LARTC - Feb 2004 - Paid project/Netfilter target to match queue backlog

Hello,

If any of kernel hackers wants to make some money:

For 300$:

What I need is a netfilter match rule that would match on depth of a queue
on a given device. Example:

iptables -t mangle -I PREROUTING -m queue_backlog --device eth0.1
--depth-gt 10 -j <chain>

If you can complete this in a reasonable time (2-3 days), let me know.

This does not need to be very complicated: The only qdiscs that this needs 
to work with is HTB and pfifo. 

I envision this as adding a function query_backlog to relevant qdiscs, and 
a simple match on dev->qdisc->query_backlog

For someone who has done this before, I''m sure should take about a day
:)

This *must* be optimized (i.e. pointer to device struct must be cached,
number of packets in backlog must be cached or otherwise computed at O(1) 
time).

To those wondering "why in the world would someone want to do that":

Consider router connected to many uplinks (possibly of varying speeds).  
Due to "political reasons" reasons (example, one link being "paid
for
completely", others being "pay per use"), I only want to send
packets over
the other link when the first link is "full". Backlog length provides
the
best measure of the "fullness" of a given device.

Now (use of ''recent'' match necessary because TCP really
dislikes
multipath within a given session):

-A PREROUTING -m recent --rdest -j MARK --seconds 60 --set-mark 1
-A PREROUTING -m queue_backlog --device eth0.1 --depth-gt 10  -j OVERFLOW
-A OVERFLOW -j MARK --set-mark 1
-A OVERFLOW -m recent --rdest --set -j ACCEPT 

...ip rule add fwmark 1 table 2

Yes, I know its somewhat fugly (essentially, I''m doing RPDB as
netfilter
rules), but seems simplest and more flexible to me. 







_______________________________________________
LARTC mailing list / LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/

> -A PREROUTING -m recent --rdest -j MARK --seconds 60 --set-mark 1
> -A PREROUTING -m queue_backlog --device eth0.1 --depth-gt 10  -j OVERFLOW
> -A OVERFLOW -j MARK --set-mark 1
> -A OVERFLOW -m recent --rdest --set -j ACCEPT
Isn''t the same possible in /lartc.org 12.3.2. Overlimit actions /  and
"reclassify" option /which i din''t used yet/.

Consider you have linux which has 2 default routes - one with TOS=0x0, and
one with tos 0x10.
so:
0.0.0.0/0 tos 0 via eth0
0.0.0.0/0 tos 0x10 via eth1
your_source=eth2

next, you must use IMQ to pass every traffic from eth2, /because as i''m
thinking there isn''t possible to do reclassify packet which is already
routed to some interface, and next put the same packet over other interface
based on some policy - PLEASE CORRECT ME!!! /

I don''t know how to put all traffic which is destined for eth0, to some
IMQ
device, next police packets there by "reclassify" options.
All the settings must be according to your full bandwidth of eth0, so all
which is overlimit eth0 speed get reclassify to other TOS.
After all packets which leaves imq, are routed again by system,
if you have differented TOS routing, /if your IGP routing domain create
such Traffic Engineered path/ then packets with such other TOS will get
outside from eth1.

/this is my teory, to correct or discuss /

This what you are trying to do is somehow Traffic Engineering or Contraint
Based Routing. I''m very pleased to be with contact with somebody who is
intend/or already done with some work with linux in this topic.

Now i elaborate how to control distribution of MPLS labels which must be
according to current interface load or even interface load with each QoS
class /in linux/.

Feel free to contact me with this topic.

Arkadiusz Binder

_______________________________________________
LARTC mailing list / LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/

> Isn''t the same possible in /lartc.org 12.3.2. Overlimit actions /
and
> "reclassify" option /which i din''t used yet/.
Interesting idea - Let me try if this works first.
> next, you must use IMQ to pass every traffic from eth2, /because as
i''m
> thinking there isn''t possible to do reclassify packet which is
already
> routed to some interface, and next put the same packet over other
> interface based on some policy - PLEASE CORRECT ME!!! /
Only way to find out is to try it, I think.
> I don''t know how to put all traffic which is destined for eth0, to
some IMQ
> device, next police packets there by "reclassify" options.
> All the settings must be according to your full bandwidth of eth0, so all
> which is overlimit eth0 speed get reclassify to other TOS.
> After all packets which leaves imq, are routed again by system,
> if you have differented TOS routing, /if your IGP routing domain create
> such Traffic Engineered path/ then packets with such other TOS will get
> outside from eth1.
> 
> /this is my teory, to correct or discuss /
> 
> This what you are trying to do is somehow Traffic Engineering or
> Contraint Based Routing. I''m very pleased to be with contact with
> somebody who is intend/or already done with some work with linux in this
> topic.
Correct - that''s exactly what I''m trying to do.
> Now i elaborate how to control distribution of MPLS labels which must be
> according to current interface load or even interface load with each QoS
> class /in linux/.
That''s slightly above the scope of what I need right now - but it is 
something I will need soon.


_______________________________________________
LARTC mailing list / LARTC@mailman.ds9a.nl
http://mailman.ds9a.nl/mailman/listinfo/lartc HOWTO: http://lartc.org/