Re: Got FPU related warning on Intel Quark during boot

[Ingo Molnar]

* Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:

> On Thu, Mar 10, 2016 at 6:59 AM, Borislav Petkov <bp@xxxxxxxxx> wrote:
> > On Thu, Mar 10, 2016 at 03:31:43PM +0200, Andy Shevchenko wrote:
> >> Looks like it lacks that one.
> >>
> >> # grep -i fxsr /proc/cpuinfo; echo $?
> >> 1
> >
> > Ok, so looking at where the warning comes from:
> >
> > [   14.714533] WARNING: CPU: 0 PID: 823 at arch/x86/include/asm/fpu/internal.h:163 fpu__clear+0x8c/0x160
> >
> > static inline void copy_kernel_to_fxregs(struct fxregs_state *fx)
> > {
> >         int err;
> >
> >         if (config_enabled(CONFIG_X86_32)) {
> >                 err = check_insn(fxrstor %[fx], "=m" (*fx), [fx] "m" (*fx));
> >                       ^^^^^^^^^^^^^^^^^
> >         } else {
> >
> >         ...
> >
> >         /* Copying from a kernel buffer to FPU registers should never fail: */
> >         WARN_ON_FPU(err);
> >
> >
> > and the stacktrace is pretty clear:
> >
> > flush_thread
> > |-> fpu__clear(&tsk->thread.fpu);
> >     |-> we are eager by default here:
> >
> >         if (!use_eager_fpu() || !static_cpu_has(X86_FEATURE_FPU)) {
> >                 /* FPU state will be reallocated lazily at the first use. */
> >                 fpu__drop(fpu);
> >         } else {
> >
> >                 --> we're in that branch.
> >
> >                 copy_init_fpstate_to_fpregs();
> >                 |-> copy_kernel_to_fxregs()
> >
> >
> > I think we should use FRSTOR on quark, i.e., copy_kernel_to_fregs().
> >
> > Does this untested wild guess even work?
> >
> > ---
> > diff --git a/arch/x86/kernel/fpu/core.c b/arch/x86/kernel/fpu/core.c
> > index dea8e76d60c6..bbafe5e8a1a6 100644
> > --- a/arch/x86/kernel/fpu/core.c
> > +++ b/arch/x86/kernel/fpu/core.c
> > @@ -474,8 +474,11 @@ static inline void copy_init_fpstate_to_fpregs(void)
> >  {
> >         if (use_xsave())
> >                 copy_kernel_to_xregs(&init_fpstate.xsave, -1);
> > -       else
> > +       else if (static_cpu_has(X86_FEATURE_FXSR))
> >                 copy_kernel_to_fxregs(&init_fpstate.fxsave);
> > +       else
> > +               copy_kernel_to_fregs(&init_fpstate.fsave);
> > +
> >  }
> >
> 
> This looks wrong, too:
> 
> /*
>  * Once per bootup FPU initialization sequences that will run on most x86 CPUs:
>  */
> static void __init fpu__init_system_generic(void)
> {
>     /*
>      * Set up the legacy init FPU context. (xstate init might overwrite this
>      * with a more modern format, if the CPU supports it.)
>      */
>     fpstate_init_fxstate(&init_fpstate.fxsave);  <-- wrong format on pre-FXSR CPUs

Indeed:

static inline void fpstate_init_fxstate(struct fxregs_state *fx)
{
        fx->cwd = 0x37f;
        fx->mxcsr = MXCSR_DEFAULT;
}

I assumed that the fxstate init is outside the legacy FPU context area, but they 
overlap: fx->cwd is the first two bytes, and fx->mxcsr overlaps the middle of the 
legacy area.

We do a later fpstate_init_fstate(), which does:

static inline void fpstate_init_fstate(struct fregs_state *fp)
{
        fp->cwd = 0xffff037fu;
        fp->swd = 0xffff0000u;
        fp->twd = 0xffffffffu;
        fp->fos = 0xffff0000u;
}

which accidentally overwrites the cwd bit - but AFAICS fx->mxcsw overlaps the 
first legacy FPU register?

So yes, this needs to be fixed too.

Thanks,

	Ingo