Line data Source code
1 : /* SPDX-License-Identifier: GPL-2.0+ */
2 : /*
3 : * RCU expedited grace periods
4 : *
5 : * Copyright IBM Corporation, 2016
6 : *
7 : * Authors: Paul E. McKenney <paulmck@linux.ibm.com>
8 : */
9 :
10 : #include <linux/lockdep.h>
11 :
12 : static void rcu_exp_handler(void *unused);
13 : static int rcu_print_task_exp_stall(struct rcu_node *rnp);
14 :
15 : /*
16 : * Record the start of an expedited grace period.
17 : */
18 161 : static void rcu_exp_gp_seq_start(void)
19 : {
20 161 : 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 : */
27 483 : static __maybe_unused unsigned long rcu_exp_gp_seq_endval(void)
28 : {
29 483 : return rcu_seq_endval(&rcu_state.expedited_sequence);
30 : }
31 :
32 : /*
33 : * Record the end of an expedited grace period.
34 : */
35 161 : static void rcu_exp_gp_seq_end(void)
36 : {
37 161 : rcu_seq_end(&rcu_state.expedited_sequence);
38 161 : 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 : */
46 161 : static unsigned long rcu_exp_gp_seq_snap(void)
47 : {
48 161 : unsigned long s;
49 :
50 161 : smp_mb(); /* Caller's modifications seen first by other CPUs. */
51 161 : s = rcu_seq_snap(&rcu_state.expedited_sequence);
52 161 : trace_rcu_exp_grace_period(rcu_state.name, s, TPS("snap"));
53 161 : 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 : */
61 640 : static bool rcu_exp_gp_seq_done(unsigned long s)
62 : {
63 640 : 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 : */
73 161 : static void sync_exp_reset_tree_hotplug(void)
74 : {
75 161 : bool done;
76 161 : unsigned long flags;
77 161 : unsigned long mask;
78 161 : unsigned long oldmask;
79 161 : int ncpus = smp_load_acquire(&rcu_state.ncpus); /* Order vs. locking. */
80 161 : struct rcu_node *rnp;
81 161 : struct rcu_node *rnp_up;
82 :
83 : /* If no new CPUs onlined since last time, nothing to do. */
84 161 : if (likely(ncpus == rcu_state.ncpus_snap))
85 : return;
86 1 : 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 2 : rcu_for_each_leaf_node(rnp) {
93 1 : raw_spin_lock_irqsave_rcu_node(rnp, flags);
94 1 : if (rnp->expmaskinit == rnp->expmaskinitnext) {
95 0 : raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
96 0 : continue; /* No new CPUs, nothing to do. */
97 : }
98 :
99 : /* Update this node's mask, track old value for propagation. */
100 1 : oldmask = rnp->expmaskinit;
101 1 : rnp->expmaskinit = rnp->expmaskinitnext;
102 2 : raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
103 :
104 : /* If was already nonzero, nothing to propagate. */
105 1 : if (oldmask)
106 0 : continue;
107 :
108 : /* Propagate the new CPU up the tree. */
109 1 : mask = rnp->grpmask;
110 1 : rnp_up = rnp->parent;
111 1 : done = false;
112 1 : while (rnp_up) {
113 0 : raw_spin_lock_irqsave_rcu_node(rnp_up, flags);
114 0 : if (rnp_up->expmaskinit)
115 0 : done = true;
116 0 : rnp_up->expmaskinit |= mask;
117 0 : raw_spin_unlock_irqrestore_rcu_node(rnp_up, flags);
118 0 : if (done)
119 : break;
120 0 : mask = rnp_up->grpmask;
121 0 : 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 : */
130 161 : static void __maybe_unused sync_exp_reset_tree(void)
131 : {
132 161 : unsigned long flags;
133 161 : struct rcu_node *rnp;
134 :
135 161 : sync_exp_reset_tree_hotplug();
136 483 : rcu_for_each_node_breadth_first(rnp) {
137 161 : raw_spin_lock_irqsave_rcu_node(rnp, flags);
138 161 : WARN_ON_ONCE(rnp->expmask);
139 161 : WRITE_ONCE(rnp->expmask, rnp->expmaskinit);
140 322 : raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
141 : }
142 161 : }
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 : */
150 925 : static bool sync_rcu_exp_done(struct rcu_node *rnp)
151 : {
152 1850 : raw_lockdep_assert_held_rcu_node(rnp);
153 925 : return READ_ONCE(rnp->exp_tasks) == NULL &&
154 925 : 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 : */
161 461 : static bool sync_rcu_exp_done_unlocked(struct rcu_node *rnp)
162 : {
163 461 : unsigned long flags;
164 461 : bool ret;
165 :
166 461 : raw_spin_lock_irqsave_rcu_node(rnp, flags);
167 461 : ret = sync_rcu_exp_done(rnp);
168 922 : raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
169 :
170 461 : 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 : */
182 464 : static void __rcu_report_exp_rnp(struct rcu_node *rnp,
183 : bool wake, unsigned long flags)
184 : __releases(rnp->lock)
185 : {
186 464 : unsigned long mask;
187 :
188 928 : raw_lockdep_assert_held_rcu_node(rnp);
189 464 : for (;;) {
190 464 : if (!sync_rcu_exp_done(rnp)) {
191 303 : if (!rnp->expmask)
192 0 : rcu_initiate_boost(rnp, flags);
193 : else
194 606 : raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
195 : break;
196 : }
197 161 : if (rnp->parent == NULL) {
198 322 : raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
199 161 : if (wake) {
200 151 : smp_mb(); /* EGP done before wake_up(). */
201 151 : swake_up_one(&rcu_state.expedited_wq);
202 : }
203 : break;
204 : }
205 0 : mask = rnp->grpmask;
206 0 : raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled */
207 0 : rnp = rnp->parent;
208 0 : raw_spin_lock_rcu_node(rnp); /* irqs already disabled */
209 0 : WARN_ON_ONCE(!(rnp->expmask & mask));
210 0 : WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask);
211 : }
212 464 : }
213 :
214 : /*
215 : * Report expedited quiescent state for specified node. This is a
216 : * lock-acquisition wrapper function for __rcu_report_exp_rnp().
217 : */
218 : static 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 : */
230 464 : static void rcu_report_exp_cpu_mult(struct rcu_node *rnp,
231 : unsigned long mask, bool wake)
232 : {
233 464 : int cpu;
234 464 : unsigned long flags;
235 464 : struct rcu_data *rdp;
236 :
237 464 : raw_spin_lock_irqsave_rcu_node(rnp, flags);
238 464 : if (!(rnp->expmask & mask)) {
239 0 : raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
240 0 : return;
241 : }
242 464 : WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask);
243 1108 : for_each_leaf_node_cpu_mask(rnp, cpu, mask) {
244 644 : rdp = per_cpu_ptr(&rcu_data, cpu);
245 644 : if (!IS_ENABLED(CONFIG_NO_HZ_FULL) || !rdp->rcu_forced_tick_exp)
246 644 : continue;
247 : rdp->rcu_forced_tick_exp = false;
248 : tick_dep_clear_cpu(cpu, TICK_DEP_BIT_RCU_EXP);
249 : }
250 464 : __rcu_report_exp_rnp(rnp, wake, flags); /* Releases rnp->lock. */
251 : }
252 :
253 : /*
254 : * Report expedited quiescent state for specified rcu_data (CPU).
255 : */
256 303 : static void rcu_report_exp_rdp(struct rcu_data *rdp)
257 : {
258 303 : WRITE_ONCE(rdp->exp_deferred_qs, false);
259 303 : rcu_report_exp_cpu_mult(rdp->mynode, rdp->grpmask, true);
260 303 : }
261 :
262 : /* Common code for work-done checking. */
263 640 : static bool sync_exp_work_done(unsigned long s)
264 : {
265 640 : if (rcu_exp_gp_seq_done(s)) {
266 161 : trace_rcu_exp_grace_period(rcu_state.name, s, TPS("done"));
267 159 : smp_mb(); /* Ensure test happens before caller kfree(). */
268 161 : 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 : */
280 161 : static bool exp_funnel_lock(unsigned long s)
281 : {
282 161 : struct rcu_data *rdp = per_cpu_ptr(&rcu_data, raw_smp_processor_id());
283 161 : struct rcu_node *rnp = rdp->mynode;
284 161 : struct rcu_node *rnp_root = rcu_get_root();
285 :
286 : /* Low-contention fastpath. */
287 161 : if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s) &&
288 0 : (rnp == rnp_root ||
289 161 : ULONG_CMP_LT(READ_ONCE(rnp_root->exp_seq_rq), s)) &&
290 161 : mutex_trylock(&rcu_state.exp_mutex))
291 161 : 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 0 : for (; rnp != NULL; rnp = rnp->parent) {
301 0 : if (sync_exp_work_done(s))
302 0 : return true;
303 :
304 : /* Work not done, either wait here or go up. */
305 0 : spin_lock(&rnp->exp_lock);
306 0 : if (ULONG_CMP_GE(rnp->exp_seq_rq, s)) {
307 :
308 : /* Someone else doing GP, so wait for them. */
309 0 : spin_unlock(&rnp->exp_lock);
310 0 : trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
311 : rnp->grplo, rnp->grphi,
312 0 : TPS("wait"));
313 0 : wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
314 : sync_exp_work_done(s));
315 0 : return true;
316 : }
317 0 : WRITE_ONCE(rnp->exp_seq_rq, s); /* Followers can wait on us. */
318 0 : spin_unlock(&rnp->exp_lock);
319 0 : trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level,
320 0 : rnp->grplo, rnp->grphi, TPS("nxtlvl"));
321 : }
322 0 : mutex_lock(&rcu_state.exp_mutex);
323 161 : fastpath:
324 161 : if (sync_exp_work_done(s)) {
325 0 : mutex_unlock(&rcu_state.exp_mutex);
326 0 : return true;
327 : }
328 161 : rcu_exp_gp_seq_start();
329 161 : trace_rcu_exp_grace_period(rcu_state.name, s, TPS("start"));
330 161 : 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 : */
337 161 : static void sync_rcu_exp_select_node_cpus(struct work_struct *wp)
338 : {
339 161 : int cpu;
340 161 : unsigned long flags;
341 161 : unsigned long mask_ofl_test;
342 161 : unsigned long mask_ofl_ipi;
343 161 : int ret;
344 161 : struct rcu_exp_work *rewp =
345 161 : container_of(wp, struct rcu_exp_work, rew_work);
346 161 : struct rcu_node *rnp = container_of(rewp, struct rcu_node, rew);
347 :
348 161 : raw_spin_lock_irqsave_rcu_node(rnp, flags);
349 :
350 : /* Each pass checks a CPU for identity, offline, and idle. */
351 161 : mask_ofl_test = 0;
352 805 : for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) {
353 644 : struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
354 644 : unsigned long mask = rdp->grpmask;
355 644 : int snap;
356 :
357 644 : if (raw_smp_processor_id() == cpu ||
358 483 : !(rnp->qsmaskinitnext & mask)) {
359 161 : mask_ofl_test |= mask;
360 : } else {
361 483 : snap = rcu_dynticks_snap(rdp);
362 483 : if (rcu_dynticks_in_eqs(snap))
363 180 : mask_ofl_test |= mask;
364 : else
365 303 : rdp->exp_dynticks_snap = snap;
366 : }
367 : }
368 161 : 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 : */
375 161 : if (rcu_preempt_has_tasks(rnp))
376 161 : WRITE_ONCE(rnp->exp_tasks, rnp->blkd_tasks.next);
377 322 : raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
378 :
379 : /* IPI the remaining CPUs for expedited quiescent state. */
380 464 : for_each_leaf_node_cpu_mask(rnp, cpu, mask_ofl_ipi) {
381 303 : struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
382 303 : unsigned long mask = rdp->grpmask;
383 :
384 303 : retry_ipi:
385 303 : if (rcu_dynticks_in_eqs_since(rdp, rdp->exp_dynticks_snap)) {
386 0 : mask_ofl_test |= mask;
387 0 : continue;
388 : }
389 303 : if (get_cpu() == cpu) {
390 0 : put_cpu();
391 0 : continue;
392 : }
393 303 : ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
394 303 : put_cpu();
395 : /* The CPU will report the QS in response to the IPI. */
396 303 : if (!ret)
397 303 : continue;
398 :
399 : /* Failed, raced with CPU hotplug operation. */
400 0 : raw_spin_lock_irqsave_rcu_node(rnp, flags);
401 0 : if ((rnp->qsmaskinitnext & mask) &&
402 0 : (rnp->expmask & mask)) {
403 : /* Online, so delay for a bit and try again. */
404 0 : raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
405 0 : trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("selectofl"));
406 0 : schedule_timeout_idle(1);
407 0 : goto retry_ipi;
408 : }
409 : /* CPU really is offline, so we must report its QS. */
410 0 : if (rnp->expmask & mask)
411 0 : mask_ofl_test |= mask;
412 303 : raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
413 : }
414 : /* Report quiescent states for those that went offline. */
415 161 : if (mask_ofl_test)
416 161 : rcu_report_exp_cpu_mult(rnp, mask_ofl_test, false);
417 161 : }
418 :
419 : /*
420 : * Select the nodes that the upcoming expedited grace period needs
421 : * to wait for.
422 : */
423 161 : static void sync_rcu_exp_select_cpus(void)
424 : {
425 161 : int cpu;
426 161 : struct rcu_node *rnp;
427 :
428 161 : trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("reset"));
429 161 : sync_exp_reset_tree();
430 161 : 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. */
433 322 : rcu_for_each_leaf_node(rnp) {
434 161 : rnp->exp_need_flush = false;
435 161 : if (!READ_ONCE(rnp->expmask))
436 0 : continue; /* Avoid early boot non-existent wq. */
437 161 : if (!READ_ONCE(rcu_par_gp_wq) ||
438 161 : rcu_scheduler_active != RCU_SCHEDULER_RUNNING ||
439 159 : rcu_is_last_leaf_node(rnp)) {
440 : /* No workqueues yet or last leaf, do direct call. */
441 161 : sync_rcu_exp_select_node_cpus(&rnp->rew.rew_work);
442 161 : continue;
443 : }
444 0 : INIT_WORK(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus);
445 0 : cpu = find_next_bit(&rnp->ffmask, BITS_PER_LONG, -1);
446 : /* If all offline, queue the work on an unbound CPU. */
447 0 : if (unlikely(cpu > rnp->grphi - rnp->grplo))
448 : cpu = WORK_CPU_UNBOUND;
449 : else
450 0 : cpu += rnp->grplo;
451 0 : queue_work_on(cpu, rcu_par_gp_wq, &rnp->rew.rew_work);
452 0 : rnp->exp_need_flush = true;
453 : }
454 :
455 : /* Wait for workqueue jobs (if any) to complete. */
456 322 : rcu_for_each_leaf_node(rnp)
457 161 : if (rnp->exp_need_flush)
458 0 : flush_work(&rnp->rew.rew_work);
459 161 : }
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 : */
466 161 : static bool synchronize_rcu_expedited_wait_once(long tlimit)
467 : {
468 161 : int t;
469 161 : struct rcu_node *rnp_root = rcu_get_root();
470 :
471 310 : 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.
475 161 : if (t > 0 || sync_rcu_exp_done_unlocked(rnp_root))
476 161 : return true;
477 0 : 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 : */
485 161 : static void synchronize_rcu_expedited_wait(void)
486 : {
487 161 : int cpu;
488 161 : unsigned long j;
489 161 : unsigned long jiffies_stall;
490 161 : unsigned long jiffies_start;
491 161 : unsigned long mask;
492 161 : int ndetected;
493 161 : struct rcu_data *rdp;
494 161 : struct rcu_node *rnp;
495 161 : struct rcu_node *rnp_root = rcu_get_root();
496 :
497 161 : trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("startwait"));
498 161 : jiffies_stall = rcu_jiffies_till_stall_check();
499 161 : jiffies_start = jiffies;
500 161 : 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 :
518 161 : for (;;) {
519 161 : if (synchronize_rcu_expedited_wait_once(jiffies_stall))
520 161 : return;
521 0 : if (rcu_stall_is_suppressed())
522 0 : continue;
523 0 : panic_on_rcu_stall();
524 0 : pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
525 : rcu_state.name);
526 0 : ndetected = 0;
527 0 : rcu_for_each_leaf_node(rnp) {
528 0 : ndetected += rcu_print_task_exp_stall(rnp);
529 0 : for_each_leaf_node_possible_cpu(rnp, cpu) {
530 0 : struct rcu_data *rdp;
531 :
532 0 : mask = leaf_node_cpu_bit(rnp, cpu);
533 0 : if (!(READ_ONCE(rnp->expmask) & mask))
534 0 : continue;
535 0 : ndetected++;
536 0 : rdp = per_cpu_ptr(&rcu_data, cpu);
537 0 : 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 0 : 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 0 : if (ndetected) {
548 0 : pr_err("blocking rcu_node structures (internal RCU debug):");
549 0 : rcu_for_each_node_breadth_first(rnp) {
550 0 : if (rnp == rnp_root)
551 0 : continue; /* printed unconditionally */
552 0 : if (sync_rcu_exp_done_unlocked(rnp))
553 0 : continue;
554 0 : 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 0 : pr_cont("\n");
560 : }
561 0 : rcu_for_each_leaf_node(rnp) {
562 0 : for_each_leaf_node_possible_cpu(rnp, cpu) {
563 0 : mask = leaf_node_cpu_bit(rnp, cpu);
564 0 : if (!(READ_ONCE(rnp->expmask) & mask))
565 0 : continue;
566 0 : dump_cpu_task(cpu);
567 : }
568 : }
569 0 : 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 : */
579 161 : static void rcu_exp_wait_wake(unsigned long s)
580 : {
581 161 : struct rcu_node *rnp;
582 :
583 161 : 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.
588 161 : mutex_lock(&rcu_state.exp_wake_mutex);
589 161 : rcu_exp_gp_seq_end();
590 161 : trace_rcu_exp_grace_period(rcu_state.name, s, TPS("end"));
591 :
592 483 : rcu_for_each_node_breadth_first(rnp) {
593 161 : if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s)) {
594 161 : spin_lock(&rnp->exp_lock);
595 : /* Recheck, avoid hang in case someone just arrived. */
596 161 : if (ULONG_CMP_LT(rnp->exp_seq_rq, s))
597 161 : WRITE_ONCE(rnp->exp_seq_rq, s);
598 161 : spin_unlock(&rnp->exp_lock);
599 : }
600 161 : smp_mb(); /* All above changes before wakeup. */
601 161 : wake_up_all(&rnp->exp_wq[rcu_seq_ctr(s) & 0x3]);
602 : }
603 161 : trace_rcu_exp_grace_period(rcu_state.name, s, TPS("endwake"));
604 161 : mutex_unlock(&rcu_state.exp_wake_mutex);
605 161 : }
606 :
607 : /*
608 : * Common code to drive an expedited grace period forward, used by
609 : * workqueues and mid-boot-time tasks.
610 : */
611 161 : static void rcu_exp_sel_wait_wake(unsigned long s)
612 : {
613 : /* Initialize the rcu_node tree in preparation for the wait. */
614 161 : sync_rcu_exp_select_cpus();
615 :
616 : /* Wait and clean up, including waking everyone. */
617 161 : rcu_exp_wait_wake(s);
618 2 : }
619 :
620 : /*
621 : * Work-queue handler to drive an expedited grace period forward.
622 : */
623 159 : static void wait_rcu_exp_gp(struct work_struct *wp)
624 : {
625 159 : struct rcu_exp_work *rewp;
626 :
627 159 : rewp = container_of(wp, struct rcu_exp_work, rew_work);
628 159 : rcu_exp_sel_wait_wake(rewp->rew_s);
629 159 : }
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 : */
640 : static 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. */
692 : static 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 : */
701 : static 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. */
723 303 : static void rcu_exp_need_qs(void)
724 : {
725 303 : __this_cpu_write(rcu_data.cpu_no_qs.b.exp, true);
726 : /* Store .exp before .rcu_urgent_qs. */
727 303 : smp_store_release(this_cpu_ptr(&rcu_data.rcu_urgent_qs), true);
728 303 : set_tsk_need_resched(current);
729 303 : set_preempt_need_resched();
730 303 : }
731 :
732 : /* Invoked on each online non-idle CPU for expedited quiescent state. */
733 303 : static void rcu_exp_handler(void *unused)
734 : {
735 303 : struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
736 303 : struct rcu_node *rnp = rdp->mynode;
737 :
738 303 : if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
739 303 : __this_cpu_read(rcu_data.cpu_no_qs.b.exp))
740 : return;
741 303 : if (rcu_is_cpu_rrupt_from_idle()) {
742 0 : rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
743 0 : return;
744 : }
745 303 : rcu_exp_need_qs();
746 : }
747 :
748 : /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
749 3 : static void sync_sched_exp_online_cleanup(int cpu)
750 : {
751 3 : unsigned long flags;
752 3 : int my_cpu;
753 3 : struct rcu_data *rdp;
754 3 : int ret;
755 3 : struct rcu_node *rnp;
756 :
757 3 : rdp = per_cpu_ptr(&rcu_data, cpu);
758 3 : rnp = rdp->mynode;
759 3 : my_cpu = get_cpu();
760 : /* Quiescent state either not needed or already requested, leave. */
761 3 : if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
762 0 : __this_cpu_read(rcu_data.cpu_no_qs.b.exp)) {
763 3 : put_cpu();
764 3 : return;
765 : }
766 : /* Quiescent state needed on current CPU, so set it up locally. */
767 0 : if (my_cpu == cpu) {
768 0 : local_irq_save(flags);
769 0 : rcu_exp_need_qs();
770 0 : local_irq_restore(flags);
771 0 : put_cpu();
772 0 : return;
773 : }
774 : /* Quiescent state needed on some other CPU, send IPI. */
775 0 : ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
776 0 : put_cpu();
777 0 : WARN_ON_ONCE(ret);
778 : }
779 :
780 : /*
781 : * Because preemptible RCU does not exist, we never have to check for
782 : * tasks blocked within RCU read-side critical sections that are
783 : * blocking the current expedited grace period.
784 : */
785 0 : static int rcu_print_task_exp_stall(struct rcu_node *rnp)
786 : {
787 0 : return 0;
788 : }
789 :
790 : #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
791 :
792 : /**
793 : * synchronize_rcu_expedited - Brute-force RCU grace period
794 : *
795 : * Wait for an RCU grace period, but expedite it. The basic idea is to
796 : * IPI all non-idle non-nohz online CPUs. The IPI handler checks whether
797 : * the CPU is in an RCU critical section, and if so, it sets a flag that
798 : * causes the outermost rcu_read_unlock() to report the quiescent state
799 : * for RCU-preempt or asks the scheduler for help for RCU-sched. On the
800 : * other hand, if the CPU is not in an RCU read-side critical section,
801 : * the IPI handler reports the quiescent state immediately.
802 : *
803 : * Although this is a great improvement over previous expedited
804 : * implementations, it is still unfriendly to real-time workloads, so is
805 : * thus not recommended for any sort of common-case code. In fact, if
806 : * you are using synchronize_rcu_expedited() in a loop, please restructure
807 : * your code to batch your updates, and then use a single synchronize_rcu()
808 : * instead.
809 : *
810 : * This has the same semantics as (but is more brutal than) synchronize_rcu().
811 : */
812 164 : void synchronize_rcu_expedited(void)
813 : {
814 164 : bool boottime = (rcu_scheduler_active == RCU_SCHEDULER_INIT);
815 164 : struct rcu_exp_work rew;
816 164 : struct rcu_node *rnp;
817 164 : unsigned long s;
818 :
819 650 : RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
820 : lock_is_held(&rcu_lock_map) ||
821 : lock_is_held(&rcu_sched_lock_map),
822 : "Illegal synchronize_rcu_expedited() in RCU read-side critical section");
823 :
824 : /* Is the state is such that the call is a grace period? */
825 164 : if (rcu_blocking_is_gp())
826 3 : return;
827 :
828 : /* If expedited grace periods are prohibited, fall back to normal. */
829 161 : if (rcu_gp_is_normal()) {
830 0 : wait_rcu_gp(call_rcu);
831 0 : return;
832 : }
833 :
834 : /* Take a snapshot of the sequence number. */
835 161 : s = rcu_exp_gp_seq_snap();
836 161 : if (exp_funnel_lock(s))
837 : return; /* Someone else did our work for us. */
838 :
839 : /* Ensure that load happens before action based on it. */
840 161 : if (unlikely(boottime)) {
841 : /* Direct call during scheduler init and early_initcalls(). */
842 2 : rcu_exp_sel_wait_wake(s);
843 : } else {
844 : /* Marshall arguments & schedule the expedited grace period. */
845 159 : rew.rew_s = s;
846 159 : INIT_WORK_ONSTACK(&rew.rew_work, wait_rcu_exp_gp);
847 159 : queue_work(rcu_gp_wq, &rew.rew_work);
848 : }
849 :
850 : /* Wait for expedited grace period to complete. */
851 161 : rnp = rcu_get_root();
852 479 : wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
853 : sync_exp_work_done(s));
854 161 : smp_mb(); /* Workqueue actions happen before return. */
855 :
856 : /* Let the next expedited grace period start. */
857 161 : mutex_unlock(&rcu_state.exp_mutex);
858 :
859 161 : if (likely(!boottime))
860 159 : destroy_work_on_stack(&rew.rew_work);
861 : }
862 : EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
|
