On Wed, Oct 31, 2018 at 2:10 AM, Peter Zijlstra [off-list ref] wrote:

On Tue, Oct 30, 2018 at 04:28:16PM +0000, Will Deacon wrote:

quoted

On Tue, Oct 30, 2018 at 04:58:41PM +0100, Peter Zijlstra wrote:

quoted

Like mentioned elsewhere; if you do write_enable() + write_disable()
thingies, it all becomes:

    write_enable();
    atomic_foo(&bar);
    write_disable();

No magic gunk infested duplication at all. Of course, ideally you'd then
teach objtool about this (or a GCC plugin I suppose) to ensure any
enable reached a disable.

Isn't the issue here that we don't want to change the page tables for the
mapping of &bar, but instead want to create a temporary writable alias
at a random virtual address?

So you'd want:

      wbar = write_enable(&bar);
      atomic_foo(wbar);
      write_disable(wbar);

which is probably better expressed as a map/unmap API. I suspect this
would also be the only way to do things for cmpxchg() loops, where you
want to create the mapping outside of the loop to minimise your time in
the critical section.

Ah, so I was thikning that the altnerative mm would have stuff in the
same location, just RW instead of RO.

I was hoping for the same location too. That allows use to use a gcc
plugin to mark, say, function pointer tables, as read-only, and
annotate their rare updates with write_rare() without any
recalculation.

-Kees

-- 
Kees Cook