Re: [PATCH] x86/mm/cpa: set PAGE_KERNEL in __set_pages_p()

["Edgecombe, Rick P" <rick.p.edgecombe@intel.com>]

On Fri, 2022-05-06 at 14:19 +0900, Hyeonggon Yoo wrote:
> __set_pages_np() not only clears _PAGE_PRESENT and _PAGE_RW, but also
> clears _PAGE_GLOBAL to avoid confusing _PAGE_GLOBAL as _PAGE_PROTNONE
> when the PTE is not present.
> 
> Common usage for __set_pages_p() is to call it after
> __set_pages_np().
> Therefore calling __set_pages_p() after __set_pages_np() clears
> _PAGE_GLOBAL, making it unable to globally shared in TLB.
> 
> As they are called by set_direct_map_{invalid,default}_noflush(),
> pages in direct map cannot be globally shared in TLB after being used
> by
> vmalloc, secretmem, and hibernation.
> 
> So set PAGE_KERNEL isntead of __pgprot(_PAGE_PRESENT | _PAGE_RW) in
> __set_pages_p().

Nice find. I think we can't always set PAGE_KERNEL also because of the
PTI code. It sometimes wants the direct map to be non global.

Maybe something like this?
(_PAGE_PRESENT | _PAGE_RW | _PAGE_GLOBAL) & __default_kernel_pte_mask

That would add back in a little of the "default global" behavior that
was removed in d1440b2, but I think it should be ok because it is
limited to the direct map. Otherwise, I wonder if the existing global
clearing logic is really needed.