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