Re: SCSI EH tidbit

[Jeff Garzik <jgarzik@pobox.com>]

Tejun Heo wrote:
> Jeff Garzik wrote:
>> scsi_finish_command unconditionally completes the command, rather than 
>> running it through scsi_decide_disposition() decision tree again.  
>> That prevents us from using the standard command-retry path, for 
>> PCI/ATA bus errors where we want to resubmit the command.
> 
> 
>  That can be done by finishing with scsi_queue_insert as done by the 
> current SCSI EH implementation.  And, yes, that also can be done by 
> generating appropriate sense data and run it through __scsi_done again.

good point.


>  The thing is that I don't really see much difference between finishing 
> with scsi_queue_insert/scsi_queue_insert and using __scsi_done.  I'm 
> opting for not using __scsi_done mainly because that's how the current 
> SCSI EH implementation is implemented and I don't really think it would 
> be a good idea to do things differently from SCSI EH without clear reason.

I like transport separation, since libata needs to be a bit closer to 
its "pure" form:  an ATA API, which is used by libata-scsi client.


>  The situation is similar when generating sense data.  We can always 
> generate sense data for which scsi_decide_disposition() would never 
> return FAILED.  But IMHO it doesn't look like a very good idea.

SCSI keeps an internal counter to make sure that commands are not 
retried forever.

	Jeff