Re: [PATCH 03/10] x86/fpu: Use validate_xstate_header() to validate the xstate_header in sanitize_restored_xstate()

[Eric Biggers]

On Sun, Sep 24, 2017 at 09:02:42PM +0200, Ingo Molnar wrote:
> 
> * Eric Biggers <ebigge...@gmail.com> wrote:
> 
> > On Sun, Sep 24, 2017 at 12:59:06PM +0200, Ingo Molnar wrote:
> > > @@ -328,10 +331,8 @@ static int __fpu__restore_sig(void __user *buf, void 
> > > __user *buf_fx, int size)
> > >                   err = copy_user_to_xstate(&fpu->state.xsave, buf_fx);
> > >           } else {
> > >                   err = __copy_from_user(&fpu->state.xsave, buf_fx, 
> > > state_size);
> > > -
> > > -                 /* xcomp_bv must be 0 when using uncompacted format */
> > > -                 if (!err && fpu->state.xsave.header.xcomp_bv)
> > > -                         err = -EINVAL;
> > > +                 if (!err)
> > > +                         err = 
> > > validate_xstate_header(&fpu->state.xsave.header);
> > >           }
> > >  
> > 
> > Sorry, this is the buggy part.  The problem is that this code runs even if 
> > XSAVE
> > isn't being used --- and in that case the state size is 512 bytes or less, 
> > so
> > the state doesn't actually include the xstate_header.  So
> > validate_xstate_header() was reading out of bounds and seeing invalid 
> > values.
> > 
> > So I think we need to check use_xsave() here, but it really needs to be in 
> > the
> > earlier patch which added the check for just ->xcomp_bv ("x86/fpu: Don't let
> > userspace set bogus xcomp_bv"), not in this one.
> > 
> > As far the split of patch 2/3 into these 10 patches, it looks fine (though 
> > it
> > suddenly became a *lot* of patches!).  One nit: the subject of this one 
> > really
> > should say "__fpu__restore_sig()", not "sanitize_restored_xstate()".
> > 
> > I can send a fixed series when I have a chance.
> 
> Could you please just send the delta patch against the whole tree to fix the 
> bug? 
> I'll worry about the patch dependencies and back-merge it to the proper place.
> 

The following diff against tip/master fixes the bug.  Note: we *could* check
'use_xsave()' instead of 'state_size > offsetof(struct xregs_state, header)',
but that might be confusing in the case where we couldn't find the xstate
information in the memory layout and only copy the fxregs_state, since then we'd
actually be validating the xsave_header which was already there, which shouldn't
ever fail.

diff --git a/arch/x86/kernel/fpu/signal.c b/arch/x86/kernel/fpu/signal.c
index afe54247cf27..fb639e70048f 100644
--- a/arch/x86/kernel/fpu/signal.c
+++ b/arch/x86/kernel/fpu/signal.c
@@ -331,7 +331,8 @@ static int __fpu__restore_sig(void __user *buf, void __user 
*buf_fx, int size)
                        err = copy_user_to_xstate(&fpu->state.xsave, buf_fx);
                } else {
                        err = __copy_from_user(&fpu->state.xsave, buf_fx, 
state_size);
-                       if (!err)
+
+                       if (!err && state_size > offsetof(struct xregs_state, 
header))
                                err = 
validate_xstate_header(&fpu->state.xsave.header);
                }