[openib-general] Re: [PATCH 03/13] [RFC] ipath copy routines

[Andi Kleen]

On Sat, Dec 17, 2005 at 09:36:29PM -0800, Robert Walsh wrote:
> On Sun, 2005-12-18 at 04:27 +0100, Andi Kleen wrote:
> > Robert Walsh <rjwalsh at pathscale.com> writes:
> > > 
> > > Any chance we could get these moved into the x86_64 arch directory,
> > > then?  We have to do double-word copies, or our chip gets unhappy.
> > 
> > Standard memcpy will do double word copies if everything is suitably
> > aligned. Just use that.
> 
> This is dealing with buffers that may be passed in from user space, so
> there's no guarantee of alignment for either the start address or the
> length.

So how can you do double word access when the length is not a multiple of four? 

The current x86-64 copy_from_user will use double work access even in that case,
except for the end of course.

But what you're doing is so deeply unportable it's not funny. I am not
sure such a unportable driver even belongs in the kernel.

If the code was really intended to run on user space addresses it 
was totally broken btw because it didn't handle exceptions.

-Andi