Those interested in getting Rust into Linux-kernel device drivers sooner rather than later should look at the short-term recommendations, while those interested in extending Rust's (and, for that matter, C's) concurrency capabilities might be more interested in the long-term recommendations.
Short-Term RecommendationsThe goal here is to allow the Rust language considerable memory-model flexibility while also providing Rust considerable freedom in what it might be used for within the Linux kernel. The recommendations are as follows:
- Provide wrappers around the existing Linux kernel's locking, MMIO, I/O-barrier, and I/O-access primitives. If I was doing this work, I would add wrappers incrementally as they were actually needed, but I freely admit that there are benefits to providing a full set of wrappers from the get-go.
- Either wrapper READ_ONCE() or be careful to take Alpha's, Itanium's, and ARMv8's requirements into account as needed. The situation with WRITE_ONCE() appears more straightforward. The same considerations apply to the atomic_read() and atomic_set() family of primitives.
- Atomic read-modify-write primitives should be made available to Rust programs via wrappers around C code. But who knows? Maybe it is once again time to try out intrinsics. Again, if I was doing the work, I would add wrappers/implementations incrementally.
- Although there has been significant discussion surrounding how sequence locking and RCU might be handled by Rust code, more work appears to be needed. For one thing, it is not clear that people proposing solutions are aware of the wide range of Linux-kernel use cases for these primitives. In the meantime, I recommend keeping direct use of sequence locking and RCU in C code, and providing Rust wrappers for the resulting higher-level APIs. In this case, it might be wise to make good use of compiler directives in order to limit Rust's ability to apply code-motion optimizations. However, if you need sequence-locking or RCU Rust-language wrappers, there are a number that have been proposed. (Except that you should first take another look or three at wrappering higher-level APIs. Yes, APIs are hard, but there are huge benefits to proper APIs!)
- Control dependencies should be confined to C code. If needed, higher-level APIs whose C-language implementations require control dependencies can be wrappered for Rust use. But my best guess is that it will be some time before Rust code needs control dependencies.
Taking this approach should avoid memory-model-mismatch issues between Rust and C code in the Linux kernel. And discussions indicate that much (and maybe all) of the wrappering work called out in the first three items above has already been done.
Of course, situations that do not fit this set of recommendations can be addressed on a case-by-case basis. I would of course be happy to help. Specifically, in my role as lead Linux-kernel RCU maintainer:
- My first reaction to submission of Rust wrappers for the RCU API will be to ask hard questions about the possibility of higher-level APIs.
- If higher-level APIs are infeasible, I will look carefully at which RCU use cases are supported by the proposed wrappering. I am working to improve documentation of the range of RCU use cases in order to help submitters to do this as well. (This work will likely be helpful elsewhere as well.)
- Again, if higher-level APIs are infeasible, I will look carefully at how the Rust wrappers are helping to find bugs. This will clearly require me to continue learning Rust, as this requires me to have a detailed view of Rust's ownership mechanisms and how things like reference-counting use cases work around limitations in these mechanisms.
This procedure should help buy the time required for me to learn more about the Rust language and for the Rust community to learn more about RCU and its many use cases.
Long-Term RecomendationsThis section takes a more utopian view. What would a perfect Rust sequence-locking implementation/wrapper look like? Here are some off-the-cuff desiderata, none of which are met by the current C-code Linux-kernel implementation:
- Use of quantities computed from sequence-lock-protected variables in a failed reader should result in a warning. But please note that reliably associating variables with sequence locks may not be easy. One approach suggested in response to this series is to supply a closure containing the read-side critical section, thus restricting such leakage to (unsafe?) side effects.
- Improper access to variables not protected by the sequence lock should result in a warning. But please note that it is quite difficult to define "improper" in this context, let alone detect it in real code.
- Data races in failed sequence-lock readers should not cause failures. But please note that this is extremely difficult in general if the data races involve non-volatile C-language accesses in the reader. For example, the compiler would be within its rights to refetch the value after the old value had been checked. (This is why the sequence-locking post suggests marked accesses to sequence-lock-protected variables.)
- Data races involving sequence-locking updaters are detected, for example, via KCSAN.
As noted elsewhere, use of a wrapper around the existing C-language implementation allows C and Rust to use the same sequence lock. This might or might not prove to be important.
Similarly, what would a perfect Rust RCU wrapper look like? Again, here are some off-the-cuff desiderata, similarly unmet by existing C-code Linux-kernel implementations:
- Make call_rcu() and friends cause the specified object to make an end-of-grace-period transition in other threads' ownership from readers to unowned. Again, not an immediate transition at the time of call_rcu() invocation, but rather a deferred transition that takes place at the end of some future grace period.
- Some Rust notion of type safety would be useful in slab caches flagged as SLAB_TYPESAFE_BY_RCU.
- Some Rust notion of existence guarantee would be useful for RCU readers.
- Detect pointers to RCU-protected objects being improperly leaked from RCU read-side critical sections, where proper leakage is possible via locks and reference counters. Perhaps an emphasis on closures is at least part of the answer.
- Detect mishandling of dependency-carrying pointers returned by rcu_dereference() and friends. Or perhaps introduce some marking such pointers to that the compiler will avoid breaking the ordering. See the address/data dependency post for a list of C++ working papers moving towards this goal.
There has recently been some work attempting to make the C compiler understand control dependencies. Perhaps Rust could make something useful happen in this area, perhaps by providing markings that allow the compiler to associate the reads with the corresponding control-dependent writes.
Perhaps Rust can better understand more of the ownership schemes that are used within the Linux kernel.
RationalePlease note that middle-term approaches are likely to be useful, for example, those that simply provide wrappers around the C-language RCU and sequence-locking implementations. However, such approaches also carry risks, especially during that time when the intersections of the set of people deeply understanding Rust with the set of people deeply understanding RCU and sequence locking is the empty set. For example, one risk that became apparent during the effort to add RCU to the C++ standard is that people new to RCU and sequence locking will latch onto the first use case that they encounter and focus solely on that use case. This has also been a recurring issue for concurrency experts who encounter sequence locking and RCU for the first time. This of course means that I need to do a better job of documenting RCU, its use cases, and the relationships between them.
This is not to say that per-use-case work is pointless. In fact, such work can be extremely valuable, if nothing else, in helping to build understanding of the overall problem. It is also quite possible that current RCU and sequence-locking use cases will resemble those of the future in the same way that the venerable "while" loop resembles the wide panoply of interator-like constructs found not only in modern languages, including those written in C in the Linux kernel, in other words, perhaps Rust will focus on specific fearless-concurrency-friendly RCU/sequence-locking use cases. Except that the Linux kernel still contains "while" loops, as does a great deal of other software, which suggests that the low-level RCU and sequence-locking APIs will always be required. There will also be questions as to whether a given new-age use case is there to help developers using RCU and sequence locking on the one hand or whether its main purpose is instead to work around a shortcoming in Rust on the other. "This should be good clean fun!" ;-)
Experience indicates that it will take significant time to sort out all of these issues. We should therefore make this time available by proceeding initially as described in the short-term recommendations.
Criteria for judging proposals include safety, performance, scalability, build time, development cost, maintenance effort, and so on, not necessarily in that order.
HistoryOctober 18, 2021: Add "Rationale" section.
October 20, 2021: Add a few expansions and clarifications.
October 21, 2021: RCU-specific recommendations.