Can the Kernel Concurrency Sanitizer Own Rust Code?

Given the data-race-freedom guarantees of Rust's non-unsafe code, one might reasonably argue that there is no point in the Kernel Concurrency Sanitizer (KCSAN) analyzing such code. However, the Linux kernel is going to need unsafe Rust code, and although there has been significant progress in formal verification of such code, one of the leading researchers in this area says “we hope to eventually develop formal methods that can be put directly in the hands of programmers”. We would be wise to take careful note of the word “eventually”. Furthermore, even given unanticipated universal acclamation of Rust within the Linux kernel community combined with equally unanticipated advances in C-to-Rust translation capabilities, a significant fraction of the existing tens of millions of lines of Linux-kernel C code will persist for some time to come. Both the unsafe Rust code and the C code can interfere with Rust non-unsafe code, and furthermore there are special cases where safe code can violate unsafe code's assumptions. Therefore, run-time analysis of Rust code (safe code included) is likely to be able to find issues that compile-time analysis cannot.

As a result, Rust seems unlikely to render KCSAN obsolete any time soon. This suggests that Rust code should include KCSAN instrumentation, both via the compiler and via atomic and volatile primitives, as is currently the case for the C compilers and primitives currently in use. The public KCSAN interface is as follows:

1. __kcsan_check_access(): This function makes a relevant memory access known to KCSAN. It takes a pointer to the object being accessed, the size of the object, and an access type (for example, zero for read, KCSAN_ACCESS_WRITE for write, and KCSAN_ACCESS_ATOMIC for atomic read-modify-write).
   
2. The __tsan_{read,write}{1,2,4,8}() family of functions serve the same purpose as __kcsan_check_access(), but the size and access type are implicit in the function name instead of being passed as arguments, which can provide better performance. This family of functions is used by KCSAN-enabled compilers.
   
3. ASSERT_EXCLUSIVE_ACCESS() is invoked from C code and causes KCSAN to complain if there is a concurrent access to the specified object. A companion ASSERT_EXCLUSIVE_ACCESS_SCOPED() expands the scope of the concurrent-access complaint to the full extent of the compound statement invoking this macro.
   
4. ASSERT_EXCLUSIVE_WRITER() is invoked from C code and causes KCSAN to complain if there is a concurrent write to the specified object. A companion ASSERT_EXCLUSIVE_WRITER_SCOPED() expands the scope of the concurrent-writer complaint to the full extent of the compound statement invoking this macro.
   
5. ASSERT_EXCLUSIVE_BITS() is similar to ASSERT_EXCLUSIVE_WRITER(), but focuses KCSAN's attention on changes to bits set in the mask passed as this macro's second argument.
   

These last three categories of interface members are designed to be directly invoked in order to check concurrency designs. See for example the direct call to ASSERT_EXCLUSIVE_WRITER() from the rcu_cpu_starting() function in Linux-kernel RCU. It would of course be quite useful for these interface members to be available from within Rust code.

KCSAN also provides a data_race() primitive that suppresses data-race diagnostics, but allows the compiler free rein on optimization. This is used in the Linux kernel for things like infrequent diagnostics that are not part of the overall concurrency design.

KCSAN has helped me fix a number of embarrassing bugs in Linux-kernel RCU that might otherwise have inconvenience end users, so they just might be helpful to people introducing Rust code into the Linux kernel. It is therefore quite encouraging that Marco Elver (KCSAN lead developer and maintainer) points out off-list that the rustc compiler already has sanitizer support, which should suffice not only for KCSAN, but for the equally helpful Kernel Address Sanitizer (KASAN) as well. He also notes that some care is required to pass in the relevant -mllvm options to rustc and to ensure that it is not attempting to link against compiler-rt because doing so is not appropriate when building the Linux kernel. Marco also noted that KCSAN relies on the front-end to properly handle volatile accesses, and Miguel Ojeda kindly confirmed that it does. So there is hope!

For more information on KCSAN and its reason for being:

1. "Concurrency bugs should fear the big bad data-race detector" (Part 1 and Part 2).
   
2. Who's afraid of a big bad optimizing compiler?.
   
3. Calibrating your fear of big bad optimizing compilers.
   
4. Sections 4.3.4 ("Accessing Shared Variables"), 15.2 ("Tricks and Traps"), and 15.3 ("Compile-Time Consternation") of perfbook.
   

History

October 7, 2021: Add current state of KCSAN/Rust integration. Clarify Rust module per Gary Guo email.
October 12, 2021: Self-review.
October 28, 2021: Add data_race() primitive.