Re: [lwip-members] question about SYS_ARCH_ defines

[David Haas]

Well, I guess I disagree.

Protection should be put where it is necessary at a low layer (although 
we should be coding to minimize those spots). Otherwise you wind up with 
your entire task protected, causing too much inter-task latency. Are we 
going to protect the entire tcp task? It's an option, but it would make 
lwip totally unsuitable for a small embedded system which uses an OS and 
has some real-time requirements.

I think lwip has the approach of pushing the protection out into the 
sys_arch stuff. This gives users lots of flexability. In fact if you 
want to protect your tasks completely, you could, by specifying that in 
sys_thread_new().

My implementation for lightweight protection for unix uses mutex and 
does not assume that they support recursive calls. A counter is 
implemented outside of the mutex. Would it work for uCOS?

I admit there are risks for deadlock. I actually fixed a deadlock 
problem I had with select() (I submitted it already).

Regards,
David.


address@hidden wrote:

> Hello,
>
> my thoughts on protection are as follows:
>
> Do not implement locking in the lwIP core files, because
> we simply cannot implement all the different protection
> schemes used by different operating systems.
>
> Implement locking in the higher layers.
>
> The benefit is that lwIP is not cluttered with different protection
> schemes.
> The drawback is that locking critical sections occurs at a more coarse
> grained level.
> The benefit is that we do not have deadlock (are your mutexes recursive??).
> For example, uCOS does not support recursive mutexes.
>
> If a task enters a critical section, then calls another function entering
> the
> same critical section (for example pbuf management), the protection mutex
> is not recursed and deadlock occurs.
>
> I would strongly opt for removing all locking primitives/macro's from the
> core lwIP API to improve stack maintainability. I think this would break
> a lot of stuff out there though...
>
> Regards,
>
> Leon Woestenberg
>
>
>
>
> _______________________________________________
> lwip-members mailing list
> address@hidden
> http://mail.nongnu.org/mailman/listinfo/lwip-members
>
>
>