PIE/PIC support on base

[Shawn Webb]

On Fri, Oct 17, 2014 at 10:05 AM, Warner Losh <imp at bsdimp.com> wrote:

> [[cc trimmed ]]
>
> On Oct 17, 2014, at 7:46 AM, Shawn Webb <lattera at gmail.com> wrote:
>
> > On Fri, Oct 17, 2014 at 9:41 AM, Warner Losh <imp at bsdimp.com> wrote:
> >
> > On Oct 17, 2014, at 2:05 AM, Shawn Webb <lattera at gmail.com> wrote:
> >
> > > On Fri, Oct 17, 2014 at 3:53 AM, Jeremie Le Hen <jlh at freebsd.org>
> wrote:
> > >
> > >> On Fri, Oct 17, 2014 at 12:15 AM, Shawn Webb <lattera at gmail.com>
> wrote:
> > >>>
> > >>>
> > >>> On Thu, Oct 16, 2014 at 5:59 PM, Jeremie Le Hen <jlh at freebsd.org>
> wrote:
> > >>>>
> > >>>> On Thu, Oct 16, 2014 at 8:21 PM, David Carlier
> > >>>> <david.carlier at hardenedbsd.org> wrote:
> > >>>>>
> > >>>>> I chose the "atomic" approach, at the moment very few binaries are
> > >>>>> concerned at the moment. So I applied INCLUDE_PIC_ARCHIVE in the
> > >> needed
> > >>>>> libraries plus created WITH_PIE which add fPIE/fpie -pie flags
> only if
> > >>>>> you
> > >>>>> include <bsd.prog.pie.mk> (which include <bsd.prog.mk>...)
> otherwise
> > >>>>> other
> > >>>>> binaries include <bsd.prog.mk> as usual hence does not apply. Look
> > >>>>> reasonable approach ?
> > >>>>
> > >>>> I think I understand what you mean.  But I think PIE is commonplace
> > >>>> nowadays and I don't understand what you win by not enabling it for
> > >>>> the whole system.  Is it a performance concern?  Is it to preserve
> > >>>> conservative minds from to much change? :)
> > >>>
> > >>>
> > >>> Looping in Kostik, Bryan Drewery, the PaX team, Hunger, and Sean
> Bruno.
> > >>>
> > >>> On i386, there is a performance cost due to not having an extra
> register
> > >>> available for the relocation work that has to happen. PIE doesn't
> carry
> > >> much
> > >>> of a performance penalty on amd64, though it still does carry some on
> > >> first
> > >>> resolution of functions (due to the extra relocation step the RTLD
> has to
> > >>> worry about). On amd64, after symbol resolution has taken place,
> there
> > >> is no
> > >>> further performance penalty due to amd64 having an extra register to
> use
> > >> for
> > >>> PIE/PIC. I'm unsure what, if any, performance penalty PIE carries on
> ARM,
> > >>> AArch64, and sparc64.
> > >>>
> > >>> Certain folk would prefer to see PIE enabled only in certain
> > >> applications.
> > >>> /bin/ls can't really make much use of PIE. But sshd can. I personally
> > >> would
> > >>> like to see all of base's applications compiled as PIEs, but that's a
> > >> long
> > >>> ways off. It took OpenBSD several years to accomplish that. Having
> > >> certain
> > >>> high-visibility applications (like sshd, inetd, etc) is a great
> start.
> > >>> Providing a framework for application developers to opt their
> application
> > >>> into PIE is another great start.
> > >>>
> > >>> Those are my two cents.
> > >>
> > >> OK.  As long as i386 is still an important architecture, it can make
> > >> sense to enable this on a per-binary basis if we don't want to have a
> > >> discrepancy between archs. Also I buy your argument on /bin/ls but I
> > >> was challenging to enable for the whole system because I wonder if
> > >> there aren't some unexpected attack surfaces, besides the obvious ones
> > >> (servers).
> > >>
> > >> Do you know what took so much time to OpenBSD?
> > >
> > >
> > > In a private conversation with Theo, I realized that my recollection
> of the
> > > time it took OpenBSD to compile all of base as PIEs was wrong. Quoting
> him:
> > >
> > > "It took 5 people approximately 3 months to debug it, activate it, and
> > > start shipping it the next release.  That was on amd64, for all
> > > dynamically linked binaries, except one (a gcc bug took some time to
> > > find).  The next architectures followed about 1 or 2 per 6-month
> > > release."
> > >
> > > Given that only one person has worked on this in the past (me) and now
> the
> > > task has been delegated to another (David Carlier), I think we're doing
> > > okay on our end. There's a lot of moving parts, and neither of us fully
> > > understand all of them completely. We're working on it in HardenedBSD,
> in
> > > the hardened/current/pie branch.
> > >
> > > I'm thinking we might try for a WITH_PIE knob (and *not* use USE_PIE)
> and
> > > have certain high-profile applications opt-in to PIE until we work out
> all
> > > the details for everything en masse. Baptiste did bring up a good point
> > > with INTERNALLIB and I'm unsure of how we should handle that.
> >
> > WITH_PIE or WITHOUT_PIE controls, on a global basis, via the MK_PIE
> > variable, whether or not the user wants to turn on this feature for those
> > program that can do PIE. Designating which programs do, or don’t,
> > use PIE simply must be done with a different variable. I posted a bit of
> a
> > rant about the current state of things that suggested a couple of
> > alternatives as well as giving some history as to why some options
> > aren’t to be used and the history behind some of my reactions. :)
> >
> > For this reason, I think WITH_PIE, as I understand your proposal,
> > likely isn’t a good fit with other WITH_xxx variables used in the src
> > tree today.
> >
> > Gotcha. To be honest, I found your email a tad bit confusing. Are you
> suggesting we create an ENABLE_feature framework? Or are you suggesting we
> have a USE_PIE flag? Or are you suggesting something different entirely
> (and if you are, what?)?
>
> I’m saying we don’t have a good framework at the moment to do this. We
> have several bad ones that all have their pitfalls. This is one reason I
> had
> the fast reaction to NO_PIE, then a minute later said “go ahead and use
> it and I’ll fix it.” I’m still cool with that position, btw.
>
> As for a name, that can be debated a  lot, but I’d like to see something
> new, easy to use and unambiguous. If you are looking for a suggestion
> for that name, let’s go with WANTS_PIE. Only Makefiles can set it.
>
> WANTS_PIE undefined means do the default behavior as defined by the
> current MK_PIE setting and perhaps system policy. “Go with this flow."
>
> WANTS_PIE=yes means that if MK_PIE is “yes”, then do PIE things for
> this thing we’re building. If MK_PIE is “no”, though PIE is disabled for
> everything.
>
> WANTS_PIE=no means that if MK_PIE is “yes”, then disable doing PIE
> things for this component. If MK_PIE is no, it is also disabled.
>
> This could also be extended to NEEDS_foo, which says “I need foo to
> build, and if MK_foo is set to no, don’t build me.” I don’t think anything
> that you are doing falls into this category though.
>
> WANTS/NEEDS also avoids the historical use of USE in the ports tree
> possibly creating confusion.
>
> If you go with WANTS_PIE, then you wouldn’t need bad.*.pie.mk.
>
> Comments?


I like that idea. I think we need buy off from Kostik. David, what are your
thoughts?

Thanks,

Shawn