4 * This file and its contents are supplied under the terms of the
5 * Common Development and Distribution License ("CDDL"), version 1.0.
6 * You may only use this file in accordance with the terms of version
9 * A full copy of the text of the CDDL should have accompanied this
10 * source. A copy of the CDDL is also available via the Internet at
11 * http://www.illumos.org/license/CDDL.
17 * Copyright (c) 2014, 2017 by Delphix. All rights reserved.
20 #include <sys/zfs_context.h>
22 #include <sys/spa_impl.h>
23 #include <sys/vdev_impl.h>
24 #include <sys/fs/zfs.h>
26 #include <sys/zio_checksum.h>
27 #include <sys/metaslab.h>
28 #include <sys/refcount.h>
30 #include <sys/vdev_indirect_mapping.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/dsl_synctask.h>
38 * An indirect vdev corresponds to a vdev that has been removed. Since
39 * we cannot rewrite block pointers of snapshots, etc., we keep a
40 * mapping from old location on the removed device to the new location
41 * on another device in the pool and use this mapping whenever we need
42 * to access the DVA. Unfortunately, this mapping did not respect
43 * logical block boundaries when it was first created, and so a DVA on
44 * this indirect vdev may be "split" into multiple sections that each
45 * map to a different location. As a consequence, not all DVAs can be
46 * translated to an equivalent new DVA. Instead we must provide a
47 * "vdev_remap" operation that executes a callback on each contiguous
48 * segment of the new location. This function is used in multiple ways:
50 * - i/os to this vdev use the callback to determine where the
51 * data is now located, and issue child i/os for each segment's new
54 * - frees and claims to this vdev use the callback to free or claim
55 * each mapped segment. (Note that we don't actually need to claim
56 * log blocks on indirect vdevs, because we don't allocate to
57 * removing vdevs. However, zdb uses zio_claim() for its leak
62 * "Big theory statement" for how we mark blocks obsolete.
64 * When a block on an indirect vdev is freed or remapped, a section of
65 * that vdev's mapping may no longer be referenced (aka "obsolete"). We
66 * keep track of how much of each mapping entry is obsolete. When
67 * an entry becomes completely obsolete, we can remove it, thus reducing
68 * the memory used by the mapping. The complete picture of obsolescence
69 * is given by the following data structures, described below:
70 * - the entry-specific obsolete count
71 * - the vdev-specific obsolete spacemap
72 * - the pool-specific obsolete bpobj
74 * == On disk data structures used ==
76 * We track the obsolete space for the pool using several objects. Each
77 * of these objects is created on demand and freed when no longer
78 * needed, and is assumed to be empty if it does not exist.
79 * SPA_FEATURE_OBSOLETE_COUNTS includes the count of these objects.
81 * - Each vic_mapping_object (associated with an indirect vdev) can
82 * have a vimp_counts_object. This is an array of uint32_t's
83 * with the same number of entries as the vic_mapping_object. When
84 * the mapping is condensed, entries from the vic_obsolete_sm_object
85 * (see below) are folded into the counts. Therefore, each
86 * obsolete_counts entry tells us the number of bytes in the
87 * corresponding mapping entry that were not referenced when the
88 * mapping was last condensed.
90 * - Each indirect or removing vdev can have a vic_obsolete_sm_object.
91 * This is a space map containing an alloc entry for every DVA that
92 * has been obsoleted since the last time this indirect vdev was
93 * condensed. We use this object in order to improve performance
94 * when marking a DVA as obsolete. Instead of modifying an arbitrary
95 * offset of the vimp_counts_object, we only need to append an entry
96 * to the end of this object. When a DVA becomes obsolete, it is
97 * added to the obsolete space map. This happens when the DVA is
98 * freed, remapped and not referenced by a snapshot, or the last
99 * snapshot referencing it is destroyed.
101 * - Each dataset can have a ds_remap_deadlist object. This is a
102 * deadlist object containing all blocks that were remapped in this
103 * dataset but referenced in a previous snapshot. Blocks can *only*
104 * appear on this list if they were remapped (dsl_dataset_block_remapped);
105 * blocks that were killed in a head dataset are put on the normal
106 * ds_deadlist and marked obsolete when they are freed.
108 * - The pool can have a dp_obsolete_bpobj. This is a list of blocks
109 * in the pool that need to be marked obsolete. When a snapshot is
110 * destroyed, we move some of the ds_remap_deadlist to the obsolete
111 * bpobj (see dsl_destroy_snapshot_handle_remaps()). We then
112 * asynchronously process the obsolete bpobj, moving its entries to
113 * the specific vdevs' obsolete space maps.
115 * == Summary of how we mark blocks as obsolete ==
117 * - When freeing a block: if any DVA is on an indirect vdev, append to
118 * vic_obsolete_sm_object.
119 * - When remapping a block, add dva to ds_remap_deadlist (if prev snap
120 * references; otherwise append to vic_obsolete_sm_object).
121 * - When freeing a snapshot: move parts of ds_remap_deadlist to
122 * dp_obsolete_bpobj (same algorithm as ds_deadlist).
123 * - When syncing the spa: process dp_obsolete_bpobj, moving ranges to
124 * individual vdev's vic_obsolete_sm_object.
128 * "Big theory statement" for how we condense indirect vdevs.
130 * Condensing an indirect vdev's mapping is the process of determining
131 * the precise counts of obsolete space for each mapping entry (by
132 * integrating the obsolete spacemap into the obsolete counts) and
133 * writing out a new mapping that contains only referenced entries.
135 * We condense a vdev when we expect the mapping to shrink (see
136 * vdev_indirect_should_condense()), but only perform one condense at a
137 * time to limit the memory usage. In addition, we use a separate
138 * open-context thread (spa_condense_indirect_thread) to incrementally
139 * create the new mapping object in a way that minimizes the impact on
140 * the rest of the system.
142 * == Generating a new mapping ==
144 * To generate a new mapping, we follow these steps:
146 * 1. Save the old obsolete space map and create a new mapping object
147 * (see spa_condense_indirect_start_sync()). This initializes the
148 * spa_condensing_indirect_phys with the "previous obsolete space map",
149 * which is now read only. Newly obsolete DVAs will be added to a
150 * new (initially empty) obsolete space map, and will not be
151 * considered as part of this condense operation.
153 * 2. Construct in memory the precise counts of obsolete space for each
154 * mapping entry, by incorporating the obsolete space map into the
155 * counts. (See vdev_indirect_mapping_load_obsolete_{counts,spacemap}().)
157 * 3. Iterate through each mapping entry, writing to the new mapping any
158 * entries that are not completely obsolete (i.e. which don't have
159 * obsolete count == mapping length). (See
160 * spa_condense_indirect_generate_new_mapping().)
162 * 4. Destroy the old mapping object and switch over to the new one
163 * (spa_condense_indirect_complete_sync).
165 * == Restarting from failure ==
167 * To restart the condense when we import/open the pool, we must start
168 * at the 2nd step above: reconstruct the precise counts in memory,
169 * based on the space map + counts. Then in the 3rd step, we start
170 * iterating where we left off: at vimp_max_offset of the new mapping
174 boolean_t zfs_condense_indirect_vdevs_enable = B_TRUE;
177 * Condense if at least this percent of the bytes in the mapping is
178 * obsolete. With the default of 25%, the amount of space mapped
179 * will be reduced to 1% of its original size after at most 16
180 * condenses. Higher values will condense less often (causing less
181 * i/o); lower values will reduce the mapping size more quickly.
183 int zfs_indirect_condense_obsolete_pct = 25;
186 * Condense if the obsolete space map takes up more than this amount of
187 * space on disk (logically). This limits the amount of disk space
188 * consumed by the obsolete space map; the default of 1GB is small enough
189 * that we typically don't mind "wasting" it.
191 uint64_t zfs_condense_max_obsolete_bytes = 1024 * 1024 * 1024;
194 * Don't bother condensing if the mapping uses less than this amount of
195 * memory. The default of 128KB is considered a "trivial" amount of
196 * memory and not worth reducing.
198 uint64_t zfs_condense_min_mapping_bytes = 128 * 1024;
201 * This is used by the test suite so that it can ensure that certain
202 * actions happen while in the middle of a condense (which might otherwise
203 * complete too quickly). If used to reduce the performance impact of
204 * condensing in production, a maximum value of 1 should be sufficient.
206 int zfs_condense_indirect_commit_entry_delay_ticks = 0;
209 * If an indirect split block contains more than this many possible unique
210 * combinations when being reconstructed, consider it too computationally
211 * expensive to check them all. Instead, try at most 100 randomly-selected
212 * combinations each time the block is accessed. This allows all segment
213 * copies to participate fairly in the reconstruction when all combinations
214 * cannot be checked and prevents repeated use of one bad copy.
216 int zfs_reconstruct_indirect_combinations_max = 256;
220 * Enable to simulate damaged segments and validate reconstruction.
223 unsigned long zfs_reconstruct_indirect_damage_fraction = 0;
226 * The indirect_child_t represents the vdev that we will read from, when we
227 * need to read all copies of the data (e.g. for scrub or reconstruction).
228 * For plain (non-mirror) top-level vdevs (i.e. is_vdev is not a mirror),
229 * ic_vdev is the same as is_vdev. However, for mirror top-level vdevs,
230 * ic_vdev is a child of the mirror.
232 typedef struct indirect_child {
237 * ic_duplicate is NULL when the ic_data contents are unique, when it
238 * is determined to be a duplicate it references the primary child.
240 struct indirect_child *ic_duplicate;
241 list_node_t ic_node; /* node on is_unique_child */
245 * The indirect_split_t represents one mapped segment of an i/o to the
246 * indirect vdev. For non-split (contiguously-mapped) blocks, there will be
247 * only one indirect_split_t, with is_split_offset==0 and is_size==io_size.
248 * For split blocks, there will be several of these.
250 typedef struct indirect_split {
251 list_node_t is_node; /* link on iv_splits */
254 * is_split_offset is the offset into the i/o.
255 * This is the sum of the previous splits' is_size's.
257 uint64_t is_split_offset;
259 vdev_t *is_vdev; /* top-level vdev */
260 uint64_t is_target_offset; /* offset on is_vdev */
262 int is_children; /* number of entries in is_child[] */
263 int is_unique_children; /* number of entries in is_unique_child */
264 list_t is_unique_child;
267 * is_good_child is the child that we are currently using to
268 * attempt reconstruction.
270 indirect_child_t *is_good_child;
272 indirect_child_t is_child[1]; /* variable-length */
276 * The indirect_vsd_t is associated with each i/o to the indirect vdev.
277 * It is the "Vdev-Specific Data" in the zio_t's io_vsd.
279 typedef struct indirect_vsd {
280 boolean_t iv_split_block;
281 boolean_t iv_reconstruct;
282 uint64_t iv_unique_combinations;
283 uint64_t iv_attempts;
284 uint64_t iv_attempts_max;
286 list_t iv_splits; /* list of indirect_split_t's */
290 vdev_indirect_map_free(zio_t *zio)
292 indirect_vsd_t *iv = zio->io_vsd;
294 indirect_split_t *is;
295 while ((is = list_head(&iv->iv_splits)) != NULL) {
296 for (int c = 0; c < is->is_children; c++) {
297 indirect_child_t *ic = &is->is_child[c];
298 if (ic->ic_data != NULL)
299 abd_free(ic->ic_data);
301 list_remove(&iv->iv_splits, is);
303 indirect_child_t *ic;
304 while ((ic = list_head(&is->is_unique_child)) != NULL)
305 list_remove(&is->is_unique_child, ic);
307 list_destroy(&is->is_unique_child);
310 offsetof(indirect_split_t, is_child[is->is_children]));
312 kmem_free(iv, sizeof (*iv));
315 static const zio_vsd_ops_t vdev_indirect_vsd_ops = {
316 vdev_indirect_map_free,
317 zio_vsd_default_cksum_report
320 * Mark the given offset and size as being obsolete.
323 vdev_indirect_mark_obsolete(vdev_t *vd, uint64_t offset, uint64_t size)
325 spa_t *spa = vd->vdev_spa;
327 ASSERT3U(vd->vdev_indirect_config.vic_mapping_object, !=, 0);
328 ASSERT(vd->vdev_removing || vd->vdev_ops == &vdev_indirect_ops);
330 VERIFY(vdev_indirect_mapping_entry_for_offset(
331 vd->vdev_indirect_mapping, offset) != NULL);
333 if (spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS)) {
334 mutex_enter(&vd->vdev_obsolete_lock);
335 range_tree_add(vd->vdev_obsolete_segments, offset, size);
336 mutex_exit(&vd->vdev_obsolete_lock);
337 vdev_dirty(vd, 0, NULL, spa_syncing_txg(spa));
342 * Mark the DVA vdev_id:offset:size as being obsolete in the given tx. This
343 * wrapper is provided because the DMU does not know about vdev_t's and
344 * cannot directly call vdev_indirect_mark_obsolete.
347 spa_vdev_indirect_mark_obsolete(spa_t *spa, uint64_t vdev_id, uint64_t offset,
348 uint64_t size, dmu_tx_t *tx)
350 vdev_t *vd = vdev_lookup_top(spa, vdev_id);
351 ASSERT(dmu_tx_is_syncing(tx));
353 /* The DMU can only remap indirect vdevs. */
354 ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
355 vdev_indirect_mark_obsolete(vd, offset, size);
358 static spa_condensing_indirect_t *
359 spa_condensing_indirect_create(spa_t *spa)
361 spa_condensing_indirect_phys_t *scip =
362 &spa->spa_condensing_indirect_phys;
363 spa_condensing_indirect_t *sci = kmem_zalloc(sizeof (*sci), KM_SLEEP);
364 objset_t *mos = spa->spa_meta_objset;
366 for (int i = 0; i < TXG_SIZE; i++) {
367 list_create(&sci->sci_new_mapping_entries[i],
368 sizeof (vdev_indirect_mapping_entry_t),
369 offsetof(vdev_indirect_mapping_entry_t, vime_node));
372 sci->sci_new_mapping =
373 vdev_indirect_mapping_open(mos, scip->scip_next_mapping_object);
379 spa_condensing_indirect_destroy(spa_condensing_indirect_t *sci)
381 for (int i = 0; i < TXG_SIZE; i++)
382 list_destroy(&sci->sci_new_mapping_entries[i]);
384 if (sci->sci_new_mapping != NULL)
385 vdev_indirect_mapping_close(sci->sci_new_mapping);
387 kmem_free(sci, sizeof (*sci));
391 vdev_indirect_should_condense(vdev_t *vd)
393 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
394 spa_t *spa = vd->vdev_spa;
396 ASSERT(dsl_pool_sync_context(spa->spa_dsl_pool));
398 if (!zfs_condense_indirect_vdevs_enable)
402 * We can only condense one indirect vdev at a time.
404 if (spa->spa_condensing_indirect != NULL)
407 if (spa_shutting_down(spa))
411 * The mapping object size must not change while we are
412 * condensing, so we can only condense indirect vdevs
413 * (not vdevs that are still in the middle of being removed).
415 if (vd->vdev_ops != &vdev_indirect_ops)
419 * If nothing new has been marked obsolete, there is no
420 * point in condensing.
422 if (vd->vdev_obsolete_sm == NULL) {
423 ASSERT0(vdev_obsolete_sm_object(vd));
427 ASSERT(vd->vdev_obsolete_sm != NULL);
429 ASSERT3U(vdev_obsolete_sm_object(vd), ==,
430 space_map_object(vd->vdev_obsolete_sm));
432 uint64_t bytes_mapped = vdev_indirect_mapping_bytes_mapped(vim);
433 uint64_t bytes_obsolete = space_map_allocated(vd->vdev_obsolete_sm);
434 uint64_t mapping_size = vdev_indirect_mapping_size(vim);
435 uint64_t obsolete_sm_size = space_map_length(vd->vdev_obsolete_sm);
437 ASSERT3U(bytes_obsolete, <=, bytes_mapped);
440 * If a high percentage of the bytes that are mapped have become
441 * obsolete, condense (unless the mapping is already small enough).
442 * This has a good chance of reducing the amount of memory used
445 if (bytes_obsolete * 100 / bytes_mapped >=
446 zfs_indirect_condense_obsolete_pct &&
447 mapping_size > zfs_condense_min_mapping_bytes) {
448 zfs_dbgmsg("should condense vdev %llu because obsolete "
449 "spacemap covers %d%% of %lluMB mapping",
450 (u_longlong_t)vd->vdev_id,
451 (int)(bytes_obsolete * 100 / bytes_mapped),
452 (u_longlong_t)bytes_mapped / 1024 / 1024);
457 * If the obsolete space map takes up too much space on disk,
458 * condense in order to free up this disk space.
460 if (obsolete_sm_size >= zfs_condense_max_obsolete_bytes) {
461 zfs_dbgmsg("should condense vdev %llu because obsolete sm "
462 "length %lluMB >= max size %lluMB",
463 (u_longlong_t)vd->vdev_id,
464 (u_longlong_t)obsolete_sm_size / 1024 / 1024,
465 (u_longlong_t)zfs_condense_max_obsolete_bytes /
474 * This sync task completes (finishes) a condense, deleting the old
475 * mapping and replacing it with the new one.
478 spa_condense_indirect_complete_sync(void *arg, dmu_tx_t *tx)
480 spa_condensing_indirect_t *sci = arg;
481 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
482 spa_condensing_indirect_phys_t *scip =
483 &spa->spa_condensing_indirect_phys;
484 vdev_t *vd = vdev_lookup_top(spa, scip->scip_vdev);
485 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
486 objset_t *mos = spa->spa_meta_objset;
487 vdev_indirect_mapping_t *old_mapping = vd->vdev_indirect_mapping;
488 uint64_t old_count = vdev_indirect_mapping_num_entries(old_mapping);
490 vdev_indirect_mapping_num_entries(sci->sci_new_mapping);
492 ASSERT(dmu_tx_is_syncing(tx));
493 ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
494 ASSERT3P(sci, ==, spa->spa_condensing_indirect);
495 for (int i = 0; i < TXG_SIZE; i++) {
496 ASSERT(list_is_empty(&sci->sci_new_mapping_entries[i]));
498 ASSERT(vic->vic_mapping_object != 0);
499 ASSERT3U(vd->vdev_id, ==, scip->scip_vdev);
500 ASSERT(scip->scip_next_mapping_object != 0);
501 ASSERT(scip->scip_prev_obsolete_sm_object != 0);
504 * Reset vdev_indirect_mapping to refer to the new object.
506 rw_enter(&vd->vdev_indirect_rwlock, RW_WRITER);
507 vdev_indirect_mapping_close(vd->vdev_indirect_mapping);
508 vd->vdev_indirect_mapping = sci->sci_new_mapping;
509 rw_exit(&vd->vdev_indirect_rwlock);
511 sci->sci_new_mapping = NULL;
512 vdev_indirect_mapping_free(mos, vic->vic_mapping_object, tx);
513 vic->vic_mapping_object = scip->scip_next_mapping_object;
514 scip->scip_next_mapping_object = 0;
516 space_map_free_obj(mos, scip->scip_prev_obsolete_sm_object, tx);
517 spa_feature_decr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
518 scip->scip_prev_obsolete_sm_object = 0;
522 VERIFY0(zap_remove(mos, DMU_POOL_DIRECTORY_OBJECT,
523 DMU_POOL_CONDENSING_INDIRECT, tx));
524 spa_condensing_indirect_destroy(spa->spa_condensing_indirect);
525 spa->spa_condensing_indirect = NULL;
527 zfs_dbgmsg("finished condense of vdev %llu in txg %llu: "
528 "new mapping object %llu has %llu entries "
529 "(was %llu entries)",
530 vd->vdev_id, dmu_tx_get_txg(tx), vic->vic_mapping_object,
531 new_count, old_count);
533 vdev_config_dirty(spa->spa_root_vdev);
537 * This sync task appends entries to the new mapping object.
540 spa_condense_indirect_commit_sync(void *arg, dmu_tx_t *tx)
542 spa_condensing_indirect_t *sci = arg;
543 uint64_t txg = dmu_tx_get_txg(tx);
544 spa_t *spa = dmu_tx_pool(tx)->dp_spa;
546 ASSERT(dmu_tx_is_syncing(tx));
547 ASSERT3P(sci, ==, spa->spa_condensing_indirect);
549 vdev_indirect_mapping_add_entries(sci->sci_new_mapping,
550 &sci->sci_new_mapping_entries[txg & TXG_MASK], tx);
551 ASSERT(list_is_empty(&sci->sci_new_mapping_entries[txg & TXG_MASK]));
555 * Open-context function to add one entry to the new mapping. The new
556 * entry will be remembered and written from syncing context.
559 spa_condense_indirect_commit_entry(spa_t *spa,
560 vdev_indirect_mapping_entry_phys_t *vimep, uint32_t count)
562 spa_condensing_indirect_t *sci = spa->spa_condensing_indirect;
564 ASSERT3U(count, <, DVA_GET_ASIZE(&vimep->vimep_dst));
566 dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
567 dmu_tx_hold_space(tx, sizeof (*vimep) + sizeof (count));
568 VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
569 int txgoff = dmu_tx_get_txg(tx) & TXG_MASK;
572 * If we are the first entry committed this txg, kick off the sync
573 * task to write to the MOS on our behalf.
575 if (list_is_empty(&sci->sci_new_mapping_entries[txgoff])) {
576 dsl_sync_task_nowait(dmu_tx_pool(tx),
577 spa_condense_indirect_commit_sync, sci,
578 0, ZFS_SPACE_CHECK_NONE, tx);
581 vdev_indirect_mapping_entry_t *vime =
582 kmem_alloc(sizeof (*vime), KM_SLEEP);
583 vime->vime_mapping = *vimep;
584 vime->vime_obsolete_count = count;
585 list_insert_tail(&sci->sci_new_mapping_entries[txgoff], vime);
591 spa_condense_indirect_generate_new_mapping(vdev_t *vd,
592 uint32_t *obsolete_counts, uint64_t start_index, zthr_t *zthr)
594 spa_t *spa = vd->vdev_spa;
595 uint64_t mapi = start_index;
596 vdev_indirect_mapping_t *old_mapping = vd->vdev_indirect_mapping;
597 uint64_t old_num_entries =
598 vdev_indirect_mapping_num_entries(old_mapping);
600 ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
601 ASSERT3U(vd->vdev_id, ==, spa->spa_condensing_indirect_phys.scip_vdev);
603 zfs_dbgmsg("starting condense of vdev %llu from index %llu",
604 (u_longlong_t)vd->vdev_id,
607 while (mapi < old_num_entries) {
609 if (zthr_iscancelled(zthr)) {
610 zfs_dbgmsg("pausing condense of vdev %llu "
611 "at index %llu", (u_longlong_t)vd->vdev_id,
616 vdev_indirect_mapping_entry_phys_t *entry =
617 &old_mapping->vim_entries[mapi];
618 uint64_t entry_size = DVA_GET_ASIZE(&entry->vimep_dst);
619 ASSERT3U(obsolete_counts[mapi], <=, entry_size);
620 if (obsolete_counts[mapi] < entry_size) {
621 spa_condense_indirect_commit_entry(spa, entry,
622 obsolete_counts[mapi]);
625 * This delay may be requested for testing, debugging,
626 * or performance reasons.
628 delay(zfs_condense_indirect_commit_entry_delay_ticks);
637 spa_condense_indirect_thread_check(void *arg, zthr_t *zthr)
641 return (spa->spa_condensing_indirect != NULL);
646 spa_condense_indirect_thread(void *arg, zthr_t *zthr)
651 ASSERT3P(spa->spa_condensing_indirect, !=, NULL);
652 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
653 vd = vdev_lookup_top(spa, spa->spa_condensing_indirect_phys.scip_vdev);
654 ASSERT3P(vd, !=, NULL);
655 spa_config_exit(spa, SCL_VDEV, FTAG);
657 spa_condensing_indirect_t *sci = spa->spa_condensing_indirect;
658 spa_condensing_indirect_phys_t *scip =
659 &spa->spa_condensing_indirect_phys;
661 uint64_t start_index;
662 vdev_indirect_mapping_t *old_mapping = vd->vdev_indirect_mapping;
663 space_map_t *prev_obsolete_sm = NULL;
665 ASSERT3U(vd->vdev_id, ==, scip->scip_vdev);
666 ASSERT(scip->scip_next_mapping_object != 0);
667 ASSERT(scip->scip_prev_obsolete_sm_object != 0);
668 ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
670 for (int i = 0; i < TXG_SIZE; i++) {
672 * The list must start out empty in order for the
673 * _commit_sync() sync task to be properly registered
674 * on the first call to _commit_entry(); so it's wise
675 * to double check and ensure we actually are starting
678 ASSERT(list_is_empty(&sci->sci_new_mapping_entries[i]));
681 VERIFY0(space_map_open(&prev_obsolete_sm, spa->spa_meta_objset,
682 scip->scip_prev_obsolete_sm_object, 0, vd->vdev_asize, 0));
683 counts = vdev_indirect_mapping_load_obsolete_counts(old_mapping);
684 if (prev_obsolete_sm != NULL) {
685 vdev_indirect_mapping_load_obsolete_spacemap(old_mapping,
686 counts, prev_obsolete_sm);
688 space_map_close(prev_obsolete_sm);
691 * Generate new mapping. Determine what index to continue from
692 * based on the max offset that we've already written in the
695 uint64_t max_offset =
696 vdev_indirect_mapping_max_offset(sci->sci_new_mapping);
697 if (max_offset == 0) {
698 /* We haven't written anything to the new mapping yet. */
702 * Pick up from where we left off. _entry_for_offset()
703 * returns a pointer into the vim_entries array. If
704 * max_offset is greater than any of the mappings
705 * contained in the table NULL will be returned and
706 * that indicates we've exhausted our iteration of the
710 vdev_indirect_mapping_entry_phys_t *entry =
711 vdev_indirect_mapping_entry_for_offset_or_next(old_mapping,
716 * We've already written the whole new mapping.
717 * This special value will cause us to skip the
718 * generate_new_mapping step and just do the sync
719 * task to complete the condense.
721 start_index = UINT64_MAX;
723 start_index = entry - old_mapping->vim_entries;
724 ASSERT3U(start_index, <,
725 vdev_indirect_mapping_num_entries(old_mapping));
729 spa_condense_indirect_generate_new_mapping(vd, counts,
732 vdev_indirect_mapping_free_obsolete_counts(old_mapping, counts);
735 * If the zthr has received a cancellation signal while running
736 * in generate_new_mapping() or at any point after that, then bail
737 * early. We don't want to complete the condense if the spa is
740 if (zthr_iscancelled(zthr))
743 VERIFY0(dsl_sync_task(spa_name(spa), NULL,
744 spa_condense_indirect_complete_sync, sci, 0,
745 ZFS_SPACE_CHECK_EXTRA_RESERVED));
749 * Sync task to begin the condensing process.
752 spa_condense_indirect_start_sync(vdev_t *vd, dmu_tx_t *tx)
754 spa_t *spa = vd->vdev_spa;
755 spa_condensing_indirect_phys_t *scip =
756 &spa->spa_condensing_indirect_phys;
758 ASSERT0(scip->scip_next_mapping_object);
759 ASSERT0(scip->scip_prev_obsolete_sm_object);
760 ASSERT0(scip->scip_vdev);
761 ASSERT(dmu_tx_is_syncing(tx));
762 ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
763 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_OBSOLETE_COUNTS));
764 ASSERT(vdev_indirect_mapping_num_entries(vd->vdev_indirect_mapping));
766 uint64_t obsolete_sm_obj = vdev_obsolete_sm_object(vd);
767 ASSERT(obsolete_sm_obj != 0);
769 scip->scip_vdev = vd->vdev_id;
770 scip->scip_next_mapping_object =
771 vdev_indirect_mapping_alloc(spa->spa_meta_objset, tx);
773 scip->scip_prev_obsolete_sm_object = obsolete_sm_obj;
776 * We don't need to allocate a new space map object, since
777 * vdev_indirect_sync_obsolete will allocate one when needed.
779 space_map_close(vd->vdev_obsolete_sm);
780 vd->vdev_obsolete_sm = NULL;
781 VERIFY0(zap_remove(spa->spa_meta_objset, vd->vdev_top_zap,
782 VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM, tx));
784 VERIFY0(zap_add(spa->spa_dsl_pool->dp_meta_objset,
785 DMU_POOL_DIRECTORY_OBJECT,
786 DMU_POOL_CONDENSING_INDIRECT, sizeof (uint64_t),
787 sizeof (*scip) / sizeof (uint64_t), scip, tx));
789 ASSERT3P(spa->spa_condensing_indirect, ==, NULL);
790 spa->spa_condensing_indirect = spa_condensing_indirect_create(spa);
792 zfs_dbgmsg("starting condense of vdev %llu in txg %llu: "
794 vd->vdev_id, dmu_tx_get_txg(tx),
795 (u_longlong_t)scip->scip_prev_obsolete_sm_object,
796 (u_longlong_t)scip->scip_next_mapping_object);
798 zthr_wakeup(spa->spa_condense_zthr);
802 * Sync to the given vdev's obsolete space map any segments that are no longer
803 * referenced as of the given txg.
805 * If the obsolete space map doesn't exist yet, create and open it.
808 vdev_indirect_sync_obsolete(vdev_t *vd, dmu_tx_t *tx)
810 spa_t *spa = vd->vdev_spa;
811 vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
813 ASSERT3U(vic->vic_mapping_object, !=, 0);
814 ASSERT(range_tree_space(vd->vdev_obsolete_segments) > 0);
815 ASSERT(vd->vdev_removing || vd->vdev_ops == &vdev_indirect_ops);
816 ASSERT(spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS));
818 if (vdev_obsolete_sm_object(vd) == 0) {
819 uint64_t obsolete_sm_object =
820 space_map_alloc(spa->spa_meta_objset,
821 vdev_standard_sm_blksz, tx);
823 ASSERT(vd->vdev_top_zap != 0);
824 VERIFY0(zap_add(vd->vdev_spa->spa_meta_objset, vd->vdev_top_zap,
825 VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM,
826 sizeof (obsolete_sm_object), 1, &obsolete_sm_object, tx));
827 ASSERT3U(vdev_obsolete_sm_object(vd), !=, 0);
829 spa_feature_incr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
830 VERIFY0(space_map_open(&vd->vdev_obsolete_sm,
831 spa->spa_meta_objset, obsolete_sm_object,
832 0, vd->vdev_asize, 0));
835 ASSERT(vd->vdev_obsolete_sm != NULL);
836 ASSERT3U(vdev_obsolete_sm_object(vd), ==,
837 space_map_object(vd->vdev_obsolete_sm));
839 space_map_write(vd->vdev_obsolete_sm,
840 vd->vdev_obsolete_segments, SM_ALLOC, SM_NO_VDEVID, tx);
841 range_tree_vacate(vd->vdev_obsolete_segments, NULL, NULL);
845 spa_condense_init(spa_t *spa)
847 int error = zap_lookup(spa->spa_meta_objset,
848 DMU_POOL_DIRECTORY_OBJECT,
849 DMU_POOL_CONDENSING_INDIRECT, sizeof (uint64_t),
850 sizeof (spa->spa_condensing_indirect_phys) / sizeof (uint64_t),
851 &spa->spa_condensing_indirect_phys);
853 if (spa_writeable(spa)) {
854 spa->spa_condensing_indirect =
855 spa_condensing_indirect_create(spa);
858 } else if (error == ENOENT) {
866 spa_condense_fini(spa_t *spa)
868 if (spa->spa_condensing_indirect != NULL) {
869 spa_condensing_indirect_destroy(spa->spa_condensing_indirect);
870 spa->spa_condensing_indirect = NULL;
875 spa_start_indirect_condensing_thread(spa_t *spa)
877 ASSERT3P(spa->spa_condense_zthr, ==, NULL);
878 spa->spa_condense_zthr = zthr_create(spa_condense_indirect_thread_check,
879 spa_condense_indirect_thread, spa);
883 * Gets the obsolete spacemap object from the vdev's ZAP.
884 * Returns the spacemap object, or 0 if it wasn't in the ZAP or the ZAP doesn't
888 vdev_obsolete_sm_object(vdev_t *vd)
890 ASSERT0(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER));
891 if (vd->vdev_top_zap == 0) {
896 int err = zap_lookup(vd->vdev_spa->spa_meta_objset, vd->vdev_top_zap,
897 VDEV_TOP_ZAP_INDIRECT_OBSOLETE_SM, sizeof (sm_obj), 1, &sm_obj);
899 ASSERT(err == 0 || err == ENOENT);
905 vdev_obsolete_counts_are_precise(vdev_t *vd)
907 ASSERT0(spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER));
908 if (vd->vdev_top_zap == 0) {
913 int err = zap_lookup(vd->vdev_spa->spa_meta_objset, vd->vdev_top_zap,
914 VDEV_TOP_ZAP_OBSOLETE_COUNTS_ARE_PRECISE, sizeof (val), 1, &val);
916 ASSERT(err == 0 || err == ENOENT);
923 vdev_indirect_close(vdev_t *vd)
929 vdev_indirect_open(vdev_t *vd, uint64_t *psize, uint64_t *max_psize,
930 uint64_t *logical_ashift, uint64_t *physical_ashift)
932 *psize = *max_psize = vd->vdev_asize +
933 VDEV_LABEL_START_SIZE + VDEV_LABEL_END_SIZE;
934 *logical_ashift = vd->vdev_ashift;
935 *physical_ashift = vd->vdev_physical_ashift;
939 typedef struct remap_segment {
943 uint64_t rs_split_offset;
948 rs_alloc(vdev_t *vd, uint64_t offset, uint64_t asize, uint64_t split_offset)
950 remap_segment_t *rs = kmem_alloc(sizeof (remap_segment_t), KM_SLEEP);
952 rs->rs_offset = offset;
953 rs->rs_asize = asize;
954 rs->rs_split_offset = split_offset;
959 * Given an indirect vdev and an extent on that vdev, it duplicates the
960 * physical entries of the indirect mapping that correspond to the extent
961 * to a new array and returns a pointer to it. In addition, copied_entries
962 * is populated with the number of mapping entries that were duplicated.
964 * Note that the function assumes that the caller holds vdev_indirect_rwlock.
965 * This ensures that the mapping won't change due to condensing as we
966 * copy over its contents.
968 * Finally, since we are doing an allocation, it is up to the caller to
969 * free the array allocated in this function.
971 vdev_indirect_mapping_entry_phys_t *
972 vdev_indirect_mapping_duplicate_adjacent_entries(vdev_t *vd, uint64_t offset,
973 uint64_t asize, uint64_t *copied_entries)
975 vdev_indirect_mapping_entry_phys_t *duplicate_mappings = NULL;
976 vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
977 uint64_t entries = 0;
979 ASSERT(RW_READ_HELD(&vd->vdev_indirect_rwlock));
981 vdev_indirect_mapping_entry_phys_t *first_mapping =
982 vdev_indirect_mapping_entry_for_offset(vim, offset);
983 ASSERT3P(first_mapping, !=, NULL);
985 vdev_indirect_mapping_entry_phys_t *m = first_mapping;
987 uint64_t size = DVA_GET_ASIZE(&m->vimep_dst);
989 ASSERT3U(offset, >=, DVA_MAPPING_GET_SRC_OFFSET(m));
990 ASSERT3U(offset, <, DVA_MAPPING_GET_SRC_OFFSET(m) + size);
992 uint64_t inner_offset = offset - DVA_MAPPING_GET_SRC_OFFSET(m);
993 uint64_t inner_size = MIN(asize, size - inner_offset);
995 offset += inner_size;
1001 size_t copy_length = entries * sizeof (*first_mapping);
1002 duplicate_mappings = kmem_alloc(copy_length, KM_SLEEP);
1003 bcopy(first_mapping, duplicate_mappings, copy_length);
1004 *copied_entries = entries;
1006 return (duplicate_mappings);
1010 * Goes through the relevant indirect mappings until it hits a concrete vdev
1011 * and issues the callback. On the way to the concrete vdev, if any other
1012 * indirect vdevs are encountered, then the callback will also be called on
1013 * each of those indirect vdevs. For example, if the segment is mapped to
1014 * segment A on indirect vdev 1, and then segment A on indirect vdev 1 is
1015 * mapped to segment B on concrete vdev 2, then the callback will be called on
1016 * both vdev 1 and vdev 2.
1018 * While the callback passed to vdev_indirect_remap() is called on every vdev
1019 * the function encounters, certain callbacks only care about concrete vdevs.
1020 * These types of callbacks should return immediately and explicitly when they
1021 * are called on an indirect vdev.
1023 * Because there is a possibility that a DVA section in the indirect device
1024 * has been split into multiple sections in our mapping, we keep track
1025 * of the relevant contiguous segments of the new location (remap_segment_t)
1026 * in a stack. This way we can call the callback for each of the new sections
1027 * created by a single section of the indirect device. Note though, that in
1028 * this scenario the callbacks in each split block won't occur in-order in
1029 * terms of offset, so callers should not make any assumptions about that.
1031 * For callbacks that don't handle split blocks and immediately return when
1032 * they encounter them (as is the case for remap_blkptr_cb), the caller can
1033 * assume that its callback will be applied from the first indirect vdev
1034 * encountered to the last one and then the concrete vdev, in that order.
1037 vdev_indirect_remap(vdev_t *vd, uint64_t offset, uint64_t asize,
1038 void (*func)(uint64_t, vdev_t *, uint64_t, uint64_t, void *), void *arg)
1041 spa_t *spa = vd->vdev_spa;
1043 list_create(&stack, sizeof (remap_segment_t),
1044 offsetof(remap_segment_t, rs_node));
1046 for (remap_segment_t *rs = rs_alloc(vd, offset, asize, 0);
1047 rs != NULL; rs = list_remove_head(&stack)) {
1048 vdev_t *v = rs->rs_vd;
1049 uint64_t num_entries = 0;
1051 ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1052 ASSERT(rs->rs_asize > 0);
1055 * Note: As this function can be called from open context
1056 * (e.g. zio_read()), we need the following rwlock to
1057 * prevent the mapping from being changed by condensing.
1059 * So we grab the lock and we make a copy of the entries
1060 * that are relevant to the extent that we are working on.
1061 * Once that is done, we drop the lock and iterate over
1062 * our copy of the mapping. Once we are done with the with
1063 * the remap segment and we free it, we also free our copy
1064 * of the indirect mapping entries that are relevant to it.
1066 * This way we don't need to wait until the function is
1067 * finished with a segment, to condense it. In addition, we
1068 * don't need a recursive rwlock for the case that a call to
1069 * vdev_indirect_remap() needs to call itself (through the
1070 * codepath of its callback) for the same vdev in the middle
1073 rw_enter(&v->vdev_indirect_rwlock, RW_READER);
1074 vdev_indirect_mapping_t *vim = v->vdev_indirect_mapping;
1075 ASSERT3P(vim, !=, NULL);
1077 vdev_indirect_mapping_entry_phys_t *mapping =
1078 vdev_indirect_mapping_duplicate_adjacent_entries(v,
1079 rs->rs_offset, rs->rs_asize, &num_entries);
1080 ASSERT3P(mapping, !=, NULL);
1081 ASSERT3U(num_entries, >, 0);
1082 rw_exit(&v->vdev_indirect_rwlock);
1084 for (uint64_t i = 0; i < num_entries; i++) {
1086 * Note: the vdev_indirect_mapping can not change
1087 * while we are running. It only changes while the
1088 * removal is in progress, and then only from syncing
1089 * context. While a removal is in progress, this
1090 * function is only called for frees, which also only
1091 * happen from syncing context.
1093 vdev_indirect_mapping_entry_phys_t *m = &mapping[i];
1095 ASSERT3P(m, !=, NULL);
1096 ASSERT3U(rs->rs_asize, >, 0);
1098 uint64_t size = DVA_GET_ASIZE(&m->vimep_dst);
1099 uint64_t dst_offset = DVA_GET_OFFSET(&m->vimep_dst);
1100 uint64_t dst_vdev = DVA_GET_VDEV(&m->vimep_dst);
1102 ASSERT3U(rs->rs_offset, >=,
1103 DVA_MAPPING_GET_SRC_OFFSET(m));
1104 ASSERT3U(rs->rs_offset, <,
1105 DVA_MAPPING_GET_SRC_OFFSET(m) + size);
1106 ASSERT3U(dst_vdev, !=, v->vdev_id);
1108 uint64_t inner_offset = rs->rs_offset -
1109 DVA_MAPPING_GET_SRC_OFFSET(m);
1110 uint64_t inner_size =
1111 MIN(rs->rs_asize, size - inner_offset);
1113 vdev_t *dst_v = vdev_lookup_top(spa, dst_vdev);
1114 ASSERT3P(dst_v, !=, NULL);
1116 if (dst_v->vdev_ops == &vdev_indirect_ops) {
1117 list_insert_head(&stack,
1118 rs_alloc(dst_v, dst_offset + inner_offset,
1119 inner_size, rs->rs_split_offset));
1123 if ((zfs_flags & ZFS_DEBUG_INDIRECT_REMAP) &&
1124 IS_P2ALIGNED(inner_size, 2 * SPA_MINBLOCKSIZE)) {
1126 * Note: This clause exists only solely for
1127 * testing purposes. We use it to ensure that
1128 * split blocks work and that the callbacks
1129 * using them yield the same result if issued
1132 uint64_t inner_half = inner_size / 2;
1134 func(rs->rs_split_offset + inner_half, dst_v,
1135 dst_offset + inner_offset + inner_half,
1138 func(rs->rs_split_offset, dst_v,
1139 dst_offset + inner_offset,
1142 func(rs->rs_split_offset, dst_v,
1143 dst_offset + inner_offset,
1147 rs->rs_offset += inner_size;
1148 rs->rs_asize -= inner_size;
1149 rs->rs_split_offset += inner_size;
1151 VERIFY0(rs->rs_asize);
1153 kmem_free(mapping, num_entries * sizeof (*mapping));
1154 kmem_free(rs, sizeof (remap_segment_t));
1156 list_destroy(&stack);
1160 vdev_indirect_child_io_done(zio_t *zio)
1162 zio_t *pio = zio->io_private;
1164 mutex_enter(&pio->io_lock);
1165 pio->io_error = zio_worst_error(pio->io_error, zio->io_error);
1166 mutex_exit(&pio->io_lock);
1169 if (zio->io_abd != NULL)
1171 abd_put(zio->io_abd);
1175 * This is a callback for vdev_indirect_remap() which allocates an
1176 * indirect_split_t for each split segment and adds it to iv_splits.
1179 vdev_indirect_gather_splits(uint64_t split_offset, vdev_t *vd, uint64_t offset,
1180 uint64_t size, void *arg)
1183 indirect_vsd_t *iv = zio->io_vsd;
1185 ASSERT3P(vd, !=, NULL);
1187 if (vd->vdev_ops == &vdev_indirect_ops)
1191 if (vd->vdev_ops == &vdev_mirror_ops)
1192 n = vd->vdev_children;
1194 indirect_split_t *is =
1195 kmem_zalloc(offsetof(indirect_split_t, is_child[n]), KM_SLEEP);
1197 is->is_children = n;
1199 is->is_split_offset = split_offset;
1200 is->is_target_offset = offset;
1202 list_create(&is->is_unique_child, sizeof (indirect_child_t),
1203 offsetof(indirect_child_t, ic_node));
1206 * Note that we only consider multiple copies of the data for
1207 * *mirror* vdevs. We don't for "replacing" or "spare" vdevs, even
1208 * though they use the same ops as mirror, because there's only one
1209 * "good" copy under the replacing/spare.
1211 if (vd->vdev_ops == &vdev_mirror_ops) {
1212 for (int i = 0; i < n; i++) {
1213 is->is_child[i].ic_vdev = vd->vdev_child[i];
1214 list_link_init(&is->is_child[i].ic_node);
1217 is->is_child[0].ic_vdev = vd;
1220 list_insert_tail(&iv->iv_splits, is);
1224 vdev_indirect_read_split_done(zio_t *zio)
1226 indirect_child_t *ic = zio->io_private;
1228 if (zio->io_error != 0) {
1230 * Clear ic_data to indicate that we do not have data for this
1233 abd_free(ic->ic_data);
1239 * Issue reads for all copies (mirror children) of all splits.
1242 vdev_indirect_read_all(zio_t *zio)
1244 indirect_vsd_t *iv = zio->io_vsd;
1246 ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ);
1248 for (indirect_split_t *is = list_head(&iv->iv_splits);
1249 is != NULL; is = list_next(&iv->iv_splits, is)) {
1250 for (int i = 0; i < is->is_children; i++) {
1251 indirect_child_t *ic = &is->is_child[i];
1253 if (!vdev_readable(ic->ic_vdev))
1257 * Note, we may read from a child whose DTL
1258 * indicates that the data may not be present here.
1259 * While this might result in a few i/os that will
1260 * likely return incorrect data, it simplifies the
1261 * code since we can treat scrub and resilver
1262 * identically. (The incorrect data will be
1263 * detected and ignored when we verify the
1267 ic->ic_data = abd_alloc_sametype(zio->io_abd,
1269 ic->ic_duplicate = NULL;
1271 zio_nowait(zio_vdev_child_io(zio, NULL,
1272 ic->ic_vdev, is->is_target_offset, ic->ic_data,
1273 is->is_size, zio->io_type, zio->io_priority, 0,
1274 vdev_indirect_read_split_done, ic));
1277 iv->iv_reconstruct = B_TRUE;
1281 vdev_indirect_io_start(zio_t *zio)
1283 spa_t *spa = zio->io_spa;
1284 indirect_vsd_t *iv = kmem_zalloc(sizeof (*iv), KM_SLEEP);
1285 list_create(&iv->iv_splits,
1286 sizeof (indirect_split_t), offsetof(indirect_split_t, is_node));
1289 zio->io_vsd_ops = &vdev_indirect_vsd_ops;
1291 ASSERT(spa_config_held(spa, SCL_ALL, RW_READER) != 0);
1293 if (zio->io_type == ZIO_TYPE_WRITE) {
1295 if (zio->io_type != ZIO_TYPE_READ) {
1296 ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE);
1299 * Note: this code can handle other kinds of writes,
1300 * but we don't expect them.
1302 ASSERT((zio->io_flags & (ZIO_FLAG_SELF_HEAL |
1303 ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE)) != 0);
1306 vdev_indirect_remap(zio->io_vd, zio->io_offset, zio->io_size,
1307 vdev_indirect_gather_splits, zio);
1309 indirect_split_t *first = list_head(&iv->iv_splits);
1310 if (first->is_size == zio->io_size) {
1312 * This is not a split block; we are pointing to the entire
1313 * data, which will checksum the same as the original data.
1314 * Pass the BP down so that the child i/o can verify the
1315 * checksum, and try a different location if available
1316 * (e.g. on a mirror).
1318 * While this special case could be handled the same as the
1319 * general (split block) case, doing it this way ensures
1320 * that the vast majority of blocks on indirect vdevs
1321 * (which are not split) are handled identically to blocks
1322 * on non-indirect vdevs. This allows us to be less strict
1323 * about performance in the general (but rare) case.
1325 ASSERT0(first->is_split_offset);
1326 ASSERT3P(list_next(&iv->iv_splits, first), ==, NULL);
1327 zio_nowait(zio_vdev_child_io(zio, zio->io_bp,
1328 first->is_vdev, first->is_target_offset,
1330 zio->io_abd == NULL ? NULL :
1332 abd_get_offset(zio->io_abd, 0),
1333 zio->io_size, zio->io_type, zio->io_priority, 0,
1334 vdev_indirect_child_io_done, zio));
1336 iv->iv_split_block = B_TRUE;
1337 if (zio->io_type == ZIO_TYPE_READ &&
1338 zio->io_flags & (ZIO_FLAG_SCRUB | ZIO_FLAG_RESILVER)) {
1340 * Read all copies. Note that for simplicity,
1341 * we don't bother consulting the DTL in the
1344 vdev_indirect_read_all(zio);
1347 * If this is a read zio, we read one copy of each
1348 * split segment, from the top-level vdev. Since
1349 * we don't know the checksum of each split
1350 * individually, the child zio can't ensure that
1351 * we get the right data. E.g. if it's a mirror,
1352 * it will just read from a random (healthy) leaf
1353 * vdev. We have to verify the checksum in
1354 * vdev_indirect_io_done().
1356 * For write zios, the vdev code will ensure we write
1359 for (indirect_split_t *is = list_head(&iv->iv_splits);
1360 is != NULL; is = list_next(&iv->iv_splits, is)) {
1361 zio_nowait(zio_vdev_child_io(zio, NULL,
1362 is->is_vdev, is->is_target_offset,
1364 zio->io_abd == NULL ? NULL :
1366 abd_get_offset(zio->io_abd,
1367 is->is_split_offset),
1368 is->is_size, zio->io_type,
1369 zio->io_priority, 0,
1370 vdev_indirect_child_io_done, zio));
1379 * Report a checksum error for a child.
1382 vdev_indirect_checksum_error(zio_t *zio,
1383 indirect_split_t *is, indirect_child_t *ic)
1385 vdev_t *vd = ic->ic_vdev;
1387 if (zio->io_flags & ZIO_FLAG_SPECULATIVE)
1390 mutex_enter(&vd->vdev_stat_lock);
1391 vd->vdev_stat.vs_checksum_errors++;
1392 mutex_exit(&vd->vdev_stat_lock);
1394 zio_bad_cksum_t zbc = { 0 };
1395 void *bad_buf = abd_borrow_buf_copy(ic->ic_data, is->is_size);
1396 abd_t *good_abd = is->is_good_child->ic_data;
1397 void *good_buf = abd_borrow_buf_copy(good_abd, is->is_size);
1398 zfs_ereport_post_checksum(zio->io_spa, vd, zio,
1399 is->is_target_offset, is->is_size, good_buf, bad_buf, &zbc);
1400 abd_return_buf(ic->ic_data, bad_buf, is->is_size);
1401 abd_return_buf(good_abd, good_buf, is->is_size);
1405 * Issue repair i/os for any incorrect copies. We do this by comparing
1406 * each split segment's correct data (is_good_child's ic_data) with each
1407 * other copy of the data. If they differ, then we overwrite the bad data
1408 * with the good copy. Note that we do this without regard for the DTL's,
1409 * which simplifies this code and also issues the optimal number of writes
1410 * (based on which copies actually read bad data, as opposed to which we
1411 * think might be wrong). For the same reason, we always use
1412 * ZIO_FLAG_SELF_HEAL, to bypass the DTL check in zio_vdev_io_start().
1415 vdev_indirect_repair(zio_t *zio)
1417 indirect_vsd_t *iv = zio->io_vsd;
1419 enum zio_flag flags = ZIO_FLAG_IO_REPAIR;
1421 if (!(zio->io_flags & (ZIO_FLAG_SCRUB | ZIO_FLAG_RESILVER)))
1422 flags |= ZIO_FLAG_SELF_HEAL;
1424 if (!spa_writeable(zio->io_spa))
1427 for (indirect_split_t *is = list_head(&iv->iv_splits);
1428 is != NULL; is = list_next(&iv->iv_splits, is)) {
1429 for (int c = 0; c < is->is_children; c++) {
1430 indirect_child_t *ic = &is->is_child[c];
1431 if (ic == is->is_good_child)
1433 if (ic->ic_data == NULL)
1435 if (ic->ic_duplicate == is->is_good_child)
1438 zio_nowait(zio_vdev_child_io(zio, NULL,
1439 ic->ic_vdev, is->is_target_offset,
1440 is->is_good_child->ic_data, is->is_size,
1441 ZIO_TYPE_WRITE, ZIO_PRIORITY_ASYNC_WRITE,
1442 ZIO_FLAG_IO_REPAIR | ZIO_FLAG_SELF_HEAL,
1445 vdev_indirect_checksum_error(zio, is, ic);
1451 * Report checksum errors on all children that we read from.
1454 vdev_indirect_all_checksum_errors(zio_t *zio)
1456 indirect_vsd_t *iv = zio->io_vsd;
1458 if (zio->io_flags & ZIO_FLAG_SPECULATIVE)
1461 for (indirect_split_t *is = list_head(&iv->iv_splits);
1462 is != NULL; is = list_next(&iv->iv_splits, is)) {
1463 for (int c = 0; c < is->is_children; c++) {
1464 indirect_child_t *ic = &is->is_child[c];
1466 if (ic->ic_data == NULL)
1469 vdev_t *vd = ic->ic_vdev;
1471 mutex_enter(&vd->vdev_stat_lock);
1472 vd->vdev_stat.vs_checksum_errors++;
1473 mutex_exit(&vd->vdev_stat_lock);
1475 zfs_ereport_post_checksum(zio->io_spa, vd, zio,
1476 is->is_target_offset, is->is_size,
1483 * Copy data from all the splits to a main zio then validate the checksum.
1484 * If then checksum is successfully validated return success.
1487 vdev_indirect_splits_checksum_validate(indirect_vsd_t *iv, zio_t *zio)
1489 zio_bad_cksum_t zbc;
1491 for (indirect_split_t *is = list_head(&iv->iv_splits);
1492 is != NULL; is = list_next(&iv->iv_splits, is)) {
1494 ASSERT3P(is->is_good_child->ic_data, !=, NULL);
1495 ASSERT3P(is->is_good_child->ic_duplicate, ==, NULL);
1497 abd_copy_off(zio->io_abd, is->is_good_child->ic_data,
1498 is->is_split_offset, 0, is->is_size);
1501 return (zio_checksum_error(zio, &zbc));
1505 * There are relatively few possible combinations making it feasible to
1506 * deterministically check them all. We do this by setting the good_child
1507 * to the next unique split version. If we reach the end of the list then
1508 * "carry over" to the next unique split version (like counting in base
1509 * is_unique_children, but each digit can have a different base).
1512 vdev_indirect_splits_enumerate_all(indirect_vsd_t *iv, zio_t *zio)
1514 boolean_t more = B_TRUE;
1516 iv->iv_attempts = 0;
1518 for (indirect_split_t *is = list_head(&iv->iv_splits);
1519 is != NULL; is = list_next(&iv->iv_splits, is))
1520 is->is_good_child = list_head(&is->is_unique_child);
1522 while (more == B_TRUE) {
1526 if (vdev_indirect_splits_checksum_validate(iv, zio) == 0)
1529 for (indirect_split_t *is = list_head(&iv->iv_splits);
1530 is != NULL; is = list_next(&iv->iv_splits, is)) {
1531 is->is_good_child = list_next(&is->is_unique_child,
1533 if (is->is_good_child != NULL) {
1538 is->is_good_child = list_head(&is->is_unique_child);
1542 ASSERT3S(iv->iv_attempts, <=, iv->iv_unique_combinations);
1544 return (SET_ERROR(ECKSUM));
1548 * There are too many combinations to try all of them in a reasonable amount
1549 * of time. So try a fixed number of random combinations from the unique
1550 * split versions, after which we'll consider the block unrecoverable.
1553 vdev_indirect_splits_enumerate_randomly(indirect_vsd_t *iv, zio_t *zio)
1555 iv->iv_attempts = 0;
1557 while (iv->iv_attempts < iv->iv_attempts_max) {
1560 for (indirect_split_t *is = list_head(&iv->iv_splits);
1561 is != NULL; is = list_next(&iv->iv_splits, is)) {
1562 indirect_child_t *ic = list_head(&is->is_unique_child);
1563 int children = is->is_unique_children;
1565 for (int i = spa_get_random(children); i > 0; i--)
1566 ic = list_next(&is->is_unique_child, ic);
1568 ASSERT3P(ic, !=, NULL);
1569 is->is_good_child = ic;
1572 if (vdev_indirect_splits_checksum_validate(iv, zio) == 0)
1576 return (SET_ERROR(ECKSUM));
1580 * This is a validation function for reconstruction. It randomly selects
1581 * a good combination, if one can be found, and then it intentionally
1582 * damages all other segment copes by zeroing them. This forces the
1583 * reconstruction algorithm to locate the one remaining known good copy.
1586 vdev_indirect_splits_damage(indirect_vsd_t *iv, zio_t *zio)
1588 /* Presume all the copies are unique for initial selection. */
1589 for (indirect_split_t *is = list_head(&iv->iv_splits);
1590 is != NULL; is = list_next(&iv->iv_splits, is)) {
1591 is->is_unique_children = 0;
1593 for (int i = 0; i < is->is_children; i++) {
1594 indirect_child_t *ic = &is->is_child[i];
1595 if (ic->ic_data != NULL) {
1596 is->is_unique_children++;
1597 list_insert_tail(&is->is_unique_child, ic);
1603 * Set each is_good_child to a randomly-selected child which
1604 * is known to contain validated data.
1606 int error = vdev_indirect_splits_enumerate_randomly(iv, zio);
1611 * Damage all but the known good copy by zeroing it. This will
1612 * result in two or less unique copies per indirect_child_t.
1613 * Both may need to be checked in order to reconstruct the block.
1614 * Set iv->iv_attempts_max such that all unique combinations will
1615 * enumerated, but limit the damage to at most 16 indirect splits.
1617 iv->iv_attempts_max = 1;
1619 for (indirect_split_t *is = list_head(&iv->iv_splits);
1620 is != NULL; is = list_next(&iv->iv_splits, is)) {
1621 for (int c = 0; c < is->is_children; c++) {
1622 indirect_child_t *ic = &is->is_child[c];
1624 if (ic == is->is_good_child)
1626 if (ic->ic_data == NULL)
1629 abd_zero(ic->ic_data, ic->ic_data->abd_size);
1632 iv->iv_attempts_max *= 2;
1633 if (iv->iv_attempts_max > (1ULL << 16)) {
1634 iv->iv_attempts_max = UINT64_MAX;
1640 /* Empty the unique children lists so they can be reconstructed. */
1641 for (indirect_split_t *is = list_head(&iv->iv_splits);
1642 is != NULL; is = list_next(&iv->iv_splits, is)) {
1643 indirect_child_t *ic;
1644 while ((ic = list_head(&is->is_unique_child)) != NULL)
1645 list_remove(&is->is_unique_child, ic);
1647 is->is_unique_children = 0;
1654 * This function is called when we have read all copies of the data and need
1655 * to try to find a combination of copies that gives us the right checksum.
1657 * If we pointed to any mirror vdevs, this effectively does the job of the
1658 * mirror. The mirror vdev code can't do its own job because we don't know
1659 * the checksum of each split segment individually.
1661 * We have to try every unique combination of copies of split segments, until
1662 * we find one that checksums correctly. Duplicate segment copies are first
1663 * identified and latter skipped during reconstruction. This optimization
1664 * reduces the search space and ensures that of the remaining combinations
1665 * at most one is correct.
1667 * When the total number of combinations is small they can all be checked.
1668 * For example, if we have 3 segments in the split, and each points to a
1669 * 2-way mirror with unique copies, we will have the following pieces of data:
1673 * ======|=====================
1674 * A | data_A_0 data_A_1
1675 * B | data_B_0 data_B_1
1676 * C | data_C_0 data_C_1
1678 * We will try the following (mirror children)^(number of splits) (2^3=8)
1679 * combinations, which is similar to bitwise-little-endian counting in
1680 * binary. In general each "digit" corresponds to a split segment, and the
1681 * base of each digit is is_children, which can be different for each
1684 * "low bit" "high bit"
1686 * data_A_0 data_B_0 data_C_0
1687 * data_A_1 data_B_0 data_C_0
1688 * data_A_0 data_B_1 data_C_0
1689 * data_A_1 data_B_1 data_C_0
1690 * data_A_0 data_B_0 data_C_1
1691 * data_A_1 data_B_0 data_C_1
1692 * data_A_0 data_B_1 data_C_1
1693 * data_A_1 data_B_1 data_C_1
1695 * Note that the split segments may be on the same or different top-level
1696 * vdevs. In either case, we may need to try lots of combinations (see
1697 * zfs_reconstruct_indirect_combinations_max). This ensures that if a mirror
1698 * has small silent errors on all of its children, we can still reconstruct
1699 * the correct data, as long as those errors are at sufficiently-separated
1700 * offsets (specifically, separated by the largest block size - default of
1701 * 128KB, but up to 16MB).
1704 vdev_indirect_reconstruct_io_done(zio_t *zio)
1706 indirect_vsd_t *iv = zio->io_vsd;
1707 boolean_t known_good = B_FALSE;
1710 iv->iv_unique_combinations = 1;
1711 iv->iv_attempts_max = UINT64_MAX;
1713 if (zfs_reconstruct_indirect_combinations_max > 0)
1714 iv->iv_attempts_max = zfs_reconstruct_indirect_combinations_max;
1717 * If nonzero, every 1/x blocks will be damaged, in order to validate
1718 * reconstruction when there are split segments with damaged copies.
1719 * Known_good will TRUE when reconstruction is known to be possible.
1721 if (zfs_reconstruct_indirect_damage_fraction != 0 &&
1722 spa_get_random(zfs_reconstruct_indirect_damage_fraction) == 0)
1723 known_good = (vdev_indirect_splits_damage(iv, zio) == 0);
1726 * Determine the unique children for a split segment and add them
1727 * to the is_unique_child list. By restricting reconstruction
1728 * to these children, only unique combinations will be considered.
1729 * This can vastly reduce the search space when there are a large
1730 * number of indirect splits.
1732 for (indirect_split_t *is = list_head(&iv->iv_splits);
1733 is != NULL; is = list_next(&iv->iv_splits, is)) {
1734 is->is_unique_children = 0;
1736 for (int i = 0; i < is->is_children; i++) {
1737 indirect_child_t *ic_i = &is->is_child[i];
1739 if (ic_i->ic_data == NULL ||
1740 ic_i->ic_duplicate != NULL)
1743 for (int j = i + 1; j < is->is_children; j++) {
1744 indirect_child_t *ic_j = &is->is_child[j];
1746 if (ic_j->ic_data == NULL ||
1747 ic_j->ic_duplicate != NULL)
1750 if (abd_cmp(ic_i->ic_data, ic_j->ic_data,
1751 is->is_size) == 0) {
1752 ic_j->ic_duplicate = ic_i;
1756 is->is_unique_children++;
1757 list_insert_tail(&is->is_unique_child, ic_i);
1760 /* Reconstruction is impossible, no valid children */
1761 EQUIV(list_is_empty(&is->is_unique_child),
1762 is->is_unique_children == 0);
1763 if (list_is_empty(&is->is_unique_child)) {
1764 zio->io_error = EIO;
1765 vdev_indirect_all_checksum_errors(zio);
1766 zio_checksum_verified(zio);
1770 iv->iv_unique_combinations *= is->is_unique_children;
1773 if (iv->iv_unique_combinations <= iv->iv_attempts_max)
1774 error = vdev_indirect_splits_enumerate_all(iv, zio);
1776 error = vdev_indirect_splits_enumerate_randomly(iv, zio);
1779 /* All attempted combinations failed. */
1780 ASSERT3B(known_good, ==, B_FALSE);
1781 zio->io_error = error;
1782 vdev_indirect_all_checksum_errors(zio);
1785 * The checksum has been successfully validated. Issue
1786 * repair I/Os to any copies of splits which don't match
1787 * the validated version.
1789 ASSERT0(vdev_indirect_splits_checksum_validate(iv, zio));
1790 vdev_indirect_repair(zio);
1791 zio_checksum_verified(zio);
1796 vdev_indirect_io_done(zio_t *zio)
1798 indirect_vsd_t *iv = zio->io_vsd;
1800 if (iv->iv_reconstruct) {
1802 * We have read all copies of the data (e.g. from mirrors),
1803 * either because this was a scrub/resilver, or because the
1804 * one-copy read didn't checksum correctly.
1806 vdev_indirect_reconstruct_io_done(zio);
1810 if (!iv->iv_split_block) {
1812 * This was not a split block, so we passed the BP down,
1813 * and the checksum was handled by the (one) child zio.
1818 zio_bad_cksum_t zbc;
1819 int ret = zio_checksum_error(zio, &zbc);
1821 zio_checksum_verified(zio);
1826 * The checksum didn't match. Read all copies of all splits, and
1827 * then we will try to reconstruct. The next time
1828 * vdev_indirect_io_done() is called, iv_reconstruct will be set.
1830 vdev_indirect_read_all(zio);
1832 zio_vdev_io_redone(zio);
1835 vdev_ops_t vdev_indirect_ops = {
1837 vdev_indirect_close,
1839 vdev_indirect_io_start,
1840 vdev_indirect_io_done,
1845 vdev_indirect_remap,
1847 VDEV_TYPE_INDIRECT, /* name of this vdev type */
1848 B_FALSE /* leaf vdev */