Re: [lwip-devel] PPP single threaded

[Simon Kallweit]

Alain M. wrote:
> Hi Simon,
>
> I am new to lwip but I hope that I can help. I have been working with 
> PSilva, I did most of the protocol decoding to find the problem...
> > I've put some work into getting PPP working in a single-threaded 
> > (NO_SYS = 1) environment.
> My first suggestion is that you make it over a Linux port, it really 
> helps A LOT to debug. Peter's test is running, but with threads, it 
> should be relatively easy to remove the threads. Al least it will make 
> it easyer FOR ME to help :)

Well, I'm actually working with an eCos port, and have my eCos running 
as a synthetic target on a linux host. I don't need real hardware to 
test, so to speak. I have a GSM modem attached on a serial port, and 
this all works so far.

> > I've also put some time into refactoring the existing PPP code to 
> > match more with the rest of lwIP, mostly renaming though. I've also 
> > implemented a really simple 'chat' module, to setup a GPRS 
> > connection. Now I wonder if the effort should be made to bring this 
> > all back into lwIP. There is still a lot to do, but it would probably 
> > be worth it.
> Please, do ...
> > Let me explain what I changed in the design:
> > * Added a simple timesys module to PPP, this is ticked by a call to 
> > ppp_tmr() like the rest of lwIP. This timesys implements a jiffies 
> > counter which PPP needs in some places.
> is jiffies used only in PPP? Usualy you will get it from the RTOS, so 
> I would suggest that a more generic function be used.

sys_jiffies() is defined in sys.h, but only used in the PPP code. It 
might be a good idea to have some standard interface to retrieve the 
system time, but it the current implementation lacks a lot, as there is 
no definition of what a jiffy is (in PPP it's assumed to be 10ms). We 
would need something like sys_get_time() which returns ms or something. 
Or we can get rid of it completely and have the application call a 
global timer function with delta time passed.

> > It also implements the timeout handlers, which are absent in lwIP for 
> > the NO_SYS = 1 case, but still necessary in PPP. I think this puts 
> > PPP more in line with the rest of the lwIP core.
> > * Added ppp_poll(pd) function, which needs to be called periodically 
> > when in single-threaded mode.
> Yes, ans yes...
> > This all seems to work fine so far. I wonder though, how to make this 
> > working again in a multi-threaded environment, where we have a 
> > separate thread to poll the serial for input.
> what usualy works just fine is to put all your code into a single 
> thread. Simple, easy and portable. Once the connection is established, 
> the thread could just close.

Yes that's true, but then we would have to have the serial input and the 
timeout handling to be running in the same thread, which will be 
impossible if the serial input is blocking.

> > I guess the PPP input and output paths are not thread-safe, so they 
> > would have to be synchronized in some way.
> I am not familiar with lwip , but from my experience with other 
> systems, I can say this: PPP has parts:
> 1) making the connection, in this part noone should use the serial 
> port and the ppp can use sio_xxx without wory
> 2) once the connection is established, it is a "packet driver" just 
> like the ethernet one. And I believe that lwip already must have 
> solved that problem ;)

I haven't dug into the PPP code enough to see through. But it seems to 
be a bit messy in parts ...

> > The current approach with calling tcpip_callback() all the time does 
> > not look like the right way to me.
> agreed, but only theoreticaly because of lack of knowledge of lwip.
> > Is anyone interested on collaborating on this?
> I can help some... specially on a Linux port. Real hardware is with 
> PSilva, but I indend to get another one soon.
>
> I have sent to the list a sio_read() that can be aborted, I can easely 
> conver it to a nonblocking timeouted one.

I think we should make blocking behaviour configurable in the sio 
interface, so we can support both threaded and polled operation.

> > I've heard from quite a few guys now, who did their own hacks to get 
> > PPP working in a single-threaded environment, and I think it's time 
> > to get it right. Any input is welcome!
> I hope that I can help, at least with chearing :)

Well it looks like quite a bit of work. One awful part of the current 
PPP implementation is the mixup of PPP over Serial and PPP over 
Ethernet. It's not cleanly separated, and the code is generally quite 
messy. I was starting to wonder if a new port on the base of a current 
pppd implementation (http://ppp.samba.org/) should be considered. But 
this would lead to even more work, and I should actually start 
developing the application, so my time on lwIP/PPP is a bit limited.

Simon