xref: /openbmc/linux/kernel/rcu/rcu_segcblist.c (revision 76c6927c)
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  * Dequeue the oldest rcu_head structure from the specified callback
28  * list.  This function assumes that the callback is non-lazy, but
29  * the caller can later invoke rcu_cblist_dequeued_lazy() if it
30  * finds otherwise (and if it cares about laziness).  This allows
31  * different users to have different ways of determining laziness.
32  */
33 struct rcu_head *rcu_cblist_dequeue(struct rcu_cblist *rclp)
34 {
35 	struct rcu_head *rhp;
36 
37 	rhp = rclp->head;
38 	if (!rhp)
39 		return NULL;
40 	rclp->len--;
41 	rclp->head = rhp->next;
42 	if (!rclp->head)
43 		rclp->tail = &rclp->head;
44 	return rhp;
45 }
46 
47 /*
48  * Initialize an rcu_segcblist structure.
49  */
50 void rcu_segcblist_init(struct rcu_segcblist *rsclp)
51 {
52 	int i;
53 
54 	BUILD_BUG_ON(RCU_NEXT_TAIL + 1 != ARRAY_SIZE(rsclp->gp_seq));
55 	BUILD_BUG_ON(ARRAY_SIZE(rsclp->tails) != ARRAY_SIZE(rsclp->gp_seq));
56 	rsclp->head = NULL;
57 	for (i = 0; i < RCU_CBLIST_NSEGS; i++)
58 		rsclp->tails[i] = &rsclp->head;
59 	rsclp->len = 0;
60 	rsclp->len_lazy = 0;
61 	rsclp->enabled = 1;
62 }
63 
64 /*
65  * Disable the specified rcu_segcblist structure, so that callbacks can
66  * no longer be posted to it.  This structure must be empty.
67  */
68 void rcu_segcblist_disable(struct rcu_segcblist *rsclp)
69 {
70 	WARN_ON_ONCE(!rcu_segcblist_empty(rsclp));
71 	WARN_ON_ONCE(rcu_segcblist_n_cbs(rsclp));
72 	WARN_ON_ONCE(rcu_segcblist_n_lazy_cbs(rsclp));
73 	rsclp->enabled = 0;
74 }
75 
76 /*
77  * Mark the specified rcu_segcblist structure as offloaded.  This
78  * structure must be empty.
79  */
80 void rcu_segcblist_offload(struct rcu_segcblist *rsclp)
81 {
82 	WARN_ON_ONCE(!rcu_segcblist_empty(rsclp));
83 	WARN_ON_ONCE(rcu_segcblist_n_cbs(rsclp));
84 	WARN_ON_ONCE(rcu_segcblist_n_lazy_cbs(rsclp));
85 	rsclp->offloaded = 1;
86 }
87 
88 /*
89  * Does the specified rcu_segcblist structure contain callbacks that
90  * are ready to be invoked?
91  */
92 bool rcu_segcblist_ready_cbs(struct rcu_segcblist *rsclp)
93 {
94 	return rcu_segcblist_is_enabled(rsclp) &&
95 	       &rsclp->head != rsclp->tails[RCU_DONE_TAIL];
96 }
97 
98 /*
99  * Does the specified rcu_segcblist structure contain callbacks that
100  * are still pending, that is, not yet ready to be invoked?
101  */
102 bool rcu_segcblist_pend_cbs(struct rcu_segcblist *rsclp)
103 {
104 	return rcu_segcblist_is_enabled(rsclp) &&
105 	       !rcu_segcblist_restempty(rsclp, RCU_DONE_TAIL);
106 }
107 
108 /*
109  * Return a pointer to the first callback in the specified rcu_segcblist
110  * structure.  This is useful for diagnostics.
111  */
112 struct rcu_head *rcu_segcblist_first_cb(struct rcu_segcblist *rsclp)
113 {
114 	if (rcu_segcblist_is_enabled(rsclp))
115 		return rsclp->head;
116 	return NULL;
117 }
118 
119 /*
120  * Return a pointer to the first pending callback in the specified
121  * rcu_segcblist structure.  This is useful just after posting a given
122  * callback -- if that callback is the first pending callback, then
123  * you cannot rely on someone else having already started up the required
124  * grace period.
125  */
126 struct rcu_head *rcu_segcblist_first_pend_cb(struct rcu_segcblist *rsclp)
127 {
128 	if (rcu_segcblist_is_enabled(rsclp))
129 		return *rsclp->tails[RCU_DONE_TAIL];
130 	return NULL;
131 }
132 
133 /*
134  * Enqueue the specified callback onto the specified rcu_segcblist
135  * structure, updating accounting as needed.  Note that the ->len
136  * field may be accessed locklessly, hence the WRITE_ONCE().
137  * The ->len field is used by rcu_barrier() and friends to determine
138  * if it must post a callback on this structure, and it is OK
139  * for rcu_barrier() to sometimes post callbacks needlessly, but
140  * absolutely not OK for it to ever miss posting a callback.
141  */
142 void rcu_segcblist_enqueue(struct rcu_segcblist *rsclp,
143 			   struct rcu_head *rhp, bool lazy)
144 {
145 	WRITE_ONCE(rsclp->len, rsclp->len + 1); /* ->len sampled locklessly. */
146 	if (lazy)
147 		rsclp->len_lazy++;
148 	smp_mb(); /* Ensure counts are updated before callback is enqueued. */
149 	rhp->next = NULL;
150 	WRITE_ONCE(*rsclp->tails[RCU_NEXT_TAIL], rhp);
151 	WRITE_ONCE(rsclp->tails[RCU_NEXT_TAIL], &rhp->next);
152 }
153 
154 /*
155  * Entrain the specified callback onto the specified rcu_segcblist at
156  * the end of the last non-empty segment.  If the entire rcu_segcblist
157  * is empty, make no change, but return false.
158  *
159  * This is intended for use by rcu_barrier()-like primitives, -not-
160  * for normal grace-period use.  IMPORTANT:  The callback you enqueue
161  * will wait for all prior callbacks, NOT necessarily for a grace
162  * period.  You have been warned.
163  */
164 bool rcu_segcblist_entrain(struct rcu_segcblist *rsclp,
165 			   struct rcu_head *rhp, bool lazy)
166 {
167 	int i;
168 
169 	if (rcu_segcblist_n_cbs(rsclp) == 0)
170 		return false;
171 	WRITE_ONCE(rsclp->len, rsclp->len + 1);
172 	if (lazy)
173 		rsclp->len_lazy++;
174 	smp_mb(); /* Ensure counts are updated before callback is entrained. */
175 	rhp->next = NULL;
176 	for (i = RCU_NEXT_TAIL; i > RCU_DONE_TAIL; i--)
177 		if (rsclp->tails[i] != rsclp->tails[i - 1])
178 			break;
179 	WRITE_ONCE(*rsclp->tails[i], rhp);
180 	for (; i <= RCU_NEXT_TAIL; i++)
181 		WRITE_ONCE(rsclp->tails[i], &rhp->next);
182 	return true;
183 }
184 
185 /*
186  * Extract only the counts from the specified rcu_segcblist structure,
187  * and place them in the specified rcu_cblist structure.  This function
188  * supports both callback orphaning and invocation, hence the separation
189  * of counts and callbacks.  (Callbacks ready for invocation must be
190  * orphaned and adopted separately from pending callbacks, but counts
191  * apply to all callbacks.  Locking must be used to make sure that
192  * both orphaned-callbacks lists are consistent.)
193  */
194 void rcu_segcblist_extract_count(struct rcu_segcblist *rsclp,
195 					       struct rcu_cblist *rclp)
196 {
197 	rclp->len_lazy += rsclp->len_lazy;
198 	rclp->len += rsclp->len;
199 	rsclp->len_lazy = 0;
200 	WRITE_ONCE(rsclp->len, 0); /* ->len sampled locklessly. */
201 }
202 
203 /*
204  * Extract only those callbacks ready to be invoked from the specified
205  * rcu_segcblist structure and place them in the specified rcu_cblist
206  * structure.
207  */
208 void rcu_segcblist_extract_done_cbs(struct rcu_segcblist *rsclp,
209 				    struct rcu_cblist *rclp)
210 {
211 	int i;
212 
213 	if (!rcu_segcblist_ready_cbs(rsclp))
214 		return; /* Nothing to do. */
215 	*rclp->tail = rsclp->head;
216 	rsclp->head = *rsclp->tails[RCU_DONE_TAIL];
217 	WRITE_ONCE(*rsclp->tails[RCU_DONE_TAIL], NULL);
218 	rclp->tail = rsclp->tails[RCU_DONE_TAIL];
219 	for (i = RCU_CBLIST_NSEGS - 1; i >= RCU_DONE_TAIL; i--)
220 		if (rsclp->tails[i] == rsclp->tails[RCU_DONE_TAIL])
221 			WRITE_ONCE(rsclp->tails[i], &rsclp->head);
222 }
223 
224 /*
225  * Extract only those callbacks still pending (not yet ready to be
226  * invoked) from the specified rcu_segcblist structure and place them in
227  * the specified rcu_cblist structure.  Note that this loses information
228  * about any callbacks that might have been partway done waiting for
229  * their grace period.  Too bad!  They will have to start over.
230  */
231 void rcu_segcblist_extract_pend_cbs(struct rcu_segcblist *rsclp,
232 				    struct rcu_cblist *rclp)
233 {
234 	int i;
235 
236 	if (!rcu_segcblist_pend_cbs(rsclp))
237 		return; /* Nothing to do. */
238 	*rclp->tail = *rsclp->tails[RCU_DONE_TAIL];
239 	rclp->tail = rsclp->tails[RCU_NEXT_TAIL];
240 	WRITE_ONCE(*rsclp->tails[RCU_DONE_TAIL], NULL);
241 	for (i = RCU_DONE_TAIL + 1; i < RCU_CBLIST_NSEGS; i++)
242 		WRITE_ONCE(rsclp->tails[i], rsclp->tails[RCU_DONE_TAIL]);
243 }
244 
245 /*
246  * Insert counts from the specified rcu_cblist structure in the
247  * specified rcu_segcblist structure.
248  */
249 void rcu_segcblist_insert_count(struct rcu_segcblist *rsclp,
250 				struct rcu_cblist *rclp)
251 {
252 	rsclp->len_lazy += rclp->len_lazy;
253 	/* ->len sampled locklessly. */
254 	WRITE_ONCE(rsclp->len, rsclp->len + rclp->len);
255 	rclp->len_lazy = 0;
256 	rclp->len = 0;
257 }
258 
259 /*
260  * Move callbacks from the specified rcu_cblist to the beginning of the
261  * done-callbacks segment of the specified rcu_segcblist.
262  */
263 void rcu_segcblist_insert_done_cbs(struct rcu_segcblist *rsclp,
264 				   struct rcu_cblist *rclp)
265 {
266 	int i;
267 
268 	if (!rclp->head)
269 		return; /* No callbacks to move. */
270 	*rclp->tail = rsclp->head;
271 	rsclp->head = rclp->head;
272 	for (i = RCU_DONE_TAIL; i < RCU_CBLIST_NSEGS; i++)
273 		if (&rsclp->head == rsclp->tails[i])
274 			WRITE_ONCE(rsclp->tails[i], rclp->tail);
275 		else
276 			break;
277 	rclp->head = NULL;
278 	rclp->tail = &rclp->head;
279 }
280 
281 /*
282  * Move callbacks from the specified rcu_cblist to the end of the
283  * new-callbacks segment of the specified rcu_segcblist.
284  */
285 void rcu_segcblist_insert_pend_cbs(struct rcu_segcblist *rsclp,
286 				   struct rcu_cblist *rclp)
287 {
288 	if (!rclp->head)
289 		return; /* Nothing to do. */
290 	WRITE_ONCE(*rsclp->tails[RCU_NEXT_TAIL], rclp->head);
291 	WRITE_ONCE(rsclp->tails[RCU_NEXT_TAIL], rclp->tail);
292 	rclp->head = NULL;
293 	rclp->tail = &rclp->head;
294 }
295 
296 /*
297  * Advance the callbacks in the specified rcu_segcblist structure based
298  * on the current value passed in for the grace-period counter.
299  */
300 void rcu_segcblist_advance(struct rcu_segcblist *rsclp, unsigned long seq)
301 {
302 	int i, j;
303 
304 	WARN_ON_ONCE(!rcu_segcblist_is_enabled(rsclp));
305 	if (rcu_segcblist_restempty(rsclp, RCU_DONE_TAIL))
306 		return;
307 
308 	/*
309 	 * Find all callbacks whose ->gp_seq numbers indicate that they
310 	 * are ready to invoke, and put them into the RCU_DONE_TAIL segment.
311 	 */
312 	for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
313 		if (ULONG_CMP_LT(seq, rsclp->gp_seq[i]))
314 			break;
315 		WRITE_ONCE(rsclp->tails[RCU_DONE_TAIL], rsclp->tails[i]);
316 	}
317 
318 	/* If no callbacks moved, nothing more need be done. */
319 	if (i == RCU_WAIT_TAIL)
320 		return;
321 
322 	/* Clean up tail pointers that might have been misordered above. */
323 	for (j = RCU_WAIT_TAIL; j < i; j++)
324 		WRITE_ONCE(rsclp->tails[j], rsclp->tails[RCU_DONE_TAIL]);
325 
326 	/*
327 	 * Callbacks moved, so clean up the misordered ->tails[] pointers
328 	 * that now point into the middle of the list of ready-to-invoke
329 	 * callbacks.  The overall effect is to copy down the later pointers
330 	 * into the gap that was created by the now-ready segments.
331 	 */
332 	for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
333 		if (rsclp->tails[j] == rsclp->tails[RCU_NEXT_TAIL])
334 			break;  /* No more callbacks. */
335 		WRITE_ONCE(rsclp->tails[j], rsclp->tails[i]);
336 		rsclp->gp_seq[j] = rsclp->gp_seq[i];
337 	}
338 }
339 
340 /*
341  * "Accelerate" callbacks based on more-accurate grace-period information.
342  * The reason for this is that RCU does not synchronize the beginnings and
343  * ends of grace periods, and that callbacks are posted locally.  This in
344  * turn means that the callbacks must be labelled conservatively early
345  * on, as getting exact information would degrade both performance and
346  * scalability.  When more accurate grace-period information becomes
347  * available, previously posted callbacks can be "accelerated", marking
348  * them to complete at the end of the earlier grace period.
349  *
350  * This function operates on an rcu_segcblist structure, and also the
351  * grace-period sequence number seq at which new callbacks would become
352  * ready to invoke.  Returns true if there are callbacks that won't be
353  * ready to invoke until seq, false otherwise.
354  */
355 bool rcu_segcblist_accelerate(struct rcu_segcblist *rsclp, unsigned long seq)
356 {
357 	int i;
358 
359 	WARN_ON_ONCE(!rcu_segcblist_is_enabled(rsclp));
360 	if (rcu_segcblist_restempty(rsclp, RCU_DONE_TAIL))
361 		return false;
362 
363 	/*
364 	 * Find the segment preceding the oldest segment of callbacks
365 	 * whose ->gp_seq[] completion is at or after that passed in via
366 	 * "seq", skipping any empty segments.  This oldest segment, along
367 	 * with any later segments, can be merged in with any newly arrived
368 	 * callbacks in the RCU_NEXT_TAIL segment, and assigned "seq"
369 	 * as their ->gp_seq[] grace-period completion sequence number.
370 	 */
371 	for (i = RCU_NEXT_READY_TAIL; i > RCU_DONE_TAIL; i--)
372 		if (rsclp->tails[i] != rsclp->tails[i - 1] &&
373 		    ULONG_CMP_LT(rsclp->gp_seq[i], seq))
374 			break;
375 
376 	/*
377 	 * If all the segments contain callbacks that correspond to
378 	 * earlier grace-period sequence numbers than "seq", leave.
379 	 * Assuming that the rcu_segcblist structure has enough
380 	 * segments in its arrays, this can only happen if some of
381 	 * the non-done segments contain callbacks that really are
382 	 * ready to invoke.  This situation will get straightened
383 	 * out by the next call to rcu_segcblist_advance().
384 	 *
385 	 * Also advance to the oldest segment of callbacks whose
386 	 * ->gp_seq[] completion is at or after that passed in via "seq",
387 	 * skipping any empty segments.
388 	 */
389 	if (++i >= RCU_NEXT_TAIL)
390 		return false;
391 
392 	/*
393 	 * Merge all later callbacks, including newly arrived callbacks,
394 	 * into the segment located by the for-loop above.  Assign "seq"
395 	 * as the ->gp_seq[] value in order to correctly handle the case
396 	 * where there were no pending callbacks in the rcu_segcblist
397 	 * structure other than in the RCU_NEXT_TAIL segment.
398 	 */
399 	for (; i < RCU_NEXT_TAIL; i++) {
400 		WRITE_ONCE(rsclp->tails[i], rsclp->tails[RCU_NEXT_TAIL]);
401 		rsclp->gp_seq[i] = seq;
402 	}
403 	return true;
404 }
405 
406 /*
407  * Merge the source rcu_segcblist structure into the destination
408  * rcu_segcblist structure, then initialize the source.  Any pending
409  * callbacks from the source get to start over.  It is best to
410  * advance and accelerate both the destination and the source
411  * before merging.
412  */
413 void rcu_segcblist_merge(struct rcu_segcblist *dst_rsclp,
414 			 struct rcu_segcblist *src_rsclp)
415 {
416 	struct rcu_cblist donecbs;
417 	struct rcu_cblist pendcbs;
418 
419 	rcu_cblist_init(&donecbs);
420 	rcu_cblist_init(&pendcbs);
421 	rcu_segcblist_extract_count(src_rsclp, &donecbs);
422 	rcu_segcblist_extract_done_cbs(src_rsclp, &donecbs);
423 	rcu_segcblist_extract_pend_cbs(src_rsclp, &pendcbs);
424 	rcu_segcblist_insert_count(dst_rsclp, &donecbs);
425 	rcu_segcblist_insert_done_cbs(dst_rsclp, &donecbs);
426 	rcu_segcblist_insert_pend_cbs(dst_rsclp, &pendcbs);
427 	rcu_segcblist_init(src_rsclp);
428 }
429