Re: [lwip-users] Making lwIP thread safe

[address@hidden]

Jonathan Larmour wrote:
> Bob McConnell wrote:
>   
> > address@hidden wrote:
> >     
> > > That said, if I understand you correctly, you want to run 200 full
> > > instances of lwIP in one process: that doesn't work (also,
> > > intentionally): many parts in lwIP use global or static variables (the
> > > whole stack is _not_ reentrant). After all, running lwIP in a process
> > > is much like running it on its own device (a completely seperate
> > > address space). Running 2 instances in 2 different threads in the same
> > > process is like running it twice on the same device: you could make it
> > > work with a high effort, but there's no use for that in normal life...
> > >       
> > I knew there was one large company in the Northwest USA that doesn't
> > believe testing is part of normal life, but I wasn't aware that everyone
> > believed that.
I didn't say I wouldn't test if I was developing a product using lwIP. 
All I said was if I would test, I wouldn't try to run software in 
multiple threads inside one process if it's normally running in 
completely seperate devices.
> >  Be that as it may, many devices need multiple stacks
> > running, for example one for a network interface and a second for ppp on
> > a serial port.
I don't really know why you would need multiple stacks to have more than 
one network interface., maybe you can explain this to me? Why don't you 
simply use 2 interfaces on one instance of the stack?
> >  This can be done with Rabbits and PICs, so it should
> > qualify as a valid target. I believe you have taken much too narrow a
> > view of normal life.
> >     
>
> lwIP supports multiple network interfaces - the normal, and reasonable
> expectation, is that nothing in lwIP prevents you having a single stack
> with both an ethernet and ppp interface. That being said....
>
>   
> > But I also think you misread my description. I need 2000 full instances
> > running simultaneously. Only the prototype may be limited to 200. Each
> >     
Got that.
> > instance must emulate one of several different devices and open a
> > permanent TCP connection with a transaction server using a unique IP
> > source address. This eliminates UML and VMWare as hosts, since they
> > cannot support anywhere near that many sessions. It is not possible to
> > set up that many devices  in our facility, even if we emptied out the
> > entire warehouse to make room for them.
> >  Perhaps you can suggest some
> > other way to implement this test bench?
> >     
I think you have two possibilities:

a) use 2000 processes (on windows or some kind of unix - e.g. linux or 
macos), each running a copy of lwIP. The advantage of this solution is 
that you have completely seperate address spaces for each device 
simulation - much like the software running on the device: that way you 
make sure one instance doesn't interfere with the others (you don't get 
that with pthreads!). Of course, processes are not as lightweight as 
threads, but an up-to-date PC should be able to run at least a few 
hundred or maybe even all (depending on the speed you need, you can use 
more than one PC for simulation).

b) use 2000 netifs in a single process. Like Jonathan said, lwIP is 
_not_ designed for this amount of netifs (lw stands for lightweight). 
The big disadvantage is that you can't completely simulate your 
software: you have to design it in a way that it can be run with its own 
stack (in the device) _or_ with a shared stack. That way, you can't 
fully test the software because it is different when run in the device.

In any case, you can take a look at the contrib folder in the CVS 
repository: there are ports for windows and unix/linux you can use to 
run lwIP on a PC (each process or netif can have its own IP; 'bridged' 
to the real network).
> Supporting that many interfaces simultaneously is not a particularly narrow
> view - it's extremely unusual. lwIP, being lightweight, would not be well
> optimised for the case of 2000 netifs. netifs are held on a list which is
> processed in a simple linear way.
>
> But on the other hand, you talk about "Rabbits and PICs", so does that mean
> each copy of lwIP would run on its own processor in its own address space?
>   
As I understand it, that's the case. Bob, please correct me if I'm wrong.

Hope that helps,

Simon