[oops] with FIB_TRIE

[Robert Olsson <robert@robur.slu.se>]

Andrei Popa writes:

 > I've used an vanilla 2.6.28.7 kernel without any additional patches with
 > the following .config and when I do in quagga a "clear ip bgp * soft"
 > when I have three full BGP sessions the kernel it oopses.
 > 
 > With FIB lookup algorithm FIB_TRIE it oopeses. With FIB_HASH it doesn't.
 > 
 > Pictures with the oops:
 > http://89.33.136.9/oops/

 > The config file:
 > CONFIG_PREEMPT=y

 Hello,

 Getting somewhat worried as we use this for infrastructure since many years. 
 I've set up test and is trying to reproduce it. 
 
 I'm running forwarding ~9.4 Gigabit/s @ 1.2 pkts sec and fib_lookups 40-200.000
 lookups per sec. Routing table has 280.000 entries this is loaded/unloaded 
 via ip route with -batch to give load for insert/delete.

 A script is continuesly adding/removing the routing table under this load. 
 The time to install the full table is ~10 sec and same time to remove
 (without netfilter ~5 sec) And this during this constant traffic load.

 The scripts and routing tables:
 ftp://robur.slu.se/pub/Linux/net-development/trie-test/

 Drivers are niu, ixgbe Netfilter modules loaded but no filters. Kernel 
 2.6.29-r2. 

 No problems seen for 3 hours but I'll let this run overnight

 One difference to your config I see is PREEMPT. We use use 
 CONFIG_PREEMPT_NONE=y with the router/servers.


 Cheers
					--ro