From: Ralf Jung Date: Fri, 25 Jul 2025 12:01:05 +0000 (+0200) Subject: mention int2ptr cast and multi-word atomics X-Git-Url: https://git.ralfj.de/web.git/commitdiff_plain/ebe98317dd54b78c4ea65cc3ecc7d8aae528a4e0 mention int2ptr cast and multi-word atomics --- diff --git a/personal/_posts/2025-07-24-memory-safety.md b/personal/_posts/2025-07-24-memory-safety.md index 74fd1b0..f940453 100644 --- a/personal/_posts/2025-07-24-memory-safety.md +++ b/personal/_posts/2025-07-24-memory-safety.md @@ -78,7 +78,7 @@ Every time `repeat_swap` stores a new value in `globalVar`, it just does two sep In `repeat_get`, there's thus a small chance that when we read `globalVar` *in between* those two stores, we get a mix of a pointer to an `Int` with the vtable for a `Ptr`. When that happens, we will run the `Ptr` version of `get`, which will dereference the `Int`'s `val` field as a pointer -- and hence the program accesses address 42, and crashes. -One could construct a similar example using Go's slices, where the data pointer, length, and capacity of the slice are stored in separate words, and reading a half-updated value can lead to an out-of-bounds access. +One could easily turn this example into a function that just casts an arbitrary integer to a pointer. ## What about other languages? @@ -92,9 +92,11 @@ In that sense, all Java programs are thread-safe.[^java-safe] [^java-safe]: Java programmers will sometimes use the terms "thread safe" and "memory safe" differently than C++ or Rust programmers would. From a Rust perspective, Java programs are memory- and thread-safe by construction. Java programmers take that so much for granted that they use the same term to refer to stronger properties, such as not having "unintended" data races or not having null pointer exceptions. However, such bugs cannot cause segfaults from invalid pointer uses, so these kinds of issues are qualitatively very different from the memory safety violation in my Go example. For the purpose of this blog post, I am using the low-level Rust and C++ meaning of these terms. Generally, there are two options a language can pursue to ensure that concurrency does not break basic invariants: -- Ensure that arbitrary concurrent programs actually behave "reasonably" in some sense. This comes at a significant cost, restricting the language to never assume consistency of multi-word values and limiting which optimizations the compiler can perform. This is the route most languages take, from Java to C#, OCaml, JavaScript, and WebAssembly. +- Ensure that arbitrary concurrent programs actually behave "reasonably" in some sense. This comes at a significant cost, restricting the language to never assume consistency of multi-word values and limiting which optimizations the compiler can perform. This is the route most languages take, from Java to C#, OCaml, JavaScript, and WebAssembly.[^multi-word] - Have a strong enough type system to fully rule out data races on most accesses, and pay the cost of having to safely deal with races for only a small subset of memory accesses. This is the approach that Rust first brought into practice, and that Swift is now also adopting with their ["strict concurrency"](https://developer.apple.com/documentation/swift/adoptingswift6). +[^multi-word]: Some hardware supports larger-than-pointer-sized atomic accesses, which could be used to ensure consistency of multi-word values. However, Go slices are three pointers large, and as far as I know no hardware supports atomic accesses which are *that* big. + Go, unfortunately, chose to do neither of these. This means it is, strictly speaking, not a memory safe language: the best the language can promise is that *if* a program has no data races (or more specifically, no data races on problematic values such as interfaces, slices, and maps), then its memory accesses will never go wrong. Now, to be fair, Go comes with out-of-the-box tooling to detect data races, which quickly finds the issue in my example.