Fast Clone Deletion

Deleting a clone requires finding blocks are clone-only, not shared
with the snapshot. This was done by traversing the entire block tree
which results in a large performance penalty for sparsely
written clones.

This is new method keeps track of clone blocks when they are
modified in a "Livelist" so that, when it’s time to delete,
the clone-specific blocks are already at hand.

We see performance improvements because now deletion work is
proportional to the number of clone-modified blocks, not the size
of the original dataset.

Reviewed-by: Sean Eric Fagan <[email protected]>
Reviewed-by: Matt Ahrens <[email protected]>
Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Serapheim Dimitropoulos <[email protected]>
Signed-off-by: Sara Hartse <[email protected]>
Closes #8416 

1 files changed, 30 insertions, 0 deletions

diff --git a/module/zfs/dbuf.c b/module/zfs/dbuf.c
index 4d347b6f4..0518205f9 100644
--- a/module/zfs/dbuf.c
+++ b/module/zfs/dbuf.c
@@ -3286,6 +3286,13 @@ dbuf_hold_impl_arg(struct dbuf_hold_arg *dh)
 
 	*(dh->dh_dbp) = NULL;
 
+	/* If the pool has been created, verify the tx_sync_lock is not held */
+	spa_t *spa = dh->dh_dn->dn_objset->os_spa;
+	dsl_pool_t *dp = spa->spa_dsl_pool;
+	if (dp != NULL) {
+		ASSERT(!MUTEX_HELD(&dp->dp_tx.tx_sync_lock));
+	}
+
 	/* dbuf_find() returns with db_mtx held */
 	dh->dh_db = dbuf_find(dh->dh_dn->dn_objset, dh->dh_dn->dn_object,
 	    dh->dh_level, dh->dh_blkid);
@@ -4480,6 +4487,29 @@ dbuf_remap_impl(dnode_t *dn, blkptr_t *bp, krwlock_t *rw, dmu_tx_t *tx)
 	if (spa_remap_blkptr(spa, &bp_copy, dbuf_remap_impl_callback,
 	    &drica)) {
 		/*
+		 * If the blkptr being remapped is tracked by a livelist,
+		 * then we need to make sure the livelist reflects the update.
+		 * First, cancel out the old blkptr by appending a 'FREE'
+		 * entry. Next, add an 'ALLOC' to track the new version. This
+		 * way we avoid trying to free an inaccurate blkptr at delete.
+		 * Note that embedded blkptrs are not tracked in livelists.
+		 */
+		if (dn->dn_objset != spa_meta_objset(spa)) {
+			dsl_dataset_t *ds = dmu_objset_ds(dn->dn_objset);
+			if (dsl_deadlist_is_open(&ds->ds_dir->dd_livelist) &&
+			    bp->blk_birth > ds->ds_dir->dd_origin_txg) {
+				ASSERT(!BP_IS_EMBEDDED(bp));
+				ASSERT(dsl_dir_is_clone(ds->ds_dir));
+				ASSERT(spa_feature_is_enabled(spa,
+				    SPA_FEATURE_LIVELIST));
+				bplist_append(&ds->ds_dir->dd_pending_frees,
+				    bp);
+				bplist_append(&ds->ds_dir->dd_pending_allocs,
+				    &bp_copy);
+			}
+		}
+
+		/*
 		 * The db_rwlock prevents dbuf_read_impl() from
 		 * dereferencing the BP while we are changing it.  To
 		 * avoid lock contention, only grab it when we are actually