Re: [patch cr 2/4] [RFC] syscalls, x86: Add __NR_kcmp syscall v7

[Cyrill Gorcunov <gorcunov@openvz.org>]

On Mon, Jan 30, 2012 at 12:58:12PM -0700, Jonathan Corbet wrote:
> Just a couple of silly little things that came to mind while I was looking
> at the code...
> 
> > +/*
> > + * We don't expose real in-memory order of objects for security
> > + * reasons, still the comparision results should be suitable for
> > + * sorting. Thus, we obfuscate kernel pointers values (using random
> > + * cookies obtaned at early boot stage) and compare the production
> > + * instead.
> > + */
> > +static unsigned long cookies[KCMP_TYPES][2] __read_mostly;
> > +
> > +static long kptr_obfuscate(long v, int type)
> > +{
> > +	return (v ^ cookies[type][0]) * cookies[type][1];
> > +}
> 
> I don't understand the purpose of this at all.  Obfuscation will cause a
> random shuffling in the ordering of the pointers - it's intended to - so
> how is the result "suitable for sorting"?  More to the point, is there
> ever a time when a user of this will care about some contrived ordering
> value?  It seems like equality is all that really matters.
> 

It won't be completely random shuffling but rather re-ordering in some
new order, which means the results might be passed to qsort or anything.
And yes, in c/r we need at least this "re-ordered" order which will help
to figure out shared file descriptors in case of huge number of files opened.

> > +
> > +/*
> > + * 0 - equal
> > + * 1 - less than
> > + * 2 - greater than
> > + * 3 - not equal but ordering unavailable (reserved for future)
> > + */
> > +static int kcmp_ptr(void *v1, void *v2, enum kcmp_type type)
> > +{
> > +	long ret;
> > +
> > +	ret = kptr_obfuscate((long)v1, type) - kptr_obfuscate((long)v2, type);
> > +
> > +	return (ret < 0) | ((ret > 0) << 1);
> > +}
> 
> That's a cute trick, but do we know that every compiler that will ever see
> this code will use 1 for a true integer comparison?  Simply spelling it
> out with an if statement might be more robust, just as efficient, and, at
> the same time, easier for others to understand.

Well, I believe if this become true, and (ret < 0) wont emit 1 -- the
number of places in kernel will be broken as well (for example see
math_div() function). But of course I don't insist and can rewrite
this code in straight fashion if needed.

	Cyrill