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 /*
27 * This file contains the code to implement file range locking in
28 * ZFS, although there isn't much specific to ZFS (all that comes to mind
29 * support for growing the blocksize).
30 *
31 * Interface
32 * ---------
33 * Defined in zfs_rlock.h but essentially:
34 * rl = zfs_range_lock(zp, off, len, lock_type);
35 * zfs_range_unlock(rl);
36 * zfs_range_reduce(rl, off, len);
37 *
38 * AVL tree
39 * --------
40 * An AVL tree is used to maintain the state of the existing ranges
41 * that are locked for exclusive (writer) or shared (reader) use.
42 * The starting range offset is used for searching and sorting the tree.
43 *
44 * Common case
45 * -----------
46 * The (hopefully) usual case is of no overlaps or contention for
47 * locks. On entry to zfs_lock_range() a rl_t is allocated; the tree
48 * searched that finds no overlap, and *this* rl_t is placed in the tree.
49 *
50 * Overlaps/Reference counting/Proxy locks
51 * ---------------------------------------
52 * The avl code only allows one node at a particular offset. Also it's very
53 * inefficient to search through all previous entries looking for overlaps
54 * (because the very 1st in the ordered list might be at offset 0 but
55 * cover the whole file).
56 * So this implementation uses reference counts and proxy range locks.
57 * Firstly, only reader locks use reference counts and proxy locks,
58 * because writer locks are exclusive.
59 * When a reader lock overlaps with another then a proxy lock is created
60 * for that range and replaces the original lock. If the overlap
61 * is exact then the reference count of the proxy is simply incremented.
62 * Otherwise, the proxy lock is split into smaller lock ranges and
63 * new proxy locks created for non overlapping ranges.
64 * The reference counts are adjusted accordingly.
65 * Meanwhile, the orginal lock is kept around (this is the callers handle)
66 * and its offset and length are used when releasing the lock.
67 *
68 * Thread coordination
69 * -------------------
70 * In order to make wakeups efficient and to ensure multiple continuous
71 * readers on a range don't starve a writer for the same range lock,
72 * two condition variables are allocated in each rl_t.
73 * If a writer (or reader) can't get a range it initialises the writer
74 * (or reader) cv; sets a flag saying there's a writer (or reader) waiting;
75 * and waits on that cv. When a thread unlocks that range it wakes up all
76 * writers then all readers before destroying the lock.
77 *
78 * Append mode writes
79 * ------------------
80 * Append mode writes need to lock a range at the end of a file.
81 * The offset of the end of the file is determined under the
82 * range locking mutex, and the lock type converted from RL_APPEND to
83 * RL_WRITER and the range locked.
84 *
85 * Grow block handling
86 * -------------------
87 * ZFS supports multiple block sizes currently upto 128K. The smallest
88 * block size is used for the file which is grown as needed. During this
89 * growth all other writers and readers must be excluded.
90 * So if the block size needs to be grown then the whole file is
91 * exclusively locked, then later the caller will reduce the lock
92 * range to just the range to be written using zfs_reduce_range.
93 */
94
95 #include <sys/zfs_rlock.h>
96
97 /*
98 * Check if a write lock can be grabbed, or wait and recheck until available.
99 */
100 static void
101 zfs_range_lock_writer(znode_t *zp, rl_t *new)
102 {
103 avl_tree_t *tree = &zp->z_range_avl;
104 rl_t *rl;
105 avl_index_t where;
106 uint64_t end_size;
107 uint64_t off = new->r_off;
108 uint64_t len = new->r_len;
109
110 for (;;) {
111 /*
112 * Range locking is also used by zvol and uses a
113 * dummied up znode. However, for zvol, we don't need to
114 * append or grow blocksize, and besides we don't have
115 * a "sa" data or z_zfsvfs - so skip that processing.
116 *
117 * Yes, this is ugly, and would be solved by not handling
118 * grow or append in range lock code. If that was done then
119 * we could make the range locking code generically available
120 * to other non-zfs consumers.
121 */
122 if (zp->z_vnode) { /* caller is ZPL */
123 /*
124 * If in append mode pick up the current end of file.
125 * This is done under z_range_lock to avoid races.
126 */
127 if (new->r_type == RL_APPEND)
128 new->r_off = zp->z_size;
129
130 /*
131 * If we need to grow the block size then grab the whole
132 * file range. This is also done under z_range_lock to
133 * avoid races.
134 */
135 end_size = MAX(zp->z_size, new->r_off + len);
136 if (end_size > zp->z_blksz && (!ISP2(zp->z_blksz) ||
137 zp->z_blksz < zp->z_zfsvfs->z_max_blksz)) {
138 new->r_off = 0;
139 new->r_len = UINT64_MAX;
140 }
141 }
142
143 /*
144 * First check for the usual case of no locks
145 */
146 if (avl_numnodes(tree) == 0) {
147 new->r_type = RL_WRITER; /* convert to writer */
148 avl_add(tree, new);
149 return;
150 }
151
152 /*
153 * Look for any locks in the range.
154 */
155 rl = avl_find(tree, new, &where);
156 if (rl)
157 goto wait; /* already locked at same offset */
158
159 rl = (rl_t *)avl_nearest(tree, where, AVL_AFTER);
160 if (rl && (rl->r_off < new->r_off + new->r_len))
161 goto wait;
162
163 rl = (rl_t *)avl_nearest(tree, where, AVL_BEFORE);
164 if (rl && rl->r_off + rl->r_len > new->r_off)
165 goto wait;
166
167 new->r_type = RL_WRITER; /* convert possible RL_APPEND */
168 avl_insert(tree, new, where);
169 return;
170 wait:
171 if (!rl->r_write_wanted) {
172 cv_init(&rl->r_wr_cv, NULL, CV_DEFAULT, NULL);
173 rl->r_write_wanted = B_TRUE;
174 }
175 cv_wait(&rl->r_wr_cv, &zp->z_range_lock);
176
177 /* reset to original */
178 new->r_off = off;
179 new->r_len = len;
180 }
181 }
182
183 /*
184 * If this is an original (non-proxy) lock then replace it by
185 * a proxy and return the proxy.
186 */
187 static rl_t *
188 zfs_range_proxify(avl_tree_t *tree, rl_t *rl)
189 {
190 rl_t *proxy;
191
192 if (rl->r_proxy)
193 return (rl); /* already a proxy */
194
195 ASSERT3U(rl->r_cnt, ==, 1);
196 ASSERT(rl->r_write_wanted == B_FALSE);
197 ASSERT(rl->r_read_wanted == B_FALSE);
198 avl_remove(tree, rl);
199 rl->r_cnt = 0;
200
201 /* create a proxy range lock */
202 proxy = kmem_alloc(sizeof (rl_t), KM_SLEEP);
203 proxy->r_off = rl->r_off;
204 proxy->r_len = rl->r_len;
205 proxy->r_cnt = 1;
206 proxy->r_type = RL_READER;
207 proxy->r_proxy = B_TRUE;
208 proxy->r_write_wanted = B_FALSE;
209 proxy->r_read_wanted = B_FALSE;
210 avl_add(tree, proxy);
211
212 return (proxy);
213 }
214
215 /*
216 * Split the range lock at the supplied offset
217 * returning the *front* proxy.
218 */
219 static rl_t *
220 zfs_range_split(avl_tree_t *tree, rl_t *rl, uint64_t off)
221 {
222 rl_t *front, *rear;
223
224 ASSERT3U(rl->r_len, >, 1);
225 ASSERT3U(off, >, rl->r_off);
226 ASSERT3U(off, <, rl->r_off + rl->r_len);
227 ASSERT(rl->r_write_wanted == B_FALSE);
228 ASSERT(rl->r_read_wanted == B_FALSE);
229
230 /* create the rear proxy range lock */
231 rear = kmem_alloc(sizeof (rl_t), KM_SLEEP);
232 rear->r_off = off;
233 rear->r_len = rl->r_off + rl->r_len - off;
234 rear->r_cnt = rl->r_cnt;
235 rear->r_type = RL_READER;
236 rear->r_proxy = B_TRUE;
237 rear->r_write_wanted = B_FALSE;
238 rear->r_read_wanted = B_FALSE;
239
240 front = zfs_range_proxify(tree, rl);
241 front->r_len = off - rl->r_off;
242
243 avl_insert_here(tree, rear, front, AVL_AFTER);
244 return (front);
245 }
246
247 /*
248 * Create and add a new proxy range lock for the supplied range.
249 */
250 static void
251 zfs_range_new_proxy(avl_tree_t *tree, uint64_t off, uint64_t len)
252 {
253 rl_t *rl;
254
255 ASSERT(len);
256 rl = kmem_alloc(sizeof (rl_t), KM_SLEEP);
257 rl->r_off = off;
258 rl->r_len = len;
259 rl->r_cnt = 1;
260 rl->r_type = RL_READER;
261 rl->r_proxy = B_TRUE;
262 rl->r_write_wanted = B_FALSE;
263 rl->r_read_wanted = B_FALSE;
264 avl_add(tree, rl);
265 }
266
267 static void
268 zfs_range_add_reader(avl_tree_t *tree, rl_t *new, rl_t *prev, avl_index_t where)
269 {
270 rl_t *next;
271 uint64_t off = new->r_off;
272 uint64_t len = new->r_len;
273
274 /*
275 * prev arrives either:
276 * - pointing to an entry at the same offset
277 * - pointing to the entry with the closest previous offset whose
278 * range may overlap with the new range
279 * - null, if there were no ranges starting before the new one
280 */
281 if (prev) {
282 if (prev->r_off + prev->r_len <= off) {
283 prev = NULL;
284 } else if (prev->r_off != off) {
285 /*
286 * convert to proxy if needed then
287 * split this entry and bump ref count
288 */
289 prev = zfs_range_split(tree, prev, off);
290 prev = AVL_NEXT(tree, prev); /* move to rear range */
291 }
292 }
293 ASSERT((prev == NULL) || (prev->r_off == off));
294
295 if (prev)
296 next = prev;
297 else
298 next = (rl_t *)avl_nearest(tree, where, AVL_AFTER);
299
300 if (next == NULL || off + len <= next->r_off) {
301 /* no overlaps, use the original new rl_t in the tree */
302 avl_insert(tree, new, where);
303 return;
304 }
305
306 if (off < next->r_off) {
307 /* Add a proxy for initial range before the overlap */
308 zfs_range_new_proxy(tree, off, next->r_off - off);
309 }
310
311 new->r_cnt = 0; /* will use proxies in tree */
312 /*
313 * We now search forward through the ranges, until we go past the end
314 * of the new range. For each entry we make it a proxy if it
315 * isn't already, then bump its reference count. If there's any
316 * gaps between the ranges then we create a new proxy range.
317 */
318 for (prev = NULL; next; prev = next, next = AVL_NEXT(tree, next)) {
319 if (off + len <= next->r_off)
320 break;
321 if (prev && prev->r_off + prev->r_len < next->r_off) {
322 /* there's a gap */
323 ASSERT3U(next->r_off, >, prev->r_off + prev->r_len);
324 zfs_range_new_proxy(tree, prev->r_off + prev->r_len,
325 next->r_off - (prev->r_off + prev->r_len));
326 }
327 if (off + len == next->r_off + next->r_len) {
328 /* exact overlap with end */
329 next = zfs_range_proxify(tree, next);
330 next->r_cnt++;
331 return;
332 }
333 if (off + len < next->r_off + next->r_len) {
334 /* new range ends in the middle of this block */
335 next = zfs_range_split(tree, next, off + len);
336 next->r_cnt++;
337 return;
338 }
339 ASSERT3U(off + len, >, next->r_off + next->r_len);
340 next = zfs_range_proxify(tree, next);
341 next->r_cnt++;
342 }
343
344 /* Add the remaining end range. */
345 zfs_range_new_proxy(tree, prev->r_off + prev->r_len,
346 (off + len) - (prev->r_off + prev->r_len));
347 }
348
349 /*
350 * Check if a reader lock can be grabbed, or wait and recheck until available.
351 */
352 static void
353 zfs_range_lock_reader(znode_t *zp, rl_t *new)
354 {
355 avl_tree_t *tree = &zp->z_range_avl;
356 rl_t *prev, *next;
357 avl_index_t where;
358 uint64_t off = new->r_off;
359 uint64_t len = new->r_len;
360
361 /*
362 * Look for any writer locks in the range.
363 */
364 retry:
365 prev = avl_find(tree, new, &where);
366 if (prev == NULL)
367 prev = (rl_t *)avl_nearest(tree, where, AVL_BEFORE);
368
369 /*
370 * Check the previous range for a writer lock overlap.
371 */
372 if (prev && (off < prev->r_off + prev->r_len)) {
373 if ((prev->r_type == RL_WRITER) || (prev->r_write_wanted)) {
374 if (!prev->r_read_wanted) {
375 cv_init(&prev->r_rd_cv, NULL, CV_DEFAULT, NULL);
376 prev->r_read_wanted = B_TRUE;
377 }
378 cv_wait(&prev->r_rd_cv, &zp->z_range_lock);
379 goto retry;
380 }
381 if (off + len < prev->r_off + prev->r_len)
382 goto got_lock;
383 }
384
385 /*
386 * Search through the following ranges to see if there's
387 * write lock any overlap.
388 */
389 if (prev)
390 next = AVL_NEXT(tree, prev);
391 else
392 next = (rl_t *)avl_nearest(tree, where, AVL_AFTER);
393 for (; next; next = AVL_NEXT(tree, next)) {
394 if (off + len <= next->r_off)
395 goto got_lock;
396 if ((next->r_type == RL_WRITER) || (next->r_write_wanted)) {
397 if (!next->r_read_wanted) {
398 cv_init(&next->r_rd_cv, NULL, CV_DEFAULT, NULL);
399 next->r_read_wanted = B_TRUE;
400 }
401 cv_wait(&next->r_rd_cv, &zp->z_range_lock);
402 goto retry;
403 }
404 if (off + len <= next->r_off + next->r_len)
405 goto got_lock;
406 }
407
408 got_lock:
409 /*
410 * Add the read lock, which may involve splitting existing
411 * locks and bumping ref counts (r_cnt).
412 */
413 zfs_range_add_reader(tree, new, prev, where);
414 }
415
416 /*
417 * Lock a range (offset, length) as either shared (RL_READER)
418 * or exclusive (RL_WRITER). Returns the range lock structure
419 * for later unlocking or reduce range (if entire file
420 * previously locked as RL_WRITER).
421 */
422 rl_t *
423 zfs_range_lock(znode_t *zp, uint64_t off, uint64_t len, rl_type_t type)
424 {
425 rl_t *new;
426
427 ASSERT(type == RL_READER || type == RL_WRITER || type == RL_APPEND);
428
429 new = kmem_alloc(sizeof (rl_t), KM_SLEEP);
430 new->r_zp = zp;
431 new->r_off = off;
432 if (len + off < off) /* overflow */
433 len = UINT64_MAX - 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 }