Linux Audio

Check our new training course

Loading...
v6.8
  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/cpu.h>
 11#include <linux/interrupt.h>
 12#include <linux/kernel.h>
 13#include <linux/types.h>
 14
 15#include "rcu_segcblist.h"
 16
 17/* Initialize simple callback list. */
 18void rcu_cblist_init(struct rcu_cblist *rclp)
 19{
 20	rclp->head = NULL;
 21	rclp->tail = &rclp->head;
 22	rclp->len = 0;
 23}
 24
 25/*
 26 * Enqueue an rcu_head structure onto the specified callback list.
 27 */
 28void rcu_cblist_enqueue(struct rcu_cblist *rclp, struct rcu_head *rhp)
 29{
 30	*rclp->tail = rhp;
 31	rclp->tail = &rhp->next;
 32	WRITE_ONCE(rclp->len, rclp->len + 1);
 33}
 34
 35/*
 36 * Flush the second rcu_cblist structure onto the first one, obliterating
 37 * any contents of the first.  If rhp is non-NULL, enqueue it as the sole
 38 * element of the second rcu_cblist structure, but ensuring that the second
 39 * rcu_cblist structure, if initially non-empty, always appears non-empty
 40 * throughout the process.  If rdp is NULL, the second rcu_cblist structure
 41 * is instead initialized to empty.
 42 */
 43void rcu_cblist_flush_enqueue(struct rcu_cblist *drclp,
 44			      struct rcu_cblist *srclp,
 45			      struct rcu_head *rhp)
 46{
 47	drclp->head = srclp->head;
 48	if (drclp->head)
 49		drclp->tail = srclp->tail;
 50	else
 51		drclp->tail = &drclp->head;
 52	drclp->len = srclp->len;
 53	if (!rhp) {
 54		rcu_cblist_init(srclp);
 55	} else {
 56		rhp->next = NULL;
 57		srclp->head = rhp;
 58		srclp->tail = &rhp->next;
 59		WRITE_ONCE(srclp->len, 1);
 60	}
 61}
 62
 63/*
 64 * Dequeue the oldest rcu_head structure from the specified callback
 65 * list.
 66 */
 67struct rcu_head *rcu_cblist_dequeue(struct rcu_cblist *rclp)
 68{
 69	struct rcu_head *rhp;
 70
 71	rhp = rclp->head;
 72	if (!rhp)
 73		return NULL;
 74	rclp->len--;
 75	rclp->head = rhp->next;
 76	if (!rclp->head)
 77		rclp->tail = &rclp->head;
 78	return rhp;
 79}
 80
 81/* Set the length of an rcu_segcblist structure. */
 82static void rcu_segcblist_set_len(struct rcu_segcblist *rsclp, long v)
 83{
 84#ifdef CONFIG_RCU_NOCB_CPU
 85	atomic_long_set(&rsclp->len, v);
 86#else
 87	WRITE_ONCE(rsclp->len, v);
 88#endif
 89}
 90
 91/* Get the length of a segment of the rcu_segcblist structure. */
 92long rcu_segcblist_get_seglen(struct rcu_segcblist *rsclp, int seg)
 93{
 94	return READ_ONCE(rsclp->seglen[seg]);
 95}
 96
 97/* Return number of callbacks in segmented callback list by summing seglen. */
 98long rcu_segcblist_n_segment_cbs(struct rcu_segcblist *rsclp)
 99{
100	long len = 0;
101	int i;
102
103	for (i = RCU_DONE_TAIL; i < RCU_CBLIST_NSEGS; i++)
104		len += rcu_segcblist_get_seglen(rsclp, i);
105
106	return len;
107}
108
109/* Set the length of a segment of the rcu_segcblist structure. */
110static void rcu_segcblist_set_seglen(struct rcu_segcblist *rsclp, int seg, long v)
111{
112	WRITE_ONCE(rsclp->seglen[seg], v);
113}
114
115/* Increase the numeric length of a segment by a specified amount. */
116static void rcu_segcblist_add_seglen(struct rcu_segcblist *rsclp, int seg, long v)
117{
118	WRITE_ONCE(rsclp->seglen[seg], rsclp->seglen[seg] + v);
119}
120
121/* Move from's segment length to to's segment. */
122static void rcu_segcblist_move_seglen(struct rcu_segcblist *rsclp, int from, int to)
123{
124	long len;
125
126	if (from == to)
127		return;
128
129	len = rcu_segcblist_get_seglen(rsclp, from);
130	if (!len)
131		return;
132
133	rcu_segcblist_add_seglen(rsclp, to, len);
134	rcu_segcblist_set_seglen(rsclp, from, 0);
135}
136
137/* Increment segment's length. */
138static void rcu_segcblist_inc_seglen(struct rcu_segcblist *rsclp, int seg)
139{
140	rcu_segcblist_add_seglen(rsclp, seg, 1);
141}
142
143/*
144 * Increase the numeric length of an rcu_segcblist structure by the
145 * specified amount, which can be negative.  This can cause the ->len
146 * field to disagree with the actual number of callbacks on the structure.
147 * This increase is fully ordered with respect to the callers accesses
148 * both before and after.
149 *
150 * So why on earth is a memory barrier required both before and after
151 * the update to the ->len field???
152 *
153 * The reason is that rcu_barrier() locklessly samples each CPU's ->len
154 * field, and if a given CPU's field is zero, avoids IPIing that CPU.
155 * This can of course race with both queuing and invoking of callbacks.
156 * Failing to correctly handle either of these races could result in
157 * rcu_barrier() failing to IPI a CPU that actually had callbacks queued
158 * which rcu_barrier() was obligated to wait on.  And if rcu_barrier()
159 * failed to wait on such a callback, unloading certain kernel modules
160 * would result in calls to functions whose code was no longer present in
161 * the kernel, for but one example.
162 *
163 * Therefore, ->len transitions from 1->0 and 0->1 have to be carefully
164 * ordered with respect with both list modifications and the rcu_barrier().
165 *
166 * The queuing case is CASE 1 and the invoking case is CASE 2.
167 *
168 * CASE 1: Suppose that CPU 0 has no callbacks queued, but invokes
169 * call_rcu() just as CPU 1 invokes rcu_barrier().  CPU 0's ->len field
170 * will transition from 0->1, which is one of the transitions that must
171 * be handled carefully.  Without the full memory barriers after the ->len
172 * update and at the beginning of rcu_barrier(), the following could happen:
173 *
174 * CPU 0				CPU 1
175 *
176 * call_rcu().
177 *					rcu_barrier() sees ->len as 0.
178 * set ->len = 1.
179 *					rcu_barrier() does nothing.
180 *					module is unloaded.
181 * callback invokes unloaded function!
182 *
183 * With the full barriers, any case where rcu_barrier() sees ->len as 0 will
184 * have unambiguously preceded the return from the racing call_rcu(), which
185 * means that this call_rcu() invocation is OK to not wait on.  After all,
186 * you are supposed to make sure that any problematic call_rcu() invocations
187 * happen before the rcu_barrier().
188 *
189 *
190 * CASE 2: Suppose that CPU 0 is invoking its last callback just as
191 * CPU 1 invokes rcu_barrier().  CPU 0's ->len field will transition from
192 * 1->0, which is one of the transitions that must be handled carefully.
193 * Without the full memory barriers before the ->len update and at the
194 * end of rcu_barrier(), the following could happen:
195 *
196 * CPU 0				CPU 1
197 *
198 * start invoking last callback
199 * set ->len = 0 (reordered)
200 *					rcu_barrier() sees ->len as 0
201 *					rcu_barrier() does nothing.
202 *					module is unloaded
203 * callback executing after unloaded!
204 *
205 * With the full barriers, any case where rcu_barrier() sees ->len as 0
206 * will be fully ordered after the completion of the callback function,
207 * so that the module unloading operation is completely safe.
208 *
209 */
210void rcu_segcblist_add_len(struct rcu_segcblist *rsclp, long v)
211{
212#ifdef CONFIG_RCU_NOCB_CPU
213	smp_mb__before_atomic(); // Read header comment above.
214	atomic_long_add(v, &rsclp->len);
215	smp_mb__after_atomic();  // Read header comment above.
216#else
217	smp_mb(); // Read header comment above.
218	WRITE_ONCE(rsclp->len, rsclp->len + v);
219	smp_mb(); // Read header comment above.
220#endif
221}
222
223/*
224 * Increase the numeric length of an rcu_segcblist structure by one.
225 * This can cause the ->len field to disagree with the actual number of
226 * callbacks on the structure.  This increase is fully ordered with respect
227 * to the callers accesses both before and after.
228 */
229void rcu_segcblist_inc_len(struct rcu_segcblist *rsclp)
230{
231	rcu_segcblist_add_len(rsclp, 1);
232}
233
234/*
235 * Initialize an rcu_segcblist structure.
236 */
237void rcu_segcblist_init(struct rcu_segcblist *rsclp)
238{
239	int i;
240
241	BUILD_BUG_ON(RCU_NEXT_TAIL + 1 != ARRAY_SIZE(rsclp->gp_seq));
242	BUILD_BUG_ON(ARRAY_SIZE(rsclp->tails) != ARRAY_SIZE(rsclp->gp_seq));
243	rsclp->head = NULL;
244	for (i = 0; i < RCU_CBLIST_NSEGS; i++) {
245		rsclp->tails[i] = &rsclp->head;
246		rcu_segcblist_set_seglen(rsclp, i, 0);
247	}
248	rcu_segcblist_set_len(rsclp, 0);
249	rcu_segcblist_set_flags(rsclp, SEGCBLIST_ENABLED);
250}
251
252/*
253 * Disable the specified rcu_segcblist structure, so that callbacks can
254 * no longer be posted to it.  This structure must be empty.
255 */
256void rcu_segcblist_disable(struct rcu_segcblist *rsclp)
257{
258	WARN_ON_ONCE(!rcu_segcblist_empty(rsclp));
259	WARN_ON_ONCE(rcu_segcblist_n_cbs(rsclp));
260	rcu_segcblist_clear_flags(rsclp, SEGCBLIST_ENABLED);
261}
262
263/*
264 * Mark the specified rcu_segcblist structure as offloaded (or not)
265 */
266void rcu_segcblist_offload(struct rcu_segcblist *rsclp, bool offload)
267{
268	if (offload)
269		rcu_segcblist_set_flags(rsclp, SEGCBLIST_LOCKING | SEGCBLIST_OFFLOADED);
270	else
 
271		rcu_segcblist_clear_flags(rsclp, SEGCBLIST_OFFLOADED);
 
272}
273
274/*
275 * Does the specified rcu_segcblist structure contain callbacks that
276 * are ready to be invoked?
277 */
278bool rcu_segcblist_ready_cbs(struct rcu_segcblist *rsclp)
279{
280	return rcu_segcblist_is_enabled(rsclp) &&
281	       &rsclp->head != READ_ONCE(rsclp->tails[RCU_DONE_TAIL]);
282}
283
284/*
285 * Does the specified rcu_segcblist structure contain callbacks that
286 * are still pending, that is, not yet ready to be invoked?
287 */
288bool rcu_segcblist_pend_cbs(struct rcu_segcblist *rsclp)
289{
290	return rcu_segcblist_is_enabled(rsclp) &&
291	       !rcu_segcblist_restempty(rsclp, RCU_DONE_TAIL);
292}
293
294/*
295 * Return a pointer to the first callback in the specified rcu_segcblist
296 * structure.  This is useful for diagnostics.
297 */
298struct rcu_head *rcu_segcblist_first_cb(struct rcu_segcblist *rsclp)
299{
300	if (rcu_segcblist_is_enabled(rsclp))
301		return rsclp->head;
302	return NULL;
303}
304
305/*
306 * Return a pointer to the first pending callback in the specified
307 * rcu_segcblist structure.  This is useful just after posting a given
308 * callback -- if that callback is the first pending callback, then
309 * you cannot rely on someone else having already started up the required
310 * grace period.
311 */
312struct rcu_head *rcu_segcblist_first_pend_cb(struct rcu_segcblist *rsclp)
313{
314	if (rcu_segcblist_is_enabled(rsclp))
315		return *rsclp->tails[RCU_DONE_TAIL];
316	return NULL;
317}
318
319/*
320 * Return false if there are no CBs awaiting grace periods, otherwise,
321 * return true and store the nearest waited-upon grace period into *lp.
322 */
323bool rcu_segcblist_nextgp(struct rcu_segcblist *rsclp, unsigned long *lp)
324{
325	if (!rcu_segcblist_pend_cbs(rsclp))
326		return false;
327	*lp = rsclp->gp_seq[RCU_WAIT_TAIL];
328	return true;
329}
330
331/*
332 * Enqueue the specified callback onto the specified rcu_segcblist
333 * structure, updating accounting as needed.  Note that the ->len
334 * field may be accessed locklessly, hence the WRITE_ONCE().
335 * The ->len field is used by rcu_barrier() and friends to determine
336 * if it must post a callback on this structure, and it is OK
337 * for rcu_barrier() to sometimes post callbacks needlessly, but
338 * absolutely not OK for it to ever miss posting a callback.
339 */
340void rcu_segcblist_enqueue(struct rcu_segcblist *rsclp,
341			   struct rcu_head *rhp)
342{
343	rcu_segcblist_inc_len(rsclp);
344	rcu_segcblist_inc_seglen(rsclp, RCU_NEXT_TAIL);
345	rhp->next = NULL;
346	WRITE_ONCE(*rsclp->tails[RCU_NEXT_TAIL], rhp);
347	WRITE_ONCE(rsclp->tails[RCU_NEXT_TAIL], &rhp->next);
348}
349
350/*
351 * Entrain the specified callback onto the specified rcu_segcblist at
352 * the end of the last non-empty segment.  If the entire rcu_segcblist
353 * is empty, make no change, but return false.
354 *
355 * This is intended for use by rcu_barrier()-like primitives, -not-
356 * for normal grace-period use.  IMPORTANT:  The callback you enqueue
357 * will wait for all prior callbacks, NOT necessarily for a grace
358 * period.  You have been warned.
359 */
360bool rcu_segcblist_entrain(struct rcu_segcblist *rsclp,
361			   struct rcu_head *rhp)
362{
363	int i;
364
365	if (rcu_segcblist_n_cbs(rsclp) == 0)
366		return false;
367	rcu_segcblist_inc_len(rsclp);
368	smp_mb(); /* Ensure counts are updated before callback is entrained. */
369	rhp->next = NULL;
370	for (i = RCU_NEXT_TAIL; i > RCU_DONE_TAIL; i--)
371		if (!rcu_segcblist_segempty(rsclp, i))
372			break;
373	rcu_segcblist_inc_seglen(rsclp, i);
374	WRITE_ONCE(*rsclp->tails[i], rhp);
375	for (; i <= RCU_NEXT_TAIL; i++)
376		WRITE_ONCE(rsclp->tails[i], &rhp->next);
377	return true;
378}
379
380/*
381 * Extract only those callbacks ready to be invoked from the specified
382 * rcu_segcblist structure and place them in the specified rcu_cblist
383 * structure.
384 */
385void rcu_segcblist_extract_done_cbs(struct rcu_segcblist *rsclp,
386				    struct rcu_cblist *rclp)
387{
388	int i;
389
390	if (!rcu_segcblist_ready_cbs(rsclp))
391		return; /* Nothing to do. */
392	rclp->len = rcu_segcblist_get_seglen(rsclp, RCU_DONE_TAIL);
393	*rclp->tail = rsclp->head;
394	WRITE_ONCE(rsclp->head, *rsclp->tails[RCU_DONE_TAIL]);
395	WRITE_ONCE(*rsclp->tails[RCU_DONE_TAIL], NULL);
396	rclp->tail = rsclp->tails[RCU_DONE_TAIL];
397	for (i = RCU_CBLIST_NSEGS - 1; i >= RCU_DONE_TAIL; i--)
398		if (rsclp->tails[i] == rsclp->tails[RCU_DONE_TAIL])
399			WRITE_ONCE(rsclp->tails[i], &rsclp->head);
400	rcu_segcblist_set_seglen(rsclp, RCU_DONE_TAIL, 0);
401}
402
403/*
404 * Extract only those callbacks still pending (not yet ready to be
405 * invoked) from the specified rcu_segcblist structure and place them in
406 * the specified rcu_cblist structure.  Note that this loses information
407 * about any callbacks that might have been partway done waiting for
408 * their grace period.  Too bad!  They will have to start over.
409 */
410void rcu_segcblist_extract_pend_cbs(struct rcu_segcblist *rsclp,
411				    struct rcu_cblist *rclp)
412{
413	int i;
414
415	if (!rcu_segcblist_pend_cbs(rsclp))
416		return; /* Nothing to do. */
417	rclp->len = 0;
418	*rclp->tail = *rsclp->tails[RCU_DONE_TAIL];
419	rclp->tail = rsclp->tails[RCU_NEXT_TAIL];
420	WRITE_ONCE(*rsclp->tails[RCU_DONE_TAIL], NULL);
421	for (i = RCU_DONE_TAIL + 1; i < RCU_CBLIST_NSEGS; i++) {
422		rclp->len += rcu_segcblist_get_seglen(rsclp, i);
423		WRITE_ONCE(rsclp->tails[i], rsclp->tails[RCU_DONE_TAIL]);
424		rcu_segcblist_set_seglen(rsclp, i, 0);
425	}
426}
427
428/*
429 * Insert counts from the specified rcu_cblist structure in the
430 * specified rcu_segcblist structure.
431 */
432void rcu_segcblist_insert_count(struct rcu_segcblist *rsclp,
433				struct rcu_cblist *rclp)
434{
435	rcu_segcblist_add_len(rsclp, rclp->len);
436}
437
438/*
439 * Move callbacks from the specified rcu_cblist to the beginning of the
440 * done-callbacks segment of the specified rcu_segcblist.
441 */
442void rcu_segcblist_insert_done_cbs(struct rcu_segcblist *rsclp,
443				   struct rcu_cblist *rclp)
444{
445	int i;
446
447	if (!rclp->head)
448		return; /* No callbacks to move. */
449	rcu_segcblist_add_seglen(rsclp, RCU_DONE_TAIL, rclp->len);
450	*rclp->tail = rsclp->head;
451	WRITE_ONCE(rsclp->head, rclp->head);
452	for (i = RCU_DONE_TAIL; i < RCU_CBLIST_NSEGS; i++)
453		if (&rsclp->head == rsclp->tails[i])
454			WRITE_ONCE(rsclp->tails[i], rclp->tail);
455		else
456			break;
457	rclp->head = NULL;
458	rclp->tail = &rclp->head;
459}
460
461/*
462 * Move callbacks from the specified rcu_cblist to the end of the
463 * new-callbacks segment of the specified rcu_segcblist.
464 */
465void rcu_segcblist_insert_pend_cbs(struct rcu_segcblist *rsclp,
466				   struct rcu_cblist *rclp)
467{
468	if (!rclp->head)
469		return; /* Nothing to do. */
470
471	rcu_segcblist_add_seglen(rsclp, RCU_NEXT_TAIL, rclp->len);
472	WRITE_ONCE(*rsclp->tails[RCU_NEXT_TAIL], rclp->head);
473	WRITE_ONCE(rsclp->tails[RCU_NEXT_TAIL], rclp->tail);
474}
475
476/*
477 * Advance the callbacks in the specified rcu_segcblist structure based
478 * on the current value passed in for the grace-period counter.
479 */
480void rcu_segcblist_advance(struct rcu_segcblist *rsclp, unsigned long seq)
481{
482	int i, j;
483
484	WARN_ON_ONCE(!rcu_segcblist_is_enabled(rsclp));
485	if (rcu_segcblist_restempty(rsclp, RCU_DONE_TAIL))
486		return;
487
488	/*
489	 * Find all callbacks whose ->gp_seq numbers indicate that they
490	 * are ready to invoke, and put them into the RCU_DONE_TAIL segment.
491	 */
492	for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
493		if (ULONG_CMP_LT(seq, rsclp->gp_seq[i]))
494			break;
495		WRITE_ONCE(rsclp->tails[RCU_DONE_TAIL], rsclp->tails[i]);
496		rcu_segcblist_move_seglen(rsclp, i, RCU_DONE_TAIL);
497	}
498
499	/* If no callbacks moved, nothing more need be done. */
500	if (i == RCU_WAIT_TAIL)
501		return;
502
503	/* Clean up tail pointers that might have been misordered above. */
504	for (j = RCU_WAIT_TAIL; j < i; j++)
505		WRITE_ONCE(rsclp->tails[j], rsclp->tails[RCU_DONE_TAIL]);
506
507	/*
508	 * Callbacks moved, so there might be an empty RCU_WAIT_TAIL
509	 * and a non-empty RCU_NEXT_READY_TAIL.  If so, copy the
510	 * RCU_NEXT_READY_TAIL segment to fill the RCU_WAIT_TAIL gap
511	 * created by the now-ready-to-invoke segments.
512	 */
513	for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
514		if (rsclp->tails[j] == rsclp->tails[RCU_NEXT_TAIL])
515			break;  /* No more callbacks. */
516		WRITE_ONCE(rsclp->tails[j], rsclp->tails[i]);
517		rcu_segcblist_move_seglen(rsclp, i, j);
518		rsclp->gp_seq[j] = rsclp->gp_seq[i];
519	}
520}
521
522/*
523 * "Accelerate" callbacks based on more-accurate grace-period information.
524 * The reason for this is that RCU does not synchronize the beginnings and
525 * ends of grace periods, and that callbacks are posted locally.  This in
526 * turn means that the callbacks must be labelled conservatively early
527 * on, as getting exact information would degrade both performance and
528 * scalability.  When more accurate grace-period information becomes
529 * available, previously posted callbacks can be "accelerated", marking
530 * them to complete at the end of the earlier grace period.
531 *
532 * This function operates on an rcu_segcblist structure, and also the
533 * grace-period sequence number seq at which new callbacks would become
534 * ready to invoke.  Returns true if there are callbacks that won't be
535 * ready to invoke until seq, false otherwise.
536 */
537bool rcu_segcblist_accelerate(struct rcu_segcblist *rsclp, unsigned long seq)
538{
539	int i, j;
540
541	WARN_ON_ONCE(!rcu_segcblist_is_enabled(rsclp));
542	if (rcu_segcblist_restempty(rsclp, RCU_DONE_TAIL))
543		return false;
544
545	/*
546	 * Find the segment preceding the oldest segment of callbacks
547	 * whose ->gp_seq[] completion is at or after that passed in via
548	 * "seq", skipping any empty segments.  This oldest segment, along
549	 * with any later segments, can be merged in with any newly arrived
550	 * callbacks in the RCU_NEXT_TAIL segment, and assigned "seq"
551	 * as their ->gp_seq[] grace-period completion sequence number.
552	 */
553	for (i = RCU_NEXT_READY_TAIL; i > RCU_DONE_TAIL; i--)
554		if (!rcu_segcblist_segempty(rsclp, i) &&
555		    ULONG_CMP_LT(rsclp->gp_seq[i], seq))
556			break;
557
558	/*
559	 * If all the segments contain callbacks that correspond to
560	 * earlier grace-period sequence numbers than "seq", leave.
561	 * Assuming that the rcu_segcblist structure has enough
562	 * segments in its arrays, this can only happen if some of
563	 * the non-done segments contain callbacks that really are
564	 * ready to invoke.  This situation will get straightened
565	 * out by the next call to rcu_segcblist_advance().
566	 *
567	 * Also advance to the oldest segment of callbacks whose
568	 * ->gp_seq[] completion is at or after that passed in via "seq",
569	 * skipping any empty segments.
570	 *
571	 * Note that segment "i" (and any lower-numbered segments
572	 * containing older callbacks) will be unaffected, and their
573	 * grace-period numbers remain unchanged.  For example, if i ==
574	 * WAIT_TAIL, then neither WAIT_TAIL nor DONE_TAIL will be touched.
575	 * Instead, the CBs in NEXT_TAIL will be merged with those in
576	 * NEXT_READY_TAIL and the grace-period number of NEXT_READY_TAIL
577	 * would be updated.  NEXT_TAIL would then be empty.
578	 */
579	if (rcu_segcblist_restempty(rsclp, i) || ++i >= RCU_NEXT_TAIL)
580		return false;
581
582	/* Accounting: everything below i is about to get merged into i. */
583	for (j = i + 1; j <= RCU_NEXT_TAIL; j++)
584		rcu_segcblist_move_seglen(rsclp, j, i);
585
586	/*
587	 * Merge all later callbacks, including newly arrived callbacks,
588	 * into the segment located by the for-loop above.  Assign "seq"
589	 * as the ->gp_seq[] value in order to correctly handle the case
590	 * where there were no pending callbacks in the rcu_segcblist
591	 * structure other than in the RCU_NEXT_TAIL segment.
592	 */
593	for (; i < RCU_NEXT_TAIL; i++) {
594		WRITE_ONCE(rsclp->tails[i], rsclp->tails[RCU_NEXT_TAIL]);
595		rsclp->gp_seq[i] = seq;
596	}
597	return true;
598}
599
600/*
601 * Merge the source rcu_segcblist structure into the destination
602 * rcu_segcblist structure, then initialize the source.  Any pending
603 * callbacks from the source get to start over.  It is best to
604 * advance and accelerate both the destination and the source
605 * before merging.
606 */
607void rcu_segcblist_merge(struct rcu_segcblist *dst_rsclp,
608			 struct rcu_segcblist *src_rsclp)
609{
610	struct rcu_cblist donecbs;
611	struct rcu_cblist pendcbs;
612
613	lockdep_assert_cpus_held();
614
615	rcu_cblist_init(&donecbs);
616	rcu_cblist_init(&pendcbs);
617
618	rcu_segcblist_extract_done_cbs(src_rsclp, &donecbs);
619	rcu_segcblist_extract_pend_cbs(src_rsclp, &pendcbs);
620
621	/*
622	 * No need smp_mb() before setting length to 0, because CPU hotplug
623	 * lock excludes rcu_barrier.
624	 */
625	rcu_segcblist_set_len(src_rsclp, 0);
626
627	rcu_segcblist_insert_count(dst_rsclp, &donecbs);
628	rcu_segcblist_insert_count(dst_rsclp, &pendcbs);
629	rcu_segcblist_insert_done_cbs(dst_rsclp, &donecbs);
630	rcu_segcblist_insert_pend_cbs(dst_rsclp, &pendcbs);
631
632	rcu_segcblist_init(src_rsclp);
633}
v5.14.15
  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/cpu.h>
 11#include <linux/interrupt.h>
 12#include <linux/kernel.h>
 13#include <linux/types.h>
 14
 15#include "rcu_segcblist.h"
 16
 17/* Initialize simple callback list. */
 18void rcu_cblist_init(struct rcu_cblist *rclp)
 19{
 20	rclp->head = NULL;
 21	rclp->tail = &rclp->head;
 22	rclp->len = 0;
 23}
 24
 25/*
 26 * Enqueue an rcu_head structure onto the specified callback list.
 27 */
 28void rcu_cblist_enqueue(struct rcu_cblist *rclp, struct rcu_head *rhp)
 29{
 30	*rclp->tail = rhp;
 31	rclp->tail = &rhp->next;
 32	WRITE_ONCE(rclp->len, rclp->len + 1);
 33}
 34
 35/*
 36 * Flush the second rcu_cblist structure onto the first one, obliterating
 37 * any contents of the first.  If rhp is non-NULL, enqueue it as the sole
 38 * element of the second rcu_cblist structure, but ensuring that the second
 39 * rcu_cblist structure, if initially non-empty, always appears non-empty
 40 * throughout the process.  If rdp is NULL, the second rcu_cblist structure
 41 * is instead initialized to empty.
 42 */
 43void rcu_cblist_flush_enqueue(struct rcu_cblist *drclp,
 44			      struct rcu_cblist *srclp,
 45			      struct rcu_head *rhp)
 46{
 47	drclp->head = srclp->head;
 48	if (drclp->head)
 49		drclp->tail = srclp->tail;
 50	else
 51		drclp->tail = &drclp->head;
 52	drclp->len = srclp->len;
 53	if (!rhp) {
 54		rcu_cblist_init(srclp);
 55	} else {
 56		rhp->next = NULL;
 57		srclp->head = rhp;
 58		srclp->tail = &rhp->next;
 59		WRITE_ONCE(srclp->len, 1);
 60	}
 61}
 62
 63/*
 64 * Dequeue the oldest rcu_head structure from the specified callback
 65 * list.
 66 */
 67struct rcu_head *rcu_cblist_dequeue(struct rcu_cblist *rclp)
 68{
 69	struct rcu_head *rhp;
 70
 71	rhp = rclp->head;
 72	if (!rhp)
 73		return NULL;
 74	rclp->len--;
 75	rclp->head = rhp->next;
 76	if (!rclp->head)
 77		rclp->tail = &rclp->head;
 78	return rhp;
 79}
 80
 81/* Set the length of an rcu_segcblist structure. */
 82static void rcu_segcblist_set_len(struct rcu_segcblist *rsclp, long v)
 83{
 84#ifdef CONFIG_RCU_NOCB_CPU
 85	atomic_long_set(&rsclp->len, v);
 86#else
 87	WRITE_ONCE(rsclp->len, v);
 88#endif
 89}
 90
 91/* Get the length of a segment of the rcu_segcblist structure. */
 92static long rcu_segcblist_get_seglen(struct rcu_segcblist *rsclp, int seg)
 93{
 94	return READ_ONCE(rsclp->seglen[seg]);
 95}
 96
 97/* Return number of callbacks in segmented callback list by summing seglen. */
 98long rcu_segcblist_n_segment_cbs(struct rcu_segcblist *rsclp)
 99{
100	long len = 0;
101	int i;
102
103	for (i = RCU_DONE_TAIL; i < RCU_CBLIST_NSEGS; i++)
104		len += rcu_segcblist_get_seglen(rsclp, i);
105
106	return len;
107}
108
109/* Set the length of a segment of the rcu_segcblist structure. */
110static void rcu_segcblist_set_seglen(struct rcu_segcblist *rsclp, int seg, long v)
111{
112	WRITE_ONCE(rsclp->seglen[seg], v);
113}
114
115/* Increase the numeric length of a segment by a specified amount. */
116static void rcu_segcblist_add_seglen(struct rcu_segcblist *rsclp, int seg, long v)
117{
118	WRITE_ONCE(rsclp->seglen[seg], rsclp->seglen[seg] + v);
119}
120
121/* Move from's segment length to to's segment. */
122static void rcu_segcblist_move_seglen(struct rcu_segcblist *rsclp, int from, int to)
123{
124	long len;
125
126	if (from == to)
127		return;
128
129	len = rcu_segcblist_get_seglen(rsclp, from);
130	if (!len)
131		return;
132
133	rcu_segcblist_add_seglen(rsclp, to, len);
134	rcu_segcblist_set_seglen(rsclp, from, 0);
135}
136
137/* Increment segment's length. */
138static void rcu_segcblist_inc_seglen(struct rcu_segcblist *rsclp, int seg)
139{
140	rcu_segcblist_add_seglen(rsclp, seg, 1);
141}
142
143/*
144 * Increase the numeric length of an rcu_segcblist structure by the
145 * specified amount, which can be negative.  This can cause the ->len
146 * field to disagree with the actual number of callbacks on the structure.
147 * This increase is fully ordered with respect to the callers accesses
148 * both before and after.
149 *
150 * So why on earth is a memory barrier required both before and after
151 * the update to the ->len field???
152 *
153 * The reason is that rcu_barrier() locklessly samples each CPU's ->len
154 * field, and if a given CPU's field is zero, avoids IPIing that CPU.
155 * This can of course race with both queuing and invoking of callbacks.
156 * Failing to correctly handle either of these races could result in
157 * rcu_barrier() failing to IPI a CPU that actually had callbacks queued
158 * which rcu_barrier() was obligated to wait on.  And if rcu_barrier()
159 * failed to wait on such a callback, unloading certain kernel modules
160 * would result in calls to functions whose code was no longer present in
161 * the kernel, for but one example.
162 *
163 * Therefore, ->len transitions from 1->0 and 0->1 have to be carefully
164 * ordered with respect with both list modifications and the rcu_barrier().
165 *
166 * The queuing case is CASE 1 and the invoking case is CASE 2.
167 *
168 * CASE 1: Suppose that CPU 0 has no callbacks queued, but invokes
169 * call_rcu() just as CPU 1 invokes rcu_barrier().  CPU 0's ->len field
170 * will transition from 0->1, which is one of the transitions that must
171 * be handled carefully.  Without the full memory barriers after the ->len
172 * update and at the beginning of rcu_barrier(), the following could happen:
173 *
174 * CPU 0				CPU 1
175 *
176 * call_rcu().
177 *					rcu_barrier() sees ->len as 0.
178 * set ->len = 1.
179 *					rcu_barrier() does nothing.
180 *					module is unloaded.
181 * callback invokes unloaded function!
182 *
183 * With the full barriers, any case where rcu_barrier() sees ->len as 0 will
184 * have unambiguously preceded the return from the racing call_rcu(), which
185 * means that this call_rcu() invocation is OK to not wait on.  After all,
186 * you are supposed to make sure that any problematic call_rcu() invocations
187 * happen before the rcu_barrier().
188 *
189 *
190 * CASE 2: Suppose that CPU 0 is invoking its last callback just as
191 * CPU 1 invokes rcu_barrier().  CPU 0's ->len field will transition from
192 * 1->0, which is one of the transitions that must be handled carefully.
193 * Without the full memory barriers before the ->len update and at the
194 * end of rcu_barrier(), the following could happen:
195 *
196 * CPU 0				CPU 1
197 *
198 * start invoking last callback
199 * set ->len = 0 (reordered)
200 *					rcu_barrier() sees ->len as 0
201 *					rcu_barrier() does nothing.
202 *					module is unloaded
203 * callback executing after unloaded!
204 *
205 * With the full barriers, any case where rcu_barrier() sees ->len as 0
206 * will be fully ordered after the completion of the callback function,
207 * so that the module unloading operation is completely safe.
208 *
209 */
210void rcu_segcblist_add_len(struct rcu_segcblist *rsclp, long v)
211{
212#ifdef CONFIG_RCU_NOCB_CPU
213	smp_mb__before_atomic(); // Read header comment above.
214	atomic_long_add(v, &rsclp->len);
215	smp_mb__after_atomic();  // Read header comment above.
216#else
217	smp_mb(); // Read header comment above.
218	WRITE_ONCE(rsclp->len, rsclp->len + v);
219	smp_mb(); // Read header comment above.
220#endif
221}
222
223/*
224 * Increase the numeric length of an rcu_segcblist structure by one.
225 * This can cause the ->len field to disagree with the actual number of
226 * callbacks on the structure.  This increase is fully ordered with respect
227 * to the callers accesses both before and after.
228 */
229void rcu_segcblist_inc_len(struct rcu_segcblist *rsclp)
230{
231	rcu_segcblist_add_len(rsclp, 1);
232}
233
234/*
235 * Initialize an rcu_segcblist structure.
236 */
237void rcu_segcblist_init(struct rcu_segcblist *rsclp)
238{
239	int i;
240
241	BUILD_BUG_ON(RCU_NEXT_TAIL + 1 != ARRAY_SIZE(rsclp->gp_seq));
242	BUILD_BUG_ON(ARRAY_SIZE(rsclp->tails) != ARRAY_SIZE(rsclp->gp_seq));
243	rsclp->head = NULL;
244	for (i = 0; i < RCU_CBLIST_NSEGS; i++) {
245		rsclp->tails[i] = &rsclp->head;
246		rcu_segcblist_set_seglen(rsclp, i, 0);
247	}
248	rcu_segcblist_set_len(rsclp, 0);
249	rcu_segcblist_set_flags(rsclp, SEGCBLIST_ENABLED);
250}
251
252/*
253 * Disable the specified rcu_segcblist structure, so that callbacks can
254 * no longer be posted to it.  This structure must be empty.
255 */
256void rcu_segcblist_disable(struct rcu_segcblist *rsclp)
257{
258	WARN_ON_ONCE(!rcu_segcblist_empty(rsclp));
259	WARN_ON_ONCE(rcu_segcblist_n_cbs(rsclp));
260	rcu_segcblist_clear_flags(rsclp, SEGCBLIST_ENABLED);
261}
262
263/*
264 * Mark the specified rcu_segcblist structure as offloaded.
265 */
266void rcu_segcblist_offload(struct rcu_segcblist *rsclp, bool offload)
267{
268	if (offload) {
269		rcu_segcblist_clear_flags(rsclp, SEGCBLIST_SOFTIRQ_ONLY);
270		rcu_segcblist_set_flags(rsclp, SEGCBLIST_OFFLOADED);
271	} else {
272		rcu_segcblist_clear_flags(rsclp, SEGCBLIST_OFFLOADED);
273	}
274}
275
276/*
277 * Does the specified rcu_segcblist structure contain callbacks that
278 * are ready to be invoked?
279 */
280bool rcu_segcblist_ready_cbs(struct rcu_segcblist *rsclp)
281{
282	return rcu_segcblist_is_enabled(rsclp) &&
283	       &rsclp->head != READ_ONCE(rsclp->tails[RCU_DONE_TAIL]);
284}
285
286/*
287 * Does the specified rcu_segcblist structure contain callbacks that
288 * are still pending, that is, not yet ready to be invoked?
289 */
290bool rcu_segcblist_pend_cbs(struct rcu_segcblist *rsclp)
291{
292	return rcu_segcblist_is_enabled(rsclp) &&
293	       !rcu_segcblist_restempty(rsclp, RCU_DONE_TAIL);
294}
295
296/*
297 * Return a pointer to the first callback in the specified rcu_segcblist
298 * structure.  This is useful for diagnostics.
299 */
300struct rcu_head *rcu_segcblist_first_cb(struct rcu_segcblist *rsclp)
301{
302	if (rcu_segcblist_is_enabled(rsclp))
303		return rsclp->head;
304	return NULL;
305}
306
307/*
308 * Return a pointer to the first pending callback in the specified
309 * rcu_segcblist structure.  This is useful just after posting a given
310 * callback -- if that callback is the first pending callback, then
311 * you cannot rely on someone else having already started up the required
312 * grace period.
313 */
314struct rcu_head *rcu_segcblist_first_pend_cb(struct rcu_segcblist *rsclp)
315{
316	if (rcu_segcblist_is_enabled(rsclp))
317		return *rsclp->tails[RCU_DONE_TAIL];
318	return NULL;
319}
320
321/*
322 * Return false if there are no CBs awaiting grace periods, otherwise,
323 * return true and store the nearest waited-upon grace period into *lp.
324 */
325bool rcu_segcblist_nextgp(struct rcu_segcblist *rsclp, unsigned long *lp)
326{
327	if (!rcu_segcblist_pend_cbs(rsclp))
328		return false;
329	*lp = rsclp->gp_seq[RCU_WAIT_TAIL];
330	return true;
331}
332
333/*
334 * Enqueue the specified callback onto the specified rcu_segcblist
335 * structure, updating accounting as needed.  Note that the ->len
336 * field may be accessed locklessly, hence the WRITE_ONCE().
337 * The ->len field is used by rcu_barrier() and friends to determine
338 * if it must post a callback on this structure, and it is OK
339 * for rcu_barrier() to sometimes post callbacks needlessly, but
340 * absolutely not OK for it to ever miss posting a callback.
341 */
342void rcu_segcblist_enqueue(struct rcu_segcblist *rsclp,
343			   struct rcu_head *rhp)
344{
345	rcu_segcblist_inc_len(rsclp);
346	rcu_segcblist_inc_seglen(rsclp, RCU_NEXT_TAIL);
347	rhp->next = NULL;
348	WRITE_ONCE(*rsclp->tails[RCU_NEXT_TAIL], rhp);
349	WRITE_ONCE(rsclp->tails[RCU_NEXT_TAIL], &rhp->next);
350}
351
352/*
353 * Entrain the specified callback onto the specified rcu_segcblist at
354 * the end of the last non-empty segment.  If the entire rcu_segcblist
355 * is empty, make no change, but return false.
356 *
357 * This is intended for use by rcu_barrier()-like primitives, -not-
358 * for normal grace-period use.  IMPORTANT:  The callback you enqueue
359 * will wait for all prior callbacks, NOT necessarily for a grace
360 * period.  You have been warned.
361 */
362bool rcu_segcblist_entrain(struct rcu_segcblist *rsclp,
363			   struct rcu_head *rhp)
364{
365	int i;
366
367	if (rcu_segcblist_n_cbs(rsclp) == 0)
368		return false;
369	rcu_segcblist_inc_len(rsclp);
370	smp_mb(); /* Ensure counts are updated before callback is entrained. */
371	rhp->next = NULL;
372	for (i = RCU_NEXT_TAIL; i > RCU_DONE_TAIL; i--)
373		if (rsclp->tails[i] != rsclp->tails[i - 1])
374			break;
375	rcu_segcblist_inc_seglen(rsclp, i);
376	WRITE_ONCE(*rsclp->tails[i], rhp);
377	for (; i <= RCU_NEXT_TAIL; i++)
378		WRITE_ONCE(rsclp->tails[i], &rhp->next);
379	return true;
380}
381
382/*
383 * Extract only those callbacks ready to be invoked from the specified
384 * rcu_segcblist structure and place them in the specified rcu_cblist
385 * structure.
386 */
387void rcu_segcblist_extract_done_cbs(struct rcu_segcblist *rsclp,
388				    struct rcu_cblist *rclp)
389{
390	int i;
391
392	if (!rcu_segcblist_ready_cbs(rsclp))
393		return; /* Nothing to do. */
394	rclp->len = rcu_segcblist_get_seglen(rsclp, RCU_DONE_TAIL);
395	*rclp->tail = rsclp->head;
396	WRITE_ONCE(rsclp->head, *rsclp->tails[RCU_DONE_TAIL]);
397	WRITE_ONCE(*rsclp->tails[RCU_DONE_TAIL], NULL);
398	rclp->tail = rsclp->tails[RCU_DONE_TAIL];
399	for (i = RCU_CBLIST_NSEGS - 1; i >= RCU_DONE_TAIL; i--)
400		if (rsclp->tails[i] == rsclp->tails[RCU_DONE_TAIL])
401			WRITE_ONCE(rsclp->tails[i], &rsclp->head);
402	rcu_segcblist_set_seglen(rsclp, RCU_DONE_TAIL, 0);
403}
404
405/*
406 * Extract only those callbacks still pending (not yet ready to be
407 * invoked) from the specified rcu_segcblist structure and place them in
408 * the specified rcu_cblist structure.  Note that this loses information
409 * about any callbacks that might have been partway done waiting for
410 * their grace period.  Too bad!  They will have to start over.
411 */
412void rcu_segcblist_extract_pend_cbs(struct rcu_segcblist *rsclp,
413				    struct rcu_cblist *rclp)
414{
415	int i;
416
417	if (!rcu_segcblist_pend_cbs(rsclp))
418		return; /* Nothing to do. */
419	rclp->len = 0;
420	*rclp->tail = *rsclp->tails[RCU_DONE_TAIL];
421	rclp->tail = rsclp->tails[RCU_NEXT_TAIL];
422	WRITE_ONCE(*rsclp->tails[RCU_DONE_TAIL], NULL);
423	for (i = RCU_DONE_TAIL + 1; i < RCU_CBLIST_NSEGS; i++) {
424		rclp->len += rcu_segcblist_get_seglen(rsclp, i);
425		WRITE_ONCE(rsclp->tails[i], rsclp->tails[RCU_DONE_TAIL]);
426		rcu_segcblist_set_seglen(rsclp, i, 0);
427	}
428}
429
430/*
431 * Insert counts from the specified rcu_cblist structure in the
432 * specified rcu_segcblist structure.
433 */
434void rcu_segcblist_insert_count(struct rcu_segcblist *rsclp,
435				struct rcu_cblist *rclp)
436{
437	rcu_segcblist_add_len(rsclp, rclp->len);
438}
439
440/*
441 * Move callbacks from the specified rcu_cblist to the beginning of the
442 * done-callbacks segment of the specified rcu_segcblist.
443 */
444void rcu_segcblist_insert_done_cbs(struct rcu_segcblist *rsclp,
445				   struct rcu_cblist *rclp)
446{
447	int i;
448
449	if (!rclp->head)
450		return; /* No callbacks to move. */
451	rcu_segcblist_add_seglen(rsclp, RCU_DONE_TAIL, rclp->len);
452	*rclp->tail = rsclp->head;
453	WRITE_ONCE(rsclp->head, rclp->head);
454	for (i = RCU_DONE_TAIL; i < RCU_CBLIST_NSEGS; i++)
455		if (&rsclp->head == rsclp->tails[i])
456			WRITE_ONCE(rsclp->tails[i], rclp->tail);
457		else
458			break;
459	rclp->head = NULL;
460	rclp->tail = &rclp->head;
461}
462
463/*
464 * Move callbacks from the specified rcu_cblist to the end of the
465 * new-callbacks segment of the specified rcu_segcblist.
466 */
467void rcu_segcblist_insert_pend_cbs(struct rcu_segcblist *rsclp,
468				   struct rcu_cblist *rclp)
469{
470	if (!rclp->head)
471		return; /* Nothing to do. */
472
473	rcu_segcblist_add_seglen(rsclp, RCU_NEXT_TAIL, rclp->len);
474	WRITE_ONCE(*rsclp->tails[RCU_NEXT_TAIL], rclp->head);
475	WRITE_ONCE(rsclp->tails[RCU_NEXT_TAIL], rclp->tail);
476}
477
478/*
479 * Advance the callbacks in the specified rcu_segcblist structure based
480 * on the current value passed in for the grace-period counter.
481 */
482void rcu_segcblist_advance(struct rcu_segcblist *rsclp, unsigned long seq)
483{
484	int i, j;
485
486	WARN_ON_ONCE(!rcu_segcblist_is_enabled(rsclp));
487	if (rcu_segcblist_restempty(rsclp, RCU_DONE_TAIL))
488		return;
489
490	/*
491	 * Find all callbacks whose ->gp_seq numbers indicate that they
492	 * are ready to invoke, and put them into the RCU_DONE_TAIL segment.
493	 */
494	for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
495		if (ULONG_CMP_LT(seq, rsclp->gp_seq[i]))
496			break;
497		WRITE_ONCE(rsclp->tails[RCU_DONE_TAIL], rsclp->tails[i]);
498		rcu_segcblist_move_seglen(rsclp, i, RCU_DONE_TAIL);
499	}
500
501	/* If no callbacks moved, nothing more need be done. */
502	if (i == RCU_WAIT_TAIL)
503		return;
504
505	/* Clean up tail pointers that might have been misordered above. */
506	for (j = RCU_WAIT_TAIL; j < i; j++)
507		WRITE_ONCE(rsclp->tails[j], rsclp->tails[RCU_DONE_TAIL]);
508
509	/*
510	 * Callbacks moved, so clean up the misordered ->tails[] pointers
511	 * that now point into the middle of the list of ready-to-invoke
512	 * callbacks.  The overall effect is to copy down the later pointers
513	 * into the gap that was created by the now-ready segments.
514	 */
515	for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
516		if (rsclp->tails[j] == rsclp->tails[RCU_NEXT_TAIL])
517			break;  /* No more callbacks. */
518		WRITE_ONCE(rsclp->tails[j], rsclp->tails[i]);
519		rcu_segcblist_move_seglen(rsclp, i, j);
520		rsclp->gp_seq[j] = rsclp->gp_seq[i];
521	}
522}
523
524/*
525 * "Accelerate" callbacks based on more-accurate grace-period information.
526 * The reason for this is that RCU does not synchronize the beginnings and
527 * ends of grace periods, and that callbacks are posted locally.  This in
528 * turn means that the callbacks must be labelled conservatively early
529 * on, as getting exact information would degrade both performance and
530 * scalability.  When more accurate grace-period information becomes
531 * available, previously posted callbacks can be "accelerated", marking
532 * them to complete at the end of the earlier grace period.
533 *
534 * This function operates on an rcu_segcblist structure, and also the
535 * grace-period sequence number seq at which new callbacks would become
536 * ready to invoke.  Returns true if there are callbacks that won't be
537 * ready to invoke until seq, false otherwise.
538 */
539bool rcu_segcblist_accelerate(struct rcu_segcblist *rsclp, unsigned long seq)
540{
541	int i, j;
542
543	WARN_ON_ONCE(!rcu_segcblist_is_enabled(rsclp));
544	if (rcu_segcblist_restempty(rsclp, RCU_DONE_TAIL))
545		return false;
546
547	/*
548	 * Find the segment preceding the oldest segment of callbacks
549	 * whose ->gp_seq[] completion is at or after that passed in via
550	 * "seq", skipping any empty segments.  This oldest segment, along
551	 * with any later segments, can be merged in with any newly arrived
552	 * callbacks in the RCU_NEXT_TAIL segment, and assigned "seq"
553	 * as their ->gp_seq[] grace-period completion sequence number.
554	 */
555	for (i = RCU_NEXT_READY_TAIL; i > RCU_DONE_TAIL; i--)
556		if (rsclp->tails[i] != rsclp->tails[i - 1] &&
557		    ULONG_CMP_LT(rsclp->gp_seq[i], seq))
558			break;
559
560	/*
561	 * If all the segments contain callbacks that correspond to
562	 * earlier grace-period sequence numbers than "seq", leave.
563	 * Assuming that the rcu_segcblist structure has enough
564	 * segments in its arrays, this can only happen if some of
565	 * the non-done segments contain callbacks that really are
566	 * ready to invoke.  This situation will get straightened
567	 * out by the next call to rcu_segcblist_advance().
568	 *
569	 * Also advance to the oldest segment of callbacks whose
570	 * ->gp_seq[] completion is at or after that passed in via "seq",
571	 * skipping any empty segments.
572	 *
573	 * Note that segment "i" (and any lower-numbered segments
574	 * containing older callbacks) will be unaffected, and their
575	 * grace-period numbers remain unchanged.  For example, if i ==
576	 * WAIT_TAIL, then neither WAIT_TAIL nor DONE_TAIL will be touched.
577	 * Instead, the CBs in NEXT_TAIL will be merged with those in
578	 * NEXT_READY_TAIL and the grace-period number of NEXT_READY_TAIL
579	 * would be updated.  NEXT_TAIL would then be empty.
580	 */
581	if (rcu_segcblist_restempty(rsclp, i) || ++i >= RCU_NEXT_TAIL)
582		return false;
583
584	/* Accounting: everything below i is about to get merged into i. */
585	for (j = i + 1; j <= RCU_NEXT_TAIL; j++)
586		rcu_segcblist_move_seglen(rsclp, j, i);
587
588	/*
589	 * Merge all later callbacks, including newly arrived callbacks,
590	 * into the segment located by the for-loop above.  Assign "seq"
591	 * as the ->gp_seq[] value in order to correctly handle the case
592	 * where there were no pending callbacks in the rcu_segcblist
593	 * structure other than in the RCU_NEXT_TAIL segment.
594	 */
595	for (; i < RCU_NEXT_TAIL; i++) {
596		WRITE_ONCE(rsclp->tails[i], rsclp->tails[RCU_NEXT_TAIL]);
597		rsclp->gp_seq[i] = seq;
598	}
599	return true;
600}
601
602/*
603 * Merge the source rcu_segcblist structure into the destination
604 * rcu_segcblist structure, then initialize the source.  Any pending
605 * callbacks from the source get to start over.  It is best to
606 * advance and accelerate both the destination and the source
607 * before merging.
608 */
609void rcu_segcblist_merge(struct rcu_segcblist *dst_rsclp,
610			 struct rcu_segcblist *src_rsclp)
611{
612	struct rcu_cblist donecbs;
613	struct rcu_cblist pendcbs;
614
615	lockdep_assert_cpus_held();
616
617	rcu_cblist_init(&donecbs);
618	rcu_cblist_init(&pendcbs);
619
620	rcu_segcblist_extract_done_cbs(src_rsclp, &donecbs);
621	rcu_segcblist_extract_pend_cbs(src_rsclp, &pendcbs);
622
623	/*
624	 * No need smp_mb() before setting length to 0, because CPU hotplug
625	 * lock excludes rcu_barrier.
626	 */
627	rcu_segcblist_set_len(src_rsclp, 0);
628
629	rcu_segcblist_insert_count(dst_rsclp, &donecbs);
630	rcu_segcblist_insert_count(dst_rsclp, &pendcbs);
631	rcu_segcblist_insert_done_cbs(dst_rsclp, &donecbs);
632	rcu_segcblist_insert_pend_cbs(dst_rsclp, &pendcbs);
633
634	rcu_segcblist_init(src_rsclp);
635}