Re: [PATCH 3/3] x86_64: Use relative per-cpu offsets

[Peter Zijlstra <peterz@infradead.org>]

On Sun, Nov 14, 2021 at 01:29:46PM -0500, Brian Gerst wrote:
> On Sun, Nov 14, 2021 at 6:03 AM Peter Zijlstra <peterz@infradead.org> wrote:
> >
> > On Sat, Nov 13, 2021 at 11:54:19PM -0500, Brian Gerst wrote:
> > > On Sat, Nov 13, 2021 at 8:18 PM Andy Lutomirski <luto@kernel.org> wrote:
> > > >
> > > >
> > > >
> > > > On Sat, Nov 13, 2021, at 4:40 AM, Brian Gerst wrote:
> > > > > The per-cpu section is currently linked at virtual address 0, because
> > > > > older compilers hardcoded the stack protector canary value at a fixed
> > > > > offset from the start of the GS segment.  Use a standard relative offset
> > > > > as the GS base when the stack protector is disabled, or a newer compiler
> > > > > is used that supports a configurable location for the stack canary.
> > > >
> > > > Can you explain the benefit?  Also, I think we should consider dropping support for the fixed model like we did on x86_32.
> > >
> > > This patch probably makes more sense if we drop the fixed model, as
> > > that gets rid of alot of code that works around having to link the
> > > percpu section differently.
> >
> > Can someone spell out these benefits please? To me having per-cpu start
> > at 0 makes perfect sense, how does not having that make things better?
> 
> The best reason is that the percpu section is currently not subject to
> KASLR.  It actually needs extra support to counter the effects of
> relocation.  There have also been a number of linker bugs over the
> years that have had to be worked around.

I'm confused.. having the variables 0-offset is related to KASLR how?
The dynamic placement of per-cpu thunks and their base address in %GS
gives plenty opportunity to move them around at boot time, no?

> If we were to decide to drop the fixed stack protector the diffstat
> would look something like:

Dropping the fixed stack protecter seems fine to me; I just don't see
why we should move away from 0-offset.