xref: /openbmc/linux/kernel/rcu/rcu_segcblist.c (revision a44e4f3a)
1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * RCU segmented callback lists, function definitions
4  *
5  * Copyright IBM Corporation, 2017
6  *
7  * Authors: Paul E. McKenney <paulmck@linux.ibm.com>
8  */
9 
10 #include <linux/types.h>
11 #include <linux/kernel.h>
12 #include <linux/interrupt.h>
13 #include <linux/rcupdate.h>
14 
15 #include "rcu_segcblist.h"
16 
17 /* Initialize simple callback list. */
18 void rcu_cblist_init(struct rcu_cblist *rclp)
19 {
20 	rclp->head = NULL;
21 	rclp->tail = &rclp->head;
22 	rclp->len = 0;
23 	rclp->len_lazy = 0;
24 }
25 
26 /*
27  * Enqueue an rcu_head structure onto the specified callback list.
28  * This function assumes that the callback is non-lazy because it
29  * is intended for use by no-CBs CPUs, which do not distinguish
30  * between lazy and non-lazy RCU callbacks.
31  */
32 void rcu_cblist_enqueue(struct rcu_cblist *rclp, struct rcu_head *rhp)
33 {
34 	*rclp->tail = rhp;
35 	rclp->tail = &rhp->next;
36 	WRITE_ONCE(rclp->len, rclp->len + 1);
37 }
38 
39 /*
40  * Flush the second rcu_cblist structure onto the first one, obliterating
41  * any contents of the first.  If rhp is non-NULL, enqueue it as the sole
42  * element of the second rcu_cblist structure, but ensuring that the second
43  * rcu_cblist structure, if initially non-empty, always appears non-empty
44  * throughout the process.  If rdp is NULL, the second rcu_cblist structure
45  * is instead initialized to empty.
46  */
47 void rcu_cblist_flush_enqueue(struct rcu_cblist *drclp,
48 			      struct rcu_cblist *srclp,
49 			      struct rcu_head *rhp)
50 {
51 	drclp->head = srclp->head;
52 	if (drclp->head)
53 		drclp->tail = srclp->tail;
54 	else
55 		drclp->tail = &drclp->head;
56 	drclp->len = srclp->len;
57 	drclp->len_lazy = srclp->len_lazy;
58 	if (!rhp) {
59 		rcu_cblist_init(srclp);
60 	} else {
61 		rhp->next = NULL;
62 		srclp->head = rhp;
63 		srclp->tail = &rhp->next;
64 		WRITE_ONCE(srclp->len, 1);
65 		srclp->len_lazy = 0;
66 	}
67 }
68 
69 /*
70  * Dequeue the oldest rcu_head structure from the specified callback
71  * list.  This function assumes that the callback is non-lazy, but
72  * the caller can later invoke rcu_cblist_dequeued_lazy() if it
73  * finds otherwise (and if it cares about laziness).  This allows
74  * different users to have different ways of determining laziness.
75  */
76 struct rcu_head *rcu_cblist_dequeue(struct rcu_cblist *rclp)
77 {
78 	struct rcu_head *rhp;
79 
80 	rhp = rclp->head;
81 	if (!rhp)
82 		return NULL;
83 	rclp->len--;
84 	rclp->head = rhp->next;
85 	if (!rclp->head)
86 		rclp->tail = &rclp->head;
87 	return rhp;
88 }
89 
90 /* Set the length of an rcu_segcblist structure. */
91 void rcu_segcblist_set_len(struct rcu_segcblist *rsclp, long v)
92 {
93 #ifdef CONFIG_RCU_NOCB_CPU
94 	atomic_long_set(&rsclp->len, v);
95 #else
96 	WRITE_ONCE(rsclp->len, v);
97 #endif
98 }
99 
100 /*
101  * Increase the numeric length of an rcu_segcblist structure by the
102  * specified amount, which can be negative.  This can cause the ->len
103  * field to disagree with the actual number of callbacks on the structure.
104  * This increase is fully ordered with respect to the callers accesses
105  * both before and after.
106  */
107 void rcu_segcblist_add_len(struct rcu_segcblist *rsclp, long v)
108 {
109 #ifdef CONFIG_RCU_NOCB_CPU
110 	smp_mb__before_atomic(); /* Up to the caller! */
111 	atomic_long_add(v, &rsclp->len);
112 	smp_mb__after_atomic(); /* Up to the caller! */
113 #else
114 	smp_mb(); /* Up to the caller! */
115 	WRITE_ONCE(rsclp->len, rsclp->len + v);
116 	smp_mb(); /* Up to the caller! */
117 #endif
118 }
119 
120 /*
121  * Increase the numeric length of an rcu_segcblist structure by one.
122  * This can cause the ->len field to disagree with the actual number of
123  * callbacks on the structure.  This increase is fully ordered with respect
124  * to the callers accesses both before and after.
125  */
126 void rcu_segcblist_inc_len(struct rcu_segcblist *rsclp)
127 {
128 	rcu_segcblist_add_len(rsclp, 1);
129 }
130 
131 /*
132  * Exchange the numeric length of the specified rcu_segcblist structure
133  * with the specified value.  This can cause the ->len field to disagree
134  * with the actual number of callbacks on the structure.  This exchange is
135  * fully ordered with respect to the callers accesses both before and after.
136  */
137 long rcu_segcblist_xchg_len(struct rcu_segcblist *rsclp, long v)
138 {
139 #ifdef CONFIG_RCU_NOCB_CPU
140 	return atomic_long_xchg(&rsclp->len, v);
141 #else
142 	long ret = rsclp->len;
143 
144 	smp_mb(); /* Up to the caller! */
145 	WRITE_ONCE(rsclp->len, v);
146 	smp_mb(); /* Up to the caller! */
147 	return ret;
148 #endif
149 }
150 
151 /*
152  * Initialize an rcu_segcblist structure.
153  */
154 void rcu_segcblist_init(struct rcu_segcblist *rsclp)
155 {
156 	int i;
157 
158 	BUILD_BUG_ON(RCU_NEXT_TAIL + 1 != ARRAY_SIZE(rsclp->gp_seq));
159 	BUILD_BUG_ON(ARRAY_SIZE(rsclp->tails) != ARRAY_SIZE(rsclp->gp_seq));
160 	rsclp->head = NULL;
161 	for (i = 0; i < RCU_CBLIST_NSEGS; i++)
162 		rsclp->tails[i] = &rsclp->head;
163 	rcu_segcblist_set_len(rsclp, 0);
164 	rsclp->len_lazy = 0;
165 	rsclp->enabled = 1;
166 }
167 
168 /*
169  * Disable the specified rcu_segcblist structure, so that callbacks can
170  * no longer be posted to it.  This structure must be empty.
171  */
172 void rcu_segcblist_disable(struct rcu_segcblist *rsclp)
173 {
174 	WARN_ON_ONCE(!rcu_segcblist_empty(rsclp));
175 	WARN_ON_ONCE(rcu_segcblist_n_cbs(rsclp));
176 	WARN_ON_ONCE(rcu_segcblist_n_lazy_cbs(rsclp));
177 	rsclp->enabled = 0;
178 }
179 
180 /*
181  * Mark the specified rcu_segcblist structure as offloaded.  This
182  * structure must be empty.
183  */
184 void rcu_segcblist_offload(struct rcu_segcblist *rsclp)
185 {
186 	rsclp->offloaded = 1;
187 }
188 
189 /*
190  * Does the specified rcu_segcblist structure contain callbacks that
191  * are ready to be invoked?
192  */
193 bool rcu_segcblist_ready_cbs(struct rcu_segcblist *rsclp)
194 {
195 	return rcu_segcblist_is_enabled(rsclp) &&
196 	       &rsclp->head != rsclp->tails[RCU_DONE_TAIL];
197 }
198 
199 /*
200  * Does the specified rcu_segcblist structure contain callbacks that
201  * are still pending, that is, not yet ready to be invoked?
202  */
203 bool rcu_segcblist_pend_cbs(struct rcu_segcblist *rsclp)
204 {
205 	return rcu_segcblist_is_enabled(rsclp) &&
206 	       !rcu_segcblist_restempty(rsclp, RCU_DONE_TAIL);
207 }
208 
209 /*
210  * Return a pointer to the first callback in the specified rcu_segcblist
211  * structure.  This is useful for diagnostics.
212  */
213 struct rcu_head *rcu_segcblist_first_cb(struct rcu_segcblist *rsclp)
214 {
215 	if (rcu_segcblist_is_enabled(rsclp))
216 		return rsclp->head;
217 	return NULL;
218 }
219 
220 /*
221  * Return a pointer to the first pending callback in the specified
222  * rcu_segcblist structure.  This is useful just after posting a given
223  * callback -- if that callback is the first pending callback, then
224  * you cannot rely on someone else having already started up the required
225  * grace period.
226  */
227 struct rcu_head *rcu_segcblist_first_pend_cb(struct rcu_segcblist *rsclp)
228 {
229 	if (rcu_segcblist_is_enabled(rsclp))
230 		return *rsclp->tails[RCU_DONE_TAIL];
231 	return NULL;
232 }
233 
234 /*
235  * Return false if there are no CBs awaiting grace periods, otherwise,
236  * return true and store the nearest waited-upon grace period into *lp.
237  */
238 bool rcu_segcblist_nextgp(struct rcu_segcblist *rsclp, unsigned long *lp)
239 {
240 	if (!rcu_segcblist_pend_cbs(rsclp))
241 		return false;
242 	*lp = rsclp->gp_seq[RCU_WAIT_TAIL];
243 	return true;
244 }
245 
246 /*
247  * Enqueue the specified callback onto the specified rcu_segcblist
248  * structure, updating accounting as needed.  Note that the ->len
249  * field may be accessed locklessly, hence the WRITE_ONCE().
250  * The ->len field is used by rcu_barrier() and friends to determine
251  * if it must post a callback on this structure, and it is OK
252  * for rcu_barrier() to sometimes post callbacks needlessly, but
253  * absolutely not OK for it to ever miss posting a callback.
254  */
255 void rcu_segcblist_enqueue(struct rcu_segcblist *rsclp,
256 			   struct rcu_head *rhp, bool lazy)
257 {
258 	rcu_segcblist_inc_len(rsclp);
259 	if (lazy)
260 		rsclp->len_lazy++;
261 	smp_mb(); /* Ensure counts are updated before callback is enqueued. */
262 	rhp->next = NULL;
263 	WRITE_ONCE(*rsclp->tails[RCU_NEXT_TAIL], rhp);
264 	WRITE_ONCE(rsclp->tails[RCU_NEXT_TAIL], &rhp->next);
265 }
266 
267 /*
268  * Entrain the specified callback onto the specified rcu_segcblist at
269  * the end of the last non-empty segment.  If the entire rcu_segcblist
270  * is empty, make no change, but return false.
271  *
272  * This is intended for use by rcu_barrier()-like primitives, -not-
273  * for normal grace-period use.  IMPORTANT:  The callback you enqueue
274  * will wait for all prior callbacks, NOT necessarily for a grace
275  * period.  You have been warned.
276  */
277 bool rcu_segcblist_entrain(struct rcu_segcblist *rsclp,
278 			   struct rcu_head *rhp, bool lazy)
279 {
280 	int i;
281 
282 	if (rcu_segcblist_n_cbs(rsclp) == 0)
283 		return false;
284 	rcu_segcblist_inc_len(rsclp);
285 	if (lazy)
286 		rsclp->len_lazy++;
287 	smp_mb(); /* Ensure counts are updated before callback is entrained. */
288 	rhp->next = NULL;
289 	for (i = RCU_NEXT_TAIL; i > RCU_DONE_TAIL; i--)
290 		if (rsclp->tails[i] != rsclp->tails[i - 1])
291 			break;
292 	WRITE_ONCE(*rsclp->tails[i], rhp);
293 	for (; i <= RCU_NEXT_TAIL; i++)
294 		WRITE_ONCE(rsclp->tails[i], &rhp->next);
295 	return true;
296 }
297 
298 /*
299  * Extract only the counts from the specified rcu_segcblist structure,
300  * and place them in the specified rcu_cblist structure.  This function
301  * supports both callback orphaning and invocation, hence the separation
302  * of counts and callbacks.  (Callbacks ready for invocation must be
303  * orphaned and adopted separately from pending callbacks, but counts
304  * apply to all callbacks.  Locking must be used to make sure that
305  * both orphaned-callbacks lists are consistent.)
306  */
307 void rcu_segcblist_extract_count(struct rcu_segcblist *rsclp,
308 					       struct rcu_cblist *rclp)
309 {
310 	rclp->len_lazy += rsclp->len_lazy;
311 	rsclp->len_lazy = 0;
312 	rclp->len = rcu_segcblist_xchg_len(rsclp, 0);
313 }
314 
315 /*
316  * Extract only those callbacks ready to be invoked from the specified
317  * rcu_segcblist structure and place them in the specified rcu_cblist
318  * structure.
319  */
320 void rcu_segcblist_extract_done_cbs(struct rcu_segcblist *rsclp,
321 				    struct rcu_cblist *rclp)
322 {
323 	int i;
324 
325 	if (!rcu_segcblist_ready_cbs(rsclp))
326 		return; /* Nothing to do. */
327 	*rclp->tail = rsclp->head;
328 	WRITE_ONCE(rsclp->head, *rsclp->tails[RCU_DONE_TAIL]);
329 	WRITE_ONCE(*rsclp->tails[RCU_DONE_TAIL], NULL);
330 	rclp->tail = rsclp->tails[RCU_DONE_TAIL];
331 	for (i = RCU_CBLIST_NSEGS - 1; i >= RCU_DONE_TAIL; i--)
332 		if (rsclp->tails[i] == rsclp->tails[RCU_DONE_TAIL])
333 			WRITE_ONCE(rsclp->tails[i], &rsclp->head);
334 }
335 
336 /*
337  * Extract only those callbacks still pending (not yet ready to be
338  * invoked) from the specified rcu_segcblist structure and place them in
339  * the specified rcu_cblist structure.  Note that this loses information
340  * about any callbacks that might have been partway done waiting for
341  * their grace period.  Too bad!  They will have to start over.
342  */
343 void rcu_segcblist_extract_pend_cbs(struct rcu_segcblist *rsclp,
344 				    struct rcu_cblist *rclp)
345 {
346 	int i;
347 
348 	if (!rcu_segcblist_pend_cbs(rsclp))
349 		return; /* Nothing to do. */
350 	*rclp->tail = *rsclp->tails[RCU_DONE_TAIL];
351 	rclp->tail = rsclp->tails[RCU_NEXT_TAIL];
352 	WRITE_ONCE(*rsclp->tails[RCU_DONE_TAIL], NULL);
353 	for (i = RCU_DONE_TAIL + 1; i < RCU_CBLIST_NSEGS; i++)
354 		WRITE_ONCE(rsclp->tails[i], rsclp->tails[RCU_DONE_TAIL]);
355 }
356 
357 /*
358  * Insert counts from the specified rcu_cblist structure in the
359  * specified rcu_segcblist structure.
360  */
361 void rcu_segcblist_insert_count(struct rcu_segcblist *rsclp,
362 				struct rcu_cblist *rclp)
363 {
364 	rsclp->len_lazy += rclp->len_lazy;
365 	rcu_segcblist_add_len(rsclp, rclp->len);
366 	rclp->len_lazy = 0;
367 	rclp->len = 0;
368 }
369 
370 /*
371  * Move callbacks from the specified rcu_cblist to the beginning of the
372  * done-callbacks segment of the specified rcu_segcblist.
373  */
374 void rcu_segcblist_insert_done_cbs(struct rcu_segcblist *rsclp,
375 				   struct rcu_cblist *rclp)
376 {
377 	int i;
378 
379 	if (!rclp->head)
380 		return; /* No callbacks to move. */
381 	*rclp->tail = rsclp->head;
382 	WRITE_ONCE(rsclp->head, rclp->head);
383 	for (i = RCU_DONE_TAIL; i < RCU_CBLIST_NSEGS; i++)
384 		if (&rsclp->head == rsclp->tails[i])
385 			WRITE_ONCE(rsclp->tails[i], rclp->tail);
386 		else
387 			break;
388 	rclp->head = NULL;
389 	rclp->tail = &rclp->head;
390 }
391 
392 /*
393  * Move callbacks from the specified rcu_cblist to the end of the
394  * new-callbacks segment of the specified rcu_segcblist.
395  */
396 void rcu_segcblist_insert_pend_cbs(struct rcu_segcblist *rsclp,
397 				   struct rcu_cblist *rclp)
398 {
399 	if (!rclp->head)
400 		return; /* Nothing to do. */
401 	WRITE_ONCE(*rsclp->tails[RCU_NEXT_TAIL], rclp->head);
402 	WRITE_ONCE(rsclp->tails[RCU_NEXT_TAIL], rclp->tail);
403 	rclp->head = NULL;
404 	rclp->tail = &rclp->head;
405 }
406 
407 /*
408  * Advance the callbacks in the specified rcu_segcblist structure based
409  * on the current value passed in for the grace-period counter.
410  */
411 void rcu_segcblist_advance(struct rcu_segcblist *rsclp, unsigned long seq)
412 {
413 	int i, j;
414 
415 	WARN_ON_ONCE(!rcu_segcblist_is_enabled(rsclp));
416 	if (rcu_segcblist_restempty(rsclp, RCU_DONE_TAIL))
417 		return;
418 
419 	/*
420 	 * Find all callbacks whose ->gp_seq numbers indicate that they
421 	 * are ready to invoke, and put them into the RCU_DONE_TAIL segment.
422 	 */
423 	for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
424 		if (ULONG_CMP_LT(seq, rsclp->gp_seq[i]))
425 			break;
426 		WRITE_ONCE(rsclp->tails[RCU_DONE_TAIL], rsclp->tails[i]);
427 	}
428 
429 	/* If no callbacks moved, nothing more need be done. */
430 	if (i == RCU_WAIT_TAIL)
431 		return;
432 
433 	/* Clean up tail pointers that might have been misordered above. */
434 	for (j = RCU_WAIT_TAIL; j < i; j++)
435 		WRITE_ONCE(rsclp->tails[j], rsclp->tails[RCU_DONE_TAIL]);
436 
437 	/*
438 	 * Callbacks moved, so clean up the misordered ->tails[] pointers
439 	 * that now point into the middle of the list of ready-to-invoke
440 	 * callbacks.  The overall effect is to copy down the later pointers
441 	 * into the gap that was created by the now-ready segments.
442 	 */
443 	for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
444 		if (rsclp->tails[j] == rsclp->tails[RCU_NEXT_TAIL])
445 			break;  /* No more callbacks. */
446 		WRITE_ONCE(rsclp->tails[j], rsclp->tails[i]);
447 		rsclp->gp_seq[j] = rsclp->gp_seq[i];
448 	}
449 }
450 
451 /*
452  * "Accelerate" callbacks based on more-accurate grace-period information.
453  * The reason for this is that RCU does not synchronize the beginnings and
454  * ends of grace periods, and that callbacks are posted locally.  This in
455  * turn means that the callbacks must be labelled conservatively early
456  * on, as getting exact information would degrade both performance and
457  * scalability.  When more accurate grace-period information becomes
458  * available, previously posted callbacks can be "accelerated", marking
459  * them to complete at the end of the earlier grace period.
460  *
461  * This function operates on an rcu_segcblist structure, and also the
462  * grace-period sequence number seq at which new callbacks would become
463  * ready to invoke.  Returns true if there are callbacks that won't be
464  * ready to invoke until seq, false otherwise.
465  */
466 bool rcu_segcblist_accelerate(struct rcu_segcblist *rsclp, unsigned long seq)
467 {
468 	int i;
469 
470 	WARN_ON_ONCE(!rcu_segcblist_is_enabled(rsclp));
471 	if (rcu_segcblist_restempty(rsclp, RCU_DONE_TAIL))
472 		return false;
473 
474 	/*
475 	 * Find the segment preceding the oldest segment of callbacks
476 	 * whose ->gp_seq[] completion is at or after that passed in via
477 	 * "seq", skipping any empty segments.  This oldest segment, along
478 	 * with any later segments, can be merged in with any newly arrived
479 	 * callbacks in the RCU_NEXT_TAIL segment, and assigned "seq"
480 	 * as their ->gp_seq[] grace-period completion sequence number.
481 	 */
482 	for (i = RCU_NEXT_READY_TAIL; i > RCU_DONE_TAIL; i--)
483 		if (rsclp->tails[i] != rsclp->tails[i - 1] &&
484 		    ULONG_CMP_LT(rsclp->gp_seq[i], seq))
485 			break;
486 
487 	/*
488 	 * If all the segments contain callbacks that correspond to
489 	 * earlier grace-period sequence numbers than "seq", leave.
490 	 * Assuming that the rcu_segcblist structure has enough
491 	 * segments in its arrays, this can only happen if some of
492 	 * the non-done segments contain callbacks that really are
493 	 * ready to invoke.  This situation will get straightened
494 	 * out by the next call to rcu_segcblist_advance().
495 	 *
496 	 * Also advance to the oldest segment of callbacks whose
497 	 * ->gp_seq[] completion is at or after that passed in via "seq",
498 	 * skipping any empty segments.
499 	 */
500 	if (++i >= RCU_NEXT_TAIL)
501 		return false;
502 
503 	/*
504 	 * Merge all later callbacks, including newly arrived callbacks,
505 	 * into the segment located by the for-loop above.  Assign "seq"
506 	 * as the ->gp_seq[] value in order to correctly handle the case
507 	 * where there were no pending callbacks in the rcu_segcblist
508 	 * structure other than in the RCU_NEXT_TAIL segment.
509 	 */
510 	for (; i < RCU_NEXT_TAIL; i++) {
511 		WRITE_ONCE(rsclp->tails[i], rsclp->tails[RCU_NEXT_TAIL]);
512 		rsclp->gp_seq[i] = seq;
513 	}
514 	return true;
515 }
516 
517 /*
518  * Merge the source rcu_segcblist structure into the destination
519  * rcu_segcblist structure, then initialize the source.  Any pending
520  * callbacks from the source get to start over.  It is best to
521  * advance and accelerate both the destination and the source
522  * before merging.
523  */
524 void rcu_segcblist_merge(struct rcu_segcblist *dst_rsclp,
525 			 struct rcu_segcblist *src_rsclp)
526 {
527 	struct rcu_cblist donecbs;
528 	struct rcu_cblist pendcbs;
529 
530 	rcu_cblist_init(&donecbs);
531 	rcu_cblist_init(&pendcbs);
532 	rcu_segcblist_extract_count(src_rsclp, &donecbs);
533 	rcu_segcblist_extract_done_cbs(src_rsclp, &donecbs);
534 	rcu_segcblist_extract_pend_cbs(src_rsclp, &pendcbs);
535 	rcu_segcblist_insert_count(dst_rsclp, &donecbs);
536 	rcu_segcblist_insert_done_cbs(dst_rsclp, &donecbs);
537 	rcu_segcblist_insert_pend_cbs(dst_rsclp, &pendcbs);
538 	rcu_segcblist_init(src_rsclp);
539 }
540