Re: [PATCH v4 0/7] arm64: untag user pointers passed to the kernel

[Catalin Marinas]

On Thu, Jun 28, 2018 at 12:46:11PM +0200, Luc Van Oostenryck wrote:
> On Thu, Jun 28, 2018 at 11:27:42AM +0100, Catalin Marinas wrote:
> > On Thu, Jun 28, 2018 at 08:17:59AM +0200, Luc Van Oostenryck wrote:
> > > On Wed, Jun 27, 2018 at 06:17:58PM +0100, Catalin Marinas wrote:
> > > > sparse is indeed an option. The current implementation doesn't warn on
> > > > an explicit cast from (void __user *) to (unsigned long) since that's a
> > > > valid thing in the kernel. I couldn't figure out if there's any other
> > > > __attribute__ that could be used to warn of such conversion.
> > > 
> > > sparse doesn't have such attribute but would an new option that would warn
> > > on such cast be a solution for your case?
> > 
> > I can't tell for sure whether such sparse option would be the full
> > solution but detecting explicit __user pointer casts to long is a good
> > starting point. So far this patchset pretty much relies on detecting
> > a syscall failure and trying to figure out why, patching the kernel. It
> > doesn't really scale.
> 
> OK, I'll add such an option this evening.

That's great, thanks. I think this should cover casting pointers to any
integer types, not just "unsigned long" (e.g. long long).

The only downside is that with this patchset the untagging can be done
after the conversion to ulong (get_user_pages()) as that's where the
problem was noticed. With a new sparse feature, we'd have to annotate
the conversion sites (not sure how many until we run the tool though).

> > As a side note, we have cases in the user-kernel ABI where the user
> > address type is "unsigned long": mmap() and friends. My feedback on an
> > early version of this patchset was to always require untagged pointers
> > coming from user space on such syscalls, so no need for explicit
> > untagging.
> 
> Mmmm yes.
> I tend to favor a sort of opposite approach. When we have an address
> that must not be dereferenced as-such (and sometimes when the address
> can be from both __user & __kernel space) I prefer to use a ulong
> which will force the use of the required operation before being
> able to do any sort of dereferencing and this won't need horrible
> casts with __force (it, of course, all depends on the full context).

I agree. That's what the kernel uses in functions like get_user_pages()
which take ulong as an argument. Similarly mmap() and friends don't
expect the pointer to be dereferenced, hence the ulong argument. The
interesting part that the man page (and the C library header
declaration) shows such address argument as void *. We could add a
syscall wrapper in the arch code, only that it doesn't feel consistent
with the "rule" that ulong addresses are not actually tagged pointers.

-- 
Catalin