Re: [PATCH 2/2] x86: Only do a single page fault for copy_from_user_nmi

[Andi Kleen <ak@linux.intel.com>]

> For now, changing the semantics of the function seems like a sure way to
> fail in the future though.

I doubt it. Nearly nobody uses the exact return value semantics.

(iirc it's mostly write() and some bizarre code in mount)

In fact it's a regular mistake to assume it returns -errno.

> > In theory we could also duplicate the whole copy_*_ path for cases
> > where the caller doesn't care about the exact bytes. But that
> > seems overkill for just this issue, and I'm not sure anyone
> > else cares about how fast this is. The simpler check works
> > as well for now.
> 
> So I don't get that code, but why not fix it in general? Taking two
> faults seems silly.

It's really complicated to reconstruct the exact bytes, as an optimized
memcpy is very complicated and has a lot of corner cases. 

I tried it originally when writing the original copy function, but
failed. That is why people came up later with this two-fault
scheme.

I think two fault is fine for most cases, just not for NMIs.

-Andi

-- 
ak@linux.intel.com -- Speaking for myself only