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  */
rcu_exp_gp_seq_start(void)18 static 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  */
rcu_exp_gp_seq_endval(void)27 static __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  */
rcu_exp_gp_seq_end(void)35 static 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  */
rcu_exp_gp_seq_snap(void)46 static 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  */
rcu_exp_gp_seq_done(unsigned long s)61 static 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  */
sync_exp_reset_tree_hotplug(void)73 static 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  */
sync_exp_reset_tree(void)130 static 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  */
sync_rcu_exp_done(struct rcu_node * rnp)150 static 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  */
sync_rcu_exp_done_unlocked(struct rcu_node * rnp)161 static 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  */
__rcu_report_exp_rnp(struct rcu_node * rnp,bool wake,unsigned long flags)182 static 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  */
rcu_report_exp_rnp(struct rcu_node * rnp,bool wake)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  */
rcu_report_exp_cpu_mult(struct rcu_node * rnp,unsigned long mask,bool wake)230 static 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  */
rcu_report_exp_rdp(struct rcu_data * rdp)256 static void rcu_report_exp_rdp(struct rcu_data *rdp)
257 {
258 	WRITE_ONCE(rdp->cpu_no_qs.b.exp, false);
259 	rcu_report_exp_cpu_mult(rdp->mynode, rdp->grpmask, true);
260 }
261 
262 /* Common code for work-done checking. */
sync_exp_work_done(unsigned long s)263 static 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  */
exp_funnel_lock(unsigned long s)280 static 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);
323 fastpath:
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  */
__sync_rcu_exp_select_node_cpus(struct rcu_exp_work * rewp)337 static void __sync_rcu_exp_select_node_cpus(struct rcu_exp_work *rewp)
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_node *rnp = container_of(rewp, struct rcu_node, rew);
345 
346 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
347 
348 	/* Each pass checks a CPU for identity, offline, and idle. */
349 	mask_ofl_test = 0;
350 	for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) {
351 		struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
352 		unsigned long mask = rdp->grpmask;
353 		int snap;
354 
355 		if (raw_smp_processor_id() == cpu ||
356 		    !(rnp->qsmaskinitnext & mask)) {
357 			mask_ofl_test |= mask;
358 		} else {
359 			snap = rcu_dynticks_snap(rdp);
360 			if (rcu_dynticks_in_eqs(snap))
361 				mask_ofl_test |= mask;
362 			else
363 				rdp->exp_dynticks_snap = snap;
364 		}
365 	}
366 	mask_ofl_ipi = rnp->expmask & ~mask_ofl_test;
367 
368 	/*
369 	 * Need to wait for any blocked tasks as well.	Note that
370 	 * additional blocking tasks will also block the expedited GP
371 	 * until such time as the ->expmask bits are cleared.
372 	 */
373 	if (rcu_preempt_has_tasks(rnp))
374 		WRITE_ONCE(rnp->exp_tasks, rnp->blkd_tasks.next);
375 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
376 
377 	/* IPI the remaining CPUs for expedited quiescent state. */
378 	for_each_leaf_node_cpu_mask(rnp, cpu, mask_ofl_ipi) {
379 		struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
380 		unsigned long mask = rdp->grpmask;
381 
382 retry_ipi:
383 		if (rcu_dynticks_in_eqs_since(rdp, rdp->exp_dynticks_snap)) {
384 			mask_ofl_test |= mask;
385 			continue;
386 		}
387 		if (get_cpu() == cpu) {
388 			mask_ofl_test |= mask;
389 			put_cpu();
390 			continue;
391 		}
392 		ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
393 		put_cpu();
394 		/* The CPU will report the QS in response to the IPI. */
395 		if (!ret)
396 			continue;
397 
398 		/* Failed, raced with CPU hotplug operation. */
399 		raw_spin_lock_irqsave_rcu_node(rnp, flags);
400 		if ((rnp->qsmaskinitnext & mask) &&
401 		    (rnp->expmask & mask)) {
402 			/* Online, so delay for a bit and try again. */
403 			raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
404 			trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("selectofl"));
405 			schedule_timeout_idle(1);
406 			goto retry_ipi;
407 		}
408 		/* CPU really is offline, so we must report its QS. */
409 		if (rnp->expmask & mask)
410 			mask_ofl_test |= mask;
411 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
412 	}
413 	/* Report quiescent states for those that went offline. */
414 	if (mask_ofl_test)
415 		rcu_report_exp_cpu_mult(rnp, mask_ofl_test, false);
416 }
417 
418 static void rcu_exp_sel_wait_wake(unsigned long s);
419 
420 #ifdef CONFIG_RCU_EXP_KTHREAD
sync_rcu_exp_select_node_cpus(struct kthread_work * wp)421 static 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 
rcu_gp_par_worker_started(void)429 static inline bool rcu_gp_par_worker_started(void)
430 {
431 	return !!READ_ONCE(rcu_exp_par_gp_kworker);
432 }
433 
sync_rcu_exp_select_cpus_queue_work(struct rcu_node * rnp)434 static inline void sync_rcu_exp_select_cpus_queue_work(struct rcu_node *rnp)
435 {
436 	kthread_init_work(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus);
437 	/*
438 	 * Use rcu_exp_par_gp_kworker, because flushing a work item from
439 	 * another work item on the same kthread worker can result in
440 	 * deadlock.
441 	 */
442 	kthread_queue_work(rcu_exp_par_gp_kworker, &rnp->rew.rew_work);
443 }
444 
sync_rcu_exp_select_cpus_flush_work(struct rcu_node * rnp)445 static inline void sync_rcu_exp_select_cpus_flush_work(struct rcu_node *rnp)
446 {
447 	kthread_flush_work(&rnp->rew.rew_work);
448 }
449 
450 /*
451  * Work-queue handler to drive an expedited grace period forward.
452  */
wait_rcu_exp_gp(struct kthread_work * wp)453 static void wait_rcu_exp_gp(struct kthread_work *wp)
454 {
455 	struct rcu_exp_work *rewp;
456 
457 	rewp = container_of(wp, struct rcu_exp_work, rew_work);
458 	rcu_exp_sel_wait_wake(rewp->rew_s);
459 }
460 
synchronize_rcu_expedited_queue_work(struct rcu_exp_work * rew)461 static inline void synchronize_rcu_expedited_queue_work(struct rcu_exp_work *rew)
462 {
463 	kthread_init_work(&rew->rew_work, wait_rcu_exp_gp);
464 	kthread_queue_work(rcu_exp_gp_kworker, &rew->rew_work);
465 }
466 
synchronize_rcu_expedited_destroy_work(struct rcu_exp_work * rew)467 static inline void synchronize_rcu_expedited_destroy_work(struct rcu_exp_work *rew)
468 {
469 }
470 #else /* !CONFIG_RCU_EXP_KTHREAD */
sync_rcu_exp_select_node_cpus(struct work_struct * wp)471 static void sync_rcu_exp_select_node_cpus(struct work_struct *wp)
472 {
473 	struct rcu_exp_work *rewp =
474 		container_of(wp, struct rcu_exp_work, rew_work);
475 
476 	__sync_rcu_exp_select_node_cpus(rewp);
477 }
478 
rcu_gp_par_worker_started(void)479 static inline bool rcu_gp_par_worker_started(void)
480 {
481 	return !!READ_ONCE(rcu_par_gp_wq);
482 }
483 
sync_rcu_exp_select_cpus_queue_work(struct rcu_node * rnp)484 static inline void sync_rcu_exp_select_cpus_queue_work(struct rcu_node *rnp)
485 {
486 	int cpu = find_next_bit(&rnp->ffmask, BITS_PER_LONG, -1);
487 
488 	INIT_WORK(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus);
489 	/* If all offline, queue the work on an unbound CPU. */
490 	if (unlikely(cpu > rnp->grphi - rnp->grplo))
491 		cpu = WORK_CPU_UNBOUND;
492 	else
493 		cpu += rnp->grplo;
494 	queue_work_on(cpu, rcu_par_gp_wq, &rnp->rew.rew_work);
495 }
496 
sync_rcu_exp_select_cpus_flush_work(struct rcu_node * rnp)497 static inline void sync_rcu_exp_select_cpus_flush_work(struct rcu_node *rnp)
498 {
499 	flush_work(&rnp->rew.rew_work);
500 }
501 
502 /*
503  * Work-queue handler to drive an expedited grace period forward.
504  */
wait_rcu_exp_gp(struct work_struct * wp)505 static void wait_rcu_exp_gp(struct work_struct *wp)
506 {
507 	struct rcu_exp_work *rewp;
508 
509 	rewp = container_of(wp, struct rcu_exp_work, rew_work);
510 	rcu_exp_sel_wait_wake(rewp->rew_s);
511 }
512 
synchronize_rcu_expedited_queue_work(struct rcu_exp_work * rew)513 static inline void synchronize_rcu_expedited_queue_work(struct rcu_exp_work *rew)
514 {
515 	INIT_WORK_ONSTACK(&rew->rew_work, wait_rcu_exp_gp);
516 	queue_work(rcu_gp_wq, &rew->rew_work);
517 }
518 
synchronize_rcu_expedited_destroy_work(struct rcu_exp_work * rew)519 static inline void synchronize_rcu_expedited_destroy_work(struct rcu_exp_work *rew)
520 {
521 	destroy_work_on_stack(&rew->rew_work);
522 }
523 #endif /* CONFIG_RCU_EXP_KTHREAD */
524 
525 /*
526  * Select the nodes that the upcoming expedited grace period needs
527  * to wait for.
528  */
sync_rcu_exp_select_cpus(void)529 static void sync_rcu_exp_select_cpus(void)
530 {
531 	struct rcu_node *rnp;
532 
533 	trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("reset"));
534 	sync_exp_reset_tree();
535 	trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("select"));
536 
537 	/* Schedule work for each leaf rcu_node structure. */
538 	rcu_for_each_leaf_node(rnp) {
539 		rnp->exp_need_flush = false;
540 		if (!READ_ONCE(rnp->expmask))
541 			continue; /* Avoid early boot non-existent wq. */
542 		if (!rcu_gp_par_worker_started() ||
543 		    rcu_scheduler_active != RCU_SCHEDULER_RUNNING ||
544 		    rcu_is_last_leaf_node(rnp)) {
545 			/* No worker started yet or last leaf, do direct call. */
546 			sync_rcu_exp_select_node_cpus(&rnp->rew.rew_work);
547 			continue;
548 		}
549 		sync_rcu_exp_select_cpus_queue_work(rnp);
550 		rnp->exp_need_flush = true;
551 	}
552 
553 	/* Wait for jobs (if any) to complete. */
554 	rcu_for_each_leaf_node(rnp)
555 		if (rnp->exp_need_flush)
556 			sync_rcu_exp_select_cpus_flush_work(rnp);
557 }
558 
559 /*
560  * Wait for the expedited grace period to elapse, within time limit.
561  * If the time limit is exceeded without the grace period elapsing,
562  * return false, otherwise return true.
563  */
synchronize_rcu_expedited_wait_once(long tlimit)564 static bool synchronize_rcu_expedited_wait_once(long tlimit)
565 {
566 	int t;
567 	struct rcu_node *rnp_root = rcu_get_root();
568 
569 	t = swait_event_timeout_exclusive(rcu_state.expedited_wq,
570 					  sync_rcu_exp_done_unlocked(rnp_root),
571 					  tlimit);
572 	// Workqueues should not be signaled.
573 	if (t > 0 || sync_rcu_exp_done_unlocked(rnp_root))
574 		return true;
575 	WARN_ON(t < 0);  /* workqueues should not be signaled. */
576 	return false;
577 }
578 
579 /*
580  * Wait for the expedited grace period to elapse, issuing any needed
581  * RCU CPU stall warnings along the way.
582  */
synchronize_rcu_expedited_wait(void)583 static void synchronize_rcu_expedited_wait(void)
584 {
585 	int cpu;
586 	unsigned long j;
587 	unsigned long jiffies_stall;
588 	unsigned long jiffies_start;
589 	unsigned long mask;
590 	int ndetected;
591 	struct rcu_data *rdp;
592 	struct rcu_node *rnp;
593 	struct rcu_node *rnp_root = rcu_get_root();
594 
595 	trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("startwait"));
596 	jiffies_stall = rcu_exp_jiffies_till_stall_check();
597 	jiffies_start = jiffies;
598 	if (tick_nohz_full_enabled() && rcu_inkernel_boot_has_ended()) {
599 		if (synchronize_rcu_expedited_wait_once(1))
600 			return;
601 		rcu_for_each_leaf_node(rnp) {
602 			mask = READ_ONCE(rnp->expmask);
603 			for_each_leaf_node_cpu_mask(rnp, cpu, mask) {
604 				rdp = per_cpu_ptr(&rcu_data, cpu);
605 				if (rdp->rcu_forced_tick_exp)
606 					continue;
607 				rdp->rcu_forced_tick_exp = true;
608 				preempt_disable();
609 				if (cpu_online(cpu))
610 					tick_dep_set_cpu(cpu, TICK_DEP_BIT_RCU_EXP);
611 				preempt_enable();
612 			}
613 		}
614 		j = READ_ONCE(jiffies_till_first_fqs);
615 		if (synchronize_rcu_expedited_wait_once(j + HZ))
616 			return;
617 	}
618 
619 	for (;;) {
620 		if (synchronize_rcu_expedited_wait_once(jiffies_stall))
621 			return;
622 		if (rcu_stall_is_suppressed())
623 			continue;
624 		panic_on_rcu_stall();
625 		trace_rcu_stall_warning(rcu_state.name, TPS("ExpeditedStall"));
626 		pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {",
627 		       rcu_state.name);
628 		ndetected = 0;
629 		rcu_for_each_leaf_node(rnp) {
630 			ndetected += rcu_print_task_exp_stall(rnp);
631 			for_each_leaf_node_possible_cpu(rnp, cpu) {
632 				struct rcu_data *rdp;
633 
634 				mask = leaf_node_cpu_bit(rnp, cpu);
635 				if (!(READ_ONCE(rnp->expmask) & mask))
636 					continue;
637 				ndetected++;
638 				rdp = per_cpu_ptr(&rcu_data, cpu);
639 				pr_cont(" %d-%c%c%c", cpu,
640 					"O."[!!cpu_online(cpu)],
641 					"o."[!!(rdp->grpmask & rnp->expmaskinit)],
642 					"N."[!!(rdp->grpmask & rnp->expmaskinitnext)]);
643 			}
644 		}
645 		pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n",
646 			jiffies - jiffies_start, rcu_state.expedited_sequence,
647 			data_race(rnp_root->expmask),
648 			".T"[!!data_race(rnp_root->exp_tasks)]);
649 		if (ndetected) {
650 			pr_err("blocking rcu_node structures (internal RCU debug):");
651 			rcu_for_each_node_breadth_first(rnp) {
652 				if (rnp == rnp_root)
653 					continue; /* printed unconditionally */
654 				if (sync_rcu_exp_done_unlocked(rnp))
655 					continue;
656 				pr_cont(" l=%u:%d-%d:%#lx/%c",
657 					rnp->level, rnp->grplo, rnp->grphi,
658 					data_race(rnp->expmask),
659 					".T"[!!data_race(rnp->exp_tasks)]);
660 			}
661 			pr_cont("\n");
662 		}
663 		rcu_for_each_leaf_node(rnp) {
664 			for_each_leaf_node_possible_cpu(rnp, cpu) {
665 				mask = leaf_node_cpu_bit(rnp, cpu);
666 				if (!(READ_ONCE(rnp->expmask) & mask))
667 					continue;
668 				dump_cpu_task(cpu);
669 			}
670 		}
671 		jiffies_stall = 3 * rcu_exp_jiffies_till_stall_check() + 3;
672 	}
673 }
674 
675 /*
676  * Wait for the current expedited grace period to complete, and then
677  * wake up everyone who piggybacked on the just-completed expedited
678  * grace period.  Also update all the ->exp_seq_rq counters as needed
679  * in order to avoid counter-wrap problems.
680  */
rcu_exp_wait_wake(unsigned long s)681 static void rcu_exp_wait_wake(unsigned long s)
682 {
683 	struct rcu_node *rnp;
684 
685 	synchronize_rcu_expedited_wait();
686 
687 	// Switch over to wakeup mode, allowing the next GP to proceed.
688 	// End the previous grace period only after acquiring the mutex
689 	// to ensure that only one GP runs concurrently with wakeups.
690 	mutex_lock(&rcu_state.exp_wake_mutex);
691 	rcu_exp_gp_seq_end();
692 	trace_rcu_exp_grace_period(rcu_state.name, s, TPS("end"));
693 
694 	rcu_for_each_node_breadth_first(rnp) {
695 		if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s)) {
696 			spin_lock(&rnp->exp_lock);
697 			/* Recheck, avoid hang in case someone just arrived. */
698 			if (ULONG_CMP_LT(rnp->exp_seq_rq, s))
699 				WRITE_ONCE(rnp->exp_seq_rq, s);
700 			spin_unlock(&rnp->exp_lock);
701 		}
702 		smp_mb(); /* All above changes before wakeup. */
703 		wake_up_all(&rnp->exp_wq[rcu_seq_ctr(s) & 0x3]);
704 	}
705 	trace_rcu_exp_grace_period(rcu_state.name, s, TPS("endwake"));
706 	mutex_unlock(&rcu_state.exp_wake_mutex);
707 }
708 
709 /*
710  * Common code to drive an expedited grace period forward, used by
711  * workqueues and mid-boot-time tasks.
712  */
rcu_exp_sel_wait_wake(unsigned long s)713 static void rcu_exp_sel_wait_wake(unsigned long s)
714 {
715 	/* Initialize the rcu_node tree in preparation for the wait. */
716 	sync_rcu_exp_select_cpus();
717 
718 	/* Wait and clean up, including waking everyone. */
719 	rcu_exp_wait_wake(s);
720 }
721 
722 #ifdef CONFIG_PREEMPT_RCU
723 
724 /*
725  * Remote handler for smp_call_function_single().  If there is an
726  * RCU read-side critical section in effect, request that the
727  * next rcu_read_unlock() record the quiescent state up the
728  * ->expmask fields in the rcu_node tree.  Otherwise, immediately
729  * report the quiescent state.
730  */
rcu_exp_handler(void * unused)731 static void rcu_exp_handler(void *unused)
732 {
733 	int depth = rcu_preempt_depth();
734 	unsigned long flags;
735 	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
736 	struct rcu_node *rnp = rdp->mynode;
737 	struct task_struct *t = current;
738 
739 	/*
740 	 * First, the common case of not being in an RCU read-side
741 	 * critical section.  If also enabled or idle, immediately
742 	 * report the quiescent state, otherwise defer.
743 	 */
744 	if (!depth) {
745 		if (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK)) ||
746 		    rcu_is_cpu_rrupt_from_idle()) {
747 			rcu_report_exp_rdp(rdp);
748 		} else {
749 			WRITE_ONCE(rdp->cpu_no_qs.b.exp, true);
750 			set_tsk_need_resched(t);
751 			set_preempt_need_resched();
752 		}
753 		return;
754 	}
755 
756 	/*
757 	 * Second, the less-common case of being in an RCU read-side
758 	 * critical section.  In this case we can count on a future
759 	 * rcu_read_unlock().  However, this rcu_read_unlock() might
760 	 * execute on some other CPU, but in that case there will be
761 	 * a future context switch.  Either way, if the expedited
762 	 * grace period is still waiting on this CPU, set ->deferred_qs
763 	 * so that the eventual quiescent state will be reported.
764 	 * Note that there is a large group of race conditions that
765 	 * can have caused this quiescent state to already have been
766 	 * reported, so we really do need to check ->expmask.
767 	 */
768 	if (depth > 0) {
769 		raw_spin_lock_irqsave_rcu_node(rnp, flags);
770 		if (rnp->expmask & rdp->grpmask) {
771 			WRITE_ONCE(rdp->cpu_no_qs.b.exp, true);
772 			t->rcu_read_unlock_special.b.exp_hint = true;
773 		}
774 		raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
775 		return;
776 	}
777 
778 	// Finally, negative nesting depth should not happen.
779 	WARN_ON_ONCE(1);
780 }
781 
782 /* PREEMPTION=y, so no PREEMPTION=n expedited grace period to clean up after. */
sync_sched_exp_online_cleanup(int cpu)783 static void sync_sched_exp_online_cleanup(int cpu)
784 {
785 }
786 
787 /*
788  * Scan the current list of tasks blocked within RCU read-side critical
789  * sections, printing out the tid of each that is blocking the current
790  * expedited grace period.
791  */
rcu_print_task_exp_stall(struct rcu_node * rnp)792 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
793 {
794 	unsigned long flags;
795 	int ndetected = 0;
796 	struct task_struct *t;
797 
798 	if (!READ_ONCE(rnp->exp_tasks))
799 		return 0;
800 	raw_spin_lock_irqsave_rcu_node(rnp, flags);
801 	t = list_entry(rnp->exp_tasks->prev,
802 		       struct task_struct, rcu_node_entry);
803 	list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
804 		pr_cont(" P%d", t->pid);
805 		ndetected++;
806 	}
807 	raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
808 	return ndetected;
809 }
810 
811 #else /* #ifdef CONFIG_PREEMPT_RCU */
812 
813 /* Request an expedited quiescent state. */
rcu_exp_need_qs(void)814 static void rcu_exp_need_qs(void)
815 {
816 	__this_cpu_write(rcu_data.cpu_no_qs.b.exp, true);
817 	/* Store .exp before .rcu_urgent_qs. */
818 	smp_store_release(this_cpu_ptr(&rcu_data.rcu_urgent_qs), true);
819 	set_tsk_need_resched(current);
820 	set_preempt_need_resched();
821 }
822 
823 /* Invoked on each online non-idle CPU for expedited quiescent state. */
rcu_exp_handler(void * unused)824 static void rcu_exp_handler(void *unused)
825 {
826 	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
827 	struct rcu_node *rnp = rdp->mynode;
828 
829 	if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
830 	    __this_cpu_read(rcu_data.cpu_no_qs.b.exp))
831 		return;
832 	if (rcu_is_cpu_rrupt_from_idle()) {
833 		rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
834 		return;
835 	}
836 	rcu_exp_need_qs();
837 }
838 
839 /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */
sync_sched_exp_online_cleanup(int cpu)840 static void sync_sched_exp_online_cleanup(int cpu)
841 {
842 	unsigned long flags;
843 	int my_cpu;
844 	struct rcu_data *rdp;
845 	int ret;
846 	struct rcu_node *rnp;
847 
848 	rdp = per_cpu_ptr(&rcu_data, cpu);
849 	rnp = rdp->mynode;
850 	my_cpu = get_cpu();
851 	/* Quiescent state either not needed or already requested, leave. */
852 	if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) ||
853 	    READ_ONCE(rdp->cpu_no_qs.b.exp)) {
854 		put_cpu();
855 		return;
856 	}
857 	/* Quiescent state needed on current CPU, so set it up locally. */
858 	if (my_cpu == cpu) {
859 		local_irq_save(flags);
860 		rcu_exp_need_qs();
861 		local_irq_restore(flags);
862 		put_cpu();
863 		return;
864 	}
865 	/* Quiescent state needed on some other CPU, send IPI. */
866 	ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0);
867 	put_cpu();
868 	WARN_ON_ONCE(ret);
869 }
870 
871 /*
872  * Because preemptible RCU does not exist, we never have to check for
873  * tasks blocked within RCU read-side critical sections that are
874  * blocking the current expedited grace period.
875  */
rcu_print_task_exp_stall(struct rcu_node * rnp)876 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
877 {
878 	return 0;
879 }
880 
881 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
882 
883 /**
884  * synchronize_rcu_expedited - Brute-force RCU grace period
885  *
886  * Wait for an RCU grace period, but expedite it.  The basic idea is to
887  * IPI all non-idle non-nohz online CPUs.  The IPI handler checks whether
888  * the CPU is in an RCU critical section, and if so, it sets a flag that
889  * causes the outermost rcu_read_unlock() to report the quiescent state
890  * for RCU-preempt or asks the scheduler for help for RCU-sched.  On the
891  * other hand, if the CPU is not in an RCU read-side critical section,
892  * the IPI handler reports the quiescent state immediately.
893  *
894  * Although this is a great improvement over previous expedited
895  * implementations, it is still unfriendly to real-time workloads, so is
896  * thus not recommended for any sort of common-case code.  In fact, if
897  * you are using synchronize_rcu_expedited() in a loop, please restructure
898  * your code to batch your updates, and then use a single synchronize_rcu()
899  * instead.
900  *
901  * This has the same semantics as (but is more brutal than) synchronize_rcu().
902  */
synchronize_rcu_expedited(void)903 void synchronize_rcu_expedited(void)
904 {
905 	bool boottime = (rcu_scheduler_active == RCU_SCHEDULER_INIT);
906 	struct rcu_exp_work rew;
907 	struct rcu_node *rnp;
908 	unsigned long s;
909 
910 	RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
911 			 lock_is_held(&rcu_lock_map) ||
912 			 lock_is_held(&rcu_sched_lock_map),
913 			 "Illegal synchronize_rcu_expedited() in RCU read-side critical section");
914 
915 	/* Is the state is such that the call is a grace period? */
916 	if (rcu_blocking_is_gp())
917 		return;
918 
919 	/* If expedited grace periods are prohibited, fall back to normal. */
920 	if (rcu_gp_is_normal()) {
921 		wait_rcu_gp(call_rcu);
922 		return;
923 	}
924 
925 	/* Take a snapshot of the sequence number.  */
926 	s = rcu_exp_gp_seq_snap();
927 	if (exp_funnel_lock(s))
928 		return;  /* Someone else did our work for us. */
929 
930 	/* Ensure that load happens before action based on it. */
931 	if (unlikely(boottime)) {
932 		/* Direct call during scheduler init and early_initcalls(). */
933 		rcu_exp_sel_wait_wake(s);
934 	} else {
935 		/* Marshall arguments & schedule the expedited grace period. */
936 		rew.rew_s = s;
937 		synchronize_rcu_expedited_queue_work(&rew);
938 	}
939 
940 	/* Wait for expedited grace period to complete. */
941 	rnp = rcu_get_root();
942 	wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3],
943 		   sync_exp_work_done(s));
944 	smp_mb(); /* Work actions happen before return. */
945 
946 	/* Let the next expedited grace period start. */
947 	mutex_unlock(&rcu_state.exp_mutex);
948 
949 	if (likely(!boottime))
950 		synchronize_rcu_expedited_destroy_work(&rew);
951 }
952 EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
953