Narrow a race, and fix a leak, in g_part_wither

A race in g_part_wither() can lead to I/O being performed with a freed GEOM
when the device disappears.  Close the race as best as we can for now,
following the code patterns from g_part_ctl_destroy() and g_part_ctl_undo().
This also fixes a leak, as g_wither_geom() does not wither providers, it
only orphans them, so the partition entries would never get destroyed in
g_wither_washer().

Note, this is not a complete fix, it can still race with g_part_start(), the
race has merely been narrowed.

Reviewed by:	markj
Sponsored by:	Dell EMC Isilon

1 files changed, 8 insertions, 5 deletions

diff --git a/sys/geom/part/g_part.c b/sys/geom/part/g_part.c
index ae76e10c6578..976ec197fe1c 100644
--- a/sys/geom/part/g_part.c
+++ b/sys/geom/part/g_part.c
@@ -1541,18 +1541,21 @@ g_part_wither(struct g_geom *gp, int error)
 {
 	struct g_part_entry *entry;
 	struct g_part_table *table;
+	struct g_provider *pp;
 
 	table = gp->softc;
 	if (table != NULL) {
-		G_PART_DESTROY(table, NULL);
+		gp->softc = NULL;
 		while ((entry = LIST_FIRST(&table->gpt_entry)) != NULL) {
 			LIST_REMOVE(entry, gpe_entry);
+			pp = entry->gpe_pp;
+			entry->gpe_pp->private = NULL;
+			entry->gpe_pp = NULL;
+			g_wither_provider(pp, error);
 			g_free(entry);
 		}
-		if (gp->softc != NULL) {
-			kobj_delete((kobj_t)gp->softc, M_GEOM);
-			gp->softc = NULL;
-		}
+		G_PART_DESTROY(table, NULL);
+		kobj_delete((kobj_t)table, M_GEOM);
 	}
 	g_wither_geom(gp, error);
 }