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

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