Re: [PATCH 2/3] x86/mm: introduce __PAGE_KERNEL_GLOBAL

[Nadav Amit]

Dave Hansen <dave.hansen@xxxxxxxxxxxxxxx> wrote:

> On 02/16/2018 09:47 AM, Nadav Amit wrote:
>>> But, this also means that we now get *no* opportunity to use
>>> global pages with PTI, even for data which is shared such as the
>>> cpu_entry_area and entry/exit text.
>> 
>> Doesnât this patch change the kernel behavior when the ânoptiâ
>> parameter is used?
> 
> I don't think so.  It takes the "nopti" behavior and effectively makes
> it apply everywhere.  So it changes the PTI behavior, not the "nopti"
> behavior.
> 
> Maybe it would help to quote the code that you think does this instead
> of the description. :)

Sorry, I thought it is obvious - so I must be missing something.

> +#ifdef CONFIG_PAGE_TABLE_ISOLATION
> +#define __PAGE_KERNEL_GLOBAL		0
> +#else
> +#define __PAGE_KERNEL_GLOBAL		_PAGE_GLOBAL
> +#endif
...
> --- a/arch/x86/mm/pageattr.c~kpti-no-global-for-kernel-mappings	2018-02-13 15:17:56.148210060 -0800
> +++ b/arch/x86/mm/pageattr.c	2018-02-13 15:17:56.153210060 -0800
> @@ -593,7 +593,8 @@ try_preserve_large_page(pte_t *kpte, uns
> 	 * different bit positions in the two formats.
> 	 */
> 	req_prot = pgprot_4k_2_large(req_prot);
> -	req_prot = pgprot_set_on_present(req_prot, _PAGE_GLOBAL | _PAGE_PSE);
> +	req_prot = pgprot_set_on_present(req_prot,
> +			__PAGE_KERNEL_GLOBAL | _PAGE_PSE);
> 	req_prot = canon_pgprot(req_prot);

From these chunks, it seems to me as req_prot will not have the global bit
on when ânoptiâ parameter is provided. What am I missing?