Re: [PATCH 8/9] x86: kvm_hv_set_msr(): use __put_user() instead of 32bit __clear_user()

[Linus Torvalds]

On Sat, May 30, 2020 at 9:20 AM Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
>
> Yes, the access_ok is done in __kvm_set_memory_region and gfn_to_hva()
> returns a page-aligned address so it's obviously ok for a u32.

It's not that it's "obviously ok for an u32".

It is _not_ obviously ok for a user address. There's actually no
access_ok() done in the lookup path at all, and what gfn_to_hva()
actually ends up doing in the end is __gfn_to_hva_memslot(), which has
zero overflow protection at all, and just does

        slot->userspace_addr + (gfn - slot->base_gfn) * PAGE_SIZE;

without us having _ever_ checked that 'gfn' parameter.

Yes, at some point in the very very distant past,
__kvm_set_memory_region() has validated
mem->{userspace_addr,memory_size}.  But even that validation is
actually very questionable, since it's not even done for all of the
memory slots, only the "user" ones.

So if at any point we have a non-user slot, of it at any point the gfn
thing was mis-calculated and {over,under}flows, there are no
protections what-so-ever.

In other words, it really looks like kvm is entirely dependent on
magic and luck and a desperate hope that there are no other bugs to
keep the end result as a user address.

Because if _any_ bug or oversight in that kvm_memory_slot handling
ever happens, you end up with random crap.

So no. I disagree. There is absolutely nothing "obviously ok" about
any of that kvm code. Quite the reverse.

I'd argue that it's very much obviously *NOT* ok, even while it might
just happen to work.

That double underscore needs to go away. It's either actively buggy
right now and I see no proof it isn't, or it's a bug just waiting to
happen in the future.

               Linus