xref: /freebsd/sys/contrib/ck/src/ck_epoch.c (revision e0c4386e)
1 /*
2  * Copyright 2011-2015 Samy Al Bahra.
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  *
14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
24  * SUCH DAMAGE.
25  */
26 
27 /*
28  * The implementation here is inspired from the work described in:
29  *   Fraser, K. 2004. Practical Lock-Freedom. PhD Thesis, University
30  *   of Cambridge Computing Laboratory.
31  */
32 
33 #include <ck_backoff.h>
34 #include <ck_cc.h>
35 #include <ck_epoch.h>
36 #include <ck_pr.h>
37 #include <ck_stack.h>
38 #include <ck_stdbool.h>
39 #include <ck_string.h>
40 
41 /*
42  * Only three distinct values are used for reclamation, but reclamation occurs
43  * at e+2 rather than e+1. Any thread in a "critical section" would have
44  * acquired some snapshot (e) of the global epoch value (e_g) and set an active
45  * flag. Any hazardous references will only occur after a full memory barrier.
46  * For example, assume an initial e_g value of 1, e value of 0 and active value
47  * of 0.
48  *
49  * ck_epoch_begin(...)
50  *   e = e_g
51  *   active = 1
52  *   memory_barrier();
53  *
54  * Any serialized reads may observe e = 0 or e = 1 with active = 0, or e = 0 or
55  * e = 1 with active = 1. The e_g value can only go from 1 to 2 if every thread
56  * has already observed the value of "1" (or the value we are incrementing
57  * from). This guarantees us that for any given value e_g, any threads with-in
58  * critical sections (referred to as "active" threads from here on) would have
59  * an e value of e_g-1 or e_g. This also means that hazardous references may be
60  * shared in both e_g-1 and e_g even if they are logically deleted in e_g.
61  *
62  * For example, assume all threads have an e value of e_g. Another thread may
63  * increment to e_g to e_g+1. Older threads may have a reference to an object
64  * which is only deleted in e_g+1. It could be that reader threads are
65  * executing some hash table look-ups, while some other writer thread (which
66  * causes epoch counter tick) actually deletes the same items that reader
67  * threads are looking up (this writer thread having an e value of e_g+1).
68  * This is possible if the writer thread re-observes the epoch after the
69  * counter tick.
70  *
71  * Psuedo-code for writer:
72  *   ck_epoch_begin()
73  *   ht_delete(x)
74  *   ck_epoch_end()
75  *   ck_epoch_begin()
76  *   ht_delete(x)
77  *   ck_epoch_end()
78  *
79  * Psuedo-code for reader:
80  *   for (;;) {
81  *      x = ht_lookup(x)
82  *      ck_pr_inc(&x->value);
83  *   }
84  *
85  * Of course, it is also possible for references logically deleted at e_g-1 to
86  * still be accessed at e_g as threads are "active" at the same time
87  * (real-world time) mutating shared objects.
88  *
89  * Now, if the epoch counter is ticked to e_g+1, then no new hazardous
90  * references could exist to objects logically deleted at e_g-1. The reason for
91  * this is that at e_g+1, all epoch read-side critical sections started at
92  * e_g-1 must have been completed. If any epoch read-side critical sections at
93  * e_g-1 were still active, then we would never increment to e_g+1 (active != 0
94  * ^ e != e_g).  Additionally, e_g may still have hazardous references to
95  * objects logically deleted at e_g-1 which means objects logically deleted at
96  * e_g-1 cannot be deleted at e_g+1 unless all threads have observed e_g+1
97  * (since it is valid for active threads to be at e_g and threads at e_g still
98  * require safe memory accesses).
99  *
100  * However, at e_g+2, all active threads must be either at e_g+1 or e_g+2.
101  * Though e_g+2 may share hazardous references with e_g+1, and e_g+1 shares
102  * hazardous references to e_g, no active threads are at e_g or e_g-1. This
103  * means no hazardous references could exist to objects deleted at e_g-1 (at
104  * e_g+2).
105  *
106  * To summarize these important points,
107  *   1) Active threads will always have a value of e_g or e_g-1.
108  *   2) Items that are logically deleted e_g or e_g-1 cannot be physically
109  *      deleted.
110  *   3) Objects logically deleted at e_g-1 can be physically destroyed at e_g+2
111  *      or at e_g+1 if no threads are at e_g.
112  *
113  * Last but not least, if we are at e_g+2, then no active thread is at e_g
114  * which means it is safe to apply modulo-3 arithmetic to e_g value in order to
115  * re-use e_g to represent the e_g+3 state. This means it is sufficient to
116  * represent e_g using only the values 0, 1 or 2. Every time a thread re-visits
117  * a e_g (which can be determined with a non-empty deferral list) it can assume
118  * objects in the e_g deferral list involved at least three e_g transitions and
119  * are thus, safe, for physical deletion.
120  *
121  * Blocking semantics for epoch reclamation have additional restrictions.
122  * Though we only require three deferral lists, reasonable blocking semantics
123  * must be able to more gracefully handle bursty write work-loads which could
124  * easily cause e_g wrap-around if modulo-3 arithmetic is used. This allows for
125  * easy-to-trigger live-lock situations. The work-around to this is to not
126  * apply modulo arithmetic to e_g but only to deferral list indexing.
127  */
128 #define CK_EPOCH_GRACE 3U
129 
130 /*
131  * CK_EPOCH_LENGTH must be a power-of-2 (because (CK_EPOCH_LENGTH - 1) is used
132  * as a mask, and it must be at least 3 (see comments above).
133  */
134 #if (CK_EPOCH_LENGTH < 3 || (CK_EPOCH_LENGTH & (CK_EPOCH_LENGTH - 1)) != 0)
135 #error "CK_EPOCH_LENGTH must be a power of 2 and >= 3"
136 #endif
137 
138 enum {
139 	CK_EPOCH_STATE_USED = 0,
140 	CK_EPOCH_STATE_FREE = 1
141 };
142 
143 CK_STACK_CONTAINER(struct ck_epoch_record, record_next,
144     ck_epoch_record_container)
145 CK_STACK_CONTAINER(struct ck_epoch_entry, stack_entry,
146     ck_epoch_entry_container)
147 
148 #define CK_EPOCH_SENSE_MASK	(CK_EPOCH_SENSE - 1)
149 
150 bool
151 _ck_epoch_delref(struct ck_epoch_record *record,
152     struct ck_epoch_section *section)
153 {
154 	struct ck_epoch_ref *current, *other;
155 	unsigned int i = section->bucket;
156 
157 	current = &record->local.bucket[i];
158 	current->count--;
159 
160 	if (current->count > 0)
161 		return false;
162 
163 	/*
164 	 * If the current bucket no longer has any references, then
165 	 * determine whether we have already transitioned into a newer
166 	 * epoch. If so, then make sure to update our shared snapshot
167 	 * to allow for forward progress.
168 	 *
169 	 * If no other active bucket exists, then the record will go
170 	 * inactive in order to allow for forward progress.
171 	 */
172 	other = &record->local.bucket[(i + 1) & CK_EPOCH_SENSE_MASK];
173 	if (other->count > 0 &&
174 	    ((int)(current->epoch - other->epoch) < 0)) {
175 		/*
176 		 * The other epoch value is actually the newest,
177 		 * transition to it.
178 		 */
179 		ck_pr_store_uint(&record->epoch, other->epoch);
180 	}
181 
182 	return true;
183 }
184 
185 void
186 _ck_epoch_addref(struct ck_epoch_record *record,
187     struct ck_epoch_section *section)
188 {
189 	struct ck_epoch *global = record->global;
190 	struct ck_epoch_ref *ref;
191 	unsigned int epoch, i;
192 
193 	epoch = ck_pr_load_uint(&global->epoch);
194 	i = epoch & CK_EPOCH_SENSE_MASK;
195 	ref = &record->local.bucket[i];
196 
197 	if (ref->count++ == 0) {
198 #ifndef CK_MD_TSO
199 		struct ck_epoch_ref *previous;
200 
201 		/*
202 		 * The system has already ticked. If another non-zero bucket
203 		 * exists, make sure to order our observations with respect
204 		 * to it. Otherwise, it is possible to acquire a reference
205 		 * from the previous epoch generation.
206 		 *
207 		 * On TSO architectures, the monoticity of the global counter
208 		 * and load-{store, load} ordering are sufficient to guarantee
209 		 * this ordering.
210 		 */
211 		previous = &record->local.bucket[(i + 1) &
212 		    CK_EPOCH_SENSE_MASK];
213 		if (previous->count > 0)
214 			ck_pr_fence_acqrel();
215 #endif /* !CK_MD_TSO */
216 
217 		/*
218 		 * If this is this is a new reference into the current
219 		 * bucket then cache the associated epoch value.
220 		 */
221 		ref->epoch = epoch;
222 	}
223 
224 	section->bucket = i;
225 	return;
226 }
227 
228 void
229 ck_epoch_init(struct ck_epoch *global)
230 {
231 
232 	ck_stack_init(&global->records);
233 	global->epoch = 1;
234 	global->n_free = 0;
235 	ck_pr_fence_store();
236 	return;
237 }
238 
239 struct ck_epoch_record *
240 ck_epoch_recycle(struct ck_epoch *global, void *ct)
241 {
242 	struct ck_epoch_record *record;
243 	ck_stack_entry_t *cursor;
244 	unsigned int state;
245 
246 	if (ck_pr_load_uint(&global->n_free) == 0)
247 		return NULL;
248 
249 	CK_STACK_FOREACH(&global->records, cursor) {
250 		record = ck_epoch_record_container(cursor);
251 
252 		if (ck_pr_load_uint(&record->state) == CK_EPOCH_STATE_FREE) {
253 			/* Serialize with respect to deferral list clean-up. */
254 			ck_pr_fence_load();
255 			state = ck_pr_fas_uint(&record->state,
256 			    CK_EPOCH_STATE_USED);
257 			if (state == CK_EPOCH_STATE_FREE) {
258 				ck_pr_dec_uint(&global->n_free);
259 				ck_pr_store_ptr(&record->ct, ct);
260 
261 				/*
262 				 * The context pointer is ordered by a
263 				 * subsequent protected section.
264 				 */
265 				return record;
266 			}
267 		}
268 	}
269 
270 	return NULL;
271 }
272 
273 void
274 ck_epoch_register(struct ck_epoch *global, struct ck_epoch_record *record,
275     void *ct)
276 {
277 	size_t i;
278 
279 	record->global = global;
280 	record->state = CK_EPOCH_STATE_USED;
281 	record->active = 0;
282 	record->epoch = 0;
283 	record->n_dispatch = 0;
284 	record->n_peak = 0;
285 	record->n_pending = 0;
286 	record->ct = ct;
287 	memset(&record->local, 0, sizeof record->local);
288 
289 	for (i = 0; i < CK_EPOCH_LENGTH; i++)
290 		ck_stack_init(&record->pending[i]);
291 
292 	ck_pr_fence_store();
293 	ck_stack_push_upmc(&global->records, &record->record_next);
294 	return;
295 }
296 
297 void
298 ck_epoch_unregister(struct ck_epoch_record *record)
299 {
300 	struct ck_epoch *global = record->global;
301 	size_t i;
302 
303 	record->active = 0;
304 	record->epoch = 0;
305 	record->n_dispatch = 0;
306 	record->n_peak = 0;
307 	record->n_pending = 0;
308 	memset(&record->local, 0, sizeof record->local);
309 
310 	for (i = 0; i < CK_EPOCH_LENGTH; i++)
311 		ck_stack_init(&record->pending[i]);
312 
313 	ck_pr_store_ptr(&record->ct, NULL);
314 	ck_pr_fence_store();
315 	ck_pr_store_uint(&record->state, CK_EPOCH_STATE_FREE);
316 	ck_pr_inc_uint(&global->n_free);
317 	return;
318 }
319 
320 static struct ck_epoch_record *
321 ck_epoch_scan(struct ck_epoch *global,
322     struct ck_epoch_record *cr,
323     unsigned int epoch,
324     bool *af)
325 {
326 	ck_stack_entry_t *cursor;
327 
328 	if (cr == NULL) {
329 		cursor = CK_STACK_FIRST(&global->records);
330 		*af = false;
331 	} else {
332 		cursor = &cr->record_next;
333 		*af = true;
334 	}
335 
336 	while (cursor != NULL) {
337 		unsigned int state, active;
338 
339 		cr = ck_epoch_record_container(cursor);
340 
341 		state = ck_pr_load_uint(&cr->state);
342 		if (state & CK_EPOCH_STATE_FREE) {
343 			cursor = CK_STACK_NEXT(cursor);
344 			continue;
345 		}
346 
347 		active = ck_pr_load_uint(&cr->active);
348 		*af |= active;
349 
350 		if (active != 0 && ck_pr_load_uint(&cr->epoch) != epoch)
351 			return cr;
352 
353 		cursor = CK_STACK_NEXT(cursor);
354 	}
355 
356 	return NULL;
357 }
358 
359 static unsigned int
360 ck_epoch_dispatch(struct ck_epoch_record *record, unsigned int e, ck_stack_t *deferred)
361 {
362 	unsigned int epoch = e & (CK_EPOCH_LENGTH - 1);
363 	ck_stack_entry_t *head, *next, *cursor;
364 	unsigned int n_pending, n_peak;
365 	unsigned int i = 0;
366 
367 	head = ck_stack_batch_pop_upmc(&record->pending[epoch]);
368 	for (cursor = head; cursor != NULL; cursor = next) {
369 		struct ck_epoch_entry *entry =
370 		    ck_epoch_entry_container(cursor);
371 
372 		next = CK_STACK_NEXT(cursor);
373 		if (deferred != NULL)
374 			ck_stack_push_spnc(deferred, &entry->stack_entry);
375 		else
376 			entry->function(entry);
377 
378 		i++;
379 	}
380 
381 	n_peak = ck_pr_load_uint(&record->n_peak);
382 	n_pending = ck_pr_load_uint(&record->n_pending);
383 
384 	/* We don't require accuracy around peak calculation. */
385 	if (n_pending > n_peak)
386 		ck_pr_store_uint(&record->n_peak, n_peak);
387 
388 	if (i > 0) {
389 		ck_pr_add_uint(&record->n_dispatch, i);
390 		ck_pr_sub_uint(&record->n_pending, i);
391 	}
392 
393 	return i;
394 }
395 
396 /*
397  * Reclaim all objects associated with a record.
398  */
399 void
400 ck_epoch_reclaim(struct ck_epoch_record *record)
401 {
402 	unsigned int epoch;
403 
404 	for (epoch = 0; epoch < CK_EPOCH_LENGTH; epoch++)
405 		ck_epoch_dispatch(record, epoch, NULL);
406 
407 	return;
408 }
409 
410 CK_CC_FORCE_INLINE static void
411 epoch_block(struct ck_epoch *global, struct ck_epoch_record *cr,
412     ck_epoch_wait_cb_t *cb, void *ct)
413 {
414 
415 	if (cb != NULL)
416 		cb(global, cr, ct);
417 
418 	return;
419 }
420 
421 /*
422  * This function must not be called with-in read section.
423  */
424 void
425 ck_epoch_synchronize_wait(struct ck_epoch *global,
426     ck_epoch_wait_cb_t *cb, void *ct)
427 {
428 	struct ck_epoch_record *cr;
429 	unsigned int delta, epoch, goal, i;
430 	bool active;
431 
432 	ck_pr_fence_memory();
433 
434 	/*
435 	 * The observation of the global epoch must be ordered with respect to
436 	 * all prior operations. The re-ordering of loads is permitted given
437 	 * monoticity of global epoch counter.
438 	 *
439 	 * If UINT_MAX concurrent mutations were to occur then it is possible
440 	 * to encounter an ABA-issue. If this is a concern, consider tuning
441 	 * write-side concurrency.
442 	 */
443 	delta = epoch = ck_pr_load_uint(&global->epoch);
444 	goal = epoch + CK_EPOCH_GRACE;
445 
446 	for (i = 0, cr = NULL; i < CK_EPOCH_GRACE - 1; cr = NULL, i++) {
447 		bool r;
448 
449 		/*
450 		 * Determine whether all threads have observed the current
451 		 * epoch with respect to the updates on invocation.
452 		 */
453 		while (cr = ck_epoch_scan(global, cr, delta, &active),
454 		    cr != NULL) {
455 			unsigned int e_d;
456 
457 			ck_pr_stall();
458 
459 			/*
460 			 * Another writer may have already observed a grace
461 			 * period.
462 			 */
463 			e_d = ck_pr_load_uint(&global->epoch);
464 			if (e_d == delta) {
465 				epoch_block(global, cr, cb, ct);
466 				continue;
467 			}
468 
469 			/*
470 			 * If the epoch has been updated, we may have already
471 			 * met our goal.
472 			 */
473 			delta = e_d;
474 			if ((goal > epoch) & (delta >= goal))
475 				goto leave;
476 
477 			epoch_block(global, cr, cb, ct);
478 
479 			/*
480 			 * If the epoch has been updated, then a grace period
481 			 * requires that all threads are observed idle at the
482 			 * same epoch.
483 			 */
484 			cr = NULL;
485 		}
486 
487 		/*
488 		 * If we have observed all threads as inactive, then we assume
489 		 * we are at a grace period.
490 		 */
491 		if (active == false)
492 			break;
493 
494 		/*
495 		 * Increment current epoch. CAS semantics are used to eliminate
496 		 * increment operations for synchronization that occurs for the
497 		 * same global epoch value snapshot.
498 		 *
499 		 * If we can guarantee there will only be one active barrier or
500 		 * epoch tick at a given time, then it is sufficient to use an
501 		 * increment operation. In a multi-barrier workload, however,
502 		 * it is possible to overflow the epoch value if we apply
503 		 * modulo-3 arithmetic.
504 		 */
505 		r = ck_pr_cas_uint_value(&global->epoch, delta, delta + 1,
506 		    &delta);
507 
508 		/* Order subsequent thread active checks. */
509 		ck_pr_fence_atomic_load();
510 
511 		/*
512 		 * If CAS has succeeded, then set delta to latest snapshot.
513 		 * Otherwise, we have just acquired latest snapshot.
514 		 */
515 		delta = delta + r;
516 	}
517 
518 	/*
519 	 * A majority of use-cases will not require full barrier semantics.
520 	 * However, if non-temporal instructions are used, full barrier
521 	 * semantics are necessary.
522 	 */
523 leave:
524 	ck_pr_fence_memory();
525 	return;
526 }
527 
528 void
529 ck_epoch_synchronize(struct ck_epoch_record *record)
530 {
531 
532 	ck_epoch_synchronize_wait(record->global, NULL, NULL);
533 	return;
534 }
535 
536 void
537 ck_epoch_barrier(struct ck_epoch_record *record)
538 {
539 
540 	ck_epoch_synchronize(record);
541 	ck_epoch_reclaim(record);
542 	return;
543 }
544 
545 void
546 ck_epoch_barrier_wait(struct ck_epoch_record *record, ck_epoch_wait_cb_t *cb,
547     void *ct)
548 {
549 
550 	ck_epoch_synchronize_wait(record->global, cb, ct);
551 	ck_epoch_reclaim(record);
552 	return;
553 }
554 
555 /*
556  * It may be worth it to actually apply these deferral semantics to an epoch
557  * that was observed at ck_epoch_call time. The problem is that the latter
558  * would require a full fence.
559  *
560  * ck_epoch_call will dispatch to the latest epoch snapshot that was observed.
561  * There are cases where it will fail to reclaim as early as it could. If this
562  * becomes a problem, we could actually use a heap for epoch buckets but that
563  * is far from ideal too.
564  */
565 bool
566 ck_epoch_poll_deferred(struct ck_epoch_record *record, ck_stack_t *deferred)
567 {
568 	bool active;
569 	unsigned int epoch;
570 	struct ck_epoch_record *cr = NULL;
571 	struct ck_epoch *global = record->global;
572 	unsigned int n_dispatch;
573 
574 	epoch = ck_pr_load_uint(&global->epoch);
575 
576 	/* Serialize epoch snapshots with respect to global epoch. */
577 	ck_pr_fence_memory();
578 
579 	/*
580 	 * At this point, epoch is the current global epoch value.
581 	 * There may or may not be active threads which observed epoch - 1.
582 	 * (ck_epoch_scan() will tell us that). However, there should be
583 	 * no active threads which observed epoch - 2.
584 	 *
585 	 * Note that checking epoch - 2 is necessary, as race conditions can
586 	 * allow another thread to increment the global epoch before this
587 	 * thread runs.
588 	 */
589 	n_dispatch = ck_epoch_dispatch(record, epoch - 2, deferred);
590 
591 	cr = ck_epoch_scan(global, cr, epoch, &active);
592 	if (cr != NULL)
593 		return (n_dispatch > 0);
594 
595 	/* We are at a grace period if all threads are inactive. */
596 	if (active == false) {
597 		record->epoch = epoch;
598 		for (epoch = 0; epoch < CK_EPOCH_LENGTH; epoch++)
599 			ck_epoch_dispatch(record, epoch, deferred);
600 
601 		return true;
602 	}
603 
604 	/*
605 	 * If an active thread exists, rely on epoch observation.
606 	 *
607 	 * All the active threads entered the epoch section during
608 	 * the current epoch. Therefore, we can now run the handlers
609 	 * for the immediately preceding epoch and attempt to
610 	 * advance the epoch if it hasn't been already.
611 	 */
612 	(void)ck_pr_cas_uint(&global->epoch, epoch, epoch + 1);
613 
614 	ck_epoch_dispatch(record, epoch - 1, deferred);
615 	return true;
616 }
617 
618 bool
619 ck_epoch_poll(struct ck_epoch_record *record)
620 {
621 
622 	return ck_epoch_poll_deferred(record, NULL);
623 }
624