[PATCH 2/5 v16] ARM: Replace string mem* functions for KASan

[Ard Biesheuvel]

On Mon, 9 Nov 2020 at 17:07, Russell King - ARM Linux admin
<linux at armlinux.org.uk> wrote:
>
> On Mon, Nov 09, 2020 at 05:02:09PM +0100, Linus Walleij wrote:
> > On Fri, Nov 6, 2020 at 4:16 PM Russell King - ARM Linux admin
> > <linux at armlinux.org.uk> wrote:
> > > On Fri, Nov 06, 2020 at 02:37:21PM +0100, Linus Walleij wrote:
> >
> > > > Aha. So shall we submit this to Russell? I figure that his git will not
> > > > build *without* the changes from mmotm?
> > > >
> > > > That tree isn't using git either is it?
> > > >
> > > > Is this one of those cases where we should ask Stephen R
> > > > to carry this patch on top of -next until the merge window?
> > >
> > > Another solution would be to drop 9017/2 ("Enable KASan for ARM")
> > > until the following merge window, and queue up the non-conflicing
> > > ARM KASan fixes in my "misc" branch along with the rest of KASan,
> > > and the conflicting patches along with 9017/2 in the following
> > > merge window.
> > >
> > > That means delaying KASan enablement another three months or so,
> > > but should result in less headaches about how to avoid build
> > > breakage with different bits going through different trees.
> > >
> > > Comments?
> >
> > I suppose I would survive deferring it. Or we could merge the
> > smaller enablement patch towards the end of the merge
> > window once the MM changes are in.
> >
> > If it is just *one* patch in the MM tree I suppose we could also
> > just apply that one patch also to the ARM tree, and then this
> > fixup on top. It does look a bit convoluted in the git history with
> > two hashes and the same patch twice, but it's what I've done
> > at times when there was no other choice that doing that or
> > deferring development. It works as long as the patches are
> > textually identical: git will cope.
>
> I thought there was a problem that if I applied the fix then my tree
> no longer builds without the changes in -mm?
>

Indeed. Someone is changing the __alias() wrappers [for no good reason
afaict] in a way that does not allow for new users of those wrappers
to come in concurrently.

Hency my suggestion to switch to the raw __attribute__((alias("..")))
notation for the time being, and switch back to __alias() somewhere
after v5.11-rc1.

Or we might add this to the file in question

#undef __alias
#define __alias(symbol) __attribute__((__alias__(symbol)))

and switch to the quoted versions of the identifier. Then we can just
drop these two lines again later, after v5.11-rc1