Re: [PATCH] scsi: sd: add runtime pm to open / release

[Alan Stern]

On Sun, Aug 09, 2020 at 11:20:22AM +0200, Martin Kepplinger wrote:
> Hey Alan, I'm really glad for that, I suspected some of this but I have
> little experience in scsi/block layers, so that is super helpful.
> 
> I'd appreciate an opinion on the below workaround that *seems* to work
> now (let's see, I had thought so before :)
> 
> Whether or not this helps to find a real solution, let's see. But
> integration of such a flag in the error handling paths is what's
> interesting for now:
> 
> 
> --- a/drivers/scsi/scsi_error.c
> +++ b/drivers/scsi/scsi_error.c
> @@ -565,6 +565,17 @@ int scsi_check_sense(struct scsi_cmnd *scmd)
>  				return NEEDS_RETRY;
>  			}
>  		}
> +		if (scmd->device->expecting_media_change) {
> +			if (sshdr.asc == 0x28 && sshdr.ascq == 0x00) {
> +				/* clear expecting_media_change in
> +				 * scsi_noretry_cmd() because we need
> +				 * to override possible "failfast" overrides
> +				 * that block readahead can cause.
> +				 */
> +				return NEEDS_RETRY;

This is a somewhat fragile approach.  You don't know for certain that 
scsi_noretry_cmd will be called.  Also, scsi_noretry_cmd can be called 
from other places.

It would be better to clear the expecting_media_change flag just before 
returning from scsi_decide_disposition.  That way its use is localized 
to one routine, not spread out between two.

Alan Stern