On Thu, Apr 30, 2009 at 10:59:10AM +0900, Tejun Heo wrote:
> Hello, James.
> 
> James Bottomley wrote:
> > This looks good (although I'd like to test it first).
> 
> Yeah, this will need quite a bit of testing.
> 
> > Might it not be better to have an accessor setting resid_len?  All
> > the other patches in the series insulate users from the actual
> > members of struct request by accessors, so this is a bit the odd man
> > out.
> 
> I actually think it's better to expose resid_len in this case as the
> semantics of the field is - initialized to zero on issue, contains
> residual count on completion and whatever it contains inbetween is
> upto the low level driver.  Request position or length are different
> as they must contain well defined values throughout request processing
> and both block layer and low level driver should agree on what they
> mean.
> 
> Fancy words aside, it basically boils down to allowing llds to do
> either "rq->resid_len = blk_rq_bytes() - xferred" on completion or
> "rq->resid_len = blk_rq_bytes()" on issue and "rq->resid_len -=
> increments" while processing.

Actually, the second one sounds more natural: resid_len == data_len on
issue and decrementing while travelling through block layer and LLDD,
while resid_len == 0 in issue might get confused somewhere.

And I like it too, we've been coming up with all sorts of hacks in
ide-atapi wrt to residual completion and accounting of what got xferred
already and rq->resid_len is much more cleaner, IMHO.

/me testing...

-- 
Regards/Gruss,
    Boris.