On Wed, Feb 23, 2022 at 12:16:24PM -0800, Bart Van Assche wrote: > On 2/23/22 04:58, Christoph Hellwig wrote: >> In general most of the zeroing here should go away entirely. Right >> now that code is a bit convolute due to the fake EH scsi command that >> also uses scsi_init_command. Once that is sorted out (I've just pinged >> Hannes for his series) scsi_init_command can be folded into >> scsi_prepare_cmd and a lot more of this can be cleaned up. > > Several SCSI LLDs rely on the SCSI core zeroing the driver-private data. > I'd be more than happy if the code for zeroing of driver-private data would > be pushed into the LLD drivers. This may require the introduction of a new > per-command flag since that zeroing should only happen if scsi_queue_rq() > decided to call scsi_prepare_cmd(). If the driver provides a init_cmd_priv method it is not zeroed. So if a driver does not want it zeroed, it can always provide this callback, including an empty one.
