Andy Gospodarek wrote:

Can you elaborate on what isn't going well with this driver/hardware?  

I have a ppc64 blade running a customized 2.6.10.  At init time, two of 
our gigE links (eth4 and eth5) are bonded together to form bond0.  This 
link has an MTU of 9000, and uses arp monitoring.  We're using an 
ethernet driver with a modified RX path for jumbo frames[1].  With the 
stock driver, it seems to work fine.

The problem is that eth5 seems to be bouncing up and down every 15 sec 
or so (see the attached log excerpt).  Also, "ifconfig" shows that only 
3 packets totalling 250 bytes have gone out eth5, when I know that the 
arp monitoring code from the bond layer is sending 10 arps/sec out the link.


eth5      Link encap:Ethernet  HWaddr 00:03:CC:51:01:3E
           inet6 addr: fe80::203:ccff:fe51:13e/64 Scope:Link
           UP BROADCAST RUNNING SLAVE MULTICAST  MTU:9000  Metric:1
           RX packets:119325 errors:90283 dropped:90283 overruns:90283 
frame:0
           TX packets:3 errors:0 dropped:0 overruns:0 carrier:0
           collisions:0 txqueuelen:1000
           RX bytes:8978310 (8.5 MiB)  TX bytes:250 (250.0 b)
           Base address:0x3840 Memory:92220000-92240000


I had initially suspected that it might be due to the "u32 jiffies" 
stuff in bonding.h, but changing that doesn't seem to fix the issue.

If I boot the system and then log in and manually create the bond link 
(rather than it happening at init time) then I don't see the problem.

If it matters at all, normally the system boots from eth4.  I'm going to 
try booting from eth6 and see if the problem still occurs.


Chris




[1] I'm not sure if I'm supposed to mention the specific driver, as it 
hasn't been officially released yet, so I'll keep this high-level. 
Normally for jumbo frames you need to allocate a large physically 
contiguous buffer.  With the modified driver, rather than receiving into 
a contiguous buffer the incoming packet is split across multiple pages 
which are then reassembled into an sk_buff and passed up the link.