Trying to get PaX RANDMMAP into the mainstream kernel
Hi everyone,
I had a brief conversation yesterday with solardiz on freenode/#openwall. The topic turned to what other hardening code could go upstream besides the stuff you guys have already been pushing. It would be nice if we could get PaX RANDMMAP in. It gives better randomization on mmap addresses, but unfortunately breaks packages which use pre-compiled headers. [1]
I wrote a little POC program to demonstrate this. [2] Try running it on vanilla ubuntu, opensuse and gentoo. Then try running it on the same with a hardened kernel with RANDMMAP enabled.
We came across this issue in a hardened gentoo bug. In one of the comments, pipacs <[email protected]> gives a very complete explanation of the situation. [3] I won't repeat it here.
If RANDMMAP does get in, then this would be incentive to the gcc people to address the limitations of their gch code. However, the logic works the other way, so this is also the barrier to getting RANDMMAP upstream. solardiz had a good idea: have some sysctl in /proc/sys/kernel either turn it on or off, or allow you to set the amount of randomization. This eases the impact in a running kernel, so its not something the user is stuck with once they configure, compile and reboot.
Also, I don't know if people here are familiar with Hedrick's work. He has broken up the grsec 50k line monolithic patch into smaller patches which address each feature individually. Critical if you want to get any of this stuff upstream.
BTW, Vasiliy, kudos on your GSoC work.
Refs:
[1] http://gcc.gnu.org/onlinedocs/gcc/Precompiled-Headers.html
[2] http://opensource.dyc.edu/pub/misc/pch-poc.tgz
[3] https://bugs.gentoo.org/show_bug.cgi?id=301299#c31
[4]
https://www.kernel.org/pub/linux/kernel/people/hedrick/security/README.grsecurity
Signature · 5 lines
Anthony G. Basile, Ph.D. Gentoo Linux Developer [Hardened] E-Mail : [email protected] GnuPG FP : 8040 5A4D 8709 21B1 1A88 33CE 979C AF40 D045 5535 GnuPG ID : D0455535
Hi Anthony,
Quoted 3 lines · Anthony G. Basile
I had a brief conversation yesterday with solardiz on freenode/#openwall. The topic turned to what other hardening code could go upstream besides the stuff you guys have already been pushing.
Sure! The previous stuff was just a part that I've tried to somehow handle during a short time window.
Quoted 4 lines
It
would be nice if we could get PaX RANDMMAP in. It gives better randomization on mmap addresses, but unfortunately breaks packages which use pre-compiled headers. [1]
Am I right that the difference is that upstream mmap() respects addr hint from userspace and PaX ignores it?
Quoted 1 line
We came across this issue in a hardened gentoo bug.
Were gcc developers informed about the issue? There is a small possibility that they don't mind to fix it :)
Quoted 5 lines
If RANDMMAP does get in, then this would be incentive to the gcc people to address the limitations of their gch code. However, the logic works the other way, so this is also the barrier to getting RANDMMAP upstream. solardiz had a good idea: have some sysctl in
/proc/sys/kerneleither turn it on or off,
Do you mean ignore addr hint or respect it? I find it too limited and upstream is likely to NACK it.
As it is only a hint for OS, it could be simply ignored, without any sysctl. But as it breaks at least one widespread application, it's unlikely to be applied by upstream :(
Quoted 1 line
or allow you to set the amount of randomization.
Agreed. sysctl is good for containers, which handle ASLR different way. E.g. some containers run apps which don't need much VA space, but another container has space critical applications.
Quoted 4 lines
Also, I don't know if people here are familiar with Hedrick's work. He has broken up the grsec 50k line monolithic patch into smaller patches which address each feature individually. Critical if you want to get any of this stuff upstream.
Great, I didn't know about it. I'll look at it very soon.
Thanks,
Signature · 1 line
Vasiliy