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 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
24 *
25 */
26
27 /*
28 * Copyright (c) 2012 by Delphix. All rights reserved.
29 */
30
31 #include <sys/zfs_context.h>
32 #include <sys/spa.h>
33 #include <sys/dmu.h>
34 #include <sys/zio.h>
35 #include <sys/space_map.h>
36
37 /*
38 * Space map routines.
39 * NOTE: caller is responsible for all locking.
40 */
41 static int
42 space_map_seg_compare(const void *x1, const void *x2)
43 {
44 const space_seg_t *s1 = x1;
45 const space_seg_t *s2 = x2;
46
47 if (s1->ss_start < s2->ss_start) {
48 if (s1->ss_end > s2->ss_start)
49 return (0);
50 return (-1);
51 }
52 if (s1->ss_start > s2->ss_start) {
53 if (s1->ss_start < s2->ss_end)
54 return (0);
55 return (1);
56 }
57 return (0);
58 }
59
60 void
61 space_map_create(space_map_t *sm, uint64_t start, uint64_t size, uint8_t shift,
62 kmutex_t *lp)
63 {
64 bzero(sm, sizeof (*sm));
65
66 cv_init(&sm->sm_load_cv, NULL, CV_DEFAULT, NULL);
67
68 avl_create(&sm->sm_root, space_map_seg_compare,
69 sizeof (space_seg_t), offsetof(struct space_seg, ss_node));
70
71 sm->sm_start = start;
72 sm->sm_size = size;
73 sm->sm_shift = shift;
74 sm->sm_lock = lp;
75 }
76
77 void
78 space_map_destroy(space_map_t *sm)
79 {
80 ASSERT(!sm->sm_loaded && !sm->sm_loading);
81 VERIFY0(sm->sm_space);
82 avl_destroy(&sm->sm_root);
83 cv_destroy(&sm->sm_load_cv);
84 }
85
86 void
87 space_map_add(space_map_t *sm, uint64_t start, uint64_t size)
88 {
89 avl_index_t where;
90 space_seg_t ssearch, *ss_before, *ss_after, *ss;
91 uint64_t end = start + size;
92 int merge_before, merge_after;
93
94 ASSERT(MUTEX_HELD(sm->sm_lock));
95 VERIFY(size != 0);
96 VERIFY3U(start, >=, sm->sm_start);
97 VERIFY3U(end, <=, sm->sm_start + sm->sm_size);
98 VERIFY(sm->sm_space + size <= sm->sm_size);
99 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
100 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
101
102 ssearch.ss_start = start;
103 ssearch.ss_end = end;
104 ss = avl_find(&sm->sm_root, &ssearch, &where);
105
106 if (ss != NULL && ss->ss_start <= start && ss->ss_end >= end) {
107 zfs_panic_recover("zfs: allocating allocated segment"
108 "(offset=%llu size=%llu)\n",
109 (longlong_t)start, (longlong_t)size);
110 return;
111 }
112
113 /* Make sure we don't overlap with either of our neighbors */
114 VERIFY(ss == NULL);
115
116 ss_before = avl_nearest(&sm->sm_root, where, AVL_BEFORE);
117 ss_after = avl_nearest(&sm->sm_root, where, AVL_AFTER);
118
119 merge_before = (ss_before != NULL && ss_before->ss_end == start);
120 merge_after = (ss_after != NULL && ss_after->ss_start == end);
121
122 if (merge_before && merge_after) {
123 avl_remove(&sm->sm_root, ss_before);
124 if (sm->sm_pp_root) {
125 avl_remove(sm->sm_pp_root, ss_before);
126 avl_remove(sm->sm_pp_root, ss_after);
127 }
128 ss_after->ss_start = ss_before->ss_start;
129 kmem_free(ss_before, sizeof (*ss_before));
130 ss = ss_after;
131 } else if (merge_before) {
132 ss_before->ss_end = end;
133 if (sm->sm_pp_root)
134 avl_remove(sm->sm_pp_root, ss_before);
135 ss = ss_before;
136 } else if (merge_after) {
137 ss_after->ss_start = start;
138 if (sm->sm_pp_root)
139 avl_remove(sm->sm_pp_root, ss_after);
140 ss = ss_after;
141 } else {
142 ss = kmem_alloc(sizeof (*ss), KM_SLEEP);
143 ss->ss_start = start;
144 ss->ss_end = end;
145 avl_insert(&sm->sm_root, ss, where);
146 }
147
148 if (sm->sm_pp_root)
149 avl_add(sm->sm_pp_root, ss);
150
151 sm->sm_space += size;
152 }
153
154 void
155 space_map_remove(space_map_t *sm, uint64_t start, uint64_t size)
156 {
157 avl_index_t where;
158 space_seg_t ssearch, *ss, *newseg;
159 uint64_t end = start + size;
160 int left_over, right_over;
161
162 ASSERT(MUTEX_HELD(sm->sm_lock));
163 VERIFY(size != 0);
164 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
165 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
166
167 ssearch.ss_start = start;
168 ssearch.ss_end = end;
169 ss = avl_find(&sm->sm_root, &ssearch, &where);
170
171 /* Make sure we completely overlap with someone */
172 if (ss == NULL) {
173 zfs_panic_recover("zfs: freeing free segment "
174 "(offset=%llu size=%llu)",
175 (longlong_t)start, (longlong_t)size);
176 return;
177 }
178 VERIFY3U(ss->ss_start, <=, start);
179 VERIFY3U(ss->ss_end, >=, end);
180 VERIFY(sm->sm_space - size <= sm->sm_size);
181
182 left_over = (ss->ss_start != start);
183 right_over = (ss->ss_end != end);
184
185 if (sm->sm_pp_root)
186 avl_remove(sm->sm_pp_root, ss);
187
188 if (left_over && right_over) {
189 newseg = kmem_alloc(sizeof (*newseg), KM_SLEEP);
190 newseg->ss_start = end;
191 newseg->ss_end = ss->ss_end;
192 ss->ss_end = start;
193 avl_insert_here(&sm->sm_root, newseg, ss, AVL_AFTER);
194 if (sm->sm_pp_root)
195 avl_add(sm->sm_pp_root, newseg);
196 } else if (left_over) {
197 ss->ss_end = start;
198 } else if (right_over) {
199 ss->ss_start = end;
200 } else {
201 avl_remove(&sm->sm_root, ss);
202 kmem_free(ss, sizeof (*ss));
203 ss = NULL;
204 }
205
206 if (sm->sm_pp_root && ss != NULL)
207 avl_add(sm->sm_pp_root, ss);
208
209 sm->sm_space -= size;
210 }
211
212 boolean_t
213 space_map_contains(space_map_t *sm, uint64_t start, uint64_t size)
214 {
215 avl_index_t where;
216 space_seg_t ssearch, *ss;
217 uint64_t end = start + size;
218
219 ASSERT(MUTEX_HELD(sm->sm_lock));
220 VERIFY(size != 0);
221 VERIFY(P2PHASE(start, 1ULL << sm->sm_shift) == 0);
222 VERIFY(P2PHASE(size, 1ULL << sm->sm_shift) == 0);
223
224 ssearch.ss_start = start;
225 ssearch.ss_end = end;
226 ss = avl_find(&sm->sm_root, &ssearch, &where);
227
228 return (ss != NULL && ss->ss_start <= start && ss->ss_end >= end);
229 }
230
231 void
232 space_map_vacate(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
233 {
234 space_seg_t *ss;
235 void *cookie = NULL;
236
237 ASSERT(MUTEX_HELD(sm->sm_lock));
238
239 while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
240 if (func != NULL)
241 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
242 kmem_free(ss, sizeof (*ss));
243 }
244 sm->sm_space = 0;
245 }
246
247 void
248 space_map_walk(space_map_t *sm, space_map_func_t *func, space_map_t *mdest)
249 {
250 space_seg_t *ss;
251
252 ASSERT(MUTEX_HELD(sm->sm_lock));
253
254 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
255 func(mdest, ss->ss_start, ss->ss_end - ss->ss_start);
256 }
257
258 /*
259 * Wait for any in-progress space_map_load() to complete.
260 */
261 void
262 space_map_load_wait(space_map_t *sm)
263 {
264 ASSERT(MUTEX_HELD(sm->sm_lock));
265
266 while (sm->sm_loading) {
267 ASSERT(!sm->sm_loaded);
268 cv_wait(&sm->sm_load_cv, sm->sm_lock);
269 }
270 }
271
272 /*
273 * Note: space_map_load() will drop sm_lock across dmu_read() calls.
274 * The caller must be OK with this.
275 */
276 int
277 space_map_load(space_map_t *sm, space_map_ops_t *ops, uint8_t maptype,
278 space_map_obj_t *smo, objset_t *os)
279 {
280 uint64_t *entry, *entry_map, *entry_map_end;
281 uint64_t bufsize, size, offset, end, space;
282 uint64_t mapstart = sm->sm_start;
283 int error = 0;
284
285 ASSERT(MUTEX_HELD(sm->sm_lock));
286 ASSERT(!sm->sm_loaded);
287 ASSERT(!sm->sm_loading);
288
289 sm->sm_loading = B_TRUE;
290 end = smo->smo_objsize;
291 space = smo->smo_alloc;
292
293 ASSERT(sm->sm_ops == NULL);
294 VERIFY0(sm->sm_space);
295
296 if (maptype == SM_FREE) {
297 space_map_add(sm, sm->sm_start, sm->sm_size);
298 space = sm->sm_size - space;
299 }
300
301 bufsize = 1ULL << SPACE_MAP_BLOCKSHIFT;
302 entry_map = zio_buf_alloc(bufsize);
303
304 mutex_exit(sm->sm_lock);
305 if (end > bufsize)
306 dmu_prefetch(os, smo->smo_object, bufsize, end - bufsize);
307 mutex_enter(sm->sm_lock);
308
309 for (offset = 0; offset < end; offset += bufsize) {
310 size = MIN(end - offset, bufsize);
311 VERIFY(P2PHASE(size, sizeof (uint64_t)) == 0);
312 VERIFY(size != 0);
313
314 dprintf("object=%llu offset=%llx size=%llx\n",
315 smo->smo_object, offset, size);
316
317 mutex_exit(sm->sm_lock);
318 error = dmu_read(os, smo->smo_object, offset, size, entry_map,
319 DMU_READ_PREFETCH);
320 mutex_enter(sm->sm_lock);
321 if (error != 0)
322 break;
323
324 entry_map_end = entry_map + (size / sizeof (uint64_t));
325 for (entry = entry_map; entry < entry_map_end; entry++) {
326 uint64_t e = *entry;
327
328 if (SM_DEBUG_DECODE(e)) /* Skip debug entries */
329 continue;
330
331 (SM_TYPE_DECODE(e) == maptype ?
332 space_map_add : space_map_remove)(sm,
333 (SM_OFFSET_DECODE(e) << sm->sm_shift) + mapstart,
334 SM_RUN_DECODE(e) << sm->sm_shift);
335 }
336 }
337
338 if (error == 0) {
339 VERIFY3U(sm->sm_space, ==, space);
340
341 sm->sm_loaded = B_TRUE;
342 sm->sm_ops = ops;
343 if (ops != NULL)
344 ops->smop_load(sm);
345 } else {
346 space_map_vacate(sm, NULL, NULL);
347 }
348
349 zio_buf_free(entry_map, bufsize);
350
351 sm->sm_loading = B_FALSE;
352
353 cv_broadcast(&sm->sm_load_cv);
354
355 return (error);
356 }
357
358 void
359 space_map_unload(space_map_t *sm)
360 {
361 ASSERT(MUTEX_HELD(sm->sm_lock));
362
363 if (sm->sm_loaded && sm->sm_ops != NULL)
364 sm->sm_ops->smop_unload(sm);
365
366 sm->sm_loaded = B_FALSE;
367 sm->sm_ops = NULL;
368
369 space_map_vacate(sm, NULL, NULL);
370 }
371
372 uint64_t
373 space_map_maxsize(space_map_t *sm)
374 {
375 ASSERT(sm->sm_ops != NULL);
376 return (sm->sm_ops->smop_max(sm));
377 }
378
379 uint64_t
380 space_map_alloc(space_map_t *sm, uint64_t size)
381 {
382 uint64_t start;
383
384 start = sm->sm_ops->smop_alloc(sm, size);
385 if (start != -1ULL)
386 space_map_remove(sm, start, size);
387 return (start);
388 }
389
390 void
391 space_map_claim(space_map_t *sm, uint64_t start, uint64_t size)
392 {
393 sm->sm_ops->smop_claim(sm, start, size);
394 space_map_remove(sm, start, size);
395 }
396
397 void
398 space_map_free(space_map_t *sm, uint64_t start, uint64_t size)
399 {
400 space_map_add(sm, start, size);
401 sm->sm_ops->smop_free(sm, start, size);
402 }
403
404 /*
405 * Note: space_map_sync() will drop sm_lock across dmu_write() calls.
406 */
407 void
408 space_map_sync(space_map_t *sm, uint8_t maptype,
409 space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
410 {
411 spa_t *spa = dmu_objset_spa(os);
412 void *cookie = NULL;
413 space_seg_t *ss;
414 uint64_t bufsize, start, size, run_len;
415 uint64_t *entry, *entry_map, *entry_map_end;
416
417 ASSERT(MUTEX_HELD(sm->sm_lock));
418
419 if (sm->sm_space == 0)
420 return;
421
422 dprintf("object %4llu, txg %llu, pass %d, %c, count %lu, space %llx\n",
423 smo->smo_object, dmu_tx_get_txg(tx), spa_sync_pass(spa),
424 maptype == SM_ALLOC ? 'A' : 'F', avl_numnodes(&sm->sm_root),
425 sm->sm_space);
426
427 if (maptype == SM_ALLOC)
428 smo->smo_alloc += sm->sm_space;
429 else
430 smo->smo_alloc -= sm->sm_space;
431
432 bufsize = (8 + avl_numnodes(&sm->sm_root)) * sizeof (uint64_t);
433 bufsize = MIN(bufsize, 1ULL << SPACE_MAP_BLOCKSHIFT);
434 entry_map = zio_buf_alloc(bufsize);
435 entry_map_end = entry_map + (bufsize / sizeof (uint64_t));
436 entry = entry_map;
437
438 *entry++ = SM_DEBUG_ENCODE(1) |
439 SM_DEBUG_ACTION_ENCODE(maptype) |
440 SM_DEBUG_SYNCPASS_ENCODE(spa_sync_pass(spa)) |
441 SM_DEBUG_TXG_ENCODE(dmu_tx_get_txg(tx));
442
443 while ((ss = avl_destroy_nodes(&sm->sm_root, &cookie)) != NULL) {
444 size = ss->ss_end - ss->ss_start;
445 start = (ss->ss_start - sm->sm_start) >> sm->sm_shift;
446
447 sm->sm_space -= size;
448 size >>= sm->sm_shift;
449
450 while (size) {
451 run_len = MIN(size, SM_RUN_MAX);
452
453 if (entry == entry_map_end) {
454 mutex_exit(sm->sm_lock);
455 dmu_write(os, smo->smo_object, smo->smo_objsize,
456 bufsize, entry_map, tx);
457 mutex_enter(sm->sm_lock);
458 smo->smo_objsize += bufsize;
459 entry = entry_map;
460 }
461
462 *entry++ = SM_OFFSET_ENCODE(start) |
463 SM_TYPE_ENCODE(maptype) |
464 SM_RUN_ENCODE(run_len);
465
466 start += run_len;
467 size -= run_len;
468 }
469 kmem_free(ss, sizeof (*ss));
470 }
471
472 if (entry != entry_map) {
473 size = (entry - entry_map) * sizeof (uint64_t);
474 mutex_exit(sm->sm_lock);
475 dmu_write(os, smo->smo_object, smo->smo_objsize,
476 size, entry_map, tx);
477 mutex_enter(sm->sm_lock);
478 smo->smo_objsize += size;
479 }
480
481 zio_buf_free(entry_map, bufsize);
482
483 VERIFY0(sm->sm_space);
484 }
485
486 void
487 space_map_truncate(space_map_obj_t *smo, objset_t *os, dmu_tx_t *tx)
488 {
489 VERIFY(dmu_free_range(os, smo->smo_object, 0, -1ULL, tx) == 0);
490
491 smo->smo_objsize = 0;
492 smo->smo_alloc = 0;
493 }
494
495 /*
496 * Space map reference trees.
497 *
498 * A space map is a collection of integers. Every integer is either
499 * in the map, or it's not. A space map reference tree generalizes
500 * the idea: it allows its members to have arbitrary reference counts,
501 * as opposed to the implicit reference count of 0 or 1 in a space map.
502 * This representation comes in handy when computing the union or
503 * intersection of multiple space maps. For example, the union of
504 * N space maps is the subset of the reference tree with refcnt >= 1.
505 * The intersection of N space maps is the subset with refcnt >= N.
506 *
507 * [It's very much like a Fourier transform. Unions and intersections
508 * are hard to perform in the 'space map domain', so we convert the maps
509 * into the 'reference count domain', where it's trivial, then invert.]
510 *
511 * vdev_dtl_reassess() uses computations of this form to determine
512 * DTL_MISSING and DTL_OUTAGE for interior vdevs -- e.g. a RAID-Z vdev
513 * has an outage wherever refcnt >= vdev_nparity + 1, and a mirror vdev
514 * has an outage wherever refcnt >= vdev_children.
515 */
516 static int
517 space_map_ref_compare(const void *x1, const void *x2)
518 {
519 const space_ref_t *sr1 = x1;
520 const space_ref_t *sr2 = x2;
521
522 if (sr1->sr_offset < sr2->sr_offset)
523 return (-1);
524 if (sr1->sr_offset > sr2->sr_offset)
525 return (1);
526
527 if (sr1 < sr2)
528 return (-1);
529 if (sr1 > sr2)
530 return (1);
531
532 return (0);
533 }
534
535 void
536 space_map_ref_create(avl_tree_t *t)
537 {
538 avl_create(t, space_map_ref_compare,
539 sizeof (space_ref_t), offsetof(space_ref_t, sr_node));
540 }
541
542 void
543 space_map_ref_destroy(avl_tree_t *t)
544 {
545 space_ref_t *sr;
546 void *cookie = NULL;
547
548 while ((sr = avl_destroy_nodes(t, &cookie)) != NULL)
549 kmem_free(sr, sizeof (*sr));
550
551 avl_destroy(t);
552 }
553
554 static void
555 space_map_ref_add_node(avl_tree_t *t, uint64_t offset, int64_t refcnt)
556 {
557 space_ref_t *sr;
558
559 sr = kmem_alloc(sizeof (*sr), KM_SLEEP);
560 sr->sr_offset = offset;
561 sr->sr_refcnt = refcnt;
562
563 avl_add(t, sr);
564 }
565
566 void
567 space_map_ref_add_seg(avl_tree_t *t, uint64_t start, uint64_t end,
568 int64_t refcnt)
569 {
570 space_map_ref_add_node(t, start, refcnt);
571 space_map_ref_add_node(t, end, -refcnt);
572 }
573
574 /*
575 * Convert (or add) a space map into a reference tree.
576 */
577 void
578 space_map_ref_add_map(avl_tree_t *t, space_map_t *sm, int64_t refcnt)
579 {
580 space_seg_t *ss;
581
582 ASSERT(MUTEX_HELD(sm->sm_lock));
583
584 for (ss = avl_first(&sm->sm_root); ss; ss = AVL_NEXT(&sm->sm_root, ss))
585 space_map_ref_add_seg(t, ss->ss_start, ss->ss_end, refcnt);
586 }
587
588 /*
589 * Convert a reference tree into a space map. The space map will contain
590 * all members of the reference tree for which refcnt >= minref.
591 */
592 void
593 space_map_ref_generate_map(avl_tree_t *t, space_map_t *sm, int64_t minref)
594 {
595 uint64_t start = -1ULL;
596 int64_t refcnt = 0;
597 space_ref_t *sr;
598
599 ASSERT(MUTEX_HELD(sm->sm_lock));
600
601 space_map_vacate(sm, NULL, NULL);
602
603 for (sr = avl_first(t); sr != NULL; sr = AVL_NEXT(t, sr)) {
604 refcnt += sr->sr_refcnt;
605 if (refcnt >= minref) {
606 if (start == -1ULL) {
607 start = sr->sr_offset;
608 }
609 } else {
610 if (start != -1ULL) {
611 uint64_t end = sr->sr_offset;
612 ASSERT(start <= end);
613 if (end > start)
614 space_map_add(sm, start, end - start);
615 start = -1ULL;
616 }
617 }
618 }
619 ASSERT(refcnt == 0);
620 ASSERT(start == -1ULL);
621 }