Loading...
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);
14
15/*
16 * Record the start of an expedited grace period.
17 */
18static void rcu_exp_gp_seq_start(void)
19{
20 rcu_seq_start(&rcu_state.expedited_sequence);
21}
22
23/*
24 * Return then value that expedited-grace-period counter will have
25 * at the end of the current grace period.
26 */
27static __maybe_unused unsigned long rcu_exp_gp_seq_endval(void)
28{
29 return rcu_seq_endval(&rcu_state.expedited_sequence);
30}
31
32/*
33 * Record the end of an expedited grace period.
34 */
35static void rcu_exp_gp_seq_end(void)
36{
37 rcu_seq_end(&rcu_state.expedited_sequence);
38 smp_mb(); /* Ensure that consecutive grace periods serialize. */
39}
40
41/*
42 * Take a snapshot of the expedited-grace-period counter.
43 */
44static unsigned long rcu_exp_gp_seq_snap(void)
45{
46 unsigned long s;
47
48 smp_mb(); /* Caller's modifications seen first by other CPUs. */
49 s = rcu_seq_snap(&rcu_state.expedited_sequence);
50 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("snap"));
51 return s;
52}
53
54/*
55 * Given a counter snapshot from rcu_exp_gp_seq_snap(), return true
56 * if a full expedited grace period has elapsed since that snapshot
57 * was taken.
58 */
59static bool rcu_exp_gp_seq_done(unsigned long s)
60{
61 return rcu_seq_done(&rcu_state.expedited_sequence, s);
62}
63
64/*
65 * Reset the ->expmaskinit values in the rcu_node tree to reflect any
66 * recent CPU-online activity. Note that these masks are not cleared
67 * when CPUs go offline, so they reflect the union of all CPUs that have
68 * ever been online. This means that this function normally takes its
69 * no-work-to-do fastpath.
70 */
71static void sync_exp_reset_tree_hotplug(void)
72{
73 bool done;
74 unsigned long flags;
75 unsigned long mask;
76 unsigned long oldmask;
77 int ncpus = smp_load_acquire(&rcu_state.ncpus); /* Order vs. locking. */
78 struct rcu_node *rnp;
79 struct rcu_node *rnp_up;
80
81 /* If no new CPUs onlined since last time, nothing to do. */
82 if (likely(ncpus == rcu_state.ncpus_snap))
83 return;
84 rcu_state.ncpus_snap = ncpus;
85
86 /*
87 * Each pass through the following loop propagates newly onlined
88 * CPUs for the current rcu_node structure up the rcu_node tree.
89 */
90 rcu_for_each_leaf_node(rnp) {
91 raw_spin_lock_irqsave_rcu_node(rnp, flags);
92 if (rnp->expmaskinit == rnp->expmaskinitnext) {
93 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
94 continue; /* No new CPUs, nothing to do. */
95 }
96
97 /* Update this node's mask, track old value for propagation. */
98 oldmask = rnp->expmaskinit;
99 rnp->expmaskinit = rnp->expmaskinitnext;
100 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
101
102 /* If was already nonzero, nothing to propagate. */
103 if (oldmask)
104 continue;
105
106 /* Propagate the new CPU up the tree. */
107 mask = rnp->grpmask;
108 rnp_up = rnp->parent;
109 done = false;
110 while (rnp_up) {
111 raw_spin_lock_irqsave_rcu_node(rnp_up, flags);
112 if (rnp_up->expmaskinit)
113 done = true;
114 rnp_up->expmaskinit |= mask;
115 raw_spin_unlock_irqrestore_rcu_node(rnp_up, flags);
116 if (done)
117 break;
118 mask = rnp_up->grpmask;
119 rnp_up = rnp_up->parent;
120 }
121 }
122}
123
124/*
125 * Reset the ->expmask values in the rcu_node tree in preparation for
126 * a new expedited grace period.
127 */
128static void __maybe_unused sync_exp_reset_tree(void)
129{
130 unsigned long flags;
131 struct rcu_node *rnp;
132
133 sync_exp_reset_tree_hotplug();
134 rcu_for_each_node_breadth_first(rnp) {
135 raw_spin_lock_irqsave_rcu_node(rnp, flags);
136 WARN_ON_ONCE(rnp->expmask);
137 rnp->expmask = rnp->expmaskinit;
138 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
139 }
140}
141
142/*
143 * Return non-zero if there is no RCU expedited grace period in progress
144 * for the specified rcu_node structure, in other words, if all CPUs and
145 * tasks covered by the specified rcu_node structure have done their bit
146 * for the current expedited grace period. Works only for preemptible
147 * RCU -- other RCU implementation use other means.
148 *
149 * Caller must hold the specificed rcu_node structure's ->lock
150 */
151static bool sync_rcu_preempt_exp_done(struct rcu_node *rnp)
152{
153 raw_lockdep_assert_held_rcu_node(rnp);
154
155 return rnp->exp_tasks == NULL &&
156 READ_ONCE(rnp->expmask) == 0;
157}
158
159/*
160 * Like sync_rcu_preempt_exp_done(), but this function assumes the caller
161 * doesn't hold the rcu_node's ->lock, and will acquire and release the lock
162 * itself
163 */
164static bool sync_rcu_preempt_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_preempt_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 *
185 * Caller must hold the specified rcu_node structure's ->lock.
186 */
187static void __rcu_report_exp_rnp(struct rcu_node *rnp,
188 bool wake, unsigned long flags)
189 __releases(rnp->lock)
190{
191 unsigned long mask;
192
193 for (;;) {
194 if (!sync_rcu_preempt_exp_done(rnp)) {
195 if (!rnp->expmask)
196 rcu_initiate_boost(rnp, flags);
197 else
198 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
199 break;
200 }
201 if (rnp->parent == NULL) {
202 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
203 if (wake) {
204 smp_mb(); /* EGP done before wake_up(). */
205 swake_up_one(&rcu_state.expedited_wq);
206 }
207 break;
208 }
209 mask = rnp->grpmask;
210 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled */
211 rnp = rnp->parent;
212 raw_spin_lock_rcu_node(rnp); /* irqs already disabled */
213 WARN_ON_ONCE(!(rnp->expmask & mask));
214 rnp->expmask &= ~mask;
215 }
216}
217
218/*
219 * Report expedited quiescent state for specified node. This is a
220 * lock-acquisition wrapper function for __rcu_report_exp_rnp().
221 */
222static void __maybe_unused rcu_report_exp_rnp(struct rcu_node *rnp, bool wake)
223{
224 unsigned long flags;
225
226 raw_spin_lock_irqsave_rcu_node(rnp, flags);
227 __rcu_report_exp_rnp(rnp, wake, flags);
228}
229
230/*
231 * Report expedited quiescent state for multiple CPUs, all covered by the
232 * specified leaf rcu_node structure.
233 */
234static void rcu_report_exp_cpu_mult(struct rcu_node *rnp,
235 unsigned long mask, bool wake)
236{
237 unsigned long flags;
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 rnp->expmask &= ~mask;
245 __rcu_report_exp_rnp(rnp, wake, flags); /* Releases rnp->lock. */
246}
247
248/*
249 * Report expedited quiescent state for specified rcu_data (CPU).
250 */
251static void rcu_report_exp_rdp(struct rcu_data *rdp)
252{
253 WRITE_ONCE(rdp->exp_deferred_qs, false);
254 rcu_report_exp_cpu_mult(rdp->mynode, rdp->grpmask, true);
255}
256
257/* Common code for work-done checking. */
258static bool sync_exp_work_done(unsigned long s)
259{
260 if (rcu_exp_gp_seq_done(s)) {
261 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("done"));
262 smp_mb(); /* Ensure test happens before caller kfree(). */
263 return true;
264 }
265 return false;
266}
267
268/*
269 * Funnel-lock acquisition for expedited grace periods. Returns true
270 * if some other task completed an expedited grace period that this task
271 * can piggy-back on, and with no mutex held. Otherwise, returns false
272 * with the mutex held, indicating that the caller must actually do the
273 * expedited grace period.
274 */
275static bool exp_funnel_lock(unsigned long s)
276{
277 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, raw_smp_processor_id());
278 struct rcu_node *rnp = rdp->mynode;
279 struct rcu_node *rnp_root = rcu_get_root();
280
281 /* Low-contention fastpath. */
282 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s) &&
283 (rnp == rnp_root ||
284 ULONG_CMP_LT(READ_ONCE(rnp_root->exp_seq_rq), s)) &&
285 mutex_trylock(&rcu_state.exp_mutex))
286 goto fastpath;
287
288 /*
289 * Each pass through the following loop works its way up
290 * the rcu_node tree, returning if others have done the work or
291 * otherwise falls through to acquire ->exp_mutex. The mapping
292 * from CPU to rcu_node structure can be inexact, as it is just
293 * promoting locality and is not strictly needed for correctness.
294 */
295 for (; rnp != NULL; rnp = rnp->parent) {
296 if (sync_exp_work_done(s))
297 return true;
298
299 /* Work not done, either wait here or go up. */
300 spin_lock(&rnp->exp_lock);
301 if (ULONG_CMP_GE(rnp->exp_seq_rq, s)) {
302
303 /* Someone else doing GP, so wait for them. */
304 spin_unlock(&rnp->exp_lock);
305 trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
306 rnp->grplo, rnp->grphi,
307 TPS("wait"));
308 wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
309 sync_exp_work_done(s));
310 return true;
311 }
312 rnp->exp_seq_rq = s; /* Followers can wait on us. */
313 spin_unlock(&rnp->exp_lock);
314 trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
315 rnp->grplo, rnp->grphi, TPS("nxtlvl"));
316 }
317 mutex_lock(&rcu_state.exp_mutex);
318fastpath:
319 if (sync_exp_work_done(s)) {
320 mutex_unlock(&rcu_state.exp_mutex);
321 return true;
322 }
323 rcu_exp_gp_seq_start();
324 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("start"));
325 return false;
326}
327
328/*
329 * Select the CPUs within the specified rcu_node that the upcoming
330 * expedited grace period needs to wait for.
331 */
332static void sync_rcu_exp_select_node_cpus(struct work_struct *wp)
333{
334 int cpu;
335 unsigned long flags;
336 unsigned long mask_ofl_test;
337 unsigned long mask_ofl_ipi;
338 int ret;
339 struct rcu_exp_work *rewp =
340 container_of(wp, struct rcu_exp_work, rew_work);
341 struct rcu_node *rnp = container_of(rewp, struct rcu_node, rew);
342
343 raw_spin_lock_irqsave_rcu_node(rnp, flags);
344
345 /* Each pass checks a CPU for identity, offline, and idle. */
346 mask_ofl_test = 0;
347 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) {
348 unsigned long mask = leaf_node_cpu_bit(rnp, cpu);
349 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
350 int snap;
351
352 if (raw_smp_processor_id() == cpu ||
353 !(rnp->qsmaskinitnext & mask)) {
354 mask_ofl_test |= mask;
355 } else {
356 snap = rcu_dynticks_snap(rdp);
357 if (rcu_dynticks_in_eqs(snap))
358 mask_ofl_test |= mask;
359 else
360 rdp->exp_dynticks_snap = snap;
361 }
362 }
363 mask_ofl_ipi = rnp->expmask & ~mask_ofl_test;
364
365 /*
366 * Need to wait for any blocked tasks as well. Note that
367 * additional blocking tasks will also block the expedited GP
368 * until such time as the ->expmask bits are cleared.
369 */
370 if (rcu_preempt_has_tasks(rnp))
371 rnp->exp_tasks = rnp->blkd_tasks.next;
372 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
373
374 /* IPI the remaining CPUs for expedited quiescent state. */
375 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) {
376 unsigned long mask = leaf_node_cpu_bit(rnp, cpu);
377 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
378
379 if (!(mask_ofl_ipi & mask))
380 continue;
381retry_ipi:
382 if (rcu_dynticks_in_eqs_since(rdp, rdp->exp_dynticks_snap)) {
383 mask_ofl_test |= mask;
384 continue;
385 }
386 if (get_cpu() == cpu) {
387 put_cpu();
388 continue;
389 }
390 ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
391 put_cpu();
392 if (!ret) {
393 mask_ofl_ipi &= ~mask;
394 continue;
395 }
396 /* Failed, raced with CPU hotplug operation. */
397 raw_spin_lock_irqsave_rcu_node(rnp, flags);
398 if ((rnp->qsmaskinitnext & mask) &&
399 (rnp->expmask & mask)) {
400 /* Online, so delay for a bit and try again. */
401 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
402 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("selectofl"));
403 schedule_timeout_uninterruptible(1);
404 goto retry_ipi;
405 }
406 /* CPU really is offline, so we can ignore it. */
407 if (!(rnp->expmask & mask))
408 mask_ofl_ipi &= ~mask;
409 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
410 }
411 /* Report quiescent states for those that went offline. */
412 mask_ofl_test |= mask_ofl_ipi;
413 if (mask_ofl_test)
414 rcu_report_exp_cpu_mult(rnp, mask_ofl_test, false);
415}
416
417/*
418 * Select the nodes that the upcoming expedited grace period needs
419 * to wait for.
420 */
421static void sync_rcu_exp_select_cpus(void)
422{
423 int cpu;
424 struct rcu_node *rnp;
425
426 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("reset"));
427 sync_exp_reset_tree();
428 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("select"));
429
430 /* Schedule work for each leaf rcu_node structure. */
431 rcu_for_each_leaf_node(rnp) {
432 rnp->exp_need_flush = false;
433 if (!READ_ONCE(rnp->expmask))
434 continue; /* Avoid early boot non-existent wq. */
435 if (!READ_ONCE(rcu_par_gp_wq) ||
436 rcu_scheduler_active != RCU_SCHEDULER_RUNNING ||
437 rcu_is_last_leaf_node(rnp)) {
438 /* No workqueues yet or last leaf, do direct call. */
439 sync_rcu_exp_select_node_cpus(&rnp->rew.rew_work);
440 continue;
441 }
442 INIT_WORK(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus);
443 cpu = find_next_bit(&rnp->ffmask, BITS_PER_LONG, -1);
444 /* If all offline, queue the work on an unbound CPU. */
445 if (unlikely(cpu > rnp->grphi - rnp->grplo))
446 cpu = WORK_CPU_UNBOUND;
447 else
448 cpu += rnp->grplo;
449 queue_work_on(cpu, rcu_par_gp_wq, &rnp->rew.rew_work);
450 rnp->exp_need_flush = true;
451 }
452
453 /* Wait for workqueue jobs (if any) to complete. */
454 rcu_for_each_leaf_node(rnp)
455 if (rnp->exp_need_flush)
456 flush_work(&rnp->rew.rew_work);
457}
458
459static void synchronize_sched_expedited_wait(void)
460{
461 int cpu;
462 unsigned long jiffies_stall;
463 unsigned long jiffies_start;
464 unsigned long mask;
465 int ndetected;
466 struct rcu_node *rnp;
467 struct rcu_node *rnp_root = rcu_get_root();
468 int ret;
469
470 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("startwait"));
471 jiffies_stall = rcu_jiffies_till_stall_check();
472 jiffies_start = jiffies;
473
474 for (;;) {
475 ret = swait_event_timeout_exclusive(
476 rcu_state.expedited_wq,
477 sync_rcu_preempt_exp_done_unlocked(rnp_root),
478 jiffies_stall);
479 if (ret > 0 || sync_rcu_preempt_exp_done_unlocked(rnp_root))
480 return;
481 WARN_ON(ret < 0); /* workqueues should not be signaled. */
482 if (rcu_cpu_stall_suppress)
483 continue;
484 panic_on_rcu_stall();
485 pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
486 rcu_state.name);
487 ndetected = 0;
488 rcu_for_each_leaf_node(rnp) {
489 ndetected += rcu_print_task_exp_stall(rnp);
490 for_each_leaf_node_possible_cpu(rnp, cpu) {
491 struct rcu_data *rdp;
492
493 mask = leaf_node_cpu_bit(rnp, cpu);
494 if (!(rnp->expmask & mask))
495 continue;
496 ndetected++;
497 rdp = per_cpu_ptr(&rcu_data, cpu);
498 pr_cont(" %d-%c%c%c", cpu,
499 "O."[!!cpu_online(cpu)],
500 "o."[!!(rdp->grpmask & rnp->expmaskinit)],
501 "N."[!!(rdp->grpmask & rnp->expmaskinitnext)]);
502 }
503 }
504 pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n",
505 jiffies - jiffies_start, rcu_state.expedited_sequence,
506 rnp_root->expmask, ".T"[!!rnp_root->exp_tasks]);
507 if (ndetected) {
508 pr_err("blocking rcu_node structures:");
509 rcu_for_each_node_breadth_first(rnp) {
510 if (rnp == rnp_root)
511 continue; /* printed unconditionally */
512 if (sync_rcu_preempt_exp_done_unlocked(rnp))
513 continue;
514 pr_cont(" l=%u:%d-%d:%#lx/%c",
515 rnp->level, rnp->grplo, rnp->grphi,
516 rnp->expmask,
517 ".T"[!!rnp->exp_tasks]);
518 }
519 pr_cont("\n");
520 }
521 rcu_for_each_leaf_node(rnp) {
522 for_each_leaf_node_possible_cpu(rnp, cpu) {
523 mask = leaf_node_cpu_bit(rnp, cpu);
524 if (!(rnp->expmask & mask))
525 continue;
526 dump_cpu_task(cpu);
527 }
528 }
529 jiffies_stall = 3 * rcu_jiffies_till_stall_check() + 3;
530 }
531}
532
533/*
534 * Wait for the current expedited grace period to complete, and then
535 * wake up everyone who piggybacked on the just-completed expedited
536 * grace period. Also update all the ->exp_seq_rq counters as needed
537 * in order to avoid counter-wrap problems.
538 */
539static void rcu_exp_wait_wake(unsigned long s)
540{
541 struct rcu_node *rnp;
542
543 synchronize_sched_expedited_wait();
544 rcu_exp_gp_seq_end();
545 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("end"));
546
547 /*
548 * Switch over to wakeup mode, allowing the next GP, but -only- the
549 * next GP, to proceed.
550 */
551 mutex_lock(&rcu_state.exp_wake_mutex);
552
553 rcu_for_each_node_breadth_first(rnp) {
554 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s)) {
555 spin_lock(&rnp->exp_lock);
556 /* Recheck, avoid hang in case someone just arrived. */
557 if (ULONG_CMP_LT(rnp->exp_seq_rq, s))
558 rnp->exp_seq_rq = s;
559 spin_unlock(&rnp->exp_lock);
560 }
561 smp_mb(); /* All above changes before wakeup. */
562 wake_up_all(&rnp->exp_wq[rcu_seq_ctr(rcu_state.expedited_sequence) & 0x3]);
563 }
564 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("endwake"));
565 mutex_unlock(&rcu_state.exp_wake_mutex);
566}
567
568/*
569 * Common code to drive an expedited grace period forward, used by
570 * workqueues and mid-boot-time tasks.
571 */
572static void rcu_exp_sel_wait_wake(unsigned long s)
573{
574 /* Initialize the rcu_node tree in preparation for the wait. */
575 sync_rcu_exp_select_cpus();
576
577 /* Wait and clean up, including waking everyone. */
578 rcu_exp_wait_wake(s);
579}
580
581/*
582 * Work-queue handler to drive an expedited grace period forward.
583 */
584static void wait_rcu_exp_gp(struct work_struct *wp)
585{
586 struct rcu_exp_work *rewp;
587
588 rewp = container_of(wp, struct rcu_exp_work, rew_work);
589 rcu_exp_sel_wait_wake(rewp->rew_s);
590}
591
592#ifdef CONFIG_PREEMPT_RCU
593
594/*
595 * Remote handler for smp_call_function_single(). If there is an
596 * RCU read-side critical section in effect, request that the
597 * next rcu_read_unlock() record the quiescent state up the
598 * ->expmask fields in the rcu_node tree. Otherwise, immediately
599 * report the quiescent state.
600 */
601static void rcu_exp_handler(void *unused)
602{
603 unsigned long flags;
604 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
605 struct rcu_node *rnp = rdp->mynode;
606 struct task_struct *t = current;
607
608 /*
609 * First, the common case of not being in an RCU read-side
610 * critical section. If also enabled or idle, immediately
611 * report the quiescent state, otherwise defer.
612 */
613 if (!t->rcu_read_lock_nesting) {
614 if (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK)) ||
615 rcu_dynticks_curr_cpu_in_eqs()) {
616 rcu_report_exp_rdp(rdp);
617 } else {
618 rdp->exp_deferred_qs = true;
619 set_tsk_need_resched(t);
620 set_preempt_need_resched();
621 }
622 return;
623 }
624
625 /*
626 * Second, the less-common case of being in an RCU read-side
627 * critical section. In this case we can count on a future
628 * rcu_read_unlock(). However, this rcu_read_unlock() might
629 * execute on some other CPU, but in that case there will be
630 * a future context switch. Either way, if the expedited
631 * grace period is still waiting on this CPU, set ->deferred_qs
632 * so that the eventual quiescent state will be reported.
633 * Note that there is a large group of race conditions that
634 * can have caused this quiescent state to already have been
635 * reported, so we really do need to check ->expmask.
636 */
637 if (t->rcu_read_lock_nesting > 0) {
638 raw_spin_lock_irqsave_rcu_node(rnp, flags);
639 if (rnp->expmask & rdp->grpmask) {
640 rdp->exp_deferred_qs = true;
641 t->rcu_read_unlock_special.b.exp_hint = true;
642 }
643 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
644 return;
645 }
646
647 /*
648 * The final and least likely case is where the interrupted
649 * code was just about to or just finished exiting the RCU-preempt
650 * read-side critical section, and no, we can't tell which.
651 * So either way, set ->deferred_qs to flag later code that
652 * a quiescent state is required.
653 *
654 * If the CPU is fully enabled (or if some buggy RCU-preempt
655 * read-side critical section is being used from idle), just
656 * invoke rcu_preempt_deferred_qs() to immediately report the
657 * quiescent state. We cannot use rcu_read_unlock_special()
658 * because we are in an interrupt handler, which will cause that
659 * function to take an early exit without doing anything.
660 *
661 * Otherwise, force a context switch after the CPU enables everything.
662 */
663 rdp->exp_deferred_qs = true;
664 if (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK)) ||
665 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs())) {
666 rcu_preempt_deferred_qs(t);
667 } else {
668 set_tsk_need_resched(t);
669 set_preempt_need_resched();
670 }
671}
672
673/* PREEMPT=y, so no PREEMPT=n expedited grace period to clean up after. */
674static void sync_sched_exp_online_cleanup(int cpu)
675{
676}
677
678/*
679 * Scan the current list of tasks blocked within RCU read-side critical
680 * sections, printing out the tid of each that is blocking the current
681 * expedited grace period.
682 */
683static int rcu_print_task_exp_stall(struct rcu_node *rnp)
684{
685 struct task_struct *t;
686 int ndetected = 0;
687
688 if (!rnp->exp_tasks)
689 return 0;
690 t = list_entry(rnp->exp_tasks->prev,
691 struct task_struct, rcu_node_entry);
692 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
693 pr_cont(" P%d", t->pid);
694 ndetected++;
695 }
696 return ndetected;
697}
698
699#else /* #ifdef CONFIG_PREEMPT_RCU */
700
701/* Request an expedited quiescent state. */
702static void rcu_exp_need_qs(void)
703{
704 __this_cpu_write(rcu_data.cpu_no_qs.b.exp, true);
705 /* Store .exp before .rcu_urgent_qs. */
706 smp_store_release(this_cpu_ptr(&rcu_data.rcu_urgent_qs), true);
707 set_tsk_need_resched(current);
708 set_preempt_need_resched();
709}
710
711/* Invoked on each online non-idle CPU for expedited quiescent state. */
712static void rcu_exp_handler(void *unused)
713{
714 struct rcu_data *rdp;
715 struct rcu_node *rnp;
716
717 rdp = this_cpu_ptr(&rcu_data);
718 rnp = rdp->mynode;
719 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
720 __this_cpu_read(rcu_data.cpu_no_qs.b.exp))
721 return;
722 if (rcu_is_cpu_rrupt_from_idle()) {
723 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
724 return;
725 }
726 rcu_exp_need_qs();
727}
728
729/* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
730static void sync_sched_exp_online_cleanup(int cpu)
731{
732 unsigned long flags;
733 int my_cpu;
734 struct rcu_data *rdp;
735 int ret;
736 struct rcu_node *rnp;
737
738 rdp = per_cpu_ptr(&rcu_data, cpu);
739 rnp = rdp->mynode;
740 my_cpu = get_cpu();
741 /* Quiescent state either not needed or already requested, leave. */
742 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
743 __this_cpu_read(rcu_data.cpu_no_qs.b.exp)) {
744 put_cpu();
745 return;
746 }
747 /* Quiescent state needed on current CPU, so set it up locally. */
748 if (my_cpu == cpu) {
749 local_irq_save(flags);
750 rcu_exp_need_qs();
751 local_irq_restore(flags);
752 put_cpu();
753 return;
754 }
755 /* Quiescent state needed on some other CPU, send IPI. */
756 ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
757 put_cpu();
758 WARN_ON_ONCE(ret);
759}
760
761/*
762 * Because preemptible RCU does not exist, we never have to check for
763 * tasks blocked within RCU read-side critical sections that are
764 * blocking the current expedited grace period.
765 */
766static int rcu_print_task_exp_stall(struct rcu_node *rnp)
767{
768 return 0;
769}
770
771#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
772
773/**
774 * synchronize_rcu_expedited - Brute-force RCU grace period
775 *
776 * Wait for an RCU grace period, but expedite it. The basic idea is to
777 * IPI all non-idle non-nohz online CPUs. The IPI handler checks whether
778 * the CPU is in an RCU critical section, and if so, it sets a flag that
779 * causes the outermost rcu_read_unlock() to report the quiescent state
780 * for RCU-preempt or asks the scheduler for help for RCU-sched. On the
781 * other hand, if the CPU is not in an RCU read-side critical section,
782 * the IPI handler reports the quiescent state immediately.
783 *
784 * Although this is a great improvement over previous expedited
785 * implementations, it is still unfriendly to real-time workloads, so is
786 * thus not recommended for any sort of common-case code. In fact, if
787 * you are using synchronize_rcu_expedited() in a loop, please restructure
788 * your code to batch your updates, and then Use a single synchronize_rcu()
789 * instead.
790 *
791 * This has the same semantics as (but is more brutal than) synchronize_rcu().
792 */
793void synchronize_rcu_expedited(void)
794{
795 bool boottime = (rcu_scheduler_active == RCU_SCHEDULER_INIT);
796 struct rcu_exp_work rew;
797 struct rcu_node *rnp;
798 unsigned long s;
799
800 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
801 lock_is_held(&rcu_lock_map) ||
802 lock_is_held(&rcu_sched_lock_map),
803 "Illegal synchronize_rcu_expedited() in RCU read-side critical section");
804
805 /* Is the state is such that the call is a grace period? */
806 if (rcu_blocking_is_gp())
807 return;
808
809 /* If expedited grace periods are prohibited, fall back to normal. */
810 if (rcu_gp_is_normal()) {
811 wait_rcu_gp(call_rcu);
812 return;
813 }
814
815 /* Take a snapshot of the sequence number. */
816 s = rcu_exp_gp_seq_snap();
817 if (exp_funnel_lock(s))
818 return; /* Someone else did our work for us. */
819
820 /* Ensure that load happens before action based on it. */
821 if (unlikely(boottime)) {
822 /* Direct call during scheduler init and early_initcalls(). */
823 rcu_exp_sel_wait_wake(s);
824 } else {
825 /* Marshall arguments & schedule the expedited grace period. */
826 rew.rew_s = s;
827 INIT_WORK_ONSTACK(&rew.rew_work, wait_rcu_exp_gp);
828 queue_work(rcu_gp_wq, &rew.rew_work);
829 }
830
831 /* Wait for expedited grace period to complete. */
832 rnp = rcu_get_root();
833 wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
834 sync_exp_work_done(s));
835 smp_mb(); /* Workqueue actions happen before return. */
836
837 /* Let the next expedited grace period start. */
838 mutex_unlock(&rcu_state.exp_mutex);
839
840 if (likely(!boottime))
841 destroy_work_on_stack(&rew.rew_work);
842}
843EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
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);
14
15/*
16 * Record the start of an expedited grace period.
17 */
18static void rcu_exp_gp_seq_start(void)
19{
20 rcu_seq_start(&rcu_state.expedited_sequence);
21}
22
23/*
24 * Return the value that the expedited-grace-period counter will have
25 * at the end of the current grace period.
26 */
27static __maybe_unused unsigned long rcu_exp_gp_seq_endval(void)
28{
29 return rcu_seq_endval(&rcu_state.expedited_sequence);
30}
31
32/*
33 * Record the end of an expedited grace period.
34 */
35static void rcu_exp_gp_seq_end(void)
36{
37 rcu_seq_end(&rcu_state.expedited_sequence);
38 smp_mb(); /* Ensure that consecutive grace periods serialize. */
39}
40
41/*
42 * Take a snapshot of the expedited-grace-period counter, which is the
43 * earliest value that will indicate that a full grace period has
44 * elapsed since the current time.
45 */
46static unsigned long rcu_exp_gp_seq_snap(void)
47{
48 unsigned long s;
49
50 smp_mb(); /* Caller's modifications seen first by other CPUs. */
51 s = rcu_seq_snap(&rcu_state.expedited_sequence);
52 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("snap"));
53 return s;
54}
55
56/*
57 * Given a counter snapshot from rcu_exp_gp_seq_snap(), return true
58 * if a full expedited grace period has elapsed since that snapshot
59 * was taken.
60 */
61static bool rcu_exp_gp_seq_done(unsigned long s)
62{
63 return rcu_seq_done(&rcu_state.expedited_sequence, s);
64}
65
66/*
67 * Reset the ->expmaskinit values in the rcu_node tree to reflect any
68 * recent CPU-online activity. Note that these masks are not cleared
69 * when CPUs go offline, so they reflect the union of all CPUs that have
70 * ever been online. This means that this function normally takes its
71 * no-work-to-do fastpath.
72 */
73static void sync_exp_reset_tree_hotplug(void)
74{
75 bool done;
76 unsigned long flags;
77 unsigned long mask;
78 unsigned long oldmask;
79 int ncpus = smp_load_acquire(&rcu_state.ncpus); /* Order vs. locking. */
80 struct rcu_node *rnp;
81 struct rcu_node *rnp_up;
82
83 /* If no new CPUs onlined since last time, nothing to do. */
84 if (likely(ncpus == rcu_state.ncpus_snap))
85 return;
86 rcu_state.ncpus_snap = ncpus;
87
88 /*
89 * Each pass through the following loop propagates newly onlined
90 * CPUs for the current rcu_node structure up the rcu_node tree.
91 */
92 rcu_for_each_leaf_node(rnp) {
93 raw_spin_lock_irqsave_rcu_node(rnp, flags);
94 if (rnp->expmaskinit == rnp->expmaskinitnext) {
95 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
96 continue; /* No new CPUs, nothing to do. */
97 }
98
99 /* Update this node's mask, track old value for propagation. */
100 oldmask = rnp->expmaskinit;
101 rnp->expmaskinit = rnp->expmaskinitnext;
102 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
103
104 /* If was already nonzero, nothing to propagate. */
105 if (oldmask)
106 continue;
107
108 /* Propagate the new CPU up the tree. */
109 mask = rnp->grpmask;
110 rnp_up = rnp->parent;
111 done = false;
112 while (rnp_up) {
113 raw_spin_lock_irqsave_rcu_node(rnp_up, flags);
114 if (rnp_up->expmaskinit)
115 done = true;
116 rnp_up->expmaskinit |= mask;
117 raw_spin_unlock_irqrestore_rcu_node(rnp_up, flags);
118 if (done)
119 break;
120 mask = rnp_up->grpmask;
121 rnp_up = rnp_up->parent;
122 }
123 }
124}
125
126/*
127 * Reset the ->expmask values in the rcu_node tree in preparation for
128 * a new expedited grace period.
129 */
130static void __maybe_unused sync_exp_reset_tree(void)
131{
132 unsigned long flags;
133 struct rcu_node *rnp;
134
135 sync_exp_reset_tree_hotplug();
136 rcu_for_each_node_breadth_first(rnp) {
137 raw_spin_lock_irqsave_rcu_node(rnp, flags);
138 WARN_ON_ONCE(rnp->expmask);
139 WRITE_ONCE(rnp->expmask, rnp->expmaskinit);
140 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
141 }
142}
143
144/*
145 * Return non-zero if there is no RCU expedited grace period in progress
146 * for the specified rcu_node structure, in other words, if all CPUs and
147 * tasks covered by the specified rcu_node structure have done their bit
148 * for the current expedited grace period.
149 */
150static bool sync_rcu_exp_done(struct rcu_node *rnp)
151{
152 raw_lockdep_assert_held_rcu_node(rnp);
153 return READ_ONCE(rnp->exp_tasks) == NULL &&
154 READ_ONCE(rnp->expmask) == 0;
155}
156
157/*
158 * Like sync_rcu_exp_done(), but where the caller does not hold the
159 * rcu_node's ->lock.
160 */
161static bool sync_rcu_exp_done_unlocked(struct rcu_node *rnp)
162{
163 unsigned long flags;
164 bool ret;
165
166 raw_spin_lock_irqsave_rcu_node(rnp, flags);
167 ret = sync_rcu_exp_done(rnp);
168 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
169
170 return ret;
171}
172
173
174/*
175 * Report the exit from RCU read-side critical section for the last task
176 * that queued itself during or before the current expedited preemptible-RCU
177 * grace period. This event is reported either to the rcu_node structure on
178 * which the task was queued or to one of that rcu_node structure's ancestors,
179 * recursively up the tree. (Calm down, calm down, we do the recursion
180 * iteratively!)
181 */
182static void __rcu_report_exp_rnp(struct rcu_node *rnp,
183 bool wake, unsigned long flags)
184 __releases(rnp->lock)
185{
186 unsigned long mask;
187
188 raw_lockdep_assert_held_rcu_node(rnp);
189 for (;;) {
190 if (!sync_rcu_exp_done(rnp)) {
191 if (!rnp->expmask)
192 rcu_initiate_boost(rnp, flags);
193 else
194 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
195 break;
196 }
197 if (rnp->parent == NULL) {
198 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
199 if (wake) {
200 smp_mb(); /* EGP done before wake_up(). */
201 swake_up_one(&rcu_state.expedited_wq);
202 }
203 break;
204 }
205 mask = rnp->grpmask;
206 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled */
207 rnp = rnp->parent;
208 raw_spin_lock_rcu_node(rnp); /* irqs already disabled */
209 WARN_ON_ONCE(!(rnp->expmask & mask));
210 WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask);
211 }
212}
213
214/*
215 * Report expedited quiescent state for specified node. This is a
216 * lock-acquisition wrapper function for __rcu_report_exp_rnp().
217 */
218static void __maybe_unused rcu_report_exp_rnp(struct rcu_node *rnp, bool wake)
219{
220 unsigned long flags;
221
222 raw_spin_lock_irqsave_rcu_node(rnp, flags);
223 __rcu_report_exp_rnp(rnp, wake, flags);
224}
225
226/*
227 * Report expedited quiescent state for multiple CPUs, all covered by the
228 * specified leaf rcu_node structure.
229 */
230static void rcu_report_exp_cpu_mult(struct rcu_node *rnp,
231 unsigned long mask, bool wake)
232{
233 int cpu;
234 unsigned long flags;
235 struct rcu_data *rdp;
236
237 raw_spin_lock_irqsave_rcu_node(rnp, flags);
238 if (!(rnp->expmask & mask)) {
239 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
240 return;
241 }
242 WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask);
243 for_each_leaf_node_cpu_mask(rnp, cpu, mask) {
244 rdp = per_cpu_ptr(&rcu_data, cpu);
245 if (!IS_ENABLED(CONFIG_NO_HZ_FULL) || !rdp->rcu_forced_tick_exp)
246 continue;
247 rdp->rcu_forced_tick_exp = false;
248 tick_dep_clear_cpu(cpu, TICK_DEP_BIT_RCU_EXP);
249 }
250 __rcu_report_exp_rnp(rnp, wake, flags); /* Releases rnp->lock. */
251}
252
253/*
254 * Report expedited quiescent state for specified rcu_data (CPU).
255 */
256static void rcu_report_exp_rdp(struct rcu_data *rdp)
257{
258 WRITE_ONCE(rdp->exp_deferred_qs, false);
259 rcu_report_exp_cpu_mult(rdp->mynode, rdp->grpmask, true);
260}
261
262/* Common code for work-done checking. */
263static bool sync_exp_work_done(unsigned long s)
264{
265 if (rcu_exp_gp_seq_done(s)) {
266 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("done"));
267 smp_mb(); /* Ensure test happens before caller kfree(). */
268 return true;
269 }
270 return false;
271}
272
273/*
274 * Funnel-lock acquisition for expedited grace periods. Returns true
275 * if some other task completed an expedited grace period that this task
276 * can piggy-back on, and with no mutex held. Otherwise, returns false
277 * with the mutex held, indicating that the caller must actually do the
278 * expedited grace period.
279 */
280static bool exp_funnel_lock(unsigned long s)
281{
282 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, raw_smp_processor_id());
283 struct rcu_node *rnp = rdp->mynode;
284 struct rcu_node *rnp_root = rcu_get_root();
285
286 /* Low-contention fastpath. */
287 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s) &&
288 (rnp == rnp_root ||
289 ULONG_CMP_LT(READ_ONCE(rnp_root->exp_seq_rq), s)) &&
290 mutex_trylock(&rcu_state.exp_mutex))
291 goto fastpath;
292
293 /*
294 * Each pass through the following loop works its way up
295 * the rcu_node tree, returning if others have done the work or
296 * otherwise falls through to acquire ->exp_mutex. The mapping
297 * from CPU to rcu_node structure can be inexact, as it is just
298 * promoting locality and is not strictly needed for correctness.
299 */
300 for (; rnp != NULL; rnp = rnp->parent) {
301 if (sync_exp_work_done(s))
302 return true;
303
304 /* Work not done, either wait here or go up. */
305 spin_lock(&rnp->exp_lock);
306 if (ULONG_CMP_GE(rnp->exp_seq_rq, s)) {
307
308 /* Someone else doing GP, so wait for them. */
309 spin_unlock(&rnp->exp_lock);
310 trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
311 rnp->grplo, rnp->grphi,
312 TPS("wait"));
313 wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
314 sync_exp_work_done(s));
315 return true;
316 }
317 WRITE_ONCE(rnp->exp_seq_rq, s); /* Followers can wait on us. */
318 spin_unlock(&rnp->exp_lock);
319 trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
320 rnp->grplo, rnp->grphi, TPS("nxtlvl"));
321 }
322 mutex_lock(&rcu_state.exp_mutex);
323fastpath:
324 if (sync_exp_work_done(s)) {
325 mutex_unlock(&rcu_state.exp_mutex);
326 return true;
327 }
328 rcu_exp_gp_seq_start();
329 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("start"));
330 return false;
331}
332
333/*
334 * Select the CPUs within the specified rcu_node that the upcoming
335 * expedited grace period needs to wait for.
336 */
337static void sync_rcu_exp_select_node_cpus(struct work_struct *wp)
338{
339 int cpu;
340 unsigned long flags;
341 unsigned long mask_ofl_test;
342 unsigned long mask_ofl_ipi;
343 int ret;
344 struct rcu_exp_work *rewp =
345 container_of(wp, struct rcu_exp_work, rew_work);
346 struct rcu_node *rnp = container_of(rewp, struct rcu_node, rew);
347
348 raw_spin_lock_irqsave_rcu_node(rnp, flags);
349
350 /* Each pass checks a CPU for identity, offline, and idle. */
351 mask_ofl_test = 0;
352 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) {
353 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
354 unsigned long mask = rdp->grpmask;
355 int snap;
356
357 if (raw_smp_processor_id() == cpu ||
358 !(rnp->qsmaskinitnext & mask)) {
359 mask_ofl_test |= mask;
360 } else {
361 snap = rcu_dynticks_snap(rdp);
362 if (rcu_dynticks_in_eqs(snap))
363 mask_ofl_test |= mask;
364 else
365 rdp->exp_dynticks_snap = snap;
366 }
367 }
368 mask_ofl_ipi = rnp->expmask & ~mask_ofl_test;
369
370 /*
371 * Need to wait for any blocked tasks as well. Note that
372 * additional blocking tasks will also block the expedited GP
373 * until such time as the ->expmask bits are cleared.
374 */
375 if (rcu_preempt_has_tasks(rnp))
376 WRITE_ONCE(rnp->exp_tasks, rnp->blkd_tasks.next);
377 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
378
379 /* IPI the remaining CPUs for expedited quiescent state. */
380 for_each_leaf_node_cpu_mask(rnp, cpu, mask_ofl_ipi) {
381 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
382 unsigned long mask = rdp->grpmask;
383
384retry_ipi:
385 if (rcu_dynticks_in_eqs_since(rdp, rdp->exp_dynticks_snap)) {
386 mask_ofl_test |= mask;
387 continue;
388 }
389 if (get_cpu() == cpu) {
390 put_cpu();
391 continue;
392 }
393 ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
394 put_cpu();
395 /* The CPU will report the QS in response to the IPI. */
396 if (!ret)
397 continue;
398
399 /* Failed, raced with CPU hotplug operation. */
400 raw_spin_lock_irqsave_rcu_node(rnp, flags);
401 if ((rnp->qsmaskinitnext & mask) &&
402 (rnp->expmask & mask)) {
403 /* Online, so delay for a bit and try again. */
404 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
405 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("selectofl"));
406 schedule_timeout_idle(1);
407 goto retry_ipi;
408 }
409 /* CPU really is offline, so we must report its QS. */
410 if (rnp->expmask & mask)
411 mask_ofl_test |= mask;
412 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
413 }
414 /* Report quiescent states for those that went offline. */
415 if (mask_ofl_test)
416 rcu_report_exp_cpu_mult(rnp, mask_ofl_test, false);
417}
418
419/*
420 * Select the nodes that the upcoming expedited grace period needs
421 * to wait for.
422 */
423static void sync_rcu_exp_select_cpus(void)
424{
425 int cpu;
426 struct rcu_node *rnp;
427
428 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("reset"));
429 sync_exp_reset_tree();
430 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("select"));
431
432 /* Schedule work for each leaf rcu_node structure. */
433 rcu_for_each_leaf_node(rnp) {
434 rnp->exp_need_flush = false;
435 if (!READ_ONCE(rnp->expmask))
436 continue; /* Avoid early boot non-existent wq. */
437 if (!READ_ONCE(rcu_par_gp_wq) ||
438 rcu_scheduler_active != RCU_SCHEDULER_RUNNING ||
439 rcu_is_last_leaf_node(rnp)) {
440 /* No workqueues yet or last leaf, do direct call. */
441 sync_rcu_exp_select_node_cpus(&rnp->rew.rew_work);
442 continue;
443 }
444 INIT_WORK(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus);
445 cpu = find_next_bit(&rnp->ffmask, BITS_PER_LONG, -1);
446 /* If all offline, queue the work on an unbound CPU. */
447 if (unlikely(cpu > rnp->grphi - rnp->grplo))
448 cpu = WORK_CPU_UNBOUND;
449 else
450 cpu += rnp->grplo;
451 queue_work_on(cpu, rcu_par_gp_wq, &rnp->rew.rew_work);
452 rnp->exp_need_flush = true;
453 }
454
455 /* Wait for workqueue jobs (if any) to complete. */
456 rcu_for_each_leaf_node(rnp)
457 if (rnp->exp_need_flush)
458 flush_work(&rnp->rew.rew_work);
459}
460
461/*
462 * Wait for the expedited grace period to elapse, within time limit.
463 * If the time limit is exceeded without the grace period elapsing,
464 * return false, otherwise return true.
465 */
466static bool synchronize_rcu_expedited_wait_once(long tlimit)
467{
468 int t;
469 struct rcu_node *rnp_root = rcu_get_root();
470
471 t = swait_event_timeout_exclusive(rcu_state.expedited_wq,
472 sync_rcu_exp_done_unlocked(rnp_root),
473 tlimit);
474 // Workqueues should not be signaled.
475 if (t > 0 || sync_rcu_exp_done_unlocked(rnp_root))
476 return true;
477 WARN_ON(t < 0); /* workqueues should not be signaled. */
478 return false;
479}
480
481/*
482 * Wait for the expedited grace period to elapse, issuing any needed
483 * RCU CPU stall warnings along the way.
484 */
485static void synchronize_rcu_expedited_wait(void)
486{
487 int cpu;
488 unsigned long j;
489 unsigned long jiffies_stall;
490 unsigned long jiffies_start;
491 unsigned long mask;
492 int ndetected;
493 struct rcu_data *rdp;
494 struct rcu_node *rnp;
495 struct rcu_node *rnp_root = rcu_get_root();
496
497 trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("startwait"));
498 jiffies_stall = rcu_jiffies_till_stall_check();
499 jiffies_start = jiffies;
500 if (tick_nohz_full_enabled() && rcu_inkernel_boot_has_ended()) {
501 if (synchronize_rcu_expedited_wait_once(1))
502 return;
503 rcu_for_each_leaf_node(rnp) {
504 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) {
505 rdp = per_cpu_ptr(&rcu_data, cpu);
506 if (rdp->rcu_forced_tick_exp)
507 continue;
508 rdp->rcu_forced_tick_exp = true;
509 tick_dep_set_cpu(cpu, TICK_DEP_BIT_RCU_EXP);
510 }
511 }
512 j = READ_ONCE(jiffies_till_first_fqs);
513 if (synchronize_rcu_expedited_wait_once(j + HZ))
514 return;
515 WARN_ON_ONCE(IS_ENABLED(CONFIG_PREEMPT_RT));
516 }
517
518 for (;;) {
519 if (synchronize_rcu_expedited_wait_once(jiffies_stall))
520 return;
521 if (rcu_stall_is_suppressed())
522 continue;
523 panic_on_rcu_stall();
524 pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
525 rcu_state.name);
526 ndetected = 0;
527 rcu_for_each_leaf_node(rnp) {
528 ndetected += rcu_print_task_exp_stall(rnp);
529 for_each_leaf_node_possible_cpu(rnp, cpu) {
530 struct rcu_data *rdp;
531
532 mask = leaf_node_cpu_bit(rnp, cpu);
533 if (!(READ_ONCE(rnp->expmask) & mask))
534 continue;
535 ndetected++;
536 rdp = per_cpu_ptr(&rcu_data, cpu);
537 pr_cont(" %d-%c%c%c", cpu,
538 "O."[!!cpu_online(cpu)],
539 "o."[!!(rdp->grpmask & rnp->expmaskinit)],
540 "N."[!!(rdp->grpmask & rnp->expmaskinitnext)]);
541 }
542 }
543 pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n",
544 jiffies - jiffies_start, rcu_state.expedited_sequence,
545 data_race(rnp_root->expmask),
546 ".T"[!!data_race(rnp_root->exp_tasks)]);
547 if (ndetected) {
548 pr_err("blocking rcu_node structures:");
549 rcu_for_each_node_breadth_first(rnp) {
550 if (rnp == rnp_root)
551 continue; /* printed unconditionally */
552 if (sync_rcu_exp_done_unlocked(rnp))
553 continue;
554 pr_cont(" l=%u:%d-%d:%#lx/%c",
555 rnp->level, rnp->grplo, rnp->grphi,
556 data_race(rnp->expmask),
557 ".T"[!!data_race(rnp->exp_tasks)]);
558 }
559 pr_cont("\n");
560 }
561 rcu_for_each_leaf_node(rnp) {
562 for_each_leaf_node_possible_cpu(rnp, cpu) {
563 mask = leaf_node_cpu_bit(rnp, cpu);
564 if (!(READ_ONCE(rnp->expmask) & mask))
565 continue;
566 dump_cpu_task(cpu);
567 }
568 }
569 jiffies_stall = 3 * rcu_jiffies_till_stall_check() + 3;
570 }
571}
572
573/*
574 * Wait for the current expedited grace period to complete, and then
575 * wake up everyone who piggybacked on the just-completed expedited
576 * grace period. Also update all the ->exp_seq_rq counters as needed
577 * in order to avoid counter-wrap problems.
578 */
579static void rcu_exp_wait_wake(unsigned long s)
580{
581 struct rcu_node *rnp;
582
583 synchronize_rcu_expedited_wait();
584
585 // Switch over to wakeup mode, allowing the next GP to proceed.
586 // End the previous grace period only after acquiring the mutex
587 // to ensure that only one GP runs concurrently with wakeups.
588 mutex_lock(&rcu_state.exp_wake_mutex);
589 rcu_exp_gp_seq_end();
590 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("end"));
591
592 rcu_for_each_node_breadth_first(rnp) {
593 if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s)) {
594 spin_lock(&rnp->exp_lock);
595 /* Recheck, avoid hang in case someone just arrived. */
596 if (ULONG_CMP_LT(rnp->exp_seq_rq, s))
597 WRITE_ONCE(rnp->exp_seq_rq, s);
598 spin_unlock(&rnp->exp_lock);
599 }
600 smp_mb(); /* All above changes before wakeup. */
601 wake_up_all(&rnp->exp_wq[rcu_seq_ctr(s) & 0x3]);
602 }
603 trace_rcu_exp_grace_period(rcu_state.name, s, TPS("endwake"));
604 mutex_unlock(&rcu_state.exp_wake_mutex);
605}
606
607/*
608 * Common code to drive an expedited grace period forward, used by
609 * workqueues and mid-boot-time tasks.
610 */
611static void rcu_exp_sel_wait_wake(unsigned long s)
612{
613 /* Initialize the rcu_node tree in preparation for the wait. */
614 sync_rcu_exp_select_cpus();
615
616 /* Wait and clean up, including waking everyone. */
617 rcu_exp_wait_wake(s);
618}
619
620/*
621 * Work-queue handler to drive an expedited grace period forward.
622 */
623static void wait_rcu_exp_gp(struct work_struct *wp)
624{
625 struct rcu_exp_work *rewp;
626
627 rewp = container_of(wp, struct rcu_exp_work, rew_work);
628 rcu_exp_sel_wait_wake(rewp->rew_s);
629}
630
631#ifdef CONFIG_PREEMPT_RCU
632
633/*
634 * Remote handler for smp_call_function_single(). If there is an
635 * RCU read-side critical section in effect, request that the
636 * next rcu_read_unlock() record the quiescent state up the
637 * ->expmask fields in the rcu_node tree. Otherwise, immediately
638 * report the quiescent state.
639 */
640static void rcu_exp_handler(void *unused)
641{
642 int depth = rcu_preempt_depth();
643 unsigned long flags;
644 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
645 struct rcu_node *rnp = rdp->mynode;
646 struct task_struct *t = current;
647
648 /*
649 * First, the common case of not being in an RCU read-side
650 * critical section. If also enabled or idle, immediately
651 * report the quiescent state, otherwise defer.
652 */
653 if (!depth) {
654 if (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK)) ||
655 rcu_dynticks_curr_cpu_in_eqs()) {
656 rcu_report_exp_rdp(rdp);
657 } else {
658 rdp->exp_deferred_qs = true;
659 set_tsk_need_resched(t);
660 set_preempt_need_resched();
661 }
662 return;
663 }
664
665 /*
666 * Second, the less-common case of being in an RCU read-side
667 * critical section. In this case we can count on a future
668 * rcu_read_unlock(). However, this rcu_read_unlock() might
669 * execute on some other CPU, but in that case there will be
670 * a future context switch. Either way, if the expedited
671 * grace period is still waiting on this CPU, set ->deferred_qs
672 * so that the eventual quiescent state will be reported.
673 * Note that there is a large group of race conditions that
674 * can have caused this quiescent state to already have been
675 * reported, so we really do need to check ->expmask.
676 */
677 if (depth > 0) {
678 raw_spin_lock_irqsave_rcu_node(rnp, flags);
679 if (rnp->expmask & rdp->grpmask) {
680 rdp->exp_deferred_qs = true;
681 t->rcu_read_unlock_special.b.exp_hint = true;
682 }
683 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
684 return;
685 }
686
687 // Finally, negative nesting depth should not happen.
688 WARN_ON_ONCE(1);
689}
690
691/* PREEMPTION=y, so no PREEMPTION=n expedited grace period to clean up after. */
692static void sync_sched_exp_online_cleanup(int cpu)
693{
694}
695
696/*
697 * Scan the current list of tasks blocked within RCU read-side critical
698 * sections, printing out the tid of each that is blocking the current
699 * expedited grace period.
700 */
701static int rcu_print_task_exp_stall(struct rcu_node *rnp)
702{
703 unsigned long flags;
704 int ndetected = 0;
705 struct task_struct *t;
706
707 if (!READ_ONCE(rnp->exp_tasks))
708 return 0;
709 raw_spin_lock_irqsave_rcu_node(rnp, flags);
710 t = list_entry(rnp->exp_tasks->prev,
711 struct task_struct, rcu_node_entry);
712 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
713 pr_cont(" P%d", t->pid);
714 ndetected++;
715 }
716 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
717 return ndetected;
718}
719
720#else /* #ifdef CONFIG_PREEMPT_RCU */
721
722/* Request an expedited quiescent state. */
723static void rcu_exp_need_qs(void)
724{
725 __this_cpu_write(rcu_data.cpu_no_qs.b.exp, true);
726 /* Store .exp before .rcu_urgent_qs. */
727 smp_store_release(this_cpu_ptr(&rcu_data.rcu_urgent_qs), true);
728 set_tsk_need_resched(current);
729 set_preempt_need_resched();
730}
731
732/* Invoked on each online non-idle CPU for expedited quiescent state. */
733static void rcu_exp_handler(void *unused)
734{
735 struct rcu_data *rdp;
736 struct rcu_node *rnp;
737
738 rdp = this_cpu_ptr(&rcu_data);
739 rnp = rdp->mynode;
740 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
741 __this_cpu_read(rcu_data.cpu_no_qs.b.exp))
742 return;
743 if (rcu_is_cpu_rrupt_from_idle()) {
744 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
745 return;
746 }
747 rcu_exp_need_qs();
748}
749
750/* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
751static void sync_sched_exp_online_cleanup(int cpu)
752{
753 unsigned long flags;
754 int my_cpu;
755 struct rcu_data *rdp;
756 int ret;
757 struct rcu_node *rnp;
758
759 rdp = per_cpu_ptr(&rcu_data, cpu);
760 rnp = rdp->mynode;
761 my_cpu = get_cpu();
762 /* Quiescent state either not needed or already requested, leave. */
763 if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
764 __this_cpu_read(rcu_data.cpu_no_qs.b.exp)) {
765 put_cpu();
766 return;
767 }
768 /* Quiescent state needed on current CPU, so set it up locally. */
769 if (my_cpu == cpu) {
770 local_irq_save(flags);
771 rcu_exp_need_qs();
772 local_irq_restore(flags);
773 put_cpu();
774 return;
775 }
776 /* Quiescent state needed on some other CPU, send IPI. */
777 ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
778 put_cpu();
779 WARN_ON_ONCE(ret);
780}
781
782/*
783 * Because preemptible RCU does not exist, we never have to check for
784 * tasks blocked within RCU read-side critical sections that are
785 * blocking the current expedited grace period.
786 */
787static int rcu_print_task_exp_stall(struct rcu_node *rnp)
788{
789 return 0;
790}
791
792#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
793
794/**
795 * synchronize_rcu_expedited - Brute-force RCU grace period
796 *
797 * Wait for an RCU grace period, but expedite it. The basic idea is to
798 * IPI all non-idle non-nohz online CPUs. The IPI handler checks whether
799 * the CPU is in an RCU critical section, and if so, it sets a flag that
800 * causes the outermost rcu_read_unlock() to report the quiescent state
801 * for RCU-preempt or asks the scheduler for help for RCU-sched. On the
802 * other hand, if the CPU is not in an RCU read-side critical section,
803 * the IPI handler reports the quiescent state immediately.
804 *
805 * Although this is a great improvement over previous expedited
806 * implementations, it is still unfriendly to real-time workloads, so is
807 * thus not recommended for any sort of common-case code. In fact, if
808 * you are using synchronize_rcu_expedited() in a loop, please restructure
809 * your code to batch your updates, and then use a single synchronize_rcu()
810 * instead.
811 *
812 * This has the same semantics as (but is more brutal than) synchronize_rcu().
813 */
814void synchronize_rcu_expedited(void)
815{
816 bool boottime = (rcu_scheduler_active == RCU_SCHEDULER_INIT);
817 struct rcu_exp_work rew;
818 struct rcu_node *rnp;
819 unsigned long s;
820
821 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
822 lock_is_held(&rcu_lock_map) ||
823 lock_is_held(&rcu_sched_lock_map),
824 "Illegal synchronize_rcu_expedited() in RCU read-side critical section");
825
826 /* Is the state is such that the call is a grace period? */
827 if (rcu_blocking_is_gp())
828 return;
829
830 /* If expedited grace periods are prohibited, fall back to normal. */
831 if (rcu_gp_is_normal()) {
832 wait_rcu_gp(call_rcu);
833 return;
834 }
835
836 /* Take a snapshot of the sequence number. */
837 s = rcu_exp_gp_seq_snap();
838 if (exp_funnel_lock(s))
839 return; /* Someone else did our work for us. */
840
841 /* Ensure that load happens before action based on it. */
842 if (unlikely(boottime)) {
843 /* Direct call during scheduler init and early_initcalls(). */
844 rcu_exp_sel_wait_wake(s);
845 } else {
846 /* Marshall arguments & schedule the expedited grace period. */
847 rew.rew_s = s;
848 INIT_WORK_ONSTACK(&rew.rew_work, wait_rcu_exp_gp);
849 queue_work(rcu_gp_wq, &rew.rew_work);
850 }
851
852 /* Wait for expedited grace period to complete. */
853 rnp = rcu_get_root();
854 wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
855 sync_exp_work_done(s));
856 smp_mb(); /* Workqueue actions happen before return. */
857
858 /* Let the next expedited grace period start. */
859 mutex_unlock(&rcu_state.exp_mutex);
860
861 if (likely(!boottime))
862 destroy_work_on_stack(&rew.rew_work);
863}
864EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);