RE: [PATCH] x86/static_call: Remove text_mutex from __static_call_fixup()

[Kaplan, David]

[AMD Official Use Only - AMD Internal Distribution Only]

> -----Original Message-----
> From: Kaplan, David
> Sent: Wednesday, December 10, 2025 1:40 PM
> To: 'Borislav Petkov' <bp@xxxxxxxxx>
> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>; Josh Poimboeuf
> <jpoimboe@xxxxxxxxxx>; Jason Baron <jbaron@xxxxxxxxxx>; Steven Rostedt
> <rostedt@xxxxxxxxxxx>; Ard Biesheuvel <ardb@xxxxxxxxxx>; Thomas Gleixner
> <tglx@xxxxxxxxxxxxx>; Ingo Molnar <mingo@xxxxxxxxxx>; Dave Hansen
> <dave.hansen@xxxxxxxxxxxxxxx>; x86@xxxxxxxxxx; H. Peter Anvin
> <hpa@xxxxxxxxx>; linux-kernel@xxxxxxxxxxxxxxx
> Subject: RE: [PATCH] x86/static_call: Remove text_mutex from
> __static_call_fixup()
>
>
>
> > -----Original Message-----
> > From: Borislav Petkov <bp@xxxxxxxxx>
> > Sent: Wednesday, December 10, 2025 1:55 PM
> > To: Kaplan, David <David.Kaplan@xxxxxxx>
> > Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>; Josh Poimboeuf
> > <jpoimboe@xxxxxxxxxx>; Jason Baron <jbaron@xxxxxxxxxx>; Steven
> Rostedt
> > <rostedt@xxxxxxxxxxx>; Ard Biesheuvel <ardb@xxxxxxxxxx>; Thomas
> Gleixner
> > <tglx@xxxxxxxxxxxxx>; Ingo Molnar <mingo@xxxxxxxxxx>; Dave Hansen
> > <dave.hansen@xxxxxxxxxxxxxxx>; x86@xxxxxxxxxx; H. Peter Anvin
> > <hpa@xxxxxxxxx>; linux-kernel@xxxxxxxxxxxxxxx
> > Subject: Re: [PATCH] x86/static_call: Remove text_mutex from
> > __static_call_fixup()
> >
> > Caution: This message originated from an External Source. Use proper
> caution
> > when opening attachments, clicking links, or responding.
> >
> >
> > Hey,
> >
> > On Thu, Nov 06, 2025 at 11:53:20AM -0600, David Kaplan wrote:
> > > __static_call_fixup() is only called from apply_returns() which modifies
> > > text either on boot when only one CPU is active, or on module load.  In
> > > both cases, it does not modify live code.  This is why this function passes
> > > 'true' for the modinit argument to __static_call_transform() which causes
> > > __static_call_transform() to use text_poke_early().
> > >
> > > text_poke_early() does not require that the text_mutex is held, as this
> > > function is only used when free of any races with other CPUs.  Therefore
> > > __static_call_fixup() does not need to take the mutex so remove it.
> > >
> > > Signed-off-by: David Kaplan <david.kaplan@xxxxxxx>
> > > ---
> > >  arch/x86/kernel/static_call.c | 2 --
> > >  1 file changed, 2 deletions(-)
> > >
> > > diff --git a/arch/x86/kernel/static_call.c b/arch/x86/kernel/static_call.c
> > > index 378c388d1b31..748fc98d416f 100644
> > > --- a/arch/x86/kernel/static_call.c
> > > +++ b/arch/x86/kernel/static_call.c
> > > @@ -212,10 +212,8 @@ bool __static_call_fixup(void *tramp, u8 op, void
> > *dest)
> > >               return false;
> > >       }
> > >
> > > -     mutex_lock(&text_mutex);
> > >       if (op == RET_INSN_OPCODE || dest == &__x86_return_thunk)
> > >               __static_call_transform(tramp, RET, NULL, true);
> > > -     mutex_unlock(&text_mutex);
> > >
> > >       return true;
> > >  }
> > >
> >
> > can you please elaborate here what the actual problem was?
> >
> > I know it had to do with dynamic mitigations, I guess something about
> > lockdep complaining... can you paste the splat here?
> >
>
> Yeah, the full splat can be found in https://lore.kernel.org/lkml/d255d8a9-
> 8e45-485e-9853-80c343bbb73b@xxxxxxxx/
>
> The problem was that patching under dynamic mitigations is done under NMI
> context, so it complained about taking a mutex in NMI context.
>
> Thing is, dynamic mitigation patching happens with all CPUs stopped and
> synchronized at a safe point.  This allows for using text_poke_early(), as we do
> during boot.  This particular code path also always ends up at
> text_poke_early(), which does not require the text_mutex being held.  So
> taking it seems unnecessary (unless I've missed something in my analysis).
>

After some discussions at LPC, I believe this patch is still needed.  In current upstream code, this function is only called by the BSP during early boot and text_mutex isn't required.  Importantly, none of the other logic that uses text_poke_early() acquires the text_mutex either.

With dynamic mitigations, I think I do need to acquire text_mutex since I'm going and modifying a bunch of text, but that needs to be done before stop_machine_nmi().  And so we can't have this function try to take it again.

--David Kaplan