Re: more build problems with "Makefile: move stackprotector availability out of Kconfig"

[Arnd Bergmann]

On Tue, Oct 17, 2017 at 5:34 PM, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> On Tue, Oct 17, 2017 at 8:26 AM, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>> On Tue, Oct 17, 2017 at 8:23 AM, Arnd Bergmann <arnd@xxxxxxxx> wrote:
>>> On Tue, Oct 17, 2017 at 1:00 PM, Arnd Bergmann <arnd@xxxxxxxx> wrote:
>>>> Hi Kees,
>>>>
>>>> On my test box, current linux-next kernels fail to build due to the
>>>> patch that introduces CONFIG_CC_STACKPROTECTOR_AUTO, with my mainline
>>>> gcc
>>>> builds up to gcc-5.5.0. gcc-6 and higher work fine, as
>>>> scripts/gcc-x86_64-has-stack-protector.sh returns 'y' for those.
>>>>
>>>> Using the compilers provided by Ubuntu (4.6/4.7/4.8/4.9), everything
>>>> also works as expected, so my interpretation is that mainline gcc did
>>>> not enable the stack protector until gcc-6, while distributions did.
>>>>
>>>> Do you agree with that interpretation?
>>>
>>> It's probably a little different. I tried bisecting the gcc commit that fixed
>>> the issue for me, and ended up with this commit
>>>
>>> https://gitlab.indel.ch/thirdparty/gcc/commit/c14bac81551d6769741c2b1cc55e04d94fe8d3a7
>>>
>>> that caused the target to change from x86_64-unknown-linux to
>>> x86_64-pc-linux, and apparently caused the compiler bootstrap
>>> to incorrectly identify the capabilities of the assembler. As a result,
>>> the assembler output inside of scripts/gcc-x86_64-has-stack-protector.sh
>>> that should be
>>> [snip]
>>
>> Yeah, %gs: vs __stack_chk_guard global.
>>
>> Do you know which gccs (of the past) had this?
>>
>> akpm's build error is different still, there are no warnings at all
>> and then the build fails with missing __stack_chks. I'm still trying
>> to figure that one out.
>
> Oh, I think I know what's happening. I'm going to try to simulate this
> and send another patch for testing...
>
> (I'm still curious about the compiler versions, since my gcc 4.4.4
> works fine for stack-protector.)

I've managed to reduce the change that fixed it to this bit in the
compiler sources:

index dbfb978..d5bc694 100755
--- a/config.guess
+++ b/config.guess
@@ -1021,7 +1021,7 @@ EOF
        echo ${UNAME_MACHINE}-dec-linux-${LIBC}
        exit ;;
     x86_64:Linux:*:*)
-       echo ${UNAME_MACHINE}-unknown-linux-${LIBC}
+       echo ${UNAME_MACHINE}-pc-linux-${LIBC}
        exit ;;
     xtensa*:Linux:*:*)
        echo ${UNAME_MACHINE}-unknown-linux-${LIBC}

I still don't know why that makes a difference, but all versions
prior to gcc-6.1 have the problem for me.

      Arnd