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
rangelock_compare(const void * arg1,const void * arg2)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 (TREE_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
rangelock_init(rangelock_t * rl,rangelock_cb_t * cb,void * arg)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
rangelock_fini(rangelock_t * rl)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
rangelock_enter_writer(rangelock_t * rl,locked_range_t * new)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 *
rangelock_proxify(avl_tree_t * tree,locked_range_t * lr)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 *
rangelock_split(avl_tree_t * tree,locked_range_t * lr,uint64_t off)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
rangelock_new_proxy(avl_tree_t * tree,uint64_t off,uint64_t len)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
rangelock_add_reader(avl_tree_t * tree,locked_range_t * new,locked_range_t * prev,avl_index_t where)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
rangelock_enter_reader(rangelock_t * rl,locked_range_t * new)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 *
rangelock_enter(rangelock_t * rl,uint64_t off,uint64_t len,rangelock_type_t type)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
rangelock_exit_reader(rangelock_t * rl,locked_range_t * remove)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
rangelock_exit(locked_range_t * lr)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
rangelock_reduce(locked_range_t * lr,uint64_t off,uint64_t len)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