Re: [PATCH] loop sendfile retval

[Hugh Dickins]

On Mon, 11 Nov 2002, Linus Torvalds wrote:
> On Mon, 11 Nov 2002, Hugh Dickins wrote:
> >
> > Buffer I/O error on device loop: its use of sendfile is (trivially)
> > broken - retval is usually count done, only an error when negative.
> 
> Hmm.. Sendfile can return other values than "count" (ie a partial read).  
> This return value change makes "do_lo_receive()" lose that information. As 
> such, the new do_lo_receive() is weaker than the old one.

True, with that patch it's passing back less info than 2.5.47 tried to do;
but no less than 2.5.46 and earlier, which always returned desc.error and
ignored the desc.written, desc.count coming back from do_generic_file_read.
So it's not a regression, but of course you're right that it's weak.

> If fixing the loop code to handle partial IO is too nasty, then I would
> suggest doing maybe something like
> 
> 	if (ret > 0 && ret != bvec->bv_len)
> 		ret = -EIO;
> 
> which at least makes a partial IO an error instead of making it a success 
> case (the code as-is seems to think that any non-negative return value 
> means that the IO was fully successful).
> 
> > Nearby spinlocking clearly bogus, delete instead of remarking on it.
> 
> I'll apply the patch, it looks better than what is there now, but it might 
> be worth fixing this _right_.

Thanks, that gets it going again.  I'll step aside and leave the correct
partial handling to those who know loop much better than I - Adam?

Hugh