Re: [PATCH v5 06/12] S.A.R.A.: WX protection

[Kees Cook]

On Sun, Jul 07, 2019 at 05:49:35PM +0200, Salvatore Mesoraca wrote:
> Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> >
> > On Sat, Jul 06, 2019 at 12:54:47PM +0200, Salvatore Mesoraca wrote:
> >
> > > +#define sara_warn_or_return(err, msg) do {           \
> > > +     if ((sara_wxp_flags & SARA_WXP_VERBOSE))        \
> > > +             pr_wxp(msg);                            \
> > > +     if (!(sara_wxp_flags & SARA_WXP_COMPLAIN))      \
> > > +             return -err;                            \
> > > +} while (0)
> > > +
> > > +#define sara_warn_or_goto(label, msg) do {           \
> > > +     if ((sara_wxp_flags & SARA_WXP_VERBOSE))        \
> > > +             pr_wxp(msg);                            \
> > > +     if (!(sara_wxp_flags & SARA_WXP_COMPLAIN))      \
> > > +             goto label;                             \
> > > +} while (0)
> >
> > No.  This kind of "style" has no place in the kernel.
> >
> > Don't hide control flow.  It's nasty enough to reviewers,
> > but it's pure hell on anyone who strays into your code while
> > chasing a bug or doing general code audit.  In effect, you
> > are creating your oh-so-private C dialect and assuming that
> > everyone who ever looks at your code will start with learning
> > that *AND* incorporating it into their mental C parser.
> > I'm sorry, but you are not that important.
> >
> > If it looks like a function call, a casual reader will assume
> > that this is exactly what it is.  And when one is scanning
> > through a function (e.g. to tell if handling of some kind
> > of refcounts is correct, with twentieth grep through the
> > tree having brought something in your code into the view),
> > the last thing one wants is to switch between the area-specific
> > C dialects.  Simply because looking at yours is sandwiched
> > between digging through some crap in drivers/target/ and that
> > weird thing in kernel/tracing/, hopefully staying limited
> > to 20 seconds of glancing through several functions in your
> > code.
> >
> > Don't Do That.  Really.
> 
> I understand your concerns.
> The first version of SARA didn't use these macros,
> they were added because I was asked[1] to do so.
> 
> I have absolutely no problems in reverting this change.
> I just want to make sure that there is agreement on this matter.
> Maybe Kees can clarify his stance.
> 
> Thank you for your suggestions.
> 
> [1] https://lkml.kernel.org/r/CAGXu5jJuQx2qOt_aDqDQDcqGOZ5kmr5rQ9Zjv=MRRCJ65ERfGw@xxxxxxxxxxxxxx

I just didn't like how difficult it was to review the repeated checking.
I thought then (and still think now) it's worth the unusual style to
improve the immediate readability. Obviously Al disagrees. I'm not
against dropping my suggestion; it's just a pain to review it and it
seems like an area that would be highly prone to subtle typos. Perhaps
some middle ground:

#define sara_warn(msg)	({				\
		if ((sara_wxp_flags & SARA_WXP_VERBOSE))	\
			pr_wxp(msg);				\
		!(sara_wxp_flags & SARA_WXP_COMPLAIN);		\
	})

...

	if (unlikely(sara_wxp_flags & SARA_WXP_WXORX &&
		     vm_flags & VM_WRITE &&
		     vm_flags & VM_EXEC &&
		     sara_warn("W^X")))
		return -EPERM;

that way the copy/pasting isn't present but the control flow is visible?

-- 
Kees Cook