Deleting incoming network packets / sk_buff

Hi all.

I'm just making my first steps in kernel programming, and I have a few 
questions.

It's necessary for me to "delete" network packets that come in on a 
ethernet device after passing them to an userspace program (through a 
packet socket that the program has opened). My first idea was to write a 
kernel module that registers on the NF_IP_PRE_ROUTING netfilter hook and 
does its job. But there are two downsides on this approach:

1. I need to delete every packet, no matter if it is IP or something 
else. As far as I understood only IP-packets will pass the netfilter 
hooks, since the netfilter framework only works with IP traffic.

2. The solution also has to work for 2.2.x kernels, as long as there is 
any way to achieve this functionality for these kernel series and 2.4.x 
as well.

I continued to dig some documentation, and found a possible solution in 
a rather old phrack article [1]: "What we do is code our own kernel 
module that registers our packet_type{} data structure to handle all 
incoming packets (sk_buff's) right after they come out of the device 
driver." While the authors use this technique to implement some 
"protocol obfuscation" or an in-kernel packet sniffer, it seems to be a 
good way to reach what I want to do.

Now, what I'm wondering about is the way to "correctly" remove an 
sk_buff so that the upper layer won't see the corresponding packet. I 
somewhere read that the code for packet sockets have to clone the 
sk_buff before passing it to the socket, else it would be "lost" and 
could not be seen by following handlers.

So, the questions I have:

1. What has to be done to delete a bypassing sk_buff, so that upper 
layer routines don't see it?

2. How can it be achieved to pass a sk_buff to opened packet sockets 
before deleting the sk_buff? I guess the handler either has to pass the 
sk_buff "manually" to the packet socket handler packet_rcv() or has to 
place itself to be called directly after packet_rcv() has been called.

I hope that my description can be understood - not enough that I'm not 
very keen with kernel programming, but I also am no native english 
speaker... :) Please let me know if you need further information to clarify.

Bye, Mike

[1] http://www.phrack.org/show.php?p=55&a=12

Michael Renzmann wrote:

Hi all.

I'm just making my first steps in kernel programming, and I have a few 
questions.

It's necessary for me to "delete" network packets that come in on a 
ethernet device after passing them to an userspace program (through a 
packet socket that the program has opened). My first idea was to write a 
kernel module that registers on the NF_IP_PRE_ROUTING netfilter hook and 
does its job. But there are two downsides on this approach:

1. I need to delete every packet, no matter if it is IP or something 
else. As far as I understood only IP-packets will pass the netfilter 
hooks, since the netfilter framework only works with IP traffic.

2. The solution also has to work for 2.2.x kernels, as long as there is 
any way to achieve this functionality for these kernel series and 2.4.x 
as well.

I continued to dig some documentation, and found a possible solution in 
a rather old phrack article [1]: "What we do is code our own kernel 
module that registers our packet_type{} data structure to handle all 
incoming packets (sk_buff's) right after they come out of the device 
driver." While the authors use this technique to implement some 
"protocol obfuscation" or an in-kernel packet sniffer, it seems to be a 
good way to reach what I want to do.

Now, what I'm wondering about is the way to "correctly" remove an 
sk_buff so that the upper layer won't see the corresponding packet. I 
somewhere read that the code for packet sockets have to clone the 
sk_buff before passing it to the socket, else it would be "lost" and 
could not be seen by following handlers.

There is no way that I am aware of to block packets from high-level protocols,
but you can disable the higher protocols (ie, set IP to 0.0.0.0).

You can also hack net/core/dev.c to stop the packets from going
up the stack, but it requires more than just a new kernel module,
you'd have to actually change dev.c.  (Search for BRIDGE).

Ben

-- 
Ben Greear [off-list ref]
Candela Technologies Inc  http://www.candelatech.com

Hi Ben.

First of all, thanks for your answer.

Ben Greear wrote:

There is no way that I am aware of to block packets from high-level 
protocols, but you can disable the higher protocols (ie, set IP to 

 > 0.0.0.0).

Unfortunately this doesn't work for my setup. I have some kind of 
userspace bridge implementation running, which copies frames incoming on 
a physical device (e.g. eth1) to a virtual (tun/tap) device (e.g. tap0) 
and vice versa. When the kernel wants to communicate with the outside 
world it does this by using tap0, and the program "bridges" the packets 
to eth1. In order to avoid to have eth1 running in promiscious mode I 
configure tap0 to have the same MAC address as eth1 - but this has the 
drawback that the kernel sees every incoming packet twice: once when it 
hits eth1, and once when it gets copied by the userspace bridge to tap0.

There is no simple way to have the incoming communication "split" over 
both interfaces without running in serious troubles regarding firewall 
rules or such things as having quagga/zebra dropping incoming OSPF 
announces from eth1 (since it self uses tap0 to communicate with the 
outside world).

Abd yes, it is necessary to have this host only using tap0 for outgoing 
communication (it would be longish to explain the reason, just believe 
me for now ;)).


Bottom line: the easiest way to solve this problem would be to have some 
way for blocking packets on eth1 from being processed (after they have 
been passed to the packet socket - else my userspace program won't be 
able to "bridge" incoming packets to tap0).

You can also hack net/core/dev.c to stop the packets from going
up the stack, but it requires more than just a new kernel module,
you'd have to actually change dev.c.  (Search for BRIDGE).

Hmm, now that you mention it: at least the 2.4.x kernels I use have the 
bridging code enabled. I'll take a look at possible hooks or other 
mechanisms that are used by the bridging code for its job - maybe this 
will help. Thanks for the hint.

Bye, Mike

Michael Renzmann wrote:

Bottom line: the easiest way to solve this problem would be to have some 
way for blocking packets on eth1 from being processed (after they have 
been passed to the packet socket - else my userspace program won't be 
able to "bridge" incoming packets to tap0).

Can't you have no IP on eth1 and have your IP on tap0 instead?  That would
at least keep any pkts received on eth1 from being processed by the
IP stack.

Ben


-- 
Ben Greear [off-list ref]
Candela Technologies Inc  http://www.candelatech.com

Hi.

Ben Greear wrote:

quoted

Bottom line: the easiest way to solve this problem would be to have 
some way for blocking packets on eth1 from being processed (after they 
have been passed to the packet socket - else my userspace program 
won't be able to "bridge" incoming packets to tap0).

Can't you have no IP on eth1 and have your IP on tap0 instead?  That would
at least keep any pkts received on eth1 from being processed by the
IP stack.

This is infact what I have done. The result is still that every packet 
will be dup'ed - which can be seen for example when pinging from that 
host, every reply is shown twice (and ping says "DUP" for every reply).

Bye, Mike