Posts Tagged ‘Linux’

ZFS with Redhat Cluster Suite

Friday, July 25th, 2014

This is a very nice project I have been working on. The hardware at hand - two servers, with a shared SAS bus containing several SAS disks. Since it's a shared bus, no RAID solution would cut it, and as I don't want to waste disks with ASM ("normal" redundancy meaning half the size...), I went to ZFS storage.

ZFS is a wonderful technology, with many advantages, but with some dangerous pitfalls. As I prefer Linux, I did not bother with any Sloaris solutions, and went directly to Centos 6. I will describe my cluster setup below.

I will disclose the entire setup, including hardware layout, Linux platform, ZFS module parameters, the Redhat Cluster Suite ZFS agent I wrote and the cluster.conf configuration file. I will also share my considerations regarding some of the choices I made. In addition, this system was designed to act as NFS storage for Citrix XenServer pool, so I will have to describe the changed I had to perform on the XenServer itself (which might make it unsupported, but I will have to live with it), to allow it to handle the timeouts resulting by server failover.

So first - the servers - each having a single CPU (quad core), 24GB RAM, and dual 1Gb/s NICs. Also - a tiny internal SATA disk is used for the OS. The shared disks - at the moment, 10 SAS disks, dual port (notice - older HP disks might mark in a very small letters that they are only a single-port SAS disks...), 72GB, 10K RPM. Zpool called 'share' with two 5 disks RaidZ1 vdevs. As I mentioned before - ZFS seemed like the best possible option allowing me to achieve my goals at minimal cost.

When I came to this project, I wanted to be able to use a native ZFS cluster agent, and not a 'script' agent, which takes a very long time to respond (30 seconds). Also - I wanted to be able to handle multiple storage pools concurrently - each floating on its own. While I have only one at the moment, I wanted the ability to have a fine-grained control over multiple pools. In addition - I am unable (or unwilling?) to handle the multiple filesystems introduced with each pool. I wanted to be able to import or export the pool silently, and with a clear head, thus I had to verify that the multiple filesystems are not in use as part of the export process.

As an agent, I wanted to comply with Redhat Cluster Suite (RHCS from now on) OCF syntax. I used the supplied script as an inspiration for my agent script, so some of it might look familiar. All credit goes to the original authors, of course.

The operating system I selected was Centos 6. Centos is based on Redhat Linux, and I find it mature and stable, which is exactly what I want when I plan a production-ready, enterprise-class storage solution. The version had to be x86_64, due to ZFS requirements, and due to the amount of RAM in the server.

To handle ZFS options, I added a file called /etc/modprobe.d/zfs.conf, with the following content

install zfs /bin/rm -f /etc/zfs/zpool.cache && /sbin/modprobe --ignore-install zfs
options zfs zfs_arc_max=12593790976
options zfs zfs_arc_min=12593790975

I had to verify there is no zpool.cache file. Since my pool was rather small (planned for 24 disks max), I was not concerned by the longer import process caused by not having the zpool.cache file. I was more concerned with automatic import process which might happen, and had to prevent it at almost any cost. In addition, I learned from other systems that the arc memory should never exceed half the RAM, and it should be given just a little under that.

Of course, when changing such module settings, you need to recreate initrd (dracut -f) to be on the safe side later on.

The agent script was placed in /usr/share/cluster directory. You must have rgmanager installed for this directory to exist, and anyhow, without rgmanager, you will have no cluster whatsoever.

This is the contents of the file. Notice that it is not compatible with Luci, so if you're using it - them kids won't play well together.


# Private return codes

. $(dirname $0)/ocf-shellfuncs

    cat < EOT


	This script will import and export ZFS storage pools
	It will make sure to mount and umount all child filesystems

        This is a ZFS pool

                Symbolic name for this zfs pool

                File System Name

		ZFS Pool name or ID

                ZFS pool name

		ZFS Pool alternate mount

                ZFS pool alternate mount

                If set, the cluster will kill all processes using 
                this file system when the resource group is 
                stopped.  Otherwise, the unmount will fail, and
                the resource group will be restarted.

                Force Unmount

                If set and unmounting the file system fails, the node will
                immediately reboot.  Generally, this is used in conjunction
                with force-unmount support, but it is not required.

                Seppuku Unmount




        echo $*

verify_driver() {
	ocf_log info "Verifying ZFS driver"
	lsmod | grep -w zfs > /dev/null >&1 && return 0
	ocf_log err "ZFS driver is not loaded"
	return $OCF_ERR_ARGS

verify_poolname() {
	ocf_log info "Verify pool name "
	if [ -z "$OCF_RESKEY_pool" ]
		ocf_log err "Missing pool name"
		return $OCF_ERR_ARGS
	zpool import | grep pool: | grep -w $OCF_RESKEY_pool > /dev/null 2>&1 && return 0
	ocf_log err "Cannot identify pool name"
	return $OCF_ERR_ARGS

verify_mounted_poolname() {
	ocf_log info "Verify pool name "
	if [ -z "$OCF_RESKEY_pool" ]
		ocf_log err "Missing pool name"
		return $OCF_ERR_ARGS
	zpool list $OCF_RESKEY_pool > /dev/null >&1 && return 0
	ocf_log err "Cannot identify pool name"
	return $OCF_ERR_ARGS

verify_mountpath() {
	ocf_log info "Verifying alternate root mount path"
	[ -z "$OCF_RESKEY_mount" ] && return 0
	declare mp="${OCF_RESKEY_mount}"
	case "$mp" in
		/*)    	# found it
        	*)      # invalid format
			ocf_log err 
"verify_mountpath: Invalid mount point format (must begin with a '/'): '$mp'"
                return $OCF_ERR_ARGS

pool_import() {
	ocf_log info "Importing pool"
	[ -n "$OCF_RESKEY_mount" ] && OPTS="-R $OCF_RESKEY_mount"
	zpool import $OCF_RESKEY_pool $OPTS
	if [ "$RET" -ne "0" ]
		ocf_log info "Cannot import without applying force"
		zpool import -f $OCF_RESKEY_pool $OPTS
	if [ "$RET" -ne "0" ]
		ocf_log err "Pool import failed for $OCF_RESKEY_pool. error=$RET"
		return 1
	ocf_log info "Imported ZFS pool"
	return $RET

check_and_release_fs() {
	ocf_log info "Checking and releasing FS"
	case ${OCF_RESKEY_force_unmount} in
        $YES_STR|on|true|1)	force_umount=$YES ;;
        *)		        force_umount="" ;;

	for i in `zfs list -t filesystem | grep ^${OCF_RESKEY_pool} | awk '{print $NF}'`
		# To be on the safe side. Why not?
		sleep 1
		# Is it mounted?
		if ! df -l | grep -w "$i" > /dev/null 2>&1
			ocf_log info "Filesystem $i is not mounted"
		if [ `lsof $i | wc -l` -gt "0" ]
			ocf_log info "Filesystem $i is in use"
			if [ "$force_umount" ]
				ocf_log info "Attempting to kill processes on $i filesystem"
				fuser -k $i
				sleep 2
				if [ `lsof $i | wc -l` -gt "0" ]
					ocf_log err "Cannot umount filesystem $i - filesystem in use"
					return 1
				ocf_log err "Cannot umount filesystem $i
 - filesystem in use"
                                return 1
	return $RET	

self_fence() {
	ocf_log info "Should we validate and call self-fence?"
	case ${OCF_RESKEY_self_fence} in
		$YES_STR|on|true|1)       self_fence=$YES ;;
       		*)              self_fence="" ;;

	if [ "$self_fence" ]; then
		ocf_log alert "umount failed - REBOOTING"
                reboot -fn

pool_export() {
	ocf_log info "Exporting zfs pool"
	check_and_release_fs || self_fence
	zpool export $OCF_RESKEY_pool
	if [ "$RET" -ne "0" ]
		ocf_log err "Pool export failed for $OCF_RESKEY_pool. error=$RET"
		return 1
	return $RET

start() {
	ocf_log info "Starting ZFS"
	verify_driver || return $OCF_ERR_ARGS 
	verify_poolname || return $OCF_ERR_ARGS
	verify_mountpath || return $OCF_ERR_ARGS
	# Handle filesystem?

stop() {
	ocf_log info "Starting ZFS"
	verify_driver || return $OCF_ERR_ARGS 
	verify_mounted_poolname || return $OCF_ERR_ARGS
	verify_mountpath || return $OCF_ERR_ARGS
	# Handle filesystem?

is_imported() {
	ocf_log debug "Checking if $OCF_RESKEY_pool is imported"
	zpool list ${OCF_RESKEY_pool} > /dev/null >&1
	return $?

is_alive() {
	ocf_log debug "Checking ZFS pool read/write"
	declare file=".writable_test.$(hostname)"
	declare TIMEOUT="10s"
	[ -z "$OCF_CHECK_LEVEL" ] && export OCF_CHECK_LEVEL=0
	mount_point=`zfs list ${OCF_RESKEY_pool} | grep ${OCF_RESKEY_pool} | awk '{print $NF}'`
	test -d "$mount_point"
        if [ $? -ne 0 ]; then
                ocf_log err "${OCF_RESOURCE_INSTANCE}: is_alive: $mount_point is not a directory"
                return $FAIL
	[ $OCF_CHECK_LEVEL -lt 10 ] && return $YES

        # depth 10 test (read test)
        timeout -s 9 $TIMEOUT ls "$mount_point" > /dev/null 2> /dev/null
        if [ $errcode -ne 0 ]; then
                ocf_log err "${OCF_RESOURCE_INSTANCE}: is_alive: failed read test on [$mount_point]. Return code: $errcode"
                return $NO

	[ $OCF_CHECK_LEVEL -lt 20 ] && return $YES

        # depth 20 check (write test)
        for o in `echo $OCF_RESKEY_options | sed -e s/,/ /g`; do
                if [ "$o" = "ro" ]; then
	if [ $rw -eq $YES ]; then
                while true; do
                        if [ -e "$file" ]; then
                timeout -s 9 $TIMEOUT touch $file > /dev/null 2> /dev/null
                if [ $errcode -ne 0 ]; then
                        ocf_log err "${OCF_RESOURCE_INSTANCE}: is_alive: failed write test on [$mount_point]. Return code: $errcode"
                        return $NO
                rm -f $file > /dev/null 2> /dev/null

	return $YES

monitor() {
	ocf_log debug "Checking ZFS pool $OCF_RESKEY_pool, Level $OCF_CHECK_LEVEL"
	verify_driver || return $OCF_ERR_ARGS 
	if [ "$RET" -ne $YES ]; then
                ocf_log err "${OCF_RESOURCE_INSTANCE}: ${OCF_RESKEY_device} is not mounted on ${OCF_RESKEY_mountpoint}"
                return $OCF_NOT_RUNNING
	return $RET

if [ -z "$OCF_CHECK_LEVEL" ]; then

case $1 in
	ocf_log info "zfs start $OCF_RESKEY_pooln"
	[ "$?" -ne "0" ] && start || ocf_log info "$OCF_RESKEY_pool is already mounted"
	exit $?
	ocf_log info "zfs stop $OCF_RESKEY_pooln"
	[ "$?" -eq "0" ] && stop || ocf_log info "$OCF_RESKEY_pool is not mounted"
	exit $?
	ocf_log debug "ZFS monitor $OCF_RESKEY_pool"
	exit $?
	echo -e "zfs metadat $OCF_RESKEY_addressn" >>/tmp/out
	exit 0
	exit 0
	echo "usage: $0 {start|stop|status|monitor|restart|meta-data|validate-all}"

All I had to do now was to build the cluster.conf file.

The reason I placed the IP address as the last to start and the first to stop was that the other way around, the NFS client would receive an ordered disconnection command, and would not bother to establish a connection with the remaining server. Abruptly taking away the clustered IP address causes the NFS clients to initiate a reconnection process, of which the systems are supposed to recover

I have left this article incomplete for a while now. It has some stuff I do like to share, so I am sharing it as-is. I will (some day) complete it.

Extracting/Recreating RHEL/Centos6 initrd.img and install.img

Tuesday, October 1st, 2013

A quick note about extracting and recreating RHEL6 or Centos6 (and their derivations) installation media components:



mv initrd.img /tmp/initrd.img.xz
cd /tmp
xz –format=lzma initrd.img.xz –decompress
mkdir initrd
cd initrd
cpio -ivdum < ../initrd.img

Archive (after you applied your changes):

cd /tmp/initrd
find . | cpio -o -H newc | xz -9 –format=lzma > ../new-initrd.img



mount -o loop install.img /mnt
mkdir /tmp/install.img.dir
cd /mnt ; tar cf – –one-file-system . | ( cd /tmp/install.img.dir ; tar xf – )
umount /mnt

Archive (after you applied your changes):

cd /tmp
mksquashfs install.img.dir/ install-new.img

Additional note for Anaconda installation parameters:

I did not test it, however there is a boot flag called stage2= which should lead to a new install.img file, other than the hardcoded one. I don’t if it will accept /images/install-new.img as its flag, but it can be a good start there.

One more thing:

Make sure that the vmlinuz and initrd used for any custom properties, in $CDROOT/isolinux do not exceed 8.3 format. Longer names didn’t work for me. I assume (without any further checks) that this is isolinux limitation.

XenServer – increase LVM over iSCSI LUN size – online

Wednesday, September 4th, 2013

The following procedure was tested by me, and was found to be working. The version of the XenServer I am using in this particular case is 6.1, however, I belive that this method is generic enough so that it could work for every version of XS, assuming you're using iSCSI and LVM (aka - not NetApp, CSLG, NFS and the likes). It might act as a general guideline for fiber channel communication, but this was not tested by me, and thus - I have no idea how it will work. It should work with some modifications when using Multipath, however, regarding multipath, you can find in this particular blog some notes on increasing multipath disks. Check the comments too - they might offer some better and simplified way of doing it.

So - let's begin.

First - increase the size of the LUN through the storage. For NetApp, it involves something like:

lun resize /vol/XenServer/luns/SR1.lun +1t

You should always make sure your storage volume, aggregate, raid group, pool or whatever is capable of holding the data, or - if using thin provisioning - that a well tested monitoring system is available to alert you when running low on storage disk space.

Now, we should identify the LUN. From now on - every action should be performed on all XS pool nodes, one after the other.

cat /proc/partitions

We should keep the output of this command somewhere. We will use it later on to identify the expanded LUN.

Now - let's scan for storage changes:

iscsiadm -m node -R

Now, running the previous command again will have a slightly different output. We can not identify the modified LUN

cat /proc/partitions

We should increase it in size. XenServer uses LVM, so we should harness it to our needs. Let's assume that the modified disk is /dev/sdd.

pvresize /dev/sdd

After completing this task on all pool hosts, we should run sr-scan command. Either by CLI, or through the GUI. When the scan operation completes, the new size would show.

Hope it helps!

Juniper NetworkConnect (NC) and 64bit Linux

Tuesday, June 25th, 2013

Due to a major disk crash, I have had to use my ‘other’ computer for VPN connections. It meant that I have had to prepare it for the operation. I attempted to login to aJuniper-based SSL-VPN connection, however, I did get a message saying that my 64bit Java was inadequate. I had a link, as part of the error message to Juniper KB, to which I must link (remembering how I have had to search for possible solutions in the past).

The nice thing about this solution is that it does not replace your default Java version on the system, which was always a problem, as I was using Java for various purposes, but it recognizes that it’s part of the (update-)alternatives list, and makes use of the correct Java version.

Juniper did it right this time!

Oh – and the link to their KB

And to Oracle Java versions, to make life slightly easier for you. You will need Oracle login, however (you can register for free).

Target-based persistent device naming

Saturday, June 22nd, 2013

When Connecting Linux to a large array of SAS disks (JBOD), udev creates default persistent names in /dev/disk/by-* . These names are based on LUN ID (all disks take lun0 by default), and by path, which includes, for a pure SAS bus – the PWWN of the disks. It means that an example to such naming would be like this (slightly trimmed for ease of view):

scsi-35000c50055924207 -> ../../sde
scsi-35000c50055c5138b -> ../../sdd
scsi-35000c50055c562eb -> ../../sda
scsi-35000c500562ffd73 -> ../../sdc
scsi-35001173100134654 -> ../../sdn
scsi-3500117310013465c -> ../../sdk
scsi-35001173100134688 -> ../../sdj
scsi-35001173100134718 -> ../../sdo
scsi-3500117310013490c -> ../../sdg
scsi-35001173100134914 -> ../../sdh
scsi-35001173100134a58 -> ../../sdp
scsi-3500117310013671c -> ../../sdm
scsi-35001173100136740 -> ../../sdl
scsi-350011731001367ac -> ../../sdi
scsi-350011731001cdd58 -> ../../sdf
wwn-0x5000c50055924207 -> ../../sde
wwn-0x5000c50055c5138b -> ../../sdd
wwn-0x5000c50055c562eb -> ../../sda
wwn-0x5000c500562ffd73 -> ../../sdc
wwn-0x5001173100134654 -> ../../sdn
wwn-0x500117310013465c -> ../../sdk
wwn-0x5001173100134688 -> ../../sdj
wwn-0x5001173100134718 -> ../../sdo
wwn-0x500117310013490c -> ../../sdg
wwn-0x5001173100134914 -> ../../sdh
wwn-0x5001173100134a58 -> ../../sdp
wwn-0x500117310013671c -> ../../sdm
wwn-0x5001173100136740 -> ../../sdl
wwn-0x50011731001367ac -> ../../sdi
wwn-0x50011731001cdd58 -> ../../sdf

pci-0000:03:00.0-sas-0x5000c50055924206-lun-0 -> ../../sde
pci-0000:03:00.0-sas-0x5000c50055c5138a-lun-0 -> ../../sdd
pci-0000:03:00.0-sas-0x5000c50055c562ea-lun-0 -> ../../sda
pci-0000:03:00.0-sas-0x5000c500562ffd72-lun-0 -> ../../sdc
pci-0000:03:00.0-sas-0x5001173100134656-lun-0 -> ../../sdn
pci-0000:03:00.0-sas-0x500117310013465e-lun-0 -> ../../sdk
pci-0000:03:00.0-sas-0x500117310013468a-lun-0 -> ../../sdj
pci-0000:03:00.0-sas-0x500117310013471a-lun-0 -> ../../sdo
pci-0000:03:00.0-sas-0x500117310013490e-lun-0 -> ../../sdg
pci-0000:03:00.0-sas-0x5001173100134916-lun-0 -> ../../sdh
pci-0000:03:00.0-sas-0x5001173100134a5a-lun-0 -> ../../sdp
pci-0000:03:00.0-sas-0x500117310013671e-lun-0 -> ../../sdm
pci-0000:03:00.0-sas-0x5001173100136742-lun-0 -> ../../sdl
pci-0000:03:00.0-sas-0x50011731001367ae-lun-0 -> ../../sdi
pci-0000:03:00.0-sas-0x50011731001cdd5a-lun-0 -> ../../sdf

Real port (connection) persistence is not possible in that manner. A map of PWWN-to-Slot is required, and handling the system in case of a disk failure by non-expert is nearly impossible. A solution for that is to create matching udev rules which will allow handling disks per-port.

While there are (absolutely) better ways of doing it, time constrains require that I get it to work quick&dirty. The solution is based on lsscsi command, as the backend engine of the system, so make sure it exists on the system. I tend to believe that the system will not be able to scale out to hundreds of disks in its current design, but for my 16 disks (and probably for several tenths as well) – it works fine.

Add 60-persistent-disk-ports.rules to /etc/udev/rules.d/ (and omit the .txt suffix)


# By Ez-Aton, based partially on the built-in udev block device rule
# forward scsi device event to corresponding block device
ACTION=="change", SUBSYSTEM=="scsi", ENV{DEVTYPE}=="scsi_device", TEST=="block", ATTR{block/*/uevent}="change"

ACTION!="add|change", GOTO="persistent_storage_end"
SUBSYSTEM!="block", GOTO="persistent_storage_end"

# skip rules for inappropriate block devices
KERNEL=="fd*|mtd*|nbd*|gnbd*|btibm*|dm-*|md*", GOTO="persistent_storage_end"

# never access non-cdrom removable ide devices, the drivers are causing event loops on open()
KERNEL=="hd*[!0-9]", ATTR{removable}=="1", SUBSYSTEMS=="ide", ATTRS{media}=="disk|floppy", GOTO="persistent_storage_end"
KERNEL=="hd*[0-9]", ATTRS{removable}=="1", GOTO="persistent_storage_end"

# ignore partitions that span the entire disk
TEST=="whole_disk", GOTO="persistent_storage_end"

# for partitions import parent information
ENV{DEVTYPE}=="partition", IMPORT{parent}="ID_*"

# Deal only with SAS disks
KERNEL=="sd*[!0-9]|sr*", ENV{ID_SERIAL}!="?*", IMPORT{program}="/usr/local/sbin/ $tempnode", ENV{ID_BUS}="scsi"
KERNEL=="sd*|sr*|cciss*", ENV{DEVTYPE}=="disk", ENV{TGT_PATH}=="?*", SYMLINK+="disk/by-target/disk-$env{TGT_PATH}"
#KERNEL=="sd*|cciss*", ENV{DEVTYPE}=="partition", ENV{ID_SERIAL}!="?*", IMPORT{program}="/usr/local/sbin/ $tempnode"
KERNEL=="sd*|cciss*", ENV{DEVTYPE}=="partition", ENV{ID_SERIAL}=="?*", IMPORT{program}="/usr/local/sbin/ $tempnode", SYMLINK+="disk/by-target/disk-$env{TGT_PATH}p%n"

ENV{DEVTYPE}=="disk", KERNEL!="xvd*|sd*|sr*", ATTR{removable}=="1", GOTO="persistent_storage_end"

You will need to add (and make executable) the script in /usr/local/sbin. Again – remove the .txt suffix

# Written by Ez-Aton to assist with disk-to-port mapping
# $1 - disk device name
# Full disk
TGT_PATH=`/usr/bin/lsscsi | grep -w /dev/$name | awk '{print $1}' | tr -d ] | tr -d [`
if [ -z "$TGT_PATH" ]
	# This is a partition, so our grep fails
	name=`echo $name | tr -d [0-9]`
	TGT_PATH=`/usr/bin/lsscsi | grep -w /dev/$name | awk '{print $1}' | tr -d ] | tr -d [`

The result of this addition to udev would be a directory called /dev/disk/by-target containing links as follow:

disk-0:0:0:0 -> ../../sda
disk-0:0:1:0 -> ../../sdb
disk-0:0:10:0 -> ../../sdk
disk-0:0:11:0 -> ../../sdl
disk-0:0:12:0 -> ../../sdm
disk-0:0:13:0 -> ../../sdn
disk-0:0:14:0 -> ../../sdo
disk-0:0:15:0 -> ../../sdp
disk-0:0:2:0 -> ../../sdc
disk-0:0:3:0 -> ../../sdd
disk-0:0:4:0 -> ../../sde
disk-0:0:5:0 -> ../../sdf
disk-0:0:6:0 -> ../../sdg
disk-0:0:7:0 -> ../../sdh
disk-0:0:8:0 -> ../../sdi
disk-0:0:9:0 -> ../../sdj

The result is a persistent naming, based on real device ports.
I hope it helps. If you get to read it and have some suggestions (or a better use of udev, which I know is far from perfect in this case), I would love to hear about it.