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. */
rcu_lockdep_is_held_nocb(struct rcu_data * rdp)19 static inline int rcu_lockdep_is_held_nocb(struct rcu_data *rdp)
20 {
21 	return lockdep_is_held(&rdp->nocb_lock);
22 }
23 
rcu_current_is_nocb_kthread(struct rcu_data * rdp)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  */
rcu_nocb_setup(char * str)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 
parse_rcu_nocb_poll(char * arg)77 static int __init parse_rcu_nocb_poll(char *arg)
78 {
79 	rcu_nocb_poll = true;
80 	return 1;
81 }
82 __setup("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  */
rcu_nocb_bypass_lock(struct rcu_data * rdp)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  */
rcu_nocb_wait_contended(struct rcu_data * rdp)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  */
rcu_nocb_bypass_trylock(struct rcu_data * rdp)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  */
rcu_nocb_bypass_unlock(struct rcu_data * rdp)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  */
rcu_nocb_lock(struct rcu_data * rdp)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  */
rcu_nocb_unlock(struct rcu_data * rdp)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  */
rcu_nocb_unlock_irqrestore(struct rcu_data * rdp,unsigned long flags)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. */
rcu_lockdep_assert_cblist_protected(struct rcu_data * rdp)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  */
rcu_nocb_gp_cleanup(struct swait_queue_head * sq)199 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
200 {
201 	swake_up_all(sq);
202 }
203 
rcu_nocb_gp_get(struct rcu_node * rnp)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 
rcu_init_one_nocb(struct rcu_node * rnp)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 
__wake_nocb_gp(struct rcu_data * rdp_gp,struct rcu_data * rdp,bool force,unsigned long flags)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  */
wake_nocb_gp(struct rcu_data * rdp,bool force)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  * LAZY_FLUSH_JIFFIES decides the maximum amount of time that
261  * can elapse before lazy callbacks are flushed. Lazy callbacks
262  * could be flushed much earlier for a number of other reasons
263  * however, LAZY_FLUSH_JIFFIES will ensure no lazy callbacks are
264  * left unsubmitted to RCU after those many jiffies.
265  */
266 #define LAZY_FLUSH_JIFFIES (10 * HZ)
267 static unsigned long jiffies_till_flush = LAZY_FLUSH_JIFFIES;
268 
269 #ifdef CONFIG_RCU_LAZY
270 // To be called only from test code.
rcu_lazy_set_jiffies_till_flush(unsigned long jif)271 void rcu_lazy_set_jiffies_till_flush(unsigned long jif)
272 {
273 	jiffies_till_flush = jif;
274 }
275 EXPORT_SYMBOL(rcu_lazy_set_jiffies_till_flush);
276 
rcu_lazy_get_jiffies_till_flush(void)277 unsigned long rcu_lazy_get_jiffies_till_flush(void)
278 {
279 	return jiffies_till_flush;
280 }
281 EXPORT_SYMBOL(rcu_lazy_get_jiffies_till_flush);
282 #endif
283 
284 /*
285  * Arrange to wake the GP kthread for this NOCB group at some future
286  * time when it is safe to do so.
287  */
wake_nocb_gp_defer(struct rcu_data * rdp,int waketype,const char * reason)288 static void wake_nocb_gp_defer(struct rcu_data *rdp, int waketype,
289 			       const char *reason)
290 {
291 	unsigned long flags;
292 	struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
293 
294 	raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
295 
296 	/*
297 	 * Bypass wakeup overrides previous deferments. In case of
298 	 * callback storms, no need to wake up too early.
299 	 */
300 	if (waketype == RCU_NOCB_WAKE_LAZY &&
301 	    rdp->nocb_defer_wakeup == RCU_NOCB_WAKE_NOT) {
302 		mod_timer(&rdp_gp->nocb_timer, jiffies + jiffies_till_flush);
303 		WRITE_ONCE(rdp_gp->nocb_defer_wakeup, waketype);
304 	} else if (waketype == RCU_NOCB_WAKE_BYPASS) {
305 		mod_timer(&rdp_gp->nocb_timer, jiffies + 2);
306 		WRITE_ONCE(rdp_gp->nocb_defer_wakeup, waketype);
307 	} else {
308 		if (rdp_gp->nocb_defer_wakeup < RCU_NOCB_WAKE)
309 			mod_timer(&rdp_gp->nocb_timer, jiffies + 1);
310 		if (rdp_gp->nocb_defer_wakeup < waketype)
311 			WRITE_ONCE(rdp_gp->nocb_defer_wakeup, waketype);
312 	}
313 
314 	raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
315 
316 	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, reason);
317 }
318 
319 /*
320  * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
321  * However, if there is a callback to be enqueued and if ->nocb_bypass
322  * proves to be initially empty, just return false because the no-CB GP
323  * kthread may need to be awakened in this case.
324  *
325  * Return true if there was something to be flushed and it succeeded, otherwise
326  * false.
327  *
328  * Note that this function always returns true if rhp is NULL.
329  */
rcu_nocb_do_flush_bypass(struct rcu_data * rdp,struct rcu_head * rhp_in,unsigned long j,bool lazy)330 static bool rcu_nocb_do_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp_in,
331 				     unsigned long j, bool lazy)
332 {
333 	struct rcu_cblist rcl;
334 	struct rcu_head *rhp = rhp_in;
335 
336 	WARN_ON_ONCE(!rcu_rdp_is_offloaded(rdp));
337 	rcu_lockdep_assert_cblist_protected(rdp);
338 	lockdep_assert_held(&rdp->nocb_bypass_lock);
339 	if (rhp && !rcu_cblist_n_cbs(&rdp->nocb_bypass)) {
340 		raw_spin_unlock(&rdp->nocb_bypass_lock);
341 		return false;
342 	}
343 	/* Note: ->cblist.len already accounts for ->nocb_bypass contents. */
344 	if (rhp)
345 		rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
346 
347 	/*
348 	 * If the new CB requested was a lazy one, queue it onto the main
349 	 * ->cblist so that we can take advantage of the grace-period that will
350 	 * happen regardless. But queue it onto the bypass list first so that
351 	 * the lazy CB is ordered with the existing CBs in the bypass list.
352 	 */
353 	if (lazy && rhp) {
354 		rcu_cblist_enqueue(&rdp->nocb_bypass, rhp);
355 		rhp = NULL;
356 	}
357 	rcu_cblist_flush_enqueue(&rcl, &rdp->nocb_bypass, rhp);
358 	WRITE_ONCE(rdp->lazy_len, 0);
359 
360 	rcu_segcblist_insert_pend_cbs(&rdp->cblist, &rcl);
361 	WRITE_ONCE(rdp->nocb_bypass_first, j);
362 	rcu_nocb_bypass_unlock(rdp);
363 	return true;
364 }
365 
366 /*
367  * Flush the ->nocb_bypass queue into ->cblist, enqueuing rhp if non-NULL.
368  * However, if there is a callback to be enqueued and if ->nocb_bypass
369  * proves to be initially empty, just return false because the no-CB GP
370  * kthread may need to be awakened in this case.
371  *
372  * Note that this function always returns true if rhp is NULL.
373  */
rcu_nocb_flush_bypass(struct rcu_data * rdp,struct rcu_head * rhp,unsigned long j,bool lazy)374 static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
375 				  unsigned long j, bool lazy)
376 {
377 	if (!rcu_rdp_is_offloaded(rdp))
378 		return true;
379 	rcu_lockdep_assert_cblist_protected(rdp);
380 	rcu_nocb_bypass_lock(rdp);
381 	return rcu_nocb_do_flush_bypass(rdp, rhp, j, lazy);
382 }
383 
384 /*
385  * If the ->nocb_bypass_lock is immediately available, flush the
386  * ->nocb_bypass queue into ->cblist.
387  */
rcu_nocb_try_flush_bypass(struct rcu_data * rdp,unsigned long j)388 static void rcu_nocb_try_flush_bypass(struct rcu_data *rdp, unsigned long j)
389 {
390 	rcu_lockdep_assert_cblist_protected(rdp);
391 	if (!rcu_rdp_is_offloaded(rdp) ||
392 	    !rcu_nocb_bypass_trylock(rdp))
393 		return;
394 	WARN_ON_ONCE(!rcu_nocb_do_flush_bypass(rdp, NULL, j, false));
395 }
396 
397 /*
398  * See whether it is appropriate to use the ->nocb_bypass list in order
399  * to control contention on ->nocb_lock.  A limited number of direct
400  * enqueues are permitted into ->cblist per jiffy.  If ->nocb_bypass
401  * is non-empty, further callbacks must be placed into ->nocb_bypass,
402  * otherwise rcu_barrier() breaks.  Use rcu_nocb_flush_bypass() to switch
403  * back to direct use of ->cblist.  However, ->nocb_bypass should not be
404  * used if ->cblist is empty, because otherwise callbacks can be stranded
405  * on ->nocb_bypass because we cannot count on the current CPU ever again
406  * invoking call_rcu().  The general rule is that if ->nocb_bypass is
407  * non-empty, the corresponding no-CBs grace-period kthread must not be
408  * in an indefinite sleep state.
409  *
410  * Finally, it is not permitted to use the bypass during early boot,
411  * as doing so would confuse the auto-initialization code.  Besides
412  * which, there is no point in worrying about lock contention while
413  * there is only one CPU in operation.
414  */
rcu_nocb_try_bypass(struct rcu_data * rdp,struct rcu_head * rhp,bool * was_alldone,unsigned long flags,bool lazy)415 static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
416 				bool *was_alldone, unsigned long flags,
417 				bool lazy)
418 {
419 	unsigned long c;
420 	unsigned long cur_gp_seq;
421 	unsigned long j = jiffies;
422 	long ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
423 	bool bypass_is_lazy = (ncbs == READ_ONCE(rdp->lazy_len));
424 
425 	lockdep_assert_irqs_disabled();
426 
427 	// Pure softirq/rcuc based processing: no bypassing, no
428 	// locking.
429 	if (!rcu_rdp_is_offloaded(rdp)) {
430 		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
431 		return false;
432 	}
433 
434 	// In the process of (de-)offloading: no bypassing, but
435 	// locking.
436 	if (!rcu_segcblist_completely_offloaded(&rdp->cblist)) {
437 		rcu_nocb_lock(rdp);
438 		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
439 		return false; /* Not offloaded, no bypassing. */
440 	}
441 
442 	// Don't use ->nocb_bypass during early boot.
443 	if (rcu_scheduler_active != RCU_SCHEDULER_RUNNING) {
444 		rcu_nocb_lock(rdp);
445 		WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
446 		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
447 		return false;
448 	}
449 
450 	// If we have advanced to a new jiffy, reset counts to allow
451 	// moving back from ->nocb_bypass to ->cblist.
452 	if (j == rdp->nocb_nobypass_last) {
453 		c = rdp->nocb_nobypass_count + 1;
454 	} else {
455 		WRITE_ONCE(rdp->nocb_nobypass_last, j);
456 		c = rdp->nocb_nobypass_count - nocb_nobypass_lim_per_jiffy;
457 		if (ULONG_CMP_LT(rdp->nocb_nobypass_count,
458 				 nocb_nobypass_lim_per_jiffy))
459 			c = 0;
460 		else if (c > nocb_nobypass_lim_per_jiffy)
461 			c = nocb_nobypass_lim_per_jiffy;
462 	}
463 	WRITE_ONCE(rdp->nocb_nobypass_count, c);
464 
465 	// If there hasn't yet been all that many ->cblist enqueues
466 	// this jiffy, tell the caller to enqueue onto ->cblist.  But flush
467 	// ->nocb_bypass first.
468 	// Lazy CBs throttle this back and do immediate bypass queuing.
469 	if (rdp->nocb_nobypass_count < nocb_nobypass_lim_per_jiffy && !lazy) {
470 		rcu_nocb_lock(rdp);
471 		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
472 		if (*was_alldone)
473 			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
474 					    TPS("FirstQ"));
475 
476 		WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, j, false));
477 		WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
478 		return false; // Caller must enqueue the callback.
479 	}
480 
481 	// If ->nocb_bypass has been used too long or is too full,
482 	// flush ->nocb_bypass to ->cblist.
483 	if ((ncbs && !bypass_is_lazy && j != READ_ONCE(rdp->nocb_bypass_first)) ||
484 	    (ncbs &&  bypass_is_lazy &&
485 	     (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + jiffies_till_flush))) ||
486 	    ncbs >= qhimark) {
487 		rcu_nocb_lock(rdp);
488 		*was_alldone = !rcu_segcblist_pend_cbs(&rdp->cblist);
489 
490 		if (!rcu_nocb_flush_bypass(rdp, rhp, j, lazy)) {
491 			if (*was_alldone)
492 				trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
493 						    TPS("FirstQ"));
494 			WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
495 			return false; // Caller must enqueue the callback.
496 		}
497 		if (j != rdp->nocb_gp_adv_time &&
498 		    rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
499 		    rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
500 			rcu_advance_cbs_nowake(rdp->mynode, rdp);
501 			rdp->nocb_gp_adv_time = j;
502 		}
503 
504 		// The flush succeeded and we moved CBs into the regular list.
505 		// Don't wait for the wake up timer as it may be too far ahead.
506 		// Wake up the GP thread now instead, if the cblist was empty.
507 		__call_rcu_nocb_wake(rdp, *was_alldone, flags);
508 
509 		return true; // Callback already enqueued.
510 	}
511 
512 	// We need to use the bypass.
513 	rcu_nocb_wait_contended(rdp);
514 	rcu_nocb_bypass_lock(rdp);
515 	ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
516 	rcu_segcblist_inc_len(&rdp->cblist); /* Must precede enqueue. */
517 	rcu_cblist_enqueue(&rdp->nocb_bypass, rhp);
518 
519 	if (lazy)
520 		WRITE_ONCE(rdp->lazy_len, rdp->lazy_len + 1);
521 
522 	if (!ncbs) {
523 		WRITE_ONCE(rdp->nocb_bypass_first, j);
524 		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("FirstBQ"));
525 	}
526 	rcu_nocb_bypass_unlock(rdp);
527 	smp_mb(); /* Order enqueue before wake. */
528 	// A wake up of the grace period kthread or timer adjustment
529 	// needs to be done only if:
530 	// 1. Bypass list was fully empty before (this is the first
531 	//    bypass list entry), or:
532 	// 2. Both of these conditions are met:
533 	//    a. The bypass list previously had only lazy CBs, and:
534 	//    b. The new CB is non-lazy.
535 	if (ncbs && (!bypass_is_lazy || lazy)) {
536 		local_irq_restore(flags);
537 	} else {
538 		// No-CBs GP kthread might be indefinitely asleep, if so, wake.
539 		rcu_nocb_lock(rdp); // Rare during call_rcu() flood.
540 		if (!rcu_segcblist_pend_cbs(&rdp->cblist)) {
541 			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
542 					    TPS("FirstBQwake"));
543 			__call_rcu_nocb_wake(rdp, true, flags);
544 		} else {
545 			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
546 					    TPS("FirstBQnoWake"));
547 			rcu_nocb_unlock_irqrestore(rdp, flags);
548 		}
549 	}
550 	return true; // Callback already enqueued.
551 }
552 
553 /*
554  * Awaken the no-CBs grace-period kthread if needed, either due to it
555  * legitimately being asleep or due to overload conditions.
556  *
557  * If warranted, also wake up the kthread servicing this CPUs queues.
558  */
__call_rcu_nocb_wake(struct rcu_data * rdp,bool was_alldone,unsigned long flags)559 static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_alldone,
560 				 unsigned long flags)
561 				 __releases(rdp->nocb_lock)
562 {
563 	long bypass_len;
564 	unsigned long cur_gp_seq;
565 	unsigned long j;
566 	long lazy_len;
567 	long len;
568 	struct task_struct *t;
569 
570 	// If we are being polled or there is no kthread, just leave.
571 	t = READ_ONCE(rdp->nocb_gp_kthread);
572 	if (rcu_nocb_poll || !t) {
573 		rcu_nocb_unlock_irqrestore(rdp, flags);
574 		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
575 				    TPS("WakeNotPoll"));
576 		return;
577 	}
578 	// Need to actually to a wakeup.
579 	len = rcu_segcblist_n_cbs(&rdp->cblist);
580 	bypass_len = rcu_cblist_n_cbs(&rdp->nocb_bypass);
581 	lazy_len = READ_ONCE(rdp->lazy_len);
582 	if (was_alldone) {
583 		rdp->qlen_last_fqs_check = len;
584 		// Only lazy CBs in bypass list
585 		if (lazy_len && bypass_len == lazy_len) {
586 			rcu_nocb_unlock_irqrestore(rdp, flags);
587 			wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE_LAZY,
588 					   TPS("WakeLazy"));
589 		} else if (!irqs_disabled_flags(flags)) {
590 			/* ... if queue was empty ... */
591 			rcu_nocb_unlock_irqrestore(rdp, flags);
592 			wake_nocb_gp(rdp, false);
593 			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
594 					    TPS("WakeEmpty"));
595 		} else {
596 			rcu_nocb_unlock_irqrestore(rdp, flags);
597 			wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE,
598 					   TPS("WakeEmptyIsDeferred"));
599 		}
600 	} else if (len > rdp->qlen_last_fqs_check + qhimark) {
601 		/* ... or if many callbacks queued. */
602 		rdp->qlen_last_fqs_check = len;
603 		j = jiffies;
604 		if (j != rdp->nocb_gp_adv_time &&
605 		    rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
606 		    rcu_seq_done(&rdp->mynode->gp_seq, cur_gp_seq)) {
607 			rcu_advance_cbs_nowake(rdp->mynode, rdp);
608 			rdp->nocb_gp_adv_time = j;
609 		}
610 		smp_mb(); /* Enqueue before timer_pending(). */
611 		if ((rdp->nocb_cb_sleep ||
612 		     !rcu_segcblist_ready_cbs(&rdp->cblist)) &&
613 		    !timer_pending(&rdp->nocb_timer)) {
614 			rcu_nocb_unlock_irqrestore(rdp, flags);
615 			wake_nocb_gp_defer(rdp, RCU_NOCB_WAKE_FORCE,
616 					   TPS("WakeOvfIsDeferred"));
617 		} else {
618 			rcu_nocb_unlock_irqrestore(rdp, flags);
619 			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WakeNot"));
620 		}
621 	} else {
622 		rcu_nocb_unlock_irqrestore(rdp, flags);
623 		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WakeNot"));
624 	}
625 }
626 
nocb_gp_toggle_rdp(struct rcu_data * rdp,bool * wake_state)627 static int nocb_gp_toggle_rdp(struct rcu_data *rdp,
628 			       bool *wake_state)
629 {
630 	struct rcu_segcblist *cblist = &rdp->cblist;
631 	unsigned long flags;
632 	int ret;
633 
634 	rcu_nocb_lock_irqsave(rdp, flags);
635 	if (rcu_segcblist_test_flags(cblist, SEGCBLIST_OFFLOADED) &&
636 	    !rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP)) {
637 		/*
638 		 * Offloading. Set our flag and notify the offload worker.
639 		 * We will handle this rdp until it ever gets de-offloaded.
640 		 */
641 		rcu_segcblist_set_flags(cblist, SEGCBLIST_KTHREAD_GP);
642 		if (rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB))
643 			*wake_state = true;
644 		ret = 1;
645 	} else if (!rcu_segcblist_test_flags(cblist, SEGCBLIST_OFFLOADED) &&
646 		   rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP)) {
647 		/*
648 		 * De-offloading. Clear our flag and notify the de-offload worker.
649 		 * We will ignore this rdp until it ever gets re-offloaded.
650 		 */
651 		rcu_segcblist_clear_flags(cblist, SEGCBLIST_KTHREAD_GP);
652 		if (!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB))
653 			*wake_state = true;
654 		ret = 0;
655 	} else {
656 		WARN_ON_ONCE(1);
657 		ret = -1;
658 	}
659 
660 	rcu_nocb_unlock_irqrestore(rdp, flags);
661 
662 	return ret;
663 }
664 
nocb_gp_sleep(struct rcu_data * my_rdp,int cpu)665 static void nocb_gp_sleep(struct rcu_data *my_rdp, int cpu)
666 {
667 	trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Sleep"));
668 	swait_event_interruptible_exclusive(my_rdp->nocb_gp_wq,
669 					!READ_ONCE(my_rdp->nocb_gp_sleep));
670 	trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("EndSleep"));
671 }
672 
673 /*
674  * No-CBs GP kthreads come here to wait for additional callbacks to show up
675  * or for grace periods to end.
676  */
nocb_gp_wait(struct rcu_data * my_rdp)677 static void nocb_gp_wait(struct rcu_data *my_rdp)
678 {
679 	bool bypass = false;
680 	int __maybe_unused cpu = my_rdp->cpu;
681 	unsigned long cur_gp_seq;
682 	unsigned long flags;
683 	bool gotcbs = false;
684 	unsigned long j = jiffies;
685 	bool lazy = false;
686 	bool needwait_gp = false; // This prevents actual uninitialized use.
687 	bool needwake;
688 	bool needwake_gp;
689 	struct rcu_data *rdp, *rdp_toggling = NULL;
690 	struct rcu_node *rnp;
691 	unsigned long wait_gp_seq = 0; // Suppress "use uninitialized" warning.
692 	bool wasempty = false;
693 
694 	/*
695 	 * Each pass through the following loop checks for CBs and for the
696 	 * nearest grace period (if any) to wait for next.  The CB kthreads
697 	 * and the global grace-period kthread are awakened if needed.
698 	 */
699 	WARN_ON_ONCE(my_rdp->nocb_gp_rdp != my_rdp);
700 	/*
701 	 * An rcu_data structure is removed from the list after its
702 	 * CPU is de-offloaded and added to the list before that CPU is
703 	 * (re-)offloaded.  If the following loop happens to be referencing
704 	 * that rcu_data structure during the time that the corresponding
705 	 * CPU is de-offloaded and then immediately re-offloaded, this
706 	 * loop's rdp pointer will be carried to the end of the list by
707 	 * the resulting pair of list operations.  This can cause the loop
708 	 * to skip over some of the rcu_data structures that were supposed
709 	 * to have been scanned.  Fortunately a new iteration through the
710 	 * entire loop is forced after a given CPU's rcu_data structure
711 	 * is added to the list, so the skipped-over rcu_data structures
712 	 * won't be ignored for long.
713 	 */
714 	list_for_each_entry(rdp, &my_rdp->nocb_head_rdp, nocb_entry_rdp) {
715 		long bypass_ncbs;
716 		bool flush_bypass = false;
717 		long lazy_ncbs;
718 
719 		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Check"));
720 		rcu_nocb_lock_irqsave(rdp, flags);
721 		lockdep_assert_held(&rdp->nocb_lock);
722 		bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
723 		lazy_ncbs = READ_ONCE(rdp->lazy_len);
724 
725 		if (bypass_ncbs && (lazy_ncbs == bypass_ncbs) &&
726 		    (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + jiffies_till_flush) ||
727 		     bypass_ncbs > 2 * qhimark)) {
728 			flush_bypass = true;
729 		} else if (bypass_ncbs && (lazy_ncbs != bypass_ncbs) &&
730 		    (time_after(j, READ_ONCE(rdp->nocb_bypass_first) + 1) ||
731 		     bypass_ncbs > 2 * qhimark)) {
732 			flush_bypass = true;
733 		} else if (!bypass_ncbs && rcu_segcblist_empty(&rdp->cblist)) {
734 			rcu_nocb_unlock_irqrestore(rdp, flags);
735 			continue; /* No callbacks here, try next. */
736 		}
737 
738 		if (flush_bypass) {
739 			// Bypass full or old, so flush it.
740 			(void)rcu_nocb_try_flush_bypass(rdp, j);
741 			bypass_ncbs = rcu_cblist_n_cbs(&rdp->nocb_bypass);
742 			lazy_ncbs = READ_ONCE(rdp->lazy_len);
743 		}
744 
745 		if (bypass_ncbs) {
746 			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
747 					    bypass_ncbs == lazy_ncbs ? TPS("Lazy") : TPS("Bypass"));
748 			if (bypass_ncbs == lazy_ncbs)
749 				lazy = true;
750 			else
751 				bypass = true;
752 		}
753 		rnp = rdp->mynode;
754 
755 		// Advance callbacks if helpful and low contention.
756 		needwake_gp = false;
757 		if (!rcu_segcblist_restempty(&rdp->cblist,
758 					     RCU_NEXT_READY_TAIL) ||
759 		    (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq) &&
760 		     rcu_seq_done(&rnp->gp_seq, cur_gp_seq))) {
761 			raw_spin_lock_rcu_node(rnp); /* irqs disabled. */
762 			needwake_gp = rcu_advance_cbs(rnp, rdp);
763 			wasempty = rcu_segcblist_restempty(&rdp->cblist,
764 							   RCU_NEXT_READY_TAIL);
765 			raw_spin_unlock_rcu_node(rnp); /* irqs disabled. */
766 		}
767 		// Need to wait on some grace period?
768 		WARN_ON_ONCE(wasempty &&
769 			     !rcu_segcblist_restempty(&rdp->cblist,
770 						      RCU_NEXT_READY_TAIL));
771 		if (rcu_segcblist_nextgp(&rdp->cblist, &cur_gp_seq)) {
772 			if (!needwait_gp ||
773 			    ULONG_CMP_LT(cur_gp_seq, wait_gp_seq))
774 				wait_gp_seq = cur_gp_seq;
775 			needwait_gp = true;
776 			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
777 					    TPS("NeedWaitGP"));
778 		}
779 		if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
780 			needwake = rdp->nocb_cb_sleep;
781 			WRITE_ONCE(rdp->nocb_cb_sleep, false);
782 			smp_mb(); /* CB invocation -after- GP end. */
783 		} else {
784 			needwake = false;
785 		}
786 		rcu_nocb_unlock_irqrestore(rdp, flags);
787 		if (needwake) {
788 			swake_up_one(&rdp->nocb_cb_wq);
789 			gotcbs = true;
790 		}
791 		if (needwake_gp)
792 			rcu_gp_kthread_wake();
793 	}
794 
795 	my_rdp->nocb_gp_bypass = bypass;
796 	my_rdp->nocb_gp_gp = needwait_gp;
797 	my_rdp->nocb_gp_seq = needwait_gp ? wait_gp_seq : 0;
798 
799 	// At least one child with non-empty ->nocb_bypass, so set
800 	// timer in order to avoid stranding its callbacks.
801 	if (!rcu_nocb_poll) {
802 		// If bypass list only has lazy CBs. Add a deferred lazy wake up.
803 		if (lazy && !bypass) {
804 			wake_nocb_gp_defer(my_rdp, RCU_NOCB_WAKE_LAZY,
805 					TPS("WakeLazyIsDeferred"));
806 		// Otherwise add a deferred bypass wake up.
807 		} else if (bypass) {
808 			wake_nocb_gp_defer(my_rdp, RCU_NOCB_WAKE_BYPASS,
809 					TPS("WakeBypassIsDeferred"));
810 		}
811 	}
812 
813 	if (rcu_nocb_poll) {
814 		/* Polling, so trace if first poll in the series. */
815 		if (gotcbs)
816 			trace_rcu_nocb_wake(rcu_state.name, cpu, TPS("Poll"));
817 		if (list_empty(&my_rdp->nocb_head_rdp)) {
818 			raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
819 			if (!my_rdp->nocb_toggling_rdp)
820 				WRITE_ONCE(my_rdp->nocb_gp_sleep, true);
821 			raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
822 			/* Wait for any offloading rdp */
823 			nocb_gp_sleep(my_rdp, cpu);
824 		} else {
825 			schedule_timeout_idle(1);
826 		}
827 	} else if (!needwait_gp) {
828 		/* Wait for callbacks to appear. */
829 		nocb_gp_sleep(my_rdp, cpu);
830 	} else {
831 		rnp = my_rdp->mynode;
832 		trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("StartWait"));
833 		swait_event_interruptible_exclusive(
834 			rnp->nocb_gp_wq[rcu_seq_ctr(wait_gp_seq) & 0x1],
835 			rcu_seq_done(&rnp->gp_seq, wait_gp_seq) ||
836 			!READ_ONCE(my_rdp->nocb_gp_sleep));
837 		trace_rcu_this_gp(rnp, my_rdp, wait_gp_seq, TPS("EndWait"));
838 	}
839 
840 	if (!rcu_nocb_poll) {
841 		raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
842 		// (De-)queue an rdp to/from the group if its nocb state is changing
843 		rdp_toggling = my_rdp->nocb_toggling_rdp;
844 		if (rdp_toggling)
845 			my_rdp->nocb_toggling_rdp = NULL;
846 
847 		if (my_rdp->nocb_defer_wakeup > RCU_NOCB_WAKE_NOT) {
848 			WRITE_ONCE(my_rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
849 			del_timer(&my_rdp->nocb_timer);
850 		}
851 		WRITE_ONCE(my_rdp->nocb_gp_sleep, true);
852 		raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
853 	} else {
854 		rdp_toggling = READ_ONCE(my_rdp->nocb_toggling_rdp);
855 		if (rdp_toggling) {
856 			/*
857 			 * Paranoid locking to make sure nocb_toggling_rdp is well
858 			 * reset *before* we (re)set SEGCBLIST_KTHREAD_GP or we could
859 			 * race with another round of nocb toggling for this rdp.
860 			 * Nocb locking should prevent from that already but we stick
861 			 * to paranoia, especially in rare path.
862 			 */
863 			raw_spin_lock_irqsave(&my_rdp->nocb_gp_lock, flags);
864 			my_rdp->nocb_toggling_rdp = NULL;
865 			raw_spin_unlock_irqrestore(&my_rdp->nocb_gp_lock, flags);
866 		}
867 	}
868 
869 	if (rdp_toggling) {
870 		bool wake_state = false;
871 		int ret;
872 
873 		ret = nocb_gp_toggle_rdp(rdp_toggling, &wake_state);
874 		if (ret == 1)
875 			list_add_tail(&rdp_toggling->nocb_entry_rdp, &my_rdp->nocb_head_rdp);
876 		else if (ret == 0)
877 			list_del(&rdp_toggling->nocb_entry_rdp);
878 		if (wake_state)
879 			swake_up_one(&rdp_toggling->nocb_state_wq);
880 	}
881 
882 	my_rdp->nocb_gp_seq = -1;
883 	WARN_ON(signal_pending(current));
884 }
885 
886 /*
887  * No-CBs grace-period-wait kthread.  There is one of these per group
888  * of CPUs, but only once at least one CPU in that group has come online
889  * at least once since boot.  This kthread checks for newly posted
890  * callbacks from any of the CPUs it is responsible for, waits for a
891  * grace period, then awakens all of the rcu_nocb_cb_kthread() instances
892  * that then have callback-invocation work to do.
893  */
rcu_nocb_gp_kthread(void * arg)894 static int rcu_nocb_gp_kthread(void *arg)
895 {
896 	struct rcu_data *rdp = arg;
897 
898 	for (;;) {
899 		WRITE_ONCE(rdp->nocb_gp_loops, rdp->nocb_gp_loops + 1);
900 		nocb_gp_wait(rdp);
901 		cond_resched_tasks_rcu_qs();
902 	}
903 	return 0;
904 }
905 
nocb_cb_can_run(struct rcu_data * rdp)906 static inline bool nocb_cb_can_run(struct rcu_data *rdp)
907 {
908 	u8 flags = SEGCBLIST_OFFLOADED | SEGCBLIST_KTHREAD_CB;
909 
910 	return rcu_segcblist_test_flags(&rdp->cblist, flags);
911 }
912 
nocb_cb_wait_cond(struct rcu_data * rdp)913 static inline bool nocb_cb_wait_cond(struct rcu_data *rdp)
914 {
915 	return nocb_cb_can_run(rdp) && !READ_ONCE(rdp->nocb_cb_sleep);
916 }
917 
918 /*
919  * Invoke any ready callbacks from the corresponding no-CBs CPU,
920  * then, if there are no more, wait for more to appear.
921  */
nocb_cb_wait(struct rcu_data * rdp)922 static void nocb_cb_wait(struct rcu_data *rdp)
923 {
924 	struct rcu_segcblist *cblist = &rdp->cblist;
925 	unsigned long cur_gp_seq;
926 	unsigned long flags;
927 	bool needwake_state = false;
928 	bool needwake_gp = false;
929 	bool can_sleep = true;
930 	struct rcu_node *rnp = rdp->mynode;
931 
932 	do {
933 		swait_event_interruptible_exclusive(rdp->nocb_cb_wq,
934 						    nocb_cb_wait_cond(rdp));
935 
936 		// VVV Ensure CB invocation follows _sleep test.
937 		if (smp_load_acquire(&rdp->nocb_cb_sleep)) { // ^^^
938 			WARN_ON(signal_pending(current));
939 			trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WokeEmpty"));
940 		}
941 	} while (!nocb_cb_can_run(rdp));
942 
943 
944 	local_irq_save(flags);
945 	rcu_momentary_dyntick_idle();
946 	local_irq_restore(flags);
947 	/*
948 	 * Disable BH to provide the expected environment.  Also, when
949 	 * transitioning to/from NOCB mode, a self-requeuing callback might
950 	 * be invoked from softirq.  A short grace period could cause both
951 	 * instances of this callback would execute concurrently.
952 	 */
953 	local_bh_disable();
954 	rcu_do_batch(rdp);
955 	local_bh_enable();
956 	lockdep_assert_irqs_enabled();
957 	rcu_nocb_lock_irqsave(rdp, flags);
958 	if (rcu_segcblist_nextgp(cblist, &cur_gp_seq) &&
959 	    rcu_seq_done(&rnp->gp_seq, cur_gp_seq) &&
960 	    raw_spin_trylock_rcu_node(rnp)) { /* irqs already disabled. */
961 		needwake_gp = rcu_advance_cbs(rdp->mynode, rdp);
962 		raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
963 	}
964 
965 	if (rcu_segcblist_test_flags(cblist, SEGCBLIST_OFFLOADED)) {
966 		if (!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB)) {
967 			rcu_segcblist_set_flags(cblist, SEGCBLIST_KTHREAD_CB);
968 			if (rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP))
969 				needwake_state = true;
970 		}
971 		if (rcu_segcblist_ready_cbs(cblist))
972 			can_sleep = false;
973 	} else {
974 		/*
975 		 * De-offloading. Clear our flag and notify the de-offload worker.
976 		 * We won't touch the callbacks and keep sleeping until we ever
977 		 * get re-offloaded.
978 		 */
979 		WARN_ON_ONCE(!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB));
980 		rcu_segcblist_clear_flags(cblist, SEGCBLIST_KTHREAD_CB);
981 		if (!rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP))
982 			needwake_state = true;
983 	}
984 
985 	WRITE_ONCE(rdp->nocb_cb_sleep, can_sleep);
986 
987 	if (rdp->nocb_cb_sleep)
988 		trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("CBSleep"));
989 
990 	rcu_nocb_unlock_irqrestore(rdp, flags);
991 	if (needwake_gp)
992 		rcu_gp_kthread_wake();
993 
994 	if (needwake_state)
995 		swake_up_one(&rdp->nocb_state_wq);
996 }
997 
998 /*
999  * Per-rcu_data kthread, but only for no-CBs CPUs.  Repeatedly invoke
1000  * nocb_cb_wait() to do the dirty work.
1001  */
rcu_nocb_cb_kthread(void * arg)1002 static int rcu_nocb_cb_kthread(void *arg)
1003 {
1004 	struct rcu_data *rdp = arg;
1005 
1006 	// Each pass through this loop does one callback batch, and,
1007 	// if there are no more ready callbacks, waits for them.
1008 	for (;;) {
1009 		nocb_cb_wait(rdp);
1010 		cond_resched_tasks_rcu_qs();
1011 	}
1012 	return 0;
1013 }
1014 
1015 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
rcu_nocb_need_deferred_wakeup(struct rcu_data * rdp,int level)1016 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp, int level)
1017 {
1018 	return READ_ONCE(rdp->nocb_defer_wakeup) >= level;
1019 }
1020 
1021 /* Do a deferred wakeup of rcu_nocb_kthread(). */
do_nocb_deferred_wakeup_common(struct rcu_data * rdp_gp,struct rcu_data * rdp,int level,unsigned long flags)1022 static bool do_nocb_deferred_wakeup_common(struct rcu_data *rdp_gp,
1023 					   struct rcu_data *rdp, int level,
1024 					   unsigned long flags)
1025 	__releases(rdp_gp->nocb_gp_lock)
1026 {
1027 	int ndw;
1028 	int ret;
1029 
1030 	if (!rcu_nocb_need_deferred_wakeup(rdp_gp, level)) {
1031 		raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
1032 		return false;
1033 	}
1034 
1035 	ndw = rdp_gp->nocb_defer_wakeup;
1036 	ret = __wake_nocb_gp(rdp_gp, rdp, ndw == RCU_NOCB_WAKE_FORCE, flags);
1037 	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DeferredWake"));
1038 
1039 	return ret;
1040 }
1041 
1042 /* Do a deferred wakeup of rcu_nocb_kthread() from a timer handler. */
do_nocb_deferred_wakeup_timer(struct timer_list * t)1043 static void do_nocb_deferred_wakeup_timer(struct timer_list *t)
1044 {
1045 	unsigned long flags;
1046 	struct rcu_data *rdp = from_timer(rdp, t, nocb_timer);
1047 
1048 	WARN_ON_ONCE(rdp->nocb_gp_rdp != rdp);
1049 	trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("Timer"));
1050 
1051 	raw_spin_lock_irqsave(&rdp->nocb_gp_lock, flags);
1052 	smp_mb__after_spinlock(); /* Timer expire before wakeup. */
1053 	do_nocb_deferred_wakeup_common(rdp, rdp, RCU_NOCB_WAKE_BYPASS, flags);
1054 }
1055 
1056 /*
1057  * Do a deferred wakeup of rcu_nocb_kthread() from fastpath.
1058  * This means we do an inexact common-case check.  Note that if
1059  * we miss, ->nocb_timer will eventually clean things up.
1060  */
do_nocb_deferred_wakeup(struct rcu_data * rdp)1061 static bool do_nocb_deferred_wakeup(struct rcu_data *rdp)
1062 {
1063 	unsigned long flags;
1064 	struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
1065 
1066 	if (!rdp_gp || !rcu_nocb_need_deferred_wakeup(rdp_gp, RCU_NOCB_WAKE))
1067 		return false;
1068 
1069 	raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
1070 	return do_nocb_deferred_wakeup_common(rdp_gp, rdp, RCU_NOCB_WAKE, flags);
1071 }
1072 
rcu_nocb_flush_deferred_wakeup(void)1073 void rcu_nocb_flush_deferred_wakeup(void)
1074 {
1075 	do_nocb_deferred_wakeup(this_cpu_ptr(&rcu_data));
1076 }
1077 EXPORT_SYMBOL_GPL(rcu_nocb_flush_deferred_wakeup);
1078 
rdp_offload_toggle(struct rcu_data * rdp,bool offload,unsigned long flags)1079 static int rdp_offload_toggle(struct rcu_data *rdp,
1080 			       bool offload, unsigned long flags)
1081 	__releases(rdp->nocb_lock)
1082 {
1083 	struct rcu_segcblist *cblist = &rdp->cblist;
1084 	struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
1085 	bool wake_gp = false;
1086 
1087 	rcu_segcblist_offload(cblist, offload);
1088 
1089 	if (rdp->nocb_cb_sleep)
1090 		rdp->nocb_cb_sleep = false;
1091 	rcu_nocb_unlock_irqrestore(rdp, flags);
1092 
1093 	/*
1094 	 * Ignore former value of nocb_cb_sleep and force wake up as it could
1095 	 * have been spuriously set to false already.
1096 	 */
1097 	swake_up_one(&rdp->nocb_cb_wq);
1098 
1099 	raw_spin_lock_irqsave(&rdp_gp->nocb_gp_lock, flags);
1100 	// Queue this rdp for add/del to/from the list to iterate on rcuog
1101 	WRITE_ONCE(rdp_gp->nocb_toggling_rdp, rdp);
1102 	if (rdp_gp->nocb_gp_sleep) {
1103 		rdp_gp->nocb_gp_sleep = false;
1104 		wake_gp = true;
1105 	}
1106 	raw_spin_unlock_irqrestore(&rdp_gp->nocb_gp_lock, flags);
1107 
1108 	return wake_gp;
1109 }
1110 
rcu_nocb_rdp_deoffload(void * arg)1111 static long rcu_nocb_rdp_deoffload(void *arg)
1112 {
1113 	struct rcu_data *rdp = arg;
1114 	struct rcu_segcblist *cblist = &rdp->cblist;
1115 	unsigned long flags;
1116 	int wake_gp;
1117 	struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
1118 
1119 	/*
1120 	 * rcu_nocb_rdp_deoffload() may be called directly if
1121 	 * rcuog/o[p] spawn failed, because at this time the rdp->cpu
1122 	 * is not online yet.
1123 	 */
1124 	WARN_ON_ONCE((rdp->cpu != raw_smp_processor_id()) && cpu_online(rdp->cpu));
1125 
1126 	pr_info("De-offloading %d\n", rdp->cpu);
1127 
1128 	rcu_nocb_lock_irqsave(rdp, flags);
1129 	/*
1130 	 * Flush once and for all now. This suffices because we are
1131 	 * running on the target CPU holding ->nocb_lock (thus having
1132 	 * interrupts disabled), and because rdp_offload_toggle()
1133 	 * invokes rcu_segcblist_offload(), which clears SEGCBLIST_OFFLOADED.
1134 	 * Thus future calls to rcu_segcblist_completely_offloaded() will
1135 	 * return false, which means that future calls to rcu_nocb_try_bypass()
1136 	 * will refuse to put anything into the bypass.
1137 	 */
1138 	WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies, false));
1139 	/*
1140 	 * Start with invoking rcu_core() early. This way if the current thread
1141 	 * happens to preempt an ongoing call to rcu_core() in the middle,
1142 	 * leaving some work dismissed because rcu_core() still thinks the rdp is
1143 	 * completely offloaded, we are guaranteed a nearby future instance of
1144 	 * rcu_core() to catch up.
1145 	 */
1146 	rcu_segcblist_set_flags(cblist, SEGCBLIST_RCU_CORE);
1147 	invoke_rcu_core();
1148 	wake_gp = rdp_offload_toggle(rdp, false, flags);
1149 
1150 	mutex_lock(&rdp_gp->nocb_gp_kthread_mutex);
1151 	if (rdp_gp->nocb_gp_kthread) {
1152 		if (wake_gp)
1153 			wake_up_process(rdp_gp->nocb_gp_kthread);
1154 
1155 		/*
1156 		 * If rcuo[p] kthread spawn failed, directly remove SEGCBLIST_KTHREAD_CB.
1157 		 * Just wait SEGCBLIST_KTHREAD_GP to be cleared by rcuog.
1158 		 */
1159 		if (!rdp->nocb_cb_kthread) {
1160 			rcu_nocb_lock_irqsave(rdp, flags);
1161 			rcu_segcblist_clear_flags(&rdp->cblist, SEGCBLIST_KTHREAD_CB);
1162 			rcu_nocb_unlock_irqrestore(rdp, flags);
1163 		}
1164 
1165 		swait_event_exclusive(rdp->nocb_state_wq,
1166 					!rcu_segcblist_test_flags(cblist,
1167 					  SEGCBLIST_KTHREAD_CB | SEGCBLIST_KTHREAD_GP));
1168 	} else {
1169 		/*
1170 		 * No kthread to clear the flags for us or remove the rdp from the nocb list
1171 		 * to iterate. Do it here instead. Locking doesn't look stricly necessary
1172 		 * but we stick to paranoia in this rare path.
1173 		 */
1174 		rcu_nocb_lock_irqsave(rdp, flags);
1175 		rcu_segcblist_clear_flags(&rdp->cblist,
1176 				SEGCBLIST_KTHREAD_CB | SEGCBLIST_KTHREAD_GP);
1177 		rcu_nocb_unlock_irqrestore(rdp, flags);
1178 
1179 		list_del(&rdp->nocb_entry_rdp);
1180 	}
1181 	mutex_unlock(&rdp_gp->nocb_gp_kthread_mutex);
1182 
1183 	/*
1184 	 * Lock one last time to acquire latest callback updates from kthreads
1185 	 * so we can later handle callbacks locally without locking.
1186 	 */
1187 	rcu_nocb_lock_irqsave(rdp, flags);
1188 	/*
1189 	 * Theoretically we could clear SEGCBLIST_LOCKING after the nocb
1190 	 * lock is released but how about being paranoid for once?
1191 	 */
1192 	rcu_segcblist_clear_flags(cblist, SEGCBLIST_LOCKING);
1193 	/*
1194 	 * Without SEGCBLIST_LOCKING, we can't use
1195 	 * rcu_nocb_unlock_irqrestore() anymore.
1196 	 */
1197 	raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1198 
1199 	/* Sanity check */
1200 	WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
1201 
1202 
1203 	return 0;
1204 }
1205 
rcu_nocb_cpu_deoffload(int cpu)1206 int rcu_nocb_cpu_deoffload(int cpu)
1207 {
1208 	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1209 	int ret = 0;
1210 
1211 	cpus_read_lock();
1212 	mutex_lock(&rcu_state.barrier_mutex);
1213 	if (rcu_rdp_is_offloaded(rdp)) {
1214 		if (cpu_online(cpu)) {
1215 			ret = work_on_cpu(cpu, rcu_nocb_rdp_deoffload, rdp);
1216 			if (!ret)
1217 				cpumask_clear_cpu(cpu, rcu_nocb_mask);
1218 		} else {
1219 			pr_info("NOCB: Cannot CB-deoffload offline CPU %d\n", rdp->cpu);
1220 			ret = -EINVAL;
1221 		}
1222 	}
1223 	mutex_unlock(&rcu_state.barrier_mutex);
1224 	cpus_read_unlock();
1225 
1226 	return ret;
1227 }
1228 EXPORT_SYMBOL_GPL(rcu_nocb_cpu_deoffload);
1229 
rcu_nocb_rdp_offload(void * arg)1230 static long rcu_nocb_rdp_offload(void *arg)
1231 {
1232 	struct rcu_data *rdp = arg;
1233 	struct rcu_segcblist *cblist = &rdp->cblist;
1234 	unsigned long flags;
1235 	int wake_gp;
1236 	struct rcu_data *rdp_gp = rdp->nocb_gp_rdp;
1237 
1238 	WARN_ON_ONCE(rdp->cpu != raw_smp_processor_id());
1239 	/*
1240 	 * For now we only support re-offload, ie: the rdp must have been
1241 	 * offloaded on boot first.
1242 	 */
1243 	if (!rdp->nocb_gp_rdp)
1244 		return -EINVAL;
1245 
1246 	if (WARN_ON_ONCE(!rdp_gp->nocb_gp_kthread))
1247 		return -EINVAL;
1248 
1249 	pr_info("Offloading %d\n", rdp->cpu);
1250 
1251 	/*
1252 	 * Can't use rcu_nocb_lock_irqsave() before SEGCBLIST_LOCKING
1253 	 * is set.
1254 	 */
1255 	raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
1256 
1257 	/*
1258 	 * We didn't take the nocb lock while working on the
1259 	 * rdp->cblist with SEGCBLIST_LOCKING cleared (pure softirq/rcuc mode).
1260 	 * Every modifications that have been done previously on
1261 	 * rdp->cblist must be visible remotely by the nocb kthreads
1262 	 * upon wake up after reading the cblist flags.
1263 	 *
1264 	 * The layout against nocb_lock enforces that ordering:
1265 	 *
1266 	 *  __rcu_nocb_rdp_offload()   nocb_cb_wait()/nocb_gp_wait()
1267 	 * -------------------------   ----------------------------
1268 	 *      WRITE callbacks           rcu_nocb_lock()
1269 	 *      rcu_nocb_lock()           READ flags
1270 	 *      WRITE flags               READ callbacks
1271 	 *      rcu_nocb_unlock()         rcu_nocb_unlock()
1272 	 */
1273 	wake_gp = rdp_offload_toggle(rdp, true, flags);
1274 	if (wake_gp)
1275 		wake_up_process(rdp_gp->nocb_gp_kthread);
1276 	swait_event_exclusive(rdp->nocb_state_wq,
1277 			      rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_CB) &&
1278 			      rcu_segcblist_test_flags(cblist, SEGCBLIST_KTHREAD_GP));
1279 
1280 	/*
1281 	 * All kthreads are ready to work, we can finally relieve rcu_core() and
1282 	 * enable nocb bypass.
1283 	 */
1284 	rcu_nocb_lock_irqsave(rdp, flags);
1285 	rcu_segcblist_clear_flags(cblist, SEGCBLIST_RCU_CORE);
1286 	rcu_nocb_unlock_irqrestore(rdp, flags);
1287 
1288 	return 0;
1289 }
1290 
rcu_nocb_cpu_offload(int cpu)1291 int rcu_nocb_cpu_offload(int cpu)
1292 {
1293 	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1294 	int ret = 0;
1295 
1296 	cpus_read_lock();
1297 	mutex_lock(&rcu_state.barrier_mutex);
1298 	if (!rcu_rdp_is_offloaded(rdp)) {
1299 		if (cpu_online(cpu)) {
1300 			ret = work_on_cpu(cpu, rcu_nocb_rdp_offload, rdp);
1301 			if (!ret)
1302 				cpumask_set_cpu(cpu, rcu_nocb_mask);
1303 		} else {
1304 			pr_info("NOCB: Cannot CB-offload offline CPU %d\n", rdp->cpu);
1305 			ret = -EINVAL;
1306 		}
1307 	}
1308 	mutex_unlock(&rcu_state.barrier_mutex);
1309 	cpus_read_unlock();
1310 
1311 	return ret;
1312 }
1313 EXPORT_SYMBOL_GPL(rcu_nocb_cpu_offload);
1314 
1315 #ifdef CONFIG_RCU_LAZY
1316 static unsigned long
lazy_rcu_shrink_count(struct shrinker * shrink,struct shrink_control * sc)1317 lazy_rcu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1318 {
1319 	int cpu;
1320 	unsigned long count = 0;
1321 
1322 	if (WARN_ON_ONCE(!cpumask_available(rcu_nocb_mask)))
1323 		return 0;
1324 
1325 	/*  Protect rcu_nocb_mask against concurrent (de-)offloading. */
1326 	if (!mutex_trylock(&rcu_state.barrier_mutex))
1327 		return 0;
1328 
1329 	/* Snapshot count of all CPUs */
1330 	for_each_cpu(cpu, rcu_nocb_mask) {
1331 		struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1332 
1333 		count +=  READ_ONCE(rdp->lazy_len);
1334 	}
1335 
1336 	mutex_unlock(&rcu_state.barrier_mutex);
1337 
1338 	return count ? count : SHRINK_EMPTY;
1339 }
1340 
1341 static unsigned long
lazy_rcu_shrink_scan(struct shrinker * shrink,struct shrink_control * sc)1342 lazy_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1343 {
1344 	int cpu;
1345 	unsigned long flags;
1346 	unsigned long count = 0;
1347 
1348 	if (WARN_ON_ONCE(!cpumask_available(rcu_nocb_mask)))
1349 		return 0;
1350 	/*
1351 	 * Protect against concurrent (de-)offloading. Otherwise nocb locking
1352 	 * may be ignored or imbalanced.
1353 	 */
1354 	if (!mutex_trylock(&rcu_state.barrier_mutex)) {
1355 		/*
1356 		 * But really don't insist if barrier_mutex is contended since we
1357 		 * can't guarantee that it will never engage in a dependency
1358 		 * chain involving memory allocation. The lock is seldom contended
1359 		 * anyway.
1360 		 */
1361 		return 0;
1362 	}
1363 
1364 	/* Snapshot count of all CPUs */
1365 	for_each_cpu(cpu, rcu_nocb_mask) {
1366 		struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1367 		int _count;
1368 
1369 		if (WARN_ON_ONCE(!rcu_rdp_is_offloaded(rdp)))
1370 			continue;
1371 
1372 		if (!READ_ONCE(rdp->lazy_len))
1373 			continue;
1374 
1375 		rcu_nocb_lock_irqsave(rdp, flags);
1376 		/*
1377 		 * Recheck under the nocb lock. Since we are not holding the bypass
1378 		 * lock we may still race with increments from the enqueuer but still
1379 		 * we know for sure if there is at least one lazy callback.
1380 		 */
1381 		_count = READ_ONCE(rdp->lazy_len);
1382 		if (!_count) {
1383 			rcu_nocb_unlock_irqrestore(rdp, flags);
1384 			continue;
1385 		}
1386 		WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies, false));
1387 		rcu_nocb_unlock_irqrestore(rdp, flags);
1388 		wake_nocb_gp(rdp, false);
1389 		sc->nr_to_scan -= _count;
1390 		count += _count;
1391 		if (sc->nr_to_scan <= 0)
1392 			break;
1393 	}
1394 
1395 	mutex_unlock(&rcu_state.barrier_mutex);
1396 
1397 	return count ? count : SHRINK_STOP;
1398 }
1399 
1400 static struct shrinker lazy_rcu_shrinker = {
1401 	.count_objects = lazy_rcu_shrink_count,
1402 	.scan_objects = lazy_rcu_shrink_scan,
1403 	.batch = 0,
1404 	.seeks = DEFAULT_SEEKS,
1405 };
1406 #endif // #ifdef CONFIG_RCU_LAZY
1407 
rcu_init_nohz(void)1408 void __init rcu_init_nohz(void)
1409 {
1410 	int cpu;
1411 	struct rcu_data *rdp;
1412 	const struct cpumask *cpumask = NULL;
1413 
1414 #if defined(CONFIG_NO_HZ_FULL)
1415 	if (tick_nohz_full_running && !cpumask_empty(tick_nohz_full_mask))
1416 		cpumask = tick_nohz_full_mask;
1417 #endif
1418 
1419 	if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_DEFAULT_ALL) &&
1420 	    !rcu_state.nocb_is_setup && !cpumask)
1421 		cpumask = cpu_possible_mask;
1422 
1423 	if (cpumask) {
1424 		if (!cpumask_available(rcu_nocb_mask)) {
1425 			if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
1426 				pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
1427 				return;
1428 			}
1429 		}
1430 
1431 		cpumask_or(rcu_nocb_mask, rcu_nocb_mask, cpumask);
1432 		rcu_state.nocb_is_setup = true;
1433 	}
1434 
1435 	if (!rcu_state.nocb_is_setup)
1436 		return;
1437 
1438 #ifdef CONFIG_RCU_LAZY
1439 	if (register_shrinker(&lazy_rcu_shrinker, "rcu-lazy"))
1440 		pr_err("Failed to register lazy_rcu shrinker!\n");
1441 #endif // #ifdef CONFIG_RCU_LAZY
1442 
1443 	if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
1444 		pr_info("\tNote: kernel parameter 'rcu_nocbs=', 'nohz_full', or 'isolcpus=' contains nonexistent CPUs.\n");
1445 		cpumask_and(rcu_nocb_mask, cpu_possible_mask,
1446 			    rcu_nocb_mask);
1447 	}
1448 	if (cpumask_empty(rcu_nocb_mask))
1449 		pr_info("\tOffload RCU callbacks from CPUs: (none).\n");
1450 	else
1451 		pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
1452 			cpumask_pr_args(rcu_nocb_mask));
1453 	if (rcu_nocb_poll)
1454 		pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
1455 
1456 	for_each_cpu(cpu, rcu_nocb_mask) {
1457 		rdp = per_cpu_ptr(&rcu_data, cpu);
1458 		if (rcu_segcblist_empty(&rdp->cblist))
1459 			rcu_segcblist_init(&rdp->cblist);
1460 		rcu_segcblist_offload(&rdp->cblist, true);
1461 		rcu_segcblist_set_flags(&rdp->cblist, SEGCBLIST_KTHREAD_CB | SEGCBLIST_KTHREAD_GP);
1462 		rcu_segcblist_clear_flags(&rdp->cblist, SEGCBLIST_RCU_CORE);
1463 	}
1464 	rcu_organize_nocb_kthreads();
1465 }
1466 
1467 /* Initialize per-rcu_data variables for no-CBs CPUs. */
rcu_boot_init_nocb_percpu_data(struct rcu_data * rdp)1468 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
1469 {
1470 	init_swait_queue_head(&rdp->nocb_cb_wq);
1471 	init_swait_queue_head(&rdp->nocb_gp_wq);
1472 	init_swait_queue_head(&rdp->nocb_state_wq);
1473 	raw_spin_lock_init(&rdp->nocb_lock);
1474 	raw_spin_lock_init(&rdp->nocb_bypass_lock);
1475 	raw_spin_lock_init(&rdp->nocb_gp_lock);
1476 	timer_setup(&rdp->nocb_timer, do_nocb_deferred_wakeup_timer, 0);
1477 	rcu_cblist_init(&rdp->nocb_bypass);
1478 	WRITE_ONCE(rdp->lazy_len, 0);
1479 	mutex_init(&rdp->nocb_gp_kthread_mutex);
1480 }
1481 
1482 /*
1483  * If the specified CPU is a no-CBs CPU that does not already have its
1484  * rcuo CB kthread, spawn it.  Additionally, if the rcuo GP kthread
1485  * for this CPU's group has not yet been created, spawn it as well.
1486  */
rcu_spawn_cpu_nocb_kthread(int cpu)1487 static void rcu_spawn_cpu_nocb_kthread(int cpu)
1488 {
1489 	struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1490 	struct rcu_data *rdp_gp;
1491 	struct task_struct *t;
1492 	struct sched_param sp;
1493 
1494 	if (!rcu_scheduler_fully_active || !rcu_state.nocb_is_setup)
1495 		return;
1496 
1497 	/* If there already is an rcuo kthread, then nothing to do. */
1498 	if (rdp->nocb_cb_kthread)
1499 		return;
1500 
1501 	/* If we didn't spawn the GP kthread first, reorganize! */
1502 	sp.sched_priority = kthread_prio;
1503 	rdp_gp = rdp->nocb_gp_rdp;
1504 	mutex_lock(&rdp_gp->nocb_gp_kthread_mutex);
1505 	if (!rdp_gp->nocb_gp_kthread) {
1506 		t = kthread_run(rcu_nocb_gp_kthread, rdp_gp,
1507 				"rcuog/%d", rdp_gp->cpu);
1508 		if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo GP kthread, OOM is now expected behavior\n", __func__)) {
1509 			mutex_unlock(&rdp_gp->nocb_gp_kthread_mutex);
1510 			goto end;
1511 		}
1512 		WRITE_ONCE(rdp_gp->nocb_gp_kthread, t);
1513 		if (kthread_prio)
1514 			sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1515 	}
1516 	mutex_unlock(&rdp_gp->nocb_gp_kthread_mutex);
1517 
1518 	/* Spawn the kthread for this CPU. */
1519 	t = kthread_run(rcu_nocb_cb_kthread, rdp,
1520 			"rcuo%c/%d", rcu_state.abbr, cpu);
1521 	if (WARN_ONCE(IS_ERR(t), "%s: Could not start rcuo CB kthread, OOM is now expected behavior\n", __func__))
1522 		goto end;
1523 
1524 	if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_CB_BOOST) && kthread_prio)
1525 		sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1526 
1527 	WRITE_ONCE(rdp->nocb_cb_kthread, t);
1528 	WRITE_ONCE(rdp->nocb_gp_kthread, rdp_gp->nocb_gp_kthread);
1529 	return;
1530 end:
1531 	mutex_lock(&rcu_state.barrier_mutex);
1532 	if (rcu_rdp_is_offloaded(rdp)) {
1533 		rcu_nocb_rdp_deoffload(rdp);
1534 		cpumask_clear_cpu(cpu, rcu_nocb_mask);
1535 	}
1536 	mutex_unlock(&rcu_state.barrier_mutex);
1537 }
1538 
1539 /* How many CB CPU IDs per GP kthread?  Default of -1 for sqrt(nr_cpu_ids). */
1540 static int rcu_nocb_gp_stride = -1;
1541 module_param(rcu_nocb_gp_stride, int, 0444);
1542 
1543 /*
1544  * Initialize GP-CB relationships for all no-CBs CPU.
1545  */
rcu_organize_nocb_kthreads(void)1546 static void __init rcu_organize_nocb_kthreads(void)
1547 {
1548 	int cpu;
1549 	bool firsttime = true;
1550 	bool gotnocbs = false;
1551 	bool gotnocbscbs = true;
1552 	int ls = rcu_nocb_gp_stride;
1553 	int nl = 0;  /* Next GP kthread. */
1554 	struct rcu_data *rdp;
1555 	struct rcu_data *rdp_gp = NULL;  /* Suppress misguided gcc warn. */
1556 
1557 	if (!cpumask_available(rcu_nocb_mask))
1558 		return;
1559 	if (ls == -1) {
1560 		ls = nr_cpu_ids / int_sqrt(nr_cpu_ids);
1561 		rcu_nocb_gp_stride = ls;
1562 	}
1563 
1564 	/*
1565 	 * Each pass through this loop sets up one rcu_data structure.
1566 	 * Should the corresponding CPU come online in the future, then
1567 	 * we will spawn the needed set of rcu_nocb_kthread() kthreads.
1568 	 */
1569 	for_each_possible_cpu(cpu) {
1570 		rdp = per_cpu_ptr(&rcu_data, cpu);
1571 		if (rdp->cpu >= nl) {
1572 			/* New GP kthread, set up for CBs & next GP. */
1573 			gotnocbs = true;
1574 			nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
1575 			rdp_gp = rdp;
1576 			INIT_LIST_HEAD(&rdp->nocb_head_rdp);
1577 			if (dump_tree) {
1578 				if (!firsttime)
1579 					pr_cont("%s\n", gotnocbscbs
1580 							? "" : " (self only)");
1581 				gotnocbscbs = false;
1582 				firsttime = false;
1583 				pr_alert("%s: No-CB GP kthread CPU %d:",
1584 					 __func__, cpu);
1585 			}
1586 		} else {
1587 			/* Another CB kthread, link to previous GP kthread. */
1588 			gotnocbscbs = true;
1589 			if (dump_tree)
1590 				pr_cont(" %d", cpu);
1591 		}
1592 		rdp->nocb_gp_rdp = rdp_gp;
1593 		if (cpumask_test_cpu(cpu, rcu_nocb_mask))
1594 			list_add_tail(&rdp->nocb_entry_rdp, &rdp_gp->nocb_head_rdp);
1595 	}
1596 	if (gotnocbs && dump_tree)
1597 		pr_cont("%s\n", gotnocbscbs ? "" : " (self only)");
1598 }
1599 
1600 /*
1601  * Bind the current task to the offloaded CPUs.  If there are no offloaded
1602  * CPUs, leave the task unbound.  Splat if the bind attempt fails.
1603  */
rcu_bind_current_to_nocb(void)1604 void rcu_bind_current_to_nocb(void)
1605 {
1606 	if (cpumask_available(rcu_nocb_mask) && !cpumask_empty(rcu_nocb_mask))
1607 		WARN_ON(sched_setaffinity(current->pid, rcu_nocb_mask));
1608 }
1609 EXPORT_SYMBOL_GPL(rcu_bind_current_to_nocb);
1610 
1611 // The ->on_cpu field is available only in CONFIG_SMP=y, so...
1612 #ifdef CONFIG_SMP
show_rcu_should_be_on_cpu(struct task_struct * tsp)1613 static char *show_rcu_should_be_on_cpu(struct task_struct *tsp)
1614 {
1615 	return tsp && task_is_running(tsp) && !tsp->on_cpu ? "!" : "";
1616 }
1617 #else // #ifdef CONFIG_SMP
show_rcu_should_be_on_cpu(struct task_struct * tsp)1618 static char *show_rcu_should_be_on_cpu(struct task_struct *tsp)
1619 {
1620 	return "";
1621 }
1622 #endif // #else #ifdef CONFIG_SMP
1623 
1624 /*
1625  * Dump out nocb grace-period kthread state for the specified rcu_data
1626  * structure.
1627  */
show_rcu_nocb_gp_state(struct rcu_data * rdp)1628 static void show_rcu_nocb_gp_state(struct rcu_data *rdp)
1629 {
1630 	struct rcu_node *rnp = rdp->mynode;
1631 
1632 	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",
1633 		rdp->cpu,
1634 		"kK"[!!rdp->nocb_gp_kthread],
1635 		"lL"[raw_spin_is_locked(&rdp->nocb_gp_lock)],
1636 		"dD"[!!rdp->nocb_defer_wakeup],
1637 		"tT"[timer_pending(&rdp->nocb_timer)],
1638 		"sS"[!!rdp->nocb_gp_sleep],
1639 		".W"[swait_active(&rdp->nocb_gp_wq)],
1640 		".W"[swait_active(&rnp->nocb_gp_wq[0])],
1641 		".W"[swait_active(&rnp->nocb_gp_wq[1])],
1642 		".B"[!!rdp->nocb_gp_bypass],
1643 		".G"[!!rdp->nocb_gp_gp],
1644 		(long)rdp->nocb_gp_seq,
1645 		rnp->grplo, rnp->grphi, READ_ONCE(rdp->nocb_gp_loops),
1646 		rdp->nocb_gp_kthread ? task_state_to_char(rdp->nocb_gp_kthread) : '.',
1647 		rdp->nocb_gp_kthread ? (int)task_cpu(rdp->nocb_gp_kthread) : -1,
1648 		show_rcu_should_be_on_cpu(rdp->nocb_gp_kthread));
1649 }
1650 
1651 /* Dump out nocb kthread state for the specified rcu_data structure. */
show_rcu_nocb_state(struct rcu_data * rdp)1652 static void show_rcu_nocb_state(struct rcu_data *rdp)
1653 {
1654 	char bufw[20];
1655 	char bufr[20];
1656 	struct rcu_data *nocb_next_rdp;
1657 	struct rcu_segcblist *rsclp = &rdp->cblist;
1658 	bool waslocked;
1659 	bool wassleep;
1660 
1661 	if (rdp->nocb_gp_rdp == rdp)
1662 		show_rcu_nocb_gp_state(rdp);
1663 
1664 	nocb_next_rdp = list_next_or_null_rcu(&rdp->nocb_gp_rdp->nocb_head_rdp,
1665 					      &rdp->nocb_entry_rdp,
1666 					      typeof(*rdp),
1667 					      nocb_entry_rdp);
1668 
1669 	sprintf(bufw, "%ld", rsclp->gp_seq[RCU_WAIT_TAIL]);
1670 	sprintf(bufr, "%ld", rsclp->gp_seq[RCU_NEXT_READY_TAIL]);
1671 	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",
1672 		rdp->cpu, rdp->nocb_gp_rdp->cpu,
1673 		nocb_next_rdp ? nocb_next_rdp->cpu : -1,
1674 		"kK"[!!rdp->nocb_cb_kthread],
1675 		"bB"[raw_spin_is_locked(&rdp->nocb_bypass_lock)],
1676 		"cC"[!!atomic_read(&rdp->nocb_lock_contended)],
1677 		"lL"[raw_spin_is_locked(&rdp->nocb_lock)],
1678 		"sS"[!!rdp->nocb_cb_sleep],
1679 		".W"[swait_active(&rdp->nocb_cb_wq)],
1680 		jiffies - rdp->nocb_bypass_first,
1681 		jiffies - rdp->nocb_nobypass_last,
1682 		rdp->nocb_nobypass_count,
1683 		".D"[rcu_segcblist_ready_cbs(rsclp)],
1684 		".W"[!rcu_segcblist_segempty(rsclp, RCU_WAIT_TAIL)],
1685 		rcu_segcblist_segempty(rsclp, RCU_WAIT_TAIL) ? "" : bufw,
1686 		".R"[!rcu_segcblist_segempty(rsclp, RCU_NEXT_READY_TAIL)],
1687 		rcu_segcblist_segempty(rsclp, RCU_NEXT_READY_TAIL) ? "" : bufr,
1688 		".N"[!rcu_segcblist_segempty(rsclp, RCU_NEXT_TAIL)],
1689 		".B"[!!rcu_cblist_n_cbs(&rdp->nocb_bypass)],
1690 		rcu_segcblist_n_cbs(&rdp->cblist),
1691 		rdp->nocb_cb_kthread ? task_state_to_char(rdp->nocb_cb_kthread) : '.',
1692 		rdp->nocb_cb_kthread ? (int)task_cpu(rdp->nocb_cb_kthread) : -1,
1693 		show_rcu_should_be_on_cpu(rdp->nocb_cb_kthread));
1694 
1695 	/* It is OK for GP kthreads to have GP state. */
1696 	if (rdp->nocb_gp_rdp == rdp)
1697 		return;
1698 
1699 	waslocked = raw_spin_is_locked(&rdp->nocb_gp_lock);
1700 	wassleep = swait_active(&rdp->nocb_gp_wq);
1701 	if (!rdp->nocb_gp_sleep && !waslocked && !wassleep)
1702 		return;  /* Nothing untoward. */
1703 
1704 	pr_info("   nocb GP activity on CB-only CPU!!! %c%c%c %c\n",
1705 		"lL"[waslocked],
1706 		"dD"[!!rdp->nocb_defer_wakeup],
1707 		"sS"[!!rdp->nocb_gp_sleep],
1708 		".W"[wassleep]);
1709 }
1710 
1711 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
1712 
rcu_lockdep_is_held_nocb(struct rcu_data * rdp)1713 static inline int rcu_lockdep_is_held_nocb(struct rcu_data *rdp)
1714 {
1715 	return 0;
1716 }
1717 
rcu_current_is_nocb_kthread(struct rcu_data * rdp)1718 static inline bool rcu_current_is_nocb_kthread(struct rcu_data *rdp)
1719 {
1720 	return false;
1721 }
1722 
1723 /* No ->nocb_lock to acquire.  */
rcu_nocb_lock(struct rcu_data * rdp)1724 static void rcu_nocb_lock(struct rcu_data *rdp)
1725 {
1726 }
1727 
1728 /* No ->nocb_lock to release.  */
rcu_nocb_unlock(struct rcu_data * rdp)1729 static void rcu_nocb_unlock(struct rcu_data *rdp)
1730 {
1731 }
1732 
1733 /* No ->nocb_lock to release.  */
rcu_nocb_unlock_irqrestore(struct rcu_data * rdp,unsigned long flags)1734 static void rcu_nocb_unlock_irqrestore(struct rcu_data *rdp,
1735 				       unsigned long flags)
1736 {
1737 	local_irq_restore(flags);
1738 }
1739 
1740 /* Lockdep check that ->cblist may be safely accessed. */
rcu_lockdep_assert_cblist_protected(struct rcu_data * rdp)1741 static void rcu_lockdep_assert_cblist_protected(struct rcu_data *rdp)
1742 {
1743 	lockdep_assert_irqs_disabled();
1744 }
1745 
rcu_nocb_gp_cleanup(struct swait_queue_head * sq)1746 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
1747 {
1748 }
1749 
rcu_nocb_gp_get(struct rcu_node * rnp)1750 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
1751 {
1752 	return NULL;
1753 }
1754 
rcu_init_one_nocb(struct rcu_node * rnp)1755 static void rcu_init_one_nocb(struct rcu_node *rnp)
1756 {
1757 }
1758 
wake_nocb_gp(struct rcu_data * rdp,bool force)1759 static bool wake_nocb_gp(struct rcu_data *rdp, bool force)
1760 {
1761 	return false;
1762 }
1763 
rcu_nocb_flush_bypass(struct rcu_data * rdp,struct rcu_head * rhp,unsigned long j,bool lazy)1764 static bool rcu_nocb_flush_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1765 				  unsigned long j, bool lazy)
1766 {
1767 	return true;
1768 }
1769 
rcu_nocb_try_bypass(struct rcu_data * rdp,struct rcu_head * rhp,bool * was_alldone,unsigned long flags,bool lazy)1770 static bool rcu_nocb_try_bypass(struct rcu_data *rdp, struct rcu_head *rhp,
1771 				bool *was_alldone, unsigned long flags, bool lazy)
1772 {
1773 	return false;
1774 }
1775 
__call_rcu_nocb_wake(struct rcu_data * rdp,bool was_empty,unsigned long flags)1776 static void __call_rcu_nocb_wake(struct rcu_data *rdp, bool was_empty,
1777 				 unsigned long flags)
1778 {
1779 	WARN_ON_ONCE(1);  /* Should be dead code! */
1780 }
1781 
rcu_boot_init_nocb_percpu_data(struct rcu_data * rdp)1782 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
1783 {
1784 }
1785 
rcu_nocb_need_deferred_wakeup(struct rcu_data * rdp,int level)1786 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp, int level)
1787 {
1788 	return false;
1789 }
1790 
do_nocb_deferred_wakeup(struct rcu_data * rdp)1791 static bool do_nocb_deferred_wakeup(struct rcu_data *rdp)
1792 {
1793 	return false;
1794 }
1795 
rcu_spawn_cpu_nocb_kthread(int cpu)1796 static void rcu_spawn_cpu_nocb_kthread(int cpu)
1797 {
1798 }
1799 
show_rcu_nocb_state(struct rcu_data * rdp)1800 static void show_rcu_nocb_state(struct rcu_data *rdp)
1801 {
1802 }
1803 
1804 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
1805