Loading...
Note: File does not exist in v3.1.
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/console.h>
11#include <linux/lockdep.h>
12
13static void rcu_exp_handler(void *unused);
14static int rcu_print_task_exp_stall(struct rcu_node *rnp);
15static void rcu_exp_print_detail_task_stall_rnp(struct rcu_node *rnp);
16
17/*
18 * Record the start of an expedited grace period.
19 */
20static void rcu_exp_gp_seq_start(void)
21{
22 rcu_seq_start(&rcu_state.expedited_sequence);
23 rcu_poll_gp_seq_start_unlocked(&rcu_state.gp_seq_polled_exp_snap);
24}
25
26/*
27 * Return the value that the expedited-grace-period counter will have
28 * at the end of the current grace period.
29 */
30static __maybe_unused unsigned long rcu_exp_gp_seq_endval(void)
31{
32 return rcu_seq_endval(&rcu_state.expedited_sequence);
33}
34
35/*
36 * Record the end of an expedited grace period.
37 */
38static void rcu_exp_gp_seq_end(void)
39{
40 rcu_poll_gp_seq_end_unlocked(&rcu_state.gp_seq_polled_exp_snap);
41 rcu_seq_end(&rcu_state.expedited_sequence);
42 smp_mb(); /* Ensure that consecutive grace periods serialize. */
43}
44
45/*
46 * Take a snapshot of the expedited-grace-period counter, which is the
47 * earliest value that will indicate that a full grace period has
48 * elapsed since the current time.
49 */
50static unsigned long rcu_exp_gp_seq_snap(void)
51{
52 unsigned long s;
53
54 smp_mb(); /* Caller's modifications seen first by other CPUs. */
55 s = rcu_seq_snap(&rcu_state.expedited_sequence);
56 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("snap"));
57 return s;
58}
59
60/*
61 * Given a counter snapshot from rcu_exp_gp_seq_snap(), return true
62 * if a full expedited grace period has elapsed since that snapshot
63 * was taken.
64 */
65static bool rcu_exp_gp_seq_done(unsigned long s)
66{
67 return rcu_seq_done(&rcu_state.expedited_sequence, s);
68}
69
70/*
71 * Reset the ->expmaskinit values in the rcu_node tree to reflect any
72 * recent CPU-online activity. Note that these masks are not cleared
73 * when CPUs go offline, so they reflect the union of all CPUs that have
74 * ever been online. This means that this function normally takes its
75 * no-work-to-do fastpath.
76 */
77static void sync_exp_reset_tree_hotplug(void)
78{
79 bool done;
80 unsigned long flags;
81 unsigned long mask;
82 unsigned long oldmask;
83 int ncpus = smp_load_acquire(&rcu_state.ncpus); /* Order vs. locking. */
84 struct rcu_node *rnp;
85 struct rcu_node *rnp_up;
86
87 /* If no new CPUs onlined since last time, nothing to do. */
88 if (likely(ncpus == rcu_state.ncpus_snap))
89 return;
90 rcu_state.ncpus_snap = ncpus;
91
92 /*
93 * Each pass through the following loop propagates newly onlined
94 * CPUs for the current rcu_node structure up the rcu_node tree.
95 */
96 rcu_for_each_leaf_node(rnp) {
97 raw_spin_lock_irqsave_rcu_node(rnp, flags);
98 if (rnp->expmaskinit == rnp->expmaskinitnext) {
99 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
100 continue; /* No new CPUs, nothing to do. */
101 }
102
103 /* Update this node's mask, track old value for propagation. */
104 oldmask = rnp->expmaskinit;
105 rnp->expmaskinit = rnp->expmaskinitnext;
106 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
107
108 /* If was already nonzero, nothing to propagate. */
109 if (oldmask)
110 continue;
111
112 /* Propagate the new CPU up the tree. */
113 mask = rnp->grpmask;
114 rnp_up = rnp->parent;
115 done = false;
116 while (rnp_up) {
117 raw_spin_lock_irqsave_rcu_node(rnp_up, flags);
118 if (rnp_up->expmaskinit)
119 done = true;
120 rnp_up->expmaskinit |= mask;
121 raw_spin_unlock_irqrestore_rcu_node(rnp_up, flags);
122 if (done)
123 break;
124 mask = rnp_up->grpmask;
125 rnp_up = rnp_up->parent;
126 }
127 }
128}
129
130/*
131 * Reset the ->expmask values in the rcu_node tree in preparation for
132 * a new expedited grace period.
133 */
134static void __maybe_unused sync_exp_reset_tree(void)
135{
136 unsigned long flags;
137 struct rcu_node *rnp;
138
139 sync_exp_reset_tree_hotplug();
140 rcu_for_each_node_breadth_first(rnp) {
141 raw_spin_lock_irqsave_rcu_node(rnp, flags);
142 WARN_ON_ONCE(rnp->expmask);
143 WRITE_ONCE(rnp->expmask, rnp->expmaskinit);
144 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
145 }
146}
147
148/*
149 * Return non-zero if there is no RCU expedited grace period in progress
150 * for the specified rcu_node structure, in other words, if all CPUs and
151 * tasks covered by the specified rcu_node structure have done their bit
152 * for the current expedited grace period.
153 */
154static bool sync_rcu_exp_done(struct rcu_node *rnp)
155{
156 raw_lockdep_assert_held_rcu_node(rnp);
157 return READ_ONCE(rnp->exp_tasks) == NULL &&
158 READ_ONCE(rnp->expmask) == 0;
159}
160
161/*
162 * Like sync_rcu_exp_done(), but where the caller does not hold the
163 * rcu_node's ->lock.
164 */
165static bool sync_rcu_exp_done_unlocked(struct rcu_node *rnp)
166{
167 unsigned long flags;
168 bool ret;
169
170 raw_spin_lock_irqsave_rcu_node(rnp, flags);
171 ret = sync_rcu_exp_done(rnp);
172 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
173
174 return ret;
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 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 /*
270 * Order GP completion with preceding accesses. Order also GP
271 * completion with post GP update side accesses. Pairs with
272 * rcu_seq_end().
273 */
274 smp_mb();
275 return true;
276 }
277 return false;
278}
279
280/*
281 * Funnel-lock acquisition for expedited grace periods. Returns true
282 * if some other task completed an expedited grace period that this task
283 * can piggy-back on, and with no mutex held. Otherwise, returns false
284 * with the mutex held, indicating that the caller must actually do the
285 * expedited grace period.
286 */
287static bool exp_funnel_lock(unsigned long s)
288{
289 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, raw_smp_processor_id());
290 struct rcu_node *rnp = rdp->mynode;
291 struct rcu_node *rnp_root = rcu_get_root();
292
293 /* Low-contention fastpath. */
294 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s) &&
295 (rnp == rnp_root ||
296 ULONG_CMP_LT(READ_ONCE(rnp_root->exp_seq_rq), s)) &&
297 mutex_trylock(&rcu_state.exp_mutex))
298 goto fastpath;
299
300 /*
301 * Each pass through the following loop works its way up
302 * the rcu_node tree, returning if others have done the work or
303 * otherwise falls through to acquire ->exp_mutex. The mapping
304 * from CPU to rcu_node structure can be inexact, as it is just
305 * promoting locality and is not strictly needed for correctness.
306 */
307 for (; rnp != NULL; rnp = rnp->parent) {
308 if (sync_exp_work_done(s))
309 return true;
310
311 /* Work not done, either wait here or go up. */
312 spin_lock(&rnp->exp_lock);
313 if (ULONG_CMP_GE(rnp->exp_seq_rq, s)) {
314
315 /* Someone else doing GP, so wait for them. */
316 spin_unlock(&rnp->exp_lock);
317 trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
318 rnp->grplo, rnp->grphi,
319 TPS("wait"));
320 wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
321 sync_exp_work_done(s));
322 return true;
323 }
324 WRITE_ONCE(rnp->exp_seq_rq, s); /* Followers can wait on us. */
325 spin_unlock(&rnp->exp_lock);
326 trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
327 rnp->grplo, rnp->grphi, TPS("nxtlvl"));
328 }
329 mutex_lock(&rcu_state.exp_mutex);
330fastpath:
331 if (sync_exp_work_done(s)) {
332 mutex_unlock(&rcu_state.exp_mutex);
333 return true;
334 }
335 rcu_exp_gp_seq_start();
336 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("start"));
337 return false;
338}
339
340/*
341 * Select the CPUs within the specified rcu_node that the upcoming
342 * expedited grace period needs to wait for.
343 */
344static void __sync_rcu_exp_select_node_cpus(struct rcu_exp_work *rewp)
345{
346 int cpu;
347 unsigned long flags;
348 unsigned long mask_ofl_test;
349 unsigned long mask_ofl_ipi;
350 int ret;
351 struct rcu_node *rnp = container_of(rewp, struct rcu_node, rew);
352
353 raw_spin_lock_irqsave_rcu_node(rnp, flags);
354
355 /* Each pass checks a CPU for identity, offline, and idle. */
356 mask_ofl_test = 0;
357 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) {
358 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
359 unsigned long mask = rdp->grpmask;
360 int snap;
361
362 if (raw_smp_processor_id() == cpu ||
363 !(rnp->qsmaskinitnext & mask)) {
364 mask_ofl_test |= mask;
365 } else {
366 /*
367 * Full ordering between remote CPU's post idle accesses
368 * and updater's accesses prior to current GP (and also
369 * the started GP sequence number) is enforced by
370 * rcu_seq_start() implicit barrier, relayed by kworkers
371 * locking and even further by smp_mb__after_unlock_lock()
372 * barriers chained all the way throughout the rnp locking
373 * tree since sync_exp_reset_tree() and up to the current
374 * leaf rnp locking.
375 *
376 * Ordering between remote CPU's pre idle accesses and
377 * post grace period updater's accesses is enforced by the
378 * below acquire semantic.
379 */
380 snap = ct_rcu_watching_cpu_acquire(cpu);
381 if (rcu_watching_snap_in_eqs(snap))
382 mask_ofl_test |= mask;
383 else
384 rdp->exp_watching_snap = snap;
385 }
386 }
387 mask_ofl_ipi = rnp->expmask & ~mask_ofl_test;
388
389 /*
390 * Need to wait for any blocked tasks as well. Note that
391 * additional blocking tasks will also block the expedited GP
392 * until such time as the ->expmask bits are cleared.
393 */
394 if (rcu_preempt_has_tasks(rnp))
395 WRITE_ONCE(rnp->exp_tasks, rnp->blkd_tasks.next);
396 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
397
398 /* IPI the remaining CPUs for expedited quiescent state. */
399 for_each_leaf_node_cpu_mask(rnp, cpu, mask_ofl_ipi) {
400 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
401 unsigned long mask = rdp->grpmask;
402
403retry_ipi:
404 if (rcu_watching_snap_stopped_since(rdp, rdp->exp_watching_snap)) {
405 mask_ofl_test |= mask;
406 continue;
407 }
408 if (get_cpu() == cpu) {
409 mask_ofl_test |= mask;
410 put_cpu();
411 continue;
412 }
413 ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
414 put_cpu();
415 /* The CPU will report the QS in response to the IPI. */
416 if (!ret)
417 continue;
418
419 /* Failed, raced with CPU hotplug operation. */
420 raw_spin_lock_irqsave_rcu_node(rnp, flags);
421 if ((rnp->qsmaskinitnext & mask) &&
422 (rnp->expmask & mask)) {
423 /* Online, so delay for a bit and try again. */
424 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
425 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("selectofl"));
426 schedule_timeout_idle(1);
427 goto retry_ipi;
428 }
429 /* CPU really is offline, so we must report its QS. */
430 if (rnp->expmask & mask)
431 mask_ofl_test |= mask;
432 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
433 }
434 /* Report quiescent states for those that went offline. */
435 if (mask_ofl_test)
436 rcu_report_exp_cpu_mult(rnp, mask_ofl_test, false);
437}
438
439static void rcu_exp_sel_wait_wake(unsigned long s);
440
441static void sync_rcu_exp_select_node_cpus(struct kthread_work *wp)
442{
443 struct rcu_exp_work *rewp =
444 container_of(wp, struct rcu_exp_work, rew_work);
445
446 __sync_rcu_exp_select_node_cpus(rewp);
447}
448
449static inline bool rcu_exp_worker_started(void)
450{
451 return !!READ_ONCE(rcu_exp_gp_kworker);
452}
453
454static inline bool rcu_exp_par_worker_started(struct rcu_node *rnp)
455{
456 return !!READ_ONCE(rnp->exp_kworker);
457}
458
459static inline void sync_rcu_exp_select_cpus_queue_work(struct rcu_node *rnp)
460{
461 kthread_init_work(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus);
462 /*
463 * Use rcu_exp_par_gp_kworker, because flushing a work item from
464 * another work item on the same kthread worker can result in
465 * deadlock.
466 */
467 kthread_queue_work(READ_ONCE(rnp->exp_kworker), &rnp->rew.rew_work);
468}
469
470static inline void sync_rcu_exp_select_cpus_flush_work(struct rcu_node *rnp)
471{
472 kthread_flush_work(&rnp->rew.rew_work);
473}
474
475/*
476 * Work-queue handler to drive an expedited grace period forward.
477 */
478static void wait_rcu_exp_gp(struct kthread_work *wp)
479{
480 struct rcu_exp_work *rewp;
481
482 rewp = container_of(wp, struct rcu_exp_work, rew_work);
483 rcu_exp_sel_wait_wake(rewp->rew_s);
484}
485
486static inline void synchronize_rcu_expedited_queue_work(struct rcu_exp_work *rew)
487{
488 kthread_init_work(&rew->rew_work, wait_rcu_exp_gp);
489 kthread_queue_work(rcu_exp_gp_kworker, &rew->rew_work);
490}
491
492/*
493 * Select the nodes that the upcoming expedited grace period needs
494 * to wait for.
495 */
496static void sync_rcu_exp_select_cpus(void)
497{
498 struct rcu_node *rnp;
499
500 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("reset"));
501 sync_exp_reset_tree();
502 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("select"));
503
504 /* Schedule work for each leaf rcu_node structure. */
505 rcu_for_each_leaf_node(rnp) {
506 rnp->exp_need_flush = false;
507 if (!READ_ONCE(rnp->expmask))
508 continue; /* Avoid early boot non-existent wq. */
509 if (!rcu_exp_par_worker_started(rnp) ||
510 rcu_scheduler_active != RCU_SCHEDULER_RUNNING ||
511 rcu_is_last_leaf_node(rnp)) {
512 /* No worker started yet or last leaf, do direct call. */
513 sync_rcu_exp_select_node_cpus(&rnp->rew.rew_work);
514 continue;
515 }
516 sync_rcu_exp_select_cpus_queue_work(rnp);
517 rnp->exp_need_flush = true;
518 }
519
520 /* Wait for jobs (if any) to complete. */
521 rcu_for_each_leaf_node(rnp)
522 if (rnp->exp_need_flush)
523 sync_rcu_exp_select_cpus_flush_work(rnp);
524}
525
526/*
527 * Wait for the expedited grace period to elapse, within time limit.
528 * If the time limit is exceeded without the grace period elapsing,
529 * return false, otherwise return true.
530 */
531static bool synchronize_rcu_expedited_wait_once(long tlimit)
532{
533 int t;
534 struct rcu_node *rnp_root = rcu_get_root();
535
536 t = swait_event_timeout_exclusive(rcu_state.expedited_wq,
537 sync_rcu_exp_done_unlocked(rnp_root),
538 tlimit);
539 // Workqueues should not be signaled.
540 if (t > 0 || sync_rcu_exp_done_unlocked(rnp_root))
541 return true;
542 WARN_ON(t < 0); /* workqueues should not be signaled. */
543 return false;
544}
545
546/*
547 * Print out an expedited RCU CPU stall warning message.
548 */
549static void synchronize_rcu_expedited_stall(unsigned long jiffies_start, unsigned long j)
550{
551 int cpu;
552 unsigned long mask;
553 int ndetected;
554 struct rcu_node *rnp;
555 struct rcu_node *rnp_root = rcu_get_root();
556
557 if (READ_ONCE(csd_lock_suppress_rcu_stall) && csd_lock_is_stuck()) {
558 pr_err("INFO: %s detected expedited stalls, but suppressed full report due to a stuck CSD-lock.\n", rcu_state.name);
559 return;
560 }
561 pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {", rcu_state.name);
562 ndetected = 0;
563 rcu_for_each_leaf_node(rnp) {
564 ndetected += rcu_print_task_exp_stall(rnp);
565 for_each_leaf_node_possible_cpu(rnp, cpu) {
566 struct rcu_data *rdp;
567
568 mask = leaf_node_cpu_bit(rnp, cpu);
569 if (!(READ_ONCE(rnp->expmask) & mask))
570 continue;
571 ndetected++;
572 rdp = per_cpu_ptr(&rcu_data, cpu);
573 pr_cont(" %d-%c%c%c%c", cpu,
574 "O."[!!cpu_online(cpu)],
575 "o."[!!(rdp->grpmask & rnp->expmaskinit)],
576 "N."[!!(rdp->grpmask & rnp->expmaskinitnext)],
577 "D."[!!data_race(rdp->cpu_no_qs.b.exp)]);
578 }
579 }
580 pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n",
581 j - jiffies_start, rcu_state.expedited_sequence, data_race(rnp_root->expmask),
582 ".T"[!!data_race(rnp_root->exp_tasks)]);
583 if (ndetected) {
584 pr_err("blocking rcu_node structures (internal RCU debug):");
585 rcu_for_each_node_breadth_first(rnp) {
586 if (rnp == rnp_root)
587 continue; /* printed unconditionally */
588 if (sync_rcu_exp_done_unlocked(rnp))
589 continue;
590 pr_cont(" l=%u:%d-%d:%#lx/%c",
591 rnp->level, rnp->grplo, rnp->grphi, data_race(rnp->expmask),
592 ".T"[!!data_race(rnp->exp_tasks)]);
593 }
594 pr_cont("\n");
595 }
596 rcu_for_each_leaf_node(rnp) {
597 for_each_leaf_node_possible_cpu(rnp, cpu) {
598 mask = leaf_node_cpu_bit(rnp, cpu);
599 if (!(READ_ONCE(rnp->expmask) & mask))
600 continue;
601 dump_cpu_task(cpu);
602 }
603 rcu_exp_print_detail_task_stall_rnp(rnp);
604 }
605}
606
607/*
608 * Wait for the expedited grace period to elapse, issuing any needed
609 * RCU CPU stall warnings along the way.
610 */
611static void synchronize_rcu_expedited_wait(void)
612{
613 int cpu;
614 unsigned long j;
615 unsigned long jiffies_stall;
616 unsigned long jiffies_start;
617 unsigned long mask;
618 struct rcu_data *rdp;
619 struct rcu_node *rnp;
620 unsigned long flags;
621
622 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("startwait"));
623 jiffies_stall = rcu_exp_jiffies_till_stall_check();
624 jiffies_start = jiffies;
625 if (tick_nohz_full_enabled() && rcu_inkernel_boot_has_ended()) {
626 if (synchronize_rcu_expedited_wait_once(1))
627 return;
628 rcu_for_each_leaf_node(rnp) {
629 raw_spin_lock_irqsave_rcu_node(rnp, flags);
630 mask = READ_ONCE(rnp->expmask);
631 for_each_leaf_node_cpu_mask(rnp, cpu, mask) {
632 rdp = per_cpu_ptr(&rcu_data, cpu);
633 if (rdp->rcu_forced_tick_exp)
634 continue;
635 rdp->rcu_forced_tick_exp = true;
636 if (cpu_online(cpu))
637 tick_dep_set_cpu(cpu, TICK_DEP_BIT_RCU_EXP);
638 }
639 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
640 }
641 j = READ_ONCE(jiffies_till_first_fqs);
642 if (synchronize_rcu_expedited_wait_once(j + HZ))
643 return;
644 }
645
646 for (;;) {
647 unsigned long j;
648
649 if (synchronize_rcu_expedited_wait_once(jiffies_stall))
650 return;
651 if (rcu_stall_is_suppressed())
652 continue;
653
654 nbcon_cpu_emergency_enter();
655
656 j = jiffies;
657 rcu_stall_notifier_call_chain(RCU_STALL_NOTIFY_EXP, (void *)(j - jiffies_start));
658 trace_rcu_stall_warning(rcu_state.name, TPS("ExpeditedStall"));
659 synchronize_rcu_expedited_stall(jiffies_start, j);
660 jiffies_stall = 3 * rcu_exp_jiffies_till_stall_check() + 3;
661
662 nbcon_cpu_emergency_exit();
663
664 panic_on_rcu_stall();
665 }
666}
667
668/*
669 * Wait for the current expedited grace period to complete, and then
670 * wake up everyone who piggybacked on the just-completed expedited
671 * grace period. Also update all the ->exp_seq_rq counters as needed
672 * in order to avoid counter-wrap problems.
673 */
674static void rcu_exp_wait_wake(unsigned long s)
675{
676 struct rcu_node *rnp;
677
678 synchronize_rcu_expedited_wait();
679
680 // Switch over to wakeup mode, allowing the next GP to proceed.
681 // End the previous grace period only after acquiring the mutex
682 // to ensure that only one GP runs concurrently with wakeups.
683 mutex_lock(&rcu_state.exp_wake_mutex);
684 rcu_exp_gp_seq_end();
685 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("end"));
686
687 rcu_for_each_node_breadth_first(rnp) {
688 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s)) {
689 spin_lock(&rnp->exp_lock);
690 /* Recheck, avoid hang in case someone just arrived. */
691 if (ULONG_CMP_LT(rnp->exp_seq_rq, s))
692 WRITE_ONCE(rnp->exp_seq_rq, s);
693 spin_unlock(&rnp->exp_lock);
694 }
695 smp_mb(); /* All above changes before wakeup. */
696 wake_up_all(&rnp->exp_wq[rcu_seq_ctr(s) & 0x3]);
697 }
698 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("endwake"));
699 mutex_unlock(&rcu_state.exp_wake_mutex);
700}
701
702/*
703 * Common code to drive an expedited grace period forward, used by
704 * workqueues and mid-boot-time tasks.
705 */
706static void rcu_exp_sel_wait_wake(unsigned long s)
707{
708 /* Initialize the rcu_node tree in preparation for the wait. */
709 sync_rcu_exp_select_cpus();
710
711 /* Wait and clean up, including waking everyone. */
712 rcu_exp_wait_wake(s);
713}
714
715#ifdef CONFIG_PREEMPT_RCU
716
717/*
718 * Remote handler for smp_call_function_single(). If there is an
719 * RCU read-side critical section in effect, request that the
720 * next rcu_read_unlock() record the quiescent state up the
721 * ->expmask fields in the rcu_node tree. Otherwise, immediately
722 * report the quiescent state.
723 */
724static void rcu_exp_handler(void *unused)
725{
726 int depth = rcu_preempt_depth();
727 unsigned long flags;
728 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
729 struct rcu_node *rnp = rdp->mynode;
730 struct task_struct *t = current;
731
732 /*
733 * First, the common case of not being in an RCU read-side
734 * critical section. If also enabled or idle, immediately
735 * report the quiescent state, otherwise defer.
736 */
737 if (!depth) {
738 if (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK)) ||
739 rcu_is_cpu_rrupt_from_idle()) {
740 rcu_report_exp_rdp(rdp);
741 } else {
742 WRITE_ONCE(rdp->cpu_no_qs.b.exp, true);
743 set_tsk_need_resched(t);
744 set_preempt_need_resched();
745 }
746 return;
747 }
748
749 /*
750 * Second, the less-common case of being in an RCU read-side
751 * critical section. In this case we can count on a future
752 * rcu_read_unlock(). However, this rcu_read_unlock() might
753 * execute on some other CPU, but in that case there will be
754 * a future context switch. Either way, if the expedited
755 * grace period is still waiting on this CPU, set ->deferred_qs
756 * so that the eventual quiescent state will be reported.
757 * Note that there is a large group of race conditions that
758 * can have caused this quiescent state to already have been
759 * reported, so we really do need to check ->expmask.
760 */
761 if (depth > 0) {
762 raw_spin_lock_irqsave_rcu_node(rnp, flags);
763 if (rnp->expmask & rdp->grpmask) {
764 WRITE_ONCE(rdp->cpu_no_qs.b.exp, true);
765 t->rcu_read_unlock_special.b.exp_hint = true;
766 }
767 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
768 return;
769 }
770
771 // Finally, negative nesting depth should not happen.
772 WARN_ON_ONCE(1);
773}
774
775/* PREEMPTION=y, so no PREEMPTION=n expedited grace period to clean up after. */
776static void sync_sched_exp_online_cleanup(int cpu)
777{
778}
779
780/*
781 * Scan the current list of tasks blocked within RCU read-side critical
782 * sections, printing out the tid of each that is blocking the current
783 * expedited grace period.
784 */
785static int rcu_print_task_exp_stall(struct rcu_node *rnp)
786{
787 unsigned long flags;
788 int ndetected = 0;
789 struct task_struct *t;
790
791 raw_spin_lock_irqsave_rcu_node(rnp, flags);
792 if (!rnp->exp_tasks) {
793 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
794 return 0;
795 }
796 t = list_entry(rnp->exp_tasks->prev,
797 struct task_struct, rcu_node_entry);
798 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
799 pr_cont(" P%d", t->pid);
800 ndetected++;
801 }
802 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
803 return ndetected;
804}
805
806/*
807 * Scan the current list of tasks blocked within RCU read-side critical
808 * sections, dumping the stack of each that is blocking the current
809 * expedited grace period.
810 */
811static void rcu_exp_print_detail_task_stall_rnp(struct rcu_node *rnp)
812{
813 unsigned long flags;
814 struct task_struct *t;
815
816 if (!rcu_exp_stall_task_details)
817 return;
818 raw_spin_lock_irqsave_rcu_node(rnp, flags);
819 if (!READ_ONCE(rnp->exp_tasks)) {
820 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
821 return;
822 }
823 t = list_entry(rnp->exp_tasks->prev,
824 struct task_struct, rcu_node_entry);
825 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
826 /*
827 * We could be printing a lot while holding a spinlock.
828 * Avoid triggering hard lockup.
829 */
830 touch_nmi_watchdog();
831 sched_show_task(t);
832 }
833 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
834}
835
836#else /* #ifdef CONFIG_PREEMPT_RCU */
837
838/* Request an expedited quiescent state. */
839static void rcu_exp_need_qs(void)
840{
841 __this_cpu_write(rcu_data.cpu_no_qs.b.exp, true);
842 /* Store .exp before .rcu_urgent_qs. */
843 smp_store_release(this_cpu_ptr(&rcu_data.rcu_urgent_qs), true);
844 set_tsk_need_resched(current);
845 set_preempt_need_resched();
846}
847
848/* Invoked on each online non-idle CPU for expedited quiescent state. */
849static void rcu_exp_handler(void *unused)
850{
851 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
852 struct rcu_node *rnp = rdp->mynode;
853 bool preempt_bh_enabled = !(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
854
855 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
856 __this_cpu_read(rcu_data.cpu_no_qs.b.exp))
857 return;
858 if (rcu_is_cpu_rrupt_from_idle() ||
859 (IS_ENABLED(CONFIG_PREEMPT_COUNT) && preempt_bh_enabled)) {
860 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
861 return;
862 }
863 rcu_exp_need_qs();
864}
865
866/* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
867static void sync_sched_exp_online_cleanup(int cpu)
868{
869 unsigned long flags;
870 int my_cpu;
871 struct rcu_data *rdp;
872 int ret;
873 struct rcu_node *rnp;
874
875 rdp = per_cpu_ptr(&rcu_data, cpu);
876 rnp = rdp->mynode;
877 my_cpu = get_cpu();
878 /* Quiescent state either not needed or already requested, leave. */
879 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
880 READ_ONCE(rdp->cpu_no_qs.b.exp)) {
881 put_cpu();
882 return;
883 }
884 /* Quiescent state needed on current CPU, so set it up locally. */
885 if (my_cpu == cpu) {
886 local_irq_save(flags);
887 rcu_exp_need_qs();
888 local_irq_restore(flags);
889 put_cpu();
890 return;
891 }
892 /* Quiescent state needed on some other CPU, send IPI. */
893 ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
894 put_cpu();
895 WARN_ON_ONCE(ret);
896}
897
898/*
899 * Because preemptible RCU does not exist, we never have to check for
900 * tasks blocked within RCU read-side critical sections that are
901 * blocking the current expedited grace period.
902 */
903static int rcu_print_task_exp_stall(struct rcu_node *rnp)
904{
905 return 0;
906}
907
908/*
909 * Because preemptible RCU does not exist, we never have to print out
910 * tasks blocked within RCU read-side critical sections that are blocking
911 * the current expedited grace period.
912 */
913static void rcu_exp_print_detail_task_stall_rnp(struct rcu_node *rnp)
914{
915}
916
917#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
918
919/**
920 * synchronize_rcu_expedited - Brute-force RCU grace period
921 *
922 * Wait for an RCU grace period, but expedite it. The basic idea is to
923 * IPI all non-idle non-nohz online CPUs. The IPI handler checks whether
924 * the CPU is in an RCU critical section, and if so, it sets a flag that
925 * causes the outermost rcu_read_unlock() to report the quiescent state
926 * for RCU-preempt or asks the scheduler for help for RCU-sched. On the
927 * other hand, if the CPU is not in an RCU read-side critical section,
928 * the IPI handler reports the quiescent state immediately.
929 *
930 * Although this is a great improvement over previous expedited
931 * implementations, it is still unfriendly to real-time workloads, so is
932 * thus not recommended for any sort of common-case code. In fact, if
933 * you are using synchronize_rcu_expedited() in a loop, please restructure
934 * your code to batch your updates, and then use a single synchronize_rcu()
935 * instead.
936 *
937 * This has the same semantics as (but is more brutal than) synchronize_rcu().
938 */
939void synchronize_rcu_expedited(void)
940{
941 unsigned long flags;
942 struct rcu_exp_work rew;
943 struct rcu_node *rnp;
944 unsigned long s;
945
946 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
947 lock_is_held(&rcu_lock_map) ||
948 lock_is_held(&rcu_sched_lock_map),
949 "Illegal synchronize_rcu_expedited() in RCU read-side critical section");
950
951 /* Is the state is such that the call is a grace period? */
952 if (rcu_blocking_is_gp()) {
953 // Note well that this code runs with !PREEMPT && !SMP.
954 // In addition, all code that advances grace periods runs
955 // at process level. Therefore, this expedited GP overlaps
956 // with other expedited GPs only by being fully nested within
957 // them, which allows reuse of ->gp_seq_polled_exp_snap.
958 rcu_poll_gp_seq_start_unlocked(&rcu_state.gp_seq_polled_exp_snap);
959 rcu_poll_gp_seq_end_unlocked(&rcu_state.gp_seq_polled_exp_snap);
960
961 local_irq_save(flags);
962 WARN_ON_ONCE(num_online_cpus() > 1);
963 rcu_state.expedited_sequence += (1 << RCU_SEQ_CTR_SHIFT);
964 local_irq_restore(flags);
965 return; // Context allows vacuous grace periods.
966 }
967
968 /* If expedited grace periods are prohibited, fall back to normal. */
969 if (rcu_gp_is_normal()) {
970 synchronize_rcu_normal();
971 return;
972 }
973
974 /* Take a snapshot of the sequence number. */
975 s = rcu_exp_gp_seq_snap();
976 if (exp_funnel_lock(s))
977 return; /* Someone else did our work for us. */
978
979 /* Ensure that load happens before action based on it. */
980 if (unlikely((rcu_scheduler_active == RCU_SCHEDULER_INIT) || !rcu_exp_worker_started())) {
981 /* Direct call during scheduler init and early_initcalls(). */
982 rcu_exp_sel_wait_wake(s);
983 } else {
984 /* Marshall arguments & schedule the expedited grace period. */
985 rew.rew_s = s;
986 synchronize_rcu_expedited_queue_work(&rew);
987 }
988
989 /* Wait for expedited grace period to complete. */
990 rnp = rcu_get_root();
991 wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
992 sync_exp_work_done(s));
993
994 /* Let the next expedited grace period start. */
995 mutex_unlock(&rcu_state.exp_mutex);
996}
997EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
998
999/*
1000 * Ensure that start_poll_synchronize_rcu_expedited() has the expedited
1001 * RCU grace periods that it needs.
1002 */
1003static void sync_rcu_do_polled_gp(struct work_struct *wp)
1004{
1005 unsigned long flags;
1006 int i = 0;
1007 struct rcu_node *rnp = container_of(wp, struct rcu_node, exp_poll_wq);
1008 unsigned long s;
1009
1010 raw_spin_lock_irqsave(&rnp->exp_poll_lock, flags);
1011 s = rnp->exp_seq_poll_rq;
1012 rnp->exp_seq_poll_rq = RCU_GET_STATE_COMPLETED;
1013 raw_spin_unlock_irqrestore(&rnp->exp_poll_lock, flags);
1014 if (s == RCU_GET_STATE_COMPLETED)
1015 return;
1016 while (!poll_state_synchronize_rcu(s)) {
1017 synchronize_rcu_expedited();
1018 if (i == 10 || i == 20)
1019 pr_info("%s: i = %d s = %lx gp_seq_polled = %lx\n", __func__, i, s, READ_ONCE(rcu_state.gp_seq_polled));
1020 i++;
1021 }
1022 raw_spin_lock_irqsave(&rnp->exp_poll_lock, flags);
1023 s = rnp->exp_seq_poll_rq;
1024 if (poll_state_synchronize_rcu(s))
1025 rnp->exp_seq_poll_rq = RCU_GET_STATE_COMPLETED;
1026 raw_spin_unlock_irqrestore(&rnp->exp_poll_lock, flags);
1027}
1028
1029/**
1030 * start_poll_synchronize_rcu_expedited - Snapshot current RCU state and start expedited grace period
1031 *
1032 * Returns a cookie to pass to a call to cond_synchronize_rcu(),
1033 * cond_synchronize_rcu_expedited(), or poll_state_synchronize_rcu(),
1034 * allowing them to determine whether or not any sort of grace period has
1035 * elapsed in the meantime. If the needed expedited grace period is not
1036 * already slated to start, initiates that grace period.
1037 */
1038unsigned long start_poll_synchronize_rcu_expedited(void)
1039{
1040 unsigned long flags;
1041 struct rcu_data *rdp;
1042 struct rcu_node *rnp;
1043 unsigned long s;
1044
1045 s = get_state_synchronize_rcu();
1046 rdp = per_cpu_ptr(&rcu_data, raw_smp_processor_id());
1047 rnp = rdp->mynode;
1048 if (rcu_init_invoked())
1049 raw_spin_lock_irqsave(&rnp->exp_poll_lock, flags);
1050 if (!poll_state_synchronize_rcu(s)) {
1051 if (rcu_init_invoked()) {
1052 rnp->exp_seq_poll_rq = s;
1053 queue_work(rcu_gp_wq, &rnp->exp_poll_wq);
1054 }
1055 }
1056 if (rcu_init_invoked())
1057 raw_spin_unlock_irqrestore(&rnp->exp_poll_lock, flags);
1058
1059 return s;
1060}
1061EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu_expedited);
1062
1063/**
1064 * start_poll_synchronize_rcu_expedited_full - Take a full snapshot and start expedited grace period
1065 * @rgosp: Place to put snapshot of grace-period state
1066 *
1067 * Places the normal and expedited grace-period states in rgosp. This
1068 * state value can be passed to a later call to cond_synchronize_rcu_full()
1069 * or poll_state_synchronize_rcu_full() to determine whether or not a
1070 * grace period (whether normal or expedited) has elapsed in the meantime.
1071 * If the needed expedited grace period is not already slated to start,
1072 * initiates that grace period.
1073 */
1074void start_poll_synchronize_rcu_expedited_full(struct rcu_gp_oldstate *rgosp)
1075{
1076 get_state_synchronize_rcu_full(rgosp);
1077 (void)start_poll_synchronize_rcu_expedited();
1078}
1079EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu_expedited_full);
1080
1081/**
1082 * cond_synchronize_rcu_expedited - Conditionally wait for an expedited RCU grace period
1083 *
1084 * @oldstate: value from get_state_synchronize_rcu(), start_poll_synchronize_rcu(), or start_poll_synchronize_rcu_expedited()
1085 *
1086 * If any type of full RCU grace period has elapsed since the earlier
1087 * call to get_state_synchronize_rcu(), start_poll_synchronize_rcu(),
1088 * or start_poll_synchronize_rcu_expedited(), just return. Otherwise,
1089 * invoke synchronize_rcu_expedited() to wait for a full grace period.
1090 *
1091 * Yes, this function does not take counter wrap into account.
1092 * But counter wrap is harmless. If the counter wraps, we have waited for
1093 * more than 2 billion grace periods (and way more on a 64-bit system!),
1094 * so waiting for a couple of additional grace periods should be just fine.
1095 *
1096 * This function provides the same memory-ordering guarantees that
1097 * would be provided by a synchronize_rcu() that was invoked at the call
1098 * to the function that provided @oldstate and that returned at the end
1099 * of this function.
1100 */
1101void cond_synchronize_rcu_expedited(unsigned long oldstate)
1102{
1103 if (!poll_state_synchronize_rcu(oldstate))
1104 synchronize_rcu_expedited();
1105}
1106EXPORT_SYMBOL_GPL(cond_synchronize_rcu_expedited);
1107
1108/**
1109 * cond_synchronize_rcu_expedited_full - Conditionally wait for an expedited RCU grace period
1110 * @rgosp: value from get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(), or start_poll_synchronize_rcu_expedited_full()
1111 *
1112 * If a full RCU grace period has elapsed since the call to
1113 * get_state_synchronize_rcu_full(), start_poll_synchronize_rcu_full(),
1114 * or start_poll_synchronize_rcu_expedited_full() from which @rgosp was
1115 * obtained, just return. Otherwise, invoke synchronize_rcu_expedited()
1116 * to wait for a full grace period.
1117 *
1118 * Yes, this function does not take counter wrap into account.
1119 * But counter wrap is harmless. If the counter wraps, we have waited for
1120 * more than 2 billion grace periods (and way more on a 64-bit system!),
1121 * so waiting for a couple of additional grace periods should be just fine.
1122 *
1123 * This function provides the same memory-ordering guarantees that
1124 * would be provided by a synchronize_rcu() that was invoked at the call
1125 * to the function that provided @rgosp and that returned at the end of
1126 * this function.
1127 */
1128void cond_synchronize_rcu_expedited_full(struct rcu_gp_oldstate *rgosp)
1129{
1130 if (!poll_state_synchronize_rcu_full(rgosp))
1131 synchronize_rcu_expedited();
1132}
1133EXPORT_SYMBOL_GPL(cond_synchronize_rcu_expedited_full);