On Wed, 2017-05-03 at 09:54 +0200, Jan Kara wrote: > Hum, since this didn't quite work out, how about storing that one bit of > information that scsi_exit_rq() needs from shost inside scsi_cmnd during > scsi_init_rq()? Hello Jan, All what's missing from the patch I posted is a execute_in_process_context() scsi_host_dev_release() call execute in a context where sleeping is allowed. What you proposed is something I had considered but that I had not yet tried to implement because it requires more changes. Anyway, I'll give that approach a try since it does not require to introduce a new work_struct. Bart.