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