[PATCHES] CTDB: improved IPv6 support

[Amitay Isaacs]

Hi Martin,

On Thu, Dec 4, 2014 at 11:27 PM, Martin Schwenke <martin at meltin.net> wrote:

> On Thu, 4 Dec 2014 17:42:51 +1100, Martin Schwenke <martin at meltin.net>
> wrote:
>
> > The attached patch series fixes most problems with IPv6 in CTDB.
> >
> > The most notable things are:
> >
> > * For machine-readable output from the "ctdb" CLI tool, effectively
> >   replace "ctdb -Y ..." with "ctdb -X ..." which uses '|' as the field
> >   delimiter.  For those that don't want '|', they can use
> >   "ctdb -x<char> ...".
> >
> >   "ctdb -Y ..." is still supported but all documentation and scripts
> >   now use "ctdb -X".
> >
> >   Mathieu, you're CC:ed because this includes a minor change to the
> >   nagios script, which I think is correct but it isn't yet tested. :-)
> >
> > * When the CTDB daemon loads the nodes file it stores a string
> >   representation of each node's IP address.  This representation is now
> >   first converted to canonical form so that IPv6 addresses will
> >   compare more reliably.
> >
> > * Some functionality has moved from the daemon to the 10.interfaces
> >   eventscript and it now works with IPv6 addresses.
> >
> >   For the past couple of years the release IP code in the daemon has
> >   depended on being able to determine which interface an IP address is
> >   on, rather than trusting vnn->interface.  This was done to more
> >   reliably be able to remove rogue IP addresses (present due to
> >   either race conditions or addresses that have been moved by hand).
> >   This is not as easy for IPv6 addresses since the SIOCGIFCONF ioctl
> >   does not support IPv6.
> >
> >   Therefore, this code has been removed and vnn->interface is now
> >   passed to the "releaseip" event, even when it is NULL (in which case
> >   "__none__" is passed).  In all cases the eventscript now uses some "ip
> >   addr show to ..." magic to determine the actual interface and netmask
> >   for an IP address being released.  If either of these differ from
> >   what was passed then a warning is logged, and the actual interface
> >   and netmask are used to delete the IP address.
> >
> >   I don't think we lost any functionality.  The main part of this is in
> >   these patches:
> >
> >     32c2eab ctdb-scripts: Make 10.interface IPv6-safe
> >     837f36a ctdb-daemon: Trust vnn->interface for an IP when releasing it
> >
> > * Gratuitous ARP equivalent for IPv6 has been changed to use neighbor
> >   advertisements.  This seems to work reliably.
> >
> > * Due to duplicate address detection in IPv6, the IP address addition
> >   code will try to wait until IPv6 addresses are no longer
> >   "tentative".  If this times out or "dadfail" occurs then the IP
> >   address is removed and the operation fails.  Note that this could
> >   cause an outage (all nodes unhealthy, one at a time) if a rogue node
> >   refuses to release IPs or if CTDB just dies.  It isn't possible to
> >   easily work around this, so it is a design limitation and may need to
> >   be looked at again in the future.
> >
> > This has been tested on a virtual cluster with IPv6 addresses only, on
> > both public and private networks.
> >
> > Please review and push if OK.
> >
> > These patches are in my ctdb-ipv6 branch:
> >
> >
> http://git.samba.org/?p=martins/samba.git;a=shortlog;h=refs/heads/ctdb-ipv6
> >
> > Still to do:
> >
> > * Update policy routing to work with IPv6.  There's some code in there
> >   that only works for IPv4.
> >
> > * Other features like LVS.
>
> Darn.  This looks like it was big enough to be moderated, so...
>
> These patches:
>
> b615d09f ctdb-tools: Produce machine readable output with new function
> printm()
> fdf0461 ctdb-tools: Add -x option to specify delimiter for machine
> readable output
> 42498ee ctdb-tools: Add -X option for machine parsable output with
> separator '|'
> 8485235 ctdb-scripts: Update eventscripts to use ctdb -X instead of ctdb -Y
> 435c200 ctdb-tools: Update onnode and ctdb-diagnostics to use ctdb -X
> 8f1ebc5 ctdb-tests: Update integration tests to use ctdb -X
> 443aa32 ctdb-tool: Fix "ctdb -Y ifaces" output to have trailing delimiters
> c2e39e4 ctdb-doc: Update examples to use ctdb -X
> 1ea9d2e ctdb-utils: Update Nagios code to use ctdb -X
> c6a448c ctdb-scripts: Add IPv6 addresses support in ip_maskbits_iface()
> 9e6d323 ctdb-scripts: New functions ip6tables() and iptables_wrapper()
> 32c2eab ctdb-scripts: Make 10.interface IPv6-safe
> 837f36a ctdb-daemon: Trust vnn->interface for an IP when releasing it
> b7debd7 ctdb-eventscripts: Specify broadcast optionally to ip addr add
> 38a5617 ctdb-scripts: Wait until IPv6 addresses are not "tentative"
> 7944a40 ctdb-daemon: Fix IP address comparisons for IPv6 addresses
> 51ec7d3 ctdb-tools: Bracket IP addresses in onnode (for IPv6)
> 9d4137d ctdb-tests: Extend regexp to match IPv6 addresses
>


> c9f3359 ctdb-tests: Try to handle IPv6 addresses for local daemons
>

In this patch, we cannot change the node IP addresses from 127.0.0.x to
127.0.234.x.  Looks like socket wrapper does not like that, socket wrapper
expects addresses in 127.0.0.0/24 range.  So here's a fixup to move node
IPs away from 127.0.0.1.

diff --git a/ctdb/tests/simple/scripts/local_daemons.bash
b/ctdb/tests/simple/scripts/local_daemons.bash
index 0131950..7d35a8f 100644
--- a/ctdb/tests/simple/scripts/local_daemons.bash
+++ b/ctdb/tests/simple/scripts/local_daemons.bash
@@ -76,7 +76,8 @@ setup_ctdb ()
                echo "fc00:10::1:$(($i + $TEST_LOCAL_DAEMONS))/64 lo"
>>"$public_addresses_all"
            fi
        else
-           echo 127.0.234.$i >>"$CTDB_NODES"
+           j=$(( $i + 10))
+           echo 127.0.0.$j >>"$CTDB_NODES"
            # 2 public addresses on most nodes, just to make things
interesting.
            if [ $(($i - 1)) -ne $no_public_ips ] ; then
                echo "192.168.234.$i/24 lo" >>"$public_addresses_all"



> 875f60e ctdb-tests: Bracket IP addresses in NFS mounts and scp command
> (for IPv6)
> befdf53 ctdb-tests: Extend regexps to handle IPv6 address matching
> b2df1a3 ctdb-tests: Use ping_wrapper to do relevant ping or ping6
> 1cc75ea0 ctdb-tests: Match IPv6 connections in netstat output
> 928b88a ctdb-tests: Generalise the gratarp and tickle sniffing code for
> IPv6
> d546cba ctdb-tests: Use ip neigh command instead of arp
> 07b8aa1 ctdb-tests: Make tcpdump output more verbose
> 2ed1f6b ctdb-tests: More debug on SSH failure
> 76e83e5 ctdb-daemon: Gratuitous ARP equivalent for IPv6 is neighbor
> advertisement
> 242442a ctdb-tests: Need to drop public IPs in kill-failover tests
>
> from my ctdb-ipv6 branch, per above gitweb link.
>
> As an aside, does someone want to add me as a list moderator?  ;-)
>
> peace & happiness,
> martin
>

I have pushed the extra patch and the rest of the patches with Reviewed-By
tag to ctdb-ipv6 branch in my tree.

Amitay.