This uses read_exact and isn't async friendly. Also, I'm not sure I consider #12 to be a bug. The best resolution to #12 is a helper type that takes a Read implementation and converts it into a Read + Seek implementation by offering some limited relative seeking.

I agree that read_exact blocks the entire thread and not async friendly, but I don't understand the difference between the original implementation. Does using read function make it async-friendly?

I don't strongly think that #12 is a bug that has to be handled right away, but something that improves the usefulness. Still, I don't think it's a good idea to enforce Seek via helper type, because that's what we can do it under the hood.

Does using read function make it async-friendly?

Yes. read itself can return a WouldBlock error. It's the pre-async model but is iirc somewhat compatible with the async/await features. Need to look up the details. Edit: see below.

that's what we can do it under the hood.

If it's done under the hood, it's done unconditionally, even though it adds a non zero overhead. So it's a bad idea.

Wait disregard that, all of PacketReader is non-async-ready.

So I don't have to care about async for this implementation? If so, what are other problems left?

My previous code returns Error when failed to find a header, which is different from what is written in the document, so I corrected. As an side effect, the code does no longer use read_exact.

Hmm thinking about this some more, it basically adds the requirement to copy data once more if you want to retain performance, which isn't optimal for performance due to the additional copy/buffering. Avoiding this was a deliberate design choice. See my comment above.

Maybe the new behaviour could be exposed under a different API. Then users can choose themselves which approach they want (lots of Read calls or occasional seeking).