)
1507 {
1508 }
1509
1510 static void rcu_prepare_kthreads(int cpu)
1511 {
1512 }
1513
1514 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1515
1516 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1517
1518 /*
1519  * Check to see if any future RCU-related work will need to be done
1520  * by the current CPU, even if none need be done immediately, returning
1521  * 1 if so.  This function is part of the RCU implementation; it is -not-
1522  * an exported member of the RCU API.
1523  *
1524  * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1525  * any flavor of RCU.
1526  */
1527 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1528 int rcu_needs_cpu(int cpu, unsigned long *delta_jiffies)
1529 {
1530         *delta_jiffies = ULONG_MAX;
1531         return rcu_cpu_has_callbacks(cpu, NULL);
1532 }
1533 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1534
1535 /*
1536  * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1537  * after it.
1538  */
1539 static void rcu_cleanup_after_idle(int cpu)
1540 {
1541 }
1542
1543 /*
1544  * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1545  * is nothing.
1546  */
1547 static void rcu_prepare_for_idle(int cpu)
1548 {
1549 }
1550
1551 /*
1552  * Don't bother keeping a running count of the number of RCU callbacks
1553  * posted because CONFIG_RCU_FAST_NO_HZ=n.
1554  */
1555 static void rcu_idle_count_callbacks_posted(void)
1556 {
1557 }
1558
1559 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1560
1561 /*
1562  * This code is invoked when a CPU goes idle, at which point we want
1563  * to have the CPU do everything required for RCU so that it can enter
1564  * the energy-efficient dyntick-idle mode.  This is handled by a
1565  * state machine implemented by rcu_prepare_for_idle() below.
1566  *
1567  * The following three proprocessor symbols control this state machine:
1568  *
1569  * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1570  *      to sleep in dyntick-idle mode with RCU callbacks pending.  This
1571  *      is sized to be roughly one RCU grace period.  Those energy-efficiency
1572  *      benchmarkers who might otherwise be tempted to set this to a large
1573  *      number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1574  *      system.  And if you are -that- concerned about energy efficiency,
1575  *      just power the system down and be done with it!
1576  * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1577  *      permitted to sleep in dyntick-idle mode with only lazy RCU
1578  *      callbacks pending.  Setting this too high can OOM your system.
1579  *
1580  * The values below work well in practice.  If future workloads require
1581  * adjustment, they can be converted into kernel config parameters, though
1582  * making the state machine smarter might be a better option.
1583  */
1584 #define RCU_IDLE_GP_DELAY 4             /* Roughly one grace period. */
1585 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1586
1587 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1588 module_param(rcu_idle_gp_delay, int, 0644);
1589 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1590 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1591
1592 extern int tick_nohz_active;
1593
1594 /*
1595  * Try to advance callbacks for all flavors of RCU on the current CPU, but
1596  * only if it has been awhile since the last time we did so.  Afterwards,
1597  * if there are any callbacks ready for immediate invocation, return true.
1598  */
1599 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1600 {
1601         bool cbs_ready = false;
1602         struct rcu_data *rdp;
1603         struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1604         struct rcu_node *rnp;
1605         struct rcu_state *rsp;
1606
1607         /* Exit early if we advanced recently. */
1608         if (jiffies == rdtp->last_advance_all)
1609                 return false;
1610         rdtp->last_advance_all = jiffies;
1611
1612         for_each_rcu_flavor(rsp) {
1613                 rdp = this_cpu_ptr(rsp->rda);
1614                 rnp = rdp->mynode;
1615
1616                 /*
1617                  * Don't bother checking unless a grace period has
1618                  * completed since we last checked and there are
1619                  * callbacks not yet ready to invoke.
1620                  */
1621                 if (rdp->completed != rnp->completed &&
1622                     rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1623                         note_gp_changes(rsp, rdp);
1624
1625                 if (cpu_has_callbacks_ready_to_invoke(rdp))
1626                         cbs_ready = true;
1627         }
1628         return cbs_ready;
1629 }
1630
1631 /*
1632  * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1633  * to invoke.  If the CPU has callbacks, try to advance them.  Tell the
1634  * caller to set the timeout based on whether or not there are non-lazy
1635  * callbacks.
1636  *
1637  * The caller must have disabled interrupts.
1638  */
1639 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1640 int rcu_needs_cpu(int cpu, unsigned long *dj)
1641 {
1642         struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1643
1644         /* Snapshot to detect later posting of non-lazy callback. */
1645         rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1646
1647         /* If no callbacks, RCU doesn't need the CPU. */
1648         if (!rcu_cpu_has_callbacks(cpu, &rdtp->all_lazy)) {
1649                 *dj = ULONG_MAX;
1650                 return 0;
1651         }
1652
1653         /* Attempt to advance callbacks. */
1654         if (rcu_try_advance_all_cbs()) {
1655                 /* Some ready to invoke, so initiate later invocation. */
1656                 invoke_rcu_core();
1657                 return 1;
1658         }
1659         rdtp->last_accelerate = jiffies;
1660
1661         /* Request timer delay depending on laziness, and round. */
1662         if (!rdtp->all_lazy) {
1663                 *dj = round_up(rcu_idle_gp_delay + jiffies,
1664                                rcu_idle_gp_delay) - jiffies;
1665         } else {
1666                 *dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1667         }
1668         return 0;
1669 }
1670 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1671
1672 /*
1673  * Prepare a CPU for idle from an RCU perspective.  The first major task
1674  * is to sense whether nohz mode has been enabled or disabled via sysfs.
1675  * The second major task is to check to see if a non-lazy callback has
1676  * arrived at a CPU that previously had only lazy callbacks.  The third
1677  * major task is to accelerate (that is, assign grace-period numbers to)
1678  * any recently arrived callbacks.
1679  *
1680  * The caller must have disabled interrupts.
1681  */
1682 static void rcu_prepare_for_idle(int cpu)
1683 {
1684 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1685         bool needwake;
1686         struct rcu_data *rdp;
1687         struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1688         struct rcu_node *rnp;
1689         struct rcu_state *rsp;
1690         int tne;
1691
1692         /* Handle nohz enablement switches conservatively. */
1693         tne = ACCESS_ONCE(tick_nohz_active);
1694         if (tne != rdtp->tick_nohz_enabled_snap) {
1695                 if (rcu_cpu_has_callbacks(cpu, NULL))
1696                         invoke_rcu_core(); /* force nohz to see update. */
1697                 rdtp->tick_nohz_enabled_snap = tne;
1698                 return;
1699         }
1700         if (!tne)
1701                 return;
1702
1703         /* If this is a no-CBs CPU, no callbacks, just return. */
1704         if (rcu_is_nocb_cpu(cpu))
1705                 return;
1706
1707         /*
1708          * If a non-lazy callback arrived at a CPU having only lazy
1709          * callbacks, invoke RCU core for the side-effect of recalculating
1710          * idle duration on re-entry to idle.
1711          */
1712         if (rdtp->all_lazy &&
1713             rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1714                 rdtp->all_lazy = false;
1715                 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1716                 invoke_rcu_core();
1717                 return;
1718         }
1719
1720         /*
1721          * If we have not yet accelerated this jiffy, accelerate all
1722          * callbacks on this CPU.
1723          */
1724         if (rdtp->last_accelerate == jiffies)
1725                 return;
1726         rdtp->last_accelerate = jiffies;
1727         for_each_rcu_flavor(rsp) {
1728                 rdp = per_cpu_ptr(rsp->rda, cpu);
1729                 if (!*rdp->nxttail[RCU_DONE_TAIL])
1730                         continue;
1731                 rnp = rdp->mynode;
1732                 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1733                 smp_mb__after_unlock_lock();
1734                 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1735                 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1736                 if (needwake)
1737                         rcu_gp_kthread_wake(rsp);
1738         }
1739 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1740 }
1741
1742 /*
1743  * Clean up for exit from idle.  Attempt to advance callbacks based on
1744  * any grace periods that elapsed while the CPU was idle, and if any
1745  * callbacks are now ready to invoke, initiate invocation.
1746  */
1747 static void rcu_cleanup_after_idle(int cpu)
1748 {
1749 #ifndef CONFIG_RCU_NOCB_CPU_ALL
1750         if (rcu_is_nocb_cpu(cpu))
1751                 return;
1752         if (rcu_try_advance_all_cbs())
1753                 invoke_rcu_core();
1754 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
1755 }
1756
1757 /*
1758  * Keep a running count of the number of non-lazy callbacks posted
1759  * on this CPU.  This running counter (which is never decremented) allows
1760  * rcu_prepare_for_idle() to detect when something out of the idle loop
1761  * posts a callback, even if an equal number of callbacks are invoked.
1762  * Of course, callbacks should only be posted from within a trace event
1763  * designed to be called from idle or from within RCU_NONIDLE().
1764  */
1765 static void rcu_idle_count_callbacks_posted(void)
1766 {
1767         __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1768 }
1769
1770 /*
1771  * Data for flushing lazy RCU callbacks at OOM time.
1772  */
1773 static atomic_t oom_callback_count;
1774 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1775
1776 /*
1777  * RCU OOM callback -- decrement the outstanding count and deliver the
1778  * wake-up if we are the last one.
1779  */
1780 static void rcu_oom_callback(struct rcu_head *rhp)
1781 {
1782         if (atomic_dec_and_test(&oom_callback_count))
1783                 wake_up(&oom_callback_wq);
1784 }
1785
1786 /*
1787  * Post an rcu_oom_notify callback on the current CPU if it has at
1788  * least one lazy callback.  This will unnecessarily post callbacks
1789  * to CPUs that already have a non-lazy callback at the end of their
1790  * callback list, but this is an infrequent operation, so accept some
1791  * extra overhead to keep things simple.
1792  */
1793 static void rcu_oom_notify_cpu(void *unused)
1794 {
1795         struct rcu_state *rsp;
1796         struct rcu_data *rdp;
1797
1798         for_each_rcu_flavor(rsp) {
1799                 rdp = raw_cpu_ptr(rsp->rda);
1800                 if (rdp->qlen_lazy != 0) {
1801                         atomic_inc(&oom_callback_count);
1802                         rsp->call(&rdp->oom_head, rcu_oom_callback);
1803                 }
1804         }
1805 }
1806
1807 /*
1808  * If low on memory, ensure that each CPU has a non-lazy callback.
1809  * This will wake up CPUs that have only lazy callbacks, in turn
1810  * ensuring that they free up the corresponding memory in a timely manner.
1811  * Because an uncertain amount of memory will be freed in some uncertain
1812  * timeframe, we do not claim to have freed anything.
1813  */
1814 static int rcu_oom_notify(struct notifier_block *self,
1815                           unsigned long notused, void *nfreed)
1816 {
1817         int cpu;
1818
1819         /* Wait for callbacks from earlier instance to complete. */
1820         wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1821         smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1822
1823         /*
1824          * Prevent premature wakeup: ensure that all increments happen
1825          * before there is a chance of the counter reaching zero.
1826          */
1827         atomic_set(&oom_callback_count, 1);
1828
1829         get_online_cpus();
1830         for_each_online_cpu(cpu) {
1831                 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1832                 cond_resched_rcu_qs();
1833         }
1834         put_online_cpus();
1835
1836         /* Unconditionally decrement: no need to wake ourselves up. */
1837         atomic_dec(&oom_callback_count);
1838
1839         return NOTIFY_OK;
1840 }
1841
1842 static struct notifier_block rcu_oom_nb = {
1843         .notifier_call = rcu_oom_notify
1844 };
1845
1846 static int __init rcu_register_oom_notifier(void)
1847 {
1848         register_oom_notifier(&rcu_oom_nb);
1849         return 0;
1850 }
1851 early_initcall(rcu_register_oom_notifier);
1852
1853 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1854
1855 #ifdef CONFIG_RCU_CPU_STALL_INFO
1856
1857 #ifdef CONFIG_RCU_FAST_NO_HZ
1858
1859 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1860 {
1861         struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1862         unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1863
1864         sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1865                 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1866                 ulong2long(nlpd),
1867                 rdtp->all_lazy ? 'L' : '.',
1868                 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1869 }
1870
1871 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1872
1873 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1874 {
1875         *cp = '\0';
1876 }
1877
1878 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1879
1880 /* Initiate the stall-info list. */
1881 static void print_cpu_stall_info_begin(void)
1882 {
1883         pr_cont("\n");
1884 }
1885
1886 /*
1887  * Print out diagnostic information for the specified stalled CPU.
1888  *
1889  * If the specified CPU is aware of the current RCU grace period
1890  * (flavor specified by rsp), then print the number of scheduling
1891  * clock interrupts the CPU has taken during the time that it has
1892  * been aware.  Otherwise, print the number of RCU grace periods
1893  * that this CPU is ignorant of, for example, "1" if the CPU was
1894  * aware of the previous grace period.
1895  *
1896  * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1897  */
1898 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1899 {
1900         char fast_no_hz[72];
1901         struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1902         struct rcu_dynticks *rdtp = rdp->dynticks;
1903         char *ticks_title;
1904         unsigned long ticks_value;
1905
1906         if (rsp->gpnum == rdp->gpnum) {
1907                 ticks_title = "ticks this GP";
1908                 ticks_value = rdp->ticks_this_gp;
1909         } else {
1910                 ticks_title = "GPs behind";
1911                 ticks_value = rsp->gpnum - rdp->gpnum;
1912         }
1913         print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1914         pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u %s\n",
1915                cpu, ticks_value, ticks_title,
1916                atomic_read(&rdtp->dynticks) & 0xfff,
1917                rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1918                rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1919                fast_no_hz);
1920 }
1921
1922 /* Terminate the stall-info list. */
1923 static void print_cpu_stall_info_end(void)
1924 {
1925         pr_err("\t");
1926 }
1927
1928 /* Zero ->ticks_this_gp for all flavors of RCU. */
1929 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1930 {
1931         rdp->ticks_this_gp = 0;
1932         rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1933 }
1934
1935 /* Increment ->ticks_this_gp for all flavors of RCU. */
1936 static void increment_cpu_stall_ticks(void)
1937 {
1938         struct rcu_state *rsp;
1939
1940         for_each_rcu_flavor(rsp)
1941                 raw_cpu_inc(rsp->rda->ticks_this_gp);
1942 }
1943
1944 #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
1945
1946 static void print_cpu_stall_info_begin(void)
1947 {
1948         pr_cont(" {");
1949 }
1950
1951 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1952 {
1953         pr_cont(" %d", cpu);
1954 }
1955
1956 static void print_cpu_stall_info_end(void)
1957 {
1958         pr_cont("} ");
1959 }
1960
1961 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1962 {
1963 }
1964
1965 static void increment_cpu_stall_ticks(void)
1966 {
1967 }
1968
1969 #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
1970
1971 #ifdef CONFIG_RCU_NOCB_CPU
1972
1973 /*
1974  * Offload callback processing from the boot-time-specified set of CPUs
1975  * specified by rcu_nocb_mask.  For each CPU in the set, there is a
1976  * kthread created that pulls the callbacks from the corresponding CPU,
1977  * waits for a grace period to elapse, and invokes the callbacks.
1978  * The no-CBs CPUs do a wake_up() on their kthread when they insert
1979  * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1980  * has been specified, in which case each kthread actively polls its
1981  * CPU.  (Which isn't so great for energy efficiency, but which does
1982  * reduce RCU's overhead on that CPU.)
1983  *
1984  * This is intended to be used in conjunction with Frederic Weisbecker's
1985  * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1986  * running CPU-bound user-mode computations.
1987  *
1988  * Offloading of callback processing could also in theory be used as
1989  * an energy-efficiency measure because CPUs with no RCU callbacks
1990  * queued are more aggressive about entering dyntick-idle mode.
1991  */
1992
1993
1994 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1995 static int __init rcu_nocb_setup(char *str)
1996 {
1997         alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1998         have_rcu_nocb_mask = true;
1999         cpulist_parse(str, rcu_nocb_mask);
2000         return 1;
2001 }
2002 __setup("rcu_nocbs=", rcu_nocb_setup);
2003
2004 static int __init parse_rcu_nocb_poll(char *arg)
2005 {
2006         rcu_nocb_poll = 1;
2007         return 0;
2008 }
2009 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
2010
2011 /*
2012  * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
2013  * grace period.
2014  */
2015 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2016 {
2017         wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]);
2018 }
2019
2020 /*
2021  * Set the root rcu_node structure's ->need_future_gp field
2022  * based on the sum of those of all rcu_node structures.  This does
2023  * double-count the root rcu_node structure's requests, but this
2024  * is necessary to handle the possibility of a rcu_nocb_kthread()
2025  * having awakened during the time that the rcu_node structures
2026  * were being updated for the end of the previous grace period.
2027  */
2028 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2029 {
2030         rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
2031 }
2032
2033 static void rcu_init_one_nocb(struct rcu_node *rnp)
2034 {
2035         init_waitqueue_head(&rnp->nocb_gp_wq[0]);
2036         init_waitqueue_head(&rnp->nocb_gp_wq[1]);
2037 }
2038
2039 #ifndef CONFIG_RCU_NOCB_CPU_ALL
2040 /* Is the specified CPU a no-CBs CPU? */
2041 bool rcu_is_nocb_cpu(int cpu)
2042 {
2043         if (have_rcu_nocb_mask)
2044                 return cpumask_test_cpu(cpu, rcu_nocb_mask);
2045         return false;
2046 }
2047 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */
2048
2049 /*
2050  * Kick the leader kthread for this NOCB group.
2051  */
2052 static void wake_nocb_leader(struct rcu_data *rdp, bool force)
2053 {
2054         struct rcu_data *rdp_leader = rdp->nocb_leader;
2055
2056         if (!ACCESS_ONCE(rdp_leader->nocb_kthread))
2057                 return;
2058         if (ACCESS_ONCE(rdp_leader->nocb_leader_sleep) || force) {
2059                 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
2060                 ACCESS_ONCE(rdp_leader->nocb_leader_sleep) = false;
2061                 wake_up(&rdp_leader->nocb_wq);
2062         }
2063 }
2064
2065 /*
2066  * Does the specified CPU need an RCU callback for the specified flavor
2067  * of rcu_barrier()?
2068  */
2069 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
2070 {
2071         struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2072         struct rcu_head *rhp;
2073
2074         /* No-CBs CPUs might have callbacks on any of three lists. */
2075         rhp = ACCESS_ONCE(rdp->nocb_head);
2076         if (!rhp)
2077                 rhp = ACCESS_ONCE(rdp->nocb_gp_head);
2078         if (!rhp)
2079                 rhp = ACCESS_ONCE(rdp->nocb_follower_head);
2080
2081         /* Having no rcuo kthread but CBs after scheduler starts is bad! */
2082         if (!ACCESS_ONCE(rdp->nocb_kthread) && rhp) {
2083                 /* RCU callback enqueued before CPU first came online??? */
2084                 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
2085                        cpu, rhp->func);
2086                 WARN_ON_ONCE(1);
2087         }
2088
2089         return !!rhp;
2090 }
2091
2092 /*
2093  * Enqueue the specified string of rcu_head structures onto the specified
2094  * CPU's no-CBs lists.  The CPU is specified by rdp, the head of the
2095  * string by rhp, and the tail of the string by rhtp.  The non-lazy/lazy
2096  * counts are supplied by rhcount and rhcount_lazy.
2097  *
2098  * If warranted, also wake up the kthread servicing this CPUs queues.
2099  */
2100 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
2101                                     struct rcu_head *rhp,
2102                                     struct rcu_head **rhtp,
2103                                     int rhcount, int rhcount_lazy,
2104                                     unsigned long flags)
2105 {
2106         int len;
2107         struct rcu_head **old_rhpp;
2108         struct task_struct *t;
2109
2110         /* Enqueue the callback on the nocb list and update counts. */
2111         old_rhpp = xchg(&rdp->nocb_tail, rhtp);
2112         ACCESS_ONCE(*old_rhpp) = rhp;
2113         atomic_long_add(rhcount, &rdp->nocb_q_count);
2114         atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
2115         smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
2116
2117         /* If we are not being polled and there is a kthread, awaken it ... */
2118         t = ACCESS_ONCE(rdp->nocb_kthread);
2119         if (rcu_nocb_poll || !t) {
2120                 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2121                                     TPS("WakeNotPoll"));
2122                 return;
2123         }
2124         len = atomic_long_read(&rdp->nocb_q_count);
2125         if (old_rhpp == &rdp->nocb_head) {
2126                 if (!irqs_disabled_flags(flags)) {
2127                         /* ... if queue was empty ... */
2128                         wake_nocb_leader(rdp, false);
2129                         trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2130                                             TPS("WakeEmpty"));
2131                 } else {
2132                         rdp->nocb_defer_wakeup = RCU_NOGP_WAKE;
2133                         trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2134                                             TPS("WakeEmptyIsDeferred"));
2135                 }
2136                 rdp->qlen_last_fqs_check = 0;
2137         } else if (len > rdp->qlen_last_fqs_check + qhimark) {
2138                 /* ... or if many callbacks queued. */
2139                 if (!irqs_disabled_flags(flags)) {
2140                         wake_nocb_leader(rdp, true);
2141                         trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2142                                             TPS("WakeOvf"));
2143                 } else {
2144                         rdp->nocb_defer_wakeup = RCU_NOGP_WAKE_FORCE;
2145                         trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2146                                             TPS("WakeOvfIsDeferred"));
2147                 }
2148                 rdp->qlen_last_fqs_check = LONG_MAX / 2;
2149         } else {
2150                 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
2151         }
2152         return;
2153 }
2154
2155 /*
2156  * This is a helper for __call_rcu(), which invokes this when the normal
2157  * callback queue is inoperable.  If this is not a no-CBs CPU, this
2158  * function returns failure back to __call_rcu(), which can complain
2159  * appropriately.
2160  *
2161  * Otherwise, this function queues the callback where the corresponding
2162  * "rcuo" kthread can find it.
2163  */
2164 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2165                             bool lazy, unsigned long flags)
2166 {
2167
2168         if (!rcu_is_nocb_cpu(rdp->cpu))
2169                 return false;
2170         __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
2171         if (__is_kfree_rcu_offset((unsigned long)rhp->func))
2172                 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
2173                                          (unsigned long)rhp->func,
2174                                          -atomic_long_read(&rdp->nocb_q_count_lazy),
2175                                          -atomic_long_read(&rdp->nocb_q_count));
2176         else
2177                 trace_rcu_callback(rdp->rsp->name, rhp,
2178                                    -atomic_long_read(&rdp->nocb_q_count_lazy),
2179                                    -atomic_long_read(&rdp->nocb_q_count));
2180
2181         /*
2182          * If called from an extended quiescent state with interrupts
2183          * disabled, invoke the RCU core in order to allow the idle-entry
2184          * deferred-wakeup check to function.
2185          */
2186         if (irqs_disabled_flags(flags) &&
2187             !rcu_is_watching() &&
2188             cpu_online(smp_processor_id()))
2189                 invoke_rcu_core();
2190
2191         return true;
2192 }
2193
2194 /*
2195  * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2196  * not a no-CBs CPU.
2197  */
2198 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2199                                                      struct rcu_data *rdp,
2200                                                      unsigned long flags)
2201 {
2202         long ql = rsp->qlen;
2203         long qll = rsp->qlen_lazy;
2204
2205         /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2206         if (!rcu_is_nocb_cpu(smp_processor_id()))
2207                 return false;
2208         rsp->qlen = 0;
2209         rsp->qlen_lazy = 0;
2210
2211         /* First, enqueue the donelist, if any.  This preserves CB ordering. */
2212         if (rsp->orphan_donelist != NULL) {
2213                 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
2214                                         rsp->orphan_donetail, ql, qll, flags);
2215                 ql = qll = 0;
2216                 rsp->orphan_donelist = NULL;
2217                 rsp->orphan_donetail = &rsp->orphan_donelist;
2218         }
2219         if (rsp->orphan_nxtlist != NULL) {
2220                 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
2221                                         rsp->orphan_nxttail, ql, qll, flags);
2222                 ql = qll = 0;
2223                 rsp->orphan_nxtlist = NULL;
2224                 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2225         }
2226         return true;
2227 }
2228
2229 /*
2230  * If necessary, kick off a new grace period, and either way wait
2231  * for a subsequent grace period to complete.
2232  */
2233 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2234 {
2235         unsigned long c;
2236         bool d;
2237         unsigned long flags;
2238         bool needwake;
2239         struct rcu_node *rnp = rdp->mynode;
2240
2241         raw_spin_lock_irqsave(&rnp->lock, flags);
2242         smp_mb__after_unlock_lock();
2243         needwake = rcu_start_future_gp(rnp, rdp, &c);
2244         raw_spin_unlock_irqrestore(&rnp->lock, flags);
2245         if (needwake)
2246                 rcu_gp_kthread_wake(rdp->rsp);
2247
2248         /*
2249          * Wait for the grace period.  Do so interruptibly to avoid messing
2250          * up the load average.
2251          */
2252         trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
2253         for (;;) {
2254                 wait_event_interruptible(
2255                         rnp->nocb_gp_wq[c & 0x1],
2256                         (d = ULONG_CMP_GE(ACCESS_ONCE(rnp->completed), c)));
2257                 if (likely(d))
2258                         break;
2259                 WARN_ON(signal_pending(current));
2260                 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
2261         }
2262         trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
2263         smp_mb(); /* Ensure that CB invocation happens after GP end. */
2264 }
2265
2266 /*
2267  * Leaders come here to wait for additional callbacks to show up.
2268  * This function does not return until callbacks appear.
2269  */
2270 static void nocb_leader_wait(struct rcu_data *my_rdp)
2271 {
2272         bool firsttime = true;
2273         bool gotcbs;
2274         struct rcu_data *rdp;
2275         struct rcu_head **tail;
2276
2277 wait_again:
2278
2279         /* Wait for callbacks to appear. */
2280         if (!rcu_nocb_poll) {
2281                 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep");
2282                 wait_event_interruptible(my_rdp->nocb_wq,
2283                                 !ACCESS_ONCE(my_rdp->nocb_leader_sleep));
2284                 /* Memory barrier handled by smp_mb() calls below and repoll. */
2285         } else if (firsttime) {
2286                 firsttime = false; /* Don't drown trace log with "Poll"! */
2287                 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll");
2288         }
2289
2290         /*
2291          * Each pass through the following loop checks a follower for CBs.
2292          * We are our own first follower.  Any CBs found are moved to
2293          * nocb_gp_head, where they await a grace period.
2294          */
2295         gotcbs = false;
2296         for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2297                 rdp->nocb_gp_head = ACCESS_ONCE(rdp->nocb_head);
2298                 if (!rdp->nocb_gp_head)
2299                         continue;  /* No CBs here, try next follower. */
2300
2301                 /* Move callbacks to wait-for-GP list, which is empty. */
2302                 ACCESS_ONCE(rdp->nocb_head) = NULL;
2303                 rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2304                 rdp->nocb_gp_count = atomic_long_xchg(&rdp->nocb_q_count, 0);
2305                 rdp->nocb_gp_count_lazy =
2306                         atomic_long_xchg(&rdp->nocb_q_count_lazy, 0);
2307                 gotcbs = true;
2308         }
2309
2310         /*
2311          * If there were no callbacks, sleep a bit, rescan after a
2312          * memory barrier, and go retry.
2313          */
2314         if (unlikely(!gotcbs)) {
2315                 if (!rcu_nocb_poll)
2316                         trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu,
2317                                             "WokeEmpty");
2318                 WARN_ON(signal_pending(current));
2319                 schedule_timeout_interruptible(1);
2320
2321                 /* Rescan in case we were a victim of memory ordering. */
2322                 my_rdp->nocb_leader_sleep = true;
2323                 smp_mb();  /* Ensure _sleep true before scan. */
2324                 for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower)
2325                         if (ACCESS_ONCE(rdp->nocb_head)) {
2326                                 /* Found CB, so short-circuit next wait. */
2327                                 my_rdp->nocb_leader_sleep = false;
2328                                 break;
2329                         }
2330                 goto wait_again;
2331         }
2332
2333         /* Wait for one grace period. */
2334         rcu_nocb_wait_gp(my_rdp);
2335
2336         /*
2337          * We left ->nocb_leader_sleep unset to reduce cache thrashing.
2338          * We set it now, but recheck for new callbacks while
2339          * traversing our follower list.
2340          */
2341         my_rdp->nocb_leader_sleep = true;
2342         smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */
2343
2344         /* Each pass through the following loop wakes a follower, if needed. */
2345         for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2346                 if (ACCESS_ONCE(rdp->nocb_head))
2347                         my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/
2348                 if (!rdp->nocb_gp_head)
2349                         continue; /* No CBs, so no need to wake follower. */
2350
2351                 /* Append callbacks to follower's "done" list. */
2352                 tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail);
2353                 *tail = rdp->nocb_gp_head;
2354                 atomic_long_add(rdp->nocb_gp_count, &rdp->nocb_follower_count);
2355                 atomic_long_add(rdp->nocb_gp_count_lazy,
2356                                 &rdp->nocb_follower_count_lazy);
2357                 smp_mb__after_atomic(); /* Store *tail before wakeup. */
2358                 if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
2359                         /*
2360                          * List was empty, wake up the follower.
2361                          * Memory barriers supplied by atomic_long_add().
2362                          */
2363                         wake_up(&rdp->nocb_wq);
2364                 }
2365         }
2366
2367         /* If we (the leader) don't have CBs, go wait some more. */
2368         if (!my_rdp->nocb_follower_head)
2369                 goto wait_again;
2370 }
2371
2372 /*
2373  * Followers come here to wait for additional callbacks to show up.
2374  * This function does not return until callbacks appear.
2375  */
2376 static void nocb_follower_wait(struct rcu_data *rdp)
2377 {
2378         bool firsttime = true;
2379
2380         for (;;) {
2381                 if (!rcu_nocb_poll) {
2382                         trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2383                                             "FollowerSleep");
2384                         wait_event_interruptible(rdp->nocb_wq,
2385                                                  ACCESS_ONCE(rdp->nocb_follower_head));
2386                 } else if (firsttime) {
2387                         /* Don't drown trace log with "Poll"! */
2388                         firsttime = false;
2389                         trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll");
2390                 }
2391                 if (smp_load_acquire(&rdp->nocb_follower_head)) {
2392                         /* ^^^ Ensure CB invocation follows _head test. */
2393                         return;
2394                 }
2395                 if (!rcu_nocb_poll)
2396                         trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2397                                             "WokeEmpty");
2398                 WARN_ON(signal_pending(current));
2399                 schedule_timeout_interruptible(1);
2400         }
2401 }
2402
2403 /*
2404  * Per-rcu_data kthread, but only for no-CBs CPUs.  Each kthread invokes
2405  * callbacks queued by the corresponding no-CBs CPU, however, there is
2406  * an optional leader-follower relationship so that the grace-period
2407  * kthreads don't have to do quite so many wakeups.
2408  */
2409 static int rcu_nocb_kthread(void *arg)
2410 {
2411         int c, cl;
2412         struct rcu_head *list;
2413         struct rcu_head *next;
2414         struct rcu_head **tail;
2415         struct rcu_data *rdp = arg;
2416
2417         /* Each pass through this loop invokes one batch of callbacks */
2418         for (;;) {
2419                 /* Wait for callbacks. */
2420                 if (rdp->nocb_leader == rdp)
2421                         nocb_leader_wait(rdp);
2422                 else
2423                         nocb_follower_wait(rdp);
2424
2425                 /* Pull the ready-to-invoke callbacks onto local list. */
2426                 list = ACCESS_ONCE(rdp->nocb_follower_head);
2427                 BUG_ON(!list);
2428                 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty");
2429                 ACCESS_ONCE(rdp->nocb_follower_head) = NULL;
2430                 tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head);
2431                 c = atomic_long_xchg(&rdp->nocb_follower_count, 0);
2432                 cl = atomic_long_xchg(&rdp->nocb_follower_count_lazy, 0);
2433                 rdp->nocb_p_count += c;
2434                 rdp->nocb_p_count_lazy += cl;
2435
2436                 /* Each pass through the following loop invokes a callback. */
2437                 trace_rcu_batch_start(rdp->rsp->name, cl, c, -1);
2438                 c = cl = 0;
2439                 while (list) {
2440                         next = list->next;
2441                         /* Wait for enqueuing to complete, if needed. */
2442                         while (next == NULL && &list->next != tail) {
2443                                 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2444                                                     TPS("WaitQueue"));
2445                                 schedule_timeout_interruptible(1);
2446                                 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2447                                                     TPS("WokeQueue"));
2448                                 next = list->next;
2449                         }
2450                         debug_rcu_head_unqueue(list);
2451                         local_bh_disable();
2452                         if (__rcu_reclaim(rdp->rsp->name, list))
2453                                 cl++;
2454                         c++;
2455                         local_bh_enable();
2456                         list = next;
2457                 }
2458                 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2459                 ACCESS_ONCE(rdp->nocb_p_count) = rdp->nocb_p_count - c;
2460                 ACCESS_ONCE(rdp->nocb_p_count_lazy) =
2461                                                 rdp->nocb_p_count_lazy - cl;
2462                 rdp->n_nocbs_invoked += c;
2463         }
2464         return 0;
2465 }
2466
2467 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2468 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2469 {
2470         return ACCESS_ONCE(rdp->nocb_defer_wakeup);
2471 }
2472
2473 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2474 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2475 {
2476         int ndw;
2477
2478         if (!rcu_nocb_need_deferred_wakeup(rdp))
2479                 return;
2480         ndw = ACCESS_ONCE(rdp->nocb_defer_wakeup);
2481         ACCESS_ONCE(rdp->nocb_defer_wakeup) = RCU_NOGP_WAKE_NOT;
2482         wake_nocb_leader(rdp, ndw == RCU_NOGP_WAKE_FORCE);
2483         trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake"));
2484 }
2485
2486 void __init rcu_init_nohz(void)
2487 {
2488         int cpu;
2489         bool need_rcu_nocb_mask = true;
2490         struct rcu_state *rsp;
2491
2492 #ifdef CONFIG_RCU_NOCB_CPU_NONE
2493         need_rcu_nocb_mask = false;
2494 #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
2495
2496 #if defined(CONFIG_NO_HZ_FULL)
2497         if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2498                 need_rcu_nocb_mask = true;
2499 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2500
2501         if (!have_rcu_nocb_mask && need_rcu_nocb_mask) {
2502                 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2503                         pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2504                         return;
2505                 }
2506                 have_rcu_nocb_mask = true;
2507         }
2508         if (!have_rcu_nocb_mask)
2509                 return;
2510
2511 #ifdef CONFIG_RCU_NOCB_CPU_ZERO
2512         pr_info("\tOffload RCU callbacks from CPU 0\n");
2513         cpumask_set_cpu(0, rcu_nocb_mask);
2514 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
2515 #ifdef CONFIG_RCU_NOCB_CPU_ALL
2516         pr_info("\tOffload RCU callbacks from all CPUs\n");
2517         cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
2518 #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
2519 #if defined(CONFIG_NO_HZ_FULL)
2520         if (tick_nohz_full_running)
2521                 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2522 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2523
2524         if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2525                 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
2526                 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2527                             rcu_nocb_mask);
2528         }
2529         cpulist_scnprintf(nocb_buf, sizeof(nocb_buf), rcu_nocb_mask);
2530         pr_info("\tOffload RCU callbacks from CPUs: %s.\n", nocb_buf);
2531         if (rcu_nocb_poll)
2532                 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2533
2534         for_each_rcu_flavor(rsp) {
2535                 for_each_cpu(cpu, rcu_nocb_mask) {
2536                         struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2537
2538                         /*
2539                          * If there are early callbacks, they will need
2540                          * to be moved to the nocb lists.
2541                          */
2542                         WARN_ON_ONCE(rdp->nxttail[RCU_NEXT_TAIL] !=
2543                                      &rdp->nxtlist &&
2544                                      rdp->nxttail[RCU_NEXT_TAIL] != NULL);
2545                         init_nocb_callback_list(rdp);
2546                 }
2547                 rcu_organize_nocb_kthreads(rsp);
2548         }
2549 }
2550
2551 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2552 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2553 {
2554         rdp->nocb_tail = &rdp->nocb_head;
2555         init_waitqueue_head(&rdp->nocb_wq);
2556         rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2557 }
2558
2559 /*
2560  * If the specified CPU is a no-CBs CPU that does not already have its
2561  * rcuo kthread for the specified RCU flavor, spawn it.  If the CPUs are
2562  * brought online out of order, this can require re-organizing the
2563  * leader-follower relationships.
2564  */
2565 static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu)
2566 {
2567         struct rcu_data *rdp;
2568         struct rcu_data *rdp_last;
2569         struct rcu_data *rdp_old_leader;
2570         struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu);
2571         struct task_struct *t;
2572
2573         /*
2574          * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2575          * then nothing to do.
2576          */
2577         if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread)
2578                 return;
2579
2580         /* If we didn't spawn the leader first, reorganize! */
2581         rdp_old_leader = rdp_spawn->nocb_leader;
2582         if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) {
2583                 rdp_last = NULL;
2584                 rdp = rdp_old_leader;
2585                 do {
2586                         rdp->nocb_leader = rdp_spawn;
2587                         if (rdp_last && rdp != rdp_spawn)
2588                                 rdp_last->nocb_next_follower = rdp;
2589                         rdp_last = rdp;
2590                         rdp = rdp->nocb_next_follower;
2591                         rdp_last->nocb_next_follower = NULL;
2592                 } while (rdp);
2593                 rdp_spawn->nocb_next_follower = rdp_old_leader;
2594         }
2595
2596         /* Spawn the kthread for this CPU and RCU flavor. */
2597         t = kthread_run(rcu_nocb_kthread, rdp_spawn,
2598                         "rcuo%c/%d", rsp->abbr, cpu);
2599         BUG_ON(IS_ERR(t));
2600         ACCESS_ONCE(rdp_spawn->nocb_kthread) = t;
2601 }
2602
2603 /*
2604  * If the specified CPU is a no-CBs CPU that does not already have its
2605  * rcuo kthreads, spawn them.
2606  */
2607 static void rcu_spawn_all_nocb_kthreads(int cpu)
2608 {
2609         struct rcu_state *rsp;
2610
2611         if (rcu_scheduler_fully_active)
2612                 for_each_rcu_flavor(rsp)
2613                         rcu_spawn_one_nocb_kthread(rsp, cpu);
2614 }
2615
2616 /*
2617  * Once the scheduler is running, spawn rcuo kthreads for all online
2618  * no-CBs CPUs.  This assumes that the early_initcall()s happen before
2619  * non-boot CPUs come online -- if this changes, we will need to add
2620  * some mutual exclusion.
2621  */
2622 static void __init rcu_spawn_nocb_kthreads(void)
2623 {
2624         int cpu;
2625
2626         for_each_online_cpu(cpu)
2627                 rcu_spawn_all_nocb_kthreads(cpu);
2628 }
2629
2630 /* How many follower CPU IDs per leader?  Default of -1 for sqrt(nr_cpu_ids). */
2631 static int rcu_nocb_leader_stride = -1;
2632 module_param(rcu_nocb_leader_stride, int, 0444);
2633
2634 /*
2635  * Initialize leader-follower relationships for all no-CBs CPU.
2636  */
2637 static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp)
2638 {
2639         int cpu;
2640         int ls = rcu_nocb_leader_stride;
2641         int nl = 0;  /* Next leader. */
2642         struct rcu_data *rdp;
2643         struct rcu_data *rdp_leader = NULL;  /* Suppress misguided gcc warn. */
2644         struct rcu_data *rdp_prev = NULL;
2645
2646         if (!have_rcu_nocb_mask)
2647                 return;
2648         if (ls == -1) {
2649                 ls = int_sqrt(nr_cpu_ids);
2650                 rcu_nocb_leader_stride = ls;
2651         }
2652
2653         /*
2654          * Each pass through this loop sets up one rcu_data structure and
2655          * spawns one rcu_nocb_kthread().
2656          */
2657         for_each_cpu(cpu, rcu_nocb_mask) {
2658                 rdp = per_cpu_ptr(rsp->rda, cpu);
2659                 if (rdp->cpu >= nl) {
2660                         /* New leader, set up for followers & next leader. */
2661                         nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2662                         rdp->nocb_leader = rdp;
2663                         rdp_leader = rdp;
2664                 } else {
2665                         /* Another follower, link to previous leader. */
2666                         rdp->nocb_leader = rdp_leader;
2667                         rdp_prev->nocb_next_follower = rdp;
2668                 }
2669                 rdp_prev = rdp;
2670         }
2671 }
2672
2673 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2674 static bool init_nocb_callback_list(struct rcu_data *rdp)
2675 {
2676         if (!rcu_is_nocb_cpu(rdp->cpu))
2677                 return false;
2678
2679         rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2680         return true;
2681 }
2682
2683 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2684
2685 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
2686 {
2687         WARN_ON_ONCE(1); /* Should be dead code. */
2688         return false;
2689 }
2690
2691 static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2692 {
2693 }
2694
2695 static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2696 {
2697 }
2698
2699 static void rcu_init_one_nocb(struct rcu_node *rnp)
2700 {
2701 }
2702
2703 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2704                             bool lazy, unsigned long flags)
2705 {
2706         return false;
2707 }
2708
2709 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2710                                                      struct rcu_data *rdp,
2711                                                      unsigned long flags)
2712 {
2713         return false;
2714 }
2715
2716 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2717 {
2718 }
2719
2720 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2721 {
2722         return false;
2723 }
2724
2725 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2726 {
2727 }
2728
2729 static void rcu_spawn_all_nocb_kthreads(int cpu)
2730 {
2731 }
2732
2733 static void __init rcu_spawn_nocb_kthreads(void)
2734 {
2735 }
2736
2737 static bool init_nocb_callback_list(struct rcu_data *rdp)
2738 {
2739         return false;
2740 }
2741
2742 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2743
2744 /*
2745  * An adaptive-ticks CPU can potentially execute in kernel mode for an
2746  * arbitrarily long period of time with the scheduling-clock tick turned
2747  * off.  RCU will be paying attention to this CPU because it is in the
2748  * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2749  * machine because the scheduling-clock tick has been disabled.  Therefore,
2750  * if an adaptive-ticks CPU is failing to respond to the current grace
2751  * period and has not be idle from an RCU perspective, kick it.
2752  */
2753 static void __maybe_unused rcu_kick_nohz_cpu(int cpu)
2754 {
2755 #ifdef CONFIG_NO_HZ_FULL
2756         if (tick_nohz_full_cpu(cpu))
2757                 smp_send_reschedule(cpu);
2758 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2759 }
2760
2761
2762 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2763
2764 static int full_sysidle_state;          /* Current system-idle state. */
2765 #define RCU_SYSIDLE_NOT         0       /* Some CPU is not idle. */
2766 #define RCU_SYSIDLE_SHORT       1       /* All CPUs idle for brief period. */
2767 #define RCU_SYSIDLE_LONG        2       /* All CPUs idle for long enough. */
2768 #define RCU_SYSIDLE_FULL        3       /* All CPUs idle, ready for sysidle. */
2769 #define RCU_SYSIDLE_FULL_NOTED  4       /* Actually entered sysidle state. */
2770
2771 /*
2772  * Invoked to note exit from irq or task transition to idle.  Note that
2773  * usermode execution does -not- count as idle here!  After all, we want
2774  * to detect full-system idle states, not RCU quiescent states and grace
2775  * periods.  The caller must have disabled interrupts.
2776  */
2777 static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq)
2778 {
2779         unsigned long j;
2780
2781         /* If there are no nohz_full= CPUs, no need to track this. */
2782         if (!tick_nohz_full_enabled())
2783                 return;
2784
2785         /* Adjust nesting, check for fully idle. */
2786         if (irq) {
2787                 rdtp->dynticks_idle_nesting--;
2788                 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2789                 if (rdtp->dynticks_idle_nesting != 0)
2790                         return;  /* Still not fully idle. */
2791         } else {
2792                 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2793                     DYNTICK_TASK_NEST_VALUE) {
2794                         rdtp->dynticks_idle_nesting = 0;
2795                 } else {
2796                         rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2797                         WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2798                         return;  /* Still not fully idle. */
2799                 }
2800         }
2801
2802         /* Record start of fully idle period. */
2803         j = jiffies;
2804         ACCESS_ONCE(rdtp->dynticks_idle_jiffies) = j;
2805         smp_mb__before_atomic();
2806         atomic_inc(&rdtp->dynticks_idle);
2807         smp_mb__after_atomic();
2808         WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2809 }
2810
2811 /*
2812  * Unconditionally force exit from full system-idle state.  This is
2813  * invoked when a normal CPU exits idle, but must be called separately
2814  * for the timekeeping CPU (tick_do_timer_cpu).  The reason for this
2815  * is that the timekeeping CPU is permitted to take scheduling-clock
2816  * interrupts while the system is in system-idle state, and of course
2817  * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2818  * interrupt from any other type of interrupt.
2819  */
2820 void rcu_sysidle_force_exit(void)
2821 {
2822         int oldstate = ACCESS_ONCE(full_sysidle_state);
2823         int newoldstate;
2824
2825         /*
2826          * Each pass through the following loop attempts to exit full
2827          * system-idle state.  If contention proves to be a problem,
2828          * a trylock-based contention tree could be used here.
2829          */
2830         while (oldstate > RCU_SYSIDLE_SHORT) {
2831                 newoldstate = cmpxchg(&full_sysidle_state,
2832                                       oldstate, RCU_SYSIDLE_NOT);
2833                 if (oldstate == newoldstate &&
2834                     oldstate == RCU_SYSIDLE_FULL_NOTED) {
2835                         rcu_kick_nohz_cpu(tick_do_timer_cpu);
2836                         return; /* We cleared it, done! */
2837                 }
2838                 oldstate = newoldstate;
2839         }
2840         smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2841 }
2842
2843 /*
2844  * Invoked to note entry to irq or task transition from idle.  Note that
2845  * usermode execution does -not- count as idle here!  The caller must
2846  * have disabled interrupts.
2847  */
2848 static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq)
2849 {
2850         /* If there are no nohz_full= CPUs, no need to track this. */
2851         if (!tick_nohz_full_enabled())
2852                 return;
2853
2854         /* Adjust nesting, check for already non-idle. */
2855         if (irq) {
2856                 rdtp->dynticks_idle_nesting++;
2857                 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2858                 if (rdtp->dynticks_idle_nesting != 1)
2859                         return; /* Already non-idle. */
2860         } else {
2861                 /*
2862                  * Allow for irq misnesting.  Yes, it really is possible
2863                  * to enter an irq handler then never leave it, and maybe
2864                  * also vice versa.  Handle both possibilities.
2865                  */
2866                 if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2867                         rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2868                         WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2869                         return; /* Already non-idle. */
2870                 } else {
2871                         rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2872                 }
2873         }
2874
2875         /* Record end of idle period. */
2876         smp_mb__before_atomic();
2877         atomic_inc(&rdtp->dynticks_idle);
2878         smp_mb__after_atomic();
2879         WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2880
2881         /*
2882          * If we are the timekeeping CPU, we are permitted to be non-idle
2883          * during a system-idle state.  This must be the case, because
2884          * the timekeeping CPU has to take scheduling-clock interrupts
2885          * during the time that the system is transitioning to full
2886          * system-idle state.  This means that the timekeeping CPU must
2887          * invoke rcu_sysidle_force_exit() directly if it does anything
2888          * more than take a scheduling-clock interrupt.
2889          */
2890         if (smp_processor_id() == tick_do_timer_cpu)
2891                 return;
2892
2893         /* Update system-idle state: We are clearly no longer fully idle! */
2894         rcu_sysidle_force_exit();
2895 }
2896
2897 /*
2898  * Check to see if the current CPU is idle.  Note that usermode execution
2899  * does not count as idle.  The caller must have disabled interrupts.
2900  */
2901 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2902                                   unsigned long *maxj)
2903 {
2904         int cur;
2905         unsigned long j;
2906         struct rcu_dynticks *rdtp = rdp->dynticks;
2907
2908         /* If there are no nohz_full= CPUs, don't check system-wide idleness. */
2909         if (!tick_nohz_full_enabled())
2910                 return;
2911
2912         /*
2913          * If some other CPU has already reported non-idle, if this is
2914          * not the flavor of RCU that tracks sysidle state, or if this
2915          * is an offline or the timekeeping CPU, nothing to do.
2916          */
2917         if (!*isidle || rdp->rsp != rcu_state_p ||
2918             cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2919                 return;
2920         if (rcu_gp_in_progress(rdp->rsp))
2921                 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2922
2923         /* Pick up current idle and NMI-nesting counter and check. */
2924         cur = atomic_read(&rdtp->dynticks_idle);
2925         if (cur & 0x1) {
2926                 *isidle = false; /* We are not idle! */
2927                 return;
2928         }
2929         smp_mb(); /* Read counters before timestamps. */
2930
2931         /* Pick up timestamps. */
2932         j = ACCESS_ONCE(rdtp->dynticks_idle_jiffies);
2933         /* If this CPU entered idle more recently, update maxj timestamp. */
2934         if (ULONG_CMP_LT(*maxj, j))
2935                 *maxj = j;
2936 }
2937
2938 /*
2939  * Is this the flavor of RCU that is handling full-system idle?
2940  */
2941 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2942 {
2943         return rsp == rcu_state_p;
2944 }
2945
2946 /*
2947  * Return a delay in jiffies based on the number of CPUs, rcu_node
2948  * leaf fanout, and jiffies tick rate.  The idea is to allow larger
2949  * systems more time to transition to full-idle state in order to
2950  * avoid the cache thrashing that otherwise occur on the state variable.
2951  * Really small systems (less than a couple of tens of CPUs) should
2952  * instead use a single global atomically incremented counter, and later
2953  * versions of this will automatically reconfigure themselves accordingly.
2954  */
2955 static unsigned long rcu_sysidle_delay(void)
2956 {
2957         if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2958                 return 0;
2959         return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2960 }
2961
2962 /*
2963  * Advance the full-system-idle state.  This is invoked when all of
2964  * the non-timekeeping CPUs are idle.
2965  */
2966 static void rcu_sysidle(unsigned long j)
2967 {
2968         /* Check the current state. */
2969         switch (ACCESS_ONCE(full_sysidle_state)) {
2970         case RCU_SYSIDLE_NOT:
2971
2972                 /* First time all are idle, so note a short idle period. */
2973                 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_SHORT;
2974                 break;
2975
2976         case RCU_SYSIDLE_SHORT:
2977
2978                 /*
2979                  * Idle for a bit, time to advance to next state?
2980                  * cmpxchg failure means race with non-idle, let them win.
2981                  */
2982                 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2983                         (void)cmpxchg(&full_sysidle_state,
2984                                       RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2985                 break;
2986
2987         case RCU_SYSIDLE_LONG:
2988
2989                 /*
2990                  * Do an additional check pass before advancing to full.
2991                  * cmpxchg failure means race with non-idle, let them win.
2992                  */
2993                 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2994                         (void)cmpxchg(&full_sysidle_state,
2995                                       RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2996                 break;
2997
2998         default:
2999                 break;
3000         }
3001 }
3002
3003 /*
3004  * Found a non-idle non-timekeeping CPU, so kick the system-idle state
3005  * back to the beginning.
3006  */
3007 static void rcu_sysidle_cancel(void)
3008 {
3009         smp_mb();
3010         if (full_sysidle_state > RCU_SYSIDLE_SHORT)
3011                 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_NOT;
3012 }
3013
3014 /*
3015  * Update the sysidle state based on the results of a force-quiescent-state
3016  * scan of the CPUs' dyntick-idle state.
3017  */
3018 static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
3019                                unsigned long maxj, bool gpkt)
3020 {
3021         if (rsp != rcu_state_p)
3022                 return;  /* Wrong flavor, ignore. */
3023         if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
3024                 return;  /* Running state machine from timekeeping CPU. */
3025         if (isidle)
3026                 rcu_sysidle(maxj);    /* More idle! */
3027         else
3028                 rcu_sysidle_cancel(); /* Idle is over. */
3029 }
3030
3031 /*
3032  * Wrapper for rcu_sysidle_report() when called from the grace-period
3033  * kthread's context.
3034  */
3035 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
3036                                   unsigned long maxj)
3037 {
3038         /* If there are no nohz_full= CPUs, no need to track this. */
3039         if (!tick_nohz_full_enabled())
3040                 return;
3041
3042         rcu_sysidle_report(rsp, isidle, maxj, true);
3043 }
3044
3045 /* Callback and function for forcing an RCU grace period. */
3046 struct rcu_sysidle_head {
3047         struct rcu_head rh;
3048         int inuse;
3049 };
3050
3051 static void rcu_sysidle_cb(struct rcu_head *rhp)
3052 {
3053         struct rcu_sysidle_head *rshp;
3054
3055         /*
3056          * The following memory barrier is needed to replace the
3057          * memory barriers that would normally be in the memory
3058          * allocator.
3059          */
3060         smp_mb();  /* grace period precedes setting inuse. */
3061
3062         rshp = container_of(rhp, struct rcu_sysidle_head, rh);
3063         ACCESS_ONCE(rshp->inuse) = 0;
3064 }
3065
3066 /*
3067  * Check to see if the system is fully idle, other than the timekeeping CPU.
3068  * The caller must have disabled interrupts.  This is not intended to be
3069  * called unless tick_nohz_full_enabled().
3070  */
3071 bool rcu_sys_is_idle(void)
3072 {
3073         static struct rcu_sysidle_head rsh;
3074         int rss = ACCESS_ONCE(full_sysidle_state);
3075
3076         if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
3077                 return false;
3078
3079         /* Handle small-system case by doing a full scan of CPUs. */
3080         if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
3081                 int oldrss = rss - 1;
3082
3083                 /*
3084                  * One pass to advance to each state up to _FULL.
3085                  * Give up if any pass fails to advance the state.
3086                  */
3087                 while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
3088                         int cpu;
3089                         bool isidle = true;
3090                         unsigned long maxj = jiffies - ULONG_MAX / 4;
3091                         struct rcu_data *rdp;
3092
3093                         /* Scan all the CPUs looking for nonidle CPUs. */
3094                         for_each_possible_cpu(cpu) {
3095                                 rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3096                                 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
3097                                 if (!isidle)
3098                                         break;
3099                         }
3100                         rcu_sysidle_report(rcu_state_p, isidle, maxj, false);
3101                         oldrss = rss;
3102                         rss = ACCESS_ONCE(full_sysidle_state);
3103                 }
3104         }
3105
3106         /* If this is the first observation of an idle period, record it. */
3107         if (rss == RCU_SYSIDLE_FULL) {
3108                 rss = cmpxchg(&full_sysidle_state,
3109                               RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
3110                 return rss == RCU_SYSIDLE_FULL;
3111         }
3112
3113         smp_mb(); /* ensure rss load happens before later caller actions. */
3114
3115         /* If already fully idle, tell the caller (in case of races). */
3116         if (rss == RCU_SYSIDLE_FULL_NOTED)
3117                 return true;
3118
3119         /*
3120          * If we aren't there yet, and a grace period is not in flight,
3121          * initiate a grace period.  Either way, tell the caller that
3122          * we are not there yet.  We use an xchg() rather than an assignment
3123          * to make up for the memory barriers that would otherwise be
3124          * provided by the memory allocator.
3125          */
3126         if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
3127             !rcu_gp_in_progress(rcu_state_p) &&
3128             !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
3129                 call_rcu(&rsh.rh, rcu_sysidle_cb);
3130         return false;
3131 }
3132
3133 /*
3134  * Initialize dynticks sysidle state for CPUs coming online.
3135  */
3136 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
3137 {
3138         rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
3139 }
3140
3141 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3142
3143 static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq)
3144 {
3145 }
3146
3147 static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq)
3148 {
3149 }
3150
3151 static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
3152                                   unsigned long *maxj)
3153 {
3154 }
3155
3156 static bool is_sysidle_rcu_state(struct rcu_state *rsp)
3157 {
3158         return false;
3159 }
3160
3161 static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
3162                                   unsigned long maxj)
3163 {
3164 }
3165
3166 static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
3167 {
3168 }
3169
3170 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3171
3172 /*
3173  * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
3174  * grace-period kthread will do force_quiescent_state() processing?
3175  * The idea is to avoid waking up RCU core processing on such a
3176  * CPU unless the grace period has extended for too long.
3177  *
3178  * This code relies on the fact that all NO_HZ_FULL CPUs are also
3179  * CONFIG_RCU_NOCB_CPU CPUs.
3180  */
3181 static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
3182 {
3183 #ifdef CONFIG_NO_HZ_FULL
3184         if (tick_nohz_full_cpu(smp_processor_id()) &&
3185             (!rcu_gp_in_progress(rsp) ||
3186              ULONG_CMP_LT(jiffies, ACCESS_ONCE(rsp->gp_start) + HZ)))
3187                 return 1;
3188 #endif /* #ifdef CONFIG_NO_HZ_FULL */
3189         return 0;
3190 }
3191
3192 /*
3193  * Bind the grace-period kthread for the sysidle flavor of RCU to the
3194  * timekeeping CPU.
3195  */
3196 static void rcu_bind_gp_kthread(void)
3197 {
3198         int __maybe_unused cpu;
3199
3200         if (!tick_nohz_full_enabled())
3201                 return;
3202 #ifdef CONFIG_NO_HZ_FULL_SYSIDLE
3203         cpu = tick_do_timer_cpu;
3204         if (cpu >= 0 && cpu < nr_cpu_ids && raw_smp_processor_id() != cpu)
3205                 set_cpus_allowed_ptr(current, cpumask_of(cpu));
3206 #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3207         if (!is_housekeeping_cpu(raw_smp_processor_id()))
3208                 housekeeping_affine(current);
3209 #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
3210 }
3211
3212 /* Record the current task on dyntick-idle entry. */
3213 static void rcu_dynticks_task_enter(void)
3214 {
3215 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3216         ACCESS_ONCE(current->rcu_tasks_idle_cpu) = smp_processor_id();
3217 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
3218 }
3219
3220 /* Record no current task on dyntick-idle exit. */
3221 static void rcu_dynticks_task_exit(void)
3222 {
3223 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
3224         ACCESS_ONCE(current->rcu_tasks_idle_cpu) = -1;
3225 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
3226 }