Re: [PATCH 2/8] block: fix SG_IO vector request data length handling

[Jens Axboe <jens.axboe@oracle.com>]

On Wed, Apr 01 2009, Tejun Heo wrote:
> FUJITA Tomonori wrote:
> > On Wed, 1 Apr 2009 13:50:58 +0200
> > Jens Axboe <jens.axboe@oracle.com> wrote:
> > 
> >> On Wed, Apr 01 2009, FUJITA Tomonori wrote:
> >>> On Wed,  1 Apr 2009 20:04:38 +0900
> >>> Tejun Heo <tj@kernel.org> wrote:
> >>>
> >>>> Impact: fix SG_IO behavior such that it matches the documentation
> >>>>
> >>>> SG_IO howto says that if ->dxfer_len and sum of iovec disagress, the
> >>>> shorter one wins.  However, the current implementation returns -EINVAL
> >>>> for such cases.  Trim iovc if it's longer than ->dxfer_len.
> >>> Is that description about sg's SG_IO?
> 
> It looks like it's the closest thing.
> 
> >> The more important question is what sg.c actually does, that's more
> >> important than the documentation.
> 
> The current code would fail it with -EINVAL but after brief look into
> 2.6.12-rc2, it seems like it would use the shorter one.  On direct
> mapping path, it builds considering both lengths and on indirect path
> it doesn't seem to look at the iov supplied till the transfer is
> actually complete using the dxfer_len and then copy out whatever can
> be copied out.
> 
> > Do you think that Doug is a person who makes such mistake? ;)
> > 
> > Seems that sg worked as the howto says. But I think that I broke it
> > when I converted sg to use the block layer. I'll fix it soon.
> > 
> > About this patch, as we know, there are lots of subtle differences
> > between sg's SG_IO and the block's. I'm not sure that it's a good idea
> > to change the behavior of the block's SG_IO.
> 
> I think it's better to make the behavior more consistent.  Using
> shorter dxfer_len can be considered a feature too, so...

It's definitely a good feature, it's how you'd probe lots of scsi
command returns. Setting up the iovec once and passing first a smaller
than a larger dxfer_len would be a perfectly reasonable way to do that.

-- 
Jens Axboe