[net-next,0/4] sockptr: Change sockptr_t to be a struct

Message

David Laight Dec. 25, 2023, 9:46 a.m. UTC

The original commit for sockptr_t tried to use the pointer value
to determine whether a pointer was user or kernel.
This can't work on some architectures and was buggy on x86.
So the is_kernel discriminator was added after the union of pointers.

However this is still open to misuse and accidents.
Replace the union with a struct and remove the is_kernel member.
The user and kernel values are now in different places.
The structure size doesn't change - it was always padded out to 'two pointers'.

The only functional difference is that NULL pointers are always 'user'.
So dereferencing will (usually) fault in copy_from_user() rather than
panic if supplied as a kernel address.

Simple driver code that uses kernel sockets still works.
I've not tested bpf - but that should work unless it is breaking
the rules.

The first three patches just change the code to use the helpers
from sockptr.h.
The functional change is in the fourth patch.


David Laight (4):
  Use sockptr_is_kernel() instead of testing is_kernel.
  Use bpfptr_is_kernel() instead of checking the is_kernel member.
  Use the sockptr_t helpers.
  Change sockptr_t to be a struct of a kernel and user pointer.

 include/linux/bpfptr.h   | 10 ++++------
 include/linux/sockptr.h  | 15 +++++----------
 kernel/bpf/bpf_iter.c    |  2 +-
 kernel/bpf/btf.c         |  2 +-
 kernel/bpf/syscall.c     | 12 ++++++------
 kernel/bpf/verifier.c    | 10 +++++-----
 net/ipv4/ip_sockglue.c   |  2 +-
 net/ipv6/ipv6_sockglue.c |  2 +-
 net/socket.c             |  2 +-
 9 files changed, 25 insertions(+), 32 deletions(-)

Comments

Jakub Kicinski Jan. 2, 2024, 10:32 p.m. UTC | #1

On Mon, 25 Dec 2023 09:46:23 +0000 David Laight wrote:
> The original commit for sockptr_t tried to use the pointer value
> to determine whether a pointer was user or kernel.
> This can't work on some architectures and was buggy on x86.
> So the is_kernel discriminator was added after the union of pointers.
> 
> However this is still open to misuse and accidents.
> Replace the union with a struct and remove the is_kernel member.
> The user and kernel values are now in different places.
> The structure size doesn't change - it was always padded out to 'two pointers'.
> 
> The only functional difference is that NULL pointers are always 'user'.
> So dereferencing will (usually) fault in copy_from_user() rather than
> panic if supplied as a kernel address.
> 
> Simple driver code that uses kernel sockets still works.
> I've not tested bpf - but that should work unless it is breaking
> the rules.

LGTM, but we either need acks from bpf folks or route this via
bpf-next. So please repost and CC bpf@ on the whole series.