Loading...
Note: File does not exist in v4.6.
1/* SPDX-License-Identifier: GPL-2.0+ */
2/*
3 * RCU expedited grace periods
4 *
5 * Copyright IBM Corporation, 2016
6 *
7 * Authors: Paul E. McKenney <paulmck@linux.ibm.com>
8 */
9
10#include <linux/lockdep.h>
11
12static void rcu_exp_handler(void *unused);
13static int rcu_print_task_exp_stall(struct rcu_node *rnp);
14static void rcu_exp_print_detail_task_stall_rnp(struct rcu_node *rnp);
15
16/*
17 * Record the start of an expedited grace period.
18 */
19static void rcu_exp_gp_seq_start(void)
20{
21 rcu_seq_start(&rcu_state.expedited_sequence);
22 rcu_poll_gp_seq_start_unlocked(&rcu_state.gp_seq_polled_exp_snap);
23}
24
25/*
26 * Return the value that the expedited-grace-period counter will have
27 * at the end of the current grace period.
28 */
29static __maybe_unused unsigned long rcu_exp_gp_seq_endval(void)
30{
31 return rcu_seq_endval(&rcu_state.expedited_sequence);
32}
33
34/*
35 * Record the end of an expedited grace period.
36 */
37static void rcu_exp_gp_seq_end(void)
38{
39 rcu_poll_gp_seq_end_unlocked(&rcu_state.gp_seq_polled_exp_snap);
40 rcu_seq_end(&rcu_state.expedited_sequence);
41 smp_mb(); /* Ensure that consecutive grace periods serialize. */
42}
43
44/*
45 * Take a snapshot of the expedited-grace-period counter, which is the
46 * earliest value that will indicate that a full grace period has
47 * elapsed since the current time.
48 */
49static unsigned long rcu_exp_gp_seq_snap(void)
50{
51 unsigned long s;
52
53 smp_mb(); /* Caller's modifications seen first by other CPUs. */
54 s = rcu_seq_snap(&rcu_state.expedited_sequence);
55 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("snap"));
56 return s;
57}
58
59/*
60 * Given a counter snapshot from rcu_exp_gp_seq_snap(), return true
61 * if a full expedited grace period has elapsed since that snapshot
62 * was taken.
63 */
64static bool rcu_exp_gp_seq_done(unsigned long s)
65{
66 return rcu_seq_done(&rcu_state.expedited_sequence, s);
67}
68
69/*
70 * Reset the ->expmaskinit values in the rcu_node tree to reflect any
71 * recent CPU-online activity. Note that these masks are not cleared
72 * when CPUs go offline, so they reflect the union of all CPUs that have
73 * ever been online. This means that this function normally takes its
74 * no-work-to-do fastpath.
75 */
76static void sync_exp_reset_tree_hotplug(void)
77{
78 bool done;
79 unsigned long flags;
80 unsigned long mask;
81 unsigned long oldmask;
82 int ncpus = smp_load_acquire(&rcu_state.ncpus); /* Order vs. locking. */
83 struct rcu_node *rnp;
84 struct rcu_node *rnp_up;
85
86 /* If no new CPUs onlined since last time, nothing to do. */
87 if (likely(ncpus == rcu_state.ncpus_snap))
88 return;
89 rcu_state.ncpus_snap = ncpus;
90
91 /*
92 * Each pass through the following loop propagates newly onlined
93 * CPUs for the current rcu_node structure up the rcu_node tree.
94 */
95 rcu_for_each_leaf_node(rnp) {
96 raw_spin_lock_irqsave_rcu_node(rnp, flags);
97 if (rnp->expmaskinit == rnp->expmaskinitnext) {
98 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
99 continue; /* No new CPUs, nothing to do. */
100 }
101
102 /* Update this node's mask, track old value for propagation. */
103 oldmask = rnp->expmaskinit;
104 rnp->expmaskinit = rnp->expmaskinitnext;
105 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
106
107 /* If was already nonzero, nothing to propagate. */
108 if (oldmask)
109 continue;
110
111 /* Propagate the new CPU up the tree. */
112 mask = rnp->grpmask;
113 rnp_up = rnp->parent;
114 done = false;
115 while (rnp_up) {
116 raw_spin_lock_irqsave_rcu_node(rnp_up, flags);
117 if (rnp_up->expmaskinit)
118 done = true;
119 rnp_up->expmaskinit |= mask;
120 raw_spin_unlock_irqrestore_rcu_node(rnp_up, flags);
121 if (done)
122 break;
123 mask = rnp_up->grpmask;
124 rnp_up = rnp_up->parent;
125 }
126 }
127}
128
129/*
130 * Reset the ->expmask values in the rcu_node tree in preparation for
131 * a new expedited grace period.
132 */
133static void __maybe_unused sync_exp_reset_tree(void)
134{
135 unsigned long flags;
136 struct rcu_node *rnp;
137
138 sync_exp_reset_tree_hotplug();
139 rcu_for_each_node_breadth_first(rnp) {
140 raw_spin_lock_irqsave_rcu_node(rnp, flags);
141 WARN_ON_ONCE(rnp->expmask);
142 WRITE_ONCE(rnp->expmask, rnp->expmaskinit);
143 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
144 }
145}
146
147/*
148 * Return non-zero if there is no RCU expedited grace period in progress
149 * for the specified rcu_node structure, in other words, if all CPUs and
150 * tasks covered by the specified rcu_node structure have done their bit
151 * for the current expedited grace period.
152 */
153static bool sync_rcu_exp_done(struct rcu_node *rnp)
154{
155 raw_lockdep_assert_held_rcu_node(rnp);
156 return READ_ONCE(rnp->exp_tasks) == NULL &&
157 READ_ONCE(rnp->expmask) == 0;
158}
159
160/*
161 * Like sync_rcu_exp_done(), but where the caller does not hold the
162 * rcu_node's ->lock.
163 */
164static bool sync_rcu_exp_done_unlocked(struct rcu_node *rnp)
165{
166 unsigned long flags;
167 bool ret;
168
169 raw_spin_lock_irqsave_rcu_node(rnp, flags);
170 ret = sync_rcu_exp_done(rnp);
171 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
172
173 return ret;
174}
175
176/*
177 * Report the exit from RCU read-side critical section for the last task
178 * that queued itself during or before the current expedited preemptible-RCU
179 * grace period. This event is reported either to the rcu_node structure on
180 * which the task was queued or to one of that rcu_node structure's ancestors,
181 * recursively up the tree. (Calm down, calm down, we do the recursion
182 * iteratively!)
183 */
184static void __rcu_report_exp_rnp(struct rcu_node *rnp,
185 bool wake, unsigned long flags)
186 __releases(rnp->lock)
187{
188 unsigned long mask;
189
190 raw_lockdep_assert_held_rcu_node(rnp);
191 for (;;) {
192 if (!sync_rcu_exp_done(rnp)) {
193 if (!rnp->expmask)
194 rcu_initiate_boost(rnp, flags);
195 else
196 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
197 break;
198 }
199 if (rnp->parent == NULL) {
200 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
201 if (wake) {
202 smp_mb(); /* EGP done before wake_up(). */
203 swake_up_one_online(&rcu_state.expedited_wq);
204 }
205 break;
206 }
207 mask = rnp->grpmask;
208 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled */
209 rnp = rnp->parent;
210 raw_spin_lock_rcu_node(rnp); /* irqs already disabled */
211 WARN_ON_ONCE(!(rnp->expmask & mask));
212 WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask);
213 }
214}
215
216/*
217 * Report expedited quiescent state for specified node. This is a
218 * lock-acquisition wrapper function for __rcu_report_exp_rnp().
219 */
220static void __maybe_unused rcu_report_exp_rnp(struct rcu_node *rnp, bool wake)
221{
222 unsigned long flags;
223
224 raw_spin_lock_irqsave_rcu_node(rnp, flags);
225 __rcu_report_exp_rnp(rnp, wake, flags);
226}
227
228/*
229 * Report expedited quiescent state for multiple CPUs, all covered by the
230 * specified leaf rcu_node structure.
231 */
232static void rcu_report_exp_cpu_mult(struct rcu_node *rnp,
233 unsigned long mask, bool wake)
234{
235 int cpu;
236 unsigned long flags;
237 struct rcu_data *rdp;
238
239 raw_spin_lock_irqsave_rcu_node(rnp, flags);
240 if (!(rnp->expmask & mask)) {
241 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
242 return;
243 }
244 WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask);
245 for_each_leaf_node_cpu_mask(rnp, cpu, mask) {
246 rdp = per_cpu_ptr(&rcu_data, cpu);
247 if (!IS_ENABLED(CONFIG_NO_HZ_FULL) || !rdp->rcu_forced_tick_exp)
248 continue;
249 rdp->rcu_forced_tick_exp = false;
250 tick_dep_clear_cpu(cpu, TICK_DEP_BIT_RCU_EXP);
251 }
252 __rcu_report_exp_rnp(rnp, wake, flags); /* Releases rnp->lock. */
253}
254
255/*
256 * Report expedited quiescent state for specified rcu_data (CPU).
257 */
258static void rcu_report_exp_rdp(struct rcu_data *rdp)
259{
260 WRITE_ONCE(rdp->cpu_no_qs.b.exp, false);
261 rcu_report_exp_cpu_mult(rdp->mynode, rdp->grpmask, true);
262}
263
264/* Common code for work-done checking. */
265static bool sync_exp_work_done(unsigned long s)
266{
267 if (rcu_exp_gp_seq_done(s)) {
268 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("done"));
269 smp_mb(); /* Ensure test happens before caller kfree(). */
270 return true;
271 }
272 return false;
273}
274
275/*
276 * Funnel-lock acquisition for expedited grace periods. Returns true
277 * if some other task completed an expedited grace period that this task
278 * can piggy-back on, and with no mutex held. Otherwise, returns false
279 * with the mutex held, indicating that the caller must actually do the
280 * expedited grace period.
281 */
282static bool exp_funnel_lock(unsigned long s)
283{
284 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, raw_smp_processor_id());
285 struct rcu_node *rnp = rdp->mynode;
286 struct rcu_node *rnp_root = rcu_get_root();
287
288 /* Low-contention fastpath. */
289 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s) &&
290 (rnp == rnp_root ||
291 ULONG_CMP_LT(READ_ONCE(rnp_root->exp_seq_rq), s)) &&
292 mutex_trylock(&rcu_state.exp_mutex))
293 goto fastpath;
294
295 /*
296 * Each pass through the following loop works its way up
297 * the rcu_node tree, returning if others have done the work or
298 * otherwise falls through to acquire ->exp_mutex. The mapping
299 * from CPU to rcu_node structure can be inexact, as it is just
300 * promoting locality and is not strictly needed for correctness.
301 */
302 for (; rnp != NULL; rnp = rnp->parent) {
303 if (sync_exp_work_done(s))
304 return true;
305
306 /* Work not done, either wait here or go up. */
307 spin_lock(&rnp->exp_lock);
308 if (ULONG_CMP_GE(rnp->exp_seq_rq, s)) {
309
310 /* Someone else doing GP, so wait for them. */
311 spin_unlock(&rnp->exp_lock);
312 trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
313 rnp->grplo, rnp->grphi,
314 TPS("wait"));
315 wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
316 sync_exp_work_done(s));
317 return true;
318 }
319 WRITE_ONCE(rnp->exp_seq_rq, s); /* Followers can wait on us. */
320 spin_unlock(&rnp->exp_lock);
321 trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
322 rnp->grplo, rnp->grphi, TPS("nxtlvl"));
323 }
324 mutex_lock(&rcu_state.exp_mutex);
325fastpath:
326 if (sync_exp_work_done(s)) {
327 mutex_unlock(&rcu_state.exp_mutex);
328 return true;
329 }
330 rcu_exp_gp_seq_start();
331 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("start"));
332 return false;
333}
334
335/*
336 * Select the CPUs within the specified rcu_node that the upcoming
337 * expedited grace period needs to wait for.
338 */
339static void __sync_rcu_exp_select_node_cpus(struct rcu_exp_work *rewp)
340{
341 int cpu;
342 unsigned long flags;
343 unsigned long mask_ofl_test;
344 unsigned long mask_ofl_ipi;
345 int ret;
346 struct rcu_node *rnp = container_of(rewp, struct rcu_node, rew);
347
348 raw_spin_lock_irqsave_rcu_node(rnp, flags);
349
350 /* Each pass checks a CPU for identity, offline, and idle. */
351 mask_ofl_test = 0;
352 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) {
353 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
354 unsigned long mask = rdp->grpmask;
355 int snap;
356
357 if (raw_smp_processor_id() == cpu ||
358 !(rnp->qsmaskinitnext & mask)) {
359 mask_ofl_test |= mask;
360 } else {
361 snap = rcu_dynticks_snap(cpu);
362 if (rcu_dynticks_in_eqs(snap))
363 mask_ofl_test |= mask;
364 else
365 rdp->exp_dynticks_snap = snap;
366 }
367 }
368 mask_ofl_ipi = rnp->expmask & ~mask_ofl_test;
369
370 /*
371 * Need to wait for any blocked tasks as well. Note that
372 * additional blocking tasks will also block the expedited GP
373 * until such time as the ->expmask bits are cleared.
374 */
375 if (rcu_preempt_has_tasks(rnp))
376 WRITE_ONCE(rnp->exp_tasks, rnp->blkd_tasks.next);
377 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
378
379 /* IPI the remaining CPUs for expedited quiescent state. */
380 for_each_leaf_node_cpu_mask(rnp, cpu, mask_ofl_ipi) {
381 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
382 unsigned long mask = rdp->grpmask;
383
384retry_ipi:
385 if (rcu_dynticks_in_eqs_since(rdp, rdp->exp_dynticks_snap)) {
386 mask_ofl_test |= mask;
387 continue;
388 }
389 if (get_cpu() == cpu) {
390 mask_ofl_test |= mask;
391 put_cpu();
392 continue;
393 }
394 ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
395 put_cpu();
396 /* The CPU will report the QS in response to the IPI. */
397 if (!ret)
398 continue;
399
400 /* Failed, raced with CPU hotplug operation. */
401 raw_spin_lock_irqsave_rcu_node(rnp, flags);
402 if ((rnp->qsmaskinitnext & mask) &&
403 (rnp->expmask & mask)) {
404 /* Online, so delay for a bit and try again. */
405 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
406 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("selectofl"));
407 schedule_timeout_idle(1);
408 goto retry_ipi;
409 }
410 /* CPU really is offline, so we must report its QS. */
411 if (rnp->expmask & mask)
412 mask_ofl_test |= mask;
413 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
414 }
415 /* Report quiescent states for those that went offline. */
416 if (mask_ofl_test)
417 rcu_report_exp_cpu_mult(rnp, mask_ofl_test, false);
418}
419
420static void rcu_exp_sel_wait_wake(unsigned long s);
421
422#ifdef CONFIG_RCU_EXP_KTHREAD
423static void sync_rcu_exp_select_node_cpus(struct kthread_work *wp)
424{
425 struct rcu_exp_work *rewp =
426 container_of(wp, struct rcu_exp_work, rew_work);
427
428 __sync_rcu_exp_select_node_cpus(rewp);
429}
430
431static inline bool rcu_gp_par_worker_started(void)
432{
433 return !!READ_ONCE(rcu_exp_par_gp_kworker);
434}
435
436static inline void sync_rcu_exp_select_cpus_queue_work(struct rcu_node *rnp)
437{
438 kthread_init_work(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus);
439 /*
440 * Use rcu_exp_par_gp_kworker, because flushing a work item from
441 * another work item on the same kthread worker can result in
442 * deadlock.
443 */
444 kthread_queue_work(rcu_exp_par_gp_kworker, &rnp->rew.rew_work);
445}
446
447static inline void sync_rcu_exp_select_cpus_flush_work(struct rcu_node *rnp)
448{
449 kthread_flush_work(&rnp->rew.rew_work);
450}
451
452/*
453 * Work-queue handler to drive an expedited grace period forward.
454 */
455static void wait_rcu_exp_gp(struct kthread_work *wp)
456{
457 struct rcu_exp_work *rewp;
458
459 rewp = container_of(wp, struct rcu_exp_work, rew_work);
460 rcu_exp_sel_wait_wake(rewp->rew_s);
461}
462
463static inline void synchronize_rcu_expedited_queue_work(struct rcu_exp_work *rew)
464{
465 kthread_init_work(&rew->rew_work, wait_rcu_exp_gp);
466 kthread_queue_work(rcu_exp_gp_kworker, &rew->rew_work);
467}
468
469static inline void synchronize_rcu_expedited_destroy_work(struct rcu_exp_work *rew)
470{
471}
472#else /* !CONFIG_RCU_EXP_KTHREAD */
473static void sync_rcu_exp_select_node_cpus(struct work_struct *wp)
474{
475 struct rcu_exp_work *rewp =
476 container_of(wp, struct rcu_exp_work, rew_work);
477
478 __sync_rcu_exp_select_node_cpus(rewp);
479}
480
481static inline bool rcu_gp_par_worker_started(void)
482{
483 return !!READ_ONCE(rcu_par_gp_wq);
484}
485
486static inline void sync_rcu_exp_select_cpus_queue_work(struct rcu_node *rnp)
487{
488 int cpu = find_next_bit(&rnp->ffmask, BITS_PER_LONG, -1);
489
490 INIT_WORK(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus);
491 /* If all offline, queue the work on an unbound CPU. */
492 if (unlikely(cpu > rnp->grphi - rnp->grplo))
493 cpu = WORK_CPU_UNBOUND;
494 else
495 cpu += rnp->grplo;
496 queue_work_on(cpu, rcu_par_gp_wq, &rnp->rew.rew_work);
497}
498
499static inline void sync_rcu_exp_select_cpus_flush_work(struct rcu_node *rnp)
500{
501 flush_work(&rnp->rew.rew_work);
502}
503
504/*
505 * Work-queue handler to drive an expedited grace period forward.
506 */
507static void wait_rcu_exp_gp(struct work_struct *wp)
508{
509 struct rcu_exp_work *rewp;
510
511 rewp = container_of(wp, struct rcu_exp_work, rew_work);
512 rcu_exp_sel_wait_wake(rewp->rew_s);
513}
514
515static inline void synchronize_rcu_expedited_queue_work(struct rcu_exp_work *rew)
516{
517 INIT_WORK_ONSTACK(&rew->rew_work, wait_rcu_exp_gp);
518 queue_work(rcu_gp_wq, &rew->rew_work);
519}
520
521static inline void synchronize_rcu_expedited_destroy_work(struct rcu_exp_work *rew)
522{
523 destroy_work_on_stack(&rew->rew_work);
524}
525#endif /* CONFIG_RCU_EXP_KTHREAD */
526
527/*
528 * Select the nodes that the upcoming expedited grace period needs
529 * to wait for.
530 */
531static void sync_rcu_exp_select_cpus(void)
532{
533 struct rcu_node *rnp;
534
535 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("reset"));
536 sync_exp_reset_tree();
537 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("select"));
538
539 /* Schedule work for each leaf rcu_node structure. */
540 rcu_for_each_leaf_node(rnp) {
541 rnp->exp_need_flush = false;
542 if (!READ_ONCE(rnp->expmask))
543 continue; /* Avoid early boot non-existent wq. */
544 if (!rcu_gp_par_worker_started() ||
545 rcu_scheduler_active != RCU_SCHEDULER_RUNNING ||
546 rcu_is_last_leaf_node(rnp)) {
547 /* No worker started yet or last leaf, do direct call. */
548 sync_rcu_exp_select_node_cpus(&rnp->rew.rew_work);
549 continue;
550 }
551 sync_rcu_exp_select_cpus_queue_work(rnp);
552 rnp->exp_need_flush = true;
553 }
554
555 /* Wait for jobs (if any) to complete. */
556 rcu_for_each_leaf_node(rnp)
557 if (rnp->exp_need_flush)
558 sync_rcu_exp_select_cpus_flush_work(rnp);
559}
560
561/*
562 * Wait for the expedited grace period to elapse, within time limit.
563 * If the time limit is exceeded without the grace period elapsing,
564 * return false, otherwise return true.
565 */
566static bool synchronize_rcu_expedited_wait_once(long tlimit)
567{
568 int t;
569 struct rcu_node *rnp_root = rcu_get_root();
570
571 t = swait_event_timeout_exclusive(rcu_state.expedited_wq,
572 sync_rcu_exp_done_unlocked(rnp_root),
573 tlimit);
574 // Workqueues should not be signaled.
575 if (t > 0 || sync_rcu_exp_done_unlocked(rnp_root))
576 return true;
577 WARN_ON(t < 0); /* workqueues should not be signaled. */
578 return false;
579}
580
581/*
582 * Wait for the expedited grace period to elapse, issuing any needed
583 * RCU CPU stall warnings along the way.
584 */
585static void synchronize_rcu_expedited_wait(void)
586{
587 int cpu;
588 unsigned long j;
589 unsigned long jiffies_stall;
590 unsigned long jiffies_start;
591 unsigned long mask;
592 int ndetected;
593 struct rcu_data *rdp;
594 struct rcu_node *rnp;
595 struct rcu_node *rnp_root = rcu_get_root();
596 unsigned long flags;
597
598 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("startwait"));
599 jiffies_stall = rcu_exp_jiffies_till_stall_check();
600 jiffies_start = jiffies;
601 if (tick_nohz_full_enabled() && rcu_inkernel_boot_has_ended()) {
602 if (synchronize_rcu_expedited_wait_once(1))
603 return;
604 rcu_for_each_leaf_node(rnp) {
605 raw_spin_lock_irqsave_rcu_node(rnp, flags);
606 mask = READ_ONCE(rnp->expmask);
607 for_each_leaf_node_cpu_mask(rnp, cpu, mask) {
608 rdp = per_cpu_ptr(&rcu_data, cpu);
609 if (rdp->rcu_forced_tick_exp)
610 continue;
611 rdp->rcu_forced_tick_exp = true;
612 if (cpu_online(cpu))
613 tick_dep_set_cpu(cpu, TICK_DEP_BIT_RCU_EXP);
614 }
615 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
616 }
617 j = READ_ONCE(jiffies_till_first_fqs);
618 if (synchronize_rcu_expedited_wait_once(j + HZ))
619 return;
620 }
621
622 for (;;) {
623 unsigned long j;
624
625 if (synchronize_rcu_expedited_wait_once(jiffies_stall))
626 return;
627 if (rcu_stall_is_suppressed())
628 continue;
629 j = jiffies;
630 rcu_stall_notifier_call_chain(RCU_STALL_NOTIFY_EXP, (void *)(j - jiffies_start));
631 trace_rcu_stall_warning(rcu_state.name, TPS("ExpeditedStall"));
632 pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
633 rcu_state.name);
634 ndetected = 0;
635 rcu_for_each_leaf_node(rnp) {
636 ndetected += rcu_print_task_exp_stall(rnp);
637 for_each_leaf_node_possible_cpu(rnp, cpu) {
638 struct rcu_data *rdp;
639
640 mask = leaf_node_cpu_bit(rnp, cpu);
641 if (!(READ_ONCE(rnp->expmask) & mask))
642 continue;
643 ndetected++;
644 rdp = per_cpu_ptr(&rcu_data, cpu);
645 pr_cont(" %d-%c%c%c%c", cpu,
646 "O."[!!cpu_online(cpu)],
647 "o."[!!(rdp->grpmask & rnp->expmaskinit)],
648 "N."[!!(rdp->grpmask & rnp->expmaskinitnext)],
649 "D."[!!data_race(rdp->cpu_no_qs.b.exp)]);
650 }
651 }
652 pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n",
653 j - jiffies_start, rcu_state.expedited_sequence,
654 data_race(rnp_root->expmask),
655 ".T"[!!data_race(rnp_root->exp_tasks)]);
656 if (ndetected) {
657 pr_err("blocking rcu_node structures (internal RCU debug):");
658 rcu_for_each_node_breadth_first(rnp) {
659 if (rnp == rnp_root)
660 continue; /* printed unconditionally */
661 if (sync_rcu_exp_done_unlocked(rnp))
662 continue;
663 pr_cont(" l=%u:%d-%d:%#lx/%c",
664 rnp->level, rnp->grplo, rnp->grphi,
665 data_race(rnp->expmask),
666 ".T"[!!data_race(rnp->exp_tasks)]);
667 }
668 pr_cont("\n");
669 }
670 rcu_for_each_leaf_node(rnp) {
671 for_each_leaf_node_possible_cpu(rnp, cpu) {
672 mask = leaf_node_cpu_bit(rnp, cpu);
673 if (!(READ_ONCE(rnp->expmask) & mask))
674 continue;
675 preempt_disable(); // For smp_processor_id() in dump_cpu_task().
676 dump_cpu_task(cpu);
677 preempt_enable();
678 }
679 rcu_exp_print_detail_task_stall_rnp(rnp);
680 }
681 jiffies_stall = 3 * rcu_exp_jiffies_till_stall_check() + 3;
682 panic_on_rcu_stall();
683 }
684}
685
686/*
687 * Wait for the current expedited grace period to complete, and then
688 * wake up everyone who piggybacked on the just-completed expedited
689 * grace period. Also update all the ->exp_seq_rq counters as needed
690 * in order to avoid counter-wrap problems.
691 */
692static void rcu_exp_wait_wake(unsigned long s)
693{
694 struct rcu_node *rnp;
695
696 synchronize_rcu_expedited_wait();
697
698 // Switch over to wakeup mode, allowing the next GP to proceed.
699 // End the previous grace period only after acquiring the mutex
700 // to ensure that only one GP runs concurrently with wakeups.
701 mutex_lock(&rcu_state.exp_wake_mutex);
702 rcu_exp_gp_seq_end();
703 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("end"));
704
705 rcu_for_each_node_breadth_first(rnp) {
706 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s)) {
707 spin_lock(&rnp->exp_lock);
708 /* Recheck, avoid hang in case someone just arrived. */
709 if (ULONG_CMP_LT(rnp->exp_seq_rq, s))
710 WRITE_ONCE(rnp->exp_seq_rq, s);
711 spin_unlock(&rnp->exp_lock);
712 }
713 smp_mb(); /* All above changes before wakeup. */
714 wake_up_all(&rnp->exp_wq[rcu_seq_ctr(s) & 0x3]);
715 }
716 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("endwake"));
717 mutex_unlock(&rcu_state.exp_wake_mutex);
718}
719
720/*
721 * Common code to drive an expedited grace period forward, used by
722 * workqueues and mid-boot-time tasks.
723 */
724static void rcu_exp_sel_wait_wake(unsigned long s)
725{
726 /* Initialize the rcu_node tree in preparation for the wait. */
727 sync_rcu_exp_select_cpus();
728
729 /* Wait and clean up, including waking everyone. */
730 rcu_exp_wait_wake(s);
731}
732
733#ifdef CONFIG_PREEMPT_RCU
734
735/*
736 * Remote handler for smp_call_function_single(). If there is an
737 * RCU read-side critical section in effect, request that the
738 * next rcu_read_unlock() record the quiescent state up the
739 * ->expmask fields in the rcu_node tree. Otherwise, immediately
740 * report the quiescent state.
741 */
742static void rcu_exp_handler(void *unused)
743{
744 int depth = rcu_preempt_depth();
745 unsigned long flags;
746 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
747 struct rcu_node *rnp = rdp->mynode;
748 struct task_struct *t = current;
749
750 /*
751 * First, the common case of not being in an RCU read-side
752 * critical section. If also enabled or idle, immediately
753 * report the quiescent state, otherwise defer.
754 */
755 if (!depth) {
756 if (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK)) ||
757 rcu_is_cpu_rrupt_from_idle()) {
758 rcu_report_exp_rdp(rdp);
759 } else {
760 WRITE_ONCE(rdp->cpu_no_qs.b.exp, true);
761 set_tsk_need_resched(t);
762 set_preempt_need_resched();
763 }
764 return;
765 }
766
767 /*
768 * Second, the less-common case of being in an RCU read-side
769 * critical section. In this case we can count on a future
770 * rcu_read_unlock(). However, this rcu_read_unlock() might
771 * execute on some other CPU, but in that case there will be
772 * a future context switch. Either way, if the expedited
773 * grace period is still waiting on this CPU, set ->deferred_qs
774 * so that the eventual quiescent state will be reported.
775 * Note that there is a large group of race conditions that
776 * can have caused this quiescent state to already have been
777 * reported, so we really do need to check ->expmask.
778 */
779 if (depth > 0) {
780 raw_spin_lock_irqsave_rcu_node(rnp, flags);
781 if (rnp->expmask & rdp->grpmask) {
782 WRITE_ONCE(rdp->cpu_no_qs.b.exp, true);
783 t->rcu_read_unlock_special.b.exp_hint = true;
784 }
785 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
786 return;
787 }
788
789 // Finally, negative nesting depth should not happen.
790 WARN_ON_ONCE(1);
791}
792
793/* PREEMPTION=y, so no PREEMPTION=n expedited grace period to clean up after. */
794static void sync_sched_exp_online_cleanup(int cpu)
795{
796}
797
798/*
799 * Scan the current list of tasks blocked within RCU read-side critical
800 * sections, printing out the tid of each that is blocking the current
801 * expedited grace period.
802 */
803static int rcu_print_task_exp_stall(struct rcu_node *rnp)
804{
805 unsigned long flags;
806 int ndetected = 0;
807 struct task_struct *t;
808
809 raw_spin_lock_irqsave_rcu_node(rnp, flags);
810 if (!rnp->exp_tasks) {
811 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
812 return 0;
813 }
814 t = list_entry(rnp->exp_tasks->prev,
815 struct task_struct, rcu_node_entry);
816 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
817 pr_cont(" P%d", t->pid);
818 ndetected++;
819 }
820 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
821 return ndetected;
822}
823
824/*
825 * Scan the current list of tasks blocked within RCU read-side critical
826 * sections, dumping the stack of each that is blocking the current
827 * expedited grace period.
828 */
829static void rcu_exp_print_detail_task_stall_rnp(struct rcu_node *rnp)
830{
831 unsigned long flags;
832 struct task_struct *t;
833
834 if (!rcu_exp_stall_task_details)
835 return;
836 raw_spin_lock_irqsave_rcu_node(rnp, flags);
837 if (!READ_ONCE(rnp->exp_tasks)) {
838 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
839 return;
840 }
841 t = list_entry(rnp->exp_tasks->prev,
842 struct task_struct, rcu_node_entry);
843 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
844 /*
845 * We could be printing a lot while holding a spinlock.
846 * Avoid triggering hard lockup.
847 */
848 touch_nmi_watchdog();
849 sched_show_task(t);
850 }
851 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
852}
853
854#else /* #ifdef CONFIG_PREEMPT_RCU */
855
856/* Request an expedited quiescent state. */
857static void rcu_exp_need_qs(void)
858{
859 __this_cpu_write(rcu_data.cpu_no_qs.b.exp, true);
860 /* Store .exp before .rcu_urgent_qs. */
861 smp_store_release(this_cpu_ptr(&rcu_data.rcu_urgent_qs), true);
862 set_tsk_need_resched(current);
863 set_preempt_need_resched();
864}
865
866/* Invoked on each online non-idle CPU for expedited quiescent state. */
867static void rcu_exp_handler(void *unused)
868{
869 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
870 struct rcu_node *rnp = rdp->mynode;
871 bool preempt_bh_enabled = !(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
872
873 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
874 __this_cpu_read(rcu_data.cpu_no_qs.b.exp))
875 return;
876 if (rcu_is_cpu_rrupt_from_idle() ||
877 (IS_ENABLED(CONFIG_PREEMPT_COUNT) && preempt_bh_enabled)) {
878 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
879 return;
880 }
881 rcu_exp_need_qs();
882}
883
884/* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
885static void sync_sched_exp_online_cleanup(int cpu)
886{
887 unsigned long flags;
888 int my_cpu;
889 struct rcu_data *rdp;
890 int ret;
891 struct rcu_node *rnp;
892
893 rdp = per_cpu_ptr(&rcu_data, cpu);
894 rnp = rdp->mynode;
895 my_cpu = get_cpu();
896 /* Quiescent state either not needed or already requested, leave. */
897 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
898 READ_ONCE(rdp->cpu_no_qs.b.exp)) {
899 put_cpu();
900 return;
901 }
902 /* Quiescent state needed on current CPU, so set it up locally. */
903 if (my_cpu == cpu) {
904 local_irq_save(flags);
905 rcu_exp_need_qs();
906 local_irq_restore(flags);
907 put_cpu();
908 return;
909 }
910 /* Quiescent state needed on some other CPU, send IPI. */
911 ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
912 put_cpu();
913 WARN_ON_ONCE(ret);
914}
915
916/*
917 * Because preemptible RCU does not exist, we never have to check for
918 * tasks blocked within RCU read-side critical sections that are
919 * blocking the current expedited grace period.
920 */
921static int rcu_print_task_exp_stall(struct rcu_node *rnp)
922{
923 return 0;
924}
925
926/*
927 * Because preemptible RCU does not exist, we never have to print out
928 * tasks blocked within RCU read-side critical sections that are blocking
929 * the current expedited grace period.
930 */
931static void rcu_exp_print_detail_task_stall_rnp(struct rcu_node *rnp)
932{
933}
934
935#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
936
937/**
938 * synchronize_rcu_expedited - Brute-force RCU grace period
939 *
940 * Wait for an RCU grace period, but expedite it. The basic idea is to
941 * IPI all non-idle non-nohz online CPUs. The IPI handler checks whether
942 * the CPU is in an RCU critical section, and if so, it sets a flag that
943 * causes the outermost rcu_read_unlock() to report the quiescent state
944 * for RCU-preempt or asks the scheduler for help for RCU-sched. On the
945 * other hand, if the CPU is not in an RCU read-side critical section,
946 * the IPI handler reports the quiescent state immediately.
947 *
948 * Although this is a great improvement over previous expedited
949 * implementations, it is still unfriendly to real-time workloads, so is
950 * thus not recommended for any sort of common-case code. In fact, if
951 * you are using synchronize_rcu_expedited() in a loop, please restructure
952 * your code to batch your updates, and then use a single synchronize_rcu()
953 * instead.
954 *
955 * This has the same semantics as (but is more brutal than) synchronize_rcu().
956 */
957void synchronize_rcu_expedited(void)
958{
959 bool boottime = (rcu_scheduler_active == RCU_SCHEDULER_INIT);
960 unsigned long flags;
961 struct rcu_exp_work rew;
962 struct rcu_node *rnp;
963 unsigned long s;
964
965 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
966 lock_is_held(&rcu_lock_map) ||
967 lock_is_held(&rcu_sched_lock_map),
968 "Illegal synchronize_rcu_expedited() in RCU read-side critical section");
969
970 /* Is the state is such that the call is a grace period? */
971 if (rcu_blocking_is_gp()) {
972 // Note well that this code runs with !PREEMPT && !SMP.
973 // In addition, all code that advances grace periods runs
974 // at process level. Therefore, this expedited GP overlaps
975 // with other expedited GPs only by being fully nested within
976 // them, which allows reuse of ->gp_seq_polled_exp_snap.
977 rcu_poll_gp_seq_start_unlocked(&rcu_state.gp_seq_polled_exp_snap);
978 rcu_poll_gp_seq_end_unlocked(&rcu_state.gp_seq_polled_exp_snap);
979
980 local_irq_save(flags);
981 WARN_ON_ONCE(num_online_cpus() > 1);
982 rcu_state.expedited_sequence += (1 << RCU_SEQ_CTR_SHIFT);
983 local_irq_restore(flags);
984 return; // Context allows vacuous grace periods.
985 }
986
987 /* If expedited grace periods are prohibited, fall back to normal. */
988 if (rcu_gp_is_normal()) {
989 wait_rcu_gp(call_rcu_hurry);
990 return;
991 }
992
993 /* Take a snapshot of the sequence number. */
994 s = rcu_exp_gp_seq_snap();
995 if (exp_funnel_lock(s))
996 return; /* Someone else did our work for us. */
997
998 /* Ensure that load happens before action based on it. */
999 if (unlikely(boottime)) {
1000 /* Direct call during scheduler init and early_initcalls(). */
1001 rcu_exp_sel_wait_wake(s);
1002 } else {
1003 /* Marshall arguments & schedule the expedited grace period. */
1004 rew.rew_s = s;
1005 synchronize_rcu_expedited_queue_work(&rew);
1006 }
1007
1008 /* Wait for expedited grace period to complete. */
1009 rnp = rcu_get_root();
1010 wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
1011 sync_exp_work_done(s));
1012 smp_mb(); /* Work actions happen before return. */
1013
1014 /* Let the next expedited grace period start. */
1015 mutex_unlock(&rcu_state.exp_mutex);
1016
1017 if (likely(!boottime))
1018 synchronize_rcu_expedited_destroy_work(&rew);
1019}
1020EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
1021
1022/*
1023 * Ensure that start_poll_synchronize_rcu_expedited() has the expedited
1024 * RCU grace periods that it needs.
1025 */
1026static void sync_rcu_do_polled_gp(struct work_struct *wp)
1027{
1028 unsigned long flags;
1029 int i = 0;
1030 struct rcu_node *rnp = container_of(wp, struct rcu_node, exp_poll_wq);
1031 unsigned long s;
1032
1033 raw_spin_lock_irqsave(&rnp->exp_poll_lock, flags);
1034 s = rnp->exp_seq_poll_rq;
1035 rnp->exp_seq_poll_rq = RCU_GET_STATE_COMPLETED;
1036 raw_spin_unlock_irqrestore(&rnp->exp_poll_lock, flags);
1037 if (s == RCU_GET_STATE_COMPLETED)
1038 return;
1039 while (!poll_state_synchronize_rcu(s)) {
1040 synchronize_rcu_expedited();
1041 if (i == 10 || i == 20)
1042 pr_info("%s: i = %d s = %lx gp_seq_polled = %lx\n", __func__, i, s, READ_ONCE(rcu_state.gp_seq_polled));
1043 i++;
1044 }
1045 raw_spin_lock_irqsave(&rnp->exp_poll_lock, flags);
1046 s = rnp->exp_seq_poll_rq;
1047 if (poll_state_synchronize_rcu(s))
1048 rnp->exp_seq_poll_rq = RCU_GET_STATE_COMPLETED;
1049 raw_spin_unlock_irqrestore(&rnp->exp_poll_lock, flags);
1050}
1051
1052/**
1053 * start_poll_synchronize_rcu_expedited - Snapshot current RCU state and start expedited grace period
1054 *
1055 * Returns a cookie to pass to a call to cond_synchronize_rcu(),
1056 * cond_synchronize_rcu_expedited(), or poll_state_synchronize_rcu(),
1057 * allowing them to determine whether or not any sort of grace period has
1058 * elapsed in the meantime. If the needed expedited grace period is not
1059 * already slated to start, initiates that grace period.
1060 */
1061unsigned long start_poll_synchronize_rcu_expedited(void)
1062{
1063 unsigned long flags;
1064 struct rcu_data *rdp;
1065 struct rcu_node *rnp;
1066 unsigned long s;
1067
1068 s = get_state_synchronize_rcu();
1069 rdp = per_cpu_ptr(&rcu_data, raw_smp_processor_id());
1070 rnp = rdp->mynode;
1071 if (rcu_init_invoked())
1072 raw_spin_lock_irqsave(&rnp->exp_poll_lock, flags);
1073 if (!poll_state_synchronize_rcu(s)) {
1074 if (rcu_init_invoked()) {
1075 rnp->exp_seq_poll_rq = s;
1076 queue_work(rcu_gp_wq, &rnp->exp_poll_wq);
1077 }
1078 }
1079 if (rcu_init_invoked())
1080 raw_spin_unlock_irqrestore(&rnp->exp_poll_lock, flags);
1081
1082 return s;
1083}
1084EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu_expedited);
1085
1086/**
1087 * start_poll_synchronize_rcu_expedited_full - Take a full snapshot and start expedited grace period
1088 * @rgosp: Place to put snapshot of grace-period state
1089 *
1090 * Places the normal and expedited grace-period states in rgosp. This
1091 * state value can be passed to a later call to cond_synchronize_rcu_full()
1092 * or poll_state_synchronize_rcu_full() to determine whether or not a
1093 * grace period (whether normal or expedited) has elapsed in the meantime.
1094 * If the needed expedited grace period is not already slated to start,
1095 * initiates that grace period.
1096 */
1097void start_poll_synchronize_rcu_expedited_full(struct rcu_gp_oldstate *rgosp)
1098{
1099 get_state_synchronize_rcu_full(rgosp);
1100 (void)start_poll_synchronize_rcu_expedited();
1101}
1102EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu_expedited_full);
1103
1104/**
1105 * cond_synchronize_rcu_expedited - Conditionally wait for an expedited RCU grace period
1106 *
1107 * @oldstate: value from get_state_synchronize_rcu(), start_poll_synchronize_rcu(), or start_poll_synchronize_rcu_expedited()
1108 *
1109 * If any type of full RCU grace period has elapsed since the earlier
1110 * call to get_state_synchronize_rcu(), start_poll_synchronize_rcu(),
1111 * or start_poll_synchronize_rcu_expedited(), just return. Otherwise,
1112 * invoke synchronize_rcu_expedited() to wait for a full grace period.
1113 *
1114 * Yes, this function does not take counter wrap into account.
1115 * But counter wrap is harmless. If the counter wraps, we have waited for
1116 * more than 2 billion grace periods (and way more on a 64-bit system!),
1117 * so waiting for a couple of additional grace periods should be just fine.
1118 *
1119 * This function provides the same memory-ordering guarantees that
1120 * would be provided by a synchronize_rcu() that was invoked at the call
1121 * to the function that provided @oldstate and that returned at the end
1122 * of this function.
1123 */
1124void cond_synchronize_rcu_expedited(unsigned long oldstate)
1125{
1126 if (!poll_state_synchronize_rcu(oldstate))
1127 synchronize_rcu_expedited();
1128}
1129EXPORT_SYMBOL_GPL(cond_synchronize_rcu_expedited);
1130
1131/**
1132 * cond_synchronize_rcu_expedited_full - Conditionally wait for an expedited RCU grace period
1133 * @rgosp: value from get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(), or start_poll_synchronize_rcu_expedited_full()
1134 *
1135 * If a full RCU grace period has elapsed since the call to
1136 * get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(),
1137 * or start_poll_synchronize_rcu_expedited_full() from which @rgosp was
1138 * obtained, just return. Otherwise, invoke synchronize_rcu_expedited()
1139 * to wait for a full grace period.
1140 *
1141 * Yes, this function does not take counter wrap into account.
1142 * But counter wrap is harmless. If the counter wraps, we have waited for
1143 * more than 2 billion grace periods (and way more on a 64-bit system!),
1144 * so waiting for a couple of additional grace periods should be just fine.
1145 *
1146 * This function provides the same memory-ordering guarantees that
1147 * would be provided by a synchronize_rcu() that was invoked at the call
1148 * to the function that provided @rgosp and that returned at the end of
1149 * this function.
1150 */
1151void cond_synchronize_rcu_expedited_full(struct rcu_gp_oldstate *rgosp)
1152{
1153 if (!poll_state_synchronize_rcu_full(rgosp))
1154 synchronize_rcu_expedited();
1155}
1156EXPORT_SYMBOL_GPL(cond_synchronize_rcu_expedited_full);