"Using Rust for kernel development": Memory model

["Paul E. McKenney" <paulmck@kernel.org>]

Hello!

This email is in response to the Rust memory-model discussion at
Maintainers Summit (https://lwn.net/Articles/870555/).

My blog series entitled "So You Want to Rust the Linux Kernel?" [1] is now
feature complete, and a big "thank you" for all the great feedback that
this series received.  I have recommendations at "TL;DR: Memory-Model
Recommendations for Rusting the Linux Kernel" [2], and this email is
therefore TL;DR(TL;DR).

Given the Rust-for-Linux's focus on device drivers, the Linux-kernel
features requiring special Linux-kernel-memory-model features can be
avoided within Rust code.  For example, code using RCU, sequence locking,
or control dependencies can remain written in C, and higher-level APIs
based on that code can be exported to Rust in manner consistent with
Rust's current ownership models.

This approach reasonably straightforwardly accommodates the more likely
short-term choices for the Rust memory model, which would presumably be
the C/C++ memory model or some stronger subset thereof, for example, one
that excludes consume and relaxed accesses.  If the Rust community chooses
a less mainstream memory model, the code that makes C-code functionality
available to Rust code would need to take up any slack.  For example,
memory barriers might need to be inserted into this wrapper code.

Longer term, I hope that the core Rust community will become interested
in supporting modern techniques, and to that end I have suggested some
longer-term goals in the TL;DR post.

There is already plenty of interest in modern techniques within the
greater Rust community, and number of people produced prototype wrappers
for various sequence-locking and RCU use cases.  I am grateful to all
who took on this challenge.

However, arriving at good wrappers requires a sufficient understanding
of Rust to be combined with sufficient knowledge of the Linux kernel's
wide variety of sequence-locking and RCU use cases, and unfortunately
this combining seems to be some ways off [3].  It therefore makes sense to
defer the need for such wrappers in order to allow time for this diffusion
of knowledge to take place.  So, in the near term, if a Rust-code project
were to request direct access to RCU APIs, I would instead ask them to
create higher-level APIs so that the RCU APIs would remain within C code.
I am also working to better document the wide range of RCU use cases
that are present in the Linux kernel, which I hope will speed up the
process of working out what a Rust-language RCU API should look like.

Of course, the corresponding choices for sequence locking are in the
capable hands of the relevant maintainers, who are CCed.

Thoughts?

						Thanx, Paul

[1] https://paulmck.livejournal.com/62436.html

[2] https://paulmck.livejournal.com/65341.html

[3] This process took four years in the C/C++ standards committees,
    but perhaps things will go faster with the Rust community.