To NX or not to NX

[Dave Jones]

On Mon, Jan 01, 2007 at 02:12:46PM -0500, Jeremy Katz wrote:

 > > doesn't anaconda default to this on NX capable machines?
 > > (all NX capable machines have PAE support as well obviously)
 >
 > We tried doing so with an FC5 test release, but things didn't go so well
 > so we reverted back for the final.

Was that NX, or was it "If we have PAE, do PAE ?".
My memory is hazy as to what exactly broke, but I wonder if we
can just special case the bad eggs.

 > CONFIG_HIGHMEM64G also isn't that
 > great for laptops.
 >
 > The work really needs to get done so that pae can be done at runtime,
 > much like smp alternatives lets us do for smp.  Separate kernels is
 > _always_ a losing battle for someone ;-)

I think it's a bigger effort to make that happen than it was for smp-alternatives,
at least in part because it needs to happen really early during boot,
possibly before we even get to C code.

Even with alternatives style runtime patching, I think we'd still have
to end up with some stuff being conditionals at runtime whereas right now
they're #defines, which the compiler can sometimes even CSE.
Introducing conditionals in places like page fault handling just screams
"performance hit" to me.

As much as we all dislike it, a separate kernel image really is the best option here.
The complexity involved in getting this stuff right is just horrific.

		Dave

-- 
http://www.codemonkey.org.uk