Re: [PATCH] x86/mm: Do not verify W^X at boot up

[Steven Rostedt]

On Mon, 24 Oct 2022 09:14:45 -0700
Dave Hansen <dave.hansen@xxxxxxxxx> wrote:

> On 10/24/22 08:45, Steven Rostedt wrote:
> > --- a/arch/x86/mm/pat/set_memory.c
> > +++ b/arch/x86/mm/pat/set_memory.c
> > @@ -587,6 +587,10 @@ static inline pgprot_t verify_rwx(pgprot_t old, pgprot_t new, unsigned long star
> >  {
> >  	unsigned long end;
> >  
> > +	/* Kernel text is rw at boot up */
> > +	if (system_state == SYSTEM_BOOTING)
> > +		return new;  
> 
> Hi Steven,
> 
> Thanks for the report and the patch.  That seems reasonable, but I'm a
> bit worried that it opens up a big hole (boot time) when a W+X mapping
> could be created *anywhere*.
> 
> Could we restrict this bypass to *only* kernel text addresses during
> boot?  Maybe something like this:
> 
> 	if ((system_state == SYSTEM_BOOTING) &&
> 	    __kernel_text_address(start))
> 		return new;

Actually, that brings back the warning, as ftrace creates a trampoline, but
text_poke() will still use memcpy on it at early boot up.

The trampolines are set to ro at the end of boot up by:

59566b0b622e3 ("x86/ftrace: Have ftrace trampolines turn read-only at the end of system boot up")

Which was added because of text_poke() doing the memcpy().

> 
> That would be safe because we know that kernel_text_address() addresses
> will be made read-only by the time userspace shows up and that
> is_kernel_inittext() addresses will be freed.
> 
> Long-term, I wonder if we could teach the early patching code that it
> can't just use memcpy().
> 

Maybe.

-- Steve