Skip to contentSkip to conversation

ppp

DISCUSSION Ryan Raasch 6 participants 13 messages Nov 12, 2009 – Nov 15, 2009

lore
Share & tools
lore
#0 lore raw

Hello,

We need a ppp interface for our modem. I was told on the chat forum that a tap/tun device was to be used. Has anyone else started with this or thought how to implement this?

After reading briefly the code from pppd, the implementation, for a first draft, is pretty straight forward (w/o encryption, etc. in userspace). "Just" dial the apn, and hand over the tty fd to ppp via ioctl.

Steps (borrowed from pppd):

1. Dial the apn (in ofonod).
2. Some synchronization code.
3. hand over tty fd to ppp driver via ioctl. the IP.

I know i am simplifying the process.

Is this the way to go?

Regards,
Ryan

#1 lore raw

Hi Ryan,

Quoted 5 lines

Hello,

We need a ppp interface for our modem. I was told on the chat forum that a tap/tun device was to be used. Has anyone else started with this or thought how to implement this?

We're currently working on a userspace ppp implementation that will integrate with GAtChat. The goal here is to make it integrate well with the current multiplexed channel capability in GAtMux/GatChat and not need to involve an external daemon.

Quoted 5 lines

After reading briefly the code from pppd, the implementation, for a first draft, is pretty straight forward (w/o encryption, etc. in userspace). "Just" dial the apn, and hand over the tty fd to ppp via ioctl.

Unfortunately it is not that simple. pppd requires a proper kernel file descriptor to hand off to the kernel ppp line discipline. In our userspace implementation that is not the case. 1 kernel file descriptor is shared by potentially many GAtChat channels using a multiplexing protocol. Hence why we can't reuse the kernel implementation.

Regards,
-Denis

#2 lore raw
Quoted 4 lines · Denis Kenzior

Hi Ryan,

Quoted 2 lines

Hello,

Quoted 9 lines
Quoted 3 lines

We need a ppp interface for our modem. I was told on the chat forum that a tap/tun device was to be used. Has anyone else started with this or thought how to implement this?

We're currently working on a userspace ppp implementation that will integrate with GAtChat. The goal here is to make it integrate well with the current multiplexed channel capability in GAtMux/GatChat and not need to involve an external daemon.

Wow. So ofono will not use the kernel space ppp and implement the ppp stack in userspace?

Quoted 11 lines
Quoted 4 lines

After reading briefly the code from pppd, the implementation, for a first draft, is pretty straight forward (w/o encryption, etc. in userspace). "Just" dial the apn, and hand over the tty fd to ppp via ioctl.

Unfortunately it is not that simple. pppd requires a proper kernel file descriptor to hand off to the kernel ppp line discipline. In our userspace implementation that is not the case. 1 kernel file descriptor is shared by potentially many GAtChat channels using a multiplexing protocol. Hence why we can't reuse the kernel implementation.

Yea. That's right.

Quoted 6 lines

Regards,
-Denis
_______________________________________________
ofono mailing list
ofono(a)ofono.org
http://lists.ofono.org/listinfo/ofono

#3 lore raw
In reply to Denis Kenzior
Quoted 7 lines · Denis Kenzior

Unfortunately it is not that simple. pppd requires a proper kernel file descriptor to hand off to the kernel ppp line discipline. In our userspace implementation that is not the case. 1 kernel file descriptor is shared by potentially many GAtChat channels using a multiplexing protocol. Hence why we can't reuse the kernel implementation.

Not sure how relevant this is, but many many years ago (1997) I used a "telnet modem" to circumvent a firewall. It was a user land process that presented a fake /dev/modem which I used using pppd and chat..

My machine was behind a firewall with a telnet proxy.. Using the software
+ pppd + chat, it was set up along the lines:

start telnet modem which presented a /dev/modem.. start pppd using /dev/modem:

ATDT 1.2.3.4
CONNECT 1.2.3.4
the-telnet-proxy>
telnet my.host 23
connected to my.host:23
login:
mypppuser
password:
mypassword
end of chat script... start pppd

In this situation, pppd and the kernel device would have been communicating with a local userland process. The TCP connection to 1.2.3.4 created by the telnet modem could not have been used by the ppp interface as the telnet modem was passing the streams in either direction essentially joining the fd of the telnet connection and fd used by pppd..

I'm not sure if the requirements of ppp in the kernel have chagned, but couldn't something like this be repeated in this situation.. allowing pppd to speak to a local socket/fifo/pair of pipes that is terminated in user space?

I'll see if I can drag the code out to see what it actually did.. and even better see if I can get it to work on a 2.6 kernel!!

#12 lore raw
In reply to Denis Kenzior
Quoted 5 lines · Denis Kenzior

Unfortunately it is not that simple. pppd requires a proper kernel file descriptor to hand off to the kernel ppp line discipline. In our userspace implementation that is not the case. 1 kernel file descriptor is shared by potentially many GAtChat channels using a multiplexing protocol.

You _could_ create a PTS/PTY pair though...

Quoted 1 line

Hence why we can't reuse the kernel implementation.

...But I am not sure if the PPP line discipline works correctly on top of /dev/pts.

Signature · 1 line
Rémi Denis-Courmont
#4 lore raw
In reply to Chris Pitchford
Quoted 2 lines · Chris Pitchford

I'm not sure if the requirements of ppp in the kernel have chagned, but couldn't something like this be repeated in this situation.. allowing

pppd

Quoted 2 lines

to speak to a local socket/fifo/pair of pipes that is terminated in user space?

In my understanding, PPP is a line discipline, meaning it can only run on a TTY character device device, not just any character device, and definitely not any random file descriptor.

Signature · 1 line
Rémi Denis-Courmont
#5 lore raw

Quoting Rémi Denis-Courmont <[email protected]>:

Quoted 5 lines
Quoted 2 lines · Chris Pitchford

I'm not sure if the requirements of ppp in the kernel have chagned, but couldn't something like this be repeated in this situation.. allowing

Quoted 7 lines

pppd

Quoted 2 lines

to speak to a local socket/fifo/pair of pipes that is terminated in user space?

In my understanding, PPP is a line discipline, meaning it can only run on a TTY character device device, not just any character device, and definitely not any random file descriptor.

turns out that the software I mentioned (virtmodem from 1995) does indeed use a pty pair. Although I've not had time during the week to get it running again, I suspect it would probably work on a recent kernel/pppd.

Is the line discipline so much of an issue when using ppp connected to a fake tty? could that not be handed on the proxy side inside ofono?

#9 lore raw
In reply to Chris Pitchford

Hi Chris,

Quoted 34 lines
Quoted 6 lines

Unfortunately it is not that simple. pppd requires a proper kernel file descriptor to hand off to the kernel ppp line discipline. In our userspace implementation that is not the case. 1 kernel file descriptor is shared by potentially many GAtChat channels using a multiplexing protocol. Hence why we can't reuse the kernel implementation.

Not sure how relevant this is, but many many years ago (1997) I used a "telnet modem" to circumvent a firewall. It was a user land process that presented a fake /dev/modem which I used using pppd and chat..

My machine was behind a firewall with a telnet proxy.. Using the software
+ pppd + chat, it was set up along the lines:

start telnet modem which presented a /dev/modem.. start pppd using /dev/modem:

ATDT 1.2.3.4
CONNECT 1.2.3.4
the-telnet-proxy>
telnet my.host 23
connected to my.host:23
login:
mypppuser
password:
mypassword
end of chat script... start pppd

In this situation, pppd and the kernel device would have been communicating with a local userland process. The TCP connection to 1.2.3.4 created by the telnet modem could not have been used by the ppp interface as the telnet modem was passing the streams in either direction essentially joining the fd of the telnet connection and fd used by pppd..

Quoted 4 lines

I'm not sure if the requirements of ppp in the kernel have chagned, but couldn't something like this be repeated in this situation.. allowing pppd to speak to a local socket/fifo/pair of pipes that is terminated in user space?

yes we could do that, but why bother with a PTY to push the data packets into the kernel. At that point we can also run PPP in userspace and use a TUN/TAP device to create our network interface. The question is when do we wanna move the data from userspace into the kernel.

Also pppd does the whole setup of PPP session. The ppp_generic in the kernel only das the data packets transport. Meaning the PPP session setup needs to be done in userspace anyway.

Regards

Marcel

#10 lore raw

Quoting Marcel Holtmann <[email protected]>:

Quoted 49 lines

Hi Chris,

Quoted 37 lines
Quoted 2 lines

Unfortunately it is not that simple. pppd requires a proper kernel file descriptor to hand off to the kernel ppp line discipline. In our

userspace

Quoted 1 line

implementation that is not the case. 1 kernel file descriptor is

shared by

Quoted 1 line

potentially many GAtChat channels using a multiplexing protocol.

Hence why we

Quoted 2 lines

can't reuse the kernel implementation.

Not sure how relevant this is, but many many years ago (1997) I used a "telnet modem" to circumvent a firewall. It was a user land process that presented a fake /dev/modem which I used using pppd and chat..

My machine was behind a firewall with a telnet proxy.. Using the software
+ pppd + chat, it was set up along the lines:

start telnet modem which presented a /dev/modem.. start pppd using /dev/modem:

ATDT 1.2.3.4
CONNECT 1.2.3.4
the-telnet-proxy>
telnet my.host 23
connected to my.host:23
login:
mypppuser
password:
mypassword
end of chat script... start pppd

In this situation, pppd and the kernel device would have been communicating with a local userland process. The TCP connection to 1.2.3.4 created by the telnet modem could not have been used by the ppp interface as the telnet modem was passing the streams in either direction essentially joining the fd of the telnet connection and fd used by pppd..

Quoted 4 lines

I'm not sure if the requirements of ppp in the kernel have chagned, but couldn't something like this be repeated in this situation.. allowing pppd to speak to a local socket/fifo/pair of pipes that is terminated in user space?

yes we could do that, but why bother with a PTY to push the data packets into the kernel. At that point we can also run PPP in userspace and use a TUN/TAP device to create our network interface. The question is when do we wanna move the data from userspace into the kernel.

Quoted 6 lines

Also pppd does the whole setup of PPP session. The ppp_generic in the kernel only das the data packets transport. Meaning the PPP session setup needs to be done in userspace anyway.

Regards

I understand your point about the setup being done in user space, but would it not be faster to present a vanilla pppd with a pty and allow it to setup the connection and pass data to the kernel interface rather than providing a new implementation? I mean from the point of view of reusing software that is already available and very well tested?

On the down side I see the "eeeew" factor of creating a pty pair just to pass data about, but if it means that far less code needs to be written (and much less code that is to interpret line traffic) I thought it might be safer solution. For example, if pppd is used, it already supports a wealth of authentication and feature negotiation.

I guess there are downsides to both approaches..

#11 lore raw

Hi Chris,

Quoted 52 lines
Quoted 42 lines
Quoted 37 lines
Quoted 2 lines

Unfortunately it is not that simple. pppd requires a proper kernel file descriptor to hand off to the kernel ppp line discipline. In our

userspace

Quoted 1 line

implementation that is not the case. 1 kernel file descriptor is

shared by

Quoted 1 line

potentially many GAtChat channels using a multiplexing protocol.

Hence why we

Quoted 2 lines

can't reuse the kernel implementation.

Not sure how relevant this is, but many many years ago (1997) I used a "telnet modem" to circumvent a firewall. It was a user land process that presented a fake /dev/modem which I used using pppd and chat..

My machine was behind a firewall with a telnet proxy.. Using the software
+ pppd + chat, it was set up along the lines:

start telnet modem which presented a /dev/modem.. start pppd using /dev/modem:

ATDT 1.2.3.4
CONNECT 1.2.3.4
the-telnet-proxy>
telnet my.host 23
connected to my.host:23
login:
mypppuser
password:
mypassword
end of chat script... start pppd

In this situation, pppd and the kernel device would have been communicating with a local userland process. The TCP connection to 1.2.3.4 created by the telnet modem could not have been used by the ppp interface as the telnet modem was passing the streams in either direction essentially joining the fd of the telnet connection and fd used by pppd..

Quoted 4 lines

I'm not sure if the requirements of ppp in the kernel have chagned, but couldn't something like this be repeated in this situation.. allowing pppd to speak to a local socket/fifo/pair of pipes that is terminated in user space?

Quoted 9 lines

yes we could do that, but why bother with a PTY to push the data packets into the kernel. At that point we can also run PPP in userspace and use a TUN/TAP device to create our network interface. The question is when do we wanna move the data from userspace into the kernel.

Also pppd does the whole setup of PPP session. The ppp_generic in the kernel only das the data packets transport. Meaning the PPP session setup needs to be done in userspace anyway.

Quoted 5 lines

I understand your point about the setup being done in user space, but would it not be faster to present a vanilla pppd with a pty and allow it to setup the connection and pass data to the kernel interface rather than providing a new implementation? I mean from the point of view of reusing software that is already available and very well tested?

if you look at the pppd code and tell us how to control it properly from within another daemon, then we can talk. That is just stupid to the end and would require massive extension of pppd. Also it ends up setting IP, routing and DNS details and some other magic. It works as a quick hack, but in the long run needs to be replaced.

Quoted 5 lines

On the down side I see the "eeeew" factor of creating a pty pair just to pass data about, but if it means that far less code needs to be written (and much less code that is to interpret line traffic) I thought it might be safer solution. For example, if pppd is used, it already supports a wealth of authentication and feature negotiation.

And the authentication etc. is not really that much with GSM cards. They are fairly limited in what they support. Especially since PPP only run between the card and the host. It doesn't go over the network.

Regards

Marcel

#6 lore raw
In reply to Chris Pitchford

Hi,

2009/11/13 Chris Pitchford <[email protected]>:

Quoted 2 lines

Is the line discipline so much of an issue when using ppp connected to a fake tty? could that not be handed on the proxy side inside ofono?

I've not tried it on a non-tty but it definitely should work modulo potential bugs that need to be squashed. It's supposed to work.

However by using a separate process to do ppp you add overhead. The original idea in ofono as well as in other projects (think OpenMoko) was to have both ppp and multiplexing in the kernel to avoid the context switches when a lot of data is being sent, but this is not very portable and causes lots of other problems. Having both in a single non-kernel process is going to be almost as fast and much more portable at the expense of reimplementing part of pppd.

Regards

#7 lore raw
Quoted 10 lines

Hi,

2009/11/13 Chris Pitchford <[email protected]>:

Quoted 2 lines

Is the line discipline so much of an issue when using ppp connected to a fake tty? could that not be handed on the proxy side inside ofono?

I've not tried it on a non-tty but it definitely should work modulo potential bugs that need to be squashed. It's supposed to work.

Quoted 8 lines

However by using a separate process to do ppp you add overhead. The original idea in ofono as well as in other projects (think OpenMoko) was to have both ppp and multiplexing in the kernel to avoid the context switches when a lot of data is being sent, but this is not very portable and causes lots of other problems. Having both in a single non-kernel process is going to be almost as fast and much more portable at the expense of reimplementing part of pppd.

I've had another thought.. and it doesn't push for pppd, but it introduces another idea.

When connecting to an APN for MMS services, there is the possibility of a clash of IP addresses both with other APN connections and with other internet connections (for example wifi)..

For example: wlan0 192.168.1.20/24 default gw 192.168.1.250

ppp0/tun0: 10.1.1.1 <=> 10.255.255.255

and the MMSC (to which the phone will need to communicate to send/receive MMS) is 192.168.1.20

bam.. it clashes with a locally assigned IP address and therefore no policy routing will allow a process to reach 192.168.1.20 via the MMS interface (or atleast not with excessive engineering)

Other clashes may cause problems too..

To solve the problem, the MMS environment could be run in its own kernel network name space. This means the ppp/tun interface would be separate from the other interfaces.. this removes the clashing IP address space problem.

If pppd were used, it could be "wrapped" in a different address space meaning the ppp interface would be localised to an environment that carried out all of the MMS communications. What's more, the work would require no change to the existing global address space.. it very neatly keeps the main environment seperate from the MMS environment that requires its own gateway and IP address space..

I appreciate the chances of a clash are slim, but with telcos using RFC-1918 and home users mainly using the same the chances of a clash are far above zero (and I hear it is a problem for iphone users.. no dig at iphones but a colleage reports seeing MMS traffic over his wireless connection.. not confirmed but still)

So, if ofonod is responsible for creating tun interfaces, I'm not sure it would be possible to isolate them within their own name space and therefore allow concurrent "internet" and "mms" access via two different connections..

I'm still formulating my thoughts on this one, but I hope it is food for thought!

cp

#8 lore raw

Le samedi 14 novembre 2009 21:26:38 Chris Pitchford, vous avez écrit :

Quoted 3 lines

When connecting to an APN for MMS services, there is the possibility of a clash of IP addresses both with other APN connections and with other internet connections (for example wifi)..

You'd have to implement TCP/IP in userspace also, which wouldn't be too nice. With recent kernels (2.6.29 and up), you'd rather want to create a new network kernel namespace, containing the MMS interface and the one or few processes sending or receiving MMS.

That avoids the IP range conflicts while not needing to reimplement TCP/IP. It also still works if the modem exposes GPRS through sone kernel driver such that you cannot cleanly send and receive IP packets in userspace anyway. Currently, there aren't any in oFono, as both PPP and Phonet can provide IP packets to userspace. But that does not mean there won't ever be any.

Signature · 2 lines
Rémi Denis-Courmont
http://www.remlab.net/