Re: [PATCH] x86: define arch_crash_save_vmcoreinfo() if CONFIG_CRASH_CORE=y

[Omar Sandoval <osandov@osandov.com>]

On Wed, Dec 18, 2019 at 11:41:13AM +0100, Borislav Petkov wrote:
> On Mon, Dec 02, 2019 at 10:02:29AM -0800, Omar Sandoval wrote:
> > From: Omar Sandoval <osandov@fb.com>
> > 
> > kernel/crash_core.c calls arch_crash_save_vmcoreinfo() to get
> > arch-specific bits for vmcoreinfo. If it is not defined, then it has a
> > no-op fallback. kernel/crash_core.c is gated behind CONFIG_CRASH_CORE.
> > However, x86 defines arch_crash_save_vmcoreinfo() in
> > arch/x86/kernel/machine_kexec_*.c, which is gated behind
> > CONFIG_KEXEC_CORE. So, a kernel with CONFIG_CRASH_CORE=y and
> > CONFIG_KEXEC_CORE=n
> 
> How does that even happen?
> 
> Symbol: KEXEC_CORE [=y]
> Type  : bool
>   Defined at arch/Kconfig:17
>   Selects: CRASH_CORE [=y]
>   Selected by [y]:
>   - KEXEC [=y]
>   - KEXEC_FILE [=y] && X86_64 [=y] && CRYPTO [=y]=y && CRYPTO_SHA256 [=y]=y
> 
> In order to do crash dumps, you need to select KEXEC, which selects
> KEXEC_CORE, which selects CRASH_CORE...
> 
> Or are you talking about the PROC_KCORE use angle where it selects
> CRASH_CORE and the crash_save_vmcoreinfo_init() initcall is then
> supposed to save arch-specific crash info?

Yes, I'm talking about reading VMCOREINFO from /proc/kcore at runtime,
no kdump involved. crash [1] and my own tool, drgn [2], use this for
live debugging.

I can set CONFIG_PROC_KCORE=y, which selects CONFIG_CRASH_CORE=y, but if
CONFIG_KEXEC_CORE=n, the VMCOREINFO is incomplete.

1: https://github.com/crash-utility/crash/commit/60a42d709280cdf38ab06327a5b4fa9d9208ef86
2: https://github.com/osandov/drgn/blob/73dd7def1217e24cc83d8ca95c995decbd9ba24c/libdrgn/program.c#L385