Re: [PATCH RFC 1/3] SCSI Userspace Target: scsi-ml changes

[Mike Christie <michaelc@cs.wisc.edu>]

James Bottomley wrote:
> On Mon, 2006-02-06 at 11:54 -0600, Mike Christie wrote:
> 
>>We actually need ths shost pointer becuase we do not represent the LU in 
>>the kernel. So we go from:
>>
>>scsi_host -> host's uspace request_queue (used to pass commands and 
>>messages from the target LLD interrup handler to our netlink 
>>code/userspace -> in userspace we represent the target, device/LU, and host.
>>
>>Tomo did you add a virtual scsi_device somewhere?
> 
> 
> OK, well, the problem is that struct scsi_cmnd is allocated in the
> critical I/O path.  The more elements it has, the more time it takes to
> set up and tear down.  Kenneth Chen of intel has already produced
> figures showing that you can get measurable I/O improvement simply by
> dropping fields from it, so I'm very, very reluctant to increase its
> size.
> 

I understand. To make this worse, for the target code since we are 
mapping data like SG_IO block layer code (we used their functions) we 
need process context for the unmap. To handle this we added a work 
struct on the scsi_cmnd and the work struct is pretty large and only 
used for targets. We can move all the target specific to a:

scsi_tgt_cmnd {
	work;
	host;
};

And store out scsi_tgt_cmnd in the scsi_cmnd->host_scribble or somewhere 
on the SCp. Is that too hacky?