Re: [PATCH] Always pass result and sense on request completion

[Boaz Harrosh <bharrosh@panasas.com>]

On 12/10/2009 06:44 PM, James Bottomley wrote:
> On Thu, 2009-12-10 at 10:49 +0100, Hannes Reinecke wrote:
>> Hi James,
>>
>> would you mind commenting on this patch?
>> We really need this if we ever want to be able to do proper error code
>> handling from multipath.
> 
> OK, not keen on the way you're setting req->errors.
> 
> Our big problem with FS requests is deferred or corrected errors
> (basically we never want the FS to think we had a problem with them).  I
> think it's currently OK because block tends to believe the returned
> error code rather than req->errors ... I'd just hate it if that changed
> and suddenly lots of stuff broke.
> 

From what I understood, req->errors is for private block-driver use and has
no meaning to the block layer. It expects a translation of req->errors to
a Linux error code passed to the blk_end_request which will be set to the
bio and passed to the async_done function. Only the block-driver understand
the format of req->errors.

Perhaps we must make sure there is an agreement between
	(returned-error == 0) == (req->errors == 0)

I know scsi-ml makes sure of that, so should the device manager.

> I think you're just looking for the sense data, so could you look at
> that and not set req->errors?
> 

I agree that the req->errors bits where not understood outside of scsi
up till now. Is multipath only compatible over scsi block devices?

> James
> 

Boaz