Re: GCC bug ? Re: [PATCH v2 10/10] powerpc/32s: Implement Kernel Userspace Access Protection

[Christophe Leroy <christophe.leroy@c-s.fr>]

Le 21/01/2020 à 20:55, Segher Boessenkool a écrit :
> On Tue, Jan 21, 2020 at 05:22:32PM +0000, Christophe Leroy wrote:
>> g1() should return 3, not 5.
> 
> What makes you say that?

What makes me say that is that NULL is obviously a constant pointer and 
I think we are all expecting gcc to see it as a constant during kernel 
build, ie at -O2

> 
> "A return of 0 does not indicate that the
>   value is _not_ a constant, but merely that GCC cannot prove it is a
>   constant with the specified value of the '-O' option."
> 
> (And the rules it uses for this are *not* the same as C "constant
> expressions" or C "integer constant expression" or C "arithmetic
> constant expression" or anything like that -- which should be already
> obvious from that it changes with different -Ox).
> 
> You can use builtin_constant_p to have the compiler do something better
> if the compiler feels like it, but not anything more.  Often people
> want stronger guarantees, but when they see how much less often it then
> returns "true", they do not want that either.
> 

in asm/book3s/64/kup-radix.h we have:

static inline void allow_user_access(void __user *to, const void __user 
*from,
				     unsigned long size)
{
	// This is written so we can resolve to a single case at build time
	if (__builtin_constant_p(to) && to == NULL)
		set_kuap(AMR_KUAP_BLOCK_WRITE);
	else if (__builtin_constant_p(from) && from == NULL)
		set_kuap(AMR_KUAP_BLOCK_READ);
	else
		set_kuap(0);
}

and in asm/kup.h we have:

static inline void allow_read_from_user(const void __user *from, 
unsigned long size)
{
	allow_user_access(NULL, from, size);
}

static inline void allow_write_to_user(void __user *to, unsigned long size)
{
	allow_user_access(to, NULL, size);
}


If GCC doesn't see NULL as a constant, then the above doesn't work as 
expected.

What's surprising and frustrating is that if you remove the 
__builtin_constant_p() and only leave the NULL check, then GCC sees it 
as a constant and drops the other leg.

So if we remove the __builtin_constant_p(to) and leave only the (to == 
NULL), it will work as expected for allow_read_from_user(). But for the 
others where (to) is not a constant, the NULL test will remain together 
with the associated leg.

Christophe