Re: [PATCH v4] [x86] detect and report lack of NX protections

["H. Peter Anvin" <hpa@zytor.com>]

On 11/10/2009 11:43 AM, Kees Cook wrote:
> 
> This is fun.  CONFIG_X86_PAE isn't defined for 64-bit, and using
> cpu_has_pae on 64-bit is considered a bug.  :)
> 

Yeah, it's somewhat obnoxious.  This stuff is a result of the 32- and
64-bit code evolving separately for too long.  All of this could and
should be cleaned up, but it takes a long time.

Either way, you can use the explicit form:

	boot_cpu_has(X86_FEATURE_PAE)

just fine, on any platform.  However, the only case for which this can
be false is for the non-PAE kernel, since the PAE kernels (32 or 64
bits) cannot boot without it.  I have personally never liked the
cpu_has_* shorthand macros, but they're occasionally useful for things
that have to be handled specially on 64 bits.  Unfortunately they have
spread and people seem to think they're the only way.

> Here is the matrix of what I want to see reported about NX at boot time.
> How do you recommend this be implemented?
> 
> kernel  cpu -> |              CPU has PAE              |  CPU lacks PAE  |
>    |           |       CPU has NX  | CPU lacks NX      |                 |
>    V           +-------------------+-------------------+-----------------+
> 32-bit non-PAE | missing in kernel | missing in kernel |    no message   |
>                +-------------------+-------------------+-----------------+
> 32-bit PAE     | active *          | missing in CPU    |    no message   |
>                +-------------------+-------------------+-----------------+
> 64-bit         | active            | missing in CPU    |    impossible   |
>                +-------------------+-------------------+-----------------+
> The box with the "*" is the only message currently reported by the kernel.

The last column should actually be "no message", "impossible", "impossible".

I also think "missing in kernel" is misleading in the 32-bit non-PAE,
no-NX case (as it would imply that another kernel could do something),
and I *really* fail to see why it is in any way different from the "CPU
lacks PAE" case -- which also means no NX.  "Unavailable in CPU" seems
to beat everything.

So the logic that makes sense would be:

if (!cpu_has_nx) {
	/* If the CPU can't do it... */
	printk(KERN_INFO "cpu: NX protection unavailable in CPU\n");
} else  {
#if defined(CONFIG_X86_32) && !defined(CONFIG_X86_PAE)
	/* Non-PAE kernel: NX unavailable */
	printk(KERN_NOTICE "cpu: NX protection missing in kernel\n");
#else
	printk(KERN_INFO "cpu: NX protection active\n");
#endif
}