Re: [GIT PULL] objtool fixes

From: Linus Torvalds <torvalds@linux-foundation.org>
To: Ingo Molnar <mingo@kernel.org>, Josh Poimboeuf <jpoimboe@redhat.com>
Cc: Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
	Thomas Gleixner <tglx@linutronix.de>,
	"H. Peter Anvin" <hpa@zytor.com>,
	Peter Zijlstra <a.p.zijlstra@chello.nl>
Subject: Re: [GIT PULL] objtool fixes
Date: Tue, 28 Feb 2017 17:55:11 -0800	[thread overview]
Message-ID: <CA+55aFw3+uoT=9Rh0LCF5fhCRtY0uaRHTs+6_K5pSjqrvZdMRw@mail.gmail.com> (raw)
In-Reply-To: <20170228075305.GA8775@gmail.com>

Guys,
 the recent 'objtool' pull request broke things.

I haven't bisected it, but I'm pretty sure that this part is pure garbage:

On Mon, Feb 27, 2017 at 11:53 PM, Ingo Molnar <mingo@kernel.org> wrote:
>
> diff --git a/arch/x86/kernel/vmlinux.lds.S b/arch/x86/kernel/vmlinux.lds.S
> index e79f15f108a8..ad0118fbce90 100644
> --- a/arch/x86/kernel/vmlinux.lds.S
> +++ b/arch/x86/kernel/vmlinux.lds.S
> @@ -346,6 +346,7 @@ SECTIONS
>         /DISCARD/ : {
>                 *(.eh_frame)
>                 *(__func_stack_frame_non_standard)
> +               *(__unreachable)
>         }
>  }
>
> diff --git a/include/linux/compiler-gcc.h b/include/linux/compiler-gcc.h
> index 0444b1336268..f457b520ead6 100644
> --- a/include/linux/compiler-gcc.h
> +++ b/include/linux/compiler-gcc.h
> @@ -195,6 +195,17 @@
>  #endif
>  #endif
>
> +#ifdef CONFIG_STACK_VALIDATION
> +#define annotate_unreachable() ({                                      \
> +       asm("%c0:\t\n"                                                  \
> +           ".pushsection __unreachable, \"a\"\t\n"                     \
> +           ".long %c0b\t\n"                                            \
> +           ".popsection\t\n" : : "i" (__LINE__));                      \
> +})
> +#else
> +#define annotate_unreachable()
> +#endif

and I think the above is what breaks module loading for me right now
on my laptop.

I get this during bootup:

    module: overflow in relocation type 10 val ffffffffc02afc81
    module: 'nvme' likely not compiled with -mcmodel=kernel

(and similar errors for other modules too), but those modules very
much *are* compiled with all the normal kernel build flags, including
-mcmodel=kernel.

Now, relocation type 10 is R_X86_64_32, so the warning is very true:
that address would fit in a _signed_ 32-bit value, but that's
supposedly a 32-bit unsigned relocation.

Trying to figure out what the hell is going on, I do:

    objdump -r nvme.ko | grep 64_32

and what do I find? I find

  RELOCATION RECORDS FOR [__unreachable]:
  OFFSET           TYPE              VALUE
  0000000000000000 R_X86_64_32       .text+0x0000000000000c81
  0000000000000004 R_X86_64_32       .text+0x0000000000000cb5
  0000000000000008 R_X86_64_32       .text+0x0000000000001a18
  000000000000000c R_X86_64_32       .text+0x0000000000001a36
  0000000000000010 R_X86_64_32       .text+0x0000000000001e38
  0000000000000014 R_X86_64_32       .text+0x0000000000001ec2
  0000000000000018 R_X86_64_32       .text+0x00000000000034e2
  000000000000001c R_X86_64_32       .text+0x0000000000003536

and then when I look more closely (objdump --disassemble), I see that
the offset 000c81 in the module refers to this:

  0000000000000c60 <nvme_admin_init_request>:
        ....
     c7f:       0f 0b                   ud2
     c81:       0f 1f 44 00 00          nopl   0x0(%rax,%rax,1)

so it very much looks like those relocations are still around on
modules, and so module loading fails.

Anyway, those annotations are completely bogus anyway, it looks. You
guys should use relative offsets in order to be able to specify a
kernel address. So doing

    .long %c0

is garbage - either it needs to be a .quad, or it needs to be relative
to the text section to fit in a .long.

Hmm? Revert or fix, but please quickly...

                   Linus