On Sun, 9 Feb 2025 at 11:48, David Laight <david.laight.linux@xxxxxxxxx> wrote:
>
> You almost need it to be 'void masked_user_access_begin(&uaddr)'.
Maybe we just need to make it a two-stage thing, with
if (!user_access_ok(uaddr, size))
return -EFAULT;
user_read_access_begin(&uaddr);
unsafe_get_user(val1, &uaddr->one, Efault);
unsafe_get_user(val2, &uaddr->two, Efault);
user_read_access_end();
... all done ..
Efault:
user_read_access_end();
return -EFAULT;
and that would actually simplify some things: right now we have
separate versions of the user address checking (for
read/write/either): user_read_access_begin() and friends.
We still need those three versions, but now we'd only need them for
the simpler non-conditional case that doesn't have to bother about the
size.
And then if you have user address masking, user_access_ok() just
unconditionally returns true and is a no-op, while
user_read_access_begin() does the masking and actually enables user
accesses.
And if you *don't* have user address masking, user_read_access_begin()
still enables user accesses and has the required speculation
synchronization, but doesn't do any address checking, because
user_access_ok() did that (and nothing else).
That seems like it might be a reasonable compromise and fairly hard to
get wrong (*)?
Linus
(*) Obviously anybody can get anything wrong, but if you forget the
user_access_ok() entirely you're being wilful about it, and if you
forget the user_read_access_begin() the code won't work, so it seems
about as safe as it can be.
