Re: [PATCH] ATAPI error handling work

[Jeff Garzik <jgarzik@pobox.com>]

Tejun Heo wrote:
>  As I wrote in the libata EH doc and associated thread, I don't think 
> doing autosensing-by-turning-qc-around is a very good idea.  I hope I 
> can make some point here.

I like autosensing, particularly because we want to get away from using 
a lot of SCSI EH infrastructure, since libata will eventually be 
separated from SCSI.


>  In the other thread, you wrote that the current ATAPI sense requisition 
> is corrupt because it finishes scmd twice and I believe this patch was 
> implemented to fix that.  I disagreed with you in that thread and gave 
> my rationale for that.  If I was wrong in that thread, please ignore 
> what follows.

I correct myself:  the corruption occurs because we obtain the scsi host 
lock (our host_set lock), and then call scsi_finish_command().  This is 
in error, because scsi_finish_command() also acquires the scsi host lock 
via a call to scsi_device_unbusy().

Multiple possible solutions:  separate scsi host lock from our own. 
batching commands until the end of EH, as SCSI EH does.  continue work 
on auto-sensing.  ...

	Jeff