[Bug ld/20113] Partial relro support for s390x ld

[krebbel at linux dot ibm.com]

https://sourceware.org/bugzilla/show_bug.cgi?id=20113

--- Comment #19 from Andreas Krebbel <krebbel at linux dot ibm.com> ---
(In reply to maamountki from comment #18)
> *patch
> 
> I just noticed that SEPARATE_GOTPLT is duplicated in the 64 bit script and
> partial relro support is already implemented so yes my patch should be
> reverted, I'm contacting Bountysource to revert the claim submission too.
> Anyway I'm not sure why you think that SEPARATE_GOTPLT must remain as 0
> hence the relro segment doesn't extend across those entries, SEPARATE_GOTPLT
> have this argument for the exact reason and architectures such as x86 and
> aarch64 take advantage of this feature and there is a reason actually there
> are no harm in making those entries non-exploitable.

For IBM Z we had to chose a slightly different scheme. Our ABI requires that
*all* GOT entries (.got + .got.plt) are accessible via _GLOBAL_OFFSET_TABLE_
symbol and a *positive* displacement. Just setting SEPARATE_GOTPLT makes
_GLOBAL_OFFSET_TABLE_ to point at the start of .got.plt which would come after
the other .got entries. Hence a negative offset would be needed to address .got
entries. More changes to the backend were required to make
_GLOBAL_OFFSET_TABLE_ end up at the start of .got.

As you say the value of SEPARATE_GOTPLT is used to make the relro segment cover
the magic entries assumed to be located at start of .got.plt. With moving
_GLOBAL_OFFSET_TABLE_ to start of .got I also had to move the 3 magic entries
to the start of .got. So they will be covered by the relro segment anyway.
Setting it to 0 is the right value for us then.

I apologize for not monitoring the bugzilla and closing it after the relro work
was finished. I know you are doing a great job for other of our bounties (e.g.
OpenBlas). Please keep up your great work!

-- 
You are receiving this mail because:
You are on the CC list for the bug.