Re: [lwip-devel] Query on Per-socket Locking and Optimal Performance Model in lwIP

[Simon Goldschmidt]

On 05.03.2024 06:29, 袁仲 wrote:
> Dear lwIP Community, I hope this message finds you well.

Yes, it found me twice. I wonder why you sent it twice?

> I am Heavy, and I have been working on a user-space network component
> built upon the lwIP library version 2.1.3-stable.In
> <http://2.1.3-stable.in/> my recent project experiences, it appears that
> lwIP（both version 2.1.13 and 2.2.0） does not inherently support
> per-socket locking. To address synchronization issues, I've resorted to
> using either the core tcpip lock provided by lwIP or creating custom
> mutexes to safeguard against concurrent access to both the lwIP stack
> operations and individual socket manipulations. This approach is
> functional, but I anticipate potential performance degradation in
> scenarios where multiple sockets are concurrently active, as these locks
> are essentially global to all socket operations.
> My question is twofold:
> 1. Does the current stable release of lwIP indeed lack native support
> for per-socket locking mechanisms?

Are you talking about sockets or about pcbs? Because there should be no
problems using different sockets from different threads. You do have the
"only one thread in core lwIP" rule, so performance is not great on
multi-core systems in that case. That's the downside of staying
"lightweight" so far (the "lw" in lwIP).

> 2. What design pattern or model would best leverage lwIP's performance,
> especially when handling multiple simultaneous socket interactions?

I don't think there has been public work in this area, so you're welcome
to provide anything if you come up with something. I could imagine
reader/writer locks, reference counting and many more. But up to now,
noone really had enough interest in this area to actually implement
this. I don't think it's a simple task to get that right!

Regards,
Simon