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 (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
26 */
27
28 #include <sys/zfs_context.h>
29 #include <sys/dmu.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/space_map.h>
32 #include <sys/metaslab_impl.h>
33 #include <sys/vdev_impl.h>
34 #include <sys/zio.h>
35 #include <sys/spa_impl.h>
36 #include <sys/zfeature.h>
37 #include <sys/vdev_indirect_mapping.h>
38
39 #define GANG_ALLOCATION(flags) \
40 ((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER))
41
42 uint64_t metaslab_aliquot = 512ULL << 10;
43 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1; /* force gang blocks */
44
45 /*
46 * The in-core space map representation is more compact than its on-disk form.
47 * The zfs_condense_pct determines how much more compact the in-core
48 * space map representation must be before we compact it on-disk.
49 * Values should be greater than or equal to 100.
50 */
51 int zfs_condense_pct = 200;
52
53 /*
54 * Condensing a metaslab is not guaranteed to actually reduce the amount of
55 * space used on disk. In particular, a space map uses data in increments of
56 * MAX(1 << ashift, space_map_blksize), so a metaslab might use the
57 * same number of blocks after condensing. Since the goal of condensing is to
150 * Enable/disable preloading of metaslab.
151 */
152 boolean_t metaslab_preload_enabled = B_TRUE;
153
154 /*
155 * Enable/disable fragmentation weighting on metaslabs.
156 */
157 boolean_t metaslab_fragmentation_factor_enabled = B_TRUE;
158
159 /*
160 * Enable/disable lba weighting (i.e. outer tracks are given preference).
161 */
162 boolean_t metaslab_lba_weighting_enabled = B_TRUE;
163
164 /*
165 * Enable/disable metaslab group biasing.
166 */
167 boolean_t metaslab_bias_enabled = B_TRUE;
168
169 /*
170 * Enable/disable remapping of indirect DVAs to their concrete vdevs.
171 */
172 boolean_t zfs_remap_blkptr_enable = B_TRUE;
173
174 /*
175 * Enable/disable segment-based metaslab selection.
176 */
177 boolean_t zfs_metaslab_segment_weight_enabled = B_TRUE;
178
179 /*
180 * When using segment-based metaslab selection, we will continue
181 * allocating from the active metaslab until we have exhausted
182 * zfs_metaslab_switch_threshold of its buckets.
183 */
184 int zfs_metaslab_switch_threshold = 2;
185
186 /*
187 * Internal switch to enable/disable the metaslab allocation tracing
188 * facility.
189 */
190 boolean_t metaslab_trace_enabled = B_TRUE;
191
192 /*
193 * Maximum entries that the metaslab allocation tracing facility will keep
194 * in a given list when running in non-debug mode. We limit the number
195 * of entries in non-debug mode to prevent us from using up too much memory.
196 * The limit should be sufficiently large that we don't expect any allocation
197 * to every exceed this value. In debug mode, the system will panic if this
198 * limit is ever reached allowing for further investigation.
199 */
200 uint64_t metaslab_trace_max_entries = 5000;
201
202 static uint64_t metaslab_weight(metaslab_t *);
203 static void metaslab_set_fragmentation(metaslab_t *);
204 static void metaslab_free_impl(vdev_t *, uint64_t, uint64_t, uint64_t);
205 static void metaslab_check_free_impl(vdev_t *, uint64_t, uint64_t);
206
207 kmem_cache_t *metaslab_alloc_trace_cache;
208
209 /*
210 * ==========================================================================
211 * Metaslab classes
212 * ==========================================================================
213 */
214 metaslab_class_t *
215 metaslab_class_create(spa_t *spa, metaslab_ops_t *ops)
216 {
217 metaslab_class_t *mc;
218
219 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
220
221 mc->mc_spa = spa;
222 mc->mc_rotor = NULL;
223 mc->mc_ops = ops;
224 mutex_init(&mc->mc_lock, NULL, MUTEX_DEFAULT, NULL);
225 refcount_create_tracked(&mc->mc_alloc_slots);
226
227 return (mc);
228 }
229
230 void
231 metaslab_class_destroy(metaslab_class_t *mc)
232 {
233 ASSERT(mc->mc_rotor == NULL);
234 ASSERT(mc->mc_alloc == 0);
235 ASSERT(mc->mc_deferred == 0);
236 ASSERT(mc->mc_space == 0);
237 ASSERT(mc->mc_dspace == 0);
238
239 refcount_destroy(&mc->mc_alloc_slots);
240 mutex_destroy(&mc->mc_lock);
241 kmem_free(mc, sizeof (metaslab_class_t));
242 }
243
244 int
245 metaslab_class_validate(metaslab_class_t *mc)
246 {
247 metaslab_group_t *mg;
248 vdev_t *vd;
249
250 /*
251 * Must hold one of the spa_config locks.
252 */
253 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
254 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
255
256 if ((mg = mc->mc_rotor) == NULL)
257 return (0);
258
305 metaslab_class_histogram_verify(metaslab_class_t *mc)
306 {
307 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
308 uint64_t *mc_hist;
309 int i;
310
311 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
312 return;
313
314 mc_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
315 KM_SLEEP);
316
317 for (int c = 0; c < rvd->vdev_children; c++) {
318 vdev_t *tvd = rvd->vdev_child[c];
319 metaslab_group_t *mg = tvd->vdev_mg;
320
321 /*
322 * Skip any holes, uninitialized top-levels, or
323 * vdevs that are not in this metalab class.
324 */
325 if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 ||
326 mg->mg_class != mc) {
327 continue;
328 }
329
330 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
331 mc_hist[i] += mg->mg_histogram[i];
332 }
333
334 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
335 VERIFY3U(mc_hist[i], ==, mc->mc_histogram[i]);
336
337 kmem_free(mc_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
338 }
339
340 /*
341 * Calculate the metaslab class's fragmentation metric. The metric
342 * is weighted based on the space contribution of each metaslab group.
343 * The return value will be a number between 0 and 100 (inclusive), or
344 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
345 * zfs_frag_table for more information about the metric.
346 */
347 uint64_t
348 metaslab_class_fragmentation(metaslab_class_t *mc)
349 {
350 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
351 uint64_t fragmentation = 0;
352
353 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
354
355 for (int c = 0; c < rvd->vdev_children; c++) {
356 vdev_t *tvd = rvd->vdev_child[c];
357 metaslab_group_t *mg = tvd->vdev_mg;
358
359 /*
360 * Skip any holes, uninitialized top-levels,
361 * or vdevs that are not in this metalab class.
362 */
363 if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 ||
364 mg->mg_class != mc) {
365 continue;
366 }
367
368 /*
369 * If a metaslab group does not contain a fragmentation
370 * metric then just bail out.
371 */
372 if (mg->mg_fragmentation == ZFS_FRAG_INVALID) {
373 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
374 return (ZFS_FRAG_INVALID);
375 }
376
377 /*
378 * Determine how much this metaslab_group is contributing
379 * to the overall pool fragmentation metric.
380 */
381 fragmentation += mg->mg_fragmentation *
382 metaslab_group_get_space(mg);
383 }
389 }
390
391 /*
392 * Calculate the amount of expandable space that is available in
393 * this metaslab class. If a device is expanded then its expandable
394 * space will be the amount of allocatable space that is currently not
395 * part of this metaslab class.
396 */
397 uint64_t
398 metaslab_class_expandable_space(metaslab_class_t *mc)
399 {
400 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
401 uint64_t space = 0;
402
403 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
404 for (int c = 0; c < rvd->vdev_children; c++) {
405 uint64_t tspace;
406 vdev_t *tvd = rvd->vdev_child[c];
407 metaslab_group_t *mg = tvd->vdev_mg;
408
409 if (!vdev_is_concrete(tvd) || tvd->vdev_ms_shift == 0 ||
410 mg->mg_class != mc) {
411 continue;
412 }
413
414 /*
415 * Calculate if we have enough space to add additional
416 * metaslabs. We report the expandable space in terms
417 * of the metaslab size since that's the unit of expansion.
418 * Adjust by efi system partition size.
419 */
420 tspace = tvd->vdev_max_asize - tvd->vdev_asize;
421 if (tspace > mc->mc_spa->spa_bootsize) {
422 tspace -= mc->mc_spa->spa_bootsize;
423 }
424 space += P2ALIGN(tspace, 1ULL << tvd->vdev_ms_shift);
425 }
426 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
427 return (space);
428 }
429
501 * ==========================================================================
502 */
503 /*
504 * Update the allocatable flag and the metaslab group's capacity.
505 * The allocatable flag is set to true if the capacity is below
506 * the zfs_mg_noalloc_threshold or has a fragmentation value that is
507 * greater than zfs_mg_fragmentation_threshold. If a metaslab group
508 * transitions from allocatable to non-allocatable or vice versa then the
509 * metaslab group's class is updated to reflect the transition.
510 */
511 static void
512 metaslab_group_alloc_update(metaslab_group_t *mg)
513 {
514 vdev_t *vd = mg->mg_vd;
515 metaslab_class_t *mc = mg->mg_class;
516 vdev_stat_t *vs = &vd->vdev_stat;
517 boolean_t was_allocatable;
518 boolean_t was_initialized;
519
520 ASSERT(vd == vd->vdev_top);
521 ASSERT3U(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_READER), ==,
522 SCL_ALLOC);
523
524 mutex_enter(&mg->mg_lock);
525 was_allocatable = mg->mg_allocatable;
526 was_initialized = mg->mg_initialized;
527
528 mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
529 (vs->vs_space + 1);
530
531 mutex_enter(&mc->mc_lock);
532
533 /*
534 * If the metaslab group was just added then it won't
535 * have any space until we finish syncing out this txg.
536 * At that point we will consider it initialized and available
537 * for allocations. We also don't consider non-activated
538 * metaslab groups (e.g. vdevs that are in the middle of being removed)
539 * to be initialized, because they can't be used for allocation.
540 */
541 mg->mg_initialized = metaslab_group_initialized(mg);
542 if (!was_initialized && mg->mg_initialized) {
600 refcount_create_tracked(&mg->mg_alloc_queue_depth);
601
602 mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct,
603 minclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT);
604
605 return (mg);
606 }
607
608 void
609 metaslab_group_destroy(metaslab_group_t *mg)
610 {
611 ASSERT(mg->mg_prev == NULL);
612 ASSERT(mg->mg_next == NULL);
613 /*
614 * We may have gone below zero with the activation count
615 * either because we never activated in the first place or
616 * because we're done, and possibly removing the vdev.
617 */
618 ASSERT(mg->mg_activation_count <= 0);
619
620 taskq_destroy(mg->mg_taskq);
621 avl_destroy(&mg->mg_metaslab_tree);
622 mutex_destroy(&mg->mg_lock);
623 refcount_destroy(&mg->mg_alloc_queue_depth);
624 kmem_free(mg, sizeof (metaslab_group_t));
625 }
626
627 void
628 metaslab_group_activate(metaslab_group_t *mg)
629 {
630 metaslab_class_t *mc = mg->mg_class;
631 metaslab_group_t *mgprev, *mgnext;
632
633 ASSERT3U(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER), !=, 0);
634
635 ASSERT(mc->mc_rotor != mg);
636 ASSERT(mg->mg_prev == NULL);
637 ASSERT(mg->mg_next == NULL);
638 ASSERT(mg->mg_activation_count <= 0);
639
640 if (++mg->mg_activation_count <= 0)
641 return;
642
643 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
644 metaslab_group_alloc_update(mg);
645
646 if ((mgprev = mc->mc_rotor) == NULL) {
647 mg->mg_prev = mg;
648 mg->mg_next = mg;
649 } else {
650 mgnext = mgprev->mg_next;
651 mg->mg_prev = mgprev;
652 mg->mg_next = mgnext;
653 mgprev->mg_next = mg;
654 mgnext->mg_prev = mg;
655 }
656 mc->mc_rotor = mg;
657 }
658
659 /*
660 * Passivate a metaslab group and remove it from the allocation rotor.
661 * Callers must hold both the SCL_ALLOC and SCL_ZIO lock prior to passivating
662 * a metaslab group. This function will momentarily drop spa_config_locks
663 * that are lower than the SCL_ALLOC lock (see comment below).
664 */
665 void
666 metaslab_group_passivate(metaslab_group_t *mg)
667 {
668 metaslab_class_t *mc = mg->mg_class;
669 spa_t *spa = mc->mc_spa;
670 metaslab_group_t *mgprev, *mgnext;
671 int locks = spa_config_held(spa, SCL_ALL, RW_WRITER);
672
673 ASSERT3U(spa_config_held(spa, SCL_ALLOC | SCL_ZIO, RW_WRITER), ==,
674 (SCL_ALLOC | SCL_ZIO));
675
676 if (--mg->mg_activation_count != 0) {
677 ASSERT(mc->mc_rotor != mg);
678 ASSERT(mg->mg_prev == NULL);
679 ASSERT(mg->mg_next == NULL);
680 ASSERT(mg->mg_activation_count < 0);
681 return;
682 }
683
684 /*
685 * The spa_config_lock is an array of rwlocks, ordered as
686 * follows (from highest to lowest):
687 * SCL_CONFIG > SCL_STATE > SCL_L2ARC > SCL_ALLOC >
688 * SCL_ZIO > SCL_FREE > SCL_VDEV
689 * (For more information about the spa_config_lock see spa_misc.c)
690 * The higher the lock, the broader its coverage. When we passivate
691 * a metaslab group, we must hold both the SCL_ALLOC and the SCL_ZIO
692 * config locks. However, the metaslab group's taskq might be trying
693 * to preload metaslabs so we must drop the SCL_ZIO lock and any
694 * lower locks to allow the I/O to complete. At a minimum,
695 * we continue to hold the SCL_ALLOC lock, which prevents any future
696 * allocations from taking place and any changes to the vdev tree.
697 */
698 spa_config_exit(spa, locks & ~(SCL_ZIO - 1), spa);
699 taskq_wait(mg->mg_taskq);
700 spa_config_enter(spa, locks & ~(SCL_ZIO - 1), spa, RW_WRITER);
701 metaslab_group_alloc_update(mg);
702
703 mgprev = mg->mg_prev;
704 mgnext = mg->mg_next;
705
706 if (mg == mgnext) {
707 mc->mc_rotor = NULL;
708 } else {
709 mc->mc_rotor = mgnext;
710 mgprev->mg_next = mgnext;
711 mgnext->mg_prev = mgprev;
712 }
713
714 mg->mg_prev = NULL;
715 mg->mg_next = NULL;
716 }
717
718 boolean_t
719 metaslab_group_initialized(metaslab_group_t *mg)
720 {
1124 range_seg_t *rs, rsearch;
1125 avl_index_t where;
1126
1127 rsearch.rs_start = start;
1128 rsearch.rs_end = start + size;
1129
1130 rs = avl_find(t, &rsearch, &where);
1131 if (rs == NULL) {
1132 rs = avl_nearest(t, where, AVL_AFTER);
1133 }
1134
1135 return (rs);
1136 }
1137
1138 /*
1139 * This is a helper function that can be used by the allocator to find
1140 * a suitable block to allocate. This will search the specified AVL
1141 * tree looking for a block that matches the specified criteria.
1142 */
1143 static uint64_t
1144 metaslab_block_picker(avl_tree_t *t, uint64_t *cursor, uint64_t size,
1145 uint64_t align)
1146 {
1147 range_seg_t *rs = metaslab_block_find(t, *cursor, size);
1148
1149 while (rs != NULL) {
1150 uint64_t offset = P2ROUNDUP(rs->rs_start, align);
1151
1152 if (offset + size <= rs->rs_end) {
1153 *cursor = offset + size;
1154 return (offset);
1155 }
1156 rs = AVL_NEXT(t, rs);
1157 }
1158
1159 /*
1160 * If we know we've searched the whole map (*cursor == 0), give up.
1161 * Otherwise, reset the cursor to the beginning and try again.
1162 */
1163 if (*cursor == 0)
1164 return (-1ULL);
1165
1166 *cursor = 0;
1167 return (metaslab_block_picker(t, cursor, size, align));
1168 }
1169
1170 /*
1171 * ==========================================================================
1172 * The first-fit block allocator
1173 * ==========================================================================
1174 */
1175 static uint64_t
1176 metaslab_ff_alloc(metaslab_t *msp, uint64_t size)
1177 {
1178 /*
1179 * Find the largest power of 2 block size that evenly divides the
1180 * requested size. This is used to try to allocate blocks with similar
1181 * alignment from the same area of the metaslab (i.e. same cursor
1182 * bucket) but it does not guarantee that other allocations sizes
1183 * may exist in the same region.
1184 */
1185 uint64_t align = size & -size;
1186 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1187 avl_tree_t *t = &msp->ms_tree->rt_root;
1188
1189 return (metaslab_block_picker(t, cursor, size, align));
1190 }
1191
1192 static metaslab_ops_t metaslab_ff_ops = {
1193 metaslab_ff_alloc
1194 };
1195
1196 /*
1197 * ==========================================================================
1198 * Dynamic block allocator -
1199 * Uses the first fit allocation scheme until space get low and then
1200 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1201 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1202 * ==========================================================================
1203 */
1204 static uint64_t
1205 metaslab_df_alloc(metaslab_t *msp, uint64_t size)
1206 {
1207 /*
1208 * Find the largest power of 2 block size that evenly divides the
1209 * requested size. This is used to try to allocate blocks with similar
1217 avl_tree_t *t = &rt->rt_root;
1218 uint64_t max_size = metaslab_block_maxsize(msp);
1219 int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
1220
1221 ASSERT(MUTEX_HELD(&msp->ms_lock));
1222 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1223
1224 if (max_size < size)
1225 return (-1ULL);
1226
1227 /*
1228 * If we're running low on space switch to using the size
1229 * sorted AVL tree (best-fit).
1230 */
1231 if (max_size < metaslab_df_alloc_threshold ||
1232 free_pct < metaslab_df_free_pct) {
1233 t = &msp->ms_size_tree;
1234 *cursor = 0;
1235 }
1236
1237 return (metaslab_block_picker(t, cursor, size, 1ULL));
1238 }
1239
1240 static metaslab_ops_t metaslab_df_ops = {
1241 metaslab_df_alloc
1242 };
1243
1244 /*
1245 * ==========================================================================
1246 * Cursor fit block allocator -
1247 * Select the largest region in the metaslab, set the cursor to the beginning
1248 * of the range and the cursor_end to the end of the range. As allocations
1249 * are made advance the cursor. Continue allocating from the cursor until
1250 * the range is exhausted and then find a new range.
1251 * ==========================================================================
1252 */
1253 static uint64_t
1254 metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
1255 {
1256 range_tree_t *rt = msp->ms_tree;
1257 avl_tree_t *t = &msp->ms_size_tree;
1258 uint64_t *cursor = &msp->ms_lbas[0];
1259 uint64_t *cursor_end = &msp->ms_lbas[1];
1260 uint64_t offset = 0;
1261
1262 ASSERT(MUTEX_HELD(&msp->ms_lock));
1263 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root));
1264
1265 ASSERT3U(*cursor_end, >=, *cursor);
1266
1267 if ((*cursor + size) > *cursor_end) {
1268 range_seg_t *rs;
1269
1270 rs = avl_last(&msp->ms_size_tree);
1271 if (rs == NULL || (rs->rs_end - rs->rs_start) < size)
1272 return (-1ULL);
1273
1274 *cursor = rs->rs_start;
1275 *cursor_end = rs->rs_end;
1276 }
1277
1278 offset = *cursor;
1279 *cursor += size;
1280
1281 return (offset);
1282 }
1283
1284 static metaslab_ops_t metaslab_cf_ops = {
1285 metaslab_cf_alloc
1286 };
1287
1288 /*
1289 * ==========================================================================
1290 * New dynamic fit allocator -
1291 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1292 * contiguous blocks. If no region is found then just use the largest segment
1293 * that remains.
1294 * ==========================================================================
1295 */
1296
1297 /*
1298 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1299 * to request from the allocator.
1300 */
1301 uint64_t metaslab_ndf_clump_shift = 4;
1302
1303 static uint64_t
1304 metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
1305 {
1306 avl_tree_t *t = &msp->ms_tree->rt_root;
1307 avl_index_t where;
1308 range_seg_t *rs, rsearch;
1309 uint64_t hbit = highbit64(size);
1310 uint64_t *cursor = &msp->ms_lbas[hbit - 1];
1311 uint64_t max_size = metaslab_block_maxsize(msp);
1312
1313 ASSERT(MUTEX_HELD(&msp->ms_lock));
1314 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1315
1316 if (max_size < size)
1317 return (-1ULL);
1318
1319 rsearch.rs_start = *cursor;
1320 rsearch.rs_end = *cursor + size;
1321
1322 rs = avl_find(t, &rsearch, &where);
1323 if (rs == NULL || (rs->rs_end - rs->rs_start) < size) {
1324 t = &msp->ms_size_tree;
1325
1326 rsearch.rs_start = 0;
1327 rsearch.rs_end = MIN(max_size,
1328 1ULL << (hbit + metaslab_ndf_clump_shift));
1329 rs = avl_find(t, &rsearch, &where);
1330 if (rs == NULL)
1331 rs = avl_nearest(t, where, AVL_AFTER);
1332 ASSERT(rs != NULL);
1333 }
1334
1335 if ((rs->rs_end - rs->rs_start) >= size) {
1336 *cursor = rs->rs_start + size;
1337 return (rs->rs_start);
1338 }
1339 return (-1ULL);
1340 }
1341
1342 static metaslab_ops_t metaslab_ndf_ops = {
1343 metaslab_ndf_alloc
1344 };
1345
1346 metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
1347
1348 /*
1349 * ==========================================================================
1350 * Metaslabs
1351 * ==========================================================================
1352 */
1353
1354 /*
1355 * Wait for any in-progress metaslab loads to complete.
1356 */
1357 void
1358 metaslab_load_wait(metaslab_t *msp)
1359 {
1360 ASSERT(MUTEX_HELD(&msp->ms_lock));
1361
1362 while (msp->ms_loading) {
1363 ASSERT(!msp->ms_loaded);
1364 cv_wait(&msp->ms_load_cv, &msp->ms_lock);
1365 }
1366 }
1367
1368 int
1369 metaslab_load(metaslab_t *msp)
1370 {
1371 int error = 0;
1372 boolean_t success = B_FALSE;
1373
1374 ASSERT(MUTEX_HELD(&msp->ms_lock));
1375 ASSERT(!msp->ms_loaded);
1376 ASSERT(!msp->ms_loading);
1377
1378 msp->ms_loading = B_TRUE;
1379 /*
1380 * Nobody else can manipulate a loading metaslab, so it's now safe
1381 * to drop the lock. This way we don't have to hold the lock while
1382 * reading the spacemap from disk.
1383 */
1384 mutex_exit(&msp->ms_lock);
1385
1386 /*
1387 * If the space map has not been allocated yet, then treat
1388 * all the space in the metaslab as free and add it to the
1389 * ms_tree.
1390 */
1391 if (msp->ms_sm != NULL)
1392 error = space_map_load(msp->ms_sm, msp->ms_tree, SM_FREE);
1393 else
1394 range_tree_add(msp->ms_tree, msp->ms_start, msp->ms_size);
1395
1396 success = (error == 0);
1397
1398 mutex_enter(&msp->ms_lock);
1399 msp->ms_loading = B_FALSE;
1400
1401 if (success) {
1402 ASSERT3P(msp->ms_group, !=, NULL);
1403 msp->ms_loaded = B_TRUE;
1404
1405 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1406 range_tree_walk(msp->ms_defertree[t],
1407 range_tree_remove, msp->ms_tree);
1408 }
1409 msp->ms_max_size = metaslab_block_maxsize(msp);
1410 }
1411 cv_broadcast(&msp->ms_load_cv);
1412 return (error);
1413 }
1414
1415 void
1416 metaslab_unload(metaslab_t *msp)
1417 {
1418 ASSERT(MUTEX_HELD(&msp->ms_lock));
1419 range_tree_vacate(msp->ms_tree, NULL, NULL);
1420 msp->ms_loaded = B_FALSE;
1421 msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
1422 msp->ms_max_size = 0;
1423 }
1424
1425 int
1426 metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg,
1427 metaslab_t **msp)
1428 {
1429 vdev_t *vd = mg->mg_vd;
1430 objset_t *mos = vd->vdev_spa->spa_meta_objset;
1431 metaslab_t *ms;
1432 int error;
1433
1434 ms = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
1435 mutex_init(&ms->ms_lock, NULL, MUTEX_DEFAULT, NULL);
1436 mutex_init(&ms->ms_sync_lock, NULL, MUTEX_DEFAULT, NULL);
1437 cv_init(&ms->ms_load_cv, NULL, CV_DEFAULT, NULL);
1438 ms->ms_id = id;
1439 ms->ms_start = id << vd->vdev_ms_shift;
1440 ms->ms_size = 1ULL << vd->vdev_ms_shift;
1441
1442 /*
1443 * We only open space map objects that already exist. All others
1444 * will be opened when we finally allocate an object for it.
1445 */
1446 if (object != 0) {
1447 error = space_map_open(&ms->ms_sm, mos, object, ms->ms_start,
1448 ms->ms_size, vd->vdev_ashift);
1449
1450 if (error != 0) {
1451 kmem_free(ms, sizeof (metaslab_t));
1452 return (error);
1453 }
1454
1455 ASSERT(ms->ms_sm != NULL);
1456 }
1457
1458 /*
1459 * We create the main range tree here, but we don't create the
1460 * other range trees until metaslab_sync_done(). This serves
1461 * two purposes: it allows metaslab_sync_done() to detect the
1462 * addition of new space; and for debugging, it ensures that we'd
1463 * data fault on any attempt to use this metaslab before it's ready.
1464 */
1465 ms->ms_tree = range_tree_create(&metaslab_rt_ops, ms);
1466 metaslab_group_add(mg, ms);
1467
1468 metaslab_set_fragmentation(ms);
1469
1470 /*
1471 * If we're opening an existing pool (txg == 0) or creating
1472 * a new one (txg == TXG_INITIAL), all space is available now.
1473 * If we're adding space to an existing pool, the new space
1474 * does not become available until after this txg has synced.
1475 * The metaslab's weight will also be initialized when we sync
1476 * out this txg. This ensures that we don't attempt to allocate
1477 * from it before we have initialized it completely.
1478 */
1479 if (txg <= TXG_INITIAL)
1480 metaslab_sync_done(ms, 0);
1481
1482 /*
1483 * If metaslab_debug_load is set and we're initializing a metaslab
1484 * that has an allocated space map object then load the its space
1485 * map so that can verify frees.
1509
1510 mutex_enter(&msp->ms_lock);
1511 VERIFY(msp->ms_group == NULL);
1512 vdev_space_update(mg->mg_vd, -space_map_allocated(msp->ms_sm),
1513 0, -msp->ms_size);
1514 space_map_close(msp->ms_sm);
1515
1516 metaslab_unload(msp);
1517 range_tree_destroy(msp->ms_tree);
1518 range_tree_destroy(msp->ms_freeingtree);
1519 range_tree_destroy(msp->ms_freedtree);
1520
1521 for (int t = 0; t < TXG_SIZE; t++) {
1522 range_tree_destroy(msp->ms_alloctree[t]);
1523 }
1524
1525 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1526 range_tree_destroy(msp->ms_defertree[t]);
1527 }
1528
1529 ASSERT0(msp->ms_deferspace);
1530
1531 mutex_exit(&msp->ms_lock);
1532 cv_destroy(&msp->ms_load_cv);
1533 mutex_destroy(&msp->ms_lock);
1534 mutex_destroy(&msp->ms_sync_lock);
1535
1536 kmem_free(msp, sizeof (metaslab_t));
1537 }
1538
1539 #define FRAGMENTATION_TABLE_SIZE 17
1540
1541 /*
1542 * This table defines a segment size based fragmentation metric that will
1543 * allow each metaslab to derive its own fragmentation value. This is done
1544 * by calculating the space in each bucket of the spacemap histogram and
1545 * multiplying that by the fragmetation metric in this table. Doing
1546 * this for all buckets and dividing it by the total amount of free
1547 * space in this metaslab (i.e. the total free space in all buckets) gives
1548 * us the fragmentation metric. This means that a high fragmentation metric
1549 * equates to most of the free space being comprised of small segments.
1550 * Conversely, if the metric is low, then most of the free space is in
1551 * large segments. A 10% change in fragmentation equates to approximately
1552 * double the number of segments.
1553 *
1554 * This table defines 0% fragmented space using 16MB segments. Testing has
1880 */
1881 should_allocate = (asize <
1882 1ULL << (WEIGHT_GET_INDEX(msp->ms_weight) + 1));
1883 } else {
1884 should_allocate = (asize <=
1885 (msp->ms_weight & ~METASLAB_WEIGHT_TYPE));
1886 }
1887 return (should_allocate);
1888 }
1889
1890 static uint64_t
1891 metaslab_weight(metaslab_t *msp)
1892 {
1893 vdev_t *vd = msp->ms_group->mg_vd;
1894 spa_t *spa = vd->vdev_spa;
1895 uint64_t weight;
1896
1897 ASSERT(MUTEX_HELD(&msp->ms_lock));
1898
1899 /*
1900 * If this vdev is in the process of being removed, there is nothing
1901 * for us to do here.
1902 */
1903 if (vd->vdev_removing)
1904 return (0);
1905
1906 metaslab_set_fragmentation(msp);
1907
1908 /*
1909 * Update the maximum size if the metaslab is loaded. This will
1910 * ensure that we get an accurate maximum size if newly freed space
1911 * has been added back into the free tree.
1912 */
1913 if (msp->ms_loaded)
1914 msp->ms_max_size = metaslab_block_maxsize(msp);
1915
1916 /*
1917 * Segment-based weighting requires space map histogram support.
1918 */
1919 if (zfs_metaslab_segment_weight_enabled &&
1920 spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
1921 (msp->ms_sm == NULL || msp->ms_sm->sm_dbuf->db_size ==
1922 sizeof (space_map_phys_t))) {
1923 weight = metaslab_segment_weight(msp);
1924 } else {
2016 if (!msp->ms_loaded)
2017 (void) metaslab_load(msp);
2018 msp->ms_selected_txg = spa_syncing_txg(spa);
2019 mutex_exit(&msp->ms_lock);
2020 }
2021
2022 static void
2023 metaslab_group_preload(metaslab_group_t *mg)
2024 {
2025 spa_t *spa = mg->mg_vd->vdev_spa;
2026 metaslab_t *msp;
2027 avl_tree_t *t = &mg->mg_metaslab_tree;
2028 int m = 0;
2029
2030 if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
2031 taskq_wait(mg->mg_taskq);
2032 return;
2033 }
2034
2035 mutex_enter(&mg->mg_lock);
2036
2037 /*
2038 * Load the next potential metaslabs
2039 */
2040 for (msp = avl_first(t); msp != NULL; msp = AVL_NEXT(t, msp)) {
2041 ASSERT3P(msp->ms_group, ==, mg);
2042
2043 /*
2044 * We preload only the maximum number of metaslabs specified
2045 * by metaslab_preload_limit. If a metaslab is being forced
2046 * to condense then we preload it too. This will ensure
2047 * that force condensing happens in the next txg.
2048 */
2049 if (++m > metaslab_preload_limit && !msp->ms_condense_wanted) {
2050 continue;
2051 }
2052
2053 VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
2054 msp, TQ_SLEEP) != NULL);
2055 }
2056 mutex_exit(&mg->mg_lock);
2057 }
2058
2059 /*
2060 * Determine if the space map's on-disk footprint is past our tolerance
2061 * for inefficiency. We would like to use the following criteria to make
2062 * our decision:
2063 *
2064 * 1. The size of the space map object should not dramatically increase as a
2065 * result of writing out the free space range tree.
2066 *
2067 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
2068 * times the size than the free space range tree representation
2069 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1MB).
2070 *
2071 * 3. The on-disk size of the space map should actually decrease.
2072 *
2073 * Checking the first condition is tricky since we don't want to walk
2074 * the entire AVL tree calculating the estimated on-disk size. Instead we
2075 * use the size-ordered range tree in the metaslab and calculate the
2076 * size required to write out the largest segment in our free tree. If the
2077 * size required to represent that segment on disk is larger than the space
2078 * map object then we avoid condensing this map.
2079 *
2080 * To determine the second criterion we use a best-case estimate and assume
2081 * each segment can be represented on-disk as a single 64-bit entry. We refer
2082 * to this best-case estimate as the space map's minimal form.
2083 *
2084 * Unfortunately, we cannot compute the on-disk size of the space map in this
2085 * context because we cannot accurately compute the effects of compression, etc.
2086 * Instead, we apply the heuristic described in the block comment for
2087 * zfs_metaslab_condense_block_threshold - we only condense if the space used
2088 * is greater than a threshold number of blocks.
2089 */
2146 ASSERT3U(spa_sync_pass(spa), ==, 1);
2147 ASSERT(msp->ms_loaded);
2148
2149
2150 spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, vdev id %llu, "
2151 "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg,
2152 msp->ms_id, msp, msp->ms_group->mg_vd->vdev_id,
2153 msp->ms_group->mg_vd->vdev_spa->spa_name,
2154 space_map_length(msp->ms_sm), avl_numnodes(&msp->ms_tree->rt_root),
2155 msp->ms_condense_wanted ? "TRUE" : "FALSE");
2156
2157 msp->ms_condense_wanted = B_FALSE;
2158
2159 /*
2160 * Create an range tree that is 100% allocated. We remove segments
2161 * that have been freed in this txg, any deferred frees that exist,
2162 * and any allocation in the future. Removing segments should be
2163 * a relatively inexpensive operation since we expect these trees to
2164 * have a small number of nodes.
2165 */
2166 condense_tree = range_tree_create(NULL, NULL);
2167 range_tree_add(condense_tree, msp->ms_start, msp->ms_size);
2168
2169 /*
2170 * Remove what's been freed in this txg from the condense_tree.
2171 * Since we're in sync_pass 1, we know that all the frees from
2172 * this txg are in the freeingtree.
2173 */
2174 range_tree_walk(msp->ms_freeingtree, range_tree_remove, condense_tree);
2175
2176 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2177 range_tree_walk(msp->ms_defertree[t],
2178 range_tree_remove, condense_tree);
2179 }
2180
2181 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
2182 range_tree_walk(msp->ms_alloctree[(txg + t) & TXG_MASK],
2183 range_tree_remove, condense_tree);
2184 }
2185
2186 /*
2187 * We're about to drop the metaslab's lock thus allowing
2188 * other consumers to change it's content. Set the
2189 * metaslab's ms_condensing flag to ensure that
2190 * allocations on this metaslab do not occur while we're
2191 * in the middle of committing it to disk. This is only critical
2192 * for the ms_tree as all other range trees use per txg
2193 * views of their content.
2194 */
2195 msp->ms_condensing = B_TRUE;
2196
2197 mutex_exit(&msp->ms_lock);
2198 space_map_truncate(sm, tx);
2199
2200 /*
2201 * While we would ideally like to create a space map representation
2202 * that consists only of allocation records, doing so can be
2203 * prohibitively expensive because the in-core free tree can be
2204 * large, and therefore computationally expensive to subtract
2205 * from the condense_tree. Instead we sync out two trees, a cheap
2206 * allocation only tree followed by the in-core free tree. While not
2207 * optimal, this is typically close to optimal, and much cheaper to
2208 * compute.
2209 */
2210 space_map_write(sm, condense_tree, SM_ALLOC, tx);
2211 range_tree_vacate(condense_tree, NULL, NULL);
2212 range_tree_destroy(condense_tree);
2213
2214 space_map_write(sm, msp->ms_tree, SM_FREE, tx);
2215 mutex_enter(&msp->ms_lock);
2216 msp->ms_condensing = B_FALSE;
2217 }
2218
2219 /*
2220 * Write a metaslab to disk in the context of the specified transaction group.
2221 */
2222 void
2223 metaslab_sync(metaslab_t *msp, uint64_t txg)
2224 {
2225 metaslab_group_t *mg = msp->ms_group;
2226 vdev_t *vd = mg->mg_vd;
2227 spa_t *spa = vd->vdev_spa;
2228 objset_t *mos = spa_meta_objset(spa);
2229 range_tree_t *alloctree = msp->ms_alloctree[txg & TXG_MASK];
2230 dmu_tx_t *tx;
2231 uint64_t object = space_map_object(msp->ms_sm);
2232
2233 ASSERT(!vd->vdev_ishole);
2234
2235 /*
2236 * This metaslab has just been added so there's no work to do now.
2237 */
2238 if (msp->ms_freeingtree == NULL) {
2239 ASSERT3P(alloctree, ==, NULL);
2240 return;
2241 }
2242
2243 ASSERT3P(alloctree, !=, NULL);
2244 ASSERT3P(msp->ms_freeingtree, !=, NULL);
2245 ASSERT3P(msp->ms_freedtree, !=, NULL);
2246
2247 /*
2248 * Normally, we don't want to process a metaslab if there
2249 * are no allocations or frees to perform. However, if the metaslab
2250 * is being forced to condense and it's loaded, we need to let it
2251 * through.
2252 */
2253 if (range_tree_space(alloctree) == 0 &&
2254 range_tree_space(msp->ms_freeingtree) == 0 &&
2255 !(msp->ms_loaded && msp->ms_condense_wanted))
2256 return;
2257
2258
2259 VERIFY(txg <= spa_final_dirty_txg(spa));
2260
2261 /*
2262 * The only state that can actually be changing concurrently with
2263 * metaslab_sync() is the metaslab's ms_tree. No other thread can
2264 * be modifying this txg's alloctree, freeingtree, freedtree, or
2265 * space_map_phys_t. We drop ms_lock whenever we could call
2266 * into the DMU, because the DMU can call down to us
2267 * (e.g. via zio_free()) at any time.
2268 *
2269 * The spa_vdev_remove_thread() can be reading metaslab state
2270 * concurrently, and it is locked out by the ms_sync_lock. Note
2271 * that the ms_lock is insufficient for this, because it is dropped
2272 * by space_map_write().
2273 */
2274
2275 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2276
2277 if (msp->ms_sm == NULL) {
2278 uint64_t new_object;
2279
2280 new_object = space_map_alloc(mos, tx);
2281 VERIFY3U(new_object, !=, 0);
2282
2283 VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
2284 msp->ms_start, msp->ms_size, vd->vdev_ashift));
2285 ASSERT(msp->ms_sm != NULL);
2286 }
2287
2288 mutex_enter(&msp->ms_sync_lock);
2289 mutex_enter(&msp->ms_lock);
2290
2291 /*
2292 * Note: metaslab_condense() clears the space map's histogram.
2293 * Therefore we must verify and remove this histogram before
2294 * condensing.
2295 */
2296 metaslab_group_histogram_verify(mg);
2297 metaslab_class_histogram_verify(mg->mg_class);
2298 metaslab_group_histogram_remove(mg, msp);
2299
2300 if (msp->ms_loaded && spa_sync_pass(spa) == 1 &&
2301 metaslab_should_condense(msp)) {
2302 metaslab_condense(msp, txg, tx);
2303 } else {
2304 mutex_exit(&msp->ms_lock);
2305 space_map_write(msp->ms_sm, alloctree, SM_ALLOC, tx);
2306 space_map_write(msp->ms_sm, msp->ms_freeingtree, SM_FREE, tx);
2307 mutex_enter(&msp->ms_lock);
2308 }
2309
2310 if (msp->ms_loaded) {
2311 /*
2312 * When the space map is loaded, we have an accurate
2313 * histogram in the range tree. This gives us an opportunity
2314 * to bring the space map's histogram up-to-date so we clear
2315 * it first before updating it.
2316 */
2317 space_map_histogram_clear(msp->ms_sm);
2318 space_map_histogram_add(msp->ms_sm, msp->ms_tree, tx);
2319
2320 /*
2321 * Since we've cleared the histogram we need to add back
2322 * any free space that has already been processed, plus
2323 * any deferred space. This allows the on-disk histogram
2324 * to accurately reflect all free space even if some space
2325 * is not yet available for allocation (i.e. deferred).
2326 */
2327 space_map_histogram_add(msp->ms_sm, msp->ms_freedtree, tx);
2328
2329 /*
2330 * Add back any deferred free space that has not been
2331 * added back into the in-core free tree yet. This will
2332 * ensure that we don't end up with a space map histogram
2360 */
2361 if (spa_sync_pass(spa) == 1) {
2362 range_tree_swap(&msp->ms_freeingtree, &msp->ms_freedtree);
2363 } else {
2364 range_tree_vacate(msp->ms_freeingtree,
2365 range_tree_add, msp->ms_freedtree);
2366 }
2367 range_tree_vacate(alloctree, NULL, NULL);
2368
2369 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
2370 ASSERT0(range_tree_space(msp->ms_alloctree[TXG_CLEAN(txg) & TXG_MASK]));
2371 ASSERT0(range_tree_space(msp->ms_freeingtree));
2372
2373 mutex_exit(&msp->ms_lock);
2374
2375 if (object != space_map_object(msp->ms_sm)) {
2376 object = space_map_object(msp->ms_sm);
2377 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
2378 msp->ms_id, sizeof (uint64_t), &object, tx);
2379 }
2380 mutex_exit(&msp->ms_sync_lock);
2381 dmu_tx_commit(tx);
2382 }
2383
2384 /*
2385 * Called after a transaction group has completely synced to mark
2386 * all of the metaslab's free space as usable.
2387 */
2388 void
2389 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
2390 {
2391 metaslab_group_t *mg = msp->ms_group;
2392 vdev_t *vd = mg->mg_vd;
2393 spa_t *spa = vd->vdev_spa;
2394 range_tree_t **defer_tree;
2395 int64_t alloc_delta, defer_delta;
2396 boolean_t defer_allowed = B_TRUE;
2397
2398 ASSERT(!vd->vdev_ishole);
2399
2400 mutex_enter(&msp->ms_lock);
2401
2402 /*
2403 * If this metaslab is just becoming available, initialize its
2404 * range trees and add its capacity to the vdev.
2405 */
2406 if (msp->ms_freedtree == NULL) {
2407 for (int t = 0; t < TXG_SIZE; t++) {
2408 ASSERT(msp->ms_alloctree[t] == NULL);
2409
2410 msp->ms_alloctree[t] = range_tree_create(NULL, NULL);
2411 }
2412
2413 ASSERT3P(msp->ms_freeingtree, ==, NULL);
2414 msp->ms_freeingtree = range_tree_create(NULL, NULL);
2415
2416 ASSERT3P(msp->ms_freedtree, ==, NULL);
2417 msp->ms_freedtree = range_tree_create(NULL, NULL);
2418
2419 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2420 ASSERT(msp->ms_defertree[t] == NULL);
2421
2422 msp->ms_defertree[t] = range_tree_create(NULL, NULL);
2423 }
2424
2425 vdev_space_update(vd, 0, 0, msp->ms_size);
2426 }
2427
2428 defer_tree = &msp->ms_defertree[txg % TXG_DEFER_SIZE];
2429
2430 uint64_t free_space = metaslab_class_get_space(spa_normal_class(spa)) -
2431 metaslab_class_get_alloc(spa_normal_class(spa));
2432 if (free_space <= spa_get_slop_space(spa) || vd->vdev_removing) {
2433 defer_allowed = B_FALSE;
2434 }
2435
2436 defer_delta = 0;
2437 alloc_delta = space_map_alloc_delta(msp->ms_sm);
2438 if (defer_allowed) {
2439 defer_delta = range_tree_space(msp->ms_freedtree) -
2440 range_tree_space(*defer_tree);
2441 } else {
2442 defer_delta -= range_tree_space(*defer_tree);
2443 }
2444
2445 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
2446
2447 /*
2448 * If there's a metaslab_load() in progress, wait for it to complete
2449 * so that we have a consistent view of the in-core space map.
2450 */
2451 metaslab_load_wait(msp);
2452
2453 /*
2454 * Move the frees from the defer_tree back to the free
2455 * range tree (if it's loaded). Swap the freed_tree and the
2456 * defer_tree -- this is safe to do because we've just emptied out
2457 * the defer_tree.
2458 */
2459 range_tree_vacate(*defer_tree,
2460 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
2461 if (defer_allowed) {
2462 range_tree_swap(&msp->ms_freedtree, defer_tree);
2463 } else {
2464 range_tree_vacate(msp->ms_freedtree,
2465 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
2466 }
2467
2468 space_map_update(msp->ms_sm);
2469
2470 msp->ms_deferspace += defer_delta;
2471 ASSERT3S(msp->ms_deferspace, >=, 0);
2472 ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
2473 if (msp->ms_deferspace != 0) {
2474 /*
2475 * Keep syncing this metaslab until all deferred frees
2476 * are back in circulation.
2477 */
2478 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
2482 * Calculate the new weights before unloading any metaslabs.
2483 * This will give us the most accurate weighting.
2484 */
2485 metaslab_group_sort(mg, msp, metaslab_weight(msp));
2486
2487 /*
2488 * If the metaslab is loaded and we've not tried to load or allocate
2489 * from it in 'metaslab_unload_delay' txgs, then unload it.
2490 */
2491 if (msp->ms_loaded &&
2492 msp->ms_selected_txg + metaslab_unload_delay < txg) {
2493 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
2494 VERIFY0(range_tree_space(
2495 msp->ms_alloctree[(txg + t) & TXG_MASK]));
2496 }
2497
2498 if (!metaslab_debug_unload)
2499 metaslab_unload(msp);
2500 }
2501
2502 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
2503 ASSERT0(range_tree_space(msp->ms_freeingtree));
2504 ASSERT0(range_tree_space(msp->ms_freedtree));
2505
2506 mutex_exit(&msp->ms_lock);
2507 }
2508
2509 void
2510 metaslab_sync_reassess(metaslab_group_t *mg)
2511 {
2512 spa_t *spa = mg->mg_class->mc_spa;
2513
2514 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
2515 metaslab_group_alloc_update(mg);
2516 mg->mg_fragmentation = metaslab_group_fragmentation(mg);
2517
2518 /*
2519 * Preload the next potential metaslabs but only on active
2520 * metaslab groups. We can get into a state where the metaslab
2521 * is no longer active since we dirty metaslabs as we remove a
2522 * a device, thus potentially making the metaslab group eligible
2523 * for preloading.
2524 */
2525 if (mg->mg_activation_count > 0) {
2526 metaslab_group_preload(mg);
2527 }
2528 spa_config_exit(spa, SCL_ALLOC, FTAG);
2529 }
2530
2531 static uint64_t
2532 metaslab_distance(metaslab_t *msp, dva_t *dva)
2533 {
2534 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
2535 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
2536 uint64_t start = msp->ms_id;
2537
2538 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
2539 return (1ULL << 63);
2540
2541 if (offset < start)
2542 return ((start - offset) << ms_shift);
2543 if (offset > start)
2544 return ((offset - start) << ms_shift);
2545 return (0);
2546 }
2547
2548 /*
2702 }
2703
2704 static uint64_t
2705 metaslab_block_alloc(metaslab_t *msp, uint64_t size, uint64_t txg)
2706 {
2707 uint64_t start;
2708 range_tree_t *rt = msp->ms_tree;
2709 metaslab_class_t *mc = msp->ms_group->mg_class;
2710
2711 VERIFY(!msp->ms_condensing);
2712
2713 start = mc->mc_ops->msop_alloc(msp, size);
2714 if (start != -1ULL) {
2715 metaslab_group_t *mg = msp->ms_group;
2716 vdev_t *vd = mg->mg_vd;
2717
2718 VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
2719 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2720 VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
2721 range_tree_remove(rt, start, size);
2722
2723 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2724 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
2725
2726 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], start, size);
2727
2728 /* Track the last successful allocation */
2729 msp->ms_alloc_txg = txg;
2730 metaslab_verify_space(msp, txg);
2731 }
2732
2733 /*
2734 * Now that we've attempted the allocation we need to update the
2735 * metaslab's maximum block size since it may have changed.
2736 */
2737 msp->ms_max_size = metaslab_block_maxsize(msp);
2738 return (start);
2739 }
2740
2741 static uint64_t
2742 metaslab_group_alloc_normal(metaslab_group_t *mg, zio_alloc_list_t *zal,
2743 uint64_t asize, uint64_t txg, uint64_t min_distance, dva_t *dva, int d)
2744 {
2745 metaslab_t *msp = NULL;
2746 uint64_t offset = -1ULL;
2747 uint64_t activation_weight;
2748 uint64_t target_distance;
2749 int i;
2750
2751 activation_weight = METASLAB_WEIGHT_PRIMARY;
2752 for (i = 0; i < d; i++) {
2753 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
2754 activation_weight = METASLAB_WEIGHT_SECONDARY;
2755 break;
2756 }
2757 }
2758
2759 metaslab_t *search = kmem_alloc(sizeof (*search), KM_SLEEP);
2760 search->ms_weight = UINT64_MAX;
2761 search->ms_start = 0;
2762 for (;;) {
2763 boolean_t was_active;
2764 avl_tree_t *t = &mg->mg_metaslab_tree;
2765 avl_index_t idx;
2766
2767 mutex_enter(&mg->mg_lock);
2768
2769 /*
2770 * Find the metaslab with the highest weight that is less
2771 * than what we've already tried. In the common case, this
2772 * means that we will examine each metaslab at most once.
2773 * Note that concurrent callers could reorder metaslabs
2774 * by activation/passivation once we have dropped the mg_lock.
2775 * If a metaslab is activated by another thread, and we fail
2776 * to allocate from the metaslab we have selected, we may
2777 * not try the newly-activated metaslab, and instead activate
2778 * another metaslab. This is not optimal, but generally
2779 * does not cause any problems (a possible exception being
2780 * if every metaslab is completely full except for the
2781 * the newly-activated metaslab which we fail to examine).
2782 */
2783 msp = avl_find(t, search, &idx);
2784 if (msp == NULL)
2785 msp = avl_nearest(t, idx, AVL_AFTER);
2786 for (; msp != NULL; msp = AVL_NEXT(t, msp)) {
2787
2788 if (!metaslab_should_allocate(msp, asize)) {
2789 metaslab_trace_add(zal, mg, msp, asize, d,
2790 TRACE_TOO_SMALL);
2791 continue;
2792 }
2793
2794 /*
2795 * If the selected metaslab is condensing, skip it.
2796 */
2797 if (msp->ms_condensing)
2798 continue;
2799
2800 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
2801 if (activation_weight == METASLAB_WEIGHT_PRIMARY)
2802 break;
2803
2804 target_distance = min_distance +
2805 (space_map_allocated(msp->ms_sm) != 0 ? 0 :
2806 min_distance >> 1);
2807
2808 for (i = 0; i < d; i++) {
2809 if (metaslab_distance(msp, &dva[i]) <
2810 target_distance)
2811 break;
2812 }
2813 if (i == d)
2814 break;
2815 }
2816 mutex_exit(&mg->mg_lock);
2817 if (msp == NULL) {
2818 kmem_free(search, sizeof (*search));
2819 return (-1ULL);
2820 }
2821 search->ms_weight = msp->ms_weight;
2822 search->ms_start = msp->ms_start + 1;
2823
2824 mutex_enter(&msp->ms_lock);
2825
2826 /*
2827 * Ensure that the metaslab we have selected is still
2828 * capable of handling our request. It's possible that
2829 * another thread may have changed the weight while we
2830 * were blocked on the metaslab lock. We check the
2831 * active status first to see if we need to reselect
2832 * a new metaslab.
2833 */
2834 if (was_active && !(msp->ms_weight & METASLAB_ACTIVE_MASK)) {
2835 mutex_exit(&msp->ms_lock);
2916 metaslab_passivate(msp,
2917 metaslab_weight_from_range_tree(msp));
2918 }
2919
2920 /*
2921 * We have just failed an allocation attempt, check
2922 * that metaslab_should_allocate() agrees. Otherwise,
2923 * we may end up in an infinite loop retrying the same
2924 * metaslab.
2925 */
2926 ASSERT(!metaslab_should_allocate(msp, asize));
2927 mutex_exit(&msp->ms_lock);
2928 }
2929 mutex_exit(&msp->ms_lock);
2930 kmem_free(search, sizeof (*search));
2931 return (offset);
2932 }
2933
2934 static uint64_t
2935 metaslab_group_alloc(metaslab_group_t *mg, zio_alloc_list_t *zal,
2936 uint64_t asize, uint64_t txg, uint64_t min_distance, dva_t *dva, int d)
2937 {
2938 uint64_t offset;
2939 ASSERT(mg->mg_initialized);
2940
2941 offset = metaslab_group_alloc_normal(mg, zal, asize, txg,
2942 min_distance, dva, d);
2943
2944 mutex_enter(&mg->mg_lock);
2945 if (offset == -1ULL) {
2946 mg->mg_failed_allocations++;
2947 metaslab_trace_add(zal, mg, NULL, asize, d,
2948 TRACE_GROUP_FAILURE);
2949 if (asize == SPA_GANGBLOCKSIZE) {
2950 /*
2951 * This metaslab group was unable to allocate
2952 * the minimum gang block size so it must be out of
2953 * space. We must notify the allocation throttle
2954 * to start skipping allocation attempts to this
2955 * metaslab group until more space becomes available.
2956 * Note: this failure cannot be caused by the
2957 * allocation throttle since the allocation throttle
2958 * is only responsible for skipping devices and
2959 * not failing block allocations.
2960 */
2961 mg->mg_no_free_space = B_TRUE;
2962 }
2963 }
2964 mg->mg_allocations++;
2965 mutex_exit(&mg->mg_lock);
2966 return (offset);
2967 }
2968
2969 /*
2970 * If we have to write a ditto block (i.e. more than one DVA for a given BP)
2971 * on the same vdev as an existing DVA of this BP, then try to allocate it
2972 * at least (vdev_asize / (2 ^ ditto_same_vdev_distance_shift)) away from the
2973 * existing DVAs.
2974 */
2975 int ditto_same_vdev_distance_shift = 3;
2976
2977 /*
2978 * Allocate a block for the specified i/o.
2979 */
2980 int
2981 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
2982 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags,
2983 zio_alloc_list_t *zal)
2984 {
2985 metaslab_group_t *mg, *rotor;
2986 vdev_t *vd;
2987 boolean_t try_hard = B_FALSE;
2988
2989 ASSERT(!DVA_IS_VALID(&dva[d]));
2990
2991 /*
2992 * For testing, make some blocks above a certain size be gang blocks.
2993 */
2994 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0) {
2995 metaslab_trace_add(zal, NULL, NULL, psize, d, TRACE_FORCE_GANG);
2996 return (SET_ERROR(ENOSPC));
2997 }
2998
2999 /*
3000 * Start at the rotor and loop through all mgs until we find something.
3006 * consecutive vdevs. If we're forced to reuse a vdev before we've
3007 * allocated all of our ditto blocks, then try and spread them out on
3008 * that vdev as much as possible. If it turns out to not be possible,
3009 * gradually lower our standards until anything becomes acceptable.
3010 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
3011 * gives us hope of containing our fault domains to something we're
3012 * able to reason about. Otherwise, any two top-level vdev failures
3013 * will guarantee the loss of data. With consecutive allocation,
3014 * only two adjacent top-level vdev failures will result in data loss.
3015 *
3016 * If we are doing gang blocks (hintdva is non-NULL), try to keep
3017 * ourselves on the same vdev as our gang block header. That
3018 * way, we can hope for locality in vdev_cache, plus it makes our
3019 * fault domains something tractable.
3020 */
3021 if (hintdva) {
3022 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
3023
3024 /*
3025 * It's possible the vdev we're using as the hint no
3026 * longer exists or its mg has been closed (e.g. by
3027 * device removal). Consult the rotor when
3028 * all else fails.
3029 */
3030 if (vd != NULL && vd->vdev_mg != NULL) {
3031 mg = vd->vdev_mg;
3032
3033 if (flags & METASLAB_HINTBP_AVOID &&
3034 mg->mg_next != NULL)
3035 mg = mg->mg_next;
3036 } else {
3037 mg = mc->mc_rotor;
3038 }
3039 } else if (d != 0) {
3040 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
3041 mg = vd->vdev_mg->mg_next;
3042 } else {
3043 mg = mc->mc_rotor;
3044 }
3045
3046 /*
3047 * If the hint put us into the wrong metaslab class, or into a
3048 * metaslab group that has been passivated, just follow the rotor.
3049 */
3050 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
3105 ASSERT(mg->mg_class == mc);
3106
3107 /*
3108 * If we don't need to try hard, then require that the
3109 * block be 1/8th of the device away from any other DVAs
3110 * in this BP. If we are trying hard, allow any offset
3111 * to be used (distance=0).
3112 */
3113 uint64_t distance = 0;
3114 if (!try_hard) {
3115 distance = vd->vdev_asize >>
3116 ditto_same_vdev_distance_shift;
3117 if (distance <= (1ULL << vd->vdev_ms_shift))
3118 distance = 0;
3119 }
3120
3121 uint64_t asize = vdev_psize_to_asize(vd, psize);
3122 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
3123
3124 uint64_t offset = metaslab_group_alloc(mg, zal, asize, txg,
3125 distance, dva, d);
3126
3127 if (offset != -1ULL) {
3128 /*
3129 * If we've just selected this metaslab group,
3130 * figure out whether the corresponding vdev is
3131 * over- or under-used relative to the pool,
3132 * and set an allocation bias to even it out.
3133 */
3134 if (mc->mc_aliquot == 0 && metaslab_bias_enabled) {
3135 vdev_stat_t *vs = &vd->vdev_stat;
3136 int64_t vu, cu;
3137
3138 vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
3139 cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
3140
3141 /*
3142 * Calculate how much more or less we should
3143 * try to allocate from this device during
3144 * this iteration around the rotor.
3145 * For example, if a device is 80% full
3146 * and the pool is 20% full then we should
3147 * reduce allocations by 60% on this device.
3148 *
3149 * mg_bias = (20 - 80) * 512K / 100 = -307K
3150 *
3151 * This reduces allocations by 307K for this
3152 * iteration.
3153 */
3154 mg->mg_bias = ((cu - vu) *
3155 (int64_t)mg->mg_aliquot) / 100;
3156 } else if (!metaslab_bias_enabled) {
3157 mg->mg_bias = 0;
3158 }
3159
3160 if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
3161 mg->mg_aliquot + mg->mg_bias) {
3162 mc->mc_rotor = mg->mg_next;
3163 mc->mc_aliquot = 0;
3164 }
3165
3166 DVA_SET_VDEV(&dva[d], vd->vdev_id);
3167 DVA_SET_OFFSET(&dva[d], offset);
3168 DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
3169 DVA_SET_ASIZE(&dva[d], asize);
3170
3171 return (0);
3172 }
3173 next:
3174 mc->mc_rotor = mg->mg_next;
3175 mc->mc_aliquot = 0;
3176 } while ((mg = mg->mg_next) != rotor);
3177
3178 /*
3179 * If we haven't tried hard, do so now.
3180 */
3181 if (!try_hard) {
3182 try_hard = B_TRUE;
3183 goto top;
3184 }
3185
3186 bzero(&dva[d], sizeof (dva_t));
3187
3188 metaslab_trace_add(zal, rotor, NULL, psize, d, TRACE_ENOSPC);
3189 return (SET_ERROR(ENOSPC));
3190 }
3191
3192 void
3193 metaslab_free_concrete(vdev_t *vd, uint64_t offset, uint64_t asize,
3194 uint64_t txg)
3195 {
3196 metaslab_t *msp;
3197 spa_t *spa = vd->vdev_spa;
3198
3199 ASSERT3U(txg, ==, spa->spa_syncing_txg);
3200 ASSERT(vdev_is_concrete(vd));
3201 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
3202 ASSERT3U(offset >> vd->vdev_ms_shift, <, vd->vdev_ms_count);
3203
3204 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3205
3206 VERIFY(!msp->ms_condensing);
3207 VERIFY3U(offset, >=, msp->ms_start);
3208 VERIFY3U(offset + asize, <=, msp->ms_start + msp->ms_size);
3209 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
3210 VERIFY0(P2PHASE(asize, 1ULL << vd->vdev_ashift));
3211
3212 metaslab_check_free_impl(vd, offset, asize);
3213 mutex_enter(&msp->ms_lock);
3214 if (range_tree_space(msp->ms_freeingtree) == 0) {
3215 vdev_dirty(vd, VDD_METASLAB, msp, txg);
3216 }
3217 range_tree_add(msp->ms_freeingtree, offset, asize);
3218 mutex_exit(&msp->ms_lock);
3219 }
3220
3221 /* ARGSUSED */
3222 void
3223 metaslab_free_impl_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset,
3224 uint64_t size, void *arg)
3225 {
3226 uint64_t *txgp = arg;
3227
3228 if (vd->vdev_ops->vdev_op_remap != NULL)
3229 vdev_indirect_mark_obsolete(vd, offset, size, *txgp);
3230 else
3231 metaslab_free_impl(vd, offset, size, *txgp);
3232 }
3233
3234 static void
3235 metaslab_free_impl(vdev_t *vd, uint64_t offset, uint64_t size,
3236 uint64_t txg)
3237 {
3238 spa_t *spa = vd->vdev_spa;
3239
3240 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
3241
3242 if (txg > spa_freeze_txg(spa))
3243 return;
3244
3245 if (spa->spa_vdev_removal != NULL &&
3246 spa->spa_vdev_removal->svr_vdev == vd &&
3247 vdev_is_concrete(vd)) {
3248 /*
3249 * Note: we check if the vdev is concrete because when
3250 * we complete the removal, we first change the vdev to be
3251 * an indirect vdev (in open context), and then (in syncing
3252 * context) clear spa_vdev_removal.
3253 */
3254 free_from_removing_vdev(vd, offset, size, txg);
3255 } else if (vd->vdev_ops->vdev_op_remap != NULL) {
3256 vdev_indirect_mark_obsolete(vd, offset, size, txg);
3257 vd->vdev_ops->vdev_op_remap(vd, offset, size,
3258 metaslab_free_impl_cb, &txg);
3259 } else {
3260 metaslab_free_concrete(vd, offset, size, txg);
3261 }
3262 }
3263
3264 typedef struct remap_blkptr_cb_arg {
3265 blkptr_t *rbca_bp;
3266 spa_remap_cb_t rbca_cb;
3267 vdev_t *rbca_remap_vd;
3268 uint64_t rbca_remap_offset;
3269 void *rbca_cb_arg;
3270 } remap_blkptr_cb_arg_t;
3271
3272 void
3273 remap_blkptr_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset,
3274 uint64_t size, void *arg)
3275 {
3276 remap_blkptr_cb_arg_t *rbca = arg;
3277 blkptr_t *bp = rbca->rbca_bp;
3278
3279 /* We can not remap split blocks. */
3280 if (size != DVA_GET_ASIZE(&bp->blk_dva[0]))
3281 return;
3282 ASSERT0(inner_offset);
3283
3284 if (rbca->rbca_cb != NULL) {
3285 /*
3286 * At this point we know that we are not handling split
3287 * blocks and we invoke the callback on the previous
3288 * vdev which must be indirect.
3289 */
3290 ASSERT3P(rbca->rbca_remap_vd->vdev_ops, ==, &vdev_indirect_ops);
3291
3292 rbca->rbca_cb(rbca->rbca_remap_vd->vdev_id,
3293 rbca->rbca_remap_offset, size, rbca->rbca_cb_arg);
3294
3295 /* set up remap_blkptr_cb_arg for the next call */
3296 rbca->rbca_remap_vd = vd;
3297 rbca->rbca_remap_offset = offset;
3298 }
3299
3300 /*
3301 * The phys birth time is that of dva[0]. This ensures that we know
3302 * when each dva was written, so that resilver can determine which
3303 * blocks need to be scrubbed (i.e. those written during the time
3304 * the vdev was offline). It also ensures that the key used in
3305 * the ARC hash table is unique (i.e. dva[0] + phys_birth). If
3306 * we didn't change the phys_birth, a lookup in the ARC for a
3307 * remapped BP could find the data that was previously stored at
3308 * this vdev + offset.
3309 */
3310 vdev_t *oldvd = vdev_lookup_top(vd->vdev_spa,
3311 DVA_GET_VDEV(&bp->blk_dva[0]));
3312 vdev_indirect_births_t *vib = oldvd->vdev_indirect_births;
3313 bp->blk_phys_birth = vdev_indirect_births_physbirth(vib,
3314 DVA_GET_OFFSET(&bp->blk_dva[0]), DVA_GET_ASIZE(&bp->blk_dva[0]));
3315
3316 DVA_SET_VDEV(&bp->blk_dva[0], vd->vdev_id);
3317 DVA_SET_OFFSET(&bp->blk_dva[0], offset);
3318 }
3319
3320 /*
3321 * If the block pointer contains any indirect DVAs, modify them to refer to
3322 * concrete DVAs. Note that this will sometimes not be possible, leaving
3323 * the indirect DVA in place. This happens if the indirect DVA spans multiple
3324 * segments in the mapping (i.e. it is a "split block").
3325 *
3326 * If the BP was remapped, calls the callback on the original dva (note the
3327 * callback can be called multiple times if the original indirect DVA refers
3328 * to another indirect DVA, etc).
3329 *
3330 * Returns TRUE if the BP was remapped.
3331 */
3332 boolean_t
3333 spa_remap_blkptr(spa_t *spa, blkptr_t *bp, spa_remap_cb_t callback, void *arg)
3334 {
3335 remap_blkptr_cb_arg_t rbca;
3336
3337 if (!zfs_remap_blkptr_enable)
3338 return (B_FALSE);
3339
3340 if (!spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS))
3341 return (B_FALSE);
3342
3343 /*
3344 * Dedup BP's can not be remapped, because ddt_phys_select() depends
3345 * on DVA[0] being the same in the BP as in the DDT (dedup table).
3346 */
3347 if (BP_GET_DEDUP(bp))
3348 return (B_FALSE);
3349
3350 /*
3351 * Gang blocks can not be remapped, because
3352 * zio_checksum_gang_verifier() depends on the DVA[0] that's in
3353 * the BP used to read the gang block header (GBH) being the same
3354 * as the DVA[0] that we allocated for the GBH.
3355 */
3356 if (BP_IS_GANG(bp))
3357 return (B_FALSE);
3358
3359 /*
3360 * Embedded BP's have no DVA to remap.
3361 */
3362 if (BP_GET_NDVAS(bp) < 1)
3363 return (B_FALSE);
3364
3365 /*
3366 * Note: we only remap dva[0]. If we remapped other dvas, we
3367 * would no longer know what their phys birth txg is.
3368 */
3369 dva_t *dva = &bp->blk_dva[0];
3370
3371 uint64_t offset = DVA_GET_OFFSET(dva);
3372 uint64_t size = DVA_GET_ASIZE(dva);
3373 vdev_t *vd = vdev_lookup_top(spa, DVA_GET_VDEV(dva));
3374
3375 if (vd->vdev_ops->vdev_op_remap == NULL)
3376 return (B_FALSE);
3377
3378 rbca.rbca_bp = bp;
3379 rbca.rbca_cb = callback;
3380 rbca.rbca_remap_vd = vd;
3381 rbca.rbca_remap_offset = offset;
3382 rbca.rbca_cb_arg = arg;
3383
3384 /*
3385 * remap_blkptr_cb() will be called in order for each level of
3386 * indirection, until a concrete vdev is reached or a split block is
3387 * encountered. old_vd and old_offset are updated within the callback
3388 * as we go from the one indirect vdev to the next one (either concrete
3389 * or indirect again) in that order.
3390 */
3391 vd->vdev_ops->vdev_op_remap(vd, offset, size, remap_blkptr_cb, &rbca);
3392
3393 /* Check if the DVA wasn't remapped because it is a split block */
3394 if (DVA_GET_VDEV(&rbca.rbca_bp->blk_dva[0]) == vd->vdev_id)
3395 return (B_FALSE);
3396
3397 return (B_TRUE);
3398 }
3399
3400 /*
3401 * Undo the allocation of a DVA which happened in the given transaction group.
3402 */
3403 void
3404 metaslab_unalloc_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
3405 {
3406 metaslab_t *msp;
3407 vdev_t *vd;
3408 uint64_t vdev = DVA_GET_VDEV(dva);
3409 uint64_t offset = DVA_GET_OFFSET(dva);
3410 uint64_t size = DVA_GET_ASIZE(dva);
3411
3412 ASSERT(DVA_IS_VALID(dva));
3413 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
3414
3415 if (txg > spa_freeze_txg(spa))
3416 return;
3417
3418 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
3419 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
3420 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
3421 (u_longlong_t)vdev, (u_longlong_t)offset);
3422 ASSERT(0);
3423 return;
3424 }
3425
3426 ASSERT(!vd->vdev_removing);
3427 ASSERT(vdev_is_concrete(vd));
3428 ASSERT0(vd->vdev_indirect_config.vic_mapping_object);
3429 ASSERT3P(vd->vdev_indirect_mapping, ==, NULL);
3430
3431 if (DVA_GET_GANG(dva))
3432 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
3433
3434 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3435
3436 mutex_enter(&msp->ms_lock);
3437 range_tree_remove(msp->ms_alloctree[txg & TXG_MASK],
3438 offset, size);
3439
3440 VERIFY(!msp->ms_condensing);
3441 VERIFY3U(offset, >=, msp->ms_start);
3442 VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
3443 VERIFY3U(range_tree_space(msp->ms_tree) + size, <=,
3444 msp->ms_size);
3445 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
3446 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
3447 range_tree_add(msp->ms_tree, offset, size);
3448 mutex_exit(&msp->ms_lock);
3449 }
3450
3451 /*
3452 * Free the block represented by DVA in the context of the specified
3453 * transaction group.
3454 */
3455 void
3456 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
3457 {
3458 uint64_t vdev = DVA_GET_VDEV(dva);
3459 uint64_t offset = DVA_GET_OFFSET(dva);
3460 uint64_t size = DVA_GET_ASIZE(dva);
3461 vdev_t *vd = vdev_lookup_top(spa, vdev);
3462
3463 ASSERT(DVA_IS_VALID(dva));
3464 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
3465
3466 if (DVA_GET_GANG(dva)) {
3467 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
3468 }
3469
3470 metaslab_free_impl(vd, offset, size, txg);
3471 }
3472
3473 /*
3474 * Reserve some allocation slots. The reservation system must be called
3475 * before we call into the allocator. If there aren't any available slots
3476 * then the I/O will be throttled until an I/O completes and its slots are
3477 * freed up. The function returns true if it was successful in placing
3478 * the reservation.
3479 */
3480 boolean_t
3481 metaslab_class_throttle_reserve(metaslab_class_t *mc, int slots, zio_t *zio,
3482 int flags)
3483 {
3484 uint64_t available_slots = 0;
3485 boolean_t slot_reserved = B_FALSE;
3486
3487 ASSERT(mc->mc_alloc_throttle_enabled);
3488 mutex_enter(&mc->mc_lock);
3489
3490 uint64_t reserved_slots = refcount_count(&mc->mc_alloc_slots);
3501 }
3502 zio->io_flags |= ZIO_FLAG_IO_ALLOCATING;
3503 slot_reserved = B_TRUE;
3504 }
3505
3506 mutex_exit(&mc->mc_lock);
3507 return (slot_reserved);
3508 }
3509
3510 void
3511 metaslab_class_throttle_unreserve(metaslab_class_t *mc, int slots, zio_t *zio)
3512 {
3513 ASSERT(mc->mc_alloc_throttle_enabled);
3514 mutex_enter(&mc->mc_lock);
3515 for (int d = 0; d < slots; d++) {
3516 (void) refcount_remove(&mc->mc_alloc_slots, zio);
3517 }
3518 mutex_exit(&mc->mc_lock);
3519 }
3520
3521 static int
3522 metaslab_claim_concrete(vdev_t *vd, uint64_t offset, uint64_t size,
3523 uint64_t txg)
3524 {
3525 metaslab_t *msp;
3526 spa_t *spa = vd->vdev_spa;
3527 int error = 0;
3528
3529 if (offset >> vd->vdev_ms_shift >= vd->vdev_ms_count)
3530 return (ENXIO);
3531
3532 ASSERT3P(vd->vdev_ms, !=, NULL);
3533 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3534
3535 mutex_enter(&msp->ms_lock);
3536
3537 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded)
3538 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
3539
3540 if (error == 0 && !range_tree_contains(msp->ms_tree, offset, size))
3541 error = SET_ERROR(ENOENT);
3542
3543 if (error || txg == 0) { /* txg == 0 indicates dry run */
3544 mutex_exit(&msp->ms_lock);
3545 return (error);
3546 }
3547
3548 VERIFY(!msp->ms_condensing);
3549 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
3550 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
3551 VERIFY3U(range_tree_space(msp->ms_tree) - size, <=, msp->ms_size);
3552 range_tree_remove(msp->ms_tree, offset, size);
3553
3554 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
3555 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
3556 vdev_dirty(vd, VDD_METASLAB, msp, txg);
3557 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, size);
3558 }
3559
3560 mutex_exit(&msp->ms_lock);
3561
3562 return (0);
3563 }
3564
3565 typedef struct metaslab_claim_cb_arg_t {
3566 uint64_t mcca_txg;
3567 int mcca_error;
3568 } metaslab_claim_cb_arg_t;
3569
3570 /* ARGSUSED */
3571 static void
3572 metaslab_claim_impl_cb(uint64_t inner_offset, vdev_t *vd, uint64_t offset,
3573 uint64_t size, void *arg)
3574 {
3575 metaslab_claim_cb_arg_t *mcca_arg = arg;
3576
3577 if (mcca_arg->mcca_error == 0) {
3578 mcca_arg->mcca_error = metaslab_claim_concrete(vd, offset,
3579 size, mcca_arg->mcca_txg);
3580 }
3581 }
3582
3583 int
3584 metaslab_claim_impl(vdev_t *vd, uint64_t offset, uint64_t size, uint64_t txg)
3585 {
3586 if (vd->vdev_ops->vdev_op_remap != NULL) {
3587 metaslab_claim_cb_arg_t arg;
3588
3589 /*
3590 * Only zdb(1M) can claim on indirect vdevs. This is used
3591 * to detect leaks of mapped space (that are not accounted
3592 * for in the obsolete counts, spacemap, or bpobj).
3593 */
3594 ASSERT(!spa_writeable(vd->vdev_spa));
3595 arg.mcca_error = 0;
3596 arg.mcca_txg = txg;
3597
3598 vd->vdev_ops->vdev_op_remap(vd, offset, size,
3599 metaslab_claim_impl_cb, &arg);
3600
3601 if (arg.mcca_error == 0) {
3602 arg.mcca_error = metaslab_claim_concrete(vd,
3603 offset, size, txg);
3604 }
3605 return (arg.mcca_error);
3606 } else {
3607 return (metaslab_claim_concrete(vd, offset, size, txg));
3608 }
3609 }
3610
3611 /*
3612 * Intent log support: upon opening the pool after a crash, notify the SPA
3613 * of blocks that the intent log has allocated for immediate write, but
3614 * which are still considered free by the SPA because the last transaction
3615 * group didn't commit yet.
3616 */
3617 static int
3618 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
3619 {
3620 uint64_t vdev = DVA_GET_VDEV(dva);
3621 uint64_t offset = DVA_GET_OFFSET(dva);
3622 uint64_t size = DVA_GET_ASIZE(dva);
3623 vdev_t *vd;
3624
3625 if ((vd = vdev_lookup_top(spa, vdev)) == NULL) {
3626 return (SET_ERROR(ENXIO));
3627 }
3628
3629 ASSERT(DVA_IS_VALID(dva));
3630
3631 if (DVA_GET_GANG(dva))
3632 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
3633
3634 return (metaslab_claim_impl(vd, offset, size, txg));
3635 }
3636
3637 int
3638 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
3639 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags,
3640 zio_alloc_list_t *zal, zio_t *zio)
3641 {
3642 dva_t *dva = bp->blk_dva;
3643 dva_t *hintdva = hintbp->blk_dva;
3644 int error = 0;
3645
3646 ASSERT(bp->blk_birth == 0);
3647 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
3648
3649 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
3650
3651 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
3652 spa_config_exit(spa, SCL_ALLOC, FTAG);
3653 return (SET_ERROR(ENOSPC));
3654 }
3655
3656 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
3657 ASSERT(BP_GET_NDVAS(bp) == 0);
3658 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
3659 ASSERT3P(zal, !=, NULL);
3660
3661 for (int d = 0; d < ndvas; d++) {
3662 error = metaslab_alloc_dva(spa, mc, psize, dva, d, hintdva,
3663 txg, flags, zal);
3664 if (error != 0) {
3665 for (d--; d >= 0; d--) {
3666 metaslab_unalloc_dva(spa, &dva[d], txg);
3667 metaslab_group_alloc_decrement(spa,
3668 DVA_GET_VDEV(&dva[d]), zio, flags);
3669 bzero(&dva[d], sizeof (dva_t));
3670 }
3671 spa_config_exit(spa, SCL_ALLOC, FTAG);
3672 return (error);
3673 } else {
3674 /*
3675 * Update the metaslab group's queue depth
3676 * based on the newly allocated dva.
3677 */
3678 metaslab_group_alloc_increment(spa,
3679 DVA_GET_VDEV(&dva[d]), zio, flags);
3680 }
3681
3682 }
3683 ASSERT(error == 0);
3684 ASSERT(BP_GET_NDVAS(bp) == ndvas);
3685
3686 spa_config_exit(spa, SCL_ALLOC, FTAG);
3687
3688 BP_SET_BIRTH(bp, txg, txg);
3689
3690 return (0);
3691 }
3692
3693 void
3694 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
3695 {
3696 const dva_t *dva = bp->blk_dva;
3697 int ndvas = BP_GET_NDVAS(bp);
3698
3699 ASSERT(!BP_IS_HOLE(bp));
3700 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
3701
3702 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
3703
3704 for (int d = 0; d < ndvas; d++) {
3705 if (now) {
3706 metaslab_unalloc_dva(spa, &dva[d], txg);
3707 } else {
3708 metaslab_free_dva(spa, &dva[d], txg);
3709 }
3710 }
3711
3712 spa_config_exit(spa, SCL_FREE, FTAG);
3713 }
3714
3715 int
3716 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
3717 {
3718 const dva_t *dva = bp->blk_dva;
3719 int ndvas = BP_GET_NDVAS(bp);
3720 int error = 0;
3721
3722 ASSERT(!BP_IS_HOLE(bp));
3723
3724 if (txg != 0) {
3725 /*
3726 * First do a dry run to make sure all DVAs are claimable,
3727 * so we don't have to unwind from partial failures below.
3728 */
3729 if ((error = metaslab_claim(spa, bp, 0)) != 0)
3730 return (error);
3731 }
3732
3733 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
3734
3735 for (int d = 0; d < ndvas; d++)
3736 if ((error = metaslab_claim_dva(spa, &dva[d], txg)) != 0)
3737 break;
3738
3739 spa_config_exit(spa, SCL_ALLOC, FTAG);
3740
3741 ASSERT(error == 0 || txg == 0);
3742
3743 return (error);
3744 }
3745
3746 /* ARGSUSED */
3747 static void
3748 metaslab_check_free_impl_cb(uint64_t inner, vdev_t *vd, uint64_t offset,
3749 uint64_t size, void *arg)
3750 {
3751 if (vd->vdev_ops == &vdev_indirect_ops)
3752 return;
3753
3754 metaslab_check_free_impl(vd, offset, size);
3755 }
3756
3757 static void
3758 metaslab_check_free_impl(vdev_t *vd, uint64_t offset, uint64_t size)
3759 {
3760 metaslab_t *msp;
3761 spa_t *spa = vd->vdev_spa;
3762
3763 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
3764 return;
3765
3766 if (vd->vdev_ops->vdev_op_remap != NULL) {
3767 vd->vdev_ops->vdev_op_remap(vd, offset, size,
3768 metaslab_check_free_impl_cb, NULL);
3769 return;
3770 }
3771
3772 ASSERT(vdev_is_concrete(vd));
3773 ASSERT3U(offset >> vd->vdev_ms_shift, <, vd->vdev_ms_count);
3774 ASSERT3U(spa_config_held(spa, SCL_ALL, RW_READER), !=, 0);
3775
3776 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3777
3778 mutex_enter(&msp->ms_lock);
3779 if (msp->ms_loaded)
3780 range_tree_verify(msp->ms_tree, offset, size);
3781
3782 range_tree_verify(msp->ms_freeingtree, offset, size);
3783 range_tree_verify(msp->ms_freedtree, offset, size);
3784 for (int j = 0; j < TXG_DEFER_SIZE; j++)
3785 range_tree_verify(msp->ms_defertree[j], offset, size);
3786 mutex_exit(&msp->ms_lock);
3787 }
3788
3789 void
3790 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
3791 {
3792 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
3793 return;
3794
3795 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
3796 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
3797 uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
3798 vdev_t *vd = vdev_lookup_top(spa, vdev);
3799 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
3800 uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
3801
3802 if (DVA_GET_GANG(&bp->blk_dva[i]))
3803 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
3804
3805 ASSERT3P(vd, !=, NULL);
3806
3807 metaslab_check_free_impl(vd, offset, size);
3808 }
3809 spa_config_exit(spa, SCL_VDEV, FTAG);
3810 }
|
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 (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2015 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
26 * Copyright 2017 Nexenta Systems, Inc. All rights reserved.
27 */
28
29 #include <sys/zfs_context.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 #include <sys/spa_impl.h>
37 #include <sys/zfeature.h>
38 #include <sys/wbc.h>
39
40 #define GANG_ALLOCATION(flags) \
41 ((flags) & (METASLAB_GANG_CHILD | METASLAB_GANG_HEADER))
42
43 uint64_t metaslab_aliquot = 512ULL << 10;
44 uint64_t metaslab_gang_bang = SPA_MAXBLOCKSIZE + 1; /* force gang blocks */
45
46 /*
47 * The in-core space map representation is more compact than its on-disk form.
48 * The zfs_condense_pct determines how much more compact the in-core
49 * space map representation must be before we compact it on-disk.
50 * Values should be greater than or equal to 100.
51 */
52 int zfs_condense_pct = 200;
53
54 /*
55 * Condensing a metaslab is not guaranteed to actually reduce the amount of
56 * space used on disk. In particular, a space map uses data in increments of
57 * MAX(1 << ashift, space_map_blksize), so a metaslab might use the
58 * same number of blocks after condensing. Since the goal of condensing is to
151 * Enable/disable preloading of metaslab.
152 */
153 boolean_t metaslab_preload_enabled = B_TRUE;
154
155 /*
156 * Enable/disable fragmentation weighting on metaslabs.
157 */
158 boolean_t metaslab_fragmentation_factor_enabled = B_TRUE;
159
160 /*
161 * Enable/disable lba weighting (i.e. outer tracks are given preference).
162 */
163 boolean_t metaslab_lba_weighting_enabled = B_TRUE;
164
165 /*
166 * Enable/disable metaslab group biasing.
167 */
168 boolean_t metaslab_bias_enabled = B_TRUE;
169
170 /*
171 * Enable/disable segment-based metaslab selection.
172 */
173 boolean_t zfs_metaslab_segment_weight_enabled = B_TRUE;
174
175 /*
176 * When using segment-based metaslab selection, we will continue
177 * allocating from the active metaslab until we have exhausted
178 * zfs_metaslab_switch_threshold of its buckets.
179 */
180 int zfs_metaslab_switch_threshold = 2;
181
182 /*
183 * Internal switch to enable/disable the metaslab allocation tracing
184 * facility.
185 */
186 boolean_t metaslab_trace_enabled = B_TRUE;
187
188 /*
189 * Maximum entries that the metaslab allocation tracing facility will keep
190 * in a given list when running in non-debug mode. We limit the number
191 * of entries in non-debug mode to prevent us from using up too much memory.
192 * The limit should be sufficiently large that we don't expect any allocation
193 * to every exceed this value. In debug mode, the system will panic if this
194 * limit is ever reached allowing for further investigation.
195 */
196 uint64_t metaslab_trace_max_entries = 5000;
197
198 static uint64_t metaslab_weight(metaslab_t *);
199 static void metaslab_set_fragmentation(metaslab_t *);
200
201 kmem_cache_t *metaslab_alloc_trace_cache;
202
203 /*
204 * Toggle between space-based DVA allocator 0, latency-based 1 or hybrid 2.
205 * A value other than 0, 1 or 2 will be considered 0 (default).
206 */
207 int metaslab_alloc_dva_algorithm = 0;
208
209 /*
210 * How many TXG's worth of updates should be aggregated per TRIM/UNMAP
211 * issued to the underlying vdev. We keep two range trees of extents
212 * (called "trim sets") to be trimmed per metaslab, the `current' and
213 * the `previous' TS. New free's are added to the current TS. Then,
214 * once `zfs_txgs_per_trim' transactions have elapsed, the `current'
215 * TS becomes the `previous' TS and a new, blank TS is created to be
216 * the new `current', which will then start accumulating any new frees.
217 * Once another zfs_txgs_per_trim TXGs have passed, the previous TS's
218 * extents are trimmed, the TS is destroyed and the current TS again
219 * becomes the previous TS.
220 * This serves to fulfill two functions: aggregate many small frees
221 * into fewer larger trim operations (which should help with devices
222 * which do not take so kindly to them) and to allow for disaster
223 * recovery (extents won't get trimmed immediately, but instead only
224 * after passing this rather long timeout, thus not preserving
225 * 'zfs import -F' functionality).
226 */
227 unsigned int zfs_txgs_per_trim = 32;
228
229 static void metaslab_trim_remove(void *arg, uint64_t offset, uint64_t size);
230 static void metaslab_trim_add(void *arg, uint64_t offset, uint64_t size);
231
232 static zio_t *metaslab_exec_trim(metaslab_t *msp);
233
234 static metaslab_trimset_t *metaslab_new_trimset(uint64_t txg, kmutex_t *lock);
235 static void metaslab_free_trimset(metaslab_trimset_t *ts);
236 static boolean_t metaslab_check_trim_conflict(metaslab_t *msp,
237 uint64_t *offset, uint64_t size, uint64_t align, uint64_t limit);
238
239 /*
240 * ==========================================================================
241 * Metaslab classes
242 * ==========================================================================
243 */
244 metaslab_class_t *
245 metaslab_class_create(spa_t *spa, metaslab_ops_t *ops)
246 {
247 metaslab_class_t *mc;
248
249 mc = kmem_zalloc(sizeof (metaslab_class_t), KM_SLEEP);
250
251 mutex_init(&mc->mc_alloc_lock, NULL, MUTEX_DEFAULT, NULL);
252 avl_create(&mc->mc_alloc_tree, zio_bookmark_compare,
253 sizeof (zio_t), offsetof(zio_t, io_alloc_node));
254
255 mc->mc_spa = spa;
256 mc->mc_rotor = NULL;
257 mc->mc_ops = ops;
258 mutex_init(&mc->mc_lock, NULL, MUTEX_DEFAULT, NULL);
259 refcount_create_tracked(&mc->mc_alloc_slots);
260
261 return (mc);
262 }
263
264 void
265 metaslab_class_destroy(metaslab_class_t *mc)
266 {
267 ASSERT(mc->mc_rotor == NULL);
268 ASSERT(mc->mc_alloc == 0);
269 ASSERT(mc->mc_deferred == 0);
270 ASSERT(mc->mc_space == 0);
271 ASSERT(mc->mc_dspace == 0);
272
273 avl_destroy(&mc->mc_alloc_tree);
274 mutex_destroy(&mc->mc_alloc_lock);
275
276 refcount_destroy(&mc->mc_alloc_slots);
277 mutex_destroy(&mc->mc_lock);
278 kmem_free(mc, sizeof (metaslab_class_t));
279 }
280
281 int
282 metaslab_class_validate(metaslab_class_t *mc)
283 {
284 metaslab_group_t *mg;
285 vdev_t *vd;
286
287 /*
288 * Must hold one of the spa_config locks.
289 */
290 ASSERT(spa_config_held(mc->mc_spa, SCL_ALL, RW_READER) ||
291 spa_config_held(mc->mc_spa, SCL_ALL, RW_WRITER));
292
293 if ((mg = mc->mc_rotor) == NULL)
294 return (0);
295
342 metaslab_class_histogram_verify(metaslab_class_t *mc)
343 {
344 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
345 uint64_t *mc_hist;
346 int i;
347
348 if ((zfs_flags & ZFS_DEBUG_HISTOGRAM_VERIFY) == 0)
349 return;
350
351 mc_hist = kmem_zalloc(sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE,
352 KM_SLEEP);
353
354 for (int c = 0; c < rvd->vdev_children; c++) {
355 vdev_t *tvd = rvd->vdev_child[c];
356 metaslab_group_t *mg = tvd->vdev_mg;
357
358 /*
359 * Skip any holes, uninitialized top-levels, or
360 * vdevs that are not in this metalab class.
361 */
362 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
363 mg->mg_class != mc) {
364 continue;
365 }
366
367 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
368 mc_hist[i] += mg->mg_histogram[i];
369 }
370
371 for (i = 0; i < RANGE_TREE_HISTOGRAM_SIZE; i++)
372 VERIFY3U(mc_hist[i], ==, mc->mc_histogram[i]);
373
374 kmem_free(mc_hist, sizeof (uint64_t) * RANGE_TREE_HISTOGRAM_SIZE);
375 }
376
377 /*
378 * Calculate the metaslab class's fragmentation metric. The metric
379 * is weighted based on the space contribution of each metaslab group.
380 * The return value will be a number between 0 and 100 (inclusive), or
381 * ZFS_FRAG_INVALID if the metric has not been set. See comment above the
382 * zfs_frag_table for more information about the metric.
383 */
384 uint64_t
385 metaslab_class_fragmentation(metaslab_class_t *mc)
386 {
387 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
388 uint64_t fragmentation = 0;
389
390 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
391
392 for (int c = 0; c < rvd->vdev_children; c++) {
393 vdev_t *tvd = rvd->vdev_child[c];
394 metaslab_group_t *mg = tvd->vdev_mg;
395
396 /*
397 * Skip any holes, uninitialized top-levels, or
398 * vdevs that are not in this metalab class.
399 */
400 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
401 mg->mg_class != mc) {
402 continue;
403 }
404
405 /*
406 * If a metaslab group does not contain a fragmentation
407 * metric then just bail out.
408 */
409 if (mg->mg_fragmentation == ZFS_FRAG_INVALID) {
410 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
411 return (ZFS_FRAG_INVALID);
412 }
413
414 /*
415 * Determine how much this metaslab_group is contributing
416 * to the overall pool fragmentation metric.
417 */
418 fragmentation += mg->mg_fragmentation *
419 metaslab_group_get_space(mg);
420 }
426 }
427
428 /*
429 * Calculate the amount of expandable space that is available in
430 * this metaslab class. If a device is expanded then its expandable
431 * space will be the amount of allocatable space that is currently not
432 * part of this metaslab class.
433 */
434 uint64_t
435 metaslab_class_expandable_space(metaslab_class_t *mc)
436 {
437 vdev_t *rvd = mc->mc_spa->spa_root_vdev;
438 uint64_t space = 0;
439
440 spa_config_enter(mc->mc_spa, SCL_VDEV, FTAG, RW_READER);
441 for (int c = 0; c < rvd->vdev_children; c++) {
442 uint64_t tspace;
443 vdev_t *tvd = rvd->vdev_child[c];
444 metaslab_group_t *mg = tvd->vdev_mg;
445
446 if (tvd->vdev_ishole || tvd->vdev_ms_shift == 0 ||
447 mg->mg_class != mc) {
448 continue;
449 }
450
451 /*
452 * Calculate if we have enough space to add additional
453 * metaslabs. We report the expandable space in terms
454 * of the metaslab size since that's the unit of expansion.
455 * Adjust by efi system partition size.
456 */
457 tspace = tvd->vdev_max_asize - tvd->vdev_asize;
458 if (tspace > mc->mc_spa->spa_bootsize) {
459 tspace -= mc->mc_spa->spa_bootsize;
460 }
461 space += P2ALIGN(tspace, 1ULL << tvd->vdev_ms_shift);
462 }
463 spa_config_exit(mc->mc_spa, SCL_VDEV, FTAG);
464 return (space);
465 }
466
538 * ==========================================================================
539 */
540 /*
541 * Update the allocatable flag and the metaslab group's capacity.
542 * The allocatable flag is set to true if the capacity is below
543 * the zfs_mg_noalloc_threshold or has a fragmentation value that is
544 * greater than zfs_mg_fragmentation_threshold. If a metaslab group
545 * transitions from allocatable to non-allocatable or vice versa then the
546 * metaslab group's class is updated to reflect the transition.
547 */
548 static void
549 metaslab_group_alloc_update(metaslab_group_t *mg)
550 {
551 vdev_t *vd = mg->mg_vd;
552 metaslab_class_t *mc = mg->mg_class;
553 vdev_stat_t *vs = &vd->vdev_stat;
554 boolean_t was_allocatable;
555 boolean_t was_initialized;
556
557 ASSERT(vd == vd->vdev_top);
558
559 mutex_enter(&mg->mg_lock);
560 was_allocatable = mg->mg_allocatable;
561 was_initialized = mg->mg_initialized;
562
563 mg->mg_free_capacity = ((vs->vs_space - vs->vs_alloc) * 100) /
564 (vs->vs_space + 1);
565
566 mutex_enter(&mc->mc_lock);
567
568 /*
569 * If the metaslab group was just added then it won't
570 * have any space until we finish syncing out this txg.
571 * At that point we will consider it initialized and available
572 * for allocations. We also don't consider non-activated
573 * metaslab groups (e.g. vdevs that are in the middle of being removed)
574 * to be initialized, because they can't be used for allocation.
575 */
576 mg->mg_initialized = metaslab_group_initialized(mg);
577 if (!was_initialized && mg->mg_initialized) {
635 refcount_create_tracked(&mg->mg_alloc_queue_depth);
636
637 mg->mg_taskq = taskq_create("metaslab_group_taskq", metaslab_load_pct,
638 minclsyspri, 10, INT_MAX, TASKQ_THREADS_CPU_PCT);
639
640 return (mg);
641 }
642
643 void
644 metaslab_group_destroy(metaslab_group_t *mg)
645 {
646 ASSERT(mg->mg_prev == NULL);
647 ASSERT(mg->mg_next == NULL);
648 /*
649 * We may have gone below zero with the activation count
650 * either because we never activated in the first place or
651 * because we're done, and possibly removing the vdev.
652 */
653 ASSERT(mg->mg_activation_count <= 0);
654
655 if (mg->mg_taskq)
656 taskq_destroy(mg->mg_taskq);
657 avl_destroy(&mg->mg_metaslab_tree);
658 mutex_destroy(&mg->mg_lock);
659 refcount_destroy(&mg->mg_alloc_queue_depth);
660 kmem_free(mg, sizeof (metaslab_group_t));
661 }
662
663 void
664 metaslab_group_activate(metaslab_group_t *mg)
665 {
666 metaslab_class_t *mc = mg->mg_class;
667 metaslab_group_t *mgprev, *mgnext;
668
669 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
670
671 ASSERT(mc->mc_rotor != mg);
672 ASSERT(mg->mg_prev == NULL);
673 ASSERT(mg->mg_next == NULL);
674 ASSERT(mg->mg_activation_count <= 0);
675
676 if (++mg->mg_activation_count <= 0)
677 return;
678
679 mg->mg_aliquot = metaslab_aliquot * MAX(1, mg->mg_vd->vdev_children);
680 metaslab_group_alloc_update(mg);
681
682 if ((mgprev = mc->mc_rotor) == NULL) {
683 mg->mg_prev = mg;
684 mg->mg_next = mg;
685 } else {
686 mgnext = mgprev->mg_next;
687 mg->mg_prev = mgprev;
688 mg->mg_next = mgnext;
689 mgprev->mg_next = mg;
690 mgnext->mg_prev = mg;
691 }
692 mc->mc_rotor = mg;
693 }
694
695 void
696 metaslab_group_passivate(metaslab_group_t *mg)
697 {
698 metaslab_class_t *mc = mg->mg_class;
699 metaslab_group_t *mgprev, *mgnext;
700
701 ASSERT(spa_config_held(mc->mc_spa, SCL_ALLOC, RW_WRITER));
702
703 if (--mg->mg_activation_count != 0) {
704 ASSERT(mc->mc_rotor != mg);
705 ASSERT(mg->mg_prev == NULL);
706 ASSERT(mg->mg_next == NULL);
707 ASSERT(mg->mg_activation_count < 0);
708 return;
709 }
710
711 taskq_wait(mg->mg_taskq);
712 metaslab_group_alloc_update(mg);
713
714 mgprev = mg->mg_prev;
715 mgnext = mg->mg_next;
716
717 if (mg == mgnext) {
718 mc->mc_rotor = NULL;
719 } else {
720 mc->mc_rotor = mgnext;
721 mgprev->mg_next = mgnext;
722 mgnext->mg_prev = mgprev;
723 }
724
725 mg->mg_prev = NULL;
726 mg->mg_next = NULL;
727 }
728
729 boolean_t
730 metaslab_group_initialized(metaslab_group_t *mg)
731 {
1135 range_seg_t *rs, rsearch;
1136 avl_index_t where;
1137
1138 rsearch.rs_start = start;
1139 rsearch.rs_end = start + size;
1140
1141 rs = avl_find(t, &rsearch, &where);
1142 if (rs == NULL) {
1143 rs = avl_nearest(t, where, AVL_AFTER);
1144 }
1145
1146 return (rs);
1147 }
1148
1149 /*
1150 * This is a helper function that can be used by the allocator to find
1151 * a suitable block to allocate. This will search the specified AVL
1152 * tree looking for a block that matches the specified criteria.
1153 */
1154 static uint64_t
1155 metaslab_block_picker(metaslab_t *msp, avl_tree_t *t, uint64_t *cursor,
1156 uint64_t size, uint64_t align)
1157 {
1158 range_seg_t *rs = metaslab_block_find(t, *cursor, size);
1159
1160 for (; rs != NULL; rs = AVL_NEXT(t, rs)) {
1161 uint64_t offset = P2ROUNDUP(rs->rs_start, align);
1162
1163 if (offset + size <= rs->rs_end &&
1164 !metaslab_check_trim_conflict(msp, &offset, size, align,
1165 rs->rs_end)) {
1166 *cursor = offset + size;
1167 return (offset);
1168 }
1169 }
1170
1171 /*
1172 * If we know we've searched the whole map (*cursor == 0), give up.
1173 * Otherwise, reset the cursor to the beginning and try again.
1174 */
1175 if (*cursor == 0)
1176 return (-1ULL);
1177
1178 *cursor = 0;
1179 return (metaslab_block_picker(msp, t, cursor, size, align));
1180 }
1181
1182 /*
1183 * ==========================================================================
1184 * The first-fit block allocator
1185 * ==========================================================================
1186 */
1187 static uint64_t
1188 metaslab_ff_alloc(metaslab_t *msp, uint64_t size)
1189 {
1190 /*
1191 * Find the largest power of 2 block size that evenly divides the
1192 * requested size. This is used to try to allocate blocks with similar
1193 * alignment from the same area of the metaslab (i.e. same cursor
1194 * bucket) but it does not guarantee that other allocations sizes
1195 * may exist in the same region.
1196 */
1197 uint64_t align = size & -size;
1198 uint64_t *cursor = &msp->ms_lbas[highbit64(align) - 1];
1199 avl_tree_t *t = &msp->ms_tree->rt_root;
1200
1201 return (metaslab_block_picker(msp, t, cursor, size, align));
1202 }
1203
1204 static metaslab_ops_t metaslab_ff_ops = {
1205 metaslab_ff_alloc
1206 };
1207
1208 /*
1209 * ==========================================================================
1210 * Dynamic block allocator -
1211 * Uses the first fit allocation scheme until space get low and then
1212 * adjusts to a best fit allocation method. Uses metaslab_df_alloc_threshold
1213 * and metaslab_df_free_pct to determine when to switch the allocation scheme.
1214 * ==========================================================================
1215 */
1216 static uint64_t
1217 metaslab_df_alloc(metaslab_t *msp, uint64_t size)
1218 {
1219 /*
1220 * Find the largest power of 2 block size that evenly divides the
1221 * requested size. This is used to try to allocate blocks with similar
1229 avl_tree_t *t = &rt->rt_root;
1230 uint64_t max_size = metaslab_block_maxsize(msp);
1231 int free_pct = range_tree_space(rt) * 100 / msp->ms_size;
1232
1233 ASSERT(MUTEX_HELD(&msp->ms_lock));
1234 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1235
1236 if (max_size < size)
1237 return (-1ULL);
1238
1239 /*
1240 * If we're running low on space switch to using the size
1241 * sorted AVL tree (best-fit).
1242 */
1243 if (max_size < metaslab_df_alloc_threshold ||
1244 free_pct < metaslab_df_free_pct) {
1245 t = &msp->ms_size_tree;
1246 *cursor = 0;
1247 }
1248
1249 return (metaslab_block_picker(msp, t, cursor, size, 1ULL));
1250 }
1251
1252 static metaslab_ops_t metaslab_df_ops = {
1253 metaslab_df_alloc
1254 };
1255
1256 /*
1257 * ==========================================================================
1258 * Cursor fit block allocator -
1259 * Select the largest region in the metaslab, set the cursor to the beginning
1260 * of the range and the cursor_end to the end of the range. As allocations
1261 * are made advance the cursor. Continue allocating from the cursor until
1262 * the range is exhausted and then find a new range.
1263 * ==========================================================================
1264 */
1265 static uint64_t
1266 metaslab_cf_alloc(metaslab_t *msp, uint64_t size)
1267 {
1268 range_tree_t *rt = msp->ms_tree;
1269 avl_tree_t *t = &msp->ms_size_tree;
1270 uint64_t *cursor = &msp->ms_lbas[0];
1271 uint64_t *cursor_end = &msp->ms_lbas[1];
1272 uint64_t offset = 0;
1273
1274 ASSERT(MUTEX_HELD(&msp->ms_lock));
1275 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&rt->rt_root));
1276
1277 ASSERT3U(*cursor_end, >=, *cursor);
1278
1279 if ((*cursor + size) > *cursor_end) {
1280 range_seg_t *rs;
1281 for (rs = avl_last(&msp->ms_size_tree);
1282 rs != NULL && rs->rs_end - rs->rs_start >= size;
1283 rs = AVL_PREV(&msp->ms_size_tree, rs)) {
1284 *cursor = rs->rs_start;
1285 *cursor_end = rs->rs_end;
1286 if (!metaslab_check_trim_conflict(msp, cursor, size,
1287 1, *cursor_end)) {
1288 /* segment appears to be acceptable */
1289 break;
1290 }
1291 }
1292 if (rs == NULL || rs->rs_end - rs->rs_start < size)
1293 return (-1ULL);
1294 }
1295
1296 offset = *cursor;
1297 *cursor += size;
1298
1299 return (offset);
1300 }
1301
1302 static metaslab_ops_t metaslab_cf_ops = {
1303 metaslab_cf_alloc
1304 };
1305
1306 /*
1307 * ==========================================================================
1308 * New dynamic fit allocator -
1309 * Select a region that is large enough to allocate 2^metaslab_ndf_clump_shift
1310 * contiguous blocks. If no region is found then just use the largest segment
1311 * that remains.
1312 * ==========================================================================
1313 */
1314
1315 /*
1316 * Determines desired number of contiguous blocks (2^metaslab_ndf_clump_shift)
1317 * to request from the allocator.
1318 */
1319 uint64_t metaslab_ndf_clump_shift = 4;
1320
1321 static uint64_t
1322 metaslab_ndf_alloc(metaslab_t *msp, uint64_t size)
1323 {
1324 avl_tree_t *t = &msp->ms_tree->rt_root;
1325 avl_index_t where;
1326 range_seg_t *rs, rsearch;
1327 uint64_t hbit = highbit64(size);
1328 uint64_t *cursor = &msp->ms_lbas[hbit - 1];
1329 uint64_t max_size = metaslab_block_maxsize(msp);
1330 /* mutable copy for adjustment by metaslab_check_trim_conflict */
1331 uint64_t adjustable_start;
1332
1333 ASSERT(MUTEX_HELD(&msp->ms_lock));
1334 ASSERT3U(avl_numnodes(t), ==, avl_numnodes(&msp->ms_size_tree));
1335
1336 if (max_size < size)
1337 return (-1ULL);
1338
1339 rsearch.rs_start = *cursor;
1340 rsearch.rs_end = *cursor + size;
1341
1342 rs = avl_find(t, &rsearch, &where);
1343 if (rs != NULL)
1344 adjustable_start = rs->rs_start;
1345 if (rs == NULL || rs->rs_end - adjustable_start < size ||
1346 metaslab_check_trim_conflict(msp, &adjustable_start, size, 1,
1347 rs->rs_end)) {
1348 /* segment not usable, try the largest remaining one */
1349 t = &msp->ms_size_tree;
1350
1351 rsearch.rs_start = 0;
1352 rsearch.rs_end = MIN(max_size,
1353 1ULL << (hbit + metaslab_ndf_clump_shift));
1354 rs = avl_find(t, &rsearch, &where);
1355 if (rs == NULL)
1356 rs = avl_nearest(t, where, AVL_AFTER);
1357 ASSERT(rs != NULL);
1358 adjustable_start = rs->rs_start;
1359 if (rs->rs_end - adjustable_start < size ||
1360 metaslab_check_trim_conflict(msp, &adjustable_start,
1361 size, 1, rs->rs_end)) {
1362 /* even largest remaining segment not usable */
1363 return (-1ULL);
1364 }
1365 }
1366
1367 *cursor = adjustable_start + size;
1368 return (*cursor);
1369 }
1370
1371 static metaslab_ops_t metaslab_ndf_ops = {
1372 metaslab_ndf_alloc
1373 };
1374
1375 metaslab_ops_t *zfs_metaslab_ops = &metaslab_df_ops;
1376
1377 /*
1378 * ==========================================================================
1379 * Metaslabs
1380 * ==========================================================================
1381 */
1382
1383 /*
1384 * Wait for any in-progress metaslab loads to complete.
1385 */
1386 void
1387 metaslab_load_wait(metaslab_t *msp)
1388 {
1389 ASSERT(MUTEX_HELD(&msp->ms_lock));
1390
1391 while (msp->ms_loading) {
1392 ASSERT(!msp->ms_loaded);
1393 cv_wait(&msp->ms_load_cv, &msp->ms_lock);
1394 }
1395 }
1396
1397 int
1398 metaslab_load(metaslab_t *msp)
1399 {
1400 int error = 0;
1401 boolean_t success = B_FALSE;
1402
1403 ASSERT(MUTEX_HELD(&msp->ms_lock));
1404 ASSERT(!msp->ms_loaded);
1405 ASSERT(!msp->ms_loading);
1406
1407 msp->ms_loading = B_TRUE;
1408
1409 /*
1410 * If the space map has not been allocated yet, then treat
1411 * all the space in the metaslab as free and add it to the
1412 * ms_tree.
1413 */
1414 if (msp->ms_sm != NULL)
1415 error = space_map_load(msp->ms_sm, msp->ms_tree, SM_FREE);
1416 else
1417 range_tree_add(msp->ms_tree, msp->ms_start, msp->ms_size);
1418
1419 success = (error == 0);
1420 msp->ms_loading = B_FALSE;
1421
1422 if (success) {
1423 ASSERT3P(msp->ms_group, !=, NULL);
1424 msp->ms_loaded = B_TRUE;
1425
1426 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1427 range_tree_walk(msp->ms_defertree[t],
1428 range_tree_remove, msp->ms_tree);
1429 range_tree_walk(msp->ms_defertree[t],
1430 metaslab_trim_remove, msp);
1431 }
1432 msp->ms_max_size = metaslab_block_maxsize(msp);
1433 }
1434 cv_broadcast(&msp->ms_load_cv);
1435 return (error);
1436 }
1437
1438 void
1439 metaslab_unload(metaslab_t *msp)
1440 {
1441 ASSERT(MUTEX_HELD(&msp->ms_lock));
1442 range_tree_vacate(msp->ms_tree, NULL, NULL);
1443 msp->ms_loaded = B_FALSE;
1444 msp->ms_weight &= ~METASLAB_ACTIVE_MASK;
1445 msp->ms_max_size = 0;
1446 }
1447
1448 int
1449 metaslab_init(metaslab_group_t *mg, uint64_t id, uint64_t object, uint64_t txg,
1450 metaslab_t **msp)
1451 {
1452 vdev_t *vd = mg->mg_vd;
1453 objset_t *mos = vd->vdev_spa->spa_meta_objset;
1454 metaslab_t *ms;
1455 int error;
1456
1457 ms = kmem_zalloc(sizeof (metaslab_t), KM_SLEEP);
1458 mutex_init(&ms->ms_lock, NULL, MUTEX_DEFAULT, NULL);
1459 cv_init(&ms->ms_load_cv, NULL, CV_DEFAULT, NULL);
1460 cv_init(&ms->ms_trim_cv, NULL, CV_DEFAULT, NULL);
1461 ms->ms_id = id;
1462 ms->ms_start = id << vd->vdev_ms_shift;
1463 ms->ms_size = 1ULL << vd->vdev_ms_shift;
1464
1465 /*
1466 * We only open space map objects that already exist. All others
1467 * will be opened when we finally allocate an object for it.
1468 */
1469 if (object != 0) {
1470 error = space_map_open(&ms->ms_sm, mos, object, ms->ms_start,
1471 ms->ms_size, vd->vdev_ashift, &ms->ms_lock);
1472
1473 if (error != 0) {
1474 kmem_free(ms, sizeof (metaslab_t));
1475 return (error);
1476 }
1477
1478 ASSERT(ms->ms_sm != NULL);
1479 }
1480
1481 ms->ms_cur_ts = metaslab_new_trimset(0, &ms->ms_lock);
1482
1483 /*
1484 * We create the main range tree here, but we don't create the
1485 * other range trees until metaslab_sync_done(). This serves
1486 * two purposes: it allows metaslab_sync_done() to detect the
1487 * addition of new space; and for debugging, it ensures that we'd
1488 * data fault on any attempt to use this metaslab before it's ready.
1489 */
1490 ms->ms_tree = range_tree_create(&metaslab_rt_ops, ms, &ms->ms_lock);
1491 metaslab_group_add(mg, ms);
1492
1493 metaslab_set_fragmentation(ms);
1494
1495 /*
1496 * If we're opening an existing pool (txg == 0) or creating
1497 * a new one (txg == TXG_INITIAL), all space is available now.
1498 * If we're adding space to an existing pool, the new space
1499 * does not become available until after this txg has synced.
1500 * The metaslab's weight will also be initialized when we sync
1501 * out this txg. This ensures that we don't attempt to allocate
1502 * from it before we have initialized it completely.
1503 */
1504 if (txg <= TXG_INITIAL)
1505 metaslab_sync_done(ms, 0);
1506
1507 /*
1508 * If metaslab_debug_load is set and we're initializing a metaslab
1509 * that has an allocated space map object then load the its space
1510 * map so that can verify frees.
1534
1535 mutex_enter(&msp->ms_lock);
1536 VERIFY(msp->ms_group == NULL);
1537 vdev_space_update(mg->mg_vd, -space_map_allocated(msp->ms_sm),
1538 0, -msp->ms_size);
1539 space_map_close(msp->ms_sm);
1540
1541 metaslab_unload(msp);
1542 range_tree_destroy(msp->ms_tree);
1543 range_tree_destroy(msp->ms_freeingtree);
1544 range_tree_destroy(msp->ms_freedtree);
1545
1546 for (int t = 0; t < TXG_SIZE; t++) {
1547 range_tree_destroy(msp->ms_alloctree[t]);
1548 }
1549
1550 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
1551 range_tree_destroy(msp->ms_defertree[t]);
1552 }
1553
1554 metaslab_free_trimset(msp->ms_cur_ts);
1555 if (msp->ms_prev_ts)
1556 metaslab_free_trimset(msp->ms_prev_ts);
1557 ASSERT3P(msp->ms_trimming_ts, ==, NULL);
1558
1559 ASSERT0(msp->ms_deferspace);
1560
1561 mutex_exit(&msp->ms_lock);
1562 cv_destroy(&msp->ms_load_cv);
1563 cv_destroy(&msp->ms_trim_cv);
1564 mutex_destroy(&msp->ms_lock);
1565
1566 kmem_free(msp, sizeof (metaslab_t));
1567 }
1568
1569 #define FRAGMENTATION_TABLE_SIZE 17
1570
1571 /*
1572 * This table defines a segment size based fragmentation metric that will
1573 * allow each metaslab to derive its own fragmentation value. This is done
1574 * by calculating the space in each bucket of the spacemap histogram and
1575 * multiplying that by the fragmetation metric in this table. Doing
1576 * this for all buckets and dividing it by the total amount of free
1577 * space in this metaslab (i.e. the total free space in all buckets) gives
1578 * us the fragmentation metric. This means that a high fragmentation metric
1579 * equates to most of the free space being comprised of small segments.
1580 * Conversely, if the metric is low, then most of the free space is in
1581 * large segments. A 10% change in fragmentation equates to approximately
1582 * double the number of segments.
1583 *
1584 * This table defines 0% fragmented space using 16MB segments. Testing has
1910 */
1911 should_allocate = (asize <
1912 1ULL << (WEIGHT_GET_INDEX(msp->ms_weight) + 1));
1913 } else {
1914 should_allocate = (asize <=
1915 (msp->ms_weight & ~METASLAB_WEIGHT_TYPE));
1916 }
1917 return (should_allocate);
1918 }
1919
1920 static uint64_t
1921 metaslab_weight(metaslab_t *msp)
1922 {
1923 vdev_t *vd = msp->ms_group->mg_vd;
1924 spa_t *spa = vd->vdev_spa;
1925 uint64_t weight;
1926
1927 ASSERT(MUTEX_HELD(&msp->ms_lock));
1928
1929 /*
1930 * This vdev is in the process of being removed so there is nothing
1931 * for us to do here.
1932 */
1933 if (vd->vdev_removing) {
1934 ASSERT0(space_map_allocated(msp->ms_sm));
1935 ASSERT0(vd->vdev_ms_shift);
1936 return (0);
1937 }
1938
1939 metaslab_set_fragmentation(msp);
1940
1941 /*
1942 * Update the maximum size if the metaslab is loaded. This will
1943 * ensure that we get an accurate maximum size if newly freed space
1944 * has been added back into the free tree.
1945 */
1946 if (msp->ms_loaded)
1947 msp->ms_max_size = metaslab_block_maxsize(msp);
1948
1949 /*
1950 * Segment-based weighting requires space map histogram support.
1951 */
1952 if (zfs_metaslab_segment_weight_enabled &&
1953 spa_feature_is_enabled(spa, SPA_FEATURE_SPACEMAP_HISTOGRAM) &&
1954 (msp->ms_sm == NULL || msp->ms_sm->sm_dbuf->db_size ==
1955 sizeof (space_map_phys_t))) {
1956 weight = metaslab_segment_weight(msp);
1957 } else {
2049 if (!msp->ms_loaded)
2050 (void) metaslab_load(msp);
2051 msp->ms_selected_txg = spa_syncing_txg(spa);
2052 mutex_exit(&msp->ms_lock);
2053 }
2054
2055 static void
2056 metaslab_group_preload(metaslab_group_t *mg)
2057 {
2058 spa_t *spa = mg->mg_vd->vdev_spa;
2059 metaslab_t *msp;
2060 avl_tree_t *t = &mg->mg_metaslab_tree;
2061 int m = 0;
2062
2063 if (spa_shutting_down(spa) || !metaslab_preload_enabled) {
2064 taskq_wait(mg->mg_taskq);
2065 return;
2066 }
2067
2068 mutex_enter(&mg->mg_lock);
2069 /*
2070 * Load the next potential metaslabs
2071 */
2072 for (msp = avl_first(t); msp != NULL; msp = AVL_NEXT(t, msp)) {
2073 /*
2074 * We preload only the maximum number of metaslabs specified
2075 * by metaslab_preload_limit. If a metaslab is being forced
2076 * to condense then we preload it too. This will ensure
2077 * that force condensing happens in the next txg.
2078 */
2079 if (++m > metaslab_preload_limit && !msp->ms_condense_wanted) {
2080 continue;
2081 }
2082
2083 VERIFY(taskq_dispatch(mg->mg_taskq, metaslab_preload,
2084 msp, TQ_SLEEP) != NULL);
2085 }
2086 mutex_exit(&mg->mg_lock);
2087 }
2088
2089 /*
2090 * Determine if the space map's on-disk footprint is past our tolerance
2091 * for inefficiency. We would like to use the following criteria to make
2092 * our decision:
2093 *
2094 * 1. The size of the space map object should not dramatically increase as a
2095 * result of writing out the free space range tree.
2096 *
2097 * 2. The minimal on-disk space map representation is zfs_condense_pct/100
2098 * times the size than the free space range tree representation
2099 * (i.e. zfs_condense_pct = 110 and in-core = 1MB, minimal = 1.1.MB).
2100 *
2101 * 3. The on-disk size of the space map should actually decrease.
2102 *
2103 * Checking the first condition is tricky since we don't want to walk
2104 * the entire AVL tree calculating the estimated on-disk size. Instead we
2105 * use the size-ordered range tree in the metaslab and calculate the
2106 * size required to write out the largest segment in our free tree. If the
2107 * size required to represent that segment on disk is larger than the space
2108 * map object then we avoid condensing this map.
2109 *
2110 * To determine the second criterion we use a best-case estimate and assume
2111 * each segment can be represented on-disk as a single 64-bit entry. We refer
2112 * to this best-case estimate as the space map's minimal form.
2113 *
2114 * Unfortunately, we cannot compute the on-disk size of the space map in this
2115 * context because we cannot accurately compute the effects of compression, etc.
2116 * Instead, we apply the heuristic described in the block comment for
2117 * zfs_metaslab_condense_block_threshold - we only condense if the space used
2118 * is greater than a threshold number of blocks.
2119 */
2176 ASSERT3U(spa_sync_pass(spa), ==, 1);
2177 ASSERT(msp->ms_loaded);
2178
2179
2180 spa_dbgmsg(spa, "condensing: txg %llu, msp[%llu] %p, vdev id %llu, "
2181 "spa %s, smp size %llu, segments %lu, forcing condense=%s", txg,
2182 msp->ms_id, msp, msp->ms_group->mg_vd->vdev_id,
2183 msp->ms_group->mg_vd->vdev_spa->spa_name,
2184 space_map_length(msp->ms_sm), avl_numnodes(&msp->ms_tree->rt_root),
2185 msp->ms_condense_wanted ? "TRUE" : "FALSE");
2186
2187 msp->ms_condense_wanted = B_FALSE;
2188
2189 /*
2190 * Create an range tree that is 100% allocated. We remove segments
2191 * that have been freed in this txg, any deferred frees that exist,
2192 * and any allocation in the future. Removing segments should be
2193 * a relatively inexpensive operation since we expect these trees to
2194 * have a small number of nodes.
2195 */
2196 condense_tree = range_tree_create(NULL, NULL, &msp->ms_lock);
2197 range_tree_add(condense_tree, msp->ms_start, msp->ms_size);
2198
2199 /*
2200 * Remove what's been freed in this txg from the condense_tree.
2201 * Since we're in sync_pass 1, we know that all the frees from
2202 * this txg are in the freeingtree.
2203 */
2204 range_tree_walk(msp->ms_freeingtree, range_tree_remove, condense_tree);
2205
2206 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2207 range_tree_walk(msp->ms_defertree[t],
2208 range_tree_remove, condense_tree);
2209 }
2210
2211 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
2212 range_tree_walk(msp->ms_alloctree[(txg + t) & TXG_MASK],
2213 range_tree_remove, condense_tree);
2214 }
2215
2216 /*
2217 * We're about to drop the metaslab's lock thus allowing
2218 * other consumers to change it's content. Set the
2219 * metaslab's ms_condensing flag to ensure that
2220 * allocations on this metaslab do not occur while we're
2221 * in the middle of committing it to disk. This is only critical
2222 * for the ms_tree as all other range trees use per txg
2223 * views of their content.
2224 */
2225 msp->ms_condensing = B_TRUE;
2226
2227 mutex_exit(&msp->ms_lock);
2228 space_map_truncate(sm, tx);
2229 mutex_enter(&msp->ms_lock);
2230
2231 /*
2232 * While we would ideally like to create a space map representation
2233 * that consists only of allocation records, doing so can be
2234 * prohibitively expensive because the in-core free tree can be
2235 * large, and therefore computationally expensive to subtract
2236 * from the condense_tree. Instead we sync out two trees, a cheap
2237 * allocation only tree followed by the in-core free tree. While not
2238 * optimal, this is typically close to optimal, and much cheaper to
2239 * compute.
2240 */
2241 space_map_write(sm, condense_tree, SM_ALLOC, tx);
2242 range_tree_vacate(condense_tree, NULL, NULL);
2243 range_tree_destroy(condense_tree);
2244
2245 space_map_write(sm, msp->ms_tree, SM_FREE, tx);
2246 msp->ms_condensing = B_FALSE;
2247 }
2248
2249 /*
2250 * Write a metaslab to disk in the context of the specified transaction group.
2251 */
2252 void
2253 metaslab_sync(metaslab_t *msp, uint64_t txg)
2254 {
2255 metaslab_group_t *mg = msp->ms_group;
2256 vdev_t *vd = mg->mg_vd;
2257 spa_t *spa = vd->vdev_spa;
2258 objset_t *mos = spa_meta_objset(spa);
2259 range_tree_t *alloctree = msp->ms_alloctree[txg & TXG_MASK];
2260 dmu_tx_t *tx;
2261 uint64_t object = space_map_object(msp->ms_sm);
2262
2263 ASSERT(!vd->vdev_ishole);
2264
2265 mutex_enter(&msp->ms_lock);
2266
2267 /*
2268 * This metaslab has just been added so there's no work to do now.
2269 */
2270 if (msp->ms_freeingtree == NULL) {
2271 ASSERT3P(alloctree, ==, NULL);
2272 mutex_exit(&msp->ms_lock);
2273 return;
2274 }
2275
2276 ASSERT3P(alloctree, !=, NULL);
2277 ASSERT3P(msp->ms_freeingtree, !=, NULL);
2278 ASSERT3P(msp->ms_freedtree, !=, NULL);
2279
2280 /*
2281 * Normally, we don't want to process a metaslab if there
2282 * are no allocations or frees to perform. However, if the metaslab
2283 * is being forced to condense and it's loaded, we need to let it
2284 * through.
2285 */
2286 if (range_tree_space(alloctree) == 0 &&
2287 range_tree_space(msp->ms_freeingtree) == 0 &&
2288 !(msp->ms_loaded && msp->ms_condense_wanted)) {
2289 mutex_exit(&msp->ms_lock);
2290 return;
2291 }
2292
2293
2294 VERIFY(txg <= spa_final_dirty_txg(spa));
2295
2296 /*
2297 * The only state that can actually be changing concurrently with
2298 * metaslab_sync() is the metaslab's ms_tree. No other thread can
2299 * be modifying this txg's alloctree, freeingtree, freedtree, or
2300 * space_map_phys_t. Therefore, we only hold ms_lock to satify
2301 * space map ASSERTs. We drop it whenever we call into the DMU,
2302 * because the DMU can call down to us (e.g. via zio_free()) at
2303 * any time.
2304 */
2305
2306 tx = dmu_tx_create_assigned(spa_get_dsl(spa), txg);
2307
2308 if (msp->ms_sm == NULL) {
2309 uint64_t new_object;
2310
2311 new_object = space_map_alloc(mos, tx);
2312 VERIFY3U(new_object, !=, 0);
2313
2314 VERIFY0(space_map_open(&msp->ms_sm, mos, new_object,
2315 msp->ms_start, msp->ms_size, vd->vdev_ashift,
2316 &msp->ms_lock));
2317 ASSERT(msp->ms_sm != NULL);
2318 }
2319
2320 /*
2321 * Note: metaslab_condense() clears the space map's histogram.
2322 * Therefore we must verify and remove this histogram before
2323 * condensing.
2324 */
2325 metaslab_group_histogram_verify(mg);
2326 metaslab_class_histogram_verify(mg->mg_class);
2327 metaslab_group_histogram_remove(mg, msp);
2328
2329 if (msp->ms_loaded && spa_sync_pass(spa) == 1 &&
2330 metaslab_should_condense(msp)) {
2331 metaslab_condense(msp, txg, tx);
2332 } else {
2333 space_map_write(msp->ms_sm, alloctree, SM_ALLOC, tx);
2334 space_map_write(msp->ms_sm, msp->ms_freeingtree, SM_FREE, tx);
2335 }
2336
2337 if (msp->ms_loaded) {
2338 /*
2339 * When the space map is loaded, we have an accruate
2340 * histogram in the range tree. This gives us an opportunity
2341 * to bring the space map's histogram up-to-date so we clear
2342 * it first before updating it.
2343 */
2344 space_map_histogram_clear(msp->ms_sm);
2345 space_map_histogram_add(msp->ms_sm, msp->ms_tree, tx);
2346
2347 /*
2348 * Since we've cleared the histogram we need to add back
2349 * any free space that has already been processed, plus
2350 * any deferred space. This allows the on-disk histogram
2351 * to accurately reflect all free space even if some space
2352 * is not yet available for allocation (i.e. deferred).
2353 */
2354 space_map_histogram_add(msp->ms_sm, msp->ms_freedtree, tx);
2355
2356 /*
2357 * Add back any deferred free space that has not been
2358 * added back into the in-core free tree yet. This will
2359 * ensure that we don't end up with a space map histogram
2387 */
2388 if (spa_sync_pass(spa) == 1) {
2389 range_tree_swap(&msp->ms_freeingtree, &msp->ms_freedtree);
2390 } else {
2391 range_tree_vacate(msp->ms_freeingtree,
2392 range_tree_add, msp->ms_freedtree);
2393 }
2394 range_tree_vacate(alloctree, NULL, NULL);
2395
2396 ASSERT0(range_tree_space(msp->ms_alloctree[txg & TXG_MASK]));
2397 ASSERT0(range_tree_space(msp->ms_alloctree[TXG_CLEAN(txg) & TXG_MASK]));
2398 ASSERT0(range_tree_space(msp->ms_freeingtree));
2399
2400 mutex_exit(&msp->ms_lock);
2401
2402 if (object != space_map_object(msp->ms_sm)) {
2403 object = space_map_object(msp->ms_sm);
2404 dmu_write(mos, vd->vdev_ms_array, sizeof (uint64_t) *
2405 msp->ms_id, sizeof (uint64_t), &object, tx);
2406 }
2407 dmu_tx_commit(tx);
2408 }
2409
2410 /*
2411 * Called after a transaction group has completely synced to mark
2412 * all of the metaslab's free space as usable.
2413 */
2414 void
2415 metaslab_sync_done(metaslab_t *msp, uint64_t txg)
2416 {
2417 metaslab_group_t *mg = msp->ms_group;
2418 vdev_t *vd = mg->mg_vd;
2419 spa_t *spa = vd->vdev_spa;
2420 range_tree_t **defer_tree;
2421 int64_t alloc_delta, defer_delta;
2422 boolean_t defer_allowed = B_TRUE;
2423
2424 ASSERT(!vd->vdev_ishole);
2425
2426 mutex_enter(&msp->ms_lock);
2427
2428 /*
2429 * If this metaslab is just becoming available, initialize its
2430 * range trees and add its capacity to the vdev.
2431 */
2432 if (msp->ms_freedtree == NULL) {
2433 for (int t = 0; t < TXG_SIZE; t++) {
2434 ASSERT(msp->ms_alloctree[t] == NULL);
2435
2436 msp->ms_alloctree[t] = range_tree_create(NULL, msp,
2437 &msp->ms_lock);
2438 }
2439
2440 ASSERT3P(msp->ms_freeingtree, ==, NULL);
2441 msp->ms_freeingtree = range_tree_create(NULL, msp,
2442 &msp->ms_lock);
2443
2444 ASSERT3P(msp->ms_freedtree, ==, NULL);
2445 msp->ms_freedtree = range_tree_create(NULL, msp,
2446 &msp->ms_lock);
2447
2448 for (int t = 0; t < TXG_DEFER_SIZE; t++) {
2449 ASSERT(msp->ms_defertree[t] == NULL);
2450
2451 msp->ms_defertree[t] = range_tree_create(NULL, msp,
2452 &msp->ms_lock);
2453 }
2454
2455 vdev_space_update(vd, 0, 0, msp->ms_size);
2456 }
2457
2458 defer_tree = &msp->ms_defertree[txg % TXG_DEFER_SIZE];
2459
2460 uint64_t free_space = metaslab_class_get_space(spa_normal_class(spa)) -
2461 metaslab_class_get_alloc(spa_normal_class(spa));
2462 if (free_space <= spa_get_slop_space(spa)) {
2463 defer_allowed = B_FALSE;
2464 }
2465
2466 defer_delta = 0;
2467 alloc_delta = space_map_alloc_delta(msp->ms_sm);
2468 if (defer_allowed) {
2469 defer_delta = range_tree_space(msp->ms_freedtree) -
2470 range_tree_space(*defer_tree);
2471 } else {
2472 defer_delta -= range_tree_space(*defer_tree);
2473 }
2474
2475 vdev_space_update(vd, alloc_delta + defer_delta, defer_delta, 0);
2476
2477 /*
2478 * If there's a metaslab_load() in progress, wait for it to complete
2479 * so that we have a consistent view of the in-core space map.
2480 */
2481 metaslab_load_wait(msp);
2482
2483 /*
2484 * Move the frees from the defer_tree back to the free
2485 * range tree (if it's loaded). Swap the freed_tree and the
2486 * defer_tree -- this is safe to do because we've just emptied out
2487 * the defer_tree.
2488 */
2489 if (spa_get_auto_trim(spa) == SPA_AUTO_TRIM_ON &&
2490 !vd->vdev_man_trimming) {
2491 range_tree_walk(*defer_tree, metaslab_trim_add, msp);
2492 if (!defer_allowed) {
2493 range_tree_walk(msp->ms_freedtree, metaslab_trim_add,
2494 msp);
2495 }
2496 }
2497 range_tree_vacate(*defer_tree,
2498 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
2499 if (defer_allowed) {
2500 range_tree_swap(&msp->ms_freedtree, defer_tree);
2501 } else {
2502 range_tree_vacate(msp->ms_freedtree,
2503 msp->ms_loaded ? range_tree_add : NULL, msp->ms_tree);
2504 }
2505
2506 space_map_update(msp->ms_sm);
2507
2508 msp->ms_deferspace += defer_delta;
2509 ASSERT3S(msp->ms_deferspace, >=, 0);
2510 ASSERT3S(msp->ms_deferspace, <=, msp->ms_size);
2511 if (msp->ms_deferspace != 0) {
2512 /*
2513 * Keep syncing this metaslab until all deferred frees
2514 * are back in circulation.
2515 */
2516 vdev_dirty(vd, VDD_METASLAB, msp, txg + 1);
2520 * Calculate the new weights before unloading any metaslabs.
2521 * This will give us the most accurate weighting.
2522 */
2523 metaslab_group_sort(mg, msp, metaslab_weight(msp));
2524
2525 /*
2526 * If the metaslab is loaded and we've not tried to load or allocate
2527 * from it in 'metaslab_unload_delay' txgs, then unload it.
2528 */
2529 if (msp->ms_loaded &&
2530 msp->ms_selected_txg + metaslab_unload_delay < txg) {
2531 for (int t = 1; t < TXG_CONCURRENT_STATES; t++) {
2532 VERIFY0(range_tree_space(
2533 msp->ms_alloctree[(txg + t) & TXG_MASK]));
2534 }
2535
2536 if (!metaslab_debug_unload)
2537 metaslab_unload(msp);
2538 }
2539
2540 mutex_exit(&msp->ms_lock);
2541 }
2542
2543 void
2544 metaslab_sync_reassess(metaslab_group_t *mg)
2545 {
2546 metaslab_group_alloc_update(mg);
2547 mg->mg_fragmentation = metaslab_group_fragmentation(mg);
2548
2549 /*
2550 * Preload the next potential metaslabs
2551 */
2552 metaslab_group_preload(mg);
2553 }
2554
2555 static uint64_t
2556 metaslab_distance(metaslab_t *msp, dva_t *dva)
2557 {
2558 uint64_t ms_shift = msp->ms_group->mg_vd->vdev_ms_shift;
2559 uint64_t offset = DVA_GET_OFFSET(dva) >> ms_shift;
2560 uint64_t start = msp->ms_id;
2561
2562 if (msp->ms_group->mg_vd->vdev_id != DVA_GET_VDEV(dva))
2563 return (1ULL << 63);
2564
2565 if (offset < start)
2566 return ((start - offset) << ms_shift);
2567 if (offset > start)
2568 return ((offset - start) << ms_shift);
2569 return (0);
2570 }
2571
2572 /*
2726 }
2727
2728 static uint64_t
2729 metaslab_block_alloc(metaslab_t *msp, uint64_t size, uint64_t txg)
2730 {
2731 uint64_t start;
2732 range_tree_t *rt = msp->ms_tree;
2733 metaslab_class_t *mc = msp->ms_group->mg_class;
2734
2735 VERIFY(!msp->ms_condensing);
2736
2737 start = mc->mc_ops->msop_alloc(msp, size);
2738 if (start != -1ULL) {
2739 metaslab_group_t *mg = msp->ms_group;
2740 vdev_t *vd = mg->mg_vd;
2741
2742 VERIFY0(P2PHASE(start, 1ULL << vd->vdev_ashift));
2743 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
2744 VERIFY3U(range_tree_space(rt) - size, <=, msp->ms_size);
2745 range_tree_remove(rt, start, size);
2746 metaslab_trim_remove(msp, start, size);
2747
2748 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
2749 vdev_dirty(mg->mg_vd, VDD_METASLAB, msp, txg);
2750
2751 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], start, size);
2752
2753 /* Track the last successful allocation */
2754 msp->ms_alloc_txg = txg;
2755 metaslab_verify_space(msp, txg);
2756 }
2757
2758 /*
2759 * Now that we've attempted the allocation we need to update the
2760 * metaslab's maximum block size since it may have changed.
2761 */
2762 msp->ms_max_size = metaslab_block_maxsize(msp);
2763 return (start);
2764 }
2765
2766 static uint64_t
2767 metaslab_group_alloc_normal(metaslab_group_t *mg, zio_alloc_list_t *zal,
2768 uint64_t asize, uint64_t txg, uint64_t min_distance, dva_t *dva, int d,
2769 int flags)
2770 {
2771 metaslab_t *msp = NULL;
2772 uint64_t offset = -1ULL;
2773 uint64_t activation_weight;
2774 uint64_t target_distance;
2775 int i;
2776
2777 activation_weight = METASLAB_WEIGHT_PRIMARY;
2778 for (i = 0; i < d; i++) {
2779 if (DVA_GET_VDEV(&dva[i]) == mg->mg_vd->vdev_id) {
2780 activation_weight = METASLAB_WEIGHT_SECONDARY;
2781 break;
2782 }
2783 }
2784
2785 metaslab_t *search = kmem_alloc(sizeof (*search), KM_SLEEP);
2786 search->ms_weight = UINT64_MAX;
2787 search->ms_start = 0;
2788 for (;;) {
2789 boolean_t was_active;
2790 boolean_t pass_primary = B_TRUE;
2791 avl_tree_t *t = &mg->mg_metaslab_tree;
2792 avl_index_t idx;
2793
2794 mutex_enter(&mg->mg_lock);
2795
2796 /*
2797 * Find the metaslab with the highest weight that is less
2798 * than what we've already tried. In the common case, this
2799 * means that we will examine each metaslab at most once.
2800 * Note that concurrent callers could reorder metaslabs
2801 * by activation/passivation once we have dropped the mg_lock.
2802 * If a metaslab is activated by another thread, and we fail
2803 * to allocate from the metaslab we have selected, we may
2804 * not try the newly-activated metaslab, and instead activate
2805 * another metaslab. This is not optimal, but generally
2806 * does not cause any problems (a possible exception being
2807 * if every metaslab is completely full except for the
2808 * the newly-activated metaslab which we fail to examine).
2809 */
2810 msp = avl_find(t, search, &idx);
2811 if (msp == NULL)
2812 msp = avl_nearest(t, idx, AVL_AFTER);
2813 for (; msp != NULL; msp = AVL_NEXT(t, msp)) {
2814
2815 if (!metaslab_should_allocate(msp, asize)) {
2816 metaslab_trace_add(zal, mg, msp, asize, d,
2817 TRACE_TOO_SMALL);
2818 continue;
2819 }
2820
2821 /*
2822 * If the selected metaslab is condensing, skip it.
2823 */
2824 if (msp->ms_condensing)
2825 continue;
2826
2827 was_active = msp->ms_weight & METASLAB_ACTIVE_MASK;
2828 if (flags & METASLAB_USE_WEIGHT_SECONDARY) {
2829 if (!pass_primary) {
2830 DTRACE_PROBE(metaslab_use_secondary);
2831 activation_weight =
2832 METASLAB_WEIGHT_SECONDARY;
2833 break;
2834 }
2835
2836 pass_primary = B_FALSE;
2837 } else {
2838 if (activation_weight ==
2839 METASLAB_WEIGHT_PRIMARY)
2840 break;
2841
2842 target_distance = min_distance +
2843 (space_map_allocated(msp->ms_sm) != 0 ? 0 :
2844 min_distance >> 1);
2845
2846 for (i = 0; i < d; i++)
2847 if (metaslab_distance(msp, &dva[i]) <
2848 target_distance)
2849 break;
2850 if (i == d)
2851 break;
2852 }
2853 }
2854 mutex_exit(&mg->mg_lock);
2855 if (msp == NULL) {
2856 kmem_free(search, sizeof (*search));
2857 return (-1ULL);
2858 }
2859 search->ms_weight = msp->ms_weight;
2860 search->ms_start = msp->ms_start + 1;
2861
2862 mutex_enter(&msp->ms_lock);
2863
2864 /*
2865 * Ensure that the metaslab we have selected is still
2866 * capable of handling our request. It's possible that
2867 * another thread may have changed the weight while we
2868 * were blocked on the metaslab lock. We check the
2869 * active status first to see if we need to reselect
2870 * a new metaslab.
2871 */
2872 if (was_active && !(msp->ms_weight & METASLAB_ACTIVE_MASK)) {
2873 mutex_exit(&msp->ms_lock);
2954 metaslab_passivate(msp,
2955 metaslab_weight_from_range_tree(msp));
2956 }
2957
2958 /*
2959 * We have just failed an allocation attempt, check
2960 * that metaslab_should_allocate() agrees. Otherwise,
2961 * we may end up in an infinite loop retrying the same
2962 * metaslab.
2963 */
2964 ASSERT(!metaslab_should_allocate(msp, asize));
2965 mutex_exit(&msp->ms_lock);
2966 }
2967 mutex_exit(&msp->ms_lock);
2968 kmem_free(search, sizeof (*search));
2969 return (offset);
2970 }
2971
2972 static uint64_t
2973 metaslab_group_alloc(metaslab_group_t *mg, zio_alloc_list_t *zal,
2974 uint64_t asize, uint64_t txg, uint64_t min_distance, dva_t *dva,
2975 int d, int flags)
2976 {
2977 uint64_t offset;
2978 ASSERT(mg->mg_initialized);
2979
2980 offset = metaslab_group_alloc_normal(mg, zal, asize, txg,
2981 min_distance, dva, d, flags);
2982
2983 mutex_enter(&mg->mg_lock);
2984 if (offset == -1ULL) {
2985 mg->mg_failed_allocations++;
2986 metaslab_trace_add(zal, mg, NULL, asize, d,
2987 TRACE_GROUP_FAILURE);
2988 if (asize == SPA_GANGBLOCKSIZE) {
2989 /*
2990 * This metaslab group was unable to allocate
2991 * the minimum gang block size so it must be out of
2992 * space. We must notify the allocation throttle
2993 * to start skipping allocation attempts to this
2994 * metaslab group until more space becomes available.
2995 * Note: this failure cannot be caused by the
2996 * allocation throttle since the allocation throttle
2997 * is only responsible for skipping devices and
2998 * not failing block allocations.
2999 */
3000 mg->mg_no_free_space = B_TRUE;
3001 }
3002 }
3003 mg->mg_allocations++;
3004 mutex_exit(&mg->mg_lock);
3005 return (offset);
3006 }
3007
3008 /*
3009 * If we have to write a ditto block (i.e. more than one DVA for a given BP)
3010 * on the same vdev as an existing DVA of this BP, then try to allocate it
3011 * at least (vdev_asize / (2 ^ ditto_same_vdev_distance_shift)) away from the
3012 * existing DVAs.
3013 */
3014 int ditto_same_vdev_distance_shift = 3;
3015
3016 /*
3017 * Allocate a block for the specified i/o.
3018 */
3019 static int
3020 metaslab_alloc_dva(spa_t *spa, metaslab_class_t *mc, uint64_t psize,
3021 dva_t *dva, int d, dva_t *hintdva, uint64_t txg, int flags,
3022 zio_alloc_list_t *zal)
3023 {
3024 metaslab_group_t *mg, *rotor;
3025 vdev_t *vd;
3026 boolean_t try_hard = B_FALSE;
3027
3028 ASSERT(!DVA_IS_VALID(&dva[d]));
3029
3030 /*
3031 * For testing, make some blocks above a certain size be gang blocks.
3032 */
3033 if (psize >= metaslab_gang_bang && (ddi_get_lbolt() & 3) == 0) {
3034 metaslab_trace_add(zal, NULL, NULL, psize, d, TRACE_FORCE_GANG);
3035 return (SET_ERROR(ENOSPC));
3036 }
3037
3038 /*
3039 * Start at the rotor and loop through all mgs until we find something.
3045 * consecutive vdevs. If we're forced to reuse a vdev before we've
3046 * allocated all of our ditto blocks, then try and spread them out on
3047 * that vdev as much as possible. If it turns out to not be possible,
3048 * gradually lower our standards until anything becomes acceptable.
3049 * Also, allocating on consecutive vdevs (as opposed to random vdevs)
3050 * gives us hope of containing our fault domains to something we're
3051 * able to reason about. Otherwise, any two top-level vdev failures
3052 * will guarantee the loss of data. With consecutive allocation,
3053 * only two adjacent top-level vdev failures will result in data loss.
3054 *
3055 * If we are doing gang blocks (hintdva is non-NULL), try to keep
3056 * ourselves on the same vdev as our gang block header. That
3057 * way, we can hope for locality in vdev_cache, plus it makes our
3058 * fault domains something tractable.
3059 */
3060 if (hintdva) {
3061 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&hintdva[d]));
3062
3063 /*
3064 * It's possible the vdev we're using as the hint no
3065 * longer exists (i.e. removed). Consult the rotor when
3066 * all else fails.
3067 */
3068 if (vd != NULL) {
3069 mg = vd->vdev_mg;
3070
3071 if (flags & METASLAB_HINTBP_AVOID &&
3072 mg->mg_next != NULL)
3073 mg = mg->mg_next;
3074 } else {
3075 mg = mc->mc_rotor;
3076 }
3077 } else if (d != 0) {
3078 vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dva[d - 1]));
3079 mg = vd->vdev_mg->mg_next;
3080 } else {
3081 mg = mc->mc_rotor;
3082 }
3083
3084 /*
3085 * If the hint put us into the wrong metaslab class, or into a
3086 * metaslab group that has been passivated, just follow the rotor.
3087 */
3088 if (mg->mg_class != mc || mg->mg_activation_count <= 0)
3143 ASSERT(mg->mg_class == mc);
3144
3145 /*
3146 * If we don't need to try hard, then require that the
3147 * block be 1/8th of the device away from any other DVAs
3148 * in this BP. If we are trying hard, allow any offset
3149 * to be used (distance=0).
3150 */
3151 uint64_t distance = 0;
3152 if (!try_hard) {
3153 distance = vd->vdev_asize >>
3154 ditto_same_vdev_distance_shift;
3155 if (distance <= (1ULL << vd->vdev_ms_shift))
3156 distance = 0;
3157 }
3158
3159 uint64_t asize = vdev_psize_to_asize(vd, psize);
3160 ASSERT(P2PHASE(asize, 1ULL << vd->vdev_ashift) == 0);
3161
3162 uint64_t offset = metaslab_group_alloc(mg, zal, asize, txg,
3163 distance, dva, d, flags);
3164
3165 if (offset != -1ULL) {
3166 /*
3167 * If we've just selected this metaslab group,
3168 * figure out whether the corresponding vdev is
3169 * over- or under-used relative to the pool,
3170 * and set an allocation bias to even it out.
3171 */
3172 if (mc->mc_aliquot == 0 && metaslab_bias_enabled) {
3173 vdev_stat_t *vs = &vd->vdev_stat;
3174 vdev_stat_t *pvs = &vd->vdev_parent->vdev_stat;
3175 int64_t vu, cu, vu_io;
3176
3177 vu = (vs->vs_alloc * 100) / (vs->vs_space + 1);
3178 cu = (mc->mc_alloc * 100) / (mc->mc_space + 1);
3179 vu_io =
3180 (((vs->vs_iotime[ZIO_TYPE_WRITE] * 100) /
3181 (pvs->vs_iotime[ZIO_TYPE_WRITE] + 1)) *
3182 (vd->vdev_parent->vdev_children)) - 100;
3183
3184 /*
3185 * Calculate how much more or less we should
3186 * try to allocate from this device during
3187 * this iteration around the rotor.
3188 * For example, if a device is 80% full
3189 * and the pool is 20% full then we should
3190 * reduce allocations by 60% on this device.
3191 *
3192 * mg_bias = (20 - 80) * 512K / 100 = -307K
3193 *
3194 * This reduces allocations by 307K for this
3195 * iteration.
3196 */
3197 mg->mg_bias = ((cu - vu) *
3198 (int64_t)mg->mg_aliquot) / 100;
3199
3200 /*
3201 * Experiment: space-based DVA allocator 0,
3202 * latency-based 1 or hybrid 2.
3203 */
3204 switch (metaslab_alloc_dva_algorithm) {
3205 case 1:
3206 mg->mg_bias =
3207 (vu_io * (int64_t)mg->mg_aliquot) /
3208 100;
3209 break;
3210 case 2:
3211 mg->mg_bias =
3212 ((((cu - vu) + vu_io) / 2) *
3213 (int64_t)mg->mg_aliquot) / 100;
3214 break;
3215 default:
3216 break;
3217 }
3218 } else if (!metaslab_bias_enabled) {
3219 mg->mg_bias = 0;
3220 }
3221
3222 if (atomic_add_64_nv(&mc->mc_aliquot, asize) >=
3223 mg->mg_aliquot + mg->mg_bias) {
3224 mc->mc_rotor = mg->mg_next;
3225 mc->mc_aliquot = 0;
3226 }
3227
3228 DVA_SET_VDEV(&dva[d], vd->vdev_id);
3229 DVA_SET_OFFSET(&dva[d], offset);
3230 DVA_SET_GANG(&dva[d], !!(flags & METASLAB_GANG_HEADER));
3231 DVA_SET_ASIZE(&dva[d], asize);
3232 DTRACE_PROBE3(alloc_dva_probe, uint64_t, vd->vdev_id,
3233 uint64_t, offset, uint64_t, psize);
3234
3235 return (0);
3236 }
3237 next:
3238 mc->mc_rotor = mg->mg_next;
3239 mc->mc_aliquot = 0;
3240 } while ((mg = mg->mg_next) != rotor);
3241
3242 /*
3243 * If we haven't tried hard, do so now.
3244 */
3245 if (!try_hard) {
3246 try_hard = B_TRUE;
3247 goto top;
3248 }
3249
3250 bzero(&dva[d], sizeof (dva_t));
3251
3252 metaslab_trace_add(zal, rotor, NULL, psize, d, TRACE_ENOSPC);
3253 return (SET_ERROR(ENOSPC));
3254 }
3255
3256 /*
3257 * Free the block represented by DVA in the context of the specified
3258 * transaction group.
3259 */
3260 void
3261 metaslab_free_dva(spa_t *spa, const dva_t *dva, uint64_t txg, boolean_t now)
3262 {
3263 uint64_t vdev = DVA_GET_VDEV(dva);
3264 uint64_t offset = DVA_GET_OFFSET(dva);
3265 uint64_t size = DVA_GET_ASIZE(dva);
3266 vdev_t *vd;
3267 metaslab_t *msp;
3268
3269 DTRACE_PROBE3(free_dva_probe, uint64_t, vdev,
3270 uint64_t, offset, uint64_t, size);
3271
3272 ASSERT(DVA_IS_VALID(dva));
3273
3274 if (txg > spa_freeze_txg(spa))
3275 return;
3276
3277 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
3278 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count) {
3279 cmn_err(CE_WARN, "metaslab_free_dva(): bad DVA %llu:%llu",
3280 (u_longlong_t)vdev, (u_longlong_t)offset);
3281 ASSERT(0);
3282 return;
3283 }
3284
3285 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3286
3287 if (DVA_GET_GANG(dva))
3288 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
3289
3290 mutex_enter(&msp->ms_lock);
3291
3292 if (now) {
3293 range_tree_remove(msp->ms_alloctree[txg & TXG_MASK],
3294 offset, size);
3295
3296 VERIFY(!msp->ms_condensing);
3297 VERIFY3U(offset, >=, msp->ms_start);
3298 VERIFY3U(offset + size, <=, msp->ms_start + msp->ms_size);
3299 VERIFY3U(range_tree_space(msp->ms_tree) + size, <=,
3300 msp->ms_size);
3301 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
3302 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
3303 range_tree_add(msp->ms_tree, offset, size);
3304 if (spa_get_auto_trim(spa) == SPA_AUTO_TRIM_ON &&
3305 !vd->vdev_man_trimming)
3306 metaslab_trim_add(msp, offset, size);
3307 msp->ms_max_size = metaslab_block_maxsize(msp);
3308 } else {
3309 VERIFY3U(txg, ==, spa->spa_syncing_txg);
3310 if (range_tree_space(msp->ms_freeingtree) == 0)
3311 vdev_dirty(vd, VDD_METASLAB, msp, txg);
3312 range_tree_add(msp->ms_freeingtree, offset, size);
3313 }
3314
3315 mutex_exit(&msp->ms_lock);
3316 }
3317
3318 /*
3319 * Intent log support: upon opening the pool after a crash, notify the SPA
3320 * of blocks that the intent log has allocated for immediate write, but
3321 * which are still considered free by the SPA because the last transaction
3322 * group didn't commit yet.
3323 */
3324 static int
3325 metaslab_claim_dva(spa_t *spa, const dva_t *dva, uint64_t txg)
3326 {
3327 uint64_t vdev = DVA_GET_VDEV(dva);
3328 uint64_t offset = DVA_GET_OFFSET(dva);
3329 uint64_t size = DVA_GET_ASIZE(dva);
3330 vdev_t *vd;
3331 metaslab_t *msp;
3332 int error = 0;
3333
3334 ASSERT(DVA_IS_VALID(dva));
3335
3336 if ((vd = vdev_lookup_top(spa, vdev)) == NULL ||
3337 (offset >> vd->vdev_ms_shift) >= vd->vdev_ms_count)
3338 return (SET_ERROR(ENXIO));
3339
3340 msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3341
3342 if (DVA_GET_GANG(dva))
3343 size = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE);
3344
3345 mutex_enter(&msp->ms_lock);
3346
3347 if ((txg != 0 && spa_writeable(spa)) || !msp->ms_loaded)
3348 error = metaslab_activate(msp, METASLAB_WEIGHT_SECONDARY);
3349
3350 if (error == 0 && !range_tree_contains(msp->ms_tree, offset, size))
3351 error = SET_ERROR(ENOENT);
3352
3353 if (error || txg == 0) { /* txg == 0 indicates dry run */
3354 mutex_exit(&msp->ms_lock);
3355 return (error);
3356 }
3357
3358 VERIFY(!msp->ms_condensing);
3359 VERIFY0(P2PHASE(offset, 1ULL << vd->vdev_ashift));
3360 VERIFY0(P2PHASE(size, 1ULL << vd->vdev_ashift));
3361 VERIFY3U(range_tree_space(msp->ms_tree) - size, <=, msp->ms_size);
3362 range_tree_remove(msp->ms_tree, offset, size);
3363 metaslab_trim_remove(msp, offset, size);
3364
3365 if (spa_writeable(spa)) { /* don't dirty if we're zdb(1M) */
3366 if (range_tree_space(msp->ms_alloctree[txg & TXG_MASK]) == 0)
3367 vdev_dirty(vd, VDD_METASLAB, msp, txg);
3368 range_tree_add(msp->ms_alloctree[txg & TXG_MASK], offset, size);
3369 }
3370
3371 mutex_exit(&msp->ms_lock);
3372
3373 return (0);
3374 }
3375
3376 /*
3377 * Reserve some allocation slots. The reservation system must be called
3378 * before we call into the allocator. If there aren't any available slots
3379 * then the I/O will be throttled until an I/O completes and its slots are
3380 * freed up. The function returns true if it was successful in placing
3381 * the reservation.
3382 */
3383 boolean_t
3384 metaslab_class_throttle_reserve(metaslab_class_t *mc, int slots, zio_t *zio,
3385 int flags)
3386 {
3387 uint64_t available_slots = 0;
3388 boolean_t slot_reserved = B_FALSE;
3389
3390 ASSERT(mc->mc_alloc_throttle_enabled);
3391 mutex_enter(&mc->mc_lock);
3392
3393 uint64_t reserved_slots = refcount_count(&mc->mc_alloc_slots);
3404 }
3405 zio->io_flags |= ZIO_FLAG_IO_ALLOCATING;
3406 slot_reserved = B_TRUE;
3407 }
3408
3409 mutex_exit(&mc->mc_lock);
3410 return (slot_reserved);
3411 }
3412
3413 void
3414 metaslab_class_throttle_unreserve(metaslab_class_t *mc, int slots, zio_t *zio)
3415 {
3416 ASSERT(mc->mc_alloc_throttle_enabled);
3417 mutex_enter(&mc->mc_lock);
3418 for (int d = 0; d < slots; d++) {
3419 (void) refcount_remove(&mc->mc_alloc_slots, zio);
3420 }
3421 mutex_exit(&mc->mc_lock);
3422 }
3423
3424 int
3425 metaslab_alloc(spa_t *spa, metaslab_class_t *mc, uint64_t psize, blkptr_t *bp,
3426 int ndvas, uint64_t txg, blkptr_t *hintbp, int flags,
3427 zio_alloc_list_t *zal, zio_t *zio)
3428 {
3429 dva_t *dva = bp->blk_dva;
3430 dva_t *hintdva = hintbp->blk_dva;
3431 int error = 0;
3432
3433 ASSERT(bp->blk_birth == 0);
3434 ASSERT(BP_PHYSICAL_BIRTH(bp) == 0);
3435
3436 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
3437
3438 if (mc->mc_rotor == NULL) { /* no vdevs in this class */
3439 spa_config_exit(spa, SCL_ALLOC, FTAG);
3440 return (SET_ERROR(ENOSPC));
3441 }
3442
3443 ASSERT(ndvas > 0 && ndvas <= spa_max_replication(spa));
3444 ASSERT(BP_GET_NDVAS(bp) == 0);
3445 ASSERT(hintbp == NULL || ndvas <= BP_GET_NDVAS(hintbp));
3446 ASSERT3P(zal, !=, NULL);
3447
3448 if (mc == spa_special_class(spa) && !BP_IS_METADATA(bp) &&
3449 !(flags & (METASLAB_GANG_HEADER)) &&
3450 !(spa->spa_meta_policy.spa_small_data_to_special &&
3451 psize <= spa->spa_meta_policy.spa_small_data_to_special)) {
3452 error = metaslab_alloc_dva(spa, spa_normal_class(spa),
3453 psize, &dva[WBC_NORMAL_DVA], 0, NULL, txg,
3454 flags | METASLAB_USE_WEIGHT_SECONDARY, zal);
3455 if (error == 0) {
3456 error = metaslab_alloc_dva(spa, mc, psize,
3457 &dva[WBC_SPECIAL_DVA], 0, NULL, txg, flags, zal);
3458 if (error != 0) {
3459 error = 0;
3460 /*
3461 * Change the place of NORMAL and cleanup the
3462 * second DVA. After that this BP is just a
3463 * regular BP with one DVA
3464 *
3465 * This operation is valid only if:
3466 * WBC_SPECIAL_DVA is dva[0]
3467 * WBC_NORMAL_DVA is dva[1]
3468 *
3469 * see wbc.h
3470 */
3471 bcopy(&dva[WBC_NORMAL_DVA],
3472 &dva[WBC_SPECIAL_DVA], sizeof (dva_t));
3473 bzero(&dva[WBC_NORMAL_DVA], sizeof (dva_t));
3474
3475 /*
3476 * Allocation of special DVA has failed,
3477 * so this BP will be a regular BP and need
3478 * to update the metaslab group's queue depth
3479 * based on the newly allocated dva.
3480 */
3481 metaslab_group_alloc_increment(spa,
3482 DVA_GET_VDEV(&dva[0]), zio, flags);
3483 } else {
3484 BP_SET_SPECIAL(bp, 1);
3485 }
3486 } else {
3487 spa_config_exit(spa, SCL_ALLOC, FTAG);
3488 return (error);
3489 }
3490 } else {
3491 for (int d = 0; d < ndvas; d++) {
3492 error = metaslab_alloc_dva(spa, mc, psize, dva, d,
3493 hintdva, txg, flags, zal);
3494 if (error != 0) {
3495 for (d--; d >= 0; d--) {
3496 metaslab_free_dva(spa, &dva[d],
3497 txg, B_TRUE);
3498 metaslab_group_alloc_decrement(spa,
3499 DVA_GET_VDEV(&dva[d]), zio, flags);
3500 bzero(&dva[d], sizeof (dva_t));
3501 }
3502 spa_config_exit(spa, SCL_ALLOC, FTAG);
3503 return (error);
3504 } else {
3505 /*
3506 * Update the metaslab group's queue depth
3507 * based on the newly allocated dva.
3508 */
3509 metaslab_group_alloc_increment(spa,
3510 DVA_GET_VDEV(&dva[d]), zio, flags);
3511 }
3512 }
3513 ASSERT(BP_GET_NDVAS(bp) == ndvas);
3514 }
3515 ASSERT(error == 0);
3516
3517 spa_config_exit(spa, SCL_ALLOC, FTAG);
3518
3519 BP_SET_BIRTH(bp, txg, txg);
3520
3521 return (0);
3522 }
3523
3524 void
3525 metaslab_free(spa_t *spa, const blkptr_t *bp, uint64_t txg, boolean_t now)
3526 {
3527 const dva_t *dva = bp->blk_dva;
3528 int ndvas = BP_GET_NDVAS(bp);
3529
3530 ASSERT(!BP_IS_HOLE(bp));
3531 ASSERT(!now || bp->blk_birth >= spa_syncing_txg(spa));
3532
3533 spa_config_enter(spa, SCL_FREE, FTAG, RW_READER);
3534
3535 if (BP_IS_SPECIAL(bp)) {
3536 int start_dva;
3537 wbc_data_t *wbc_data = spa_get_wbc_data(spa);
3538
3539 mutex_enter(&wbc_data->wbc_lock);
3540 start_dva = wbc_first_valid_dva(bp, wbc_data, B_TRUE);
3541 mutex_exit(&wbc_data->wbc_lock);
3542
3543 /*
3544 * Actual freeing should not be locked as
3545 * the block is already exempted from WBC
3546 * trees, and thus will not be moved
3547 */
3548 metaslab_free_dva(spa, &dva[WBC_NORMAL_DVA], txg, now);
3549 if (start_dva == 0) {
3550 metaslab_free_dva(spa, &dva[WBC_SPECIAL_DVA],
3551 txg, now);
3552 }
3553 } else {
3554 for (int d = 0; d < ndvas; d++)
3555 metaslab_free_dva(spa, &dva[d], txg, now);
3556 }
3557
3558 spa_config_exit(spa, SCL_FREE, FTAG);
3559 }
3560
3561 int
3562 metaslab_claim(spa_t *spa, const blkptr_t *bp, uint64_t txg)
3563 {
3564 const dva_t *dva = bp->blk_dva;
3565 int ndvas = BP_GET_NDVAS(bp);
3566 int error = 0;
3567
3568 ASSERT(!BP_IS_HOLE(bp));
3569
3570 if (txg != 0) {
3571 /*
3572 * First do a dry run to make sure all DVAs are claimable,
3573 * so we don't have to unwind from partial failures below.
3574 */
3575 if ((error = metaslab_claim(spa, bp, 0)) != 0)
3576 return (error);
3577 }
3578
3579 spa_config_enter(spa, SCL_ALLOC, FTAG, RW_READER);
3580
3581 if (BP_IS_SPECIAL(bp)) {
3582 int start_dva;
3583 wbc_data_t *wbc_data = spa_get_wbc_data(spa);
3584
3585 mutex_enter(&wbc_data->wbc_lock);
3586 start_dva = wbc_first_valid_dva(bp, wbc_data, B_FALSE);
3587
3588 /*
3589 * Actual claiming should be under lock for WBC blocks. It must
3590 * be done to ensure zdb will not fail. The only other user of
3591 * the claiming is ZIL whose blocks can not be WBC ones, and
3592 * thus the lock will not be held for them.
3593 */
3594 error = metaslab_claim_dva(spa,
3595 &dva[WBC_NORMAL_DVA], txg);
3596 if (error == 0 && start_dva == 0) {
3597 error = metaslab_claim_dva(spa,
3598 &dva[WBC_SPECIAL_DVA], txg);
3599 }
3600
3601 mutex_exit(&wbc_data->wbc_lock);
3602 } else {
3603 for (int d = 0; d < ndvas; d++)
3604 if ((error = metaslab_claim_dva(spa,
3605 &dva[d], txg)) != 0)
3606 break;
3607 }
3608
3609 spa_config_exit(spa, SCL_ALLOC, FTAG);
3610
3611 ASSERT(error == 0 || txg == 0);
3612
3613 return (error);
3614 }
3615
3616 void
3617 metaslab_check_free(spa_t *spa, const blkptr_t *bp)
3618 {
3619 if ((zfs_flags & ZFS_DEBUG_ZIO_FREE) == 0)
3620 return;
3621
3622 if (BP_IS_SPECIAL(bp)) {
3623 /* Do not check frees for WBC blocks */
3624 return;
3625 }
3626
3627 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
3628 for (int i = 0; i < BP_GET_NDVAS(bp); i++) {
3629 uint64_t vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
3630 vdev_t *vd = vdev_lookup_top(spa, vdev);
3631 uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]);
3632 uint64_t size = DVA_GET_ASIZE(&bp->blk_dva[i]);
3633 metaslab_t *msp = vd->vdev_ms[offset >> vd->vdev_ms_shift];
3634
3635 if (msp->ms_loaded) {
3636 range_tree_verify(msp->ms_tree, offset, size);
3637 range_tree_verify(msp->ms_cur_ts->ts_tree,
3638 offset, size);
3639 if (msp->ms_prev_ts != NULL) {
3640 range_tree_verify(msp->ms_prev_ts->ts_tree,
3641 offset, size);
3642 }
3643 }
3644
3645 range_tree_verify(msp->ms_freeingtree, offset, size);
3646 range_tree_verify(msp->ms_freedtree, offset, size);
3647 for (int j = 0; j < TXG_DEFER_SIZE; j++)
3648 range_tree_verify(msp->ms_defertree[j], offset, size);
3649 }
3650 spa_config_exit(spa, SCL_VDEV, FTAG);
3651 }
3652
3653 /*
3654 * Trims all free space in the metaslab. Returns the root TRIM zio (that the
3655 * caller should zio_wait() for) and the amount of space in the metaslab that
3656 * has been scheduled for trimming in the `delta' return argument.
3657 */
3658 zio_t *
3659 metaslab_trim_all(metaslab_t *msp, uint64_t *delta)
3660 {
3661 boolean_t was_loaded;
3662 uint64_t trimmed_space;
3663 zio_t *trim_io;
3664
3665 ASSERT(!MUTEX_HELD(&msp->ms_group->mg_lock));
3666
3667 mutex_enter(&msp->ms_lock);
3668
3669 while (msp->ms_loading)
3670 metaslab_load_wait(msp);
3671 /* If we loaded the metaslab, unload it when we're done. */
3672 was_loaded = msp->ms_loaded;
3673 if (!was_loaded) {
3674 if (metaslab_load(msp) != 0) {
3675 mutex_exit(&msp->ms_lock);
3676 return (0);
3677 }
3678 }
3679 /* Flush out any scheduled extents and add everything in ms_tree. */
3680 range_tree_vacate(msp->ms_cur_ts->ts_tree, NULL, NULL);
3681 range_tree_walk(msp->ms_tree, metaslab_trim_add, msp);
3682
3683 /* Force this trim to take place ASAP. */
3684 if (msp->ms_prev_ts != NULL)
3685 metaslab_free_trimset(msp->ms_prev_ts);
3686 msp->ms_prev_ts = msp->ms_cur_ts;
3687 msp->ms_cur_ts = metaslab_new_trimset(0, &msp->ms_lock);
3688 trimmed_space = range_tree_space(msp->ms_tree);
3689 if (!was_loaded)
3690 metaslab_unload(msp);
3691
3692 trim_io = metaslab_exec_trim(msp);
3693 mutex_exit(&msp->ms_lock);
3694 *delta = trimmed_space;
3695
3696 return (trim_io);
3697 }
3698
3699 /*
3700 * Notifies the trimsets in a metaslab that an extent has been allocated.
3701 * This removes the segment from the queues of extents awaiting to be trimmed.
3702 */
3703 static void
3704 metaslab_trim_remove(void *arg, uint64_t offset, uint64_t size)
3705 {
3706 metaslab_t *msp = arg;
3707
3708 range_tree_remove_overlap(msp->ms_cur_ts->ts_tree, offset, size);
3709 if (msp->ms_prev_ts != NULL) {
3710 range_tree_remove_overlap(msp->ms_prev_ts->ts_tree, offset,
3711 size);
3712 }
3713 }
3714
3715 /*
3716 * Notifies the trimsets in a metaslab that an extent has been freed.
3717 * This adds the segment to the currently open queue of extents awaiting
3718 * to be trimmed.
3719 */
3720 static void
3721 metaslab_trim_add(void *arg, uint64_t offset, uint64_t size)
3722 {
3723 metaslab_t *msp = arg;
3724 ASSERT(msp->ms_cur_ts != NULL);
3725 range_tree_add(msp->ms_cur_ts->ts_tree, offset, size);
3726 }
3727
3728 /*
3729 * Does a metaslab's automatic trim operation processing. This must be
3730 * called from metaslab_sync, with the txg number of the txg. This function
3731 * issues trims in intervals as dictated by the zfs_txgs_per_trim tunable.
3732 */
3733 void
3734 metaslab_auto_trim(metaslab_t *msp, uint64_t txg)
3735 {
3736 /* for atomicity */
3737 uint64_t txgs_per_trim = zfs_txgs_per_trim;
3738
3739 ASSERT(!MUTEX_HELD(&msp->ms_lock));
3740 mutex_enter(&msp->ms_lock);
3741
3742 /*
3743 * Since we typically have hundreds of metaslabs per vdev, but we only
3744 * trim them once every zfs_txgs_per_trim txgs, it'd be best if we
3745 * could sequence the TRIM commands from all metaslabs so that they
3746 * don't all always pound the device in the same txg. We do so by
3747 * artificially inflating the birth txg of the first trim set by a
3748 * sequence number derived from the metaslab's starting offset
3749 * (modulo zfs_txgs_per_trim). Thus, for the default 200 metaslabs and
3750 * 32 txgs per trim, we'll only be trimming ~6.25 metaslabs per txg.
3751 *
3752 * If we detect that the txg has advanced too far ahead of ts_birth,
3753 * it means our birth txg is out of lockstep. Recompute it by
3754 * rounding down to the nearest zfs_txgs_per_trim multiple and adding
3755 * our metaslab id modulo zfs_txgs_per_trim.
3756 */
3757 if (txg > msp->ms_cur_ts->ts_birth + txgs_per_trim) {
3758 msp->ms_cur_ts->ts_birth = (txg / txgs_per_trim) *
3759 txgs_per_trim + (msp->ms_id % txgs_per_trim);
3760 }
3761
3762 /* Time to swap out the current and previous trimsets */
3763 if (txg == msp->ms_cur_ts->ts_birth + txgs_per_trim) {
3764 if (msp->ms_prev_ts != NULL) {
3765 if (msp->ms_trimming_ts != NULL) {
3766 spa_t *spa = msp->ms_group->mg_class->mc_spa;
3767 /*
3768 * The previous trim run is still ongoing, so
3769 * the device is reacting slowly to our trim
3770 * requests. Drop this trimset, so as not to
3771 * back the device up with trim requests.
3772 */
3773 spa_trimstats_auto_slow_incr(spa);
3774 metaslab_free_trimset(msp->ms_prev_ts);
3775 } else if (msp->ms_group->mg_vd->vdev_man_trimming) {
3776 /*
3777 * If a manual trim is ongoing, we want to
3778 * inhibit autotrim temporarily so it doesn't
3779 * slow down the manual trim.
3780 */
3781 metaslab_free_trimset(msp->ms_prev_ts);
3782 } else {
3783 /*
3784 * Trim out aged extents on the vdevs - these
3785 * are safe to be destroyed now. We'll keep
3786 * the trimset around to deny allocations from
3787 * these regions while the trims are ongoing.
3788 */
3789 zio_nowait(metaslab_exec_trim(msp));
3790 }
3791 }
3792 msp->ms_prev_ts = msp->ms_cur_ts;
3793 msp->ms_cur_ts = metaslab_new_trimset(txg, &msp->ms_lock);
3794 }
3795 mutex_exit(&msp->ms_lock);
3796 }
3797
3798 static void
3799 metaslab_trim_done(zio_t *zio)
3800 {
3801 metaslab_t *msp = zio->io_private;
3802 boolean_t held;
3803
3804 ASSERT(msp != NULL);
3805 ASSERT(msp->ms_trimming_ts != NULL);
3806 held = MUTEX_HELD(&msp->ms_lock);
3807 if (!held)
3808 mutex_enter(&msp->ms_lock);
3809 metaslab_free_trimset(msp->ms_trimming_ts);
3810 msp->ms_trimming_ts = NULL;
3811 cv_signal(&msp->ms_trim_cv);
3812 if (!held)
3813 mutex_exit(&msp->ms_lock);
3814 }
3815
3816 /*
3817 * Executes a zio_trim on a range tree holding freed extents in the metaslab.
3818 */
3819 static zio_t *
3820 metaslab_exec_trim(metaslab_t *msp)
3821 {
3822 metaslab_group_t *mg = msp->ms_group;
3823 spa_t *spa = mg->mg_class->mc_spa;
3824 vdev_t *vd = mg->mg_vd;
3825 range_tree_t *trim_tree;
3826 zio_t *zio;
3827
3828 ASSERT(MUTEX_HELD(&msp->ms_lock));
3829
3830 /* wait for a preceding trim to finish */
3831 while (msp->ms_trimming_ts != NULL)
3832 cv_wait(&msp->ms_trim_cv, &msp->ms_lock);
3833 msp->ms_trimming_ts = msp->ms_prev_ts;
3834 msp->ms_prev_ts = NULL;
3835 trim_tree = msp->ms_trimming_ts->ts_tree;
3836 #ifdef DEBUG
3837 if (msp->ms_loaded) {
3838 for (range_seg_t *rs = avl_first(&trim_tree->rt_root);
3839 rs != NULL; rs = AVL_NEXT(&trim_tree->rt_root, rs)) {
3840 if (!range_tree_contains(msp->ms_tree,
3841 rs->rs_start, rs->rs_end - rs->rs_start)) {
3842 panic("trimming allocated region; mss=%p",
3843 (void*)rs);
3844 }
3845 }
3846 }
3847 #endif
3848
3849 /* Nothing to trim */
3850 if (range_tree_space(trim_tree) == 0) {
3851 metaslab_free_trimset(msp->ms_trimming_ts);
3852 msp->ms_trimming_ts = 0;
3853 return (zio_root(spa, NULL, NULL, 0));
3854 }
3855 zio = zio_trim(spa, vd, trim_tree, metaslab_trim_done, msp, 0,
3856 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY |
3857 ZIO_FLAG_CONFIG_WRITER, msp);
3858
3859 return (zio);
3860 }
3861
3862 /*
3863 * Allocates and initializes a new trimset structure. The `txg' argument
3864 * indicates when this trimset was born and `lock' indicates the lock to
3865 * link to the range tree.
3866 */
3867 static metaslab_trimset_t *
3868 metaslab_new_trimset(uint64_t txg, kmutex_t *lock)
3869 {
3870 metaslab_trimset_t *ts;
3871
3872 ts = kmem_zalloc(sizeof (*ts), KM_SLEEP);
3873 ts->ts_birth = txg;
3874 ts->ts_tree = range_tree_create(NULL, NULL, lock);
3875
3876 return (ts);
3877 }
3878
3879 /*
3880 * Destroys and frees a trim set previously allocated by metaslab_new_trimset.
3881 */
3882 static void
3883 metaslab_free_trimset(metaslab_trimset_t *ts)
3884 {
3885 range_tree_vacate(ts->ts_tree, NULL, NULL);
3886 range_tree_destroy(ts->ts_tree);
3887 kmem_free(ts, sizeof (*ts));
3888 }
3889
3890 /*
3891 * Checks whether an allocation conflicts with an ongoing trim operation in
3892 * the given metaslab. This function takes a segment starting at `*offset'
3893 * of `size' and checks whether it hits any region in the metaslab currently
3894 * being trimmed. If yes, it tries to adjust the allocation to the end of
3895 * the region being trimmed (P2ROUNDUP aligned by `align'), but only up to
3896 * `limit' (no part of the allocation is allowed to go past this point).
3897 *
3898 * Returns B_FALSE if either the original allocation wasn't in conflict, or
3899 * the conflict could be resolved by adjusting the value stored in `offset'
3900 * such that the whole allocation still fits below `limit'. Returns B_TRUE
3901 * if the allocation conflict couldn't be resolved.
3902 */
3903 static boolean_t metaslab_check_trim_conflict(metaslab_t *msp,
3904 uint64_t *offset, uint64_t size, uint64_t align, uint64_t limit)
3905 {
3906 uint64_t new_offset;
3907
3908 if (msp->ms_trimming_ts == NULL)
3909 /* no trim conflict, original offset is OK */
3910 return (B_FALSE);
3911
3912 new_offset = P2ROUNDUP(range_tree_find_gap(msp->ms_trimming_ts->ts_tree,
3913 *offset, size), align);
3914 if (new_offset != *offset && new_offset + size > limit)
3915 /* trim conflict and adjustment not possible */
3916 return (B_TRUE);
3917
3918 /* trim conflict, but adjusted offset still within limit */
3919 *offset = new_offset;
3920 return (B_FALSE);
3921 }
|