Question about the usage of SANLock as the cluster lock

[Qixiaozhen]

Hi, all

I have an idea of cluster lock with sanlock. However, I am not familiar with the principle detail of sanlock. This mail is sent to the expert of sanlock to discuss its feasibility.

My idea about the cluster lock:

1)       Initialize the lockspace and resource at first.

2)       Add the hostId of nodes into the lockspace after step 1.

3)       Loop:

4)          If the node wants to acquire the cluster lock, "sanlock.acquire()" can be called. "sanlock.release()" will be called once the operation finished.

What about the feasibility of this idea?

I hava debuged the demo code of sanlock named 'python/example.py' step by step in two nodes(Node1, Node2). Demo code likes the following segment:

Demo code:
L1     import sanlock
L2     sd_path = "/dev/pool/"
L3     ids_file = sd_path + "ids_file"
L4     lease_file = sd_path + "lease_file"
L5     LEASE_NAME = "LEASE"

L6     sanlock.init_lockspace(sd_path, ids_file)
L7     sanlock.init_resource(sd_path, LEASE_NAME, [lease_file])

L8     host_id = 2                                            <-----different node has its own host_id here.
L9     sanlock.add_lockspace(sd_path, host_id, ids_file)

L10   _sanlock_fd = sanlock.register()

L11   sanlock.acquire(sd_path, LEASE_NAME,[lease_file], slkfd=_sanlock_fd)

L1~L8 are executed in node1 and node2.

Execute in same time : L9 cannot be executed in the same time of these two nodes and one node would fail.
Execute one by one  :When node1 was added its hostId into the lockspace successfully first, the add_lockspace of node2 would overwrite the lockspace in "ids_file" without any exception. There are only one node which writes its timestamp into its sector.

Appreciate for your reply.

Thank you very much.

Qi

[David Teigland]

On Tue, Jun 18, 2013 at 09:20:07AM +0000, Qixiaozhen wrote:
> My idea about the cluster lock:
>   Initialize the lockspace and resource at first.
>
>   Add the hostId of nodes into the lockspace after step 1.
>
>   Loop:
>
>      If the node wants to acquire the cluster lock, "sanlock.acquire()" can be called. "sanlock.release()" will be called once the operation finished.
>
>
>
> What about the feasibility of this idea?
>
> I hava debuged the demo code of sanlock named 'python/example.py' step by step in two nodes(Node1, Node2). Demo code likes the following segment:

I'm not sure exactly how you are running this, but you're probably doing
it wrong.

Here is an example using the command line and a simple test program I've
attached (compile with -lsanlock).

1. set up shared storage (/dev/sdb)

host1: vgcreate test /dev/sdb
host1: lvcreate -n leases -L 1G test
host2: vgscan
host2: lvchange -ay /dev/test/leases

2. start the daemons

host1: modprobe softdog
host2: modprobe softdog
host1: wdmd
host2: wdmd
host1: sanlock daemon
host2: sanlock daemon

(it's best to use a real watchdog driver instead of softdog)

3. initialize the lockspace (named "LS") and the resource (named "RX")

host1: sanlock client init -s LS:0:/dev/test/leases:0
host1: sanlock client init -r LS:RX:/dev/test/leases:1048576

(done from only one host)

4. add the lockspace

host1: sanlock client add_lockspace -s LS:1:/dev/test/leases:0
host2: sanlock client add_lockspace -s LS:2:/dev/test/leases:0

(this will take 20+ seconds)
(each host uses a different host id)

5. verify that both hosts have joined the lockspace

host1: sanlock client host_status
lockspace LS
1 timestamp 1687
2 timestamp 1582

host2: sanlock client host_status
lockspace LS
1 timestamp 1687
2 timestamp 1561

6. acquire/release lock on RX

host1: sanlk_lockr LS RX /dev/test/leases 1048576 5
host2: sanlk_lockr LS RX /dev/test/leases 1048576 5