This post has been discussed here a couple times, but AMA :)
edit, the most commented version of this post was the original:
https://news.ycombinator.com/item?id=20096946.
This is what I'm up to these days:
I did a bit of link diving to the various blog posts and sites but haven't been able to find it. Would be nice, if it exists, to have it front and centre.
Andrea Lattuada (https://andrea.lattuada.me/) is more likely to have done work in his implementation of the algorithm than I did on mine.
I have run the testing with various dynamic analysis tools, but for sure this isn't formal verification. If someone is interested in doing it, happy to chat with them!
Lock-free here does not mean "without any form of synchronization primitive". That is impossible. Some synchronization must occur.
Instead, it is a term of art that means:
1. Thread failure can't cause other threads to get blocked. 2. Forward progress in the algorithm is guaranteed.
Basically: threads can't block each other and death of threads does not screw things up.
https://en.wikipedia.org/wiki/Non-blocking_algorithm
CPU atomics often work by blocking (they have to) but they are non-interruptible instructions and there is a fixed maximum execution time.
(So you can still guarantee forward progress if you use them).
Any ideas ?
I've used tokio's broadcast channel quite a bit before, but it is also bounded. After talking to Eliza from the tokio project, I'm fairly convinced that unbounded queues are a scary thing to have around, operationally :).
But again - this is a bit out of my actual expertise!
So, everyone has a pointer to the current one, and there's an atomic pointer to the next.
When the log fills up, you allocate a new one and set the next pointer to it using compare_and_swap. If someone beat you to it, you can walk the list and add yours to the end so that the allocation isn't wasted.
This way, most of the time you're using the efficient bounded log.
So the linked list would involve going through all the links and following end tail refs/pointers. It would make reading O(n) and that's a Nope.
However, you could imagine having a Vec index that contains a ref to all allocated inner-logs, and query the index first in order to obtain the buffers' location. That works, but then the index has to go through a lock (either a RWLock or a mutex) as the CaS operation isn't enough if we get to the end of the index and it needs to be reallocated. It's fine, and I think that's the most appropriate solution.
PS : In fact, there is a sweet spot where you'd like to have a semi-unbounded log. If your index is big enough to contain something like 4Md entries, you'd probably end-up splitting the log in several pieces for archiving and performances purposes. Loading the log (fully or partially) from disk efficiently is then more important than having a real unbounded log. Then you would not necessarily use a lock and could CaS in the index.
The alternative is to dive into the literature on lock-free mpmc queues, which are kind of gnarly to implement. A lot of the literature handwaves ABA problems with stuff like "use hazard pointers and RCU" without correct pseudo code to help you.
That's why locking in unbounded queues is popular, imo. It's not actually inefficient, and the alternatives are arcane enough to avoid. No one is going to understand or trust the code anyway.
It's worth mentioning that "lock free" means "one task does not prevent other tasks from making progress" and in the case of a bounded queue, you can trivially accomplish that by busy-waiting when the queue is full. This isn't appropriate if you need consumers to receive events in the order in which they happen, but you can kind of fix that using an atomic counter (to paraphrase Mike Acton, if you have a safe counter, you have a safe queue) and time stamp events/sort on which ones are consumed.
Yes and I do think they are fundamentally incorrect or they cannot be translated to the log use case.
I also agree that, performance-wise, the lock make little difference. In fact my old benchmarks tend to point that locks were MORE EFFICIENT than atomics on ARM64 (MBP M1), for example. It's more like a fun little exercise, and also to confirm that I'm not completely dumb and that the problem is not solvable with the simple use of 2 atomics and a counter.
I see a lot of critique in the previous (2019) thread, but no summary in key points. What are the reasons that this is hard?
In multiprocessor systems, are memory writes not guaranteed to be sequential? e.g. can data be written out-of-order, after the synchronization bit is written?
Or is it more about the use case, that it is optimized for minimal number of instructions, e.g. avoiding additional memory reads? (e.g. by writing data to the buffer, instead writing storing pointers)?
Or is it that you're trying to use constant memory (irrespective of number of threads)?
Because to me, it seems like a trivial problem to solve, if you have sequential writes, store data separately from the queue, and may linearly scale memory with number of threads.
(... and probably wasn't aware that the assumptions made in the specification can be encoded in the API signature in Rust. This wouldn't be possible in C, which is why C forces you to solve the harder problem or rely on your users to not accidentally misuse the data structure.)
Instead of pushing and popping one byte/thing at a time (inefficient, a lot of overhead), we want to have the ability to push or pop big chunks at a time.
We also don't want to pay to copy these chunks in or out, because we're going to have the hardware do it for us autonomously. This means we have to be very careful that the "one side" doesn't look or touch a chunk of memory currently being used by the "other side".
The ideal flow is that we have the CPU get some "writing space", the CPU asks DMA to fill it up, when DMA is done, the CPU marks it as "ready to read". Then at some later time, the CPU (maybe another core or thread) asks for some "ready to read space", the CPU then either uses it, or maybe asks DMA to copy it somewhere else. Then the CPU marks that data as "successfully read", and the space can be recycled the next time something wants to send.
The trick is how you coordinate access to shared memory between two CPUs, correctly, and using the least overhead, so neither side prevents the other from reading or writing, as long as theres a little slack in the line.
if buffer.len.saturating_sub(buffer.write.load()) >= write_len {
// not shown: check `read` to make sure there's enough free room
buffer.watermark.store(buffer.write.load() + write_len);
buffer.write.store(buffer.write.load() + write_len);
}
It seems to me that watermark need not be atomic anyway: it is owned by the writer when read <= write, and by the reader when write < read. Whoever wraps around implicitly passes the ownership to the other thread. More precisely:
* In the initial state, and until the writer reaches the end of the buffer, read <= write, and the writer owns watermark, whose value is irrelevant. In particular, the reader makes sure not to overtake write, and knows that it must not access watermark.
* When the writer needs to wrap around, it first sets the watermark (possibly at len, if the hole is empty), then updates write. With its usual release memory ordering constraint, the atomic store to write transfers to the reader the ownership of the written messages, but also of watermark.
* Eventually, the reader notices that write has wrapped around, and then uses watermark to know where to stop reading. Meanwhile, the writer may fill the beginning of the buffer, carefully not overtaking read - 1, and never touching watermark.
* When the reader finally wraps around, the atomic write to read transfers to the writer the ownership of the consumed message slots, but also of watermark, and we are back in the first phase.
IOW, watermark has the same owner as the last element of the array.
What do you think? What have I missed?