Re: Kernel Panic with bonding + IPoIB on 3.2.9

Joseph Glanville [off-list ref] wrote:

On 20 March 2012 06:05, Roland Dreier [off-list ref] wrote:

quoted

On Sun, Mar 18, 2012 at 1:21 PM, Joseph Glanville
[off-list ref] wrote:

quoted

[  422.047024] kernel BUG at net/core/dev.c:1896!

So this line is

       BUG_ON(offset >= skb_headlen(skb));

right?  No paritcular idea how we hit this, though...

Yep... I have looked through most of /drivers/net/bonding and I can't
really see why it should be blowing up there.. it really should cause
the BUG_ON under normal IPoIB if the MTU was the cause - yet I have
not experienced this.
The bonding code doesn't seem to do anything special with the MTU
other than propagating changes to the slaves.

	For IPoIB, though, there is some extra initialization stuff in
bond_setup_by_slave(), and the hard_header_len will end up being set to
something different from the usual Ethernet value.

	In looking at ipoib_setup, I see that hard_header_len appears to
be set to 4 (IPOIB_ENCAP_LEN).  My recollection was that the IPoIB
hard_header_len was quite a bit larger than that; it looks like it
changed very recently from IPOIB_ENCAP_LEN + INFINIBAND_ALEN to what it
is now:

commit afd87adacb5de00768b2e54f0bd851278f2e6179
Author: Roland Dreier [off-list ref]
Date:   Tue Feb 7 14:51:21 2012 +0000

    IPoIB: Stop lying about hard_header_len and use skb->cb to stash LL addresses
    
    [ Upstream commit 936d7de3d736e0737542641269436f4b5968e9ef ]
    
    Commit a0417fa3a18a ("net: Make qdisc_skb_cb upper size bound
    explicit.") made it possible for a netdev driver to use skb->cb
    between its header_ops.create method and its .ndo_start_xmit
    method.  Use this in ipoib_hard_header() to stash away the LL address
    (GID + QPN), instead of the "ipoib_pseudoheader" hack.  This allows
    IPoIB to stop lying about its hard_header_len, which will let us fix
    the L2 check for GRO.


	I don't know if this change could be causing the problem (it
appears to be new in 3.2.9), but the hard_header_len is one of the few
areas in the TX path of bonding that IPoIB ends up being different from
regular Ethernet.

	-J

---
	-Jay Vosburgh, IBM Linux Technology Center, fubar@us.ibm.com