Bug 459879

What exactly do you think is going wrong?  bond1 should be passed into find_activate_slaves as arg 1, and as a result we should find all physical interfaces for which we are attached to  that bond.  Your patch seems like a hack to make your environment work properly (especially since you hardcode bond0  into your definition of OLD_MASTER.  If you would please send me your kdump.conf, and the initramfs image you produce from it.  Thanks!

Created attachment 314951 [details]
kdump.conf

Created attachment 314953 [details]
inird for the crash kernel

Created attachment 314954 [details]
Correct one: initrd for the crash kernel

Attached 314953 is incorrect. Please ignore it. This is the correct one.

Please re-check how the "init" script in the initrd for the crach kernel (initrd-<Version>kdump.img) handles bonding devices. What I found was that:

1. bondX (a bond master in the normal kernel. bond1 in my case) is forcefully converted to bond0 (that is hardcoded) with the "map_interface" script.

* Corresponding part of "init" (in initrd-<Version>kdump.img)
------------
for i in `ls /etc/ifcfg-*`
do
   NETDEV=`echo $i | cut -d"-" -f2`
   map_interface $NETDEV
done
rename_interfaces
IFACE=`cat /etc/iface_to_activate`
ifup $IFACE
------------

* "scriptfns/map_interface" called from "init".
------------
. /etc/ifcfg-$NETDEV
for j in `ifconfig -a | awk '/.*Link encap.*/ {print $1}'`
do
    case "$BUS_ID" in
    Bonding)
        REAL_DEV=bond0
        RENAMED="yes"
        ;;
...
#build the interface rename map
echo $NETDEV $REAL_DEV tmp$TMPCNT >> /etc/iface_map
TMPCNT=`echo $TMPCNT 1 + p | dc`
echo $TMPCNT > /tmp/tmpcnt
echo mapping $NETDEV to $REAL_DEV
-------------

2. As a result, bond0 (instead of the original bond master, bond1 in my case) is always activated as a master, and "bond0" is passed into find_activate_slaves as arg 1.

3. However, since "ifcfg-eth1, ifcfg-eth2" still contains "MASTER=bond1", it fails to be enslaved to the active bond0.

Note that, another problem is that find_activate_slaves tries to read ifcfg-eth0 first and failes (as ifcfg-eth0 doesn't exist in the crash kernel environment.) Then, find_activate_slaves is aborted there. This is why I added 
+    if [ -f /etc/ifcfg-\$j ];
in the patch.

See the attached (314951, 314954) for kdump.conf and initrd (for crash kernel) which was built with the original (non-patched) mkdumprd.

The concerns you have in note 2 is supposed to be taken care of in the rename_interfaces function.  Although looking at it, it seems like that function misses handling slaves on bonded interfaces.  I'll write a patch...

Created attachment 315787 [details]
patch to update MASTER pointers for slaves on bonded interfaces

I've not tested it yet, but I think this will update what your missing.  Please let me know if it solves your problems.  Thanks!

Created attachment 315851 [details]
fixing typo and the second problem for /sbin/mkdumprd

Thanks Neil. 

I found two typos in your patch:

+    IS_BOND=\`echo /etc/ifcfg-\$NEW | grep bond\`
+    if [ -n "$IS_BOND" ] #### shoud be =====> if [ -n "\$IS_BIND" ]
+    then
+        for i in \`ls /etc/ifcfg-*\`
+        do
+            sed -e"s/.*MASTER=\$CURRENT.*/MASTER=\$NEW/" \$i > /tmp/ifcfg-tmp
+            mv /sbin/ifcfg-tmp \$i  #### shold be =====> mv /tmp/ifcfg-tmp \$i
+        done
+    fi

And there still remains the second problem:

find_activate_slaves finds device names ethXX from 'ifconfig' output and tries to read the corresponding /etc/if-ethXX, but it doesn't always exist the corresponding one. Hence, when find_activate_slaves tries to read non-existing /etc/ifcfg-ethX, if failes and find_activate_slaves is aborted there. (This may be a particular behaviour of the busybox shell.) 

See the attached (id=315851) for the patch fixing the typos and this problem. It worked in my environment.

Like I said, I hadn't tested it.  What is the result of your testing after fixing the typos?

As for your patch, As I've noted, I still don't like it for the reasons I gave previously. However, given that you seem adamant on it, I'll consider it if you can test it and show that it works in other cases as well.  Speciically if it works in the trivial case (where there is only one bonded interface bond0 in the entire system runing under a normal kernel), and in the case where there is only one bonded interface without the normal name (say bondtest).  Then I'll look into taking it.

No, no. I'm not sticking to the original patch (id=314866), let's throw it away.

Please look at the contents of my last patch (id=315851).

https://bugzilla.redhat.com/attachment.cgi?id=315851

It's just a slight modification of your patch (id=315787) in the follwing ways.

- Fixing typos.
- Add a fix to the another problem below (other than the one fixed by your patch.)
------------------
find_activate_slaves finds device names ethXX from 'ifconfig' output and tries
to read the corresponding /etc/if-ethXX, but it doesn't always exist the
corresponding one. Hence, when find_activate_slaves tries to read non-existing
/etc/ifcfg-ethX, if failes and find_activate_slaves is aborted there. (This may
be a particular behaviour of the busybox shell.) 
------------------

And my testing result is that.

Your patch (id=315787) didn't work becase of the typos and the another problem.
The modified version of your patch (id=315851) worked well.

Does it make sense? Thanks.

Ahh, sorry, I missed that attachment.  Yes, what you have there makes sense to me, I'll check that it as soon as I can.

Thanks!

An advisory has been issued which should help the problem
described in this bug report. This report is therefore being
closed with a resolution of ERRATA. For more information
on therefore solution and/or where to find the updated files,
please follow the link below. You may reopen this bug report
if the solution does not work for you.

http://rhn.redhat.com/errata/RHBA-2009-0105.html