Re: [patch] scsi: revert "[SCSI] Get rid of scsi_cmnd->done"

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Sun, 2008-01-06 at 18:19 +0200, Boaz Harrosh wrote:
> On Sun, Jan 06 2008 at 5:43 +0200, Linus Torvalds <torvalds@linux-foundation.org> wrote:
> > 
> > 
> > This all still leaves the question unanswered why that commit 
> > 6f5391c283d7fdcf24bf40786ea79061919d1e1d changed any behaviour at all. 
> > Because the thing that Peter is describing has nothing to do with any 
> > low-level drivers what-so-ever.
> > 
> > 		Linus
> > 
> 
> James Matthew.
> I have a (very) wild guess at what maybe have changed with the cmnd->done
> patch:
> 
> Do you remember the effective loop in scsi_lib:scsi_end_request() where
> if bufflen was smaller then original request size, do to truncation
> of bufflen by ULD, then the remaining of the request is re-queued again
> as a new scsi-command. Well I think that the old system would call
> cmnd->done for every iteration, and the new system, since the done is
> called by the block-Q, does not see the resubmit of the new command.

Actually, this is cmnd->done, not req->done we're removing.
cmnd->done() isn't seen by the block layer; all its uses are in the SCSI
mid-layer.

> I have not followed all code path of the matter, but I know that sr does
> alters bufflen in some cases. 
> All this is not a bug in itself, but it is a change in behavior that might 
> cause the current sr hack to fail.

It's a good thought.  You're right, the old code calls done for every
iteration.  However, it calls it in scsi_finish_completion.  The new
code will actually call drv->done() in that same spot for every
iteration as well.

The requeue is done via scsi_requeue_request which calls
blk_requeue_request, which resets the START flag and sends the command
right back through the system (including the prep function because
scsi_requeue_request unpreps the command), so even with the new code
we'll go back through all the same done paths.

James