Re: Is _PAGE_PROTNONE set only for user mappings?

[Sean Christopherson <seanjc@google.com>]

On Sun, May 22, 2022, Hyeonggon Yoo wrote:
> On Mon, May 16, 2022 at 07:04:32AM -0700, Dave Hansen wrote:
> > I was thinking of something more along the lines of taking the
> > set_memory.c code and ensuring that it never sets (or even observes)
> > _PAGE_BIT_GLOBAL on a _PAGE_USER mapping.
> 
> Yeah that would be a bit more explicit solution.
> 
> > There was also a question of
> > if set_memory.c is ever used on userspace mappings.  It would be good to
> > validate whether it's possible in-tree today and if not, enforce that
> > _PAGE_USER PTEs should never even be observed with set_memory.c.
> 
> Simply adding dump_stack() tells me my kernel on my machine does not use
> set_memory.c for userspace mappings but Hmm I'll take a look.

vc_slow_virt_to_phys() uses lookup_address_in_pgd() with user mappings, but that
code is all but guaranteed to be buggy, e.g. doesn't guard against concurrent
modifications to user mappings.

show_fault_oops() can also call into lookup_address_in_pgd() with a user mapping,
though at that point the kernel has bigger problems since it's executing from user
memory.

And up until commits 44187235cbcc ("KVM: x86/mmu: fix potential races when walking
host page table") and 643d95aac59a ("Revert "x86/mm: Introduce lookup_address_in_mm()""),
KVM had a similar bug.

Generally speaking, set_memory.c is not equipped to play nice with user mappings.
It mostly "works", but there are races galore.  IMO, hardening set_memory.c to scream
if it's handed a user address or encounters _PAGE_USER PTEs would be a very good thing.