Manpages

NAME

sbd - STONITH Block Device daemon

SYNOPSIS

sbd <-d /dev/...> [options] "command"

SUMMARY

SBD provides a node fencing mechanism (Shoot the other node in the head, STONITH ) for Pacemaker-based clusters through the exchange of messages via shared block storage such as for example a SAN, iSCSI, FCoE. This isolates the fencing mechanism from changes in firmware version or dependencies on specific firmware controllers, and it can be used as a STONITH mechanism in all configurations that have reliable shared storage.

SBD can also be used without any shared storage. In this mode, the watchdog device will be used to reset the node if it loses quorum, if any monitored daemon is lost and not recovered or if Pacemaker decides that the node requires fencing.

The sbd binary implements both the daemon that watches the message slots as well as the management tool for interacting with the block storage device(s). This mode of operation is specified via the "command" parameter; some of these modes take additional parameters.

To use SBD with shared storage, you must first "create" the messaging layout on one to three block devices. Second, configure /etc/default/sbd to list those devices (and possibly adjust other options), and restart the cluster stack on each node to ensure that "sbd" is started. Third, configure the "external/sbd" fencing resource in the Pacemaker CIB.

Each of these steps is documented in more detail below the description of the command options.

"sbd" can only be used as root.

GENERAL OPTIONS
-d /dev/...

Specify the block device(s) to be used. If you have more than one, specify this option up to three times. This parameter is mandatory for all modes, since SBD always needs a block device to interact with.

This man page uses /dev/sda1, /dev/sdb1, and /dev/sdc1 as example device names for brevity. However, in your production environment, you should instead always refer to them by using the long, stable device name (e.g., /dev/disk/by-id/dm-uuid-part1-mpath-3600508b400105b5a0001500000250000).

-v|-vv|-vvv

Enable verbose|debug|debug-library logging (optional)

-n node

Set local node name; defaults to "uname -n". This should not need to be set.

-I N

Async IO timeout (defaults to 3 seconds, optional). You should not need to adjust this unless your IO setup is really very slow.

(In daemon mode, the watchdog is refreshed when the majority of devices could be read within this time.)

create
Example usage:

sbd -d /dev/sdc2 -d /dev/sdd3 create

If you specify the create command, sbd will write a metadata header to the device(s) specified and also initialize the messaging slots for up to 255 nodes.

Warning: This command will not prompt for confirmation. Roughly the first megabyte of the specified block device(s) will be overwritten immediately and without backup.

This command accepts a few options to adjust the default timings that are written to the metadata (to ensure they are identical across all nodes accessing the device).
-1 N

Set watchdog timeout to N seconds. This depends mostly on your storage latency; the majority of devices must be successfully read within this time, or else the node will self-fence.

If your sbd device(s) reside on a multipath setup or iSCSI, this should be the time required to detect a path failure. You may be able to reduce this if your device outages are independent, or if you are using the Pacemaker integration.

-2 N

Set slot allocation timeout to N seconds. You should not need to tune this.

-3 N

Set daemon loop timeout to N seconds. You should not need to tune this.

-4 N

Set msgwait timeout to N seconds. This should be twice the watchdog timeout. This is the time after which a message written to a node’s slot will be considered delivered. (Or long enough for the node to detect that it needed to self-fence.)

This also affects the stonith-timeout in Pacemaker’s CIB ; see below.

list
Example usage:

# sbd -d /dev/sda1 list
0 hex-0 clear
1 hex-7 clear
2 hex-9 clear

List all allocated slots on device, and messages. You should see all cluster nodes that have ever been started against this device. Nodes that are currently running should have a clear state; nodes that have been fenced, but not yet restarted, will show the appropriate fencing message.

dump
Example usage:

# sbd -d /dev/sda1 dump
==Dumping header on disk /dev/sda1
Header version : 2
Number of slots : 255
Sector size : 512
Timeout (watchdog) : 15
Timeout (allocate) : 2
Timeout (loop) : 1
Timeout (msgwait) : 30
==Header on disk /dev/sda1 is dumped

Dump meta-data header from device.

watch
Example usage:

sbd -d /dev/sdc2 -d /dev/sdd3 -P watch

This command will make "sbd" start in daemon mode. It will constantly monitor the message slot of the local node for incoming messages, reachability, and optionally take Pacemaker’s state into account.

"sbd" must be started on boot before the cluster stack! See below for enabling this according to your boot environment.

The options for this mode are rarely specified directly on the commandline directly, but most frequently set via /etc/default/sbd.

It also constantly monitors connectivity to the storage device, and self-fences in case the partition becomes unreachable, guaranteeing that it does not disconnect from fencing messages.

A node slot is automatically allocated on the device(s) the first time the daemon starts watching the device; hence, manual allocation is not usually required.

If a watchdog is used together with the "sbd" as is strongly recommended, the watchdog is activated at initial start of the sbd daemon. The watchdog is refreshed every time the majority of SBD devices has been successfully read. Using a watchdog provides additional protection against "sbd" crashing.

If the Pacemaker integration is activated, "sbd" will not self-fence if device majority is lost, if:

This allows "sbd" to survive temporary outages of the majority of devices. However, while the cluster is in such a degraded state, it can neither successfully fence nor be shutdown cleanly (as taking the cluster below the quorum threshold will immediately cause all remaining nodes to self-fence). In short, it will not tolerate any further faults. Please repair the system before continuing.

There is one "sbd" process that acts as a master to which all watchers report; one per device to monitor the node’s slot; and, optionally, one that handles the Pacemaker integration.

Defaults to enabled.

-w /dev/watchdog

This can be used to override the default watchdog device used and should not usually be necessary.

-p /run/sbd.pid

This option can be used to specify a pidfile for the main sbd process.

-F N

Number of failures before a failing servant process will not be restarted immediately until the dampening delay has expired. If set to zero, servants will be restarted immediately and indefinitely. If set to one, a failed servant will be restarted once every -t seconds. If set to a different value, the servant will be restarted that many times within the dampening period and then delay.

Defaults to 1.

-t N

Dampening delay before faulty servants are restarted. Combined with "-F 1", the most logical way to tune the restart frequency of servant processes. Default is 5 seconds.

If set to zero, processes will be restarted indefinitely and immediately.

Defaults to enabled.

-S N

Set the start mode. (Defaults to 0.)

If this is set to zero, sbd will always start up unconditionally, regardless of whether the node was previously fenced or not.

If set to one, sbd will only start if the node was previously shutdown cleanly (as indicated by an exit request message in the slot), or if the slot is empty. A reset, crashdump, or power-off request in any slot will halt the start up.

This is useful to prevent nodes from rejoining if they were faulty. The node must be manually "unfenced" by sending an empty message to it:

sbd -d /dev/sda1 message node1 clear

-s N

Set the start-up wait time for devices. (Defaults to 120.)

Dynamic block devices such as iSCSI might not be fully initialized and present yet. This allows one to set a timeout for waiting for devices to appear on start-up. If set to 0, start-up will be aborted immediately if no devices are available.

-5 N

Warn if the time interval for tickling the watchdog exceeds this many seconds. Since the node is unable to log the watchdog expiry (it reboots immediately without a chance to write its logs to disk), this is very useful for getting an indication that the watchdog timeout is too short for the IO load of the system.

Default is about 3/5 of watchdog timeout, set to zero to disable.

-C N

Watchdog timeout to set before crashdumping. If SBD is set to crashdump instead of reboot - either via the trace mode settings or the external/sbd fencing agent’s parameter -, SBD will adjust the watchdog timeout to this setting before triggering the dump. Otherwise, the watchdog might trigger and prevent a successful crashdump from ever being written.

Set to zero (= default) to disable.

-r N

Actions to be executed when the watchers don’t timely report to the sbd master process or one of the watchers detects that the master process has died.

Set timeout-action to comma-separated combination of noflush|flush plus reboot|crashdump|off. If just one of both is given the other stays at the default.

This doesn’t affect actions like off, crashdump, reboot explicitly triggered via message slots. And it does as well not configure the action a watchdog would trigger should it run off (there is no generic interface).

Defaults to flush,reboot.

allocate
Example usage:

sbd -d /dev/sda1 allocate node1

Explicitly allocates a slot for the specified node name. This should rarely be necessary, as every node will automatically allocate itself a slot the first time it starts up on watch mode.

message
Example usage:

sbd -d /dev/sda1 message node1 test

Writes the specified message to node’s slot. This is rarely done directly, but rather abstracted via the "external/sbd" fencing agent configured as a cluster resource.

Supported message types are:
test

This only generates a log message on the receiving node and can be used to check if SBD is seeing the device. Note that this could overwrite a fencing request send by the cluster, so should not be used during production.

reset

Reset the target upon receipt of this message.

crashdump

Cause the target node to crashdump.

exit

This will make the "sbd" daemon exit cleanly on the target. You should not send this message manually; this is handled properly during shutdown of the cluster stack. Manually stopping the daemon means the node is unprotected!

clear

This message indicates that no real message has been sent to the node. You should not set this manually; "sbd" will clear the message slot automatically during start-up, and setting this manually could overwrite a fencing message by the cluster.

query-watchdog
Example usage:

sbd query-watchdog

Check for available watchdog devices and print some info.

Warning: This command will arm the watchdog during query, and if your watchdog refuses disarming (for example, if its kernel module has the ’nowayout’ parameter set) this will reset your system.

test-watchdog
Example usage:

sbd test-watchdog [-w /dev/watchdog3]

Test specified watchdog device (/dev/watchdog by default).

Warning: This command will arm the watchdog and have your system reset in case your watchdog is working properly! If issued from an interactive session, it will prompt for confirmation.

Base system configuration

Configure a watchdog
It is highly recommended that you configure your Linux system to load a watchdog driver with hardware assistance (as is available on most modern systems), such as hpwdt, iTCO_wdt, or others. As a fall-back, you can use the softdog module.

No other software must access the watchdog timer; it can only be accessed by one process at any given time. Some hardware vendors ship systems management software that use the watchdog for system resets (f.e. HP ASR daemon). Such software has to be disabled if the watchdog is to be used by SBD.

Choosing and initializing the block device(s)
First, you have to decide if you want to use one, two, or three devices.

If you are using multiple ones, they should reside on independent storage setups. Putting all three of them on the same logical unit for example would not provide any additional redundancy.

The SBD device can be connected via Fibre Channel, Fibre Channel over Ethernet, or even iSCSI. Thus, an iSCSI target can become a sort-of network-based quorum server; the advantage is that it does not require a smart host at your third location, just block storage.

The SBD partitions themselves must not be mirrored (via MD, DRBD, or the storage layer itself), since this could result in a split-mirror scenario. Nor can they reside on cLVM2 volume groups, since they must be accessed by the cluster stack before it has started the cLVM2 daemons; hence, these should be either raw partitions or logical units on (multipath) storage.

The block device(s) must be accessible from all nodes. (While it is not necessary that they share the same path name on all nodes, this is considered a very good idea.)

SBD will only use about one megabyte per device, so you can easily create a small partition, or very small logical units. (The size of the SBD device depends on the block size of the underlying device. Thus, 1MB is fine on plain SCSI devices and SAN storage with 512 byte blocks. On the IBM s390x architecture in particular, disks default to 4k blocks, and thus require roughly 4MB.)

The number of devices will affect the operation of SBD as follows:
One device

In its most simple implementation, you use one device only. This is appropriate for clusters where all your data is on the same shared storage (with internal redundancy) anyway; the SBD device does not introduce an additional single point of failure then.

If the SBD device is not accessible, the daemon will fail to start and inhibit startup of cluster services.

Two devices

This configuration is a trade-off, primarily aimed at environments where host-based mirroring is used, but no third storage device is available.

SBD will not commit suicide if it loses access to one mirror leg; this allows the cluster to continue to function even in the face of one outage.

However, SBD will not fence the other side while only one mirror leg is available, since it does not have enough knowledge to detect an asymmetric split of the storage. So it will not be able to automatically tolerate a second failure while one of the storage arrays is down. (Though you can use the appropriate crm command to acknowledge the fence manually.)

It will not start unless both devices are accessible on boot.

Three devices

In this most reliable and recommended configuration, SBD will only self-fence if more than one device is lost; hence, this configuration is resilient against temporary single device outages (be it due to failures or maintenance). Fencing messages can still be successfully relayed if at least two devices remain accessible.

This configuration is appropriate for more complex scenarios where storage is not confined to a single array. For example, host-based mirroring solutions could have one SBD per mirror leg (not mirrored itself), and an additional tie-breaker on iSCSI.

It will only start if at least two devices are accessible on boot.

After you have chosen the devices and created the appropriate partitions and perhaps multipath alias names to ease management, use the "sbd create" command described above to initialize the SBD metadata on them.

Sharing the block device(s) between multiple clusters

It is possible to share the block devices between multiple clusters, provided the total number of nodes accessing them does not exceed 255 nodes, and they all must share the same SBD timeouts (since these are part of the metadata).

If you are using multiple devices this can reduce the setup overhead required. However, you should not share devices between clusters in different security domains.

Configure SBD to start on boot
On systems using "sysvinit", the "openais" or "corosync" system start-up scripts must handle starting or stopping "sbd" as required before starting the rest of the cluster stack.

For "systemd", sbd simply has to be enabled using

systemctl enable sbd.service

The daemon is brought online on each node before corosync and Pacemaker are started, and terminated only after all other cluster components have been shut down - ensuring that cluster resources are never activated without SBD supervision.

Configuration via sysconfig
The system instance of "sbd" is configured via /etc/default/sbd. In this file, you must specify the device(s) used, as well as any options to pass to the daemon:

SBD_DEVICE="/dev/sda1;/dev/sdb1;/dev/sdc1"
SBD_PACEMAKER="true"

"sbd" will fail to start if no "SBD_DEVICE" is specified. See the installed template or section for configuration via environment for more options that can be configured here. In general configuration done via parameters takes precedence over the configuration from the configuration file.

Configuration via environment
SBD_DEVICE

Allows "string" defaulting to ""

SBD_DEVICE specifies the devices to use for exchanging sbd messages and to monitor. If specifying more than one path, use ";" as separator.

SBD_PACEMAKER

Allows "yesno" defaulting to "yes"

Whether to enable the pacemaker integration.

SBD_STARTMODE

Allows "always / clean" defaulting to "always"

Specify the start mode for sbd. Setting this to "clean" will only allow sbd to start if it was not previously fenced. See the -S option in the man page.

SBD_DELAY_START

Allows "yesno / integer" defaulting to "no"

Whether to delay after starting sbd on boot for "msgwait" seconds. This may be necessary if your cluster nodes reboot so fast that the other nodes are still waiting in the fence acknowledgement phase. This is an occasional issue with virtual machines.

This can also be enabled by being set to a specific delay value, in seconds. Sometimes a longer delay than the default, "msgwait", is needed, for example in the cases where it’s considered to be safer to wait longer than: corosync token timeout + consensus timeout + pcmk_delay_max + msgwait

Be aware that the special value "1" means "yes" rather than "1s".

Consider that you might have to adapt the startup-timeout accordingly if the default isn’t sufficient. (TimeoutStartSec for systemd)

This option may be ignored at a later point, once pacemaker handles this case better.

SBD_WATCHDOG_DEV

Allows "string" defaulting to "/dev/watchdog"

Watchdog device to use. If set to /dev/null, no watchdog device will be used.

SBD_WATCHDOG_TIMEOUT

Allows "integer" defaulting to 5

How long, in seconds, the watchdog will wait before panicking the node if no-one tickles it.

This depends mostly on your storage latency; the majority of devices must be successfully read within this time, or else the node will self-fence.

If your sbd device(s) reside on a multipath setup or iSCSI, this should be the time required to detect a path failure.

Be aware that watchdog timeout set in the on-disk metadata takes precedence.

SBD_TIMEOUT_ACTION

Allows "string" defaulting to "flush,reboot"

Actions to be executed when the watchers don’t timely report to the sbd master process or one of the watchers detects that the master process has died.

Set timeout-action to comma-separated combination of noflush|flush plus reboot|crashdump|off. If just one of both is given the other stays at the default.

This doesn’t affect actions like off, crashdump, reboot explicitly triggered via message slots. And it does as well not configure the action a watchdog would trigger should it run off (there is no generic interface).

SBD_MOVE_TO_ROOT_CGROUP

Allows "yesno / auto" defaulting to "auto"

If CPUAccounting is enabled default is not to assign any RT-budget to the system.slice which prevents sbd from running RR-scheduled.

One way to escape that issue is to move sbd-processes from the slice they were originally started to root-slice. Of course starting sbd in a certain slice might be intentional. Thus in auto-mode sbd will check if the slice has RT-budget assigned. If that is the case sbd will stay in that slice while it will be moved to root-slice otherwise.

SBD_SYNC_RESOURCE_STARTUP

Allows "yesno" defaulting to "yes"

If resource startup syncing is enabled then pacemakerd is gonna wait to be pinged via IPC before it starts resources. On shutdown pacemakerd is going to wait in a state where it has cleanly shutdown resources till sbd fetches that state.

The default is set when building SBD and Pacemaker from source. Going for ’no’ is safer if it can’t be assured that SBD and Pacemaker installed do both support the synchronization feature. When going with ’yes’ - also using package dependencies to assure SBD & Pacemaker both support the synchronization feature and are assuming the same default - an SBD configuration inherited via an upgrade doesn’t have to be altered to still benefit from the new feature.

SBD_OPTS

Allows "string" defaulting to ""

Additional options for starting sbd

Testing the sbd installation
After a restart of the cluster stack on this node, you can now try sending a test message to it as root, from this or any other node:

sbd -d /dev/sda1 message node1 test

The node will acknowledge the receipt of the message in the system logs:

Aug 29 14:10:00 node1 sbd: [13412]: info: Received command test from node2

This confirms that SBD is indeed up and running on the node, and that it is ready to receive messages.

Make sure that /etc/default/sbd is identical on all cluster nodes, and that all cluster nodes are running the daemon.

Pacemaker CIB integration

Fencing resource
Pacemaker can only interact with SBD to issue a node fence if there is a configure fencing resource. This should be a primitive, not a clone, as follows:

primitive fencing-sbd stonith:external/sbd \
params pcmk_delay_max=30

This will automatically use the same devices as configured in /etc/default/sbd.

While you should not configure this as a clone (as Pacemaker will register the fencing device on each node automatically), the pcmk_delay_max setting enables random fencing delay which ensures, in a scenario where a split-brain scenario did occur in a two node cluster, that one of the nodes has a better chance to survive to avoid double fencing.

SBD also supports turning the reset request into a crash request, which may be helpful for debugging if you have kernel crashdumping configured; then, every fence request will cause the node to dump core. You can enable this via the "crashdump="true"" parameter on the fencing resource. This is not recommended for production use, but only for debugging phases.

General cluster properties
You must also enable STONITH in general, and set the STONITH timeout to be at least twice the msgwait timeout you have configured, to allow enough time for the fencing message to be delivered. If your msgwait timeout is 60 seconds, this is a possible configuration:

property stonith-enabled="true"
property stonith-timeout="120s"

Caution: if stonith-timeout is too low for msgwait and the system overhead, sbd will never be able to successfully complete a fence request. This will create a fencing loop.

Note that the sbd fencing agent will try to detect this and automatically extend the stonith-timeout setting to a reasonable value, on the assumption that sbd modifying your configuration is preferable to not fencing.

Management tasks

Recovering from temporary SBD device outage
If you have multiple devices, failure of a single device is not immediately fatal. "sbd" will retry to restart the monitor for the device every 5 seconds by default. However, you can tune this via the options to the watch command.

In case you wish the immediately force a restart of all currently disabled monitor processes, you can send a SIGUSR1 to the SBD inquisitor process.

LICENSE

Copyright (C) 2008-2013 Lars Marowsky-Bree

This program is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version.

This software is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY ; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.

For details see the GNU General Public License at http://www.gnu.org/licenses/gpl-2.0.html (version 2) and/or http://www.gnu.org/licenses/gpl.html (the newest as per "any later").