Safe Lock-free Primitives with iceoryx2's ByteAtomic

eelfenpiff about 5 hours ago 27 commentsRead Article on ekxide.io

RU version is available. Content is displayed in original English for accuracy.

⚡ Community Insights

Discussion Sentiment

14% Positive

Analyzed from 991 words in the discussion.

Discussion (27 Comments)Read Original on HackerNews

danbruc•about 3 hours ago

A common approach to mitigating the described data race without using blocking locks is to utilize a sequence lock.

A sequence lock is a blocking lock. If the writer dies between the two increment operations, then the readers will spin forever waiting for the counter to become even again.

rbranson•about 2 hours ago

This is why wait-free and lock-free are separate concepts. Author is not claiming wait-free.

danbruc•about 1 hour ago

Obstruction-free, lock-free, and wait-free are increasingly strong guarantees for non-blocking synchronization mechanisms. A sequence lock is a blocking synchronization mechanism and therefore provides none of the aforementioned guarantees.

eqvinox•about 1 hour ago

No, it's neither wait-free nor lock-free. Wait-free guarantees all threads make progress, lock free guarantees at least one thread makes progress even if another thread is suspended or dead.

jnwatson•about 2 hours ago

Also, there's no guarantee of progress. The writer can starve the reader forever.

adrian_b•about 1 hour ago

That is always possible, even in any algorithm that claims to be wait-free, if some writer just keeps writing the shared data.

All the claims about something being lock-free and/or wait-free depend on a rational behavior of the writers.

If any writer acts crazy, progress becomes impossible regardless of what all others do, unless someone kills the rogue thread or process.

In practice, the algorithm from TFA is much more likely to guarantee progress than any of the algorithms that are theoretically proven to guarantee progress, because it has an extremely small overhead, while the alternatives are much more complex and they waste a lot of time.

Moreover, most wait-free algorithms guarantee progress only for the whole system, in the sense that one random thread will progress, but they do not guarantee anything for a given thread, which may be blocked forever or stuck in an infinite loop, if unlucky.

danbruc•about 1 hour ago

Not true, a wait-free algorithm guarantees that a read will complete in a bounded amount of time. And it guarantees that all threads make progress, it is lock-free that only guarantees progress for one thread.

If the value changes frequently, it will get outdated quickly, but that has nothing to do with the synchronization mechanism used. And even if writes happen rarely, there is always a chance that the value you read will be outdated a nanosecond later.

eqvinox•about 1 hour ago

I think you missed the point; seqlock based approaches will lock dead if you suspend/abort a thread in the wrong place. Other lock-free approaches don't have this issue. This isn't about a thread writing garbage, it's about guarantees applicable within the constraints.

danbruc•about 1 hour ago

[...] then the readers will spin forever waiting for the counter to become even again.

jimaway123•about 2 hours ago

From the article: >>>The Problem: Even if the reader detects that the data was modified and discards the copy before use, the act of copying the non-atomic data itself still triggers undefined behavior. While a sequence lock can detect that a data race occurred, it does not prevent it.

Why is this a problem? Isn't the correct way to deal with a sequence lock failure to just retry? A torn read yes means you get undefined behavior as far as the result of your read, but you throw it all away and start again anyway so what is this solving?

adrian_b•36 minutes ago

The problem is that a reader should not do anything with the data that is read, which depends on its value, before validating the copy, by counter comparison.

As another poster has said, the high-level languages leave undefined what happens when you copy non-atomically data that is written concurrently, but in fact the computer cannot catch fire when you do that, and the only thing that can happen is that the data may have values that are invalid for its type, e.g. an integer that is defined to belong in a range may have a value outside that range.

A much more serious problem that is not mentioned in TFA is that on computers that do not use Intel/AMD CPUs, this algorithm needs write barriers and read barriers. The writer must use 2 write barriers, after incrementing the counter before accessing the shared data, and before incrementing the counter after finishing with the shared data. Similarly the reader needs read barriers after the first reading of the counter and before the final reading of the counter.

jimaway123•25 minutes ago

>>>The problem is that a reader should not do anything with the data that is read, which depends on its value, before validating the copy, by counter comparison.

The description of the problem explicitly says that this is not happening. The data is being thrown away if the counter comparison fails. So the only undefined behavior that I can think they might be referring to is the act of loading the data itself, even though it is thrown away if it is wrong. Is that what they mean?

charleslmunger•about 2 hours ago

From a hardware perspective this is correct. From a language perspective it's UB, and unless your compiler has defined that UB, it doesn't matter what the hardware's behavior is unless you're writing assembly.

jimaway123•29 minutes ago

So you're saying the problem that this article is solving is just preventing the language from generating code that can possibly load some incorrect bytes from main memory, even though these incorrect bytes will in every case be ignored? Why is this considered a problem at all?

elfenpiff•about 5 hours ago

iceoryx2 provides zero-copy inter-process communication mechanisms based on shared memory and data structures that are modified concurrently by multiple processes.

One of the key operations in these algorithms is a memory copy using core::ptr::copy. However, this results in undefined behavior if one process reads the data while another process writes to it concurrently. Even if our lock-free algorithm reliably detects such a race, iceoryx2 cannot depend on undefined behavior in a safety-critical system. This blog post introduces our solution: a byte-wise atomic wrapper that enables well-defined concurrent copy operations. It also shows how it can be used to implement a simple sequence lock.

eqvinox•about 3 hours ago

You're fixing a theoretical problem (mismatch between CPU and compiler memory models, the CPU is perfectly fine doing these reads and writes, it's only the compiler declaring them "UB") by throwing away a shitton of performance, forcing everything into bytewise accesses. Considering this is Rust, I would at minimum expect this be written to be generic over access size to allow using 64-bit reads/writes.

I'm also missing any acquire/release barrier annotations in your code snippets. If you're using sequentially consistent accesses you might as well just single thread your code, performance wise.

Lastly, in almost all cases it's way more efficient and appropriate to shuffle things on the whole-object level, posting and retrieving pointers, and not poke around inside objects (especially on the byte level). Check how rare the use of seqlocks in the Linux kernel is, compared to other RCU primitives. (and regarding "appropriate", cf. top-level comment by danbruc https://news.ycombinator.com/item?id=49168283 )

wbl•about 1 hour ago

The compiler declares them UB because very useful transformations would change the semantics of races. Races are ok with specially marked variables!

eqvinox•about 1 hour ago

That's not the point of my argument. The compiler declaring things UB needs to be addressed by telling the compiler to not be silly, not by forcing every single access to be on the byte level. And especially not if those are SeqCst atomics.