Re: [tip: x86/fpu] x86/fpu/xstate: Define new functions for clearing fpregs and xstates

[Yu, Yu-cheng]

On 5/24/2021 9:34 AM, Andy Lutomirski wrote:
> On Sat, May 16, 2020 at 8:10 AM tip-bot2 for Fenghua Yu
> <tip-bot2@xxxxxxxxxxxxx> wrote:
> > The following commit has been merged into the x86/fpu branch of tip:
> >
> > Commit-ID:     b860eb8dce5906b14e3a7f3c771e0b3d6ef61b94
> > Gitweb:        https://git.kernel.org/tip/b860eb8dce5906b14e3a7f3c771e0b3d6ef61b94
> > Author:        Fenghua Yu <fenghua.yu@xxxxxxxxx>
> > AuthorDate:    Tue, 12 May 2020 07:54:39 -07:00
> > Committer:     Borislav Petkov <bp@xxxxxxx>
> > CommitterDate: Wed, 13 May 2020 13:41:50 +02:00
> >
> > x86/fpu/xstate: Define new functions for clearing fpregs and xstates
>
> syzbot says this is busted.  I've made no effort to identify the
> precise bug that is making syzbot complain, but:
>
> >   /*
> > - * Clear FPU registers by setting them up from
> > - * the init fpstate:
> > + * Clear FPU registers by setting them up from the init fpstate.
> > + * Caller must do fpregs_[un]lock() around it.
> >    */
> > -static inline void copy_init_fpstate_to_fpregs(void)
> > +static inline void copy_init_fpstate_to_fpregs(u64 features_mask)
> >   {
> > -       fpregs_lock();
> > -
>
>
>
> >          if (use_xsave())
> > -               copy_kernel_to_xregs(&init_fpstate.xsave, -1);
> > +               copy_kernel_to_xregs(&init_fpstate.xsave, features_mask);
> >          else if (static_cpu_has(X86_FEATURE_FXSR))
> >                  copy_kernel_to_fxregs(&init_fpstate.fxsave);
> >          else
> > @@ -307,9 +305,6 @@ static inline void copy_init_fpstate_to_fpregs(void)
> >
> >          if (boot_cpu_has(X86_FEATURE_OSPKE))
> >                  copy_init_pkru_to_fpregs();
>
> if (boot_cpu_has(X86_FEATURE_OSPKE) && (features_mask & PKRU)), perhaps?
>
> > -
> > -       fpregs_mark_activate();
> > -       fpregs_unlock();
> >   }
> >
> >   /*
> > @@ -318,18 +313,40 @@ static inline void copy_init_fpstate_to_fpregs(void)
> >    * Called by sys_execve(), by the signal handler code and by various
> >    * error paths.
> >    */
> > -void fpu__clear(struct fpu *fpu)
> > +static void fpu__clear(struct fpu *fpu, bool user_only)
> >   {
> > -       WARN_ON_FPU(fpu != &current->thread.fpu); /* Almost certainly an anomaly */
> > +       WARN_ON_FPU(fpu != &current->thread.fpu);
> >
> > -       fpu__drop(fpu);
> > +       if (!static_cpu_has(X86_FEATURE_FPU)) {
> > +               fpu__drop(fpu);
> > +               fpu__initialize(fpu);
> > +               return;
> > +       }
> >
> > -       /*
> > -        * Make sure fpstate is cleared and initialized.
> > -        */
> > -       fpu__initialize(fpu);
> > -       if (static_cpu_has(X86_FEATURE_FPU))
> > -               copy_init_fpstate_to_fpregs();
> > +       fpregs_lock();
> > +
> > +       if (user_only) {
> > +               if (!fpregs_state_valid(fpu, smp_processor_id()) &&
> > +                   xfeatures_mask_supervisor())
> > +                       copy_kernel_to_xregs(&fpu->state.xsave,
> > +                                            xfeatures_mask_supervisor());
>
> This looks correct to me.
>
> So I'm guessing that syzbot may have misattributed the problem.  But
> we definitely need to clean up the XRSTOR #GP handling before CET
> lands.

From the crash dump, the system is doing syscall_exit_to_user_mode() 
for __x64_sys_futex().  The futex syscall does not seem to modify 
xstates, but upon returning to user mode, XRSTORS gets a GP.  Can this 
be some memory corruption?  fpu__clear() is merely helping to clear the 
mess and seems to be innocent.

I also run the syz repro on my Tiger Lake machine, and it only produces 
segfaults (no Bad FPU state, etc.).

Yu-cheng