Re: [PATCH v11 01/11] x86/hw_breakpoints: Make DR7 updates NMI safe

On Tue, Aug 04, 2026 at 07:55:17AM +0900, Masami Hiramatsu wrote:
> On Mon, 3 Aug 2026 08:57:44 +0200
> Peter Zijlstra <[email protected]> wrote:
> 
> > On Sun, Aug 02, 2026 at 05:18:37PM +0900, Masami Hiramatsu (Google) wrote:
> > > From: Jinchao Wang <[email protected]>
> > > 
> > > Hardware breakpoint installation and removal run with IRQs disabled, but
> > > an NMI can still enter the same code through KGDB. The interrupted
> > > operation and the NMI can consequently claim the same slot or overwrite
> > > each other's DR7 state.
> > 
> > Is KGDB really the only way to trip this? Mostly I think we let KGDB
> > have the pieces if it does something 'funny'.
> 
> Hmm, I think even if so, this detection is better to be handled in
> hw_breakpoint layer. And I plan to use this from kprobe events,
> which is also a kind of NMI.
> 
> > 
> > > diff --git a/arch/x86/kernel/hw_breakpoint.c b/arch/x86/kernel/hw_breakpoint.c
> > > index f846c15f21ca..9ef24b55737f 100644
> > > --- a/arch/x86/kernel/hw_breakpoint.c
> > > +++ b/arch/x86/kernel/hw_breakpoint.c
> > > @@ -40,6 +40,9 @@
> > >  DEFINE_PER_CPU(unsigned long, cpu_dr7);
> > >  EXPORT_PER_CPU_SYMBOL(cpu_dr7);
> > >  
> > > +/* Sequence number of the per-CPU DR7 state. */
> > > +DEFINE_PER_CPU(unsigned int, cpu_dr7_seq);
> > 
> > > diff --git a/arch/x86/kernel/nmi.c b/arch/x86/kernel/nmi.c
> > > index 3c9f60d6ca5a..f55a0cbd5927 100644
> > > --- a/arch/x86/kernel/nmi.c
> > > +++ b/arch/x86/kernel/nmi.c
> > > @@ -532,10 +532,13 @@ enum nmi_states {
> > >  static DEFINE_PER_CPU(enum nmi_states, nmi_state);
> > >  static DEFINE_PER_CPU(unsigned long, nmi_cr2);
> > >  static DEFINE_PER_CPU(unsigned long, nmi_dr7);
> > > +static DEFINE_PER_CPU(unsigned int, nmi_dr7_seq);
> > 
> > This is weird, why have two distinct sequence numbers for dr7?
> 
> nmi_dr7_seq is for sequence number of operation, which is for
> detecting dr7 overwrite in NMI.
> nmi_dr7 is for saving the DR7.

I mean cpu_dr7_seq and nmi_dr7_seq.