Re: [PATCH v12 01/11] x86: text_poke() may access uninitialized struct pages

[Peter Zijlstra]

On Mon, Jun 25, 2018 at 10:39:24AM +0200, Thomas Gleixner wrote:
> On Mon, 25 Jun 2018, Peter Zijlstra wrote:
> > On Thu, Jun 21, 2018 at 05:25:08PM -0400, Pavel Tatashin wrote:
> > > -void *text_poke(void *addr, const void *opcode, size_t len)
> > > +void __ref *text_poke(void *addr, const void *opcode, size_t len)
> > >  {
> > >  	unsigned long flags;
> > >  	char *vaddr;
> > >  	struct page *pages[2];
> > >  	int i;
> > >  
> > > +	/* While boot memory allocator is runnig we cannot use struct
> > 
> > Broken comment style..
> > 
> > > +	 * pages as they are not yet initialized. However, we also know
> > > +	 * that this is early in boot, and it is safe to fallback to
> > > +	 * text_poke_early.
> > > +	 */
> > > +	if (unlikely(!after_bootmem))
> > > +		return text_poke_early(addr, opcode, len);
> > 
> > I'm not entirely sure this is right.. Because not only do we need the
> > whole fixmap stuff working, we also need #DB and the IPI handlers set-up
> > and working.
> 
> IPI? That's early UP boot why would you need an IPI?

Because the way this is called is from __jump_label_transform() ->
text_poke_bp() -> text_poke() -> text_poke_early().

And if you look at text_poke_bp(), you'll note it relies on #DB and
on_each_cpu() IPIs.

Which is of course exactly the reason I avoided text_poke_bp() entirely
in the first place.