Bug 1543902

Hello,

The "dracut_args --include /etc/hosts /etc/nsswitch.conf" workaround you mentioned above does not work,mainly because the side effect of that entry line results in the content of file /etc/hosts being installed as /etc/nsswitch.conf. 

After inspecting how /usr/sbin/mkdumprd build the dracut_args array variable it receives as first argument to dracut I realized that probably you intended to use --install instead of --include.

The former allows you to install the content of several files into the target kdump ramdisk, while the latter can only be used once, assuming that what is stated in dracut man pages is correct.

Indeed, I checked that using several --include options (it does not matter if they appear in an unique line or in several dracut_args entries in /etc/kdump.conf), results in dracut complaining about syntax issues and aborting.

On the other side, using --install instead of --include takes you to a correct, consistent and complete ramdisk-kdump image:

This is the line I used in /etc/kdump.conf:

dracut_args --install "/etc/nsswitch.conf" --install "/etc/hosts"


Jesús.

(In reply to Josef Zimek from comment #0)
> Description of problem:
> 
> fence_kdump_nodes is required parameter in case fence_kdump is used. If user
> doesn't specify fence_kdump_nodes in /etc/kdump.conf the list of cluster
> nodes is obtained from cluster automatically and that value is passed to
> fence_kdump_nodes:
> 
> 
> # retrieves fence_kdump nodes from Pacemaker cluster configuration
> get_pcs_fence_kdump_nodes() {
>     local nodes
> 
>     # get cluster nodes from cluster cib, get interface and ip address
>     nodelist=`pcs cluster cib | xmllint --xpath
> "/cib/status/node_state/@uname" -`
> 
>     # nodelist is formed as 'uname="node1" uname="node2" ... uname="nodeX"'
>     # we need to convert each to node1, node2 ... nodeX in each iteration
>     for node in ${nodelist}; do
>         # convert $node from 'uname="nodeX"' to 'nodeX'
>         eval $node
>         nodename=$uname
>         # Skip its own node name
>         if [ "$nodename" = `hostname` -o "$nodename" = `hostname -s` ]; then
>             continue
>         fi
>         nodes="$nodes $nodename"
>     done
> 
>     echo $nodes
> }
> 
> 
> Problem with above approach is that cluster node names from pacemaker doen't
> necessarily have to be FQDN but can be user defined aliases. So this way we
> feed fence_kdump_nodes with list of aliases which are typically mapped in
> /etc/hosts but /etc/hosts is not included in initramfs by dracut by default.
> There are multiple reasons why it is not included and there is low change
> get it there by default. However it make sense to include /etc/hosts (and
> /etc/nsswitch.conf to make it work) when fence_kdump is used because without
> that hostname aliases are not resolvable from crashkernel which leads to
> fence_kdump failure (fence_kdump_send fails to send notifications and
> fencing results in timeout = vmcore doesn't get captured).
> 
Hi, do you hit this issue? During mkdumprd, it stores the ipaddr, instead of the hostname. Hence alias will not cause the problem.

Thanks,
Pingfan

But, it is harmless to include hosts and nsswitch.conf. If debugging in kdump shell, we try to scp something to alias.

My test environment is broken. After adding rd.debug in cmdline, I found that fence_kdump_send cmd has no node as its input param. Hence during my test, I did not hit the condition (fence_kdump_send fails to send notifications and fencing results in timeout = vmcore doesn't get captured).

It turns out that we need this fix.

Thanks,
Pingfan

Since the problem described in this bug report should be
resolved in a recent advisory, it has been closed with a
resolution of ERRATA.

For information on the advisory, and where to find the updated
files, follow the link below.

If the solution does not work for you, open a new bug report.

https://access.redhat.com/errata/RHBA-2018:3240