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(sp: &rcu_state.expedited_sequence);
22 rcu_poll_gp_seq_start_unlocked(snap: &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(sp: &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(snap: &rcu_state.gp_seq_polled_exp_snap);
40 rcu_seq_end(sp: &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(sp: &rcu_state.expedited_sequence);
55 trace_rcu_exp_grace_period(rcuname: rcu_state.name, gpseq: 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(sp: &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/*
178 * Report the exit from RCU read-side critical section for the last task
179 * that queued itself during or before the current expedited preemptible-RCU
180 * grace period. This event is reported either to the rcu_node structure on
181 * which the task was queued or to one of that rcu_node structure's ancestors,
182 * recursively up the tree. (Calm down, calm down, we do the recursion
183 * iteratively!)
184 */
185static void __rcu_report_exp_rnp(struct rcu_node *rnp,
186 bool wake, unsigned long flags)
187 __releases(rnp->lock)
188{
189 unsigned long mask;
190
191 raw_lockdep_assert_held_rcu_node(rnp);
192 for (;;) {
193 if (!sync_rcu_exp_done(rnp)) {
194 if (!rnp->expmask)
195 rcu_initiate_boost(rnp, flags);
196 else
197 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
198 break;
199 }
200 if (rnp->parent == NULL) {
201 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
202 if (wake) {
203 smp_mb(); /* EGP done before wake_up(). */
204 swake_up_one(q: &rcu_state.expedited_wq);
205 }
206 break;
207 }
208 mask = rnp->grpmask;
209 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled */
210 rnp = rnp->parent;
211 raw_spin_lock_rcu_node(rnp); /* irqs already disabled */
212 WARN_ON_ONCE(!(rnp->expmask & mask));
213 WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask);
214 }
215}
216
217/*
218 * Report expedited quiescent state for specified node. This is a
219 * lock-acquisition wrapper function for __rcu_report_exp_rnp().
220 */
221static void __maybe_unused rcu_report_exp_rnp(struct rcu_node *rnp, bool wake)
222{
223 unsigned long flags;
224
225 raw_spin_lock_irqsave_rcu_node(rnp, flags);
226 __rcu_report_exp_rnp(rnp, wake, flags);
227}
228
229/*
230 * Report expedited quiescent state for multiple CPUs, all covered by the
231 * specified leaf rcu_node structure.
232 */
233static void rcu_report_exp_cpu_mult(struct rcu_node *rnp,
234 unsigned long mask, bool wake)
235{
236 int cpu;
237 unsigned long flags;
238 struct rcu_data *rdp;
239
240 raw_spin_lock_irqsave_rcu_node(rnp, flags);
241 if (!(rnp->expmask & mask)) {
242 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
243 return;
244 }
245 WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask);
246 for_each_leaf_node_cpu_mask(rnp, cpu, mask) {
247 rdp = per_cpu_ptr(&rcu_data, cpu);
248 if (!IS_ENABLED(CONFIG_NO_HZ_FULL) || !rdp->rcu_forced_tick_exp)
249 continue;
250 rdp->rcu_forced_tick_exp = false;
251 tick_dep_clear_cpu(cpu, bit: TICK_DEP_BIT_RCU_EXP);
252 }
253 __rcu_report_exp_rnp(rnp, wake, flags); /* Releases rnp->lock. */
254}
255
256/*
257 * Report expedited quiescent state for specified rcu_data (CPU).
258 */
259static void rcu_report_exp_rdp(struct rcu_data *rdp)
260{
261 WRITE_ONCE(rdp->cpu_no_qs.b.exp, false);
262 rcu_report_exp_cpu_mult(rnp: rdp->mynode, mask: rdp->grpmask, wake: true);
263}
264
265/* Common code for work-done checking. */
266static bool sync_exp_work_done(unsigned long s)
267{
268 if (rcu_exp_gp_seq_done(s)) {
269 trace_rcu_exp_grace_period(rcuname: rcu_state.name, gpseq: s, TPS("done"));
270 smp_mb(); /* Ensure test happens before caller kfree(). */
271 return true;
272 }
273 return false;
274}
275
276/*
277 * Funnel-lock acquisition for expedited grace periods. Returns true
278 * if some other task completed an expedited grace period that this task
279 * can piggy-back on, and with no mutex held. Otherwise, returns false
280 * with the mutex held, indicating that the caller must actually do the
281 * expedited grace period.
282 */
283static bool exp_funnel_lock(unsigned long s)
284{
285 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, raw_smp_processor_id());
286 struct rcu_node *rnp = rdp->mynode;
287 struct rcu_node *rnp_root = rcu_get_root();
288
289 /* Low-contention fastpath. */
290 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s) &&
291 (rnp == rnp_root ||
292 ULONG_CMP_LT(READ_ONCE(rnp_root->exp_seq_rq), s)) &&
293 mutex_trylock(lock: &rcu_state.exp_mutex))
294 goto fastpath;
295
296 /*
297 * Each pass through the following loop works its way up
298 * the rcu_node tree, returning if others have done the work or
299 * otherwise falls through to acquire ->exp_mutex. The mapping
300 * from CPU to rcu_node structure can be inexact, as it is just
301 * promoting locality and is not strictly needed for correctness.
302 */
303 for (; rnp != NULL; rnp = rnp->parent) {
304 if (sync_exp_work_done(s))
305 return true;
306
307 /* Work not done, either wait here or go up. */
308 spin_lock(lock: &rnp->exp_lock);
309 if (ULONG_CMP_GE(rnp->exp_seq_rq, s)) {
310
311 /* Someone else doing GP, so wait for them. */
312 spin_unlock(lock: &rnp->exp_lock);
313 trace_rcu_exp_funnel_lock(rcuname: rcu_state.name, level: rnp->level,
314 grplo: rnp->grplo, grphi: rnp->grphi,
315 TPS("wait"));
316 wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
317 sync_exp_work_done(s));
318 return true;
319 }
320 WRITE_ONCE(rnp->exp_seq_rq, s); /* Followers can wait on us. */
321 spin_unlock(lock: &rnp->exp_lock);
322 trace_rcu_exp_funnel_lock(rcuname: rcu_state.name, level: rnp->level,
323 grplo: rnp->grplo, grphi: rnp->grphi, TPS("nxtlvl"));
324 }
325 mutex_lock(&rcu_state.exp_mutex);
326fastpath:
327 if (sync_exp_work_done(s)) {
328 mutex_unlock(lock: &rcu_state.exp_mutex);
329 return true;
330 }
331 rcu_exp_gp_seq_start();
332 trace_rcu_exp_grace_period(rcuname: rcu_state.name, gpseq: s, TPS("start"));
333 return false;
334}
335
336/*
337 * Select the CPUs within the specified rcu_node that the upcoming
338 * expedited grace period needs to wait for.
339 */
340static void __sync_rcu_exp_select_node_cpus(struct rcu_exp_work *rewp)
341{
342 int cpu;
343 unsigned long flags;
344 unsigned long mask_ofl_test;
345 unsigned long mask_ofl_ipi;
346 int ret;
347 struct rcu_node *rnp = container_of(rewp, struct rcu_node, rew);
348
349 raw_spin_lock_irqsave_rcu_node(rnp, flags);
350
351 /* Each pass checks a CPU for identity, offline, and idle. */
352 mask_ofl_test = 0;
353 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) {
354 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
355 unsigned long mask = rdp->grpmask;
356 int snap;
357
358 if (raw_smp_processor_id() == cpu ||
359 !(rnp->qsmaskinitnext & mask)) {
360 mask_ofl_test |= mask;
361 } else {
362 snap = rcu_dynticks_snap(cpu);
363 if (rcu_dynticks_in_eqs(snap))
364 mask_ofl_test |= mask;
365 else
366 rdp->exp_dynticks_snap = snap;
367 }
368 }
369 mask_ofl_ipi = rnp->expmask & ~mask_ofl_test;
370
371 /*
372 * Need to wait for any blocked tasks as well. Note that
373 * additional blocking tasks will also block the expedited GP
374 * until such time as the ->expmask bits are cleared.
375 */
376 if (rcu_preempt_has_tasks(rnp))
377 WRITE_ONCE(rnp->exp_tasks, rnp->blkd_tasks.next);
378 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
379
380 /* IPI the remaining CPUs for expedited quiescent state. */
381 for_each_leaf_node_cpu_mask(rnp, cpu, mask_ofl_ipi) {
382 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
383 unsigned long mask = rdp->grpmask;
384
385retry_ipi:
386 if (rcu_dynticks_in_eqs_since(rdp, snap: rdp->exp_dynticks_snap)) {
387 mask_ofl_test |= mask;
388 continue;
389 }
390 if (get_cpu() == cpu) {
391 mask_ofl_test |= mask;
392 put_cpu();
393 continue;
394 }
395 ret = smp_call_function_single(cpuid: cpu, func: rcu_exp_handler, NULL, wait: 0);
396 put_cpu();
397 /* The CPU will report the QS in response to the IPI. */
398 if (!ret)
399 continue;
400
401 /* Failed, raced with CPU hotplug operation. */
402 raw_spin_lock_irqsave_rcu_node(rnp, flags);
403 if ((rnp->qsmaskinitnext & mask) &&
404 (rnp->expmask & mask)) {
405 /* Online, so delay for a bit and try again. */
406 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
407 trace_rcu_exp_grace_period(rcuname: rcu_state.name, gpseq: rcu_exp_gp_seq_endval(), TPS("selectofl"));
408 schedule_timeout_idle(timeout: 1);
409 goto retry_ipi;
410 }
411 /* CPU really is offline, so we must report its QS. */
412 if (rnp->expmask & mask)
413 mask_ofl_test |= mask;
414 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
415 }
416 /* Report quiescent states for those that went offline. */
417 if (mask_ofl_test)
418 rcu_report_exp_cpu_mult(rnp, mask: mask_ofl_test, wake: false);
419}
420
421static void rcu_exp_sel_wait_wake(unsigned long s);
422
423#ifdef CONFIG_RCU_EXP_KTHREAD
424static void sync_rcu_exp_select_node_cpus(struct kthread_work *wp)
425{
426 struct rcu_exp_work *rewp =
427 container_of(wp, struct rcu_exp_work, rew_work);
428
429 __sync_rcu_exp_select_node_cpus(rewp);
430}
431
432static inline bool rcu_gp_par_worker_started(void)
433{
434 return !!READ_ONCE(rcu_exp_par_gp_kworker);
435}
436
437static inline void sync_rcu_exp_select_cpus_queue_work(struct rcu_node *rnp)
438{
439 kthread_init_work(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus);
440 /*
441 * Use rcu_exp_par_gp_kworker, because flushing a work item from
442 * another work item on the same kthread worker can result in
443 * deadlock.
444 */
445 kthread_queue_work(worker: rcu_exp_par_gp_kworker, work: &rnp->rew.rew_work);
446}
447
448static inline void sync_rcu_exp_select_cpus_flush_work(struct rcu_node *rnp)
449{
450 kthread_flush_work(work: &rnp->rew.rew_work);
451}
452
453/*
454 * Work-queue handler to drive an expedited grace period forward.
455 */
456static void wait_rcu_exp_gp(struct kthread_work *wp)
457{
458 struct rcu_exp_work *rewp;
459
460 rewp = container_of(wp, struct rcu_exp_work, rew_work);
461 rcu_exp_sel_wait_wake(s: rewp->rew_s);
462}
463
464static inline void synchronize_rcu_expedited_queue_work(struct rcu_exp_work *rew)
465{
466 kthread_init_work(&rew->rew_work, wait_rcu_exp_gp);
467 kthread_queue_work(worker: rcu_exp_gp_kworker, work: &rew->rew_work);
468}
469
470static inline void synchronize_rcu_expedited_destroy_work(struct rcu_exp_work *rew)
471{
472}
473#else /* !CONFIG_RCU_EXP_KTHREAD */
474static void sync_rcu_exp_select_node_cpus(struct work_struct *wp)
475{
476 struct rcu_exp_work *rewp =
477 container_of(wp, struct rcu_exp_work, rew_work);
478
479 __sync_rcu_exp_select_node_cpus(rewp);
480}
481
482static inline bool rcu_gp_par_worker_started(void)
483{
484 return !!READ_ONCE(rcu_par_gp_wq);
485}
486
487static inline void sync_rcu_exp_select_cpus_queue_work(struct rcu_node *rnp)
488{
489 int cpu = find_next_bit(&rnp->ffmask, BITS_PER_LONG, -1);
490
491 INIT_WORK(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus);
492 /* If all offline, queue the work on an unbound CPU. */
493 if (unlikely(cpu > rnp->grphi - rnp->grplo))
494 cpu = WORK_CPU_UNBOUND;
495 else
496 cpu += rnp->grplo;
497 queue_work_on(cpu, rcu_par_gp_wq, &rnp->rew.rew_work);
498}
499
500static inline void sync_rcu_exp_select_cpus_flush_work(struct rcu_node *rnp)
501{
502 flush_work(&rnp->rew.rew_work);
503}
504
505/*
506 * Work-queue handler to drive an expedited grace period forward.
507 */
508static void wait_rcu_exp_gp(struct work_struct *wp)
509{
510 struct rcu_exp_work *rewp;
511
512 rewp = container_of(wp, struct rcu_exp_work, rew_work);
513 rcu_exp_sel_wait_wake(rewp->rew_s);
514}
515
516static inline void synchronize_rcu_expedited_queue_work(struct rcu_exp_work *rew)
517{
518 INIT_WORK_ONSTACK(&rew->rew_work, wait_rcu_exp_gp);
519 queue_work(rcu_gp_wq, &rew->rew_work);
520}
521
522static inline void synchronize_rcu_expedited_destroy_work(struct rcu_exp_work *rew)
523{
524 destroy_work_on_stack(&rew->rew_work);
525}
526#endif /* CONFIG_RCU_EXP_KTHREAD */
527
528/*
529 * Select the nodes that the upcoming expedited grace period needs
530 * to wait for.
531 */
532static void sync_rcu_exp_select_cpus(void)
533{
534 struct rcu_node *rnp;
535
536 trace_rcu_exp_grace_period(rcuname: rcu_state.name, gpseq: rcu_exp_gp_seq_endval(), TPS("reset"));
537 sync_exp_reset_tree();
538 trace_rcu_exp_grace_period(rcuname: rcu_state.name, gpseq: rcu_exp_gp_seq_endval(), TPS("select"));
539
540 /* Schedule work for each leaf rcu_node structure. */
541 rcu_for_each_leaf_node(rnp) {
542 rnp->exp_need_flush = false;
543 if (!READ_ONCE(rnp->expmask))
544 continue; /* Avoid early boot non-existent wq. */
545 if (!rcu_gp_par_worker_started() ||
546 rcu_scheduler_active != RCU_SCHEDULER_RUNNING ||
547 rcu_is_last_leaf_node(rnp)) {
548 /* No worker started yet or last leaf, do direct call. */
549 sync_rcu_exp_select_node_cpus(wp: &rnp->rew.rew_work);
550 continue;
551 }
552 sync_rcu_exp_select_cpus_queue_work(rnp);
553 rnp->exp_need_flush = true;
554 }
555
556 /* Wait for jobs (if any) to complete. */
557 rcu_for_each_leaf_node(rnp)
558 if (rnp->exp_need_flush)
559 sync_rcu_exp_select_cpus_flush_work(rnp);
560}
561
562/*
563 * Wait for the expedited grace period to elapse, within time limit.
564 * If the time limit is exceeded without the grace period elapsing,
565 * return false, otherwise return true.
566 */
567static bool synchronize_rcu_expedited_wait_once(long tlimit)
568{
569 int t;
570 struct rcu_node *rnp_root = rcu_get_root();
571
572 t = swait_event_timeout_exclusive(rcu_state.expedited_wq,
573 sync_rcu_exp_done_unlocked(rnp_root),
574 tlimit);
575 // Workqueues should not be signaled.
576 if (t > 0 || sync_rcu_exp_done_unlocked(rnp: rnp_root))
577 return true;
578 WARN_ON(t < 0); /* workqueues should not be signaled. */
579 return false;
580}
581
582/*
583 * Wait for the expedited grace period to elapse, issuing any needed
584 * RCU CPU stall warnings along the way.
585 */
586static void synchronize_rcu_expedited_wait(void)
587{
588 int cpu;
589 unsigned long j;
590 unsigned long jiffies_stall;
591 unsigned long jiffies_start;
592 unsigned long mask;
593 int ndetected;
594 struct rcu_data *rdp;
595 struct rcu_node *rnp;
596 struct rcu_node *rnp_root = rcu_get_root();
597 unsigned long flags;
598
599 trace_rcu_exp_grace_period(rcuname: rcu_state.name, gpseq: rcu_exp_gp_seq_endval(), TPS("startwait"));
600 jiffies_stall = rcu_exp_jiffies_till_stall_check();
601 jiffies_start = jiffies;
602 if (tick_nohz_full_enabled() && rcu_inkernel_boot_has_ended()) {
603 if (synchronize_rcu_expedited_wait_once(tlimit: 1))
604 return;
605 rcu_for_each_leaf_node(rnp) {
606 raw_spin_lock_irqsave_rcu_node(rnp, flags);
607 mask = READ_ONCE(rnp->expmask);
608 for_each_leaf_node_cpu_mask(rnp, cpu, mask) {
609 rdp = per_cpu_ptr(&rcu_data, cpu);
610 if (rdp->rcu_forced_tick_exp)
611 continue;
612 rdp->rcu_forced_tick_exp = true;
613 if (cpu_online(cpu))
614 tick_dep_set_cpu(cpu, bit: TICK_DEP_BIT_RCU_EXP);
615 }
616 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
617 }
618 j = READ_ONCE(jiffies_till_first_fqs);
619 if (synchronize_rcu_expedited_wait_once(tlimit: j + HZ))
620 return;
621 }
622
623 for (;;) {
624 unsigned long j;
625
626 if (synchronize_rcu_expedited_wait_once(tlimit: jiffies_stall))
627 return;
628 if (rcu_stall_is_suppressed())
629 continue;
630 j = jiffies;
631 rcu_stall_notifier_call_chain(RCU_STALL_NOTIFY_EXP, v: (void *)(j - jiffies_start));
632 trace_rcu_stall_warning(rcuname: rcu_state.name, TPS("ExpeditedStall"));
633 pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
634 rcu_state.name);
635 ndetected = 0;
636 rcu_for_each_leaf_node(rnp) {
637 ndetected += rcu_print_task_exp_stall(rnp);
638 for_each_leaf_node_possible_cpu(rnp, cpu) {
639 struct rcu_data *rdp;
640
641 mask = leaf_node_cpu_bit(rnp, cpu);
642 if (!(READ_ONCE(rnp->expmask) & mask))
643 continue;
644 ndetected++;
645 rdp = per_cpu_ptr(&rcu_data, cpu);
646 pr_cont(" %d-%c%c%c%c", cpu,
647 "O."[!!cpu_online(cpu)],
648 "o."[!!(rdp->grpmask & rnp->expmaskinit)],
649 "N."[!!(rdp->grpmask & rnp->expmaskinitnext)],
650 "D."[!!data_race(rdp->cpu_no_qs.b.exp)]);
651 }
652 }
653 pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n",
654 j - jiffies_start, rcu_state.expedited_sequence,
655 data_race(rnp_root->expmask),
656 ".T"[!!data_race(rnp_root->exp_tasks)]);
657 if (ndetected) {
658 pr_err("blocking rcu_node structures (internal RCU debug):");
659 rcu_for_each_node_breadth_first(rnp) {
660 if (rnp == rnp_root)
661 continue; /* printed unconditionally */
662 if (sync_rcu_exp_done_unlocked(rnp))
663 continue;
664 pr_cont(" l=%u:%d-%d:%#lx/%c",
665 rnp->level, rnp->grplo, rnp->grphi,
666 data_race(rnp->expmask),
667 ".T"[!!data_race(rnp->exp_tasks)]);
668 }
669 pr_cont("\n");
670 }
671 rcu_for_each_leaf_node(rnp) {
672 for_each_leaf_node_possible_cpu(rnp, cpu) {
673 mask = leaf_node_cpu_bit(rnp, cpu);
674 if (!(READ_ONCE(rnp->expmask) & mask))
675 continue;
676 preempt_disable(); // For smp_processor_id() in dump_cpu_task().
677 dump_cpu_task(cpu);
678 preempt_enable();
679 }
680 rcu_exp_print_detail_task_stall_rnp(rnp);
681 }
682 jiffies_stall = 3 * rcu_exp_jiffies_till_stall_check() + 3;
683 panic_on_rcu_stall();
684 }
685}
686
687/*
688 * Wait for the current expedited grace period to complete, and then
689 * wake up everyone who piggybacked on the just-completed expedited
690 * grace period. Also update all the ->exp_seq_rq counters as needed
691 * in order to avoid counter-wrap problems.
692 */
693static void rcu_exp_wait_wake(unsigned long s)
694{
695 struct rcu_node *rnp;
696
697 synchronize_rcu_expedited_wait();
698
699 // Switch over to wakeup mode, allowing the next GP to proceed.
700 // End the previous grace period only after acquiring the mutex
701 // to ensure that only one GP runs concurrently with wakeups.
702 mutex_lock(&rcu_state.exp_wake_mutex);
703 rcu_exp_gp_seq_end();
704 trace_rcu_exp_grace_period(rcuname: rcu_state.name, gpseq: s, TPS("end"));
705
706 rcu_for_each_node_breadth_first(rnp) {
707 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s)) {
708 spin_lock(lock: &rnp->exp_lock);
709 /* Recheck, avoid hang in case someone just arrived. */
710 if (ULONG_CMP_LT(rnp->exp_seq_rq, s))
711 WRITE_ONCE(rnp->exp_seq_rq, s);
712 spin_unlock(lock: &rnp->exp_lock);
713 }
714 smp_mb(); /* All above changes before wakeup. */
715 wake_up_all(&rnp->exp_wq[rcu_seq_ctr(s) & 0x3]);
716 }
717 trace_rcu_exp_grace_period(rcuname: rcu_state.name, gpseq: s, TPS("endwake"));
718 mutex_unlock(lock: &rcu_state.exp_wake_mutex);
719}
720
721/*
722 * Common code to drive an expedited grace period forward, used by
723 * workqueues and mid-boot-time tasks.
724 */
725static void rcu_exp_sel_wait_wake(unsigned long s)
726{
727 /* Initialize the rcu_node tree in preparation for the wait. */
728 sync_rcu_exp_select_cpus();
729
730 /* Wait and clean up, including waking everyone. */
731 rcu_exp_wait_wake(s);
732}
733
734#ifdef CONFIG_PREEMPT_RCU
735
736/*
737 * Remote handler for smp_call_function_single(). If there is an
738 * RCU read-side critical section in effect, request that the
739 * next rcu_read_unlock() record the quiescent state up the
740 * ->expmask fields in the rcu_node tree. Otherwise, immediately
741 * report the quiescent state.
742 */
743static void rcu_exp_handler(void *unused)
744{
745 int depth = rcu_preempt_depth();
746 unsigned long flags;
747 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
748 struct rcu_node *rnp = rdp->mynode;
749 struct task_struct *t = current;
750
751 /*
752 * First, the common case of not being in an RCU read-side
753 * critical section. If also enabled or idle, immediately
754 * report the quiescent state, otherwise defer.
755 */
756 if (!depth) {
757 if (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK)) ||
758 rcu_is_cpu_rrupt_from_idle()) {
759 rcu_report_exp_rdp(rdp);
760 } else {
761 WRITE_ONCE(rdp->cpu_no_qs.b.exp, true);
762 set_tsk_need_resched(t);
763 set_preempt_need_resched();
764 }
765 return;
766 }
767
768 /*
769 * Second, the less-common case of being in an RCU read-side
770 * critical section. In this case we can count on a future
771 * rcu_read_unlock(). However, this rcu_read_unlock() might
772 * execute on some other CPU, but in that case there will be
773 * a future context switch. Either way, if the expedited
774 * grace period is still waiting on this CPU, set ->deferred_qs
775 * so that the eventual quiescent state will be reported.
776 * Note that there is a large group of race conditions that
777 * can have caused this quiescent state to already have been
778 * reported, so we really do need to check ->expmask.
779 */
780 if (depth > 0) {
781 raw_spin_lock_irqsave_rcu_node(rnp, flags);
782 if (rnp->expmask & rdp->grpmask) {
783 WRITE_ONCE(rdp->cpu_no_qs.b.exp, true);
784 t->rcu_read_unlock_special.b.exp_hint = true;
785 }
786 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
787 return;
788 }
789
790 // Finally, negative nesting depth should not happen.
791 WARN_ON_ONCE(1);
792}
793
794/* PREEMPTION=y, so no PREEMPTION=n expedited grace period to clean up after. */
795static void sync_sched_exp_online_cleanup(int cpu)
796{
797}
798
799/*
800 * Scan the current list of tasks blocked within RCU read-side critical
801 * sections, printing out the tid of each that is blocking the current
802 * expedited grace period.
803 */
804static int rcu_print_task_exp_stall(struct rcu_node *rnp)
805{
806 unsigned long flags;
807 int ndetected = 0;
808 struct task_struct *t;
809
810 raw_spin_lock_irqsave_rcu_node(rnp, flags);
811 if (!rnp->exp_tasks) {
812 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
813 return 0;
814 }
815 t = list_entry(rnp->exp_tasks->prev,
816 struct task_struct, rcu_node_entry);
817 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
818 pr_cont(" P%d", t->pid);
819 ndetected++;
820 }
821 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
822 return ndetected;
823}
824
825/*
826 * Scan the current list of tasks blocked within RCU read-side critical
827 * sections, dumping the stack of each that is blocking the current
828 * expedited grace period.
829 */
830static void rcu_exp_print_detail_task_stall_rnp(struct rcu_node *rnp)
831{
832 unsigned long flags;
833 struct task_struct *t;
834
835 if (!rcu_exp_stall_task_details)
836 return;
837 raw_spin_lock_irqsave_rcu_node(rnp, flags);
838 if (!READ_ONCE(rnp->exp_tasks)) {
839 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
840 return;
841 }
842 t = list_entry(rnp->exp_tasks->prev,
843 struct task_struct, rcu_node_entry);
844 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
845 /*
846 * We could be printing a lot while holding a spinlock.
847 * Avoid triggering hard lockup.
848 */
849 touch_nmi_watchdog();
850 sched_show_task(p: t);
851 }
852 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
853}
854
855#else /* #ifdef CONFIG_PREEMPT_RCU */
856
857/* Request an expedited quiescent state. */
858static void rcu_exp_need_qs(void)
859{
860 __this_cpu_write(rcu_data.cpu_no_qs.b.exp, true);
861 /* Store .exp before .rcu_urgent_qs. */
862 smp_store_release(this_cpu_ptr(&rcu_data.rcu_urgent_qs), true);
863 set_tsk_need_resched(current);
864 set_preempt_need_resched();
865}
866
867/* Invoked on each online non-idle CPU for expedited quiescent state. */
868static void rcu_exp_handler(void *unused)
869{
870 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
871 struct rcu_node *rnp = rdp->mynode;
872 bool preempt_bh_enabled = !(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
873
874 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
875 __this_cpu_read(rcu_data.cpu_no_qs.b.exp))
876 return;
877 if (rcu_is_cpu_rrupt_from_idle() ||
878 (IS_ENABLED(CONFIG_PREEMPT_COUNT) && preempt_bh_enabled)) {
879 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
880 return;
881 }
882 rcu_exp_need_qs();
883}
884
885/* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
886static void sync_sched_exp_online_cleanup(int cpu)
887{
888 unsigned long flags;
889 int my_cpu;
890 struct rcu_data *rdp;
891 int ret;
892 struct rcu_node *rnp;
893
894 rdp = per_cpu_ptr(&rcu_data, cpu);
895 rnp = rdp->mynode;
896 my_cpu = get_cpu();
897 /* Quiescent state either not needed or already requested, leave. */
898 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
899 READ_ONCE(rdp->cpu_no_qs.b.exp)) {
900 put_cpu();
901 return;
902 }
903 /* Quiescent state needed on current CPU, so set it up locally. */
904 if (my_cpu == cpu) {
905 local_irq_save(flags);
906 rcu_exp_need_qs();
907 local_irq_restore(flags);
908 put_cpu();
909 return;
910 }
911 /* Quiescent state needed on some other CPU, send IPI. */
912 ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
913 put_cpu();
914 WARN_ON_ONCE(ret);
915}
916
917/*
918 * Because preemptible RCU does not exist, we never have to check for
919 * tasks blocked within RCU read-side critical sections that are
920 * blocking the current expedited grace period.
921 */
922static int rcu_print_task_exp_stall(struct rcu_node *rnp)
923{
924 return 0;
925}
926
927/*
928 * Because preemptible RCU does not exist, we never have to print out
929 * tasks blocked within RCU read-side critical sections that are blocking
930 * the current expedited grace period.
931 */
932static void rcu_exp_print_detail_task_stall_rnp(struct rcu_node *rnp)
933{
934}
935
936#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
937
938/**
939 * synchronize_rcu_expedited - Brute-force RCU grace period
940 *
941 * Wait for an RCU grace period, but expedite it. The basic idea is to
942 * IPI all non-idle non-nohz online CPUs. The IPI handler checks whether
943 * the CPU is in an RCU critical section, and if so, it sets a flag that
944 * causes the outermost rcu_read_unlock() to report the quiescent state
945 * for RCU-preempt or asks the scheduler for help for RCU-sched. On the
946 * other hand, if the CPU is not in an RCU read-side critical section,
947 * the IPI handler reports the quiescent state immediately.
948 *
949 * Although this is a great improvement over previous expedited
950 * implementations, it is still unfriendly to real-time workloads, so is
951 * thus not recommended for any sort of common-case code. In fact, if
952 * you are using synchronize_rcu_expedited() in a loop, please restructure
953 * your code to batch your updates, and then use a single synchronize_rcu()
954 * instead.
955 *
956 * This has the same semantics as (but is more brutal than) synchronize_rcu().
957 */
958void synchronize_rcu_expedited(void)
959{
960 bool boottime = (rcu_scheduler_active == RCU_SCHEDULER_INIT);
961 unsigned long flags;
962 struct rcu_exp_work rew;
963 struct rcu_node *rnp;
964 unsigned long s;
965
966 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
967 lock_is_held(&rcu_lock_map) ||
968 lock_is_held(&rcu_sched_lock_map),
969 "Illegal synchronize_rcu_expedited() in RCU read-side critical section");
970
971 /* Is the state is such that the call is a grace period? */
972 if (rcu_blocking_is_gp()) {
973 // Note well that this code runs with !PREEMPT && !SMP.
974 // In addition, all code that advances grace periods runs
975 // at process level. Therefore, this expedited GP overlaps
976 // with other expedited GPs only by being fully nested within
977 // them, which allows reuse of ->gp_seq_polled_exp_snap.
978 rcu_poll_gp_seq_start_unlocked(snap: &rcu_state.gp_seq_polled_exp_snap);
979 rcu_poll_gp_seq_end_unlocked(snap: &rcu_state.gp_seq_polled_exp_snap);
980
981 local_irq_save(flags);
982 WARN_ON_ONCE(num_online_cpus() > 1);
983 rcu_state.expedited_sequence += (1 << RCU_SEQ_CTR_SHIFT);
984 local_irq_restore(flags);
985 return; // Context allows vacuous grace periods.
986 }
987
988 /* If expedited grace periods are prohibited, fall back to normal. */
989 if (rcu_gp_is_normal()) {
990 wait_rcu_gp(call_rcu_hurry);
991 return;
992 }
993
994 /* Take a snapshot of the sequence number. */
995 s = rcu_exp_gp_seq_snap();
996 if (exp_funnel_lock(s))
997 return; /* Someone else did our work for us. */
998
999 /* Ensure that load happens before action based on it. */
1000 if (unlikely(boottime)) {
1001 /* Direct call during scheduler init and early_initcalls(). */
1002 rcu_exp_sel_wait_wake(s);
1003 } else {
1004 /* Marshall arguments & schedule the expedited grace period. */
1005 rew.rew_s = s;
1006 synchronize_rcu_expedited_queue_work(rew: &rew);
1007 }
1008
1009 /* Wait for expedited grace period to complete. */
1010 rnp = rcu_get_root();
1011 wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
1012 sync_exp_work_done(s));
1013 smp_mb(); /* Work actions happen before return. */
1014
1015 /* Let the next expedited grace period start. */
1016 mutex_unlock(lock: &rcu_state.exp_mutex);
1017
1018 if (likely(!boottime))
1019 synchronize_rcu_expedited_destroy_work(rew: &rew);
1020}
1021EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
1022
1023/*
1024 * Ensure that start_poll_synchronize_rcu_expedited() has the expedited
1025 * RCU grace periods that it needs.
1026 */
1027static void sync_rcu_do_polled_gp(struct work_struct *wp)
1028{
1029 unsigned long flags;
1030 int i = 0;
1031 struct rcu_node *rnp = container_of(wp, struct rcu_node, exp_poll_wq);
1032 unsigned long s;
1033
1034 raw_spin_lock_irqsave(&rnp->exp_poll_lock, flags);
1035 s = rnp->exp_seq_poll_rq;
1036 rnp->exp_seq_poll_rq = RCU_GET_STATE_COMPLETED;
1037 raw_spin_unlock_irqrestore(&rnp->exp_poll_lock, flags);
1038 if (s == RCU_GET_STATE_COMPLETED)
1039 return;
1040 while (!poll_state_synchronize_rcu(s)) {
1041 synchronize_rcu_expedited();
1042 if (i == 10 || i == 20)
1043 pr_info("%s: i = %d s = %lx gp_seq_polled = %lx\n", __func__, i, s, READ_ONCE(rcu_state.gp_seq_polled));
1044 i++;
1045 }
1046 raw_spin_lock_irqsave(&rnp->exp_poll_lock, flags);
1047 s = rnp->exp_seq_poll_rq;
1048 if (poll_state_synchronize_rcu(s))
1049 rnp->exp_seq_poll_rq = RCU_GET_STATE_COMPLETED;
1050 raw_spin_unlock_irqrestore(&rnp->exp_poll_lock, flags);
1051}
1052
1053/**
1054 * start_poll_synchronize_rcu_expedited - Snapshot current RCU state and start expedited grace period
1055 *
1056 * Returns a cookie to pass to a call to cond_synchronize_rcu(),
1057 * cond_synchronize_rcu_expedited(), or poll_state_synchronize_rcu(),
1058 * allowing them to determine whether or not any sort of grace period has
1059 * elapsed in the meantime. If the needed expedited grace period is not
1060 * already slated to start, initiates that grace period.
1061 */
1062unsigned long start_poll_synchronize_rcu_expedited(void)
1063{
1064 unsigned long flags;
1065 struct rcu_data *rdp;
1066 struct rcu_node *rnp;
1067 unsigned long s;
1068
1069 s = get_state_synchronize_rcu();
1070 rdp = per_cpu_ptr(&rcu_data, raw_smp_processor_id());
1071 rnp = rdp->mynode;
1072 if (rcu_init_invoked())
1073 raw_spin_lock_irqsave(&rnp->exp_poll_lock, flags);
1074 if (!poll_state_synchronize_rcu(s)) {
1075 if (rcu_init_invoked()) {
1076 rnp->exp_seq_poll_rq = s;
1077 queue_work(wq: rcu_gp_wq, work: &rnp->exp_poll_wq);
1078 }
1079 }
1080 if (rcu_init_invoked())
1081 raw_spin_unlock_irqrestore(&rnp->exp_poll_lock, flags);
1082
1083 return s;
1084}
1085EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu_expedited);
1086
1087/**
1088 * start_poll_synchronize_rcu_expedited_full - Take a full snapshot and start expedited grace period
1089 * @rgosp: Place to put snapshot of grace-period state
1090 *
1091 * Places the normal and expedited grace-period states in rgosp. This
1092 * state value can be passed to a later call to cond_synchronize_rcu_full()
1093 * or poll_state_synchronize_rcu_full() to determine whether or not a
1094 * grace period (whether normal or expedited) has elapsed in the meantime.
1095 * If the needed expedited grace period is not already slated to start,
1096 * initiates that grace period.
1097 */
1098void start_poll_synchronize_rcu_expedited_full(struct rcu_gp_oldstate *rgosp)
1099{
1100 get_state_synchronize_rcu_full(rgosp);
1101 (void)start_poll_synchronize_rcu_expedited();
1102}
1103EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu_expedited_full);
1104
1105/**
1106 * cond_synchronize_rcu_expedited - Conditionally wait for an expedited RCU grace period
1107 *
1108 * @oldstate: value from get_state_synchronize_rcu(), start_poll_synchronize_rcu(), or start_poll_synchronize_rcu_expedited()
1109 *
1110 * If any type of full RCU grace period has elapsed since the earlier
1111 * call to get_state_synchronize_rcu(), start_poll_synchronize_rcu(),
1112 * or start_poll_synchronize_rcu_expedited(), just return. Otherwise,
1113 * invoke synchronize_rcu_expedited() to wait for a full grace period.
1114 *
1115 * Yes, this function does not take counter wrap into account.
1116 * But counter wrap is harmless. If the counter wraps, we have waited for
1117 * more than 2 billion grace periods (and way more on a 64-bit system!),
1118 * so waiting for a couple of additional grace periods should be just fine.
1119 *
1120 * This function provides the same memory-ordering guarantees that
1121 * would be provided by a synchronize_rcu() that was invoked at the call
1122 * to the function that provided @oldstate and that returned at the end
1123 * of this function.
1124 */
1125void cond_synchronize_rcu_expedited(unsigned long oldstate)
1126{
1127 if (!poll_state_synchronize_rcu(oldstate))
1128 synchronize_rcu_expedited();
1129}
1130EXPORT_SYMBOL_GPL(cond_synchronize_rcu_expedited);
1131
1132/**
1133 * cond_synchronize_rcu_expedited_full - Conditionally wait for an expedited RCU grace period
1134 * @rgosp: value from get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(), or start_poll_synchronize_rcu_expedited_full()
1135 *
1136 * If a full RCU grace period has elapsed since the call to
1137 * get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(),
1138 * or start_poll_synchronize_rcu_expedited_full() from which @rgosp was
1139 * obtained, just return. Otherwise, invoke synchronize_rcu_expedited()
1140 * to wait for a full grace period.
1141 *
1142 * Yes, this function does not take counter wrap into account.
1143 * But counter wrap is harmless. If the counter wraps, we have waited for
1144 * more than 2 billion grace periods (and way more on a 64-bit system!),
1145 * so waiting for a couple of additional grace periods should be just fine.
1146 *
1147 * This function provides the same memory-ordering guarantees that
1148 * would be provided by a synchronize_rcu() that was invoked at the call
1149 * to the function that provided @rgosp and that returned at the end of
1150 * this function.
1151 */
1152void cond_synchronize_rcu_expedited_full(struct rcu_gp_oldstate *rgosp)
1153{
1154 if (!poll_state_synchronize_rcu_full(rgosp))
1155 synchronize_rcu_expedited();
1156}
1157EXPORT_SYMBOL_GPL(cond_synchronize_rcu_expedited_full);
1158

source code of linux/kernel/rcu/tree_exp.h