Douglas Gilbert wrote:
> After the comments on the RFC for this thread, here is
> a patch against lk 2.6.9-rc1 to kick around.
> 
> The patch only touches two files: scsi.h and scsi_lib.c
> It adds the proposed facility and then uses it in scsi_lib
> in roughly 4 locations. IMO there were 3 sense processing errors:
>    - block SG_IO did not get passed back deferred errors
>      [SG_IO is a __pass-through__ interface!!]
>    - MEDIUM_ERRORs _do_ get processed for deferred sense errors
>      in scsi_io_completion() which seems unintended
>      [I did not fix this one.]
>    - invalid command operation code handling in
>      __scsi_mode_sense() was just wrong
> 
> If people think this is a reasonable approach, then the rest of
> the scsi mid-level and upper level driver could be converted.
> As Kai pointed we may need some general routines to pick up the
> sense data "extras".
> 
> The benefit of doing this conversion is that it may well
> highlight a lot more sense data processing errors (if the
> above is any guide).
> 
> Changes:
>     - add structure to receive normalized sense data from either
>       fixed or descriptor format sense data
>     - add scsi_normalize_sense() function to populate that structure
>     - add scsi_sense_is_deferred() function so deferred errors can
>       be ignored in many contexts
>     - apply new mechanism to sense data processing within the
>       scsi_lib.c file

Here is a second cut of the patch.

If James accepts it naturally he is free to enforce the
appropriate coding style.

I believe that cleaning up scsi error processing is going
to be a "work in progress" for a while.

Doug Gilbert