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 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
32 * support for growing the blocksize).
33 *
34 * Interface
35 * ---------
36 * Defined in zfs_rlock.h but essentially:
37 * rl = zfs_range_lock(zp, off, len, lock_type);
38 * zfs_range_unlock(rl);
39 * zfs_range_reduce(rl, off, len);
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
50 * locks. On entry to zfs_lock_range() a rl_t is allocated; the tree
51 * searched that finds no overlap, and *this* rl_t is placed in the tree.
52 *
53 * Overlaps/Reference counting/Proxy locks
54 * ---------------------------------------
55 * The avl code only allows one node at a particular offset. Also it's very
56 * inefficient to search through all previous entries looking for overlaps
57 * (because the very 1st in the ordered list might be at offset 0 but
58 * cover the whole file).
59 * So this implementation uses reference counts and proxy range locks.
60 * Firstly, only reader locks use reference counts and proxy locks,
61 * because writer locks are exclusive.
62 * When a reader lock overlaps with another then a proxy lock is created
63 * for that range and replaces the original lock. If the overlap
64 * is exact then the reference count of the proxy is simply incremented.
65 * Otherwise, the proxy lock is split into smaller lock ranges and
66 * new proxy locks created for non overlapping ranges.
67 * The reference counts are adjusted accordingly.
68 * Meanwhile, the orginal lock is kept around (this is the callers handle)
69 * and its offset and length are used when releasing the lock.
70 *
71 * Thread coordination
72 * -------------------
73 * In order to make wakeups efficient and to ensure multiple continuous
74 * readers on a range don't starve a writer for the same range lock,
75 * two condition variables are allocated in each rl_t.
76 * If a writer (or reader) can't get a range it initialises the writer
77 * (or reader) cv; sets a flag saying there's a writer (or reader) waiting;
78 * and waits on that cv. When a thread unlocks that range it wakes up all
79 * writers then all readers before destroying the lock.
80 *
81 * Append mode writes
82 * ------------------
83 * Append mode writes need to lock a range at the end of a file.
84 * The offset of the end of the file is determined under the
85 * range locking mutex, and the lock type converted from RL_APPEND to
86 * RL_WRITER and the range locked.
87 *
88 * Grow block handling
89 * -------------------
90 * ZFS supports multiple block sizes currently upto 128K. The smallest
91 * block size is used for the file which is grown as needed. During this
92 * growth all other writers and readers must be excluded.
93 * So if the block size needs to be grown then the whole file is
94 * exclusively locked, then later the caller will reduce the lock
95 * range to just the range to be written using zfs_reduce_range.
96 */
97
98 #include <sys/zfs_rlock.h>
99
100 /*
101 * Check if a write lock can be grabbed, or wait and recheck until available.
102 */
103 static void
104 zfs_range_lock_writer(znode_t *zp, rl_t *new)
105 {
106 avl_tree_t *tree = &zp->z_range_avl;
107 rl_t *rl;
108 avl_index_t where;
109 uint64_t end_size;
110 uint64_t off = new->r_off;
111 uint64_t len = new->r_len;
112
113 for (;;) {
114 /*
115 * Range locking is also used by zvol and uses a
116 * dummied up znode. However, for zvol, we don't need to
117 * append or grow blocksize, and besides we don't have
118 * a "sa" data or z_zfsvfs - so skip that processing.
119 *
120 * Yes, this is ugly, and would be solved by not handling
121 * grow or append in range lock code. If that was done then
122 * we could make the range locking code generically available
123 * to other non-zfs consumers.
124 */
125 if (zp->z_vnode) { /* caller is ZPL */
126 /*
127 * If in append mode pick up the current end of file.
128 * This is done under z_range_lock to avoid races.
129 */
130 if (new->r_type == RL_APPEND)
131 new->r_off = zp->z_size;
132
133 /*
134 * If we need to grow the block size then grab the whole
135 * file range. This is also done under z_range_lock to
136 * avoid races.
137 */
138 end_size = MAX(zp->z_size, new->r_off + len);
139 if (end_size > zp->z_blksz && (!ISP2(zp->z_blksz) ||
140 zp->z_blksz < zp->z_zfsvfs->z_max_blksz)) {
141 new->r_off = 0;
142 new->r_len = UINT64_MAX;
143 }
144 }
145
146 /*
147 * First check for the usual case of no locks
148 */
149 if (avl_numnodes(tree) == 0) {
150 new->r_type = RL_WRITER; /* convert to writer */
151 avl_add(tree, new);
152 return;
153 }
154
155 /*
156 * Look for any locks in the range.
157 */
158 rl = avl_find(tree, new, &where);
159 if (rl)
160 goto wait; /* already locked at same offset */
161
162 rl = (rl_t *)avl_nearest(tree, where, AVL_AFTER);
163 if (rl && (rl->r_off < new->r_off + new->r_len))
164 goto wait;
165
166 rl = (rl_t *)avl_nearest(tree, where, AVL_BEFORE);
167 if (rl && rl->r_off + rl->r_len > new->r_off)
168 goto wait;
169
170 new->r_type = RL_WRITER; /* convert possible RL_APPEND */
171 avl_insert(tree, new, where);
172 return;
173 wait:
174 if (!rl->r_write_wanted) {
175 cv_init(&rl->r_wr_cv, NULL, CV_DEFAULT, NULL);
176 rl->r_write_wanted = B_TRUE;
177 }
178 cv_wait(&rl->r_wr_cv, &zp->z_range_lock);
179
180 /* reset to original */
181 new->r_off = off;
182 new->r_len = len;
183 }
184 }
185
186 /*
187 * If this is an original (non-proxy) lock then replace it by
188 * a proxy and return the proxy.
189 */
190 static rl_t *
191 zfs_range_proxify(avl_tree_t *tree, rl_t *rl)
192 {
193 rl_t *proxy;
194
195 if (rl->r_proxy)
196 return (rl); /* already a proxy */
197
198 ASSERT3U(rl->r_cnt, ==, 1);
199 ASSERT(rl->r_write_wanted == B_FALSE);
200 ASSERT(rl->r_read_wanted == B_FALSE);
201 avl_remove(tree, rl);
202 rl->r_cnt = 0;
203
204 /* create a proxy range lock */
205 proxy = kmem_alloc(sizeof (rl_t), KM_SLEEP);
206 proxy->r_off = rl->r_off;
207 proxy->r_len = rl->r_len;
208 proxy->r_cnt = 1;
209 proxy->r_type = RL_READER;
210 proxy->r_proxy = B_TRUE;
211 proxy->r_write_wanted = B_FALSE;
212 proxy->r_read_wanted = B_FALSE;
213 avl_add(tree, proxy);
214
215 return (proxy);
216 }
217
218 /*
219 * Split the range lock at the supplied offset
220 * returning the *front* proxy.
221 */
222 static rl_t *
223 zfs_range_split(avl_tree_t *tree, rl_t *rl, uint64_t off)
224 {
225 rl_t *front, *rear;
226
227 ASSERT3U(rl->r_len, >, 1);
228 ASSERT3U(off, >, rl->r_off);
229 ASSERT3U(off, <, rl->r_off + rl->r_len);
230 ASSERT(rl->r_write_wanted == B_FALSE);
231 ASSERT(rl->r_read_wanted == B_FALSE);
232
233 /* create the rear proxy range lock */
234 rear = kmem_alloc(sizeof (rl_t), KM_SLEEP);
235 rear->r_off = off;
236 rear->r_len = rl->r_off + rl->r_len - off;
237 rear->r_cnt = rl->r_cnt;
238 rear->r_type = RL_READER;
239 rear->r_proxy = B_TRUE;
240 rear->r_write_wanted = B_FALSE;
241 rear->r_read_wanted = B_FALSE;
242
243 front = zfs_range_proxify(tree, rl);
244 front->r_len = off - rl->r_off;
245
246 avl_insert_here(tree, rear, front, AVL_AFTER);
247 return (front);
248 }
249
250 /*
251 * Create and add a new proxy range lock for the supplied range.
252 */
253 static void
254 zfs_range_new_proxy(avl_tree_t *tree, uint64_t off, uint64_t len)
255 {
256 rl_t *rl;
257
258 ASSERT(len);
259 rl = kmem_alloc(sizeof (rl_t), KM_SLEEP);
260 rl->r_off = off;
261 rl->r_len = len;
262 rl->r_cnt = 1;
263 rl->r_type = RL_READER;
264 rl->r_proxy = B_TRUE;
265 rl->r_write_wanted = B_FALSE;
266 rl->r_read_wanted = B_FALSE;
267 avl_add(tree, rl);
268 }
269
270 static void
271 zfs_range_add_reader(avl_tree_t *tree, rl_t *new, rl_t *prev, avl_index_t where)
272 {
273 rl_t *next;
274 uint64_t off = new->r_off;
275 uint64_t len = new->r_len;
276
277 /*
278 * prev arrives either:
279 * - pointing to an entry at the same offset
280 * - pointing to the entry with the closest previous offset whose
281 * range may overlap with the new range
282 * - null, if there were no ranges starting before the new one
283 */
284 if (prev) {
285 if (prev->r_off + prev->r_len <= off) {
286 prev = NULL;
287 } else if (prev->r_off != off) {
288 /*
289 * convert to proxy if needed then
290 * split this entry and bump ref count
291 */
292 prev = zfs_range_split(tree, prev, off);
293 prev = AVL_NEXT(tree, prev); /* move to rear range */
294 }
295 }
296 ASSERT((prev == NULL) || (prev->r_off == off));
297
298 if (prev)
299 next = prev;
300 else
301 next = (rl_t *)avl_nearest(tree, where, AVL_AFTER);
302
303 if (next == NULL || off + len <= next->r_off) {
304 /* no overlaps, use the original new rl_t in the tree */
305 avl_insert(tree, new, where);
306 return;
307 }
308
309 if (off < next->r_off) {
310 /* Add a proxy for initial range before the overlap */
311 zfs_range_new_proxy(tree, off, next->r_off - off);
312 }
313
314 new->r_cnt = 0; /* will use proxies in tree */
315 /*
316 * We now search forward through the ranges, until we go past the end
317 * of the new range. For each entry we make it a proxy if it
318 * isn't already, then bump its reference count. If there's any
319 * gaps between the ranges then we create a new proxy range.
320 */
321 for (prev = NULL; next; prev = next, next = AVL_NEXT(tree, next)) {
322 if (off + len <= next->r_off)
323 break;
324 if (prev && prev->r_off + prev->r_len < next->r_off) {
325 /* there's a gap */
326 ASSERT3U(next->r_off, >, prev->r_off + prev->r_len);
327 zfs_range_new_proxy(tree, prev->r_off + prev->r_len,
328 next->r_off - (prev->r_off + prev->r_len));
329 }
330 if (off + len == next->r_off + next->r_len) {
331 /* exact overlap with end */
332 next = zfs_range_proxify(tree, next);
333 next->r_cnt++;
334 return;
335 }
336 if (off + len < next->r_off + next->r_len) {
337 /* new range ends in the middle of this block */
338 next = zfs_range_split(tree, next, off + len);
339 next->r_cnt++;
340 return;
341 }
342 ASSERT3U(off + len, >, next->r_off + next->r_len);
343 next = zfs_range_proxify(tree, next);
344 next->r_cnt++;
345 }
346
347 /* Add the remaining end range. */
348 zfs_range_new_proxy(tree, prev->r_off + prev->r_len,
349 (off + len) - (prev->r_off + prev->r_len));
350 }
351
352 /*
353 * Check if a reader lock can be grabbed, or wait and recheck until available.
354 */
355 static void
356 zfs_range_lock_reader(znode_t *zp, rl_t *new)
357 {
358 avl_tree_t *tree = &zp->z_range_avl;
359 rl_t *prev, *next;
360 avl_index_t where;
361 uint64_t off = new->r_off;
362 uint64_t len = new->r_len;
363
364 /*
365 * Look for any writer locks in the range.
366 */
367 retry:
368 prev = avl_find(tree, new, &where);
369 if (prev == NULL)
370 prev = (rl_t *)avl_nearest(tree, where, AVL_BEFORE);
371
372 /*
373 * Check the previous range for a writer lock overlap.
374 */
375 if (prev && (off < prev->r_off + prev->r_len)) {
376 if ((prev->r_type == RL_WRITER) || (prev->r_write_wanted)) {
377 if (!prev->r_read_wanted) {
378 cv_init(&prev->r_rd_cv, NULL, CV_DEFAULT, NULL);
379 prev->r_read_wanted = B_TRUE;
380 }
381 cv_wait(&prev->r_rd_cv, &zp->z_range_lock);
382 goto retry;
383 }
384 if (off + len < prev->r_off + prev->r_len)
385 goto got_lock;
386 }
387
388 /*
389 * Search through the following ranges to see if there's
390 * write lock any overlap.
391 */
392 if (prev)
393 next = AVL_NEXT(tree, prev);
394 else
395 next = (rl_t *)avl_nearest(tree, where, AVL_AFTER);
396 for (; next; next = AVL_NEXT(tree, next)) {
397 if (off + len <= next->r_off)
398 goto got_lock;
399 if ((next->r_type == RL_WRITER) || (next->r_write_wanted)) {
400 if (!next->r_read_wanted) {
401 cv_init(&next->r_rd_cv, NULL, CV_DEFAULT, NULL);
402 next->r_read_wanted = B_TRUE;
403 }
404 cv_wait(&next->r_rd_cv, &zp->z_range_lock);
405 goto retry;
406 }
407 if (off + len <= next->r_off + next->r_len)
408 goto got_lock;
409 }
410
411 got_lock:
412 /*
413 * Add the read lock, which may involve splitting existing
414 * locks and bumping ref counts (r_cnt).
415 */
416 zfs_range_add_reader(tree, new, prev, where);
417 }
418
419 /*
420 * Lock a range (offset, length) as either shared (RL_READER)
421 * or exclusive (RL_WRITER). Returns the range lock structure
422 * for later unlocking or reduce range (if entire file
423 * previously locked as RL_WRITER).
424 */
425 rl_t *
426 zfs_range_lock(znode_t *zp, uint64_t off, uint64_t len, rl_type_t type)
427 {
428 rl_t *new;
429
430 ASSERT(type == RL_READER || type == RL_WRITER || type == RL_APPEND);
431
432 new = kmem_alloc(sizeof (rl_t), KM_SLEEP);
433 new->r_zp = zp;
434 new->r_off = off;
435 if (len + off < off) /* overflow */
436 len = UINT64_MAX - off;
437 new->r_len = len;
438 new->r_cnt = 1; /* assume it's going to be in the tree */
439 new->r_type = type;
440 new->r_proxy = B_FALSE;
441 new->r_write_wanted = B_FALSE;
442 new->r_read_wanted = B_FALSE;
443
444 mutex_enter(&zp->z_range_lock);
445 if (type == RL_READER) {
446 /*
447 * First check for the usual case of no locks
448 */
449 if (avl_numnodes(&zp->z_range_avl) == 0)
450 avl_add(&zp->z_range_avl, new);
451 else
452 zfs_range_lock_reader(zp, new);
453 } else
454 zfs_range_lock_writer(zp, new); /* RL_WRITER or RL_APPEND */
455 mutex_exit(&zp->z_range_lock);
456 return (new);
457 }
458
459 /*
460 * Unlock a reader lock
461 */
462 static void
463 zfs_range_unlock_reader(znode_t *zp, rl_t *remove)
464 {
465 avl_tree_t *tree = &zp->z_range_avl;
466 rl_t *rl, *next = NULL;
467 uint64_t len;
468
469 /*
470 * The common case is when the remove entry is in the tree
471 * (cnt == 1) meaning there's been no other reader locks overlapping
472 * with this one. Otherwise the remove entry will have been
473 * removed from the tree and replaced by proxies (one or
474 * more ranges mapping to the entire range).
475 */
476 if (remove->r_cnt == 1) {
477 avl_remove(tree, remove);
478 if (remove->r_write_wanted) {
479 cv_broadcast(&remove->r_wr_cv);
480 cv_destroy(&remove->r_wr_cv);
481 }
482 if (remove->r_read_wanted) {
483 cv_broadcast(&remove->r_rd_cv);
484 cv_destroy(&remove->r_rd_cv);
485 }
486 } else {
487 ASSERT0(remove->r_cnt);
488 ASSERT0(remove->r_write_wanted);
489 ASSERT0(remove->r_read_wanted);
490 /*
491 * Find start proxy representing this reader lock,
492 * then decrement ref count on all proxies
493 * that make up this range, freeing them as needed.
494 */
495 rl = avl_find(tree, remove, NULL);
496 ASSERT(rl);
497 ASSERT(rl->r_cnt);
498 ASSERT(rl->r_type == RL_READER);
499 for (len = remove->r_len; len != 0; rl = next) {
500 len -= rl->r_len;
501 if (len) {
502 next = AVL_NEXT(tree, rl);
503 ASSERT(next);
504 ASSERT(rl->r_off + rl->r_len == next->r_off);
505 ASSERT(next->r_cnt);
506 ASSERT(next->r_type == RL_READER);
507 }
508 rl->r_cnt--;
509 if (rl->r_cnt == 0) {
510 avl_remove(tree, rl);
511 if (rl->r_write_wanted) {
512 cv_broadcast(&rl->r_wr_cv);
513 cv_destroy(&rl->r_wr_cv);
514 }
515 if (rl->r_read_wanted) {
516 cv_broadcast(&rl->r_rd_cv);
517 cv_destroy(&rl->r_rd_cv);
518 }
519 kmem_free(rl, sizeof (rl_t));
520 }
521 }
522 }
523 kmem_free(remove, sizeof (rl_t));
524 }
525
526 /*
527 * Unlock range and destroy range lock structure.
528 */
529 void
530 zfs_range_unlock(rl_t *rl)
531 {
532 znode_t *zp = rl->r_zp;
533
534 ASSERT(rl->r_type == RL_WRITER || rl->r_type == RL_READER);
535 ASSERT(rl->r_cnt == 1 || rl->r_cnt == 0);
536 ASSERT(!rl->r_proxy);
537
538 mutex_enter(&zp->z_range_lock);
539 if (rl->r_type == RL_WRITER) {
540 /* writer locks can't be shared or split */
541 avl_remove(&zp->z_range_avl, rl);
542 mutex_exit(&zp->z_range_lock);
543 if (rl->r_write_wanted) {
544 cv_broadcast(&rl->r_wr_cv);
545 cv_destroy(&rl->r_wr_cv);
546 }
547 if (rl->r_read_wanted) {
548 cv_broadcast(&rl->r_rd_cv);
549 cv_destroy(&rl->r_rd_cv);
550 }
551 kmem_free(rl, sizeof (rl_t));
552 } else {
553 /*
554 * lock may be shared, let zfs_range_unlock_reader()
555 * release the lock and free the rl_t
556 */
557 zfs_range_unlock_reader(zp, rl);
558 mutex_exit(&zp->z_range_lock);
559 }
560 }
561
562 /*
563 * Reduce range locked as RL_WRITER from whole file to specified range.
564 * Asserts the whole file is exclusivly locked and so there's only one
565 * entry in the tree.
566 */
567 void
568 zfs_range_reduce(rl_t *rl, uint64_t off, uint64_t len)
569 {
570 znode_t *zp = rl->r_zp;
571
572 /* Ensure there are no other locks */
573 ASSERT(avl_numnodes(&zp->z_range_avl) == 1);
574 ASSERT(rl->r_off == 0);
575 ASSERT(rl->r_type == RL_WRITER);
576 ASSERT(!rl->r_proxy);
577 ASSERT3U(rl->r_len, ==, UINT64_MAX);
578 ASSERT3U(rl->r_cnt, ==, 1);
579
580 mutex_enter(&zp->z_range_lock);
581 rl->r_off = off;
582 rl->r_len = len;
583 mutex_exit(&zp->z_range_lock);
584 if (rl->r_write_wanted)
585 cv_broadcast(&rl->r_wr_cv);
586 if (rl->r_read_wanted)
587 cv_broadcast(&rl->r_rd_cv);
588 }
589
590 /*
591 * AVL comparison function used to order range locks
592 * Locks are ordered on the start offset of the range.
593 */
594 int
595 zfs_range_compare(const void *arg1, const void *arg2)
596 {
597 const rl_t *rl1 = arg1;
598 const rl_t *rl2 = arg2;
599
600 if (rl1->r_off > rl2->r_off)
601 return (1);
602 if (rl1->r_off < rl2->r_off)
603 return (-1);
604 return (0);
605 }