Earlier today I was asked how one would go about debugging early startup of KVM under GDB, when launched by libvirtd. It was not possible to simply attach to KVM after it had been launched by libvirtd, since that was too late. In addition running the same KVM command outside libvirt did not exhibit the problem that was being investigated.
Fortunately, with a little cleverness, it is actually possible to debug a KVM guest launched by libvirtd, right from the start. The key is to combine a couple of breakpoints with use of follow-fork-mode. When libvirtd starts up a KVM guest, it runs QEMU a couple of times in order to detect which command line arguments are supported. This means the follow-fork-mode setting cannot be changed too early, otherwise GDB will end up following the wrong process.
I happen to know that there is only one place in the libvirt code which calls virCommandSetPreExecHook, and that is immediately before launching the real QEMU process. A nice thing about GDB is that when following forked/exec’d children, it will apply any existing breakpoints in the child, even if it is a new binary. So a break point set on ‘main’, while still in libvirtd will happily catch ‘main’ in the QEMU process. The only remaining problem is that if QEMU does not setup and activate the monitor quickly enough, libvirtd will try to kill it off again. Fortunately GDB lets you ignore SIGTERM, and even SIGKILL :-)
The start of the trick is this:
# pgrep libvirtd
12345
# gdb
(gdb) attach 12345
(gdb) break virCommandSetPreExecHook
(gdb) cont
Now in a separate shell
# virsh start $GUESTNAME
Back in the GDB shell the breakpoint should have triggered, allowing the trick to be finished:
(gdb) break main
(gdb) handle SIGKILL nopass noprint nostop
Signal Stop Print Pass to program Description
SIGKILL No No No Killed
(gdb) handle SIGTERM nopass noprint nostop
Signal Stop Print Pass to program Description
SIGTERM No No No Terminated
(gdb) set follow-fork-mode child
(gdb) cont
process 3020 is executing new program: /usr/bin/qemu-kvm
[Thread debugging using libthread_db enabled]
[Switching to Thread 0x7f2a4064c700 (LWP 3020)]
Breakpoint 2, main (argc=38, argv=0x7fff71f85af8, envp=0x7fff71f85c30)
at /usr/src/debug/qemu-kvm-0.14.0/vl.c:1968
1968 {
(gdb)
Bingo, you can now debug QEMU startup at your leisure
Yesterday I talked about setting up Sheepdog with KVM, so today is it is time to discuss use of Ceph and RBD with KVM.
Host Cluster Setup, the easy way
Fedora has included Ceph for a couple of releases, but since my hosts are on Fedora 14/15, I grabbed the latest ceph 0.3.1 sRPMs from Fedora 16 and rebuilt those to get something reasonably up2date. In the end I have the following packages installed, though to be honest I don’t really need anything except the base ‘ceph’ RPM:
# rpm -qa | grep ceph | sort
ceph-0.31-4.fc17.x86_64
ceph-debuginfo-0.31-4.fc17.x86_64
ceph-devel-0.31-4.fc17.x86_64
ceph-fuse-0.31-4.fc17.x86_64
ceph-gcephtool-0.31-4.fc17.x86_64
ceph-obsync-0.31-4.fc17.x86_64
ceph-radosgw-0.31-4.fc17.x86_64
Installing the software is the easy bit, configuring the cluster is where the fun begins. I had three hosts available for testing all of which are virtualization hosts. Ceph has at least 3 daemons it needs to run, which should all be replicated across several hosts for redundancy. There’s no requirement to use the same hosts for each daemon, but for simplicity I decided to run every Ceph daemon on every virtualization host.
My hosts are called lettuce, avocado and mustard. Following the Ceph wiki instructions, I settled on a configuration file that looks like this:
[global]
auth supported = cephx
keyring = /etc/ceph/keyring.admin
[mds]
keyring = /etc/ceph/keyring.$name
[mds.lettuce]
host = lettuce
[mds.avocado]
host = avocado
[mds.mustard]
host = mustard
[osd]
osd data = /srv/ceph/osd$id
osd journal = /srv/ceph/osd$id/journal
osd journal size = 512
osd class dir = /usr/lib64/rados-classes
keyring = /etc/ceph/keyring.$name
[osd.0]
host = lettuce
[osd.1]
host = avocado
[osd.2]
host = mustard
[mon]
mon data = /srv/ceph/mon$id
[mon.0]
host = lettuce
mon addr = 192.168.1.1:6789
[mon.1]
host = avocado
mon addr = 192.168.1.2:6789
[mon.2]
host = mustard
mon addr = 192.168.1.3:6789
The osd class dir bit should not actually be required, but the OSD code looks in the wrong place (/usr/lib instead of /usr/lib64) on x86_64 arches.
With the configuration file written, it is time to actually initialize the cluster filesystem / object store. This is the really fun bit. The Ceph wiki has a very basic page which talks about the mkcephfs tool, along with a scary warning about how it’ll ‘rm -rf’ all the data on the filesystem it is initializing. It turns out that it didn’t mean your entire host filesystem, AFAICT, it only the blows away the contents of the directory configured for ‘osd data‘ and ‘mon data‘, in my case both under /srv/ceph.
The recommended way is to let mkcephfs ssh into each of your hosts and run all the configuration tasks automatically. Having tried the non-recommended way and failed several times before finally getting it right, I can recommend following the recommended way :-P There are some caveats not mentioned in the wiki page though:
- The configuration file above must be copied to
/etc/ceph/ceph.conf on every node before attempting to run mkcephfs.
- The configuration file on the host where you run
mkcephfsmust be in /etc/ceph/ceph.conf or it will get rather confused about where it is in the other nodes.
- The
mkcephfscommand must be run as root since, it doesn’t specify ‘-l root’ to ssh, leading to an inability to setup the nodes.
- The directories
/srv/ceph/osd$i must be pre-created, since it is unable to do that itself, despite being able to creat the /srv/ceph/mon$idirectories.
- The Fedora RPMs have also forgotten to create
/etc/ceph
With that in mind, I ran the following commands from my laptop, as root
# n=0
# for host in lettuce avocado mustard ; \
do \
ssh root@$host mkdir -p /etc/ceph /srv/ceph/mon$n; \
n=$(expr $n + 1; \
scp /etc/ceph/ceph.conf root@$host:/etc/ceph/ceph.conf
done
# mkcephfs -a -c /etc/ceph/ceph.conf -k /etc/ceph/keyring.bin
On the host where you ran mkcephfs there should now be a file /etc/ceph/keyring.admin. This will be needed for mounting filesystems. I copied it across to all my virtualization hosts
# for host in lettuce avocado mustard ; \
do \
scp /etc/ceph/keyring.admin root@$host:/etc/ceph/keyring.admin; \
done
Host Cluster Usage
Assuming the setup phase all went to plan, the cluster can now be started. A word of warning though, Ceph really wants your clocks VERY well synchronized. If your NTP server is a long way away, the synchronization might not be good enough to stop Ceph complaining. You really want a NTP server on your local LAN for hosts to sync against. Sort this out before trying to start the cluster.
# for host in lettuce avocado mustard ; \
do \
ssh root@$host service ceph start; \
done
The ceph tool can show the status of everything. The ‘mon’, ‘osd’ and ‘msd’ lines in the status ought to show all 3 host present & correct
# ceph -s
2011-10-12 14:49:39.085764 pg v235: 594 pgs: 594 active+clean; 24 KB data, 94212 MB used, 92036 MB / 191 GB avail
2011-10-12 14:49:39.086585 mds e6: 1/1/1 up {0=lettuce=up:active}, 2 up:standby
2011-10-12 14:49:39.086622 osd e5: 3 osds: 3 up, 3 in
2011-10-12 14:49:39.086908 log 2011-10-12 14:38:50.263058 osd1 192.168.1.1:6801/8637 197 : [INF] 2.1p1 scrub ok
2011-10-12 14:49:39.086977 mon e1: 3 mons at {0=192.168.1.1:6789/0,1=192.168.1.2:6789/0,2=192.168.1.3:6789/0}
The cluster configuration I chose has authentication enabled, so to actually mount the ceph filesystem requires a secret key. This key is stored in the /etc/ceph/keyring.admin file that was created earlier. To view the keyring contents, the cauthtool program must be used
# cauthtool -l /etc/ceph/keyring.admin
[client.admin]
key = AQDLk5VOeHkHLxAAfGjcaUsOXOhJr7hZCNjXSQ==
auid = 18446744073709551615
The base64 key there will be passed to the mount command, repeating on every host needing a filesystem present:
# mount -t ceph 192.168.1.1:6789:/ /mnt/ -o name=admin,secret=AQDLk5VOeHkHLxAAfGjcaUsOXOhJr7hZCNjXSQ==
error adding secret to kernel, key name client.admin: No such device
For some reason, that error message is always printed on my Fedora hosts, and despite that, the mount has actually succeeded
# grep /mnt /proc/mounts
192.168.1.1:6789:/ /mnt ceph rw,relatime,name=admin,secret= 0 0
Congratulations, /mnt is now a distributed filesystem. If you create a file on one host, it should appear on the other hosts & vica-verca.
RBD Volume setup
A shared filesystem is very nice, and can be used to hold regular virtual disk images in a variety of formats (raw, qcow2, etc). What I really wanted to try was the RBD virtual block device functionality in QEMU. Ceph includes a tool called rbd for manipulating those. The syntax of this tool is pretty self-explanatory
# rbd create --size 100 demo
# rbd ls
demo
# rbd info demo
rbd image 'demo':
size 102400 KB in 25 objects
order 22 (4096 KB objects)
block_name_prefix: rb.0.0
parent: (pool -1)
Alternatively RBD volume creation can be done using qemu-img …. at least once the Fedora QEMU package is fixed to enable RBD support.
# qemu-img create -f rbd rbd:rbd/demo 100M
Formatting 'rbd:rbd/foo', fmt=rbd size=104857600 cluster_size=0
# qemu-img info rbd:rbd/demo
image: rbd:rbd/foo
file format: raw
virtual size: 100M (104857600 bytes)
disk size: unavailable
KVM guest setup
The syntax for configuring a RBD block device in libvirt, is very similar to that used for Sheepdog. In Sheepdog, every single virtualization node is also a storage node, so there is no hostname required. Not so for RBD. Here it is necessary to specify one or more host names, for the RBD servers.
<disk type='network' device='disk'>
<driver name='qemu' type='raw'/>
<source protocol='rbd' name='demo/wibble'>
<host name='lettuce.example.org' port='6798'/>
<host name='mustard.example.org' port='6798'/>
<host name='avocado.example.org' port='6798'/>
</source>
<target dev='vdb' bus='virtio'/>
</disk>
More observant people might be wondering how QEMU gets permission to connect to the RBD server, given that the configuration earlier enabled authentication. This is thanks to the magic of the /etc/ceph/keyring.admin file which must exist on any virtualization server. Patches are currently being discussed which will allow authentication credentials to be set via libvirt, avoiding the need to store the credentials on the virtualization hosts permanently.