Re: [PATCH WIP] x86/kgdb: trampolines for shadowed instructions

[Thomas Gleixner]

On Wed, Aug 14 2024 at 11:29, Daniel Thompson wrote:
> On Wed, Aug 14, 2024 at 10:51:41AM +0200, Florian Rommel wrote:
> That's enough to eventuallyremove the int3 instructions but it relies
> on entering the debug trap handler and there's no limit on how long
> could take before that happens. For that reason I think the core should
> also attempt to transition BP_REMOVE_PENDING breakpoints to BP_REMOVE
> after kgdb_skipexception() returns true. That means if we keep trapping
> on a disabled breakpoint eventually we will hit a window where the
> text_mutex is free and clean things up.

Even when text_mutex is uncontended then text_poke_kgdb() is completely
broken in the KGDB NMI context when the NMI hit into anything related to
mm switching and tlb flushing, which is utilized in __text_poke().

The same problem is obviously true for installing a breakpoint from that
context.

I'm starting to be more convinced that the only sane solution for all of
this is to disable CET when KGDB is on and use CRO.WP to work around all
of this.

Thanks,

        tglx