xref: /linux/kernel/rcu/tree_nocb.h (revision 24bce201d79807b668bf9d9e0aca801c5c0d5f78)
1 /* SPDX-License-Identifier: GPL-2.0+ */
2 /*
3  * Read-Copy Update mechanism for mutual exclusion (tree-based version)
4  * Internal non-public definitions that provide either classic
5  * or preemptible semantics.
6  *
7  * Copyright Red Hat, 2009
8  * Copyright IBM Corporation, 2009
9  * Copyright SUSE, 2021
10  *
11  * Author: Ingo Molnar <mingo@elte.hu>
12  *	   Paul E. McKenney <paulmck@linux.ibm.com>
13  *	   Frederic Weisbecker <frederic@kernel.org>
14  */
15 
16 #ifdef CONFIG_RCU_NOCB_CPU
17 static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
18 static bool __read_mostly rcu_nocb_poll;    /* Offload kthread are to poll. */
19 static inline int rcu_lockdep_is_held_nocb(struct rcu_data *rdp)
20 {
21 	return lockdep_is_held(&rdp->nocb_lock);
22 }
23 
24 static inline bool rcu_current_is_nocb_kthread(struct rcu_data *rdp)
25 {
26 	/* Race on early boot between thread creation and assignment */
27 	if (!rdp->nocb_cb_kthread || !rdp->nocb_gp_kthread)
28 		return true;
29 
30 	if (current == rdp->nocb_cb_kthread || current == rdp->nocb_gp_kthread)
31 		if (in_task())
32 			return true;
33 	return false;
34 }
35 
36 /*
37  * Offload callback processing from the boot-time-specified set of CPUs
38  * specified by rcu_nocb_mask.  For the CPUs in the set, there are kthreads
39  * created that pull the callbacks from the corresponding CPU, wait for
40  * a grace period to elapse, and invoke the callbacks.  These kthreads
41  * are organized into GP kthreads, which manage incoming callbacks, wait for
42  * grace periods, and awaken CB kthreads, and the CB kthreads, which only
43  * invoke callbacks.  Each GP kthread invokes its own CBs.  The no-CBs CPUs
44  * do a wake_up() on their GP kthread when they insert a callback into any
45  * empty list, unless the rcu_nocb_poll boot parameter has been specified,
46  * in which case each kthread actively polls its CPU.  (Which isn't so great
47  * for energy efficiency, but which does reduce RCU's overhead on that CPU.)
48  *
49  * This is intended to be used in conjunction with Frederic Weisbecker's
50  * adaptive-idle work, which would seriously reduce OS jitter on CPUs
51  * running CPU-bound user-mode computations.
52  *
53  * Offloading of callbacks can also be used as an energy-efficiency
54  * measure because CPUs with no RCU callbacks queued are more aggressive
55  * about entering dyntick-idle mode.
56  */
57 
58 
59 /*
60  * Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters.
61  * If the list is invalid, a warning is emitted and all CPUs are offloaded.
62  */
63 static int __init rcu_nocb_setup(char *str)
64 {
65 	alloc_bootmem_cpumask_var(&rcu_nocb_mask);
66 	if (*str == '=') {
67 		if (cpulist_parse(++str, rcu_nocb_mask)) {
68 			pr_warn("rcu_nocbs= bad CPU range, all CPUs set\n");
69 			cpumask_setall(rcu_nocb_mask);
70 		}
71 	}
72 	rcu_state.nocb_is_setup = true;
73 	return 1;
74 }
75 __setup("rcu_nocbs", rcu_nocb_setup);
76 
77 static int __init parse_rcu_nocb_poll(char *arg)
78 {
79 	rcu_nocb_poll = true;
80 	return 0;
81 }
82 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
83 
84 /*
85  * Don't bother bypassing ->cblist if the call_rcu() rate is low.
86  * After all, the main point of bypassing is to avoid lock contention
87  * on ->nocb_lock, which only can happen at high call_rcu() rates.
88  */
89 static int nocb_nobypass_lim_per_jiffy = 16 * 1000 / HZ;
90 module_param(nocb_nobypass_lim_per_jiffy, int, 0);
91 
92 /*
93  * Acquire the specified rcu_data structure's ->nocb_bypass_lock.  If the
94  * lock isn't immediately available, increment ->nocb_lock_contended to
95  * flag the contention.
96  */
97 static void rcu_nocb_bypass_lock(struct rcu_data *rdp)
98 	__acquires(&rdp->nocb_bypass_lock)
99 {
100 	lockdep_assert_irqs_disabled();
101 	if (raw_spin_trylock(&rdp->nocb_bypass_lock))
102 		return;
103 	atomic_inc(&rdp->nocb_lock_contended);
104 	WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
105 	smp_mb__after_atomic(); /* atomic_inc() before lock. */
106 	raw_spin_lock(&rdp->nocb_bypass_lock);
107 	smp_mb__before_atomic(); /* atomic_dec() after lock. */
108 	atomic_dec(&rdp->nocb_lock_contended);
109 }
110 
111 /*
112  * Spinwait until the specified rcu_data structure's ->nocb_lock is
113  * not contended.  Please note that this is extremely special-purpose,
114  * relying on the fact that at most two kthreads and one CPU contend for
115  * this lock, and also that the two kthreads are guaranteed to have frequent
116  * grace-period-duration time intervals between successive acquisitions
117  * of the lock.  This allows us to use an extremely simple throttling
118  * mechanism, and further to apply it only to the CPU doing floods of
119  * call_rcu() invocations.  Don't try this at home!
120  */
121 static void rcu_nocb_wait_contended(struct rcu_data *rdp)
122 {
123 	WARN_ON_ONCE(smp_processor_id() != rdp->cpu);
124 	while (WARN_ON_ONCE(atomic_read(&rdp->nocb_lock_contended)))
125 		cpu_relax();
126 }
127 
128 /*
129  * Conditionally acquire the specified rcu_data structure's
130  * ->nocb_bypass_lock.
131  */
132 static bool rcu_nocb_bypass_trylock(struct rcu_data *rdp)
133 {
134 	lockdep_assert_irqs_disabled();
135 	return raw_spin_trylock(&rdp->nocb_bypass_lock);
136 }
137 
138 /*
139  * Release the specified rcu_data structure's ->nocb_bypass_lock.
140  */
141 static void rcu_nocb_bypass_unlock(struct rcu_data *rdp)
142 	__releases(&rdp->nocb_bypass_lock)
143 {
144 	lockdep_assert_irqs_disabled();
145 	raw_spin_unlock(&rdp->nocb_bypass_lock);
146 }
147 
148 /*
149  * Acquire the specified rcu_data structure's ->nocb_lock, but only
150  * if it corresponds to a no-CBs CPU.
151  */
152 static void rcu_nocb_lock(struct rcu_data *rdp)
153 {
154 	lockdep_assert_irqs_disabled();
155 	if (!rcu_rdp_is_offloaded(rdp))
156 		return;
157 	raw_spin_lock(&rdp->nocb_lock);
158 }
159 
160 /*
161  * Release the specified rcu_data structure's ->nocb_lock, but only
162  * if it corresponds to a no-CBs CPU.
163  */
164 static void rcu_nocb_unlock(struct rcu_data *rdp)
165 {
166 	if (rcu_rdp_is_offloaded(rdp)) {
167 		lockdep_assert_irqs_disabled();
168 		raw_spin_unlock(&rdp->nocb_lock);
169 	}
170 }
171 
172 /*
173  * Release the specified rcu_data structure's ->nocb_lock and restore
174  * interrupts, but only if it corresponds to a no-CBs CPU.
175  */
176 static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
177 				       unsigned long flags)
178 {
179 	if (rcu_rdp_is_offloaded(rdp)) {
180 		lockdep_assert_irqs_disabled();
181 		raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
182 	} else {
183 		local_irq_restore(flags);
184 	}
185 }
186 
187 /* Lockdep check that ->cblist may be safely accessed. */
188 static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
189 {
190 	lockdep_assert_irqs_disabled();
191 	if (rcu_rdp_is_offloaded(rdp))
192 		lockdep_assert_held(&rdp->nocb_lock);
193 }
194 
195 /*
196  * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
197  * grace period.
198  */
199 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
200 {
201 	swake_up_all(sq);
202 }
203 
204 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
205 {
206 	return &rnp->nocb_gp_wq[rcu_seq_ctr(rnp->gp_seq) & 0x1];
207 }
208 
209 static void rcu_init_one_nocb(struct rcu_node *rnp)
210 {
211 	init_swait_queue_head(&rnp->nocb_gp_wq[0]);
212 	init_swait_queue_head(&rnp->nocb_gp_wq[1]);
213 }
214 
215 static bool __wake_nocb_gp(struct rcu_data *rdp_gp,
216 			   struct rcu_data *rdp,
217 			   bool force, unsigned long flags)
218 	__releases(rdp_gp->nocb_gp_lock)
219 {
220 	bool needwake = false;
221 
222 	if (!READ_ONCE(rdp_gp->nocb_gp_kthread)) {
223 		raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
224 		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
225 				    TPS("AlreadyAwake"));
226 		return false;
227 	}
228 
229 	if (rdp_gp->nocb_defer_wakeup > RCU_NOCB_WAKE_NOT) {
230 		WRITE_ONCE(rdp_gp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
231 		del_timer(&rdp_gp->nocb_timer);
232 	}
233 
234 	if (force || READ_ONCE(rdp_gp->nocb_gp_sleep)) {
235 		WRITE_ONCE(rdp_gp->nocb_gp_sleep, false);
236 		needwake = true;
237 	}
238 	raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
239 	if (needwake) {
240 		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DoWake"));
241 		wake_up_process(rdp_gp->nocb_gp_kthread);
242 	}
243 
244 	return needwake;
245 }
246 
247 /*
248  * Kick the GP kthread for this NOCB group.
249  */
250 static bool wake_nocb_gp(struct rcu_data *rdp, bool force)
251 {
252 	unsigned long flags;
253 	struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
254 
255 	raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
256 	return __wake_nocb_gp(rdp_gp, rdp, force, flags);
257 }
258 
259 /*
260  * Arrange to wake the GP kthread for this NOCB group at some future
261  * time when it is safe to do so.
262  */
263 static void wake_nocb_gp_defer(struct rcu_data *rdp, int waketype,
264 			       const char *reason)
265 {
266 	unsigned long flags;
267 	struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
268 
269 	raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
270 
271 	/*
272 	 * Bypass wakeup overrides previous deferments. In case
273 	 * of callback storm, no need to wake up too early.
274 	 */
275 	if (waketype == RCU_NOCB_WAKE_BYPASS) {
276 		mod_timer(&rdp_gp->nocb_timer, jiffies + 2);
277 		WRITE_ONCE(rdp_gp->nocb_defer_wakeup, waketype);
278 	} else {
279 		if (rdp_gp->nocb_defer_wakeup < RCU_NOCB_WAKE)
280 			mod_timer(&rdp_gp->nocb_timer, jiffies + 1);
281 		if (rdp_gp->nocb_defer_wakeup < waketype)
282 			WRITE_ONCE(rdp_gp->nocb_defer_wakeup, waketype);
283 	}
284 
285 	raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
286 
287 	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, reason);
288 }
289 
290 /*
291  * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
292  * However, if there is a callback to be enqueued and if ->nocb_bypass
293  * proves to be initially empty, just return false because the no-CB GP
294  * kthread may need to be awakened in this case.
295  *
296  * Note that this function always returns true if rhp is NULL.
297  */
298 static bool rcu_nocb_do_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
299 				     unsigned long j)
300 {
301 	struct rcu_cblist rcl;
302 
303 	WARN_ON_ONCE(!rcu_rdp_is_offloaded(rdp));
304 	rcu_lockdep_assert_cblist_protected(rdp);
305 	lockdep_assert_held(&rdp->nocb_bypass_lock);
306 	if (rhp && !rcu_cblist_n_cbs(&rdp->nocb_bypass)) {
307 		raw_spin_unlock(&rdp->nocb_bypass_lock);
308 		return false;
309 	}
310 	/* Note: ->cblist.len already accounts for ->nocb_bypass contents. */
311 	if (rhp)
312 		rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
313 	rcu_cblist_flush_enqueue(&rcl, &rdp->nocb_bypass, rhp);
314 	rcu_segcblist_insert_pend_cbs(&rdp->cblist, &rcl);
315 	WRITE_ONCE(rdp->nocb_bypass_first, j);
316 	rcu_nocb_bypass_unlock(rdp);
317 	return true;
318 }
319 
320 /*
321  * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
322  * However, if there is a callback to be enqueued and if ->nocb_bypass
323  * proves to be initially empty, just return false because the no-CB GP
324  * kthread may need to be awakened in this case.
325  *
326  * Note that this function always returns true if rhp is NULL.
327  */
328 static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
329 				  unsigned long j)
330 {
331 	if (!rcu_rdp_is_offloaded(rdp))
332 		return true;
333 	rcu_lockdep_assert_cblist_protected(rdp);
334 	rcu_nocb_bypass_lock(rdp);
335 	return rcu_nocb_do_flush_bypass(rdp, rhp, j);
336 }
337 
338 /*
339  * If the ->nocb_bypass_lock is immediately available, flush the
340  * ->nocb_bypass queue into ->cblist.
341  */
342 static void rcu_nocb_try_flush_bypass(struct rcu_data *rdp, unsigned long j)
343 {
344 	rcu_lockdep_assert_cblist_protected(rdp);
345 	if (!rcu_rdp_is_offloaded(rdp) ||
346 	    !rcu_nocb_bypass_trylock(rdp))
347 		return;
348 	WARN_ON_ONCE(!rcu_nocb_do_flush_bypass(rdp, NULL, j));
349 }
350 
351 /*
352  * See whether it is appropriate to use the ->nocb_bypass list in order
353  * to control contention on ->nocb_lock.  A limited number of direct
354  * enqueues are permitted into ->cblist per jiffy.  If ->nocb_bypass
355  * is non-empty, further callbacks must be placed into ->nocb_bypass,
356  * otherwise rcu_barrier() breaks.  Use rcu_nocb_flush_bypass() to switch
357  * back to direct use of ->cblist.  However, ->nocb_bypass should not be
358  * used if ->cblist is empty, because otherwise callbacks can be stranded
359  * on ->nocb_bypass because we cannot count on the current CPU ever again
360  * invoking call_rcu().  The general rule is that if ->nocb_bypass is
361  * non-empty, the corresponding no-CBs grace-period kthread must not be
362  * in an indefinite sleep state.
363  *
364  * Finally, it is not permitted to use the bypass during early boot,
365  * as doing so would confuse the auto-initialization code.  Besides
366  * which, there is no point in worrying about lock contention while
367  * there is only one CPU in operation.
368  */
369 static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
370 				bool *was_alldone, unsigned long flags)
371 {
372 	unsigned long c;
373 	unsigned long cur_gp_seq;
374 	unsigned long j = jiffies;
375 	long ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
376 
377 	lockdep_assert_irqs_disabled();
378 
379 	// Pure softirq/rcuc based processing: no bypassing, no
380 	// locking.
381 	if (!rcu_rdp_is_offloaded(rdp)) {
382 		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
383 		return false;
384 	}
385 
386 	// In the process of (de-)offloading: no bypassing, but
387 	// locking.
388 	if (!rcu_segcblist_completely_offloaded(&rdp->cblist)) {
389 		rcu_nocb_lock(rdp);
390 		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
391 		return false; /* Not offloaded, no bypassing. */
392 	}
393 
394 	// Don't use ->nocb_bypass during early boot.
395 	if (rcu_scheduler_active != RCU_SCHEDULER_RUNNING) {
396 		rcu_nocb_lock(rdp);
397 		WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
398 		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
399 		return false;
400 	}
401 
402 	// If we have advanced to a new jiffy, reset counts to allow
403 	// moving back from ->nocb_bypass to ->cblist.
404 	if (j == rdp->nocb_nobypass_last) {
405 		c = rdp->nocb_nobypass_count + 1;
406 	} else {
407 		WRITE_ONCE(rdp->nocb_nobypass_last, j);
408 		c = rdp->nocb_nobypass_count - nocb_nobypass_lim_per_jiffy;
409 		if (ULONG_CMP_LT(rdp->nocb_nobypass_count,
410 				 nocb_nobypass_lim_per_jiffy))
411 			c = 0;
412 		else if (c > nocb_nobypass_lim_per_jiffy)
413 			c = nocb_nobypass_lim_per_jiffy;
414 	}
415 	WRITE_ONCE(rdp->nocb_nobypass_count, c);
416 
417 	// If there hasn't yet been all that many ->cblist enqueues
418 	// this jiffy, tell the caller to enqueue onto ->cblist.  But flush
419 	// ->nocb_bypass first.
420 	if (rdp->nocb_nobypass_count < nocb_nobypass_lim_per_jiffy) {
421 		rcu_nocb_lock(rdp);
422 		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
423 		if (*was_alldone)
424 			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
425 					    TPS("FirstQ"));
426 		WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, j));
427 		WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
428 		return false; // Caller must enqueue the callback.
429 	}
430 
431 	// If ->nocb_bypass has been used too long or is too full,
432 	// flush ->nocb_bypass to ->cblist.
433 	if ((ncbs && j != READ_ONCE(rdp->nocb_bypass_first)) ||
434 	    ncbs >= qhimark) {
435 		rcu_nocb_lock(rdp);
436 		if (!rcu_nocb_flush_bypass(rdp, rhp, j)) {
437 			*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
438 			if (*was_alldone)
439 				trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
440 						    TPS("FirstQ"));
441 			WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
442 			return false; // Caller must enqueue the callback.
443 		}
444 		if (j != rdp->nocb_gp_adv_time &&
445 		    rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
446 		    rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
447 			rcu_advance_cbs_nowake(rdp->mynode, rdp);
448 			rdp->nocb_gp_adv_time = j;
449 		}
450 		rcu_nocb_unlock_irqrestore(rdp, flags);
451 		return true; // Callback already enqueued.
452 	}
453 
454 	// We need to use the bypass.
455 	rcu_nocb_wait_contended(rdp);
456 	rcu_nocb_bypass_lock(rdp);
457 	ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
458 	rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
459 	rcu_cblist_enqueue(&rdp->nocb_bypass, rhp);
460 	if (!ncbs) {
461 		WRITE_ONCE(rdp->nocb_bypass_first, j);
462 		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("FirstBQ"));
463 	}
464 	rcu_nocb_bypass_unlock(rdp);
465 	smp_mb(); /* Order enqueue before wake. */
466 	if (ncbs) {
467 		local_irq_restore(flags);
468 	} else {
469 		// No-CBs GP kthread might be indefinitely asleep, if so, wake.
470 		rcu_nocb_lock(rdp); // Rare during call_rcu() flood.
471 		if (!rcu_segcblist_pend_cbs(&rdp->cblist)) {
472 			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
473 					    TPS("FirstBQwake"));
474 			__call_rcu_nocb_wake(rdp, true, flags);
475 		} else {
476 			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
477 					    TPS("FirstBQnoWake"));
478 			rcu_nocb_unlock_irqrestore(rdp, flags);
479 		}
480 	}
481 	return true; // Callback already enqueued.
482 }
483 
484 /*
485  * Awaken the no-CBs grace-period kthread if needed, either due to it
486  * legitimately being asleep or due to overload conditions.
487  *
488  * If warranted, also wake up the kthread servicing this CPUs queues.
489  */
490 static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_alldone,
491 				 unsigned long flags)
492 				 __releases(rdp->nocb_lock)
493 {
494 	unsigned long cur_gp_seq;
495 	unsigned long j;
496 	long len;
497 	struct task_struct *t;
498 
499 	// If we are being polled or there is no kthread, just leave.
500 	t = READ_ONCE(rdp->nocb_gp_kthread);
501 	if (rcu_nocb_poll || !t) {
502 		rcu_nocb_unlock_irqrestore(rdp, flags);
503 		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
504 				    TPS("WakeNotPoll"));
505 		return;
506 	}
507 	// Need to actually to a wakeup.
508 	len = rcu_segcblist_n_cbs(&rdp->cblist);
509 	if (was_alldone) {
510 		rdp->qlen_last_fqs_check = len;
511 		if (!irqs_disabled_flags(flags)) {
512 			/* ... if queue was empty ... */
513 			rcu_nocb_unlock_irqrestore(rdp, flags);
514 			wake_nocb_gp(rdp, false);
515 			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
516 					    TPS("WakeEmpty"));
517 		} else {
518 			rcu_nocb_unlock_irqrestore(rdp, flags);
519 			wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE,
520 					   TPS("WakeEmptyIsDeferred"));
521 		}
522 	} else if (len > rdp->qlen_last_fqs_check + qhimark) {
523 		/* ... or if many callbacks queued. */
524 		rdp->qlen_last_fqs_check = len;
525 		j = jiffies;
526 		if (j != rdp->nocb_gp_adv_time &&
527 		    rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
528 		    rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
529 			rcu_advance_cbs_nowake(rdp->mynode, rdp);
530 			rdp->nocb_gp_adv_time = j;
531 		}
532 		smp_mb(); /* Enqueue before timer_pending(). */
533 		if ((rdp->nocb_cb_sleep ||
534 		     !rcu_segcblist_ready_cbs(&rdp->cblist)) &&
535 		    !timer_pending(&rdp->nocb_timer)) {
536 			rcu_nocb_unlock_irqrestore(rdp, flags);
537 			wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE_FORCE,
538 					   TPS("WakeOvfIsDeferred"));
539 		} else {
540 			rcu_nocb_unlock_irqrestore(rdp, flags);
541 			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WakeNot"));
542 		}
543 	} else {
544 		rcu_nocb_unlock_irqrestore(rdp, flags);
545 		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WakeNot"));
546 	}
547 }
548 
549 /*
550  * Check if we ignore this rdp.
551  *
552  * We check that without holding the nocb lock but
553  * we make sure not to miss a freshly offloaded rdp
554  * with the current ordering:
555  *
556  *  rdp_offload_toggle()        nocb_gp_enabled_cb()
557  * -------------------------   ----------------------------
558  *    WRITE flags                 LOCK nocb_gp_lock
559  *    LOCK nocb_gp_lock           READ/WRITE nocb_gp_sleep
560  *    READ/WRITE nocb_gp_sleep    UNLOCK nocb_gp_lock
561  *    UNLOCK nocb_gp_lock         READ flags
562  */
563 static inline bool nocb_gp_enabled_cb(struct rcu_data *rdp)
564 {
565 	u8 flags = SEGCBLIST_OFFLOADED | SEGCBLIST_KTHREAD_GP;
566 
567 	return rcu_segcblist_test_flags(&rdp->cblist, flags);
568 }
569 
570 static inline bool nocb_gp_update_state_deoffloading(struct rcu_data *rdp,
571 						     bool *needwake_state)
572 {
573 	struct rcu_segcblist *cblist = &rdp->cblist;
574 
575 	if (rcu_segcblist_test_flags(cblist, SEGCBLIST_OFFLOADED)) {
576 		if (!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP)) {
577 			rcu_segcblist_set_flags(cblist, SEGCBLIST_KTHREAD_GP);
578 			if (rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB))
579 				*needwake_state = true;
580 		}
581 		return false;
582 	}
583 
584 	/*
585 	 * De-offloading. Clear our flag and notify the de-offload worker.
586 	 * We will ignore this rdp until it ever gets re-offloaded.
587 	 */
588 	WARN_ON_ONCE(!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP));
589 	rcu_segcblist_clear_flags(cblist, SEGCBLIST_KTHREAD_GP);
590 	if (!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB))
591 		*needwake_state = true;
592 	return true;
593 }
594 
595 
596 /*
597  * No-CBs GP kthreads come here to wait for additional callbacks to show up
598  * or for grace periods to end.
599  */
600 static void nocb_gp_wait(struct rcu_data *my_rdp)
601 {
602 	bool bypass = false;
603 	long bypass_ncbs;
604 	int __maybe_unused cpu = my_rdp->cpu;
605 	unsigned long cur_gp_seq;
606 	unsigned long flags;
607 	bool gotcbs = false;
608 	unsigned long j = jiffies;
609 	bool needwait_gp = false; // This prevents actual uninitialized use.
610 	bool needwake;
611 	bool needwake_gp;
612 	struct rcu_data *rdp;
613 	struct rcu_node *rnp;
614 	unsigned long wait_gp_seq = 0; // Suppress "use uninitialized" warning.
615 	bool wasempty = false;
616 
617 	/*
618 	 * Each pass through the following loop checks for CBs and for the
619 	 * nearest grace period (if any) to wait for next.  The CB kthreads
620 	 * and the global grace-period kthread are awakened if needed.
621 	 */
622 	WARN_ON_ONCE(my_rdp->nocb_gp_rdp != my_rdp);
623 	/*
624 	 * An rcu_data structure is removed from the list after its
625 	 * CPU is de-offloaded and added to the list before that CPU is
626 	 * (re-)offloaded.  If the following loop happens to be referencing
627 	 * that rcu_data structure during the time that the corresponding
628 	 * CPU is de-offloaded and then immediately re-offloaded, this
629 	 * loop's rdp pointer will be carried to the end of the list by
630 	 * the resulting pair of list operations.  This can cause the loop
631 	 * to skip over some of the rcu_data structures that were supposed
632 	 * to have been scanned.  Fortunately a new iteration through the
633 	 * entire loop is forced after a given CPU's rcu_data structure
634 	 * is added to the list, so the skipped-over rcu_data structures
635 	 * won't be ignored for long.
636 	 */
637 	list_for_each_entry_rcu(rdp, &my_rdp->nocb_head_rdp, nocb_entry_rdp, 1) {
638 		bool needwake_state = false;
639 
640 		if (!nocb_gp_enabled_cb(rdp))
641 			continue;
642 		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Check"));
643 		rcu_nocb_lock_irqsave(rdp, flags);
644 		if (nocb_gp_update_state_deoffloading(rdp, &needwake_state)) {
645 			rcu_nocb_unlock_irqrestore(rdp, flags);
646 			if (needwake_state)
647 				swake_up_one(&rdp->nocb_state_wq);
648 			continue;
649 		}
650 		bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
651 		if (bypass_ncbs &&
652 		    (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + 1) ||
653 		     bypass_ncbs > 2 * qhimark)) {
654 			// Bypass full or old, so flush it.
655 			(void)rcu_nocb_try_flush_bypass(rdp, j);
656 			bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
657 		} else if (!bypass_ncbs && rcu_segcblist_empty(&rdp->cblist)) {
658 			rcu_nocb_unlock_irqrestore(rdp, flags);
659 			if (needwake_state)
660 				swake_up_one(&rdp->nocb_state_wq);
661 			continue; /* No callbacks here, try next. */
662 		}
663 		if (bypass_ncbs) {
664 			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
665 					    TPS("Bypass"));
666 			bypass = true;
667 		}
668 		rnp = rdp->mynode;
669 
670 		// Advance callbacks if helpful and low contention.
671 		needwake_gp = false;
672 		if (!rcu_segcblist_restempty(&rdp->cblist,
673 					     RCU_NEXT_READY_TAIL) ||
674 		    (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
675 		     rcu_seq_done(&rnp->gp_seq, cur_gp_seq))) {
676 			raw_spin_lock_rcu_node(rnp); /* irqs disabled. */
677 			needwake_gp = rcu_advance_cbs(rnp, rdp);
678 			wasempty = rcu_segcblist_restempty(&rdp->cblist,
679 							   RCU_NEXT_READY_TAIL);
680 			raw_spin_unlock_rcu_node(rnp); /* irqs disabled. */
681 		}
682 		// Need to wait on some grace period?
683 		WARN_ON_ONCE(wasempty &&
684 			     !rcu_segcblist_restempty(&rdp->cblist,
685 						      RCU_NEXT_READY_TAIL));
686 		if (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq)) {
687 			if (!needwait_gp ||
688 			    ULONG_CMP_LT(cur_gp_seq, wait_gp_seq))
689 				wait_gp_seq = cur_gp_seq;
690 			needwait_gp = true;
691 			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
692 					    TPS("NeedWaitGP"));
693 		}
694 		if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
695 			needwake = rdp->nocb_cb_sleep;
696 			WRITE_ONCE(rdp->nocb_cb_sleep, false);
697 			smp_mb(); /* CB invocation -after- GP end. */
698 		} else {
699 			needwake = false;
700 		}
701 		rcu_nocb_unlock_irqrestore(rdp, flags);
702 		if (needwake) {
703 			swake_up_one(&rdp->nocb_cb_wq);
704 			gotcbs = true;
705 		}
706 		if (needwake_gp)
707 			rcu_gp_kthread_wake();
708 		if (needwake_state)
709 			swake_up_one(&rdp->nocb_state_wq);
710 	}
711 
712 	my_rdp->nocb_gp_bypass = bypass;
713 	my_rdp->nocb_gp_gp = needwait_gp;
714 	my_rdp->nocb_gp_seq = needwait_gp ? wait_gp_seq : 0;
715 
716 	if (bypass && !rcu_nocb_poll) {
717 		// At least one child with non-empty ->nocb_bypass, so set
718 		// timer in order to avoid stranding its callbacks.
719 		wake_nocb_gp_defer(my_rdp, RCU_NOCB_WAKE_BYPASS,
720 				   TPS("WakeBypassIsDeferred"));
721 	}
722 	if (rcu_nocb_poll) {
723 		/* Polling, so trace if first poll in the series. */
724 		if (gotcbs)
725 			trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Poll"));
726 		schedule_timeout_idle(1);
727 	} else if (!needwait_gp) {
728 		/* Wait for callbacks to appear. */
729 		trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Sleep"));
730 		swait_event_interruptible_exclusive(my_rdp->nocb_gp_wq,
731 				!READ_ONCE(my_rdp->nocb_gp_sleep));
732 		trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("EndSleep"));
733 	} else {
734 		rnp = my_rdp->mynode;
735 		trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("StartWait"));
736 		swait_event_interruptible_exclusive(
737 			rnp->nocb_gp_wq[rcu_seq_ctr(wait_gp_seq) & 0x1],
738 			rcu_seq_done(&rnp->gp_seq, wait_gp_seq) ||
739 			!READ_ONCE(my_rdp->nocb_gp_sleep));
740 		trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("EndWait"));
741 	}
742 	if (!rcu_nocb_poll) {
743 		raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
744 		if (my_rdp->nocb_defer_wakeup > RCU_NOCB_WAKE_NOT) {
745 			WRITE_ONCE(my_rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
746 			del_timer(&my_rdp->nocb_timer);
747 		}
748 		WRITE_ONCE(my_rdp->nocb_gp_sleep, true);
749 		raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
750 	}
751 	my_rdp->nocb_gp_seq = -1;
752 	WARN_ON(signal_pending(current));
753 }
754 
755 /*
756  * No-CBs grace-period-wait kthread.  There is one of these per group
757  * of CPUs, but only once at least one CPU in that group has come online
758  * at least once since boot.  This kthread checks for newly posted
759  * callbacks from any of the CPUs it is responsible for, waits for a
760  * grace period, then awakens all of the rcu_nocb_cb_kthread() instances
761  * that then have callback-invocation work to do.
762  */
763 static int rcu_nocb_gp_kthread(void *arg)
764 {
765 	struct rcu_data *rdp = arg;
766 
767 	for (;;) {
768 		WRITE_ONCE(rdp->nocb_gp_loops, rdp->nocb_gp_loops + 1);
769 		nocb_gp_wait(rdp);
770 		cond_resched_tasks_rcu_qs();
771 	}
772 	return 0;
773 }
774 
775 static inline bool nocb_cb_can_run(struct rcu_data *rdp)
776 {
777 	u8 flags = SEGCBLIST_OFFLOADED | SEGCBLIST_KTHREAD_CB;
778 
779 	return rcu_segcblist_test_flags(&rdp->cblist, flags);
780 }
781 
782 static inline bool nocb_cb_wait_cond(struct rcu_data *rdp)
783 {
784 	return nocb_cb_can_run(rdp) && !READ_ONCE(rdp->nocb_cb_sleep);
785 }
786 
787 /*
788  * Invoke any ready callbacks from the corresponding no-CBs CPU,
789  * then, if there are no more, wait for more to appear.
790  */
791 static void nocb_cb_wait(struct rcu_data *rdp)
792 {
793 	struct rcu_segcblist *cblist = &rdp->cblist;
794 	unsigned long cur_gp_seq;
795 	unsigned long flags;
796 	bool needwake_state = false;
797 	bool needwake_gp = false;
798 	bool can_sleep = true;
799 	struct rcu_node *rnp = rdp->mynode;
800 
801 	do {
802 		swait_event_interruptible_exclusive(rdp->nocb_cb_wq,
803 						    nocb_cb_wait_cond(rdp));
804 
805 		// VVV Ensure CB invocation follows _sleep test.
806 		if (smp_load_acquire(&rdp->nocb_cb_sleep)) { // ^^^
807 			WARN_ON(signal_pending(current));
808 			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WokeEmpty"));
809 		}
810 	} while (!nocb_cb_can_run(rdp));
811 
812 
813 	local_irq_save(flags);
814 	rcu_momentary_dyntick_idle();
815 	local_irq_restore(flags);
816 	/*
817 	 * Disable BH to provide the expected environment.  Also, when
818 	 * transitioning to/from NOCB mode, a self-requeuing callback might
819 	 * be invoked from softirq.  A short grace period could cause both
820 	 * instances of this callback would execute concurrently.
821 	 */
822 	local_bh_disable();
823 	rcu_do_batch(rdp);
824 	local_bh_enable();
825 	lockdep_assert_irqs_enabled();
826 	rcu_nocb_lock_irqsave(rdp, flags);
827 	if (rcu_segcblist_nextgp(cblist, &cur_gp_seq) &&
828 	    rcu_seq_done(&rnp->gp_seq, cur_gp_seq) &&
829 	    raw_spin_trylock_rcu_node(rnp)) { /* irqs already disabled. */
830 		needwake_gp = rcu_advance_cbs(rdp->mynode, rdp);
831 		raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
832 	}
833 
834 	if (rcu_segcblist_test_flags(cblist, SEGCBLIST_OFFLOADED)) {
835 		if (!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB)) {
836 			rcu_segcblist_set_flags(cblist, SEGCBLIST_KTHREAD_CB);
837 			if (rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP))
838 				needwake_state = true;
839 		}
840 		if (rcu_segcblist_ready_cbs(cblist))
841 			can_sleep = false;
842 	} else {
843 		/*
844 		 * De-offloading. Clear our flag and notify the de-offload worker.
845 		 * We won't touch the callbacks and keep sleeping until we ever
846 		 * get re-offloaded.
847 		 */
848 		WARN_ON_ONCE(!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB));
849 		rcu_segcblist_clear_flags(cblist, SEGCBLIST_KTHREAD_CB);
850 		if (!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP))
851 			needwake_state = true;
852 	}
853 
854 	WRITE_ONCE(rdp->nocb_cb_sleep, can_sleep);
855 
856 	if (rdp->nocb_cb_sleep)
857 		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("CBSleep"));
858 
859 	rcu_nocb_unlock_irqrestore(rdp, flags);
860 	if (needwake_gp)
861 		rcu_gp_kthread_wake();
862 
863 	if (needwake_state)
864 		swake_up_one(&rdp->nocb_state_wq);
865 }
866 
867 /*
868  * Per-rcu_data kthread, but only for no-CBs CPUs.  Repeatedly invoke
869  * nocb_cb_wait() to do the dirty work.
870  */
871 static int rcu_nocb_cb_kthread(void *arg)
872 {
873 	struct rcu_data *rdp = arg;
874 
875 	// Each pass through this loop does one callback batch, and,
876 	// if there are no more ready callbacks, waits for them.
877 	for (;;) {
878 		nocb_cb_wait(rdp);
879 		cond_resched_tasks_rcu_qs();
880 	}
881 	return 0;
882 }
883 
884 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
885 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp, int level)
886 {
887 	return READ_ONCE(rdp->nocb_defer_wakeup) >= level;
888 }
889 
890 /* Do a deferred wakeup of rcu_nocb_kthread(). */
891 static bool do_nocb_deferred_wakeup_common(struct rcu_data *rdp_gp,
892 					   struct rcu_data *rdp, int level,
893 					   unsigned long flags)
894 	__releases(rdp_gp->nocb_gp_lock)
895 {
896 	int ndw;
897 	int ret;
898 
899 	if (!rcu_nocb_need_deferred_wakeup(rdp_gp, level)) {
900 		raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
901 		return false;
902 	}
903 
904 	ndw = rdp_gp->nocb_defer_wakeup;
905 	ret = __wake_nocb_gp(rdp_gp, rdp, ndw == RCU_NOCB_WAKE_FORCE, flags);
906 	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DeferredWake"));
907 
908 	return ret;
909 }
910 
911 /* Do a deferred wakeup of rcu_nocb_kthread() from a timer handler. */
912 static void do_nocb_deferred_wakeup_timer(struct timer_list *t)
913 {
914 	unsigned long flags;
915 	struct rcu_data *rdp = from_timer(rdp, t, nocb_timer);
916 
917 	WARN_ON_ONCE(rdp->nocb_gp_rdp != rdp);
918 	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Timer"));
919 
920 	raw_spin_lock_irqsave(&rdp->nocb_gp_lock, flags);
921 	smp_mb__after_spinlock(); /* Timer expire before wakeup. */
922 	do_nocb_deferred_wakeup_common(rdp, rdp, RCU_NOCB_WAKE_BYPASS, flags);
923 }
924 
925 /*
926  * Do a deferred wakeup of rcu_nocb_kthread() from fastpath.
927  * This means we do an inexact common-case check.  Note that if
928  * we miss, ->nocb_timer will eventually clean things up.
929  */
930 static bool do_nocb_deferred_wakeup(struct rcu_data *rdp)
931 {
932 	unsigned long flags;
933 	struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
934 
935 	if (!rdp_gp || !rcu_nocb_need_deferred_wakeup(rdp_gp, RCU_NOCB_WAKE))
936 		return false;
937 
938 	raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
939 	return do_nocb_deferred_wakeup_common(rdp_gp, rdp, RCU_NOCB_WAKE, flags);
940 }
941 
942 void rcu_nocb_flush_deferred_wakeup(void)
943 {
944 	do_nocb_deferred_wakeup(this_cpu_ptr(&rcu_data));
945 }
946 EXPORT_SYMBOL_GPL(rcu_nocb_flush_deferred_wakeup);
947 
948 static int rdp_offload_toggle(struct rcu_data *rdp,
949 			       bool offload, unsigned long flags)
950 	__releases(rdp->nocb_lock)
951 {
952 	struct rcu_segcblist *cblist = &rdp->cblist;
953 	struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
954 	bool wake_gp = false;
955 
956 	rcu_segcblist_offload(cblist, offload);
957 
958 	if (rdp->nocb_cb_sleep)
959 		rdp->nocb_cb_sleep = false;
960 	rcu_nocb_unlock_irqrestore(rdp, flags);
961 
962 	/*
963 	 * Ignore former value of nocb_cb_sleep and force wake up as it could
964 	 * have been spuriously set to false already.
965 	 */
966 	swake_up_one(&rdp->nocb_cb_wq);
967 
968 	raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
969 	if (rdp_gp->nocb_gp_sleep) {
970 		rdp_gp->nocb_gp_sleep = false;
971 		wake_gp = true;
972 	}
973 	raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
974 
975 	if (wake_gp)
976 		wake_up_process(rdp_gp->nocb_gp_kthread);
977 
978 	return 0;
979 }
980 
981 static long rcu_nocb_rdp_deoffload(void *arg)
982 {
983 	struct rcu_data *rdp = arg;
984 	struct rcu_segcblist *cblist = &rdp->cblist;
985 	unsigned long flags;
986 	int ret;
987 
988 	WARN_ON_ONCE(rdp->cpu != raw_smp_processor_id());
989 
990 	pr_info("De-offloading %d\n", rdp->cpu);
991 
992 	rcu_nocb_lock_irqsave(rdp, flags);
993 	/*
994 	 * Flush once and for all now. This suffices because we are
995 	 * running on the target CPU holding ->nocb_lock (thus having
996 	 * interrupts disabled), and because rdp_offload_toggle()
997 	 * invokes rcu_segcblist_offload(), which clears SEGCBLIST_OFFLOADED.
998 	 * Thus future calls to rcu_segcblist_completely_offloaded() will
999 	 * return false, which means that future calls to rcu_nocb_try_bypass()
1000 	 * will refuse to put anything into the bypass.
1001 	 */
1002 	WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies));
1003 	/*
1004 	 * Start with invoking rcu_core() early. This way if the current thread
1005 	 * happens to preempt an ongoing call to rcu_core() in the middle,
1006 	 * leaving some work dismissed because rcu_core() still thinks the rdp is
1007 	 * completely offloaded, we are guaranteed a nearby future instance of
1008 	 * rcu_core() to catch up.
1009 	 */
1010 	rcu_segcblist_set_flags(cblist, SEGCBLIST_RCU_CORE);
1011 	invoke_rcu_core();
1012 	ret = rdp_offload_toggle(rdp, false, flags);
1013 	swait_event_exclusive(rdp->nocb_state_wq,
1014 			      !rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB |
1015 							SEGCBLIST_KTHREAD_GP));
1016 	/* Stop nocb_gp_wait() from iterating over this structure. */
1017 	list_del_rcu(&rdp->nocb_entry_rdp);
1018 	/*
1019 	 * Lock one last time to acquire latest callback updates from kthreads
1020 	 * so we can later handle callbacks locally without locking.
1021 	 */
1022 	rcu_nocb_lock_irqsave(rdp, flags);
1023 	/*
1024 	 * Theoretically we could clear SEGCBLIST_LOCKING after the nocb
1025 	 * lock is released but how about being paranoid for once?
1026 	 */
1027 	rcu_segcblist_clear_flags(cblist, SEGCBLIST_LOCKING);
1028 	/*
1029 	 * Without SEGCBLIST_LOCKING, we can't use
1030 	 * rcu_nocb_unlock_irqrestore() anymore.
1031 	 */
1032 	raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1033 
1034 	/* Sanity check */
1035 	WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1036 
1037 
1038 	return ret;
1039 }
1040 
1041 int rcu_nocb_cpu_deoffload(int cpu)
1042 {
1043 	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1044 	int ret = 0;
1045 
1046 	mutex_lock(&rcu_state.barrier_mutex);
1047 	cpus_read_lock();
1048 	if (rcu_rdp_is_offloaded(rdp)) {
1049 		if (cpu_online(cpu)) {
1050 			ret = work_on_cpu(cpu, rcu_nocb_rdp_deoffload, rdp);
1051 			if (!ret)
1052 				cpumask_clear_cpu(cpu, rcu_nocb_mask);
1053 		} else {
1054 			pr_info("NOCB: Can't CB-deoffload an offline CPU\n");
1055 			ret = -EINVAL;
1056 		}
1057 	}
1058 	cpus_read_unlock();
1059 	mutex_unlock(&rcu_state.barrier_mutex);
1060 
1061 	return ret;
1062 }
1063 EXPORT_SYMBOL_GPL(rcu_nocb_cpu_deoffload);
1064 
1065 static long rcu_nocb_rdp_offload(void *arg)
1066 {
1067 	struct rcu_data *rdp = arg;
1068 	struct rcu_segcblist *cblist = &rdp->cblist;
1069 	unsigned long flags;
1070 	int ret;
1071 
1072 	WARN_ON_ONCE(rdp->cpu != raw_smp_processor_id());
1073 	/*
1074 	 * For now we only support re-offload, ie: the rdp must have been
1075 	 * offloaded on boot first.
1076 	 */
1077 	if (!rdp->nocb_gp_rdp)
1078 		return -EINVAL;
1079 
1080 	pr_info("Offloading %d\n", rdp->cpu);
1081 
1082 	/*
1083 	 * Cause future nocb_gp_wait() invocations to iterate over
1084 	 * structure, resetting ->nocb_gp_sleep and waking up the related
1085 	 * "rcuog".  Since nocb_gp_wait() in turn locks ->nocb_gp_lock
1086 	 * before setting ->nocb_gp_sleep again, we are guaranteed to
1087 	 * iterate this newly added structure before "rcuog" goes to
1088 	 * sleep again.
1089 	 */
1090 	list_add_tail_rcu(&rdp->nocb_entry_rdp, &rdp->nocb_gp_rdp->nocb_head_rdp);
1091 
1092 	/*
1093 	 * Can't use rcu_nocb_lock_irqsave() before SEGCBLIST_LOCKING
1094 	 * is set.
1095 	 */
1096 	raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
1097 
1098 	/*
1099 	 * We didn't take the nocb lock while working on the
1100 	 * rdp->cblist with SEGCBLIST_LOCKING cleared (pure softirq/rcuc mode).
1101 	 * Every modifications that have been done previously on
1102 	 * rdp->cblist must be visible remotely by the nocb kthreads
1103 	 * upon wake up after reading the cblist flags.
1104 	 *
1105 	 * The layout against nocb_lock enforces that ordering:
1106 	 *
1107 	 *  __rcu_nocb_rdp_offload()   nocb_cb_wait()/nocb_gp_wait()
1108 	 * -------------------------   ----------------------------
1109 	 *      WRITE callbacks           rcu_nocb_lock()
1110 	 *      rcu_nocb_lock()           READ flags
1111 	 *      WRITE flags               READ callbacks
1112 	 *      rcu_nocb_unlock()         rcu_nocb_unlock()
1113 	 */
1114 	ret = rdp_offload_toggle(rdp, true, flags);
1115 	swait_event_exclusive(rdp->nocb_state_wq,
1116 			      rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB) &&
1117 			      rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP));
1118 
1119 	/*
1120 	 * All kthreads are ready to work, we can finally relieve rcu_core() and
1121 	 * enable nocb bypass.
1122 	 */
1123 	rcu_nocb_lock_irqsave(rdp, flags);
1124 	rcu_segcblist_clear_flags(cblist, SEGCBLIST_RCU_CORE);
1125 	rcu_nocb_unlock_irqrestore(rdp, flags);
1126 
1127 	return ret;
1128 }
1129 
1130 int rcu_nocb_cpu_offload(int cpu)
1131 {
1132 	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1133 	int ret = 0;
1134 
1135 	mutex_lock(&rcu_state.barrier_mutex);
1136 	cpus_read_lock();
1137 	if (!rcu_rdp_is_offloaded(rdp)) {
1138 		if (cpu_online(cpu)) {
1139 			ret = work_on_cpu(cpu, rcu_nocb_rdp_offload, rdp);
1140 			if (!ret)
1141 				cpumask_set_cpu(cpu, rcu_nocb_mask);
1142 		} else {
1143 			pr_info("NOCB: Can't CB-offload an offline CPU\n");
1144 			ret = -EINVAL;
1145 		}
1146 	}
1147 	cpus_read_unlock();
1148 	mutex_unlock(&rcu_state.barrier_mutex);
1149 
1150 	return ret;
1151 }
1152 EXPORT_SYMBOL_GPL(rcu_nocb_cpu_offload);
1153 
1154 void __init rcu_init_nohz(void)
1155 {
1156 	int cpu;
1157 	bool need_rcu_nocb_mask = false;
1158 	struct rcu_data *rdp;
1159 
1160 #if defined(CONFIG_NO_HZ_FULL)
1161 	if (tick_nohz_full_running && !cpumask_empty(tick_nohz_full_mask))
1162 		need_rcu_nocb_mask = true;
1163 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
1164 
1165 	if (need_rcu_nocb_mask) {
1166 		if (!cpumask_available(rcu_nocb_mask)) {
1167 			if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
1168 				pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
1169 				return;
1170 			}
1171 		}
1172 		rcu_state.nocb_is_setup = true;
1173 	}
1174 
1175 	if (!rcu_state.nocb_is_setup)
1176 		return;
1177 
1178 #if defined(CONFIG_NO_HZ_FULL)
1179 	if (tick_nohz_full_running)
1180 		cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
1181 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
1182 
1183 	if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
1184 		pr_info("\tNote: kernel parameter 'rcu_nocbs=', 'nohz_full', or 'isolcpus=' contains nonexistent CPUs.\n");
1185 		cpumask_and(rcu_nocb_mask, cpu_possible_mask,
1186 			    rcu_nocb_mask);
1187 	}
1188 	if (cpumask_empty(rcu_nocb_mask))
1189 		pr_info("\tOffload RCU callbacks from CPUs: (none).\n");
1190 	else
1191 		pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
1192 			cpumask_pr_args(rcu_nocb_mask));
1193 	if (rcu_nocb_poll)
1194 		pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
1195 
1196 	for_each_cpu(cpu, rcu_nocb_mask) {
1197 		rdp = per_cpu_ptr(&rcu_data, cpu);
1198 		if (rcu_segcblist_empty(&rdp->cblist))
1199 			rcu_segcblist_init(&rdp->cblist);
1200 		rcu_segcblist_offload(&rdp->cblist, true);
1201 		rcu_segcblist_set_flags(&rdp->cblist, SEGCBLIST_KTHREAD_CB | SEGCBLIST_KTHREAD_GP);
1202 		rcu_segcblist_clear_flags(&rdp->cblist, SEGCBLIST_RCU_CORE);
1203 	}
1204 	rcu_organize_nocb_kthreads();
1205 }
1206 
1207 /* Initialize per-rcu_data variables for no-CBs CPUs. */
1208 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
1209 {
1210 	init_swait_queue_head(&rdp->nocb_cb_wq);
1211 	init_swait_queue_head(&rdp->nocb_gp_wq);
1212 	init_swait_queue_head(&rdp->nocb_state_wq);
1213 	raw_spin_lock_init(&rdp->nocb_lock);
1214 	raw_spin_lock_init(&rdp->nocb_bypass_lock);
1215 	raw_spin_lock_init(&rdp->nocb_gp_lock);
1216 	timer_setup(&rdp->nocb_timer, do_nocb_deferred_wakeup_timer, 0);
1217 	rcu_cblist_init(&rdp->nocb_bypass);
1218 	mutex_init(&rdp->nocb_gp_kthread_mutex);
1219 }
1220 
1221 /*
1222  * If the specified CPU is a no-CBs CPU that does not already have its
1223  * rcuo CB kthread, spawn it.  Additionally, if the rcuo GP kthread
1224  * for this CPU's group has not yet been created, spawn it as well.
1225  */
1226 static void rcu_spawn_cpu_nocb_kthread(int cpu)
1227 {
1228 	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1229 	struct rcu_data *rdp_gp;
1230 	struct task_struct *t;
1231 	struct sched_param sp;
1232 
1233 	if (!rcu_scheduler_fully_active || !rcu_state.nocb_is_setup)
1234 		return;
1235 
1236 	/* If there already is an rcuo kthread, then nothing to do. */
1237 	if (rdp->nocb_cb_kthread)
1238 		return;
1239 
1240 	/* If we didn't spawn the GP kthread first, reorganize! */
1241 	sp.sched_priority = kthread_prio;
1242 	rdp_gp = rdp->nocb_gp_rdp;
1243 	mutex_lock(&rdp_gp->nocb_gp_kthread_mutex);
1244 	if (!rdp_gp->nocb_gp_kthread) {
1245 		t = kthread_run(rcu_nocb_gp_kthread, rdp_gp,
1246 				"rcuog/%d", rdp_gp->cpu);
1247 		if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo GP kthread, OOM is now expected behavior\n", __func__)) {
1248 			mutex_unlock(&rdp_gp->nocb_gp_kthread_mutex);
1249 			return;
1250 		}
1251 		WRITE_ONCE(rdp_gp->nocb_gp_kthread, t);
1252 		if (kthread_prio)
1253 			sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1254 	}
1255 	mutex_unlock(&rdp_gp->nocb_gp_kthread_mutex);
1256 
1257 	/* Spawn the kthread for this CPU. */
1258 	t = kthread_run(rcu_nocb_cb_kthread, rdp,
1259 			"rcuo%c/%d", rcu_state.abbr, cpu);
1260 	if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo CB kthread, OOM is now expected behavior\n", __func__))
1261 		return;
1262 
1263 	if (kthread_prio)
1264 		sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1265 	WRITE_ONCE(rdp->nocb_cb_kthread, t);
1266 	WRITE_ONCE(rdp->nocb_gp_kthread, rdp_gp->nocb_gp_kthread);
1267 }
1268 
1269 /* How many CB CPU IDs per GP kthread?  Default of -1 for sqrt(nr_cpu_ids). */
1270 static int rcu_nocb_gp_stride = -1;
1271 module_param(rcu_nocb_gp_stride, int, 0444);
1272 
1273 /*
1274  * Initialize GP-CB relationships for all no-CBs CPU.
1275  */
1276 static void __init rcu_organize_nocb_kthreads(void)
1277 {
1278 	int cpu;
1279 	bool firsttime = true;
1280 	bool gotnocbs = false;
1281 	bool gotnocbscbs = true;
1282 	int ls = rcu_nocb_gp_stride;
1283 	int nl = 0;  /* Next GP kthread. */
1284 	struct rcu_data *rdp;
1285 	struct rcu_data *rdp_gp = NULL;  /* Suppress misguided gcc warn. */
1286 
1287 	if (!cpumask_available(rcu_nocb_mask))
1288 		return;
1289 	if (ls == -1) {
1290 		ls = nr_cpu_ids / int_sqrt(nr_cpu_ids);
1291 		rcu_nocb_gp_stride = ls;
1292 	}
1293 
1294 	/*
1295 	 * Each pass through this loop sets up one rcu_data structure.
1296 	 * Should the corresponding CPU come online in the future, then
1297 	 * we will spawn the needed set of rcu_nocb_kthread() kthreads.
1298 	 */
1299 	for_each_possible_cpu(cpu) {
1300 		rdp = per_cpu_ptr(&rcu_data, cpu);
1301 		if (rdp->cpu >= nl) {
1302 			/* New GP kthread, set up for CBs & next GP. */
1303 			gotnocbs = true;
1304 			nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
1305 			rdp_gp = rdp;
1306 			INIT_LIST_HEAD(&rdp->nocb_head_rdp);
1307 			if (dump_tree) {
1308 				if (!firsttime)
1309 					pr_cont("%s\n", gotnocbscbs
1310 							? "" : " (self only)");
1311 				gotnocbscbs = false;
1312 				firsttime = false;
1313 				pr_alert("%s: No-CB GP kthread CPU %d:",
1314 					 __func__, cpu);
1315 			}
1316 		} else {
1317 			/* Another CB kthread, link to previous GP kthread. */
1318 			gotnocbscbs = true;
1319 			if (dump_tree)
1320 				pr_cont(" %d", cpu);
1321 		}
1322 		rdp->nocb_gp_rdp = rdp_gp;
1323 		if (cpumask_test_cpu(cpu, rcu_nocb_mask))
1324 			list_add_tail(&rdp->nocb_entry_rdp, &rdp_gp->nocb_head_rdp);
1325 	}
1326 	if (gotnocbs && dump_tree)
1327 		pr_cont("%s\n", gotnocbscbs ? "" : " (self only)");
1328 }
1329 
1330 /*
1331  * Bind the current task to the offloaded CPUs.  If there are no offloaded
1332  * CPUs, leave the task unbound.  Splat if the bind attempt fails.
1333  */
1334 void rcu_bind_current_to_nocb(void)
1335 {
1336 	if (cpumask_available(rcu_nocb_mask) && !cpumask_empty(rcu_nocb_mask))
1337 		WARN_ON(sched_setaffinity(current->pid, rcu_nocb_mask));
1338 }
1339 EXPORT_SYMBOL_GPL(rcu_bind_current_to_nocb);
1340 
1341 // The ->on_cpu field is available only in CONFIG_SMP=y, so...
1342 #ifdef CONFIG_SMP
1343 static char *show_rcu_should_be_on_cpu(struct task_struct *tsp)
1344 {
1345 	return tsp && task_is_running(tsp) && !tsp->on_cpu ? "!" : "";
1346 }
1347 #else // #ifdef CONFIG_SMP
1348 static char *show_rcu_should_be_on_cpu(struct task_struct *tsp)
1349 {
1350 	return "";
1351 }
1352 #endif // #else #ifdef CONFIG_SMP
1353 
1354 /*
1355  * Dump out nocb grace-period kthread state for the specified rcu_data
1356  * structure.
1357  */
1358 static void show_rcu_nocb_gp_state(struct rcu_data *rdp)
1359 {
1360 	struct rcu_node *rnp = rdp->mynode;
1361 
1362 	pr_info("nocb GP %d %c%c%c%c%c %c[%c%c] %c%c:%ld rnp %d:%d %lu %c CPU %d%s\n",
1363 		rdp->cpu,
1364 		"kK"[!!rdp->nocb_gp_kthread],
1365 		"lL"[raw_spin_is_locked(&rdp->nocb_gp_lock)],
1366 		"dD"[!!rdp->nocb_defer_wakeup],
1367 		"tT"[timer_pending(&rdp->nocb_timer)],
1368 		"sS"[!!rdp->nocb_gp_sleep],
1369 		".W"[swait_active(&rdp->nocb_gp_wq)],
1370 		".W"[swait_active(&rnp->nocb_gp_wq[0])],
1371 		".W"[swait_active(&rnp->nocb_gp_wq[1])],
1372 		".B"[!!rdp->nocb_gp_bypass],
1373 		".G"[!!rdp->nocb_gp_gp],
1374 		(long)rdp->nocb_gp_seq,
1375 		rnp->grplo, rnp->grphi, READ_ONCE(rdp->nocb_gp_loops),
1376 		rdp->nocb_gp_kthread ? task_state_to_char(rdp->nocb_gp_kthread) : '.',
1377 		rdp->nocb_cb_kthread ? (int)task_cpu(rdp->nocb_gp_kthread) : -1,
1378 		show_rcu_should_be_on_cpu(rdp->nocb_cb_kthread));
1379 }
1380 
1381 /* Dump out nocb kthread state for the specified rcu_data structure. */
1382 static void show_rcu_nocb_state(struct rcu_data *rdp)
1383 {
1384 	char bufw[20];
1385 	char bufr[20];
1386 	struct rcu_data *nocb_next_rdp;
1387 	struct rcu_segcblist *rsclp = &rdp->cblist;
1388 	bool waslocked;
1389 	bool wassleep;
1390 
1391 	if (rdp->nocb_gp_rdp == rdp)
1392 		show_rcu_nocb_gp_state(rdp);
1393 
1394 	nocb_next_rdp = list_next_or_null_rcu(&rdp->nocb_gp_rdp->nocb_head_rdp,
1395 					      &rdp->nocb_entry_rdp,
1396 					      typeof(*rdp),
1397 					      nocb_entry_rdp);
1398 
1399 	sprintf(bufw, "%ld", rsclp->gp_seq[RCU_WAIT_TAIL]);
1400 	sprintf(bufr, "%ld", rsclp->gp_seq[RCU_NEXT_READY_TAIL]);
1401 	pr_info("   CB %d^%d->%d %c%c%c%c%c%c F%ld L%ld C%d %c%c%s%c%s%c%c q%ld %c CPU %d%s\n",
1402 		rdp->cpu, rdp->nocb_gp_rdp->cpu,
1403 		nocb_next_rdp ? nocb_next_rdp->cpu : -1,
1404 		"kK"[!!rdp->nocb_cb_kthread],
1405 		"bB"[raw_spin_is_locked(&rdp->nocb_bypass_lock)],
1406 		"cC"[!!atomic_read(&rdp->nocb_lock_contended)],
1407 		"lL"[raw_spin_is_locked(&rdp->nocb_lock)],
1408 		"sS"[!!rdp->nocb_cb_sleep],
1409 		".W"[swait_active(&rdp->nocb_cb_wq)],
1410 		jiffies - rdp->nocb_bypass_first,
1411 		jiffies - rdp->nocb_nobypass_last,
1412 		rdp->nocb_nobypass_count,
1413 		".D"[rcu_segcblist_ready_cbs(rsclp)],
1414 		".W"[!rcu_segcblist_segempty(rsclp, RCU_WAIT_TAIL)],
1415 		rcu_segcblist_segempty(rsclp, RCU_WAIT_TAIL) ? "" : bufw,
1416 		".R"[!rcu_segcblist_segempty(rsclp, RCU_NEXT_READY_TAIL)],
1417 		rcu_segcblist_segempty(rsclp, RCU_NEXT_READY_TAIL) ? "" : bufr,
1418 		".N"[!rcu_segcblist_segempty(rsclp, RCU_NEXT_TAIL)],
1419 		".B"[!!rcu_cblist_n_cbs(&rdp->nocb_bypass)],
1420 		rcu_segcblist_n_cbs(&rdp->cblist),
1421 		rdp->nocb_cb_kthread ? task_state_to_char(rdp->nocb_cb_kthread) : '.',
1422 		rdp->nocb_cb_kthread ? (int)task_cpu(rdp->nocb_gp_kthread) : -1,
1423 		show_rcu_should_be_on_cpu(rdp->nocb_cb_kthread));
1424 
1425 	/* It is OK for GP kthreads to have GP state. */
1426 	if (rdp->nocb_gp_rdp == rdp)
1427 		return;
1428 
1429 	waslocked = raw_spin_is_locked(&rdp->nocb_gp_lock);
1430 	wassleep = swait_active(&rdp->nocb_gp_wq);
1431 	if (!rdp->nocb_gp_sleep && !waslocked && !wassleep)
1432 		return;  /* Nothing untoward. */
1433 
1434 	pr_info("   nocb GP activity on CB-only CPU!!! %c%c%c %c\n",
1435 		"lL"[waslocked],
1436 		"dD"[!!rdp->nocb_defer_wakeup],
1437 		"sS"[!!rdp->nocb_gp_sleep],
1438 		".W"[wassleep]);
1439 }
1440 
1441 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
1442 
1443 static inline int rcu_lockdep_is_held_nocb(struct rcu_data *rdp)
1444 {
1445 	return 0;
1446 }
1447 
1448 static inline bool rcu_current_is_nocb_kthread(struct rcu_data *rdp)
1449 {
1450 	return false;
1451 }
1452 
1453 /* No ->nocb_lock to acquire.  */
1454 static void rcu_nocb_lock(struct rcu_data *rdp)
1455 {
1456 }
1457 
1458 /* No ->nocb_lock to release.  */
1459 static void rcu_nocb_unlock(struct rcu_data *rdp)
1460 {
1461 }
1462 
1463 /* No ->nocb_lock to release.  */
1464 static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
1465 				       unsigned long flags)
1466 {
1467 	local_irq_restore(flags);
1468 }
1469 
1470 /* Lockdep check that ->cblist may be safely accessed. */
1471 static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
1472 {
1473 	lockdep_assert_irqs_disabled();
1474 }
1475 
1476 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
1477 {
1478 }
1479 
1480 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
1481 {
1482 	return NULL;
1483 }
1484 
1485 static void rcu_init_one_nocb(struct rcu_node *rnp)
1486 {
1487 }
1488 
1489 static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1490 				  unsigned long j)
1491 {
1492 	return true;
1493 }
1494 
1495 static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1496 				bool *was_alldone, unsigned long flags)
1497 {
1498 	return false;
1499 }
1500 
1501 static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_empty,
1502 				 unsigned long flags)
1503 {
1504 	WARN_ON_ONCE(1);  /* Should be dead code! */
1505 }
1506 
1507 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
1508 {
1509 }
1510 
1511 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp, int level)
1512 {
1513 	return false;
1514 }
1515 
1516 static bool do_nocb_deferred_wakeup(struct rcu_data *rdp)
1517 {
1518 	return false;
1519 }
1520 
1521 static void rcu_spawn_cpu_nocb_kthread(int cpu)
1522 {
1523 }
1524 
1525 static void show_rcu_nocb_state(struct rcu_data *rdp)
1526 {
1527 }
1528 
1529 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
1530