Re: [PATCH 3/3] tcm ibmvscsis driver

["Nicholas A. Bellinger" <nab@linux-iscsi.org>]

On Mon, 2011-02-14 at 18:29 +0900, FUJITA Tomonori wrote:
> On Mon, 14 Feb 2011 01:01:12 -0800
> "Nicholas A. Bellinger" <nab@linux-iscsi.org> wrote:
> 
> > > btw, can we kill the non scatter/gather data path? I think that we
> > > should always use the scatter/gather data transfer.
> > 
> > Unfortuately it's not that easy.  The main reason why CDB type
> > SCF_SCSI_CONTROL_NONSG_IO_CDB was originally added (back in 2.2/2.4
> > days) was because certain LLDs had a problem with basic control CDBs
> > using SGLs..
> > 
> > Obviously we are way past that point with drivers/scsi today, but the
> > main reason today why SCF_SCSI_CONTROL_NONSG_IO_CDB still exists is
> > because of CDB emulation for complex stuff in target_core_cdb.c.  It has
> > historically proven much easier to code complex CDB emulation using a
> > contigious buffer, than with walking SGL formatted memory.
> > 
> > Converting over the more complex CDB emulation stuff to SGLs would
> > somewhat painful, at least without adding an extra location allocation +
> > copy into SGLs (not a big deal for CONTROL CDB stuff),
> 
> lib/scatterlist.c provides nice helper functions to copy data between
> a linear buffer and an SG list. I think that at least, it's more clear
> than now.
> 

Yeah, I think this is going to make the most sense for a proper long
term conversion and removal of SCF_SCSI_CONTROL_NONSG_IO_CDB.

> 
> > or something else
> > to obtain a virtually contigious buffer for building the emulated
> > response.
> 
> Hmm, I don't think that very large contiguous buffer for CONTROL CDB
> so why can't we allocate a physically contiguous buffer for it?
> 

Well, that depends what you consider large..  ;)

> We could play with vmap for a virtually contiguous buffer but it adds
> the complicated cache issue so I don't like that.

Yes, most likely overkill I think..

Best Regards,

--nab