lock_gulmd - Grand Unified Lock Manager

lock_gulmd --cluster_name <string> [options]

lock_gulmd is a lock manager for GFS that was designed to take advantage of the way GFS uses locks, and the way data is transferred over TCP/IP.

lock_gulmd supports failover so that your gfs cluster can keep running if the lockserver machine dies (or if one machine of a lockserver cluster dies). You can also run lock_gulmd on the same nodes that mount your gfs filesystem(s).

lock_gulmd is really three servers in one: It contains the core, locktable interface, and locktable servers. Each of these gets its own process, and the locktable server may get more than one process depending on your config. Core is responsible for client membership and heartbeats. Locktable and locktable interface handle the locking. Multiple locktable processes can be run to improve performance on SMP systems via the lt_partitions option in the configuration.

OPTIONS You can completely configure gulm from the command line. If you do this, you need to use the same options on every node in the cluster. You should always specify the --cluster_name option, and then either the --use_ccs or the --servers options. Then other options should follow.

-h --help Print usage information, then exit.

-V --version Print version information, then exit.

-c --use_ccs Make the calls out to ccsd load config from cluster.conf. Will override any previous options. Likewise, options that follow this will override settings from ccs.

-v --verbosity verbose flags Sets which types of messages can be logged.

verbose flags is a comma separated list of the possible flags. If a flag is prefixed with a '-', it is unset, other wise it is set. The special flag 'clear' unsets all verbosity flags. Any flag that is not recognized is ignored.

The verbosity flags for gulm:

l l. Network Basic network related messages Network2 T{ More specific network messages T} Network3 Nothing currently Fencing T{ When calling out to the fencing sub-system T} Heartbeat T{ Every heartbeat sent and received T} Locking T{ Various internal informational messages about the locks T} Forking T{ Anytime a child processes is spawned T} ServerState T{ Print out a message when ever the server changes state, saying what state it is now in. T} JIDMap T{ Details of each JID Mapping request T} JIDUpdates T{ JID Mapping updates sent to slaves T} LockUpdates T{ Lock requests sent to slaves T} LoginLoops T{ Messages related to searching for and becoming the Master T}

ReallyAll All messages above Default same as -v "Network,Fencing,Forking" All T{ same as -v "ReallyAll,-Network3,-JIDUpdates,-LockUpdates" T}

The verbose flags can be changed while lock_gulmd is running with gulm_tool.

Default is Default. (witty no?)

-s --servers server list Comma seperated list of nodes that can be master servers. No default.

You can use either IPs or names. Either IPv4 or IPv6 addresses can be used. Node names are looked up via libresolv into IPs. IPv6 addresses will be used over IPv4 addresses.

-n --cluster_name string The name of this cluster. No default.

--heartbeat_rate number Number of seconds to wait before checking for missed heartbeats. 2/3 of this value is the rate at which nodes send heartbeats to the master server. You can specify this as a floating point number to get less than a second times.

Use subsecond values at your own risk, as varying network loads can cause false node expirations.

Default is 15.

--allowed_misses number How many heartbeats can be missed before the node is considered to have expired. Default is 2.

--new_connection_timeout number How many seconds to wait before deciding a new socket is bogus and dropping it. Can use floating point for sub second values. Default is 15.

--master_scan_delay number How many seconds between each probe for a new master server. Can use floating point for sub second values. Default is 1.

--coreport number Which port does the core server listen on and connect to. Default is 40040.

--ltpxport number Which port does the ltpx server listen on and lock clients connect to. Default is 40042.

--ltport number Which port does the LT server listen on, and LT and LTPX clients connect to. If you have multiple lt_partitions, the LT's id is added to this to get its port. (Using the default, LT000 is at 41040, LT001 is at 41041, ect) Default is 41040.

--fence_bin string The name of the program that handles fencing nodes for gulm. This needs to be a full path.

The program takes a single argument, the name of the node to be fenced. If the program returns an exit status of 0, then the fencing was succesful. Otherwise gulm waits 5 seconds, and calls it again.

Default is /sbin/fence_node

--run_as string User to switch into and run as. Default is root. (which is not good.)

--lock_dir string The directory to place and store the pid lock files. Default is /var/run/sistina/

--lt_partitions number Number of Lock Tables to run. If more than one there will be multiple LTs, and the LTPXes will stripe the locks across the LTs. This is for preformance on servers with multiple CPUs. Default is 1.

--name string Set the name of this gulm server to string instead of the default. Default is output of `uname -n`.

--ip ip addr Set the IP of this gulm server to ip addr instead of trying to resolve it from the name. Default is to resolve this gulm servers name into an IP.

--ifdev network device name Use the IP that is configured for this network device name. If there is an IPv6 address, that is used. Otherwise the IPv4 will be used. No default.

-e When switching into daemon mode, leave stderr and stdout open.

-d Do not damonize. (will still fork each server.)

-C Load all config items (command line arguments and ccs data), and print configuration as we see it and exit.

SIGNALS

SIGTERM This signal is ignored. To stop gulm you should use the shutdown command to gulm_tool.

SIGUSR1 Dump out internal tables for debugging. This creates a bunch of files in /tmp (or whatever you have TMPDIR set to). All of these start with the prefix Gulm_ and will be appended to if the file already exists.

Much of the information in these dump files is available via gulm_tool, and gulm_tool is the preferred method of getting this information; the action of dumping these tables out stops all other activity and thus can have negative affects on the performance of lock_gulmd. You should not send this signal unless you really want those dump files and know what to do with them.

PRIVATE NETWORKS

Getting gulm to use a private network is a matter of not relying on IP lookups. Sepcify all of the <lockserver/> entries with IP addresses. Then when starting lock_gulmd, make use of either the --ip option or the --ifdev option.

EXAMPLES

When using ccs to configure gulm, start the daemon with: lock_gulmd --cluster_name foo --use_ccs

or: lock_gulmd -n foo -c

lock_gulmd can be run without CCS: lock_gulmd -n foo -s 192.168.1.1,192.168.1.2,192.168.1.3

This adds the following two verbose flags to the default set: lock_gulmd -n foo -c -v "Heartbeat,Locking"

Show only the Network messages: lock_gulmd -n foo -c -v "clear,Network"

Use the ip on the second ethernet device, and call me bar: lock_gulmd -n foo -c --name bar --ifdev eth1

Stopping the server: gulm_tool shutdown localhost

FILES /var/run/sistina/lock_gulmd_core.pid

/var/run/sistina/lock_gulmd_LTPX.pid

/var/run/sistina/lock_gulmd_LT000.pid These are the pid lock files to keep more than one instance of the servers running per node. They can be put elsewhere via a configuration option.

SEE ALSO gulm_tool(8), lock_gulmd(5), ccs(7)