Re: [PATCH 4.9 18/22] x86/fpu: Disable bottom halves while loading FPU registers

[Sasha Levin]

On Fri, Jul 24, 2020 at 07:44:37PM +0200, Greg Kroah-Hartman wrote:
> On Fri, Jul 24, 2020 at 07:07:06PM +0200, Jan Kiszka wrote:
> > On 28.12.18 12:52, Greg Kroah-Hartman wrote:
> > > 4.9-stable review patch.  If anyone has any objections, please let me know.
> > >
> > > ------------------
> > >
> > > From: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
> > >
> > > commit 68239654acafe6aad5a3c1dc7237e60accfebc03 upstream.
> > >
> > > The sequence
> > >
> > >    fpu->initialized = 1;		/* step A */
> > >    preempt_disable();		/* step B */
> > >    fpu__restore(fpu);
> > >    preempt_enable();
> > >
> > > in __fpu__restore_sig() is racy in regard to a context switch.
> > >
> > > For 32bit frames, __fpu__restore_sig() prepares the FPU state within
> > > fpu->state. To ensure that a context switch (switch_fpu_prepare() in
> > > particular) does not modify fpu->state it uses fpu__drop() which sets
> > > fpu->initialized to 0.
> > >
> > > After fpu->initialized is cleared, the CPU's FPU state is not saved
> > > to fpu->state during a context switch. The new state is loaded via
> > > fpu__restore(). It gets loaded into fpu->state from userland and
> > > ensured it is sane. fpu->initialized is then set to 1 in order to avoid
> > > fpu__initialize() doing anything (overwrite the new state) which is part
> > > of fpu__restore().
> > >
> > > A context switch between step A and B above would save CPU's current FPU
> > > registers to fpu->state and overwrite the newly prepared state. This
> > > looks like a tiny race window but the Kernel Test Robot reported this
> > > back in 2016 while we had lazy FPU support. Borislav Petkov made the
> > > link between that report and another patch that has been posted. Since
> > > the removal of the lazy FPU support, this race goes unnoticed because
> > > the warning has been removed.
> > >
> > > Disable bottom halves around the restore sequence to avoid the race. BH
> > > need to be disabled because BH is allowed to run (even with preemption
> > > disabled) and might invoke kernel_fpu_begin() by doing IPsec.
> > >
> > >   [ bp: massage commit message a bit. ]
> > >
> > > Signed-off-by: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
> > > Signed-off-by: Borislav Petkov <bp@xxxxxxx>
> > > Acked-by: Ingo Molnar <mingo@xxxxxxxxxx>
> > > Acked-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> > > Cc: Andy Lutomirski <luto@xxxxxxxxxx>
> > > Cc: Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx>
> > > Cc: "H. Peter Anvin" <hpa@xxxxxxxxx>
> > > Cc: "Jason A. Donenfeld" <Jason@xxxxxxxxx>
> > > Cc: kvm ML <kvm@xxxxxxxxxxxxxxx>
> > > Cc: Paolo Bonzini <pbonzini@xxxxxxxxxx>
> > > Cc: Radim KrÄmÃÅ <rkrcmar@xxxxxxxxxx>
> > > Cc: Rik van Riel <riel@xxxxxxxxxxx>
> > > Cc: stable@xxxxxxxxxxxxxxx
> > > Cc: x86-ml <x86@xxxxxxxxxx>
> > > Link: http://lkml.kernel.org/r/20181120102635.ddv3fvavxajjlfqk@xxxxxxxxxxxxx
> > > Link: https://lkml.kernel.org/r/20160226074940.GA28911@xxxxxxx
> > > Signed-off-by: Sebastian Andrzej Siewior <bigeasy@xxxxxxxxxxxxx>
> > > Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx>
> > > ---
> > >   arch/x86/kernel/fpu/signal.c |    4 ++--
> > >   1 file changed, 2 insertions(+), 2 deletions(-)
> > >
> > > --- a/arch/x86/kernel/fpu/signal.c
> > > +++ b/arch/x86/kernel/fpu/signal.c
> > > @@ -342,10 +342,10 @@ static int __fpu__restore_sig(void __use
> > >   			sanitize_restored_xstate(tsk, &env, xfeatures, fx_only);
> > >   		}
> > > +		local_bh_disable();
> > >   		fpu->fpstate_active = 1;
> > > -		preempt_disable();
> > >   		fpu__restore(fpu);
> > > -		preempt_enable();
> > > +		local_bh_enable();
> > >   		return err;
> > >   	} else {
> > >
> > >
> >
> > Any reason why the backport stopped back than at 4.9? I just debugged this
> > out of a 4.4 kernel, and it is needed there as well. I'm happy to propose a
> > backport, would just appreciate a hint if the BH protection is needed also
> > there (my case was without BH).
>
> You are asking about something we did back in 2018.  I can't remember
> what I did last week :)
>
> If you provide a backport that works, I'll be glad to take it.  The
> current patch does not apply cleanly there at all.

The conflict was due to a missing rename back in 4.4: e4a81bfcaae1
("x86/fpu: Rename fpu::fpstate_active to fpu::initialized").

I've fixed up the patch and queued it for 4.4, thanks for pointing it
out Jan!

--
Thanks,
Sasha