David Miller wrote:

Another point worth making is that for function calls you
can fix things up lazily if you want.
[...]
In fact forget I mentioned this idea :)
  

OK :)  I think we'll only ever want to bind to a hypervisor once, since
the underlying hypervisor can't change on the fly (well, in principle it
might if you migrate, but you'll have more problems than just dealing
with the hook points).  And lazy binding doesn't buy much; I think its
much better to get it all out of the way at once, and then throw the
reloc info away.

As another note, I do agree with Linus about the register usage
arguments.  It is important.  I think it's been mentioned but what you
could do is save nothing (so that "sti" and "cli" are just that and
cost nothing), but the more complicated versions save and restore
enough registers to operate.
  

Right, that's pretty much what we do now.

It all depends upon what you're trying to do.  For example, it's
easy to use patching to make different PTE layouts be supportable
in the same kernel image.  We do this on sparc64 since sun4v
has a different PTE layout than sun4u, you can see the code in
asm-sparc64/pgtable.h for details (search for "sun4v_*_patch")
  

I see.  We want to do something conceptually like this, but we need to
handle more than just adjusting which of two constants to use as a
mask.  For example, Xen needs to run pfns through a pfn->mfn (machine
frame number) conversion and back when making/unpacking pagetable entries.

    J