When an IO fails then zio_vdev_io_done() can call vdev_probe() to determine the health of the vdev. This is safe as long as the original zio was submitted with zio_wait() and holds the SCL_STATE_ALL lock over the operation. If zio_no_wait() was used then the done callback will submit the probe IO outside the SCL_STATE_ALL lock and hit this ASSERT in zio_create() ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); Resolve the issue by only allowing vdev_probe() to be called when there's a waiter indicating the caller is using zio_wait(). This assumes that caller is still holding SCL_STATE_ALL. This issue isn't MMP specific but was surfaced when testing. Without this patch it can be reproduced by running: zpool set multihost on <pool> zinject -d <vdev> -e io -T write -f 50 <pool> -L uber Reviewed-by: Olaf Faaland <faaland1@llnl.gov> Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Don Brady <don.brady@intel.com> Closes #745 Closes #6279