Suspend/resume zvol for recv and rollback

When doing recv and rollback, dsl_dataset_clone_swap_sync_impl will be
called to swap out the ds_objset and do dmu_objset_evict on the old one.
However, currently zv->zv_objset will not be swapped out accordingly, so
if anyone currently holds a fd on the zvol, we risk hitting a use-after-free.

We fix this by introducing the suspend and resume mechanism of zsb to
zv.  Before recv or rollback, we use zvol_suspend to block all access to
zv_objset and shut it down. After the recv or rollback, we use zvol_resume
to swap in zv_objset with the new ds_objset and unblock the access.

Reviewed-by: Brian Behlendorf <[email protected]>
Signed-off-by: Chunwei Chen <[email protected]>
Closes #4866 
Closes #5609 

1 files changed, 2 insertions, 3 deletions

diff --git a/module/zfs/dsl_dataset.c b/module/zfs/dsl_dataset.c
index 43fecf28b..ddab3eddb 100644
--- a/module/zfs/dsl_dataset.c
+++ b/module/zfs/dsl_dataset.c
@@ -2124,8 +2124,7 @@ dsl_dataset_rename_snapshot(const char *fsname,
  * only one long hold on the dataset.  We're not allowed to change anything here
  * so we don't permanently release the long hold or regular hold here.  We want
  * to do this only when syncing to avoid the dataset unexpectedly going away
- * when we release the long hold.  Allow a long hold to exist for volumes, this
- * may occur when asynchronously registering the minor with the kernel.
+ * when we release the long hold.
  */
 static int
 dsl_dataset_handoff_check(dsl_dataset_t *ds, void *owner, dmu_tx_t *tx)
@@ -2140,7 +2139,7 @@ dsl_dataset_handoff_check(dsl_dataset_t *ds, void *owner, dmu_tx_t *tx)
 		dsl_dataset_long_rele(ds, owner);
 	}
 
-	held = (dsl_dataset_long_held(ds) && (ds->ds_owner != zvol_tag));
+	held = dsl_dataset_long_held(ds);
 
 	if (owner != NULL)
 		dsl_dataset_long_hold(ds, owner);