Linux-insides: sequential locks in the Linux kernel
github.comI wrote an implementation of sequential locks in Rust [1]. It is extensively commented, which makes it a good way to learn how the algorithm works.
[1] https://github.com/Amanieu/seqlock/blob/master/src/lib.rs
On a related note, take a look a Hans Boehm's paper "Can Seqlocks Get Along With Programming Language Memory Models?" [1].
I enjoyed footnote 3:
... in spite of my earlier claims in the previously mentioned mailing list discussion.
This is in error: "So, no one reader or writer can't access protected data. " (in reference to read_seqlock_excl()).
Readers using read_seqbegin() can still access the protected data.
Not an expert but RCU is more complex with respect to checking the CPU states for activity to ensure consistency. Sequential locks use spinlocks and a counter and allows for concurrent reads. This link [1] actually answers it specifically (sorry about the quora link).
[1] https://www.quora.com/What-is-difference-between-the-RCU-rea...
LWN's coverage is pretty good https://lwn.net/Articles/262464/
Where do these stand with respect to RCU locks? Are they a replacement, or are these used in different situations?
The main downside of seqlocks is that you usually can't have pointers in them. Since the basic scheme is read, do stuff, check if sequence counter changed, you need to make sure that you don't dereference a potentially invalid pointer in the "do stuff" phase if another thread is concurrently modifying the data.
RCU avoids this issue by essentially delaying freeing of objects until there are no longer any threads accessing it. But it requires very complicated mechanisms to track all active threads, whereas a seqlock is a very simple and self-contained.
In both cases the read-side overhead is tiny: for example on x86, you don't need any memory barrier or atomic instructions.