Re: [RFC PATCH] binfmt_elf: Fix PIE execution with randomization disabled

[Josh Boyer <jwboyer@redhat.com>]

On Thu, Sep 29, 2011 at 08:41:33PM -0400, Nicolas Pitre wrote:
> On Thu, 29 Sep 2011, Andrew Morton wrote:
> 
> > On Thu, 29 Sep 2011 15:53:59 -0400
> > Josh Boyer <jwboyer@redhat.com> wrote:
> > 
> > > We've had a bug report[1] of some PIE programs getting a SIGKILL upon exec
> > > if you disable address randomization with:
> > > 
> > > 	echo 0 > /proc/sys/kernel/randomize_va_space
> > > 
> > > I tracked this down to get_unmapped_area_prot returning -ENOMEM because
> > > the address being passed in is larger than TASK_SIZE - len for the bss
> > > section of the test executable.  That filters back to set_brk returning
> > > an error to load_elf_binary and the SIGKILL being sent around line 872
> > > of binfmt_elf.c.
> > > 
> > > H.J. submitted an upstream bug report [2] as well, but got no feedback
> > > and we can't view it with kernel.org being down anyway.  He came up with
> > > the patch below as well, which is what I'm sending on for comments.  The
> > > changelog is my addition, so if that is wrong yell at me.
> > > 
> > > I wanted to get some more eyes on this, because the current code sets
> > > load_bias to 0 unconditionally on CONFIG_X86 or CONFIG_ARM.  I have no
> > > idea why that is.  The original execshield patches had an #ifdef on
> > > __i386__ but the patch that was commited to add PIE support has the
> > > CONFIG_X86 setting.
> > > 
> > 
> > It appears that Nicolas understood what's going on in there when he
> > wrote e4eab08d6050ad0 ("ARM: 6342/1: fix ASLR of PIE executables"). 
> > Alas, that patch's changelog is rather useless.
> > 
> > Help?
> 
> Well, in order to obtain randomization, the addr argument to elf_map() 
> must be zero to eventually let arch_get_unmapped_area() do its job of 
> selecting a random address.

Hm.  But it doesn't do that if ASLR is disabled at runtime, so the
address causes issues?

> Since only X86 supported ASLR at the time, I simply did the same for ARM 
> i.e. let load_bias be set to 0 so elf_map() would get a zero address.

So as of right now, only X86 and ARM support ASLR?  Maybe we could
change the define to be more descriptive.  Something like ARCH_HAS_ASLR?

> > Also, please: review and test?
> > 
> > >  #if defined(CONFIG_X86) || defined(CONFIG_ARM)
> > > -			load_bias = 0;
> > > +			if (vaddr)
> > > +				load_bias = 0;
> > > +			else
> > > +				load_bias = ELF_PAGESTART(ELF_ET_DYN_BASE);
> > >  #else
> > >  			load_bias = ELF_PAGESTART(ELF_ET_DYN_BASE - vaddr);
> > >  #endif
> 
> Simply looking at this patch, I don't see how the second argument to 
> elf_map() called as follows could ever be zero anymore, effectively 
> breaking ASLR.

Perhaps another check here for randomize?  Something like:

#if defined(CONFIG_X86) || defined(CONFIG_ARM)
		if (current->flags & PF_RANDOMIZE)
			load_bias = 0;
		else if (vaddr)
			load_bias = 0;
		else
			load_bias = ELF_PAGESTART(ELF_ET_DYN_BASE);
#else
		load_bias = ELF_PAGESTART(ELF_ET_DYN_BASE - vaddr);
#endif

If that's stupid, then feel free to tell me.  I won't pretend like I
understand what is going on here yet, but based on the explanation you
provided that might work.

josh