Re: [CFT] ASLR, PIE, and segvguard on 11-current and 10-stable

On Wed, May 14, 2014 at 09:58:52AM -0400, Shawn Webb wrote:
> Hey All,
> 
> [NOTE: crossposting between freebsd-current_at_, freebsd-security_at_, and
> freebsd-stable_at_. Please forgive me if crossposting is frowned upon.]
> 
> Address Space Layout Randomization, or ASLR for short, is an exploit
> mitigation technology. It helps secure applications against low-level
> exploits. A popular secure implementation is known as PaX ASLR, which is
> a third-party patch for Linux. Our implementation is based off of PaX's.
> 
> Oliver Pinter, Danilo Egea, and I have been working hard to bring more
> features and robust stability to our ASLR patches. We've done extensive
> testing on amd64. We'd like to get as many people testing these patches.
> Given the nature of them, we'd also like as many eyeballs reviewing the
> code as well.
> 
> I have a Raspberry Pi and have noticed a few bugs. On ARM (at least, on
> the RPI), when a parent forks a child, and the child gracefully exits,
> the parent segfaults with the pc register pointing to 0xc0000000. That
> address is always the same, no matter the application. If anyone knows
> the ARM architecture well, and how FreeBSD ties into it, I'd like a
> little guidance.
> 
> I also have a sparc64 box, but I'm having trouble getting a vanilla
> 11-current system to be stable on it. I ought to file a few PRs.
> 
> You can find links to the patches below.
> 
> Patch for 11-current:
> http://www.crysys.hu/~op/freebsd/patches/20140514091132-freebsd-current-aslr-segvguard-SNAPSHOT.diff
> 
> Patch for 10-stable:
> http://www.crysys.hu/~op/freebsd/patches/20140514091132-freebsd-stable-10-aslr-segvguard-SNAPSHOT.diff
> 

Shawn

I appreciate you working on this. We must have this in FreeBSD.

I looked at the patch and I read, but not run it.  Comments below.

My personal opinion is that kern_pax.c should be compiled in by default.  If
it adds a lot of size, it'd be better to provide empty stub calls instead of
#ifdef'ing everything. But security is very important especially in
embeddded systems, so you can imagine you're writing the code that everybody
wants and must have enabled for decent level of security.

All modern systems run with ASLR turned on.

I skipped user-space stuff. I don't think it's necessary in this commit and
should be separated.

There's a lot of lines of code for status showing. Not sure if we care that
much: ASLR is either on or off. Not sure about more granularity. More below.

Lots of files:

You conditionally make .sv_pax_aslr_init method point to something else. I'd
assume PAX function _pax_aslr_init32() always gets called and based on
whether ASLR is on or not, it does something or not. This will simplify the
code a lot, and the difference probably won't be measurable.

You have:

	int	a;
	int	b;

instead of:

	int	a, b;

And you miss spaces around "=" sometimes.

kern_jail.c:

something looks wrong here. Sounds like you need "pr->pax".  But I don't
understand why you need to have these pr_* values here. It seems
unnecessary.

kern_pax.c:

I can't quickly tell what locking is using. Some ASSERTS() in pax_ function
would help.

pax_aslr_active():

I don't see why you need to pass "td" and "proc" (I looked at usage: you
pass proc only once). I think you could always pass proc to it, with
td->td_proc passed typically.

kern_pax_*:

There's so many SYSCTLs I think people will have problem configuring it.
Pick reasonable value for all values and let users change them via
SYSCTL_INT (static sysctls) only for debugging.

I can imagine we won't want ASLR only temporarily, for ports which break and
must be fixed. So we probably just need per-process ASLR on/off switch and a
wrapper which could be used like:

	aslr off program ....

The debug stuff I'd remove too. We could have additional CTR stubs used
there, if necessary.

segvguard part I didn't understand. Why do you keep a list of programs that
failed? There was no ASSERTs, thus it was hard to understand the locking
too.

I'm trying to understand if randomization is done correctly. Do you think
you could post the results?

Program:

	http://pastebin.com/XTRHLhMg

Results:
	cat > aslr.c
	<paste>
	gcc aslr.c -o aslr
	echo 1 2 3 4 5 | xargs -I % -n 1 echo "./aslr > aslr.%" | sh
	paste aslr.[12345] | column -t

Linux with ASLR:
	
	http://pastebin.com/UuwW1JMN

MacOSX:

	http://pastebin.com/kuQnYS4e

Thanks,

-- 
Wojciech A. Koszek
wkoszek_at_FreeBSD.czest.pl
http://FreeBSD.czest.pl/~wkoszek/