Refcounted DSL Crypto Key Mappings

Since native ZFS encryption was merged, we have been fighting
against a series of bugs that come down to the same problem: Key
mappings (which must be present during all I/O operations) are
created and destroyed based on dataset ownership, but I/Os can
have traditionally been allowed to "leak" into the next txg after
the dataset is disowned.

In the past we have attempted to solve this problem by trying to
ensure that datasets are disowned ater all I/O is finished by
calling txg_wait_synced(), but we have repeatedly found edge cases
that need to be squashed and code paths that might incur a high
number of txg syncs. This patch attempts to resolve this issue
differently, by adding a reference to the key mapping for each txg
it is dirtied in. By doing so, we can remove many of the
unnecessary calls to txg_wait_synced() we have added in the past
and ensure we don't need to deal with this problem in the future.

Reviewed-by: Jorgen Lundman <[email protected]>
Reviewed by: Matthew Ahrens <[email protected]>
Reviewed-by: Brian Behlendorf <[email protected]>
Signed-off-by: Tom Caputi <[email protected]>
Closes #7949 

6 files changed, 186 insertions, 94 deletions

diff --git a/module/zfs/dmu_objset.c b/module/zfs/dmu_objset.c
index 3c9a817f7..086a65114 100644
--- a/module/zfs/dmu_objset.c
+++ b/module/zfs/dmu_objset.c
@@ -1157,6 +1157,7 @@ dmu_objset_create_sync(void *arg, dmu_tx_t *tx)
 {
 	dmu_objset_create_arg_t *doca = arg;
 	dsl_pool_t *dp = dmu_tx_pool(tx);
+	spa_t *spa = dp->dp_spa;
 	dsl_dir_t *pdd;
 	const char *tail;
 	dsl_dataset_t *ds;
@@ -1174,8 +1175,7 @@ dmu_objset_create_sync(void *arg, dmu_tx_t *tx)
 	    DS_HOLD_FLAG_DECRYPT, FTAG, &ds));
 	rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
 	bp = dsl_dataset_get_blkptr(ds);
-	os = dmu_objset_create_impl(pdd->dd_pool->dp_spa,
-	    ds, bp, doca->doca_type, tx);
+	os = dmu_objset_create_impl(spa, ds, bp, doca->doca_type, tx);
 	rrw_exit(&ds->ds_bp_rwlock, FTAG);
 
 	if (doca->doca_userfunc != NULL) {
@@ -1199,7 +1199,7 @@ dmu_objset_create_sync(void *arg, dmu_tx_t *tx)
 		ds->ds_owner = FTAG;
 		mutex_exit(&ds->ds_lock);
 
-		rzio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
+		rzio = zio_root(spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
 		tmpds = txg_list_remove_this(&dp->dp_dirty_datasets, ds,
 		    tx->tx_txg);
 		if (tmpds != NULL) {
@@ -1210,8 +1210,12 @@ dmu_objset_create_sync(void *arg, dmu_tx_t *tx)
 
 		dmu_objset_do_userquota_updates(os, tx);
 		taskq_wait(dp->dp_sync_taskq);
+		if (txg_list_member(&dp->dp_dirty_datasets, ds, tx->tx_txg)) {
+			ASSERT3P(ds->ds_key_mapping, !=, NULL);
+			key_mapping_rele(spa, ds->ds_key_mapping, ds);
+		}
 
-		rzio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
+		rzio = zio_root(spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
 		tmpds = txg_list_remove_this(&dp->dp_dirty_datasets, ds,
 		    tx->tx_txg);
 		if (tmpds != NULL) {
@@ -1220,8 +1224,11 @@ dmu_objset_create_sync(void *arg, dmu_tx_t *tx)
 		}
 		VERIFY0(zio_wait(rzio));
 
-		if (need_sync_done)
+		if (need_sync_done) {
+			ASSERT3P(ds->ds_key_mapping, !=, NULL);
+			key_mapping_rele(spa, ds->ds_key_mapping, ds);
 			dsl_dataset_sync_done(ds, tx);
+		}
 
 		mutex_enter(&ds->ds_lock);
 		ds->ds_owner = NULL;
@@ -1229,7 +1236,7 @@ dmu_objset_create_sync(void *arg, dmu_tx_t *tx)
 	}
 
 	spa_history_log_internal_ds(ds, "create", tx, "");
-	zvol_create_minors(dp->dp_spa, doca->doca_name, B_TRUE);
+	zvol_create_minors(spa, doca->doca_name, B_TRUE);
 
 	dsl_dataset_rele_flags(ds, DS_HOLD_FLAG_DECRYPT, FTAG);
 	dsl_dir_rele(pdd, FTAG);
@@ -1702,7 +1709,6 @@ dmu_objset_sync(objset_t *os, zio_t *pio, dmu_tx_t *tx)
 	if (os->os_raw_receive ||
 	    os->os_next_write_raw[tx->tx_txg & TXG_MASK]) {
 		ASSERT(os->os_encrypted);
-		os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_FALSE;
 		arc_convert_to_raw(os->os_phys_buf,
 		    os->os_dsl_dataset->ds_object, ZFS_HOST_BYTEORDER,
 		    DMU_OT_OBJSET, NULL, NULL, NULL);
diff --git a/module/zfs/dsl_crypt.c b/module/zfs/dsl_crypt.c
index 6937fc9e1..0eeb260c3 100644
--- a/module/zfs/dsl_crypt.c
+++ b/module/zfs/dsl_crypt.c
@@ -896,6 +896,20 @@ spa_keystore_unload_wkey(const char *dsname)
 	int ret = 0;
 	dsl_dir_t *dd = NULL;
 	dsl_pool_t *dp = NULL;
+	spa_t *spa = NULL;
+
+	ret = spa_open(dsname, &spa, FTAG);
+	if (ret != 0)
+		return (ret);
+
+	/*
+	 * Wait for any outstanding txg IO to complete, releasing any
+	 * remaining references on the wkey.
+	 */
+	if (spa_mode(spa) != FREAD)
+		txg_wait_synced(spa->spa_dsl_pool, 0);
+
+	spa_close(spa, FTAG);
 
 	/* hold the dsl dir */
 	ret = dsl_pool_hold(dsname, FTAG, &dp);
@@ -935,9 +949,56 @@ error:
 	return (ret);
 }
 
+void
+key_mapping_add_ref(dsl_key_mapping_t *km, void *tag)
+{
+	ASSERT3U(zfs_refcount_count(&km->km_refcnt), >=, 1);
+	zfs_refcount_add(&km->km_refcnt, tag);
+}
+
+/*
+ * The locking here is a little tricky to ensure we don't cause unnecessary
+ * performance problems. We want to release a key mapping whenever someone
+ * decrements the refcount to 0, but freeing the mapping requires removing
+ * it from the spa_keystore, which requires holding sk_km_lock as a writer.
+ * Most of the time we don't want to hold this lock as a writer, since the
+ * same lock is held as a reader for each IO that needs to encrypt / decrypt
+ * data for any dataset and in practice we will only actually free the
+ * mapping after unmounting a dataset.
+ */
+void
+key_mapping_rele(spa_t *spa, dsl_key_mapping_t *km, void *tag)
+{
+	ASSERT3U(zfs_refcount_count(&km->km_refcnt), >=, 1);
+
+	if (zfs_refcount_remove(&km->km_refcnt, tag) != 0)
+		return;
+
+	/*
+	 * We think we are going to need to free the mapping. Add a
+	 * reference to prevent most other releasers from thinking
+	 * this might be their responsibility. This is inherently
+	 * racy, so we will confirm that we are legitimately the
+	 * last holder once we have the sk_km_lock as a writer.
+	 */
+	zfs_refcount_add(&km->km_refcnt, FTAG);
+
+	rw_enter(&spa->spa_keystore.sk_km_lock, RW_WRITER);
+	if (zfs_refcount_remove(&km->km_refcnt, FTAG) != 0) {
+		rw_exit(&spa->spa_keystore.sk_km_lock);
+		return;
+	}
+
+	avl_remove(&spa->spa_keystore.sk_key_mappings, km);
+	rw_exit(&spa->spa_keystore.sk_km_lock);
+
+	spa_keystore_dsl_key_rele(spa, km->km_key, km);
+	kmem_free(km, sizeof (dsl_key_mapping_t));
+}
+
 int
-spa_keystore_create_mapping_impl(spa_t *spa, uint64_t dsobj,
-    dsl_dir_t *dd, void *tag)
+spa_keystore_create_mapping(spa_t *spa, dsl_dataset_t *ds, void *tag,
+    dsl_key_mapping_t **km_out)
 {
 	int ret;
 	avl_index_t where;
@@ -948,14 +1009,17 @@ spa_keystore_create_mapping_impl(spa_t *spa, uint64_t dsobj,
 	km = kmem_zalloc(sizeof (dsl_key_mapping_t), KM_SLEEP);
 	zfs_refcount_create(&km->km_refcnt);
 
-	ret = spa_keystore_dsl_key_hold_dd(spa, dd, km, &km->km_key);
+	ret = spa_keystore_dsl_key_hold_dd(spa, ds->ds_dir, km, &km->km_key);
 	if (ret != 0) {
 		zfs_refcount_destroy(&km->km_refcnt);
 		kmem_free(km, sizeof (dsl_key_mapping_t));
+
+		if (km_out != NULL)
+			*km_out = NULL;
 		return (ret);
 	}
 
-	km->km_dsobj = dsobj;
+	km->km_dsobj = ds->ds_object;
 
 	rw_enter(&spa->spa_keystore.sk_km_lock, RW_WRITER);
 
@@ -971,9 +1035,13 @@ spa_keystore_create_mapping_impl(spa_t *spa, uint64_t dsobj,
 	if (found_km != NULL) {
 		should_free = B_TRUE;
 		zfs_refcount_add(&found_km->km_refcnt, tag);
+		if (km_out != NULL)
+			*km_out = found_km;
 	} else {
 		zfs_refcount_add(&km->km_refcnt, tag);
 		avl_insert(&spa->spa_keystore.sk_key_mappings, km, where);
+		if (km_out != NULL)
+			*km_out = km;
 	}
 
 	rw_exit(&spa->spa_keystore.sk_km_lock);
@@ -988,24 +1056,16 @@ spa_keystore_create_mapping_impl(spa_t *spa, uint64_t dsobj,
 }
 
 int
-spa_keystore_create_mapping(spa_t *spa, dsl_dataset_t *ds, void *tag)
-{
-	return (spa_keystore_create_mapping_impl(spa, ds->ds_object,
-	    ds->ds_dir, tag));
-}
-
-int
 spa_keystore_remove_mapping(spa_t *spa, uint64_t dsobj, void *tag)
 {
 	int ret;
 	dsl_key_mapping_t search_km;
 	dsl_key_mapping_t *found_km;
-	boolean_t should_free = B_FALSE;
 
 	/* init the search key mapping */
 	search_km.km_dsobj = dsobj;
 
-	rw_enter(&spa->spa_keystore.sk_km_lock, RW_WRITER);
+	rw_enter(&spa->spa_keystore.sk_km_lock, RW_READER);
 
 	/* find the matching mapping */
 	found_km = avl_find(&spa->spa_keystore.sk_key_mappings,
@@ -1015,23 +1075,9 @@ spa_keystore_remove_mapping(spa_t *spa, uint64_t dsobj, void *tag)
 		goto error_unlock;
 	}
 
-	/*
-	 * Decrement the refcount on the mapping and remove it from the tree if
-	 * it is zero. Try to minimize time spent in this lock by deferring
-	 * cleanup work.
-	 */
-	if (zfs_refcount_remove(&found_km->km_refcnt, tag) == 0) {
-		should_free = B_TRUE;
-		avl_remove(&spa->spa_keystore.sk_key_mappings, found_km);
-	}
-
 	rw_exit(&spa->spa_keystore.sk_km_lock);
 
-	/* destroy the key mapping */
-	if (should_free) {
-		spa_keystore_dsl_key_rele(spa, found_km->km_key, found_km);
-		kmem_free(found_km, sizeof (dsl_key_mapping_t));
-	}
+	key_mapping_rele(spa, found_km, tag);
 
 	return (0);
 
diff --git a/module/zfs/dsl_dataset.c b/module/zfs/dsl_dataset.c
index 36715b8a2..757b2922e 100644
--- a/module/zfs/dsl_dataset.c
+++ b/module/zfs/dsl_dataset.c
@@ -438,8 +438,8 @@ dsl_dataset_try_add_ref(dsl_pool_t *dp, dsl_dataset_t *ds, void *tag)
 }
 
 int
-dsl_dataset_hold_obj_flags(dsl_pool_t *dp, uint64_t dsobj,
-    ds_hold_flags_t flags, void *tag, dsl_dataset_t **dsp)
+dsl_dataset_hold_obj(dsl_pool_t *dp, uint64_t dsobj, void *tag,
+    dsl_dataset_t **dsp)
 {
 	objset_t *mos = dp->dp_meta_objset;
 	dmu_buf_t *dbuf;
@@ -599,6 +599,7 @@ dsl_dataset_hold_obj_flags(dsl_pool_t *dp, uint64_t dsobj,
 			}
 		}
 	}
+
 	ASSERT3P(ds->ds_dbuf, ==, dbuf);
 	ASSERT3P(dsl_dataset_phys(ds), ==, dbuf->db_data);
 	ASSERT(dsl_dataset_phys(ds)->ds_prev_snap_obj != 0 ||
@@ -606,22 +607,40 @@ dsl_dataset_hold_obj_flags(dsl_pool_t *dp, uint64_t dsobj,
 	    dp->dp_origin_snap == NULL || ds == dp->dp_origin_snap);
 	*dsp = ds;
 
-	if ((flags & DS_HOLD_FLAG_DECRYPT) && ds->ds_dir->dd_crypto_obj != 0) {
-		err = spa_keystore_create_mapping(dp->dp_spa, ds, ds);
-		if (err != 0) {
-			dsl_dataset_rele(ds, tag);
-			return (SET_ERROR(EACCES));
-		}
-	}
-
 	return (0);
 }
 
 int
-dsl_dataset_hold_obj(dsl_pool_t *dp, uint64_t dsobj, void *tag,
-    dsl_dataset_t **dsp)
+dsl_dataset_create_key_mapping(dsl_dataset_t *ds)
 {
-	return (dsl_dataset_hold_obj_flags(dp, dsobj, 0, tag, dsp));
+	dsl_dir_t *dd = ds->ds_dir;
+
+	if (dd->dd_crypto_obj == 0)
+		return (0);
+
+	return (spa_keystore_create_mapping(dd->dd_pool->dp_spa,
+	    ds, ds, &ds->ds_key_mapping));
+}
+
+int
+dsl_dataset_hold_obj_flags(dsl_pool_t *dp, uint64_t dsobj,
+    ds_hold_flags_t flags, void *tag, dsl_dataset_t **dsp)
+{
+	int err;
+
+	err = dsl_dataset_hold_obj(dp, dsobj, tag, dsp);
+	if (err != 0)
+		return (err);
+
+	ASSERT3P(*dsp, !=, NULL);
+
+	if (flags & DS_HOLD_FLAG_DECRYPT) {
+		err = dsl_dataset_create_key_mapping(*dsp);
+		if (err != 0)
+			dsl_dataset_rele(*dsp, tag);
+	}
+
+	return (err);
 }
 
 int
@@ -788,21 +807,30 @@ dsl_dataset_namelen(dsl_dataset_t *ds)
 }
 
 void
-dsl_dataset_rele_flags(dsl_dataset_t *ds, ds_hold_flags_t flags, void *tag)
+dsl_dataset_rele(dsl_dataset_t *ds, void *tag)
 {
-	if (ds->ds_dir != NULL && ds->ds_dir->dd_crypto_obj != 0 &&
-	    (flags & DS_HOLD_FLAG_DECRYPT)) {
-		(void) spa_keystore_remove_mapping(ds->ds_dir->dd_pool->dp_spa,
-		    ds->ds_object, ds);
-	}
-
 	dmu_buf_rele(ds->ds_dbuf, tag);
 }
 
 void
-dsl_dataset_rele(dsl_dataset_t *ds, void *tag)
+dsl_dataset_remove_key_mapping(dsl_dataset_t *ds)
+{
+	dsl_dir_t *dd = ds->ds_dir;
+
+	if (dd == NULL || dd->dd_crypto_obj == 0)
+		return;
+
+	(void) spa_keystore_remove_mapping(dd->dd_pool->dp_spa,
+	    ds->ds_object, ds);
+}
+
+void
+dsl_dataset_rele_flags(dsl_dataset_t *ds, ds_hold_flags_t flags, void *tag)
 {
-	dsl_dataset_rele_flags(ds, 0, tag);
+	if (flags & DS_HOLD_FLAG_DECRYPT)
+		dsl_dataset_remove_key_mapping(ds);
+
+	dsl_dataset_rele(ds, tag);
 }
 
 void
@@ -1154,8 +1182,18 @@ dsl_dataset_dirty(dsl_dataset_t *ds, dmu_tx_t *tx)
 
 	dp = ds->ds_dir->dd_pool;
 	if (txg_list_add(&dp->dp_dirty_datasets, ds, tx->tx_txg)) {
+		objset_t *os = ds->ds_objset;
+
 		/* up the hold count until we can be written out */
 		dmu_buf_add_ref(ds->ds_dbuf, ds);
+
+		/* if this dataset is encrypted, grab a reference to the DCK */
+		if (ds->ds_dir->dd_crypto_obj != 0 &&
+		    !os->os_raw_receive &&
+		    !os->os_next_write_raw[tx->tx_txg & TXG_MASK]) {
+			ASSERT3P(ds->ds_key_mapping, !=, NULL);
+			key_mapping_add_ref(ds->ds_key_mapping, ds);
+		}
 	}
 }
 
@@ -1800,6 +1838,11 @@ dsl_dataset_sync_done(dsl_dataset_t *ds, dmu_tx_t *tx)
 		os->os_synced_dnodes = NULL;
 	}
 
+	if (os->os_encrypted)
+		os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_FALSE;
+	else
+		ASSERT0(os->os_next_write_raw[tx->tx_txg & TXG_MASK]);
+
 	ASSERT(!dmu_objset_is_dirty(os, dmu_tx_get_txg(tx)));
 
 	dmu_buf_rele(ds->ds_dbuf, ds);
diff --git a/module/zfs/dsl_pool.c b/module/zfs/dsl_pool.c
index e8f519b18..bd2055710 100644
--- a/module/zfs/dsl_pool.c
+++ b/module/zfs/dsl_pool.c
@@ -516,7 +516,8 @@ dsl_pool_create(spa_t *spa, nvlist_t *zplprops, dsl_crypto_params_t *dcp,
 	obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, dcp, 0, tx);
 
 	/* create the root objset */
-	VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
+	VERIFY0(dsl_dataset_hold_obj_flags(dp, obj,
+	    DS_HOLD_FLAG_DECRYPT, FTAG, &ds));
 #ifdef _KERNEL
 	{
 		objset_t *os;
@@ -527,7 +528,7 @@ dsl_pool_create(spa_t *spa, nvlist_t *zplprops, dsl_crypto_params_t *dcp,
 		zfs_create_fs(os, kcred, zplprops, tx);
 	}
 #endif
-	dsl_dataset_rele(ds, FTAG);
+	dsl_dataset_rele_flags(ds, DS_HOLD_FLAG_DECRYPT, FTAG);
 
 	dmu_tx_commit(tx);
 
@@ -690,9 +691,22 @@ dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
 	 */
 	zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
 	while ((ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) != NULL) {
+		objset_t *os = ds->ds_objset;
+
 		ASSERT(list_link_active(&ds->ds_synced_link));
 		dmu_buf_rele(ds->ds_dbuf, ds);
 		dsl_dataset_sync(ds, zio, tx);
+
+		/*
+		 * Release any key mappings created by calls to
+		 * dsl_dataset_dirty() from the userquota accounting
+		 * code paths.
+		 */
+		if (os->os_encrypted && !os->os_raw_receive &&
+		    !os->os_next_write_raw[txg & TXG_MASK]) {
+			ASSERT3P(ds->ds_key_mapping, !=, NULL);
+			key_mapping_rele(dp->dp_spa, ds->ds_key_mapping, ds);
+		}
 	}
 	VERIFY0(zio_wait(zio));
 
@@ -702,8 +716,17 @@ dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
 	 *
 	 *  - move dead blocks from the pending deadlist to the on-disk deadlist
 	 *  - release hold from dsl_dataset_dirty()
+	 *  - release key mapping hold from dsl_dataset_dirty()
 	 */
 	while ((ds = list_remove_head(&synced_datasets)) != NULL) {
+		objset_t *os = ds->ds_objset;
+
+		if (os->os_encrypted && !os->os_raw_receive &&
+		    !os->os_next_write_raw[txg & TXG_MASK]) {
+			ASSERT3P(ds->ds_key_mapping, !=, NULL);
+			key_mapping_rele(dp->dp_spa, ds->ds_key_mapping, ds);
+		}
+
 		dsl_dataset_sync_done(ds, tx);
 	}
 
diff --git a/module/zfs/spa.c b/module/zfs/spa.c
index ad8000914..a1851bca2 100644
--- a/module/zfs/spa.c
+++ b/module/zfs/spa.c
@@ -4995,7 +4995,7 @@ spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props,
 	uint64_t txg = TXG_INITIAL;
 	nvlist_t **spares, **l2cache;
 	uint_t nspares, nl2cache;
-	uint64_t version, obj, root_dsobj = 0;
+	uint64_t version, obj;
 	boolean_t has_features;
 	boolean_t has_encryption;
 	spa_feature_t feat;
@@ -5249,27 +5249,11 @@ spa_create(const char *pool, nvlist_t *nvroot, nvlist_t *props,
 
 	dmu_tx_commit(tx);
 
-	/*
-	 * If the root dataset is encrypted we will need to create key mappings
-	 * for the zio layer before we start to write any data to disk and hold
-	 * them until after the first txg has been synced. Waiting for the first
-	 * transaction to complete also ensures that our bean counters are
-	 * appropriately updated.
-	 */
-	if (dp->dp_root_dir->dd_crypto_obj != 0) {
-		root_dsobj = dsl_dir_phys(dp->dp_root_dir)->dd_head_dataset_obj;
-		VERIFY0(spa_keystore_create_mapping_impl(spa, root_dsobj,
-		    dp->dp_root_dir, FTAG));
-	}
-
 	spa->spa_sync_on = B_TRUE;
 	txg_sync_start(dp);
 	mmp_thread_start(spa);
 	txg_wait_synced(dp, txg);
 
-	if (dp->dp_root_dir->dd_crypto_obj != 0)
-		VERIFY0(spa_keystore_remove_mapping(spa, root_dsobj, FTAG));
-
 	spa_spawn_aux_threads(spa);
 
 	spa_write_cachefile(spa, B_FALSE, B_TRUE);
diff --git a/module/zfs/zil.c b/module/zfs/zil.c
index 8b7aeb5c3..e247c1e4f 100644
--- a/module/zfs/zil.c
+++ b/module/zfs/zil.c
@@ -3236,8 +3236,8 @@ zil_suspend(const char *osname, void **cookiep)
 	 * grabbing a reference to it. If the key isn't loaded we have no
 	 * choice but to return an error until the wrapping key is loaded.
 	 */
-	if (os->os_encrypted && spa_keystore_create_mapping(os->os_spa,
-	    dmu_objset_ds(os), FTAG) != 0) {
+	if (os->os_encrypted &&
+	    dsl_dataset_create_key_mapping(dmu_objset_ds(os)) != 0) {
 		zilog->zl_suspend--;
 		mutex_exit(&zilog->zl_lock);
 		dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
@@ -3259,9 +3259,10 @@ zil_suspend(const char *osname, void **cookiep)
 	zil_commit_impl(zilog, 0);
 
 	/*
-	 * Now that we've ensured all lwb's are LWB_STATE_DONE, we use
-	 * txg_wait_synced() to ensure the data from the zilog has
-	 * migrated to the main pool before calling zil_destroy().
+	 * Now that we've ensured all lwb's are LWB_STATE_DONE,
+	 * txg_wait_synced() will be called from within zil_destroy(),
+	 * which will ensure the data from the zilog has migrated to the
+	 * main pool before it returns.
 	 */
 	txg_wait_synced(zilog->zl_dmu_pool, 0);
 
@@ -3272,19 +3273,8 @@ zil_suspend(const char *osname, void **cookiep)
 	cv_broadcast(&zilog->zl_cv_suspend);
 	mutex_exit(&zilog->zl_lock);
 
-	if (os->os_encrypted) {
-		/*
-		 * Encrypted datasets need to wait for all data to be
-		 * synced out before removing the mapping.
-		 *
-		 * XXX: Depending on the number of datasets with
-		 * outstanding ZIL data on a given log device, this
-		 * might cause spa_offline_log() to take a long time.
-		 */
-		txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
-		VERIFY0(spa_keystore_remove_mapping(os->os_spa,
-		    dmu_objset_id(os), FTAG));
-	}
+	if (os->os_encrypted)
+		dsl_dataset_remove_key_mapping(dmu_objset_ds(os));
 
 	if (cookiep == NULL)
 		zil_resume(os);