xref: /titanic_44/usr/src/uts/common/fs/zfs/metaslab.c (revision a576ab5b6e08c47732b3dedca9eaa8a8cbb85720)
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 #include <sys/zfs_context.h>
29 #include <sys/spa_impl.h>
30 #include <sys/dmu.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/space_map.h>
33 #include <sys/metaslab_impl.h>
34 #include <sys/vdev_impl.h>
35 #include <sys/zio.h>
36 
37 uint64_t metaslab_aliquot = 512ULL << 10;
38 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1;	/* force gang blocks */
39 
40 /*
41  * ==========================================================================
42  * Metaslab classes
43  * ==========================================================================
44  */
45 metaslab_class_t *
46 metaslab_class_create(void)
47 {
48 	metaslab_class_t *mc;
49 
50 	mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
51 
52 	mc->mc_rotor = NULL;
53 
54 	return (mc);
55 }
56 
57 void
58 metaslab_class_destroy(metaslab_class_t *mc)
59 {
60 	metaslab_group_t *mg;
61 
62 	while ((mg = mc->mc_rotor) != NULL) {
63 		metaslab_class_remove(mc, mg);
64 		metaslab_group_destroy(mg);
65 	}
66 
67 	kmem_free(mc, sizeof (metaslab_class_t));
68 }
69 
70 void
71 metaslab_class_add(metaslab_class_t *mc, metaslab_group_t *mg)
72 {
73 	metaslab_group_t *mgprev, *mgnext;
74 
75 	ASSERT(mg->mg_class == NULL);
76 
77 	if ((mgprev = mc->mc_rotor) == NULL) {
78 		mg->mg_prev = mg;
79 		mg->mg_next = mg;
80 	} else {
81 		mgnext = mgprev->mg_next;
82 		mg->mg_prev = mgprev;
83 		mg->mg_next = mgnext;
84 		mgprev->mg_next = mg;
85 		mgnext->mg_prev = mg;
86 	}
87 	mc->mc_rotor = mg;
88 	mg->mg_class = mc;
89 }
90 
91 void
92 metaslab_class_remove(metaslab_class_t *mc, metaslab_group_t *mg)
93 {
94 	metaslab_group_t *mgprev, *mgnext;
95 
96 	ASSERT(mg->mg_class == mc);
97 
98 	mgprev = mg->mg_prev;
99 	mgnext = mg->mg_next;
100 
101 	if (mg == mgnext) {
102 		mc->mc_rotor = NULL;
103 	} else {
104 		mc->mc_rotor = mgnext;
105 		mgprev->mg_next = mgnext;
106 		mgnext->mg_prev = mgprev;
107 	}
108 
109 	mg->mg_prev = NULL;
110 	mg->mg_next = NULL;
111 	mg->mg_class = NULL;
112 }
113 
114 /*
115  * ==========================================================================
116  * Metaslab groups
117  * ==========================================================================
118  */
119 static int
120 metaslab_compare(const void *x1, const void *x2)
121 {
122 	const metaslab_t *m1 = x1;
123 	const metaslab_t *m2 = x2;
124 
125 	if (m1->ms_weight < m2->ms_weight)
126 		return (1);
127 	if (m1->ms_weight > m2->ms_weight)
128 		return (-1);
129 
130 	/*
131 	 * If the weights are identical, use the offset to force uniqueness.
132 	 */
133 	if (m1->ms_map.sm_start < m2->ms_map.sm_start)
134 		return (-1);
135 	if (m1->ms_map.sm_start > m2->ms_map.sm_start)
136 		return (1);
137 
138 	ASSERT3P(m1, ==, m2);
139 
140 	return (0);
141 }
142 
143 metaslab_group_t *
144 metaslab_group_create(metaslab_class_t *mc, vdev_t *vd)
145 {
146 	metaslab_group_t *mg;
147 
148 	mg = kmem_zalloc(sizeof (metaslab_group_t), KM_SLEEP);
149 	mutex_init(&mg->mg_lock, NULL, MUTEX_DEFAULT, NULL);
150 	avl_create(&mg->mg_metaslab_tree, metaslab_compare,
151 	    sizeof (metaslab_t), offsetof(struct metaslab, ms_group_node));
152 	mg->mg_aliquot = metaslab_aliquot * MAX(1, vd->vdev_children);
153 	mg->mg_vd = vd;
154 	metaslab_class_add(mc, mg);
155 
156 	return (mg);
157 }
158 
159 void
160 metaslab_group_destroy(metaslab_group_t *mg)
161 {
162 	avl_destroy(&mg->mg_metaslab_tree);
163 	mutex_destroy(&mg->mg_lock);
164 	kmem_free(mg, sizeof (metaslab_group_t));
165 }
166 
167 static void
168 metaslab_group_add(metaslab_group_t *mg, metaslab_t *msp)
169 {
170 	mutex_enter(&mg->mg_lock);
171 	ASSERT(msp->ms_group == NULL);
172 	msp->ms_group = mg;
173 	msp->ms_weight = 0;
174 	avl_add(&mg->mg_metaslab_tree, msp);
175 	mutex_exit(&mg->mg_lock);
176 }
177 
178 static void
179 metaslab_group_remove(metaslab_group_t *mg, metaslab_t *msp)
180 {
181 	mutex_enter(&mg->mg_lock);
182 	ASSERT(msp->ms_group == mg);
183 	avl_remove(&mg->mg_metaslab_tree, msp);
184 	msp->ms_group = NULL;
185 	mutex_exit(&mg->mg_lock);
186 }
187 
188 static void
189 metaslab_group_sort(metaslab_group_t *mg, metaslab_t *msp, uint64_t weight)
190 {
191 	/*
192 	 * Although in principle the weight can be any value, in
193 	 * practice we do not use values in the range [1, 510].
194 	 */
195 	ASSERT(weight >= SPA_MINBLOCKSIZE-1 || weight == 0);
196 	ASSERT(MUTEX_HELD(&msp->ms_lock));
197 
198 	mutex_enter(&mg->mg_lock);
199 	ASSERT(msp->ms_group == mg);
200 	avl_remove(&mg->mg_metaslab_tree, msp);
201 	msp->ms_weight = weight;
202 	avl_add(&mg->mg_metaslab_tree, msp);
203 	mutex_exit(&mg->mg_lock);
204 }
205 
206 /*
207  * ==========================================================================
208  * The first-fit block allocator
209  * ==========================================================================
210  */
211 static void
212 metaslab_ff_load(space_map_t *sm)
213 {
214 	ASSERT(sm->sm_ppd == NULL);
215 	sm->sm_ppd = kmem_zalloc(64 * sizeof (uint64_t), KM_SLEEP);
216 }
217 
218 static void
219 metaslab_ff_unload(space_map_t *sm)
220 {
221 	kmem_free(sm->sm_ppd, 64 * sizeof (uint64_t));
222 	sm->sm_ppd = NULL;
223 }
224 
225 static uint64_t
226 metaslab_ff_alloc(space_map_t *sm, uint64_t size)
227 {
228 	avl_tree_t *t = &sm->sm_root;
229 	uint64_t align = size & -size;
230 	uint64_t *cursor = (uint64_t *)sm->sm_ppd + highbit(align) - 1;
231 	space_seg_t *ss, ssearch;
232 	avl_index_t where;
233 
234 	ssearch.ss_start = *cursor;
235 	ssearch.ss_end = *cursor + size;
236 
237 	ss = avl_find(t, &ssearch, &where);
238 	if (ss == NULL)
239 		ss = avl_nearest(t, where, AVL_AFTER);
240 
241 	while (ss != NULL) {
242 		uint64_t offset = P2ROUNDUP(ss->ss_start, align);
243 
244 		if (offset + size <= ss->ss_end) {
245 			*cursor = offset + size;
246 			return (offset);
247 		}
248 		ss = AVL_NEXT(t, ss);
249 	}
250 
251 	/*
252 	 * If we know we've searched the whole map (*cursor == 0), give up.
253 	 * Otherwise, reset the cursor to the beginning and try again.
254 	 */
255 	if (*cursor == 0)
256 		return (-1ULL);
257 
258 	*cursor = 0;
259 	return (metaslab_ff_alloc(sm, size));
260 }
261 
262 /* ARGSUSED */
263 static void
264 metaslab_ff_claim(space_map_t *sm, uint64_t start, uint64_t size)
265 {
266 	/* No need to update cursor */
267 }
268 
269 /* ARGSUSED */
270 static void
271 metaslab_ff_free(space_map_t *sm, uint64_t start, uint64_t size)
272 {
273 	/* No need to update cursor */
274 }
275 
276 static space_map_ops_t metaslab_ff_ops = {
277 	metaslab_ff_load,
278 	metaslab_ff_unload,
279 	metaslab_ff_alloc,
280 	metaslab_ff_claim,
281 	metaslab_ff_free
282 };
283 
284 /*
285  * ==========================================================================
286  * Metaslabs
287  * ==========================================================================
288  */
289 metaslab_t *
290 metaslab_init(metaslab_group_t *mg, space_map_obj_t *smo,
291 	uint64_t start, uint64_t size, uint64_t txg)
292 {
293 	vdev_t *vd = mg->mg_vd;
294 	metaslab_t *msp;
295 
296 	msp = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
297 	mutex_init(&msp->ms_lock, NULL, MUTEX_DEFAULT, NULL);
298 
299 	msp->ms_smo_syncing = *smo;
300 
301 	/*
302 	 * We create the main space map here, but we don't create the
303 	 * allocmaps and freemaps until metaslab_sync_done().  This serves
304 	 * two purposes: it allows metaslab_sync_done() to detect the
305 	 * addition of new space; and for debugging, it ensures that we'd
306 	 * data fault on any attempt to use this metaslab before it's ready.
307 	 */
308 	space_map_create(&msp->ms_map, start, size,
309 	    vd->vdev_ashift, &msp->ms_lock);
310 
311 	metaslab_group_add(mg, msp);
312 
313 	/*
314 	 * If we're opening an existing pool (txg == 0) or creating
315 	 * a new one (txg == TXG_INITIAL), all space is available now.
316 	 * If we're adding space to an existing pool, the new space
317 	 * does not become available until after this txg has synced.
318 	 */
319 	if (txg <= TXG_INITIAL)
320 		metaslab_sync_done(msp, 0);
321 
322 	if (txg != 0) {
323 		/*
324 		 * The vdev is dirty, but the metaslab isn't -- it just needs
325 		 * to have metaslab_sync_done() invoked from vdev_sync_done().
326 		 * [We could just dirty the metaslab, but that would cause us
327 		 * to allocate a space map object for it, which is wasteful
328 		 * and would mess up the locality logic in metaslab_weight().]
329 		 */
330 		ASSERT(TXG_CLEAN(txg) == spa_last_synced_txg(vd->vdev_spa));
331 		vdev_dirty(vd, 0, NULL, txg);
332 		vdev_dirty(vd, VDD_METASLAB, msp, TXG_CLEAN(txg));
333 	}
334 
335 	return (msp);
336 }
337 
338 void
339 metaslab_fini(metaslab_t *msp)
340 {
341 	metaslab_group_t *mg = msp->ms_group;
342 	int t;
343 
344 	vdev_space_update(mg->mg_vd, -msp->ms_map.sm_size,
345 	    -msp->ms_smo.smo_alloc, B_TRUE);
346 
347 	metaslab_group_remove(mg, msp);
348 
349 	mutex_enter(&msp->ms_lock);
350 
351 	space_map_unload(&msp->ms_map);
352 	space_map_destroy(&msp->ms_map);
353 
354 	for (t = 0; t < TXG_SIZE; t++) {
355 		space_map_destroy(&msp->ms_allocmap[t]);
356 		space_map_destroy(&msp->ms_freemap[t]);
357 	}
358 
359 	mutex_exit(&msp->ms_lock);
360 	mutex_destroy(&msp->ms_lock);
361 
362 	kmem_free(msp, sizeof (metaslab_t));
363 }
364 
365 #define	METASLAB_WEIGHT_PRIMARY		(1ULL << 63)
366 #define	METASLAB_WEIGHT_SECONDARY	(1ULL << 62)
367 #define	METASLAB_ACTIVE_MASK		\
368 	(METASLAB_WEIGHT_PRIMARY | METASLAB_WEIGHT_SECONDARY)
369 #define	METASLAB_SMO_BONUS_MULTIPLIER	2
370 
371 static uint64_t
372 metaslab_weight(metaslab_t *msp)
373 {
374 	metaslab_group_t *mg = msp->ms_group;
375 	space_map_t *sm = &msp->ms_map;
376 	space_map_obj_t *smo = &msp->ms_smo;
377 	vdev_t *vd = mg->mg_vd;
378 	uint64_t weight, space;
379 
380 	ASSERT(MUTEX_HELD(&msp->ms_lock));
381 
382 	/*
383 	 * The baseline weight is the metaslab's free space.
384 	 */
385 	space = sm->sm_size - smo->smo_alloc;
386 	weight = space;
387 
388 	/*
389 	 * Modern disks have uniform bit density and constant angular velocity.
390 	 * Therefore, the outer recording zones are faster (higher bandwidth)
391 	 * than the inner zones by the ratio of outer to inner track diameter,
392 	 * which is typically around 2:1.  We account for this by assigning
393 	 * higher weight to lower metaslabs (multiplier ranging from 2x to 1x).
394 	 * In effect, this means that we'll select the metaslab with the most
395 	 * free bandwidth rather than simply the one with the most free space.
396 	 */
397 	weight = 2 * weight -
398 	    ((sm->sm_start >> vd->vdev_ms_shift) * weight) / vd->vdev_ms_count;
399 	ASSERT(weight >= space && weight <= 2 * space);
400 
401 	/*
402 	 * For locality, assign higher weight to metaslabs we've used before.
403 	 */
404 	if (smo->smo_object != 0)
405 		weight *= METASLAB_SMO_BONUS_MULTIPLIER;
406 	ASSERT(weight >= space &&
407 	    weight <= 2 * METASLAB_SMO_BONUS_MULTIPLIER * space);
408 
409 	/*
410 	 * If this metaslab is one we're actively using, adjust its weight to
411 	 * make it preferable to any inactive metaslab so we'll polish it off.
412 	 */
413 	weight |= (msp->ms_weight & METASLAB_ACTIVE_MASK);
414 
415 	return (weight);
416 }
417 
418 static int
419 metaslab_activate(metaslab_t *msp, uint64_t activation_weight)
420 {
421 	space_map_t *sm = &msp->ms_map;
422 
423 	ASSERT(MUTEX_HELD(&msp->ms_lock));
424 
425 	if ((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
426 		int error = space_map_load(sm, &metaslab_ff_ops,
427 		    SM_FREE, &msp->ms_smo,
428 		    msp->ms_group->mg_vd->vdev_spa->spa_meta_objset);
429 		if (error) {
430 			metaslab_group_sort(msp->ms_group, msp, 0);
431 			return (error);
432 		}
433 		metaslab_group_sort(msp->ms_group, msp,
434 		    msp->ms_weight | activation_weight);
435 	}
436 	ASSERT(sm->sm_loaded);
437 	ASSERT(msp->ms_weight & METASLAB_ACTIVE_MASK);
438 
439 	return (0);
440 }
441 
442 static void
443 metaslab_passivate(metaslab_t *msp, uint64_t size)
444 {
445 	/*
446 	 * If size < SPA_MINBLOCKSIZE, then we will not allocate from
447 	 * this metaslab again.  In that case, it had better be empty,
448 	 * or we would be leaving space on the table.
449 	 */
450 	ASSERT(size >= SPA_MINBLOCKSIZE || msp->ms_map.sm_space == 0);
451 	metaslab_group_sort(msp->ms_group, msp, MIN(msp->ms_weight, size));
452 	ASSERT((msp->ms_weight & METASLAB_ACTIVE_MASK) == 0);
453 }
454 
455 /*
456  * Write a metaslab to disk in the context of the specified transaction group.
457  */
458 void
459 metaslab_sync(metaslab_t *msp, uint64_t txg)
460 {
461 	vdev_t *vd = msp->ms_group->mg_vd;
462 	spa_t *spa = vd->vdev_spa;
463 	objset_t *mos = spa->spa_meta_objset;
464 	space_map_t *allocmap = &msp->ms_allocmap[txg & TXG_MASK];
465 	space_map_t *freemap = &msp->ms_freemap[txg & TXG_MASK];
466 	space_map_t *freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
467 	space_map_t *sm = &msp->ms_map;
468 	space_map_obj_t *smo = &msp->ms_smo_syncing;
469 	dmu_buf_t *db;
470 	dmu_tx_t *tx;
471 	int t;
472 
473 	tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
474 
475 	/*
476 	 * The only state that can actually be changing concurrently with
477 	 * metaslab_sync() is the metaslab's ms_map.  No other thread can
478 	 * be modifying this txg's allocmap, freemap, freed_map, or smo.
479 	 * Therefore, we only hold ms_lock to satify space_map ASSERTs.
480 	 * We drop it whenever we call into the DMU, because the DMU
481 	 * can call down to us (e.g. via zio_free()) at any time.
482 	 */
483 	mutex_enter(&msp->ms_lock);
484 
485 	if (smo->smo_object == 0) {
486 		ASSERT(smo->smo_objsize == 0);
487 		ASSERT(smo->smo_alloc == 0);
488 		mutex_exit(&msp->ms_lock);
489 		smo->smo_object = dmu_object_alloc(mos,
490 		    DMU_OT_SPACE_MAP, 1 << SPACE_MAP_BLOCKSHIFT,
491 		    DMU_OT_SPACE_MAP_HEADER, sizeof (*smo), tx);
492 		ASSERT(smo->smo_object != 0);
493 		dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
494 		    (sm->sm_start >> vd->vdev_ms_shift),
495 		    sizeof (uint64_t), &smo->smo_object, tx);
496 		mutex_enter(&msp->ms_lock);
497 	}
498 
499 	space_map_walk(freemap, space_map_add, freed_map);
500 
501 	if (sm->sm_loaded && spa_sync_pass(spa) == 1 && smo->smo_objsize >=
502 	    2 * sizeof (uint64_t) * avl_numnodes(&sm->sm_root)) {
503 		/*
504 		 * The in-core space map representation is twice as compact
505 		 * as the on-disk one, so it's time to condense the latter
506 		 * by generating a pure allocmap from first principles.
507 		 *
508 		 * This metaslab is 100% allocated,
509 		 * minus the content of the in-core map (sm),
510 		 * minus what's been freed this txg (freed_map),
511 		 * minus allocations from txgs in the future
512 		 * (because they haven't been committed yet).
513 		 */
514 		space_map_vacate(allocmap, NULL, NULL);
515 		space_map_vacate(freemap, NULL, NULL);
516 
517 		space_map_add(allocmap, allocmap->sm_start, allocmap->sm_size);
518 
519 		space_map_walk(sm, space_map_remove, allocmap);
520 		space_map_walk(freed_map, space_map_remove, allocmap);
521 
522 		for (t = 1; t < TXG_CONCURRENT_STATES; t++)
523 			space_map_walk(&msp->ms_allocmap[(txg + t) & TXG_MASK],
524 			    space_map_remove, allocmap);
525 
526 		mutex_exit(&msp->ms_lock);
527 		space_map_truncate(smo, mos, tx);
528 		mutex_enter(&msp->ms_lock);
529 	}
530 
531 	space_map_sync(allocmap, SM_ALLOC, smo, mos, tx);
532 	space_map_sync(freemap, SM_FREE, smo, mos, tx);
533 
534 	mutex_exit(&msp->ms_lock);
535 
536 	VERIFY(0 == dmu_bonus_hold(mos, smo->smo_object, FTAG, &db));
537 	dmu_buf_will_dirty(db, tx);
538 	ASSERT3U(db->db_size, >=, sizeof (*smo));
539 	bcopy(smo, db->db_data, sizeof (*smo));
540 	dmu_buf_rele(db, FTAG);
541 
542 	dmu_tx_commit(tx);
543 }
544 
545 /*
546  * Called after a transaction group has completely synced to mark
547  * all of the metaslab's free space as usable.
548  */
549 void
550 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
551 {
552 	space_map_obj_t *smo = &msp->ms_smo;
553 	space_map_obj_t *smosync = &msp->ms_smo_syncing;
554 	space_map_t *sm = &msp->ms_map;
555 	space_map_t *freed_map = &msp->ms_freemap[TXG_CLEAN(txg) & TXG_MASK];
556 	metaslab_group_t *mg = msp->ms_group;
557 	vdev_t *vd = mg->mg_vd;
558 	int t;
559 
560 	mutex_enter(&msp->ms_lock);
561 
562 	/*
563 	 * If this metaslab is just becoming available, initialize its
564 	 * allocmaps and freemaps and add its capacity to the vdev.
565 	 */
566 	if (freed_map->sm_size == 0) {
567 		for (t = 0; t < TXG_SIZE; t++) {
568 			space_map_create(&msp->ms_allocmap[t], sm->sm_start,
569 			    sm->sm_size, sm->sm_shift, sm->sm_lock);
570 			space_map_create(&msp->ms_freemap[t], sm->sm_start,
571 			    sm->sm_size, sm->sm_shift, sm->sm_lock);
572 		}
573 		vdev_space_update(vd, sm->sm_size, 0, B_TRUE);
574 	}
575 
576 	vdev_space_update(vd, 0, smosync->smo_alloc - smo->smo_alloc, B_TRUE);
577 
578 	ASSERT(msp->ms_allocmap[txg & TXG_MASK].sm_space == 0);
579 	ASSERT(msp->ms_freemap[txg & TXG_MASK].sm_space == 0);
580 
581 	/*
582 	 * If there's a space_map_load() in progress, wait for it to complete
583 	 * so that we have a consistent view of the in-core space map.
584 	 * Then, add everything we freed in this txg to the map.
585 	 */
586 	space_map_load_wait(sm);
587 	space_map_vacate(freed_map, sm->sm_loaded ? space_map_free : NULL, sm);
588 
589 	*smo = *smosync;
590 
591 	/*
592 	 * If the map is loaded but no longer active, evict it as soon as all
593 	 * future allocations have synced.  (If we unloaded it now and then
594 	 * loaded a moment later, the map wouldn't reflect those allocations.)
595 	 */
596 	if (sm->sm_loaded && (msp->ms_weight & METASLAB_ACTIVE_MASK) == 0) {
597 		int evictable = 1;
598 
599 		for (t = 1; t < TXG_CONCURRENT_STATES; t++)
600 			if (msp->ms_allocmap[(txg + t) & TXG_MASK].sm_space)
601 				evictable = 0;
602 
603 		if (evictable)
604 			space_map_unload(sm);
605 	}
606 
607 	metaslab_group_sort(mg, msp, metaslab_weight(msp));
608 
609 	mutex_exit(&msp->ms_lock);
610 }
611 
612 static uint64_t
613 metaslab_distance(metaslab_t *msp, dva_t *dva)
614 {
615 	uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
616 	uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
617 	uint64_t start = msp->ms_map.sm_start >> ms_shift;
618 
619 	if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
620 		return (1ULL << 63);
621 
622 	if (offset < start)
623 		return ((start - offset) << ms_shift);
624 	if (offset > start)
625 		return ((offset - start) << ms_shift);
626 	return (0);
627 }
628 
629 static uint64_t
630 metaslab_group_alloc(metaslab_group_t *mg, uint64_t size, uint64_t txg,
631     uint64_t min_distance, dva_t *dva, int d)
632 {
633 	metaslab_t *msp = NULL;
634 	uint64_t offset = -1ULL;
635 	avl_tree_t *t = &mg->mg_metaslab_tree;
636 	uint64_t activation_weight;
637 	uint64_t target_distance;
638 	int i;
639 
640 	activation_weight = METASLAB_WEIGHT_PRIMARY;
641 	for (i = 0; i < d; i++)
642 		if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id)
643 			activation_weight = METASLAB_WEIGHT_SECONDARY;
644 
645 	for (;;) {
646 		mutex_enter(&mg->mg_lock);
647 		for (msp = avl_first(t); msp; msp = AVL_NEXT(t, msp)) {
648 			if (msp->ms_weight < size) {
649 				mutex_exit(&mg->mg_lock);
650 				return (-1ULL);
651 			}
652 
653 			if (activation_weight == METASLAB_WEIGHT_PRIMARY)
654 				break;
655 
656 			target_distance = min_distance +
657 			    (msp->ms_smo.smo_alloc ? 0 : min_distance >> 1);
658 
659 			for (i = 0; i < d; i++)
660 				if (metaslab_distance(msp, &dva[i]) <
661 				    target_distance)
662 					break;
663 			if (i == d)
664 				break;
665 		}
666 		mutex_exit(&mg->mg_lock);
667 		if (msp == NULL)
668 			return (-1ULL);
669 
670 		mutex_enter(&msp->ms_lock);
671 
672 		/*
673 		 * Ensure that the metaslab we have selected is still
674 		 * capable of handling our request. It's possible that
675 		 * another thread may have changed the weight while we
676 		 * were blocked on the metaslab lock.
677 		 */
678 		if (msp->ms_weight < size) {
679 			mutex_exit(&msp->ms_lock);
680 			continue;
681 		}
682 
683 		if ((msp->ms_weight & METASLAB_WEIGHT_SECONDARY) &&
684 		    activation_weight == METASLAB_WEIGHT_PRIMARY) {
685 			metaslab_passivate(msp,
686 			    msp->ms_weight & ~METASLAB_ACTIVE_MASK);
687 			mutex_exit(&msp->ms_lock);
688 			continue;
689 		}
690 
691 		if (metaslab_activate(msp, activation_weight) != 0) {
692 			mutex_exit(&msp->ms_lock);
693 			continue;
694 		}
695 
696 		if ((offset = space_map_alloc(&msp->ms_map, size)) != -1ULL)
697 			break;
698 
699 		metaslab_passivate(msp, size - 1);
700 
701 		mutex_exit(&msp->ms_lock);
702 	}
703 
704 	if (msp->ms_allocmap[txg & TXG_MASK].sm_space == 0)
705 		vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
706 
707 	space_map_add(&msp->ms_allocmap[txg & TXG_MASK], offset, size);
708 
709 	mutex_exit(&msp->ms_lock);
710 
711 	return (offset);
712 }
713 
714 /*
715  * Allocate a block for the specified i/o.
716  */
717 static int
718 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
719     dva_t *dva, int d, dva_t *hintdva, uint64_t txg, boolean_t hintdva_avoid)
720 {
721 	metaslab_group_t *mg, *rotor;
722 	vdev_t *vd;
723 	int dshift = 3;
724 	int all_zero;
725 	uint64_t offset = -1ULL;
726 	uint64_t asize;
727 	uint64_t distance;
728 
729 	ASSERT(!DVA_IS_VALID(&dva[d]));
730 
731 	/*
732 	 * For testing, make some blocks above a certain size be gang blocks.
733 	 */
734 	if (psize >= metaslab_gang_bang && (lbolt & 3) == 0)
735 		return (ENOSPC);
736 
737 	/*
738 	 * Start at the rotor and loop through all mgs until we find something.
739 	 * Note that there's no locking on mc_rotor or mc_allocated because
740 	 * nothing actually breaks if we miss a few updates -- we just won't
741 	 * allocate quite as evenly.  It all balances out over time.
742 	 *
743 	 * If we are doing ditto or log blocks, try to spread them across
744 	 * consecutive vdevs.  If we're forced to reuse a vdev before we've
745 	 * allocated all of our ditto blocks, then try and spread them out on
746 	 * that vdev as much as possible.  If it turns out to not be possible,
747 	 * gradually lower our standards until anything becomes acceptable.
748 	 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
749 	 * gives us hope of containing our fault domains to something we're
750 	 * able to reason about.  Otherwise, any two top-level vdev failures
751 	 * will guarantee the loss of data.  With consecutive allocation,
752 	 * only two adjacent top-level vdev failures will result in data loss.
753 	 *
754 	 * If we are doing gang blocks (hintdva is non-NULL), try to keep
755 	 * ourselves on the same vdev as our gang block header.  That
756 	 * way, we can hope for locality in vdev_cache, plus it makes our
757 	 * fault domains something tractable.
758 	 */
759 	if (hintdva) {
760 		vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
761 		if (hintdva_avoid)
762 			mg = vd->vdev_mg->mg_next;
763 		else
764 			mg = vd->vdev_mg;
765 	} else if (d != 0) {
766 		vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
767 		mg = vd->vdev_mg->mg_next;
768 	} else {
769 		mg = mc->mc_rotor;
770 	}
771 
772 	/*
773 	 * If the hint put us into the wrong class, just follow the rotor.
774 	 */
775 	if (mg->mg_class != mc)
776 		mg = mc->mc_rotor;
777 
778 	rotor = mg;
779 top:
780 	all_zero = B_TRUE;
781 	do {
782 		vd = mg->mg_vd;
783 		/*
784 		 * Dont allocate from faulted devices
785 		 */
786 		if (!vdev_writeable(vd))
787 			goto next;
788 		/*
789 		 * Avoid writing single-copy data to a failing vdev
790 		 */
791 		if ((vd->vdev_stat.vs_write_errors > 0 ||
792 		    vd->vdev_state < VDEV_STATE_HEALTHY) &&
793 		    d == 0 && dshift == 3) {
794 			all_zero = B_FALSE;
795 			goto next;
796 		}
797 
798 		ASSERT(mg->mg_class == mc);
799 
800 		distance = vd->vdev_asize >> dshift;
801 		if (distance <= (1ULL << vd->vdev_ms_shift))
802 			distance = 0;
803 		else
804 			all_zero = B_FALSE;
805 
806 		asize = vdev_psize_to_asize(vd, psize);
807 		ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
808 
809 		offset = metaslab_group_alloc(mg, asize, txg, distance, dva, d);
810 		if (offset != -1ULL) {
811 			/*
812 			 * If we've just selected this metaslab group,
813 			 * figure out whether the corresponding vdev is
814 			 * over- or under-used relative to the pool,
815 			 * and set an allocation bias to even it out.
816 			 */
817 			if (mc->mc_allocated == 0) {
818 				vdev_stat_t *vs = &vd->vdev_stat;
819 				uint64_t alloc, space;
820 				int64_t vu, su;
821 
822 				alloc = spa_get_alloc(spa);
823 				space = spa_get_space(spa);
824 
825 				/*
826 				 * Determine percent used in units of 0..1024.
827 				 * (This is just to avoid floating point.)
828 				 */
829 				vu = (vs->vs_alloc << 10) / (vs->vs_space + 1);
830 				su = (alloc << 10) / (space + 1);
831 
832 				/*
833 				 * Bias by at most +/- 25% of the aliquot.
834 				 */
835 				mg->mg_bias = ((su - vu) *
836 				    (int64_t)mg->mg_aliquot) / (1024 * 4);
837 			}
838 
839 			if (atomic_add_64_nv(&mc->mc_allocated, asize) >=
840 			    mg->mg_aliquot + mg->mg_bias) {
841 				mc->mc_rotor = mg->mg_next;
842 				mc->mc_allocated = 0;
843 			}
844 
845 			DVA_SET_VDEV(&dva[d], vd->vdev_id);
846 			DVA_SET_OFFSET(&dva[d], offset);
847 			DVA_SET_GANG(&dva[d], 0);
848 			DVA_SET_ASIZE(&dva[d], asize);
849 
850 			return (0);
851 		}
852 next:
853 		mc->mc_rotor = mg->mg_next;
854 		mc->mc_allocated = 0;
855 	} while ((mg = mg->mg_next) != rotor);
856 
857 	if (!all_zero) {
858 		dshift++;
859 		ASSERT(dshift < 64);
860 		goto top;
861 	}
862 
863 	bzero(&dva[d], sizeof (dva_t));
864 
865 	return (ENOSPC);
866 }
867 
868 /*
869  * Free the block represented by DVA in the context of the specified
870  * transaction group.
871  */
872 static void
873 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
874 {
875 	uint64_t vdev = DVA_GET_VDEV(dva);
876 	uint64_t offset = DVA_GET_OFFSET(dva);
877 	uint64_t size = DVA_GET_ASIZE(dva);
878 	vdev_t *vd;
879 	metaslab_t *msp;
880 
881 	ASSERT(DVA_IS_VALID(dva));
882 
883 	if (txg > spa_freeze_txg(spa))
884 		return;
885 
886 	if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
887 	    (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
888 		cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
889 		    (u_longlong_t)vdev, (u_longlong_t)offset);
890 		ASSERT(0);
891 		return;
892 	}
893 
894 	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
895 
896 	if (DVA_GET_GANG(dva))
897 		size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
898 
899 	mutex_enter(&msp->ms_lock);
900 
901 	if (now) {
902 		space_map_remove(&msp->ms_allocmap[txg & TXG_MASK],
903 		    offset, size);
904 		space_map_free(&msp->ms_map, offset, size);
905 	} else {
906 		if (msp->ms_freemap[txg & TXG_MASK].sm_space == 0)
907 			vdev_dirty(vd, VDD_METASLAB, msp, txg);
908 		space_map_add(&msp->ms_freemap[txg & TXG_MASK], offset, size);
909 
910 		/*
911 		 * verify that this region is actually allocated in
912 		 * either a ms_allocmap or the ms_map
913 		 */
914 		if (msp->ms_map.sm_loaded) {
915 			boolean_t allocd = B_FALSE;
916 			int i;
917 
918 			if (!space_map_contains(&msp->ms_map, offset, size)) {
919 				allocd = B_TRUE;
920 			} else {
921 				for (i = 0; i < TXG_CONCURRENT_STATES; i++) {
922 					space_map_t *sm = &msp->ms_allocmap
923 					    [(txg - i) & TXG_MASK];
924 					if (space_map_contains(sm,
925 					    offset, size)) {
926 						allocd = B_TRUE;
927 						break;
928 					}
929 				}
930 			}
931 
932 			if (!allocd) {
933 				zfs_panic_recover("freeing free segment "
934 				    "(vdev=%llu offset=%llx size=%llx)",
935 				    (longlong_t)vdev, (longlong_t)offset,
936 				    (longlong_t)size);
937 			}
938 		}
939 
940 
941 	}
942 
943 	mutex_exit(&msp->ms_lock);
944 }
945 
946 /*
947  * Intent log support: upon opening the pool after a crash, notify the SPA
948  * of blocks that the intent log has allocated for immediate write, but
949  * which are still considered free by the SPA because the last transaction
950  * group didn't commit yet.
951  */
952 static int
953 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
954 {
955 	uint64_t vdev = DVA_GET_VDEV(dva);
956 	uint64_t offset = DVA_GET_OFFSET(dva);
957 	uint64_t size = DVA_GET_ASIZE(dva);
958 	vdev_t *vd;
959 	metaslab_t *msp;
960 	int error;
961 
962 	ASSERT(DVA_IS_VALID(dva));
963 
964 	if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
965 	    (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
966 		return (ENXIO);
967 
968 	msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
969 
970 	if (DVA_GET_GANG(dva))
971 		size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
972 
973 	mutex_enter(&msp->ms_lock);
974 
975 	error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
976 	if (error) {
977 		mutex_exit(&msp->ms_lock);
978 		return (error);
979 	}
980 
981 	if (msp->ms_allocmap[txg & TXG_MASK].sm_space == 0)
982 		vdev_dirty(vd, VDD_METASLAB, msp, txg);
983 
984 	space_map_claim(&msp->ms_map, offset, size);
985 	space_map_add(&msp->ms_allocmap[txg & TXG_MASK], offset, size);
986 
987 	mutex_exit(&msp->ms_lock);
988 
989 	return (0);
990 }
991 
992 int
993 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
994     int ndvas, uint64_t txg, blkptr_t *hintbp, boolean_t hintbp_avoid)
995 {
996 	dva_t *dva = bp->blk_dva;
997 	dva_t *hintdva = hintbp->blk_dva;
998 	int d;
999 	int error = 0;
1000 
1001 	if (mc->mc_rotor == NULL)	/* no vdevs in this class */
1002 		return (ENOSPC);
1003 
1004 	ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
1005 	ASSERT(BP_GET_NDVAS(bp) == 0);
1006 	ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
1007 
1008 	for (d = 0; d < ndvas; d++) {
1009 		error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
1010 		    txg, hintbp_avoid);
1011 		if (error) {
1012 			for (d--; d >= 0; d--) {
1013 				metaslab_free_dva(spa, &dva[d], txg, B_TRUE);
1014 				bzero(&dva[d], sizeof (dva_t));
1015 			}
1016 			return (error);
1017 		}
1018 	}
1019 	ASSERT(error == 0);
1020 	ASSERT(BP_GET_NDVAS(bp) == ndvas);
1021 
1022 	return (0);
1023 }
1024 
1025 void
1026 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
1027 {
1028 	const dva_t *dva = bp->blk_dva;
1029 	int ndvas = BP_GET_NDVAS(bp);
1030 	int d;
1031 
1032 	ASSERT(!BP_IS_HOLE(bp));
1033 
1034 	for (d = 0; d < ndvas; d++)
1035 		metaslab_free_dva(spa, &dva[d], txg, now);
1036 }
1037 
1038 int
1039 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
1040 {
1041 	const dva_t *dva = bp->blk_dva;
1042 	int ndvas = BP_GET_NDVAS(bp);
1043 	int d, error;
1044 	int last_error = 0;
1045 
1046 	ASSERT(!BP_IS_HOLE(bp));
1047 
1048 	for (d = 0; d < ndvas; d++)
1049 		if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
1050 			last_error = error;
1051 
1052 	return (last_error);
1053 }
1054