Switch allocations from KM_SLEEP to KM_PUSHPAGE

A couple of kmem_alloc() allocations were using KM_SLEEP in
the sync thread context.  These were accidentally introduced
by the recent set of Illumos patches.  The solution is to
switch to KM_PUSHPAGE.

dsl_dataset_promote_sync() -> promote_hold() -> snaplist_make() ->
kmem_alloc(sizeof (*snap), KM_SLEEP);

dsl_dataset_user_hold_sync() -> dsl_onexit_hold_cleanup() ->
kmem_alloc(sizeof (*ca), KM_SLEEP)

Signed-off-by: Richard Yao <ryao@gentoo.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue openzfs#1775

module/zfs/dsl_dataset.c

@@ -2196,7 +2196,7 @@ snaplist_make(dsl_pool_t *dp,
 		if (first_obj == 0)
 			first_obj = ds->ds_dir->dd_phys->dd_origin_obj;
 
-		snap = kmem_alloc(sizeof (*snap), KM_SLEEP);
+		snap = kmem_alloc(sizeof (*snap), KM_PUSHPAGE);
 		snap->ds = ds;
 		list_insert_tail(l, snap);
 		obj = ds->ds_phys->ds_prev_snap_obj;

module/zfs/dsl_userhold.c

@@ -225,7 +225,7 @@ dsl_onexit_hold_cleanup(spa_t *spa, nvlist_t *holds, minor_t minor)
 	}
 
 	ASSERT(spa != NULL);
-	ca = kmem_alloc(sizeof (*ca), KM_SLEEP);
+	ca = kmem_alloc(sizeof (*ca), KM_PUSHPAGE);
 
 	(void) strlcpy(ca->zhca_spaname, spa_name(spa),
 	    sizeof (ca->zhca_spaname));