Re: [PATCH 2/2] kbuild: add -fno-PIE

From: Austin S. Hemmelgarn
Date: Fri Nov 04 2016 - 10:47:52 EST

On 2016-11-04 10:24, Sebastian Andrzej Siewior wrote:
On 2016-11-04 07:37:02 [-0400], Austin S. Hemmelgarn wrote:
clued enough to have known better. Reassigning bug reports in question
from gcc-6 to linux is beyond stupid; Balint is either being deliberately
obtuse, or geniunely unable to imagine that somebody might be using the
compiler _not_ for debian package builds.

If it helps, you could point out that Gentoo's hardened profile's GCC builds
use PIE by default and have absolutely zero issues building the Linux kernel
without any special kernel patches to turn it off (and has been doing so for
years).

Interesting. So I took a look at Gentoo. They ship gcc 4.9.3 by default.
They have their own PIE patch since it is not yet supported by gcc. And
let me quote why that works for them:

| This file will add -fstack-protector-all, -fstack-check, -fPIE, -pie and -z now
| as default if the defines and the spec allow it.
| Added a hack for gcc-specs-* in toolchain-funcs.eclass and _filter-hardened in flag-o-matic.eclass
| to support older hardened GCC patches and we don't need to change the code on gcc-specs-* and _filter-hardened.
| This will add some unsupported upstream commands options as -nopie and -nonow.
| -D__KERNEL__ is added so we don't have -fPIE, -pie and -fstack-protector-all and -fstack-check when building kernels.
| ESP_CC1_SPEC is added to CC1_SPEC.
| ESP_CC1_STRICT_OVERFLOW_SPEC is added so we don't disable the strict-overflow check.
| ESP_LINK_PIE_CHECK_SPEC check for -pie, -p, -pg, -profile and -static.
| ENABLE_CRTBEGINP add support for crtbeginP.o, build -static with -fPIE or -fpie.

I was thinking about asking doko for something similar but no. Looking at
portage they have a few patches where they add -fno-PIE to some packages.
Also disabling PIE based on __KERNEL__ does not look right. So no, Gentoo
did not better.
And according to Google, there are also people in the ARCH Linux camp
with the same problem. Gentoo's 6 gcc is completely masked and it does
not reference the patch I quote above so Gentoo will run into this
problem once they enable gcc 6 and don't add the -D__KERNEL__ hack.
Eventually Fedora and SUSE will migrate to PIE by default and by then we
should cover all major distros so even Al should be affected unless he
decides not to update or is using something else.
While I don't agree with _how_ they worked around it, it still works correctly with no user intervention for pretty much every important case, and my point was more that it is possible to make this work without a kernel patch than 'Hey, it works over here, lets do what they're doing'.

I would still argue that the root of the issue is how GCC handles options specified on the command line that conflict with it's compile time defaults. This is at least the second kernel related case where things broke because GCC doesn't do the sensible thing and override defaults based on command line options (there is (or was, not sure if it's been resolved yet or not) an issue too on MIPS with some other option that I can't remember right now). If option X and option Y are mutually exclusive, and option X is specified on the command line while option Y isn't, they should not use option Y regardless of whether or not it's the default and possibly spit out a warning if it is the default (for PIC, yes, there probably should be a warning), not die.