Make Send+Sync impls optional for compatibility with AVR #70
Conversation
In practice, making `sync` a default feature makes AVR compatibility a pain, as all the other crates extending this core crate (such as `usbd-hid`) would have to explicitly opt out of the default feature.
|
Hi @agausmann , it's cool to see Rust coming to AVR! The thread safety actually is still necessary on single-core chips, because interrupts provide additional execution contexts. The ATSAMD21 for instance also lacks a compare-and-swap (CAS) instruction, the usb-device implementation in the HAL uses an interrupt-free block (here) to implement a "mutex". However, that approach can be problematic, as I ramble on about in this issue. |
|
@agausmann, you can make "atomics" work on AVR by setting "max-atomic-width": 16,in your rustc-target JSON. This works because there are intrinsics that implement the atomic operations by disabling interrupts for their duration. However, I am not certain that they are always compiled correctly at this time. |
|
Sorry, I meant to respond to this a long time ago. Just want to clear some things up:
Okay, so my comment about "chips are single-core anyway" wasn't accurate in hindsight. Don't even know why I wrote that. What I should have said: The motivation for this is to use (or at least try to use) usb-device only on the main execution thread without accessing it in ISRs; I've had success with that approach on other MCUs, and I wanted to try that on AVR. This PR was a potential solution for that, I wanted to opt-in to !Send + !Sync to work around the lack of atomics in AVR. (also thanks @Rahix for helping with that) |
|
No longer interested in this |
The AVR architecture doesn't have atomic operations, but the chips are generally single-core anyway so thread safety isn't necessary.
This crate can't be used on AVR in official Rust yet, since this crate uses
dyn/ trait objects (rust-lang/rust#79889). But there is a pending fix, and I'm testing it out with my own patched version of Rust/LLVM.