xref: /illumos-gate/usr/src/uts/common/fs/zfs/vdev_indirect.c

/*
 * CDDL HEADER START
 *
 * This file and its contents are supplied under the terms of the
 * Common Development and Distribution License ("CDDL"), version 1.0.
 * You may only use this file in accordance with the terms of version
 * 1.0 of the CDDL.
 *
 * A full copy of the text of the CDDL should have accompanied this
 * source.  A copy of the CDDL is also available via the Internet at
 * http://www.illumos.org/license/CDDL.
 *
 * CDDL HEADER END
 */

/*
 * Copyright (c) 2014, 2019 by Delphix. All rights reserved.
 */

#include <sys/zfs_context.h>
#include <sys/spa.h>
#include <sys/spa_impl.h>
#include <sys/vdev_impl.h>
#include <sys/fs/zfs.h>
#include <sys/zio.h>
#include <sys/zio_checksum.h>
#include <sys/metaslab.h>
#include <sys/refcount.h>
#include <sys/dmu.h>
#include <sys/vdev_indirect_mapping.h>
#include <sys/dmu_tx.h>
#include <sys/dsl_synctask.h>
#include <sys/zap.h>
#include <sys/abd.h>
#include <sys/zthr.h>

/*
 * An indirect vdev corresponds to a vdev that has been removed.  Since
 * we cannot rewrite block pointers of snapshots, etc., we keep a
 * mapping from old location on the removed device to the new location
 * on another device in the pool and use this mapping whenever we need
 * to access the DVA.  Unfortunately, this mapping did not respect
 * logical block boundaries when it was first created, and so a DVA on
 * this indirect vdev may be "split" into multiple sections that each
 * map to a different location.  As a consequence, not all DVAs can be
 * translated to an equivalent new DVA.  Instead we must provide a
 * "vdev_remap" operation that executes a callback on each contiguous
 * segment of the new location.  This function is used in multiple ways:
 *
 *  - i/os to this vdev use the callback to determine where the
 *    data is now located, and issue child i/os for each segment's new
 *    location.
 *
 *  - frees and claims to this vdev use the callback to free or claim
 *    each mapped segment.  (Note that we don't actually need to claim
 *    log blocks on indirect vdevs, because we don't allocate to
 *    removing vdevs.  However, zdb uses zio_claim() for its leak
 *    detection.)
 */

/*
 * "Big theory statement" for how we mark blocks obsolete.
 *
 * When a block on an indirect vdev is freed or remapped, a section of
 * that vdev's mapping may no longer be referenced (aka "obsolete").  We
 * keep track of how much of each mapping entry is obsolete.  When
 * an entry becomes completely obsolete, we can remove it, thus reducing
 * the memory used by the mapping.  The complete picture of obsolescence
 * is given by the following data structures, described below:
 *  - the entry-specific obsolete count
 *  - the vdev-specific obsolete spacemap
 *  - the pool-specific obsolete bpobj
 *
 * == On disk data structures used ==
 *
 * We track the obsolete space for the pool using several objects.  Each
 * of these objects is created on demand and freed when no longer
 * needed, and is assumed to be empty if it does not exist.
 * SPA_FEATURE_OBSOLETE_COUNTS includes the count of these objects.
 *
 *  - Each vic_mapping_object (associated with an indirect vdev) can
 *    have a vimp_counts_object.  This is an array of uint32_t's
 *    with the same number of entries as the vic_mapping_object.  When
 *    the mapping is condensed, entries from the vic_obsolete_sm_object
 *    (see below) are folded into the counts.  Therefore, each
 *    obsolete_counts entry tells us the number of bytes in the
 *    corresponding mapping entry that were not referenced when the
 *    mapping was last condensed.
 *
 *  - Each indirect or removing vdev can have a vic_obsolete_sm_object.
 *    This is a space map containing an alloc entry for every DVA that
 *    has been obsoleted since the last time this indirect vdev was
 *    condensed.  We use this object in order to improve performance
 *    when marking a DVA as obsolete.  Instead of modifying an arbitrary
 *    offset of the vimp_counts_object, we only need to append an entry
 *    to the end of this object.  When a DVA becomes obsolete, it is
 *    added to the obsolete space map.  This happens when the DVA is
 *    freed, remapped and not referenced by a snapshot, or the last
 *    snapshot referencing it is destroyed.
 *
 *  - Each dataset can have a ds_remap_deadlist object.  This is a
 *    deadlist object containing all blocks that were remapped in this
 *    dataset but referenced in a previous snapshot.  Blocks can *only*
 *    appear on this list if they were remapped (dsl_dataset_block_remapped);
 *    blocks that were killed in a head dataset are put on the normal
 *    ds_deadlist and marked obsolete when they are freed.
 *
 *  - The pool can have a dp_obsolete_bpobj.  This is a list of blocks
 *    in the pool that need to be marked obsolete.  When a snapshot is
 *    destroyed, we move some of the ds_remap_deadlist to the obsolete
 *    bpobj (see dsl_destroy_snapshot_handle_remaps()).  We then
 *    asynchronously process the obsolete bpobj, moving its entries to
 *    the specific vdevs' obsolete space maps.
 *
 * == Summary of how we mark blocks as obsolete ==
 *
 * - When freeing a block: if any DVA is on an indirect vdev, append to
 *   vic_obsolete_sm_object.
 * - When remapping a block, add dva to ds_remap_deadlist (if prev snap
 *   references; otherwise append to vic_obsolete_sm_object).
 * - When freeing a snapshot: move parts of ds_remap_deadlist to
 *   dp_obsolete_bpobj (same algorithm as ds_deadlist).
 * - When syncing the spa: process dp_obsolete_bpobj, moving ranges to
 *   individual vdev's vic_obsolete_sm_object.
 */

/*
 * "Big theory statement" for how we condense indirect vdevs.
 *
 * Condensing an indirect vdev's mapping is the process of determining
 * the precise counts of obsolete space for each mapping entry (by
 * integrating the obsolete spacemap into the obsolete counts) and
 * writing out a new mapping that contains only referenced entries.
 *
 * We condense a vdev when we expect the mapping to shrink (see
 * vdev_indirect_should_condense()), but only perform one condense at a
 * time to limit the memory usage.  In addition, we use a separate
 * open-context thread (spa_condense_indirect_thread) to incrementally
 * create the new mapping object in a way that minimizes the impact on
 * the rest of the system.
 *
 * == Generating a new mapping ==
 *
 * To generate a new mapping, we follow these steps:
 *
 * 1. Save the old obsolete space map and create a new mapping object
 *    (see spa_condense_indirect_start_sync()).  This initializes the
 *    spa_condensing_indirect_phys with the "previous obsolete space map",
 *    which is now read only.  Newly obsolete DVAs will be added to a
 *    new (initially empty) obsolete space map, and will not be
 *    considered as part of this condense operation.
 *
 * 2. Construct in memory the precise counts of obsolete space for each
 *    mapping entry, by incorporating the obsolete space map into the
 *    counts.  (See vdev_indirect_mapping_load_obsolete_{counts,spacemap}().)
 *
 * 3. Iterate through each mapping entry, writing to the new mapping any
 *    entries that are not completely obsolete (i.e. which don't have
 *    obsolete count == mapping length).  (See
 *    spa_condense_indirect_generate_new_mapping().)
 *
 * 4. Destroy the old mapping object and switch over to the new one
 *    (spa_condense_indirect_complete_sync).
 *
 * == Restarting from failure ==
 *
 * To restart the condense when we import/open the pool, we must start
 * at the 2nd step above: reconstruct the precise counts in memory,
 * based on the space map + counts.  Then in the 3rd step, we start
 * iterating where we left off: at vimp_max_offset of the new mapping
 * object.
 */

boolean_t zfs_condense_indirect_vdevs_enable = B_TRUE;

/*
 * Condense if at least this percent of the bytes in the mapping is
 * obsolete.  With the default of 25%, the amount of space mapped
 * will be reduced to 1% of its original size after at most 16
 * condenses.  Higher values will condense less often (causing less
 * i/o); lower values will reduce the mapping size more quickly.
 */
int zfs_indirect_condense_obsolete_pct = 25;

/*
 * Condense if the obsolete space map takes up more than this amount of
 * space on disk (logically).  This limits the amount of disk space
 * consumed by the obsolete space map; the default of 1GB is small enough
 * that we typically don't mind "wasting" it.
 */
uint64_t zfs_condense_max_obsolete_bytes = 1024 * 1024 * 1024;

/*
 * Don't bother condensing if the mapping uses less than this amount of
 * memory.  The default of 128KB is considered a "trivial" amount of
 * memory and not worth reducing.
 */
uint64_t zfs_condense_min_mapping_bytes = 128 * 1024;

/*
 * This is used by the test suite so that it can ensure that certain
 * actions happen while in the middle of a condense (which might otherwise
 * complete too quickly).  If used to reduce the performance impact of
 * condensing in production, a maximum value of 1 should be sufficient.
 */
int zfs_condense_indirect_commit_entry_delay_ticks = 0;

/*
 * If an indirect split block contains more than this many possible unique
 * combinations when being reconstructed, consider it too computationally
 * expensive to check them all. Instead, try at most 100 randomly-selected
 * combinations each time the block is accessed.  This allows all segment
 * copies to participate fairly in the reconstruction when all combinations
 * cannot be checked and prevents repeated use of one bad copy.
 */
int zfs_reconstruct_indirect_combinations_max = 256;


/*
 * Enable to simulate damaged segments and validate reconstruction.
 * Used by ztest
 */
unsigned long zfs_reconstruct_indirect_damage_fraction = 0;

/*
 * The indirect_child_t represents the vdev that we will read from, when we
 * need to read all copies of the data (e.g. for scrub or reconstruction).
 * For plain (non-mirror) top-level vdevs (i.e. is_vdev is not a mirror),
 * ic_vdev is the same as is_vdev.  However, for mirror top-level vdevs,
 * ic_vdev is a child of the mirror.
 */
typedef struct indirect_child {
	abd_t *ic_data;
	vdev_t *ic_vdev;

	/*
	 * ic_duplicate is NULL when the ic_data contents are unique, when it
	 * is determined to be a duplicate it references the primary child.
	 */
	struct indirect_child *ic_duplicate;
	list_node_t ic_node; /* node on is_unique_child */
} indirect_child_t;

/*
 * The indirect_split_t represents one mapped segment of an i/o to the
 * indirect vdev. For non-split (contiguously-mapped) blocks, there will be
 * only one indirect_split_t, with is_split_offset==0 and is_size==io_size.
 * For split blocks, there will be several of these.
 */
typedef struct indirect_split {
	list_node_t is_node; /* link on iv_splits */

	/*
	 * is_split_offset is the offset into the i/o.
	 * This is the sum of the previous splits' is_size's.
	 */
	uint64_t is_split_offset;

	vdev_t *is_vdev; /* top-level vdev */
	uint64_t is_target_offset; /* offset on is_vdev */
	uint64_t is_size;
	int is_children; /* number of entries in is_child[] */
	int is_unique_children; /* number of entries in is_unique_child */
	list_t is_unique_child;

	/*
	 * is_good_child is the child that we are currently using to
	 * attempt reconstruction.
	 */
	indirect_child_t *is_good_child;

	indirect_child_t is_child[1]; /* variable-length */
} indirect_split_t;

/*
 * The indirect_vsd_t is associated with each i/o to the indirect vdev.
 * It is the "Vdev-Specific Data" in the zio_t's io_vsd.
 */
typedef struct indirect_vsd {
	boolean_t iv_split_block;
	boolean_t iv_reconstruct;
	uint64_t iv_unique_combinations;
	uint64_t iv_attempts;
	uint64_t iv_attempts_max;

	list_t iv_splits; /* list of indirect_split_t's */
} indirect_vsd_t;

static void
vdev_indirect_map_free(zio_t *zio)
{
	indirect_vsd_t *iv = zio->io_vsd;

	indirect_split_t *is;
	while ((is = list_head(&iv->iv_splits)) != NULL) {
		for (int c = 0; c < is->is_children; c++) {
			indirect_child_t *ic = &is->is_child[c];
			if (ic->ic_data != NULL)
				abd_free(ic->ic_data);
		}
		list_remove(&iv->iv_splits, is);

		indirect_child_t *ic;
		while ((ic = list_head(&is->is_unique_child)) != NULL)
			list_remove(&is->is_unique_child, ic);

		list_destroy(&is->is_unique_child);

		kmem_free(is,
		    offsetof(indirect_split_t, is_child[is->is_children]));
	}
	kmem_free(iv, sizeof (*iv));
}

static const zio_vsd_ops_t vdev_indirect_vsd_ops = {
	vdev_indirect_map_free,
	zio_vsd_default_cksum_report
};
/*
 * Mark the given offset and size as being obsolete.
 */
void
vdev_indirect_mark_obsolete(vdev_t *vd, uint64_t offset, uint64_t size)
{
	spa_t *spa = vd->vdev_spa;

	ASSERT3U(vd->vdev_indirect_config.vic_mapping_object, !=, 0);
	ASSERT(vd->vdev_removing || vd->vdev_ops == &vdev_indirect_ops);
	ASSERT(size > 0);
	VERIFY(vdev_indirect_mapping_entry_for_offset(
	    vd->vdev_indirect_mapping, offset) != NULL);

	if (spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS)) {
		mutex_enter(&vd->vdev_obsolete_lock);
		range_tree_add(vd->vdev_obsolete_segments, offset, size);
		mutex_exit(&vd->vdev_obsolete_lock);
		vdev_dirty(vd, 0, NULL, spa_syncing_txg(spa));
	}
}

/*
 * Mark the DVA vdev_id:offset:size as being obsolete in the given tx. This
 * wrapper is provided because the DMU does not know about vdev_t's and
 * cannot directly call vdev_indirect_mark_obsolete.
 */
void
spa_vdev_indirect_mark_obsolete(spa_t *spa, uint64_t vdev_id, uint64_t offset,
    uint64_t size, dmu_tx_t *tx)
{
	vdev_t *vd = vdev_lookup_top(spa, vdev_id);
	ASSERT(dmu_tx_is_syncing(tx));

	/* The DMU can only remap indirect vdevs. */
	ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
	vdev_indirect_mark_obsolete(vd, offset, size);
}

static spa_condensing_indirect_t *
spa_condensing_indirect_create(spa_t *spa)
{
	spa_condensing_indirect_phys_t *scip =
	    &spa->spa_condensing_indirect_phys;
	spa_condensing_indirect_t *sci = kmem_zalloc(sizeof (*sci), KM_SLEEP);
	objset_t *mos = spa->spa_meta_objset;

	for (int i = 0; i < TXG_SIZE; i++) {
		list_create(&sci->sci_new_mapping_entries[i],
		    sizeof (vdev_indirect_mapping_entry_t),
		    offsetof(vdev_indirect_mapping_entry_t, vime_node));
	}

	sci->sci_new_mapping =
	    vdev_indirect_mapping_open(mos, scip->scip_next_mapping_object);

	return (sci);
}

static void
spa_condensing_indirect_destroy(spa_condensing_indirect_t *sci)
{
	for (int i = 0; i < TXG_SIZE; i++)
		list_destroy(&sci->sci_new_mapping_entries[i]);

	if (sci->sci_new_mapping != NULL)
		vdev_indirect_mapping_close(sci->sci_new_mapping);

	kmem_free(sci, sizeof (*sci));
}

boolean_t
vdev_indirect_should_condense(vdev_t *vd)
{
	vdev_indirect_mapping_t *vim = vd->vdev_indirect_mapping;
	spa_t *spa = vd->vdev_spa;

	ASSERT(dsl_pool_sync_context(spa->spa_dsl_pool));

	if (!zfs_condense_indirect_vdevs_enable)
		return (B_FALSE);

	/*
	 * We can only condense one indirect vdev at a time.
	 */
	if (spa->spa_condensing_indirect != NULL)
		return (B_FALSE);

	if (spa_shutting_down(spa))
		return (B_FALSE);

	/*
	 * The mapping object size must not change while we are
	 * condensing, so we can only condense indirect vdevs
	 * (not vdevs that are still in the middle of being removed).
	 */
	if (vd->vdev_ops != &vdev_indirect_ops)
		return (B_FALSE);

	/*
	 * If nothing new has been marked obsolete, there is no
	 * point in condensing.
	 */
	if (vd->vdev_obsolete_sm == NULL) {
		ASSERT0(vdev_obsolete_sm_object(vd));
		return (B_FALSE);
	}

	ASSERT(vd->vdev_obsolete_sm != NULL);

	ASSERT3U(vdev_obsolete_sm_object(vd), ==,
	    space_map_object(vd->vdev_obsolete_sm));

	uint64_t bytes_mapped = vdev_indirect_mapping_bytes_mapped(vim);
	uint64_t bytes_obsolete = space_map_allocated(vd->vdev_obsolete_sm);
	uint64_t mapping_size = vdev_indirect_mapping_size(vim);
	uint64_t obsolete_sm_size = space_map_length(vd->vdev_obsolete_sm);

	ASSERT3U(bytes_obsolete, <=, bytes_mapped);

	/*
	 * If a high percentage of the bytes that are mapped have become
	 * obsolete, condense (unless the mapping is already small enough).
	 * This has a good chance of reducing the amount of memory used
	 * by the mapping.
	 */
	if (bytes_obsolete * 100 / bytes_mapped >=
	    zfs_indirect_condense_obsolete_pct &&
	    mapping_size > zfs_condense_min_mapping_bytes) {
		zfs_dbgmsg("should condense vdev %llu because obsolete "
		    "spacemap covers %d%% of %lluMB mapping",
		    (u_longlong_t)vd->vdev_id,
		    (int)(bytes_obsolete * 100 / bytes_mapped),
		    (u_longlong_t)bytes_mapped / 1024 / 1024);
		return (B_TRUE);
	}

	/*
	 * If the obsolete space map takes up too much space on disk,
	 * condense in order to free up this disk space.
	 */
	if (obsolete_sm_size >= zfs_condense_max_obsolete_bytes) {
		zfs_dbgmsg("should condense vdev %llu because obsolete sm "
		    "length %lluMB >= max size %lluMB",
		    (u_longlong_t)vd->vdev_id,
		    (u_longlong_t)obsolete_sm_size / 1024 / 1024,
		    (u_longlong_t)zfs_condense_max_obsolete_bytes /
		    1024 / 1024);
		return (B_TRUE);
	}

	return (B_FALSE);
}

/*
 * This sync task completes (finishes) a condense, deleting the old
 * mapping and replacing it with the new one.
 */
static void
spa_condense_indirect_complete_sync(void *arg, dmu_tx_t *tx)
{
	spa_condensing_indirect_t *sci = arg;
	spa_t *spa = dmu_tx_pool(tx)->dp_spa;
	spa_condensing_indirect_phys_t *scip =
	    &spa->spa_condensing_indirect_phys;
	vdev_t *vd = vdev_lookup_top(spa, scip->scip_vdev);
	vdev_indirect_config_t *vic = &vd->vdev_indirect_config;
	objset_t *mos = spa->spa_meta_objset;
	vdev_indirect_mapping_t *old_mapping = vd->vdev_indirect_mapping;
	uint64_t old_count = vdev_indirect_mapping_num_entries(old_mapping);
	uint64_t new_count =
	    vdev_indirect_mapping_num_entries(sci->sci_new_mapping);

	ASSERT(dmu_tx_is_syncing(tx));
	ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
	ASSERT3P(sci, ==, spa->spa_condensing_indirect);
	for (int i = 0; i < TXG_SIZE; i++) {
		ASSERT(list_is_empty(&sci->sci_new_mapping_entries[i]));
	}
	ASSERT(vic->vic_mapping_object != 0);
	ASSERT3U(vd->vdev_id, ==, scip->scip_vdev);
	ASSERT(scip->scip_next_mapping_object != 0);
	ASSERT(scip->scip_prev_obsolete_sm_object != 0);

	/*
	 * Reset vdev_indirect_mapping to refer to the new object.
	 */
	rw_enter(&vd->vdev_indirect_rwlock, RW_WRITER);
	vdev_indirect_mapping_close(vd->vdev_indirect_mapping);
	vd->vdev_indirect_mapping = sci->sci_new_mapping;
	rw_exit(&vd->vdev_indirect_rwlock);

	sci->sci_new_mapping = NULL;
	vdev_indirect_mapping_free(mos, vic->vic_mapping_object, tx);
	vic->vic_mapping_object = scip->scip_next_mapping_object;
	scip->scip_next_mapping_object = 0;

	space_map_free_obj(mos, scip->scip_prev_obsolete_sm_object, tx);
	spa_feature_decr(spa, SPA_FEATURE_OBSOLETE_COUNTS, tx);
	scip->scip_prev_obsolete_sm_object = 0;

	scip->scip_vdev = 0;

	VERIFY0(zap_remove(mos, DMU_POOL_DIRECTORY_OBJECT,
	    DMU_POOL_CONDENSING_INDIRECT, tx));
	spa_condensing_indirect_destroy(spa->spa_condensing_indirect);
	spa->spa_condensing_indirect = NULL;

	zfs_dbgmsg("finished condense of vdev %llu in txg %llu: "
	    "new mapping object %llu has %llu entries "
	    "(was %llu entries)",
	    vd->vdev_id, dmu_tx_get_txg(tx), vic->vic_mapping_object,
	    new_count, old_count);

	vdev_config_dirty(spa->spa_root_vdev);
}

/*
 * This sync task appends entries to the new mapping object.
 */
static void
spa_condense_indirect_commit_sync(void *arg, dmu_tx_t *tx)
{
	spa_condensing_indirect_t *sci = arg;
	uint64_t txg = dmu_tx_get_txg(tx);
	spa_t *spa = dmu_tx_pool(tx)->dp_spa;

	ASSERT(dmu_tx_is_syncing(tx));
	ASSERT3P(sci, ==, spa->spa_condensing_indirect);

	vdev_indirect_mapping_add_entries(sci->sci_new_mapping,
	    &sci->sci_new_mapping_entries[txg & TXG_MASK], tx);
	ASSERT(list_is_empty(&sci->sci_new_mapping_entries[txg & TXG_MASK]));
}

/*
 * Open-context function to add one entry to the new mapping.  The new
 * entry will be remembered and written from syncing context.
 */
static void
spa_condense_indirect_commit_entry(spa_t *spa,
    vdev_indirect_mapping_entry_phys_t *vimep, uint32_t count)
{
	spa_condensing_indirect_t *sci = spa->spa_condensing_indirect;

	ASSERT3U(count, <, DVA_GET_ASIZE(&vimep->vimep_dst));

	dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
	dmu_tx_hold_space(tx, sizeof (*vimep) + sizeof (count));
	VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
	int txgoff = dmu_tx_get_txg(tx) & TXG_MASK;

	/*
	 * If we are the first entry committed this txg, kick off the sync
	 * task to write to the MOS on our behalf.
	 */
	if (list_is_empty(&sci->sci_new_mapping_entries[txgoff])) {
		dsl_sync_task_nowait(dmu_tx_pool(tx),
		    spa_condense_indirect_commit_sync, sci,
		    0, ZFS_SPACE_CHECK_NONE, tx);
	}

	vdev_indirect_mapping_entry_t *vime =
	    kmem_alloc(sizeof (*vime), KM_SLEEP);
	vime->vime_mapping = *vimep;
	vime->vime_obsolete_count = count;
	list_insert_tail(&sci->sci_new_mapping_entries[txgoff], vime);

	dmu_tx_commit(tx);
}

static void
spa_condense_indirect_generate_new_mapping(vdev_t *vd,
    uint32_t *obsolete_counts, uint64_t start_index, zthr_t *zthr)
{
	spa_t *spa = vd->vdev_spa;
	uint64_t mapi = start_index;
	vdev_indirect_mapping_t *old_mapping = vd->vdev_indirect_mapping;
	uint64_t old_num_entries =
	    vdev_indirect_mapping_num_entries(old_mapping);

	ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);
	ASSERT3U(vd->vdev_id, ==, spa->spa_condensing_indirect_phys.scip_vdev);

	zfs_dbgmsg("starting condense of vdev %llu from index %llu",
	    (u_longlong_t)vd->vdev_id,
	    (u_longlong_t)mapi);

	while (mapi < old_num_entries) {

		if (zthr_iscancelled(zthr)) {
			zfs_dbgmsg("pausing condense of vdev %llu "
			    "at index %llu", (u_longlong_t)vd->vdev_id,
			    (u_longlong_t)mapi);
			break;
		}

		vdev_indirect_mapping_entry_phys_t *entry =
		    &old_mapping->vim_entries[mapi];
		uint64_t entry_size = DVA_GET_ASIZE(&entry->vimep_dst);
		ASSERT3U(obsolete_counts[mapi], <=, entry_size);
		if (obsolete_counts[mapi] < entry_size) {
			spa_condense_indirect_commit_entry(spa, entry,
			    obsolete_counts[mapi]);

			/*
			 * This delay may be requested for testing, debugging,
			 * or performance reasons.
			 */
			delay(zfs_condense_indirect_commit_entry_delay_ticks);
		}

		mapi++;
	}
}

/* ARGSUSED */
static boolean_t
spa_condense_indirect_thread_check(void *arg, zthr_t *zthr)
{
	spa_t *spa = arg;

	return (spa->spa_condensing_indirect != NULL);
}

/* ARGSUSED */
static void
spa_condense_indirect_thread(void *arg, zthr_t *zthr)
{
	spa_t *spa = arg;
	vdev_t *vd;

	ASSERT3P(spa->spa_condensing_indirect, !=, NULL);
	spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
	vd = vdev_lookup_top(spa, spa->spa_condensing_indirect_phys.scip_vdev);
	ASSERT3P(vd, !=, NULL);
	spa_config_exit(spa, SCL_VDEV, FTAG);

	spa_condensing_indirect_t *sci = spa->spa_condensing_indirect;
	spa_condensing_indirect_phys_t *scip =
	    &spa->spa_condensing_indirect_phys;
	uint32_t *counts;
	uint64_t start_index;
	vdev_indirect_mapping_t *old_mapping = vd->vdev_indirect_mapping;
	space_map_t *prev_obsolete_sm = NULL;

	ASSERT3U(vd->vdev_id, ==, scip->scip_vdev);
	ASSERT(scip->scip_next_mapping_object != 0);
	ASSERT(scip->scip_prev_obsolete_sm_object != 0);
	ASSERT3P(vd->vdev_ops, ==, &vdev_indirect_ops);

	for (int i = 0; i < TXG_SIZE; i++) {
		/*
		 * The list must start out empty in order for the
		 * _commit_sync() sync task to be properly registered
		 * on the first call to _commit_entry(); so it's wise
		 * to double check and ensure we actually are starting
		 * with empty lists.
		 */
		ASSERT(list_is_empty(&sci->sci_new_mapping_entries[i]));
	}

	VERIFY0(space_map_open(&prev_obsolete_sm, spa->spa_meta_objset,
	    scip->scip_prev_obsolete_sm_object, 0, vd->vdev_asize, 0));
	counts = vdev_indirect_mapping_load_obsolete_counts(old_mapping);
	if (prev_obsolete_sm != NULL) {
		vdev_indirect_mapping_load_obsolete_spacemap(old_mapping,
		    counts, prev_obsolete_sm);