Re: [PATCH v2 7/8] x86/segments/32: Introduce CPU_NUMBER segment
From: Andy Lutomirski
Date: Wed Jun 06 2018 - 15:24:54 EST
On Wed, Jun 6, 2018 at 12:07 PM Brian Gerst <brgerst@xxxxxxxxx> wrote:
>
> On Wed, Jun 6, 2018 at 1:16 PM, Andy Lutomirski <luto@xxxxxxxxxx> wrote:
> > On Wed, Jun 6, 2018 at 9:23 AM Chang S. Bae <chang.seok.bae@xxxxxxxxx> wrote:
> >>
> >> The new entry will be equivalent to that of x86-64 which
> >> stores CPU number. The entry is placed in segment 23 in GDT
> >> by bumping down 23-28 by one, which are all kernel-internal
> >> segments and so have no impact on user space.
> >>
> >> CPU_NUMBER segment will always be at '%ss (USER_DS) + 80'
> >> for the default (flat, initial) user space %ss.
> >
> > No, it won't :( This is because, on Xen PV, user code very frequently
> > sees a different, Xen-supplied "flat" SS value. This is definitely
> > true right now on 64-bit, and I'm reasonably confident it's also the
> > case on 32-bit.
> >
> > As it stands, as far as I can tell, we don't have a "cpu number"
> > segment on 32-bit kernels. I see no compelling reason to add one, and
> > we should definitely not add one as part of the FSGSBASE series. I
> > think the right solution is to rename the 64-bit segment to
> > "CPU_NUMBER" and then have the rearrangement of the initialization
> > code as a followup patch. The goal is to make the patches
> > individually reviewable. As it stands, this patch adds some #defines
> > without making them work, which is extra confusing.
> >
> > Given how many times we screwed it up, I really want the patch that
> > moves the initialization of the 64-bit CPU number to be obviously
> > correct and to avoid changing the sematics of anything except the
> > actual CPU number fields during boot.
> >
> > So NAK to this patch, at least as part of the FSGSBASE series.
> >
> > (My apologies -- a bunch of this is because I along with everyone else
> > misunderstood the existing code.)
>
> The sole purpose of this segment is for the getcpu() function in the
> VDSO. No other userspace code can rely on its presence or location.
>
Agreed. But this means that no code whatsoever should use it on a
32-bit kernel, so let's not add support.\