Re: [PATCH v2 0/5] mm/memfd: MFD_NOEXEC for memfd_create

[Kees Cook <keescook@chromium.org>]

On Fri, Aug 05, 2022 at 10:21:21PM +0000, jeffxu@google.com wrote:
> This v2 series MFD_NOEXEC, this series includes:
> 1> address comments in V1
> 2> add sysctl (vm.mfd_noexec) to change the default file permissions
>     of memfd_create to be non-executable.
> 
> Below are cover-level for v1:
> 
> The default file permissions on a memfd include execute bits, which
> means that such a memfd can be filled with a executable and passed to
> the exec() family of functions. This is undesirable on systems where all
> code is verified and all filesystems are intended to be mounted noexec,
> since an attacker may be able to use a memfd to load unverified code and
> execute it.

I would absolutely like to see some kind of protection here. However,
I'd like a more specific threat model. What are the cases where the X
bit has been abused (e.g.[1])? What are the cases where the X bit is
needed (e.g.[2])? With those in mind, it should be possible to draw
a clear line between the two cases. (e.g. we need to avoid a confused
deputy attack where an "unprivileged" user can pass an executable memfd
to a "privileged" user. How those privileges are defined may matter a
lot based on how memfds are being used. For example, can runc's use of
executable memfds be distinguished from an attacker's?)

> Additionally, execution via memfd is a common way to avoid scrutiny for
> malicious code, since it allows execution of a program without a file
> ever appearing on disk. This attack vector is not totally mitigated with
> this new flag, since the default memfd file permissions must remain
> executable to avoid breaking existing legitimate uses, but it should be
> possible to use other security mechanisms to prevent memfd_create calls
> without MFD_NOEXEC on systems where it is known that executable memfds
> are not necessary.

This reminds me of dealing with non-executable stacks. There ended up
being three states:

- requested to be executable (PT_GNU_STACK X)
- requested to be non-executable (PT_GNU_STACK NX)
- undefined (no PT_GNU_STACK)

The first two are clearly defined, but the third needed a lot of special
handling. For a "safe by default" world, the third should be "NX", but
old stuff depended on it being "X".

Here, we have a bit being present or not, so we only have a binary
state. I'd much rather the default be NX (no bit set) instead of making
every future (safe) user of memfd have to specify MFD_NOEXEC.

It's also easier on a filtering side to say "disallow memfd_create with
MFD_EXEC", but how do we deal with the older software?

If the default perms of memfd_create()'s exec bit is controlled by a
sysctl and the sysctl is set to "leave it executable", how does a user
create an NX memfd? (i.e. setting MFD_EXEC means "exec" and not setting
it means "exec" also.) Are two bits needed? Seems wasteful.
MFD_I_KNOW_HOW_TO_SET_EXEC | MFD_EXEC, etc...

For F_SEAL_EXEC, it seems this should imply F_SEAL_WRITE if forced
executable to avoid WX mappings (i.e. provide W^X from the start).

-Kees

[1] https://bugs.chromium.org/p/chromium/issues/list?q=type%3Dbug-security%20memfd%20escalation&can=1
[2] https://lwn.net/Articles/781013/

-- 
Kees Cook