Our great sponsors
-
WorkOS
The modern identity platform for B2B SaaS. The APIs are flexible and easy-to-use, supporting authentication, user identity, and complex enterprise features like SSO and SCIM provisioning.
If you want more than a spsc queue, I've written `lockfree`, a collection of SPSC and MPMC data structures along the same principles the author here used:https://github.com/DNedic/lockfree.
The library is written in standard C++11 (but additional API's for higher C++ versions have been added), uses no dynamic allocation and is configurable so it is both big metal and deeply embedded friendly.
It's just amortizing for large transfers. For frequent near-empty cases it still has the shared pointers problem. NVMe have a really clever way to avoid this, but it depends on the fact that there's [occasional] communication on the reverse stream (the SQ/CQ pair): The cache line sized entries in the submit queue has a "phase" bit in the last word. The reader can just read that to know when the entry has updated (the polarity of the bit toggles each time we loop around and you can just use the top+1 index bit if power-of-two sized).
The back pressure is handled with a credit scheme and the producer gets an updated copy of most recently-know consumer read counter with every completion message back.
Using this scheme you can achieve the optimal performance with just a single cache line of traffic for each cache-line sized message.
Unfortunately I haven't found a SPSC Rust crate that does this, but https://github.com/polyfractal/bounded-spsc-queue [abandoned] comes close.