Regression with arm in next with stack protector

[Rich Felker]

On Tue, Mar 27, 2018 at 06:20:27PM +0100, Russell King - ARM Linux wrote:
> On Tue, Mar 27, 2018 at 11:35:25AM -0400, Rich Felker wrote:
> > On Tue, Mar 27, 2018 at 10:04:10AM +0100, Russell King - ARM Linux wrote:
> > > On Fri, Mar 23, 2018 at 11:14:53AM -0700, Tony Lindgren wrote:
> > > > Hi,
> > > > 
> > > > Looks like commit 5638790dadae ("zboot: fix stack protector in
> > > > compressed boot phase") breaks booting on arm.
> > > > 
> > > > This is all I get from the bootloader on omap3:
> > > > 
> > > > Starting kernel ...
> > > > 
> > > > data abort
> > > > pc : [<810002d0>]          lr : [<100110a8>]
> > > > reloc pc : [<9d6002d0>]    lr : [<2c6110a8>]
> > > > sp : 81467c18  ip : 81466bf0     fp : 81466bf0
> > > > r10: 80fc2c40  r9 : 81000258     r8 : 86fec000
> > > > r7 : ffffffff  r6 : 81466bf8     r5 : 00000000  r4 : 80008000
> > > > r3 : 81466c14  r2 : 81466c18     r1 : 000a0dff  r0 : 00466bf8
> > > > Flags: nZCv  IRQs off  FIQs off  Mode SVC_32
> > > > Resetting CPU ...
> > > > 
> > > > resetting ...
> > > 
> > > The reason for this is the following code that was introduced by the
> > > referenced patch:
> > > 
> > > +               ldr     r0, =__stack_chk_guard
> > > +               ldr     r1, =0x000a0dff
> > > +               str     r1, [r0]
> > > 
> > > This uses the absolute address of __stack_chk_guard in the decompressor,
> > > which is a self-relocatable image.  As with all constructs like the
> > > above, this absolute address doesn't get fixed up, and so it ends up
> > > pointing at invalid memory (in this case 0x466bf8) vs RAM at 0x80000000,
> > > and the decompressor looks to be around 0x81000000.
> > > 
> > > Such constructs can not be used in the decompressor for exactly this
> > > reason - they need to use PC-relative addressing instead just like
> > > everything else does in head.S.
> > 
> > Can someone please answer why this is even needed to begin with? I
> > don't see any compelling reason __stack_chk_guard needs a particular
> > value in the decompressor, which is not dealing with any non-constant
> > input.
> 
> Untrue - it can do some parsing of the DT and updating/appending
> information from ATAGs.  However, all that should be coming from
> a trusted environment, so I don't see much of a "trust" issue here.
> (If the parent environment is not trusted, then the environment we're
> running in is not trusted.)

OK, I was considering DT constant, but it doesn't really matter as you
say since the input comes from a trusted environment and could subvert
the system in much more direct ways than blowing away the
decompressor's stack buffers if it wanted to.

> > Just putting __stack_chk_guard in its bss should be fine and
> > would eliminate all the risks of wrong code to load a value into it.
> > Alternatively put it in initialized data with the desired value.
> 
> I'm no expert with this, so I can't comment.  I build my kernels
> with gcc 4.7.4, which I don't think supports this feature.

By "this feature" do you mean stack protector? I still have a 4.7.3
for x86 around and -fstack-protector-all works fine on it. Not sure if
there are issues using stack protector with kernel, or on ARM, for
older GCCs. In any case defining __stack_chk_guard as initialized data
should work on any gcc version regardless of whether stack protector
is actually used; it doesn't require any compiler features just basic
C.

Rich