xref: /titanic_52/usr/src/uts/common/fs/zfs/zfs_rlock.c (revision c7158ae983f5a04c4a998f468ecefba6d23ba721)
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 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 /*
29  * This file contains the code to implement file range locking in
30  * ZFS, although there isn't much specific to ZFS (all that comes to mind
31  * support for growing the blocksize).
32  *
33  * Interface
34  * ---------
35  * Defined in zfs_rlock.h but essentially:
36  *	rl = zfs_range_lock(zp, off, len, lock_type);
37  *	zfs_range_unlock(rl);
38  *	zfs_range_reduce(rl, off, len);
39  *
40  * AVL tree
41  * --------
42  * An AVL tree is used to maintain the state of the existing ranges
43  * that are locked for exclusive (writer) or shared (reader) use.
44  * The starting range offset is used for searching and sorting the tree.
45  *
46  * Common case
47  * -----------
48  * The (hopefully) usual case is of no overlaps or contention for
49  * locks. On entry to zfs_lock_range() a rl_t is allocated; the tree
50  * searched that finds no overlap, and *this* rl_t is placed in the tree.
51  *
52  * Overlaps/Reference counting/Proxy locks
53  * ---------------------------------------
54  * The avl code only allows one node at a particular offset. Also it's very
55  * inefficient to search through all previous entries looking for overlaps
56  * (because the very 1st in the ordered list might be at offset 0 but
57  * cover the whole file).
58  * So this implementation uses reference counts and proxy range locks.
59  * Firstly, only reader locks use reference counts and proxy locks,
60  * because writer locks are exclusive.
61  * When a reader lock overlaps with another then a proxy lock is created
62  * for that range and replaces the original lock. If the overlap
63  * is exact then the reference count of the proxy is simply incremented.
64  * Otherwise, the proxy lock is split into smaller lock ranges and
65  * new proxy locks created for non overlapping ranges.
66  * The reference counts are adjusted accordingly.
67  * Meanwhile, the orginal lock is kept around (this is the callers handle)
68  * and its offset and length are used when releasing the lock.
69  *
70  * Thread coordination
71  * -------------------
72  * In order to make wakeups efficient and to ensure multiple continuous
73  * readers on a range don't starve a writer for the same range lock,
74  * two condition variables are allocated in each rl_t.
75  * If a writer (or reader) can't get a range it initialises the writer
76  * (or reader) cv; sets a flag saying there's a writer (or reader) waiting;
77  * and waits on that cv. When a thread unlocks that range it wakes up all
78  * writers then all readers before destroying the lock.
79  *
80  * Append mode writes
81  * ------------------
82  * Append mode writes need to lock a range at the end of a file.
83  * The offset of the end of the file is determined under the
84  * range locking mutex, and the lock type converted from RL_APPEND to
85  * RL_WRITER and the range locked.
86  *
87  * Grow block handling
88  * -------------------
89  * ZFS supports multiple block sizes currently upto 128K. The smallest
90  * block size is used for the file which is grown as needed. During this
91  * growth all other writers and readers must be excluded.
92  * So if the block size needs to be grown then the whole file is
93  * exclusively locked, then later the caller will reduce the lock
94  * range to just the range to be written using zfs_reduce_range.
95  */
96 
97 #include <sys/zfs_rlock.h>
98 
99 /*
100  * Check if a write lock can be grabbed, or wait and recheck until available.
101  */
102 static void
103 zfs_range_lock_writer(znode_t *zp, rl_t *new)
104 {
105 	avl_tree_t *tree = &zp->z_range_avl;
106 	rl_t *rl;
107 	avl_index_t where;
108 	uint64_t end_size;
109 	uint64_t off = new->r_off;
110 	uint64_t len = new->r_len;
111 
112 	for (;;) {
113 		/*
114 		 * Range locking is also used by zvol and uses a
115 		 * dummied up znode. However, for zvol, we don't need to
116 		 * append or grow blocksize, and besides we don't have
117 		 * a z_phys or z_zfsvfs - so skip that processing.
118 		 *
119 		 * Yes, this is ugly, and would be solved by not handling
120 		 * grow or append in range lock code. If that was done then
121 		 * we could make the range locking code generically available
122 		 * to other non-zfs consumers.
123 		 */
124 		if (zp->z_vnode) { /* caller is ZPL */
125 			/*
126 			 * If in append mode pick up the current end of file.
127 			 * This is done under z_range_lock to avoid races.
128 			 */
129 			if (new->r_type == RL_APPEND)
130 				new->r_off = zp->z_phys->zp_size;
131 
132 			/*
133 			 * If we need to grow the block size then grab the whole
134 			 * file range. This is also done under z_range_lock to
135 			 * avoid races.
136 			 */
137 			end_size = MAX(zp->z_phys->zp_size, new->r_off + len);
138 			if (end_size > zp->z_blksz && (!ISP2(zp->z_blksz) ||
139 			    zp->z_blksz < zp->z_zfsvfs->z_max_blksz)) {
140 				new->r_off = 0;
141 				new->r_len = UINT64_MAX;
142 			}
143 		}
144 
145 		/*
146 		 * First check for the usual case of no locks
147 		 */
148 		if (avl_numnodes(tree) == 0) {
149 			new->r_type = RL_WRITER; /* convert to writer */
150 			avl_add(tree, new);
151 			return;
152 		}
153 
154 		/*
155 		 * Look for any locks in the range.
156 		 */
157 		rl = avl_find(tree, new, &where);
158 		if (rl)
159 			goto wait; /* already locked at same offset */
160 
161 		rl = (rl_t *)avl_nearest(tree, where, AVL_AFTER);
162 		if (rl && (rl->r_off < new->r_off + new->r_len))
163 			goto wait;
164 
165 		rl = (rl_t *)avl_nearest(tree, where, AVL_BEFORE);
166 		if (rl && rl->r_off + rl->r_len > new->r_off)
167 			goto wait;
168 
169 		new->r_type = RL_WRITER; /* convert possible RL_APPEND */
170 		avl_insert(tree, new, where);
171 		return;
172 wait:
173 		if (!rl->r_write_wanted) {
174 			cv_init(&rl->r_wr_cv, NULL, CV_DEFAULT, NULL);
175 			rl->r_write_wanted = B_TRUE;
176 		}
177 		cv_wait(&rl->r_wr_cv, &zp->z_range_lock);
178 
179 		/* reset to original */
180 		new->r_off = off;
181 		new->r_len = len;
182 	}
183 }
184 
185 /*
186  * If this is an original (non-proxy) lock then replace it by
187  * a proxy and return the proxy.
188  */
189 static rl_t *
190 zfs_range_proxify(avl_tree_t *tree, rl_t *rl)
191 {
192 	rl_t *proxy;
193 
194 	if (rl->r_proxy)
195 		return (rl); /* already a proxy */
196 
197 	ASSERT3U(rl->r_cnt, ==, 1);
198 	ASSERT(rl->r_write_wanted == B_FALSE);
199 	ASSERT(rl->r_read_wanted == B_FALSE);
200 	avl_remove(tree, rl);
201 	rl->r_cnt = 0;
202 
203 	/* create a proxy range lock */
204 	proxy = kmem_alloc(sizeof (rl_t), KM_SLEEP);
205 	proxy->r_off = rl->r_off;
206 	proxy->r_len = rl->r_len;
207 	proxy->r_cnt = 1;
208 	proxy->r_type = RL_READER;
209 	proxy->r_proxy = B_TRUE;
210 	proxy->r_write_wanted = B_FALSE;
211 	proxy->r_read_wanted = B_FALSE;
212 	avl_add(tree, proxy);
213 
214 	return (proxy);
215 }
216 
217 /*
218  * Split the range lock at the supplied offset
219  * returning the *front* proxy.
220  */
221 static rl_t *
222 zfs_range_split(avl_tree_t *tree, rl_t *rl, uint64_t off)
223 {
224 	rl_t *front, *rear;
225 
226 	ASSERT3U(rl->r_len, >, 1);
227 	ASSERT3U(off, >, rl->r_off);
228 	ASSERT3U(off, <, rl->r_off + rl->r_len);
229 	ASSERT(rl->r_write_wanted == B_FALSE);
230 	ASSERT(rl->r_read_wanted == B_FALSE);
231 
232 	/* create the rear proxy range lock */
233 	rear = kmem_alloc(sizeof (rl_t), KM_SLEEP);
234 	rear->r_off = off;
235 	rear->r_len = rl->r_off + rl->r_len - off;
236 	rear->r_cnt = rl->r_cnt;
237 	rear->r_type = RL_READER;
238 	rear->r_proxy = B_TRUE;
239 	rear->r_write_wanted = B_FALSE;
240 	rear->r_read_wanted = B_FALSE;
241 
242 	front = zfs_range_proxify(tree, rl);
243 	front->r_len = off - rl->r_off;
244 
245 	avl_insert_here(tree, rear, front, AVL_AFTER);
246 	return (front);
247 }
248 
249 /*
250  * Create and add a new proxy range lock for the supplied range.
251  */
252 static void
253 zfs_range_new_proxy(avl_tree_t *tree, uint64_t off, uint64_t len)
254 {
255 	rl_t *rl;
256 
257 	ASSERT(len);
258 	rl = kmem_alloc(sizeof (rl_t), KM_SLEEP);
259 	rl->r_off = off;
260 	rl->r_len = len;
261 	rl->r_cnt = 1;
262 	rl->r_type = RL_READER;
263 	rl->r_proxy = B_TRUE;
264 	rl->r_write_wanted = B_FALSE;
265 	rl->r_read_wanted = B_FALSE;
266 	avl_add(tree, rl);
267 }
268 
269 static void
270 zfs_range_add_reader(avl_tree_t *tree, rl_t *new, rl_t *prev, avl_index_t where)
271 {
272 	rl_t *next;
273 	uint64_t off = new->r_off;
274 	uint64_t len = new->r_len;
275 
276 	/*
277 	 * prev arrives either:
278 	 * - pointing to an entry at the same offset
279 	 * - pointing to the entry with the closest previous offset whose
280 	 *   range may overlap with the new range
281 	 * - null, if there were no ranges starting before the new one
282 	 */
283 	if (prev) {
284 		if (prev->r_off + prev->r_len <= off) {
285 			prev = NULL;
286 		} else if (prev->r_off != off) {
287 			/*
288 			 * convert to proxy if needed then
289 			 * split this entry and bump ref count
290 			 */
291 			prev = zfs_range_split(tree, prev, off);
292 			prev = AVL_NEXT(tree, prev); /* move to rear range */
293 		}
294 	}
295 	ASSERT((prev == NULL) || (prev->r_off == off));
296 
297 	if (prev)
298 		next = prev;
299 	else
300 		next = (rl_t *)avl_nearest(tree, where, AVL_AFTER);
301 
302 	if (next == NULL || off + len <= next->r_off) {
303 		/* no overlaps, use the original new rl_t in the tree */
304 		avl_insert(tree, new, where);
305 		return;
306 	}
307 
308 	if (off < next->r_off) {
309 		/* Add a proxy for initial range before the overlap */
310 		zfs_range_new_proxy(tree, off, next->r_off - off);
311 	}
312 
313 	new->r_cnt = 0; /* will use proxies in tree */
314 	/*
315 	 * We now search forward through the ranges, until we go past the end
316 	 * of the new range. For each entry we make it a proxy if it
317 	 * isn't already, then bump its reference count. If there's any
318 	 * gaps between the ranges then we create a new proxy range.
319 	 */
320 	for (prev = NULL; next; prev = next, next = AVL_NEXT(tree, next)) {
321 		if (off + len <= next->r_off)
322 			break;
323 		if (prev && prev->r_off + prev->r_len < next->r_off) {
324 			/* there's a gap */
325 			ASSERT3U(next->r_off, >, prev->r_off + prev->r_len);
326 			zfs_range_new_proxy(tree, prev->r_off + prev->r_len,
327 			    next->r_off - (prev->r_off + prev->r_len));
328 		}
329 		if (off + len == next->r_off + next->r_len) {
330 			/* exact overlap with end */
331 			next = zfs_range_proxify(tree, next);
332 			next->r_cnt++;
333 			return;
334 		}
335 		if (off + len < next->r_off + next->r_len) {
336 			/* new range ends in the middle of this block */
337 			next = zfs_range_split(tree, next, off + len);
338 			next->r_cnt++;
339 			return;
340 		}
341 		ASSERT3U(off + len, >, next->r_off + next->r_len);
342 		next = zfs_range_proxify(tree, next);
343 		next->r_cnt++;
344 	}
345 
346 	/* Add the remaining end range. */
347 	zfs_range_new_proxy(tree, prev->r_off + prev->r_len,
348 	    (off + len) - (prev->r_off + prev->r_len));
349 }
350 
351 /*
352  * Check if a reader lock can be grabbed, or wait and recheck until available.
353  */
354 static void
355 zfs_range_lock_reader(znode_t *zp, rl_t *new)
356 {
357 	avl_tree_t *tree = &zp->z_range_avl;
358 	rl_t *prev, *next;
359 	avl_index_t where;
360 	uint64_t off = new->r_off;
361 	uint64_t len = new->r_len;
362 
363 	/*
364 	 * Look for any writer locks in the range.
365 	 */
366 retry:
367 	prev = avl_find(tree, new, &where);
368 	if (prev == NULL)
369 		prev = (rl_t *)avl_nearest(tree, where, AVL_BEFORE);
370 
371 	/*
372 	 * Check the previous range for a writer lock overlap.
373 	 */
374 	if (prev && (off < prev->r_off + prev->r_len)) {
375 		if ((prev->r_type == RL_WRITER) || (prev->r_write_wanted)) {
376 			if (!prev->r_read_wanted) {
377 				cv_init(&prev->r_rd_cv, NULL, CV_DEFAULT, NULL);
378 				prev->r_read_wanted = B_TRUE;
379 			}
380 			cv_wait(&prev->r_rd_cv, &zp->z_range_lock);
381 			goto retry;
382 		}
383 		if (off + len < prev->r_off + prev->r_len)
384 			goto got_lock;
385 	}
386 
387 	/*
388 	 * Search through the following ranges to see if there's
389 	 * write lock any overlap.
390 	 */
391 	if (prev)
392 		next = AVL_NEXT(tree, prev);
393 	else
394 		next = (rl_t *)avl_nearest(tree, where, AVL_AFTER);
395 	for (; next; next = AVL_NEXT(tree, next)) {
396 		if (off + len <= next->r_off)
397 			goto got_lock;
398 		if ((next->r_type == RL_WRITER) || (next->r_write_wanted)) {
399 			if (!next->r_read_wanted) {
400 				cv_init(&next->r_rd_cv, NULL, CV_DEFAULT, NULL);
401 				next->r_read_wanted = B_TRUE;
402 			}
403 			cv_wait(&next->r_rd_cv, &zp->z_range_lock);
404 			goto retry;
405 		}
406 		if (off + len <= next->r_off + next->r_len)
407 			goto got_lock;
408 	}
409 
410 got_lock:
411 	/*
412 	 * Add the read lock, which may involve splitting existing
413 	 * locks and bumping ref counts (r_cnt).
414 	 */
415 	zfs_range_add_reader(tree, new, prev, where);
416 }
417 
418 /*
419  * Lock a range (offset, length) as either shared (RL_READER)
420  * or exclusive (RL_WRITER). Returns the range lock structure
421  * for later unlocking or reduce range (if entire file
422  * previously locked as RL_WRITER).
423  */
424 rl_t *
425 zfs_range_lock(znode_t *zp, uint64_t off, uint64_t len, rl_type_t type)
426 {
427 	rl_t *new;
428 
429 	ASSERT(type == RL_READER || type == RL_WRITER || type == RL_APPEND);
430 
431 	new = kmem_alloc(sizeof (rl_t), KM_SLEEP);
432 	new->r_zp = zp;
433 	new->r_off = off;
434 	new->r_len = len;
435 	new->r_cnt = 1; /* assume it's going to be in the tree */
436 	new->r_type = type;
437 	new->r_proxy = B_FALSE;
438 	new->r_write_wanted = B_FALSE;
439 	new->r_read_wanted = B_FALSE;
440 
441 	mutex_enter(&zp->z_range_lock);
442 	if (type == RL_READER) {
443 		/*
444 		 * First check for the usual case of no locks
445 		 */
446 		if (avl_numnodes(&zp->z_range_avl) == 0)
447 			avl_add(&zp->z_range_avl, new);
448 		else
449 			zfs_range_lock_reader(zp, new);
450 	} else
451 		zfs_range_lock_writer(zp, new); /* RL_WRITER or RL_APPEND */
452 	mutex_exit(&zp->z_range_lock);
453 	return (new);
454 }
455 
456 /*
457  * Unlock a reader lock
458  */
459 static void
460 zfs_range_unlock_reader(znode_t *zp, rl_t *remove)
461 {
462 	avl_tree_t *tree = &zp->z_range_avl;
463 	rl_t *rl, *next;
464 	uint64_t len;
465 
466 	/*
467 	 * The common case is when the remove entry is in the tree
468 	 * (cnt == 1) meaning there's been no other reader locks overlapping
469 	 * with this one. Otherwise the remove entry will have been
470 	 * removed from the tree and replaced by proxies (one or
471 	 * more ranges mapping to the entire range).
472 	 */
473 	if (remove->r_cnt == 1) {
474 		avl_remove(tree, remove);
475 		if (remove->r_write_wanted) {
476 			cv_broadcast(&remove->r_wr_cv);
477 			cv_destroy(&remove->r_wr_cv);
478 		}
479 		if (remove->r_read_wanted) {
480 			cv_broadcast(&remove->r_rd_cv);
481 			cv_destroy(&remove->r_rd_cv);
482 		}
483 	} else {
484 		ASSERT3U(remove->r_cnt, ==, 0);
485 		ASSERT3U(remove->r_write_wanted, ==, 0);
486 		ASSERT3U(remove->r_read_wanted, ==, 0);
487 		/*
488 		 * Find start proxy representing this reader lock,
489 		 * then decrement ref count on all proxies
490 		 * that make up this range, freeing them as needed.
491 		 */
492 		rl = avl_find(tree, remove, NULL);
493 		ASSERT(rl);
494 		ASSERT(rl->r_cnt);
495 		ASSERT(rl->r_type == RL_READER);
496 		for (len = remove->r_len; len != 0; rl = next) {
497 			len -= rl->r_len;
498 			if (len) {
499 				next = AVL_NEXT(tree, rl);
500 				ASSERT(next);
501 				ASSERT(rl->r_off + rl->r_len == next->r_off);
502 				ASSERT(next->r_cnt);
503 				ASSERT(next->r_type == RL_READER);
504 			}
505 			rl->r_cnt--;
506 			if (rl->r_cnt == 0) {
507 				avl_remove(tree, rl);
508 				if (rl->r_write_wanted) {
509 					cv_broadcast(&rl->r_wr_cv);
510 					cv_destroy(&rl->r_wr_cv);
511 				}
512 				if (rl->r_read_wanted) {
513 					cv_broadcast(&rl->r_rd_cv);
514 					cv_destroy(&rl->r_rd_cv);
515 				}
516 				kmem_free(rl, sizeof (rl_t));
517 			}
518 		}
519 	}
520 	kmem_free(remove, sizeof (rl_t));
521 }
522 
523 /*
524  * Unlock range and destroy range lock structure.
525  */
526 void
527 zfs_range_unlock(rl_t *rl)
528 {
529 	znode_t *zp = rl->r_zp;
530 
531 	ASSERT(rl->r_type == RL_WRITER || rl->r_type == RL_READER);
532 	ASSERT(rl->r_cnt == 1 || rl->r_cnt == 0);
533 	ASSERT(!rl->r_proxy);
534 
535 	mutex_enter(&zp->z_range_lock);
536 	if (rl->r_type == RL_WRITER) {
537 		/* writer locks can't be shared or split */
538 		avl_remove(&zp->z_range_avl, rl);
539 		mutex_exit(&zp->z_range_lock);
540 		if (rl->r_write_wanted) {
541 			cv_broadcast(&rl->r_wr_cv);
542 			cv_destroy(&rl->r_wr_cv);
543 		}
544 		if (rl->r_read_wanted) {
545 			cv_broadcast(&rl->r_rd_cv);
546 			cv_destroy(&rl->r_rd_cv);
547 		}
548 		kmem_free(rl, sizeof (rl_t));
549 	} else {
550 		/*
551 		 * lock may be shared, let zfs_range_unlock_reader()
552 		 * release the lock and free the rl_t
553 		 */
554 		zfs_range_unlock_reader(zp, rl);
555 		mutex_exit(&zp->z_range_lock);
556 	}
557 }
558 
559 /*
560  * Reduce range locked as RL_WRITER from whole file to specified range.
561  * Asserts the whole file is exclusivly locked and so there's only one
562  * entry in the tree.
563  */
564 void
565 zfs_range_reduce(rl_t *rl, uint64_t off, uint64_t len)
566 {
567 	znode_t *zp = rl->r_zp;
568 
569 	/* Ensure there are no other locks */
570 	ASSERT(avl_numnodes(&zp->z_range_avl) == 1);
571 	ASSERT(rl->r_off == 0);
572 	ASSERT(rl->r_type == RL_WRITER);
573 	ASSERT(!rl->r_proxy);
574 	ASSERT3U(rl->r_len, ==, UINT64_MAX);
575 	ASSERT3U(rl->r_cnt, ==, 1);
576 
577 	mutex_enter(&zp->z_range_lock);
578 	rl->r_off = off;
579 	rl->r_len = len;
580 	mutex_exit(&zp->z_range_lock);
581 	if (rl->r_write_wanted)
582 		cv_broadcast(&rl->r_wr_cv);
583 	if (rl->r_read_wanted)
584 		cv_broadcast(&rl->r_rd_cv);
585 }
586 
587 /*
588  * AVL comparison function used to order range locks
589  * Locks are ordered on the start offset of the range.
590  */
591 int
592 zfs_range_compare(const void *arg1, const void *arg2)
593 {
594 	const rl_t *rl1 = arg1;
595 	const rl_t *rl2 = arg2;
596 
597 	if (rl1->r_off > rl2->r_off)
598 		return (1);
599 	if (rl1->r_off < rl2->r_off)
600 		return (-1);
601 	return (0);
602 }
603