xref: /illumos-gate/usr/src/uts/common/fs/zfs/zfs_rlock.c (revision 142d813a06c6f9a6142e2c276b62129a17a31a65)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 /*
26  * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
27  */
28 
29 /*
30  * This file contains the code to implement file range locking in
31  * ZFS, although there isn't much specific to ZFS (all that comes to mind is
32  * support for growing the blocksize).
33  *
34  * Interface
35  * ---------
36  * Defined in zfs_rlock.h but essentially:
37  *	lr = rangelock_enter(zp, off, len, lock_type);
38  *	rangelock_reduce(lr, off, len); // optional
39  *	rangelock_exit(lr);
40  *
41  * AVL tree
42  * --------
43  * An AVL tree is used to maintain the state of the existing ranges
44  * that are locked for exclusive (writer) or shared (reader) use.
45  * The starting range offset is used for searching and sorting the tree.
46  *
47  * Common case
48  * -----------
49  * The (hopefully) usual case is of no overlaps or contention for locks. On
50  * entry to rangelock_enter(), a locked_range_t is allocated; the tree
51  * searched that finds no overlap, and *this* locked_range_t is placed in the
52  * tree.
53  *
54  * Overlaps/Reference counting/Proxy locks
55  * ---------------------------------------
56  * The avl code only allows one node at a particular offset. Also it's very
57  * inefficient to search through all previous entries looking for overlaps
58  * (because the very 1st in the ordered list might be at offset 0 but
59  * cover the whole file).
60  * So this implementation uses reference counts and proxy range locks.
61  * Firstly, only reader locks use reference counts and proxy locks,
62  * because writer locks are exclusive.
63  * When a reader lock overlaps with another then a proxy lock is created
64  * for that range and replaces the original lock. If the overlap
65  * is exact then the reference count of the proxy is simply incremented.
66  * Otherwise, the proxy lock is split into smaller lock ranges and
67  * new proxy locks created for non overlapping ranges.
68  * The reference counts are adjusted accordingly.
69  * Meanwhile, the orginal lock is kept around (this is the callers handle)
70  * and its offset and length are used when releasing the lock.
71  *
72  * Thread coordination
73  * -------------------
74  * In order to make wakeups efficient and to ensure multiple continuous
75  * readers on a range don't starve a writer for the same range lock,
76  * two condition variables are allocated in each rl_t.
77  * If a writer (or reader) can't get a range it initialises the writer
78  * (or reader) cv; sets a flag saying there's a writer (or reader) waiting;
79  * and waits on that cv. When a thread unlocks that range it wakes up all
80  * writers then all readers before destroying the lock.
81  *
82  * Append mode writes
83  * ------------------
84  * Append mode writes need to lock a range at the end of a file.
85  * The offset of the end of the file is determined under the
86  * range locking mutex, and the lock type converted from RL_APPEND to
87  * RL_WRITER and the range locked.
88  *
89  * Grow block handling
90  * -------------------
91  * ZFS supports multiple block sizes, up to 16MB. The smallest
92  * block size is used for the file which is grown as needed. During this
93  * growth all other writers and readers must be excluded.
94  * So if the block size needs to be grown then the whole file is
95  * exclusively locked, then later the caller will reduce the lock
96  * range to just the range to be written using rangelock_reduce().
97  */
98 
99 #include <sys/zfs_context.h>
100 #include <sys/zfs_rlock.h>
101 
102 /*
103  * AVL comparison function used to order range locks
104  * Locks are ordered on the start offset of the range.
105  */
106 static int
107 rangelock_compare(const void *arg1, const void *arg2)
108 {
109 	const locked_range_t *rl1 = (const locked_range_t *)arg1;
110 	const locked_range_t *rl2 = (const locked_range_t *)arg2;
111 
112 	return (AVL_CMP(rl1->lr_offset, rl2->lr_offset));
113 }
114 
115 /*
116  * The callback is invoked when acquiring a RL_WRITER or RL_APPEND lock.
117  * It must convert RL_APPEND to RL_WRITER (starting at the end of the file),
118  * and may increase the range that's locked for RL_WRITER.
119  */
120 void
121 rangelock_init(rangelock_t *rl, rangelock_cb_t *cb, void *arg)
122 {
123 	mutex_init(&rl->rl_lock, NULL, MUTEX_DEFAULT, NULL);
124 	avl_create(&rl->rl_tree, rangelock_compare,
125 	    sizeof (locked_range_t), offsetof(locked_range_t, lr_node));
126 	rl->rl_cb = cb;
127 	rl->rl_arg = arg;
128 }
129 
130 void
131 rangelock_fini(rangelock_t *rl)
132 {
133 	mutex_destroy(&rl->rl_lock);
134 	avl_destroy(&rl->rl_tree);
135 }
136 
137 /*
138  * Check if a write lock can be grabbed, or wait and recheck until available.
139  */
140 static void
141 rangelock_enter_writer(rangelock_t *rl, locked_range_t *new)
142 {
143 	avl_tree_t *tree = &rl->rl_tree;
144 	locked_range_t *lr;
145 	avl_index_t where;
146 	uint64_t orig_off = new->lr_offset;
147 	uint64_t orig_len = new->lr_length;
148 	rangelock_type_t orig_type = new->lr_type;
149 
150 	for (;;) {
151 		/*
152 		 * Call callback which can modify new->r_off,len,type.
153 		 * Note, the callback is used by the ZPL to handle appending
154 		 * and changing blocksizes.  It isn't needed for zvols.
155 		 */
156 		if (rl->rl_cb != NULL) {
157 			rl->rl_cb(new, rl->rl_arg);
158 		}
159 
160 		/*
161 		 * If the type was APPEND, the callback must convert it to
162 		 * WRITER.
163 		 */
164 		ASSERT3U(new->lr_type, ==, RL_WRITER);
165 
166 		/*
167 		 * First check for the usual case of no locks
168 		 */
169 		if (avl_numnodes(tree) == 0) {
170 			avl_add(tree, new);
171 			return;
172 		}
173 
174 		/*
175 		 * Look for any locks in the range.
176 		 */
177 		lr = avl_find(tree, new, &where);
178 		if (lr != NULL)
179 			goto wait; /* already locked at same offset */
180 
181 		lr = (locked_range_t *)avl_nearest(tree, where, AVL_AFTER);
182 		if (lr != NULL &&
183 		    lr->lr_offset < new->lr_offset + new->lr_length)
184 			goto wait;
185 
186 		lr = (locked_range_t *)avl_nearest(tree, where, AVL_BEFORE);
187 		if (lr != NULL &&
188 		    lr->lr_offset + lr->lr_length > new->lr_offset)
189 			goto wait;
190 
191 		avl_insert(tree, new, where);
192 		return;
193 wait:
194 		if (!lr->lr_write_wanted) {
195 			cv_init(&lr->lr_write_cv, NULL, CV_DEFAULT, NULL);
196 			lr->lr_write_wanted = B_TRUE;
197 		}
198 		cv_wait(&lr->lr_write_cv, &rl->rl_lock);
199 
200 		/* reset to original */
201 		new->lr_offset = orig_off;
202 		new->lr_length = orig_len;
203 		new->lr_type = orig_type;
204 	}
205 }
206 
207 /*
208  * If this is an original (non-proxy) lock then replace it by
209  * a proxy and return the proxy.
210  */
211 static locked_range_t *
212 rangelock_proxify(avl_tree_t *tree, locked_range_t *lr)
213 {
214 	locked_range_t *proxy;
215 
216 	if (lr->lr_proxy)
217 		return (lr); /* already a proxy */
218 
219 	ASSERT3U(lr->lr_count, ==, 1);
220 	ASSERT(lr->lr_write_wanted == B_FALSE);
221 	ASSERT(lr->lr_read_wanted == B_FALSE);
222 	avl_remove(tree, lr);
223 	lr->lr_count = 0;
224 
225 	/* create a proxy range lock */
226 	proxy = kmem_alloc(sizeof (locked_range_t), KM_SLEEP);
227 	proxy->lr_offset = lr->lr_offset;
228 	proxy->lr_length = lr->lr_length;
229 	proxy->lr_count = 1;
230 	proxy->lr_type = RL_READER;
231 	proxy->lr_proxy = B_TRUE;
232 	proxy->lr_write_wanted = B_FALSE;
233 	proxy->lr_read_wanted = B_FALSE;
234 	avl_add(tree, proxy);
235 
236 	return (proxy);
237 }
238 
239 /*
240  * Split the range lock at the supplied offset
241  * returning the *front* proxy.
242  */
243 static locked_range_t *
244 rangelock_split(avl_tree_t *tree, locked_range_t *lr, uint64_t off)
245 {
246 	ASSERT3U(lr->lr_length, >, 1);
247 	ASSERT3U(off, >, lr->lr_offset);
248 	ASSERT3U(off, <, lr->lr_offset + lr->lr_length);
249 	ASSERT(lr->lr_write_wanted == B_FALSE);
250 	ASSERT(lr->lr_read_wanted == B_FALSE);
251 
252 	/* create the rear proxy range lock */
253 	locked_range_t *rear = kmem_alloc(sizeof (locked_range_t), KM_SLEEP);
254 	rear->lr_offset = off;
255 	rear->lr_length = lr->lr_offset + lr->lr_length - off;
256 	rear->lr_count = lr->lr_count;
257 	rear->lr_type = RL_READER;
258 	rear->lr_proxy = B_TRUE;
259 	rear->lr_write_wanted = B_FALSE;
260 	rear->lr_read_wanted = B_FALSE;
261 
262 	locked_range_t *front = rangelock_proxify(tree, lr);
263 	front->lr_length = off - lr->lr_offset;
264 
265 	avl_insert_here(tree, rear, front, AVL_AFTER);
266 	return (front);
267 }
268 
269 /*
270  * Create and add a new proxy range lock for the supplied range.
271  */
272 static void
273 rangelock_new_proxy(avl_tree_t *tree, uint64_t off, uint64_t len)
274 {
275 	ASSERT(len != 0);
276 	locked_range_t *lr = kmem_alloc(sizeof (locked_range_t), KM_SLEEP);
277 	lr->lr_offset = off;
278 	lr->lr_length = len;
279 	lr->lr_count = 1;
280 	lr->lr_type = RL_READER;
281 	lr->lr_proxy = B_TRUE;
282 	lr->lr_write_wanted = B_FALSE;
283 	lr->lr_read_wanted = B_FALSE;
284 	avl_add(tree, lr);
285 }
286 
287 static void
288 rangelock_add_reader(avl_tree_t *tree, locked_range_t *new,
289     locked_range_t *prev, avl_index_t where)
290 {
291 	locked_range_t *next;
292 	uint64_t off = new->lr_offset;
293 	uint64_t len = new->lr_length;
294 
295 	/*
296 	 * prev arrives either:
297 	 * - pointing to an entry at the same offset
298 	 * - pointing to the entry with the closest previous offset whose
299 	 *   range may overlap with the new range
300 	 * - null, if there were no ranges starting before the new one
301 	 */
302 	if (prev != NULL) {
303 		if (prev->lr_offset + prev->lr_length <= off) {
304 			prev = NULL;
305 		} else if (prev->lr_offset != off) {
306 			/*
307 			 * convert to proxy if needed then
308 			 * split this entry and bump ref count
309 			 */
310 			prev = rangelock_split(tree, prev, off);
311 			prev = AVL_NEXT(tree, prev); /* move to rear range */
312 		}
313 	}
314 	ASSERT((prev == NULL) || (prev->lr_offset == off));
315 
316 	if (prev != NULL)
317 		next = prev;
318 	else
319 		next = avl_nearest(tree, where, AVL_AFTER);
320 
321 	if (next == NULL || off + len <= next->lr_offset) {
322 		/* no overlaps, use the original new rl_t in the tree */
323 		avl_insert(tree, new, where);
324 		return;
325 	}
326 
327 	if (off < next->lr_offset) {
328 		/* Add a proxy for initial range before the overlap */
329 		rangelock_new_proxy(tree, off, next->lr_offset - off);
330 	}
331 
332 	new->lr_count = 0; /* will use proxies in tree */
333 	/*
334 	 * We now search forward through the ranges, until we go past the end
335 	 * of the new range. For each entry we make it a proxy if it
336 	 * isn't already, then bump its reference count. If there's any
337 	 * gaps between the ranges then we create a new proxy range.
338 	 */
339 	for (prev = NULL; next; prev = next, next = AVL_NEXT(tree, next)) {
340 		if (off + len <= next->lr_offset)
341 			break;
342 		if (prev != NULL && prev->lr_offset + prev->lr_length <
343 		    next->lr_offset) {
344 			/* there's a gap */
345 			ASSERT3U(next->lr_offset, >,
346 			    prev->lr_offset + prev->lr_length);
347 			rangelock_new_proxy(tree,
348 			    prev->lr_offset + prev->lr_length,
349 			    next->lr_offset -
350 			    (prev->lr_offset + prev->lr_length));
351 		}
352 		if (off + len == next->lr_offset + next->lr_length) {
353 			/* exact overlap with end */
354 			next = rangelock_proxify(tree, next);
355 			next->lr_count++;
356 			return;
357 		}
358 		if (off + len < next->lr_offset + next->lr_length) {
359 			/* new range ends in the middle of this block */
360 			next = rangelock_split(tree, next, off + len);
361 			next->lr_count++;
362 			return;
363 		}
364 		ASSERT3U(off + len, >, next->lr_offset + next->lr_length);
365 		next = rangelock_proxify(tree, next);
366 		next->lr_count++;
367 	}
368 
369 	/* Add the remaining end range. */
370 	rangelock_new_proxy(tree, prev->lr_offset + prev->lr_length,
371 	    (off + len) - (prev->lr_offset + prev->lr_length));
372 }
373 
374 /*
375  * Check if a reader lock can be grabbed, or wait and recheck until available.
376  */
377 static void
378 rangelock_enter_reader(rangelock_t *rl, locked_range_t *new)
379 {
380 	avl_tree_t *tree = &rl->rl_tree;
381 	locked_range_t *prev, *next;
382 	avl_index_t where;
383 	uint64_t off = new->lr_offset;
384 	uint64_t len = new->lr_length;
385 
386 	/*
387 	 * Look for any writer locks in the range.
388 	 */
389 retry:
390 	prev = avl_find(tree, new, &where);
391 	if (prev == NULL)
392 		prev = (locked_range_t *)avl_nearest(tree, where, AVL_BEFORE);
393 
394 	/*
395 	 * Check the previous range for a writer lock overlap.
396 	 */
397 	if (prev && (off < prev->lr_offset + prev->lr_length)) {
398 		if ((prev->lr_type == RL_WRITER) || (prev->lr_write_wanted)) {
399 			if (!prev->lr_read_wanted) {
400 				cv_init(&prev->lr_read_cv,
401 				    NULL, CV_DEFAULT, NULL);
402 				prev->lr_read_wanted = B_TRUE;
403 			}
404 			cv_wait(&prev->lr_read_cv, &rl->rl_lock);
405 			goto retry;
406 		}
407 		if (off + len < prev->lr_offset + prev->lr_length)
408 			goto got_lock;
409 	}
410 
411 	/*
412 	 * Search through the following ranges to see if there's
413 	 * write lock any overlap.
414 	 */
415 	if (prev != NULL)
416 		next = AVL_NEXT(tree, prev);
417 	else
418 		next = (locked_range_t *)avl_nearest(tree, where, AVL_AFTER);
419 	for (; next != NULL; next = AVL_NEXT(tree, next)) {
420 		if (off + len <= next->lr_offset)
421 			goto got_lock;
422 		if ((next->lr_type == RL_WRITER) || (next->lr_write_wanted)) {
423 			if (!next->lr_read_wanted) {
424 				cv_init(&next->lr_read_cv,
425 				    NULL, CV_DEFAULT, NULL);
426 				next->lr_read_wanted = B_TRUE;
427 			}
428 			cv_wait(&next->lr_read_cv, &rl->rl_lock);
429 			goto retry;
430 		}
431 		if (off + len <= next->lr_offset + next->lr_length)
432 			goto got_lock;
433 	}
434 
435 got_lock:
436 	/*
437 	 * Add the read lock, which may involve splitting existing
438 	 * locks and bumping ref counts (r_count).
439 	 */
440 	rangelock_add_reader(tree, new, prev, where);
441 }
442 
443 /*
444  * Lock a range (offset, length) as either shared (RL_READER) or exclusive
445  * (RL_WRITER or RL_APPEND).  If RL_APPEND is specified, rl_cb() will convert
446  * it to a RL_WRITER lock (with the offset at the end of the file).  Returns
447  * the range lock structure for later unlocking (or reduce range if the
448  * entire file is locked as RL_WRITER).
449  */
450 locked_range_t *
451 rangelock_enter(rangelock_t *rl, uint64_t off, uint64_t len,
452     rangelock_type_t type)
453 {
454 	ASSERT(type == RL_READER || type == RL_WRITER || type == RL_APPEND);
455 
456 	locked_range_t *new = kmem_alloc(sizeof (locked_range_t), KM_SLEEP);
457 	new->lr_rangelock = rl;
458 	new->lr_offset = off;
459 	if (len + off < off)	/* overflow */
460 		len = UINT64_MAX - off;
461 	new->lr_length = len;
462 	new->lr_count = 1; /* assume it's going to be in the tree */
463 	new->lr_type = type;
464 	new->lr_proxy = B_FALSE;
465 	new->lr_write_wanted = B_FALSE;
466 	new->lr_read_wanted = B_FALSE;
467 
468 	mutex_enter(&rl->rl_lock);
469 	if (type == RL_READER) {
470 		/*
471 		 * First check for the usual case of no locks
472 		 */
473 		if (avl_numnodes(&rl->rl_tree) == 0)
474 			avl_add(&rl->rl_tree, new);
475 		else
476 			rangelock_enter_reader(rl, new);
477 	} else
478 		rangelock_enter_writer(rl, new); /* RL_WRITER or RL_APPEND */
479 	mutex_exit(&rl->rl_lock);
480 	return (new);
481 }
482 
483 /*
484  * Unlock a reader lock
485  */
486 static void
487 rangelock_exit_reader(rangelock_t *rl, locked_range_t *remove)
488 {
489 	avl_tree_t *tree = &rl->rl_tree;
490 	uint64_t len;
491 
492 	/*
493 	 * The common case is when the remove entry is in the tree
494 	 * (cnt == 1) meaning there's been no other reader locks overlapping
495 	 * with this one. Otherwise the remove entry will have been
496 	 * removed from the tree and replaced by proxies (one or
497 	 * more ranges mapping to the entire range).
498 	 */
499 	if (remove->lr_count == 1) {
500 		avl_remove(tree, remove);
501 		if (remove->lr_write_wanted) {
502 			cv_broadcast(&remove->lr_write_cv);
503 			cv_destroy(&remove->lr_write_cv);
504 		}
505 		if (remove->lr_read_wanted) {
506 			cv_broadcast(&remove->lr_read_cv);
507 			cv_destroy(&remove->lr_read_cv);
508 		}
509 	} else {
510 		ASSERT0(remove->lr_count);
511 		ASSERT0(remove->lr_write_wanted);
512 		ASSERT0(remove->lr_read_wanted);
513 		/*
514 		 * Find start proxy representing this reader lock,
515 		 * then decrement ref count on all proxies
516 		 * that make up this range, freeing them as needed.
517 		 */
518 		locked_range_t *lr = avl_find(tree, remove, NULL);
519 		ASSERT3P(lr, !=, NULL);
520 		ASSERT3U(lr->lr_count, !=, 0);
521 		ASSERT3U(lr->lr_type, ==, RL_READER);
522 		locked_range_t *next = NULL;
523 		for (len = remove->lr_length; len != 0; lr = next) {
524 			len -= lr->lr_length;
525 			if (len != 0) {
526 				next = AVL_NEXT(tree, lr);
527 				ASSERT3P(next, !=, NULL);
528 				ASSERT3U(lr->lr_offset + lr->lr_length, ==,
529 				    next->lr_offset);
530 				ASSERT3U(next->lr_count, !=, 0);
531 				ASSERT3U(next->lr_type, ==, RL_READER);
532 			}
533 			lr->lr_count--;
534 			if (lr->lr_count == 0) {
535 				avl_remove(tree, lr);
536 				if (lr->lr_write_wanted) {
537 					cv_broadcast(&lr->lr_write_cv);
538 					cv_destroy(&lr->lr_write_cv);
539 				}
540 				if (lr->lr_read_wanted) {
541 					cv_broadcast(&lr->lr_read_cv);
542 					cv_destroy(&lr->lr_read_cv);
543 				}
544 				kmem_free(lr, sizeof (locked_range_t));
545 			}
546 		}
547 	}
548 	kmem_free(remove, sizeof (locked_range_t));
549 }
550 
551 /*
552  * Unlock range and destroy range lock structure.
553  */
554 void
555 rangelock_exit(locked_range_t *lr)
556 {
557 	rangelock_t *rl = lr->lr_rangelock;
558 
559 	ASSERT(lr->lr_type == RL_WRITER || lr->lr_type == RL_READER);
560 	ASSERT(lr->lr_count == 1 || lr->lr_count == 0);
561 	ASSERT(!lr->lr_proxy);
562 
563 	mutex_enter(&rl->rl_lock);
564 	if (lr->lr_type == RL_WRITER) {
565 		/* writer locks can't be shared or split */
566 		avl_remove(&rl->rl_tree, lr);
567 		mutex_exit(&rl->rl_lock);
568 		if (lr->lr_write_wanted) {
569 			cv_broadcast(&lr->lr_write_cv);
570 			cv_destroy(&lr->lr_write_cv);
571 		}
572 		if (lr->lr_read_wanted) {
573 			cv_broadcast(&lr->lr_read_cv);
574 			cv_destroy(&lr->lr_read_cv);
575 		}
576 		kmem_free(lr, sizeof (locked_range_t));
577 	} else {
578 		/*
579 		 * lock may be shared, let rangelock_exit_reader()
580 		 * release the lock and free the rl_t
581 		 */
582 		rangelock_exit_reader(rl, lr);
583 		mutex_exit(&rl->rl_lock);
584 	}
585 }
586 
587 /*
588  * Reduce range locked as RL_WRITER from whole file to specified range.
589  * Asserts the whole file is exclusively locked and so there's only one
590  * entry in the tree.
591  */
592 void
593 rangelock_reduce(locked_range_t *lr, uint64_t off, uint64_t len)
594 {
595 	rangelock_t *rl = lr->lr_rangelock;
596 
597 	/* Ensure there are no other locks */
598 	ASSERT3U(avl_numnodes(&rl->rl_tree), ==, 1);
599 	ASSERT3U(lr->lr_offset, ==, 0);
600 	ASSERT3U(lr->lr_type, ==, RL_WRITER);
601 	ASSERT(!lr->lr_proxy);
602 	ASSERT3U(lr->lr_length, ==, UINT64_MAX);
603 	ASSERT3U(lr->lr_count, ==, 1);
604 
605 	mutex_enter(&rl->rl_lock);
606 	lr->lr_offset = off;
607 	lr->lr_length = len;
608 	mutex_exit(&rl->rl_lock);
609 	if (lr->lr_write_wanted)
610 		cv_broadcast(&lr->lr_write_cv);
611 	if (lr->lr_read_wanted)
612 		cv_broadcast(&lr->lr_read_cv);
613 }
614