[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

[dm-devel] Re: lockspace interface



The first UUID is for the lockspace. You can have multiple lockspaces. Taking out a lock with the exact same name in two different lockspaces would not conflict.

You need to identify your lockspace across the cluster. In some cases, it makes sense to have the lockspace named according to the module that uses it - like "dm-clusterized-exception-module". In these cases, there are infrequently used locks - one per device - that are named according to the device UUID. So, if you have 4 cluster snapshots, you would have one lockspace with 4 infrequently used locks.

On the other hand, you may have a module (dm-raid45 comes to mind) where you are locking on a per sector basis. You certainly don't want locks to conflict from one device to another. So, you have a unique lockspace for each /device/. The lockspace name could be the UUID of the dm device. The locks would be named by their stripe, e.g. 1, 2, 3, 576, etc.

So, it is not 2 level locking. You merely create a cluster-wide known name for the lockspace and, once you have the lockspace, create locks by name (dm_cluster_lock_by_str) or by number (dm_cluster_lock).

The extra complexity of making the lock ops faster has not yet been done because it hasn't been needed. The implementation I had used the locks as little as possible. I figure I can introduce the lock concept in its simple form and improve things when/as necessary.

 brassow

On Sep 28, 2009, at 1:53 PM, Mikulas Patocka wrote:

Hi

What's the purpose of two-level uuid-locking?
I.e. you call
lockspace_handle = dm_cluster_lock_init(char *uuid1)
and then
dm_cluster_lock_by_str(void *lockspace_handle, char *uuid2, ...)

Why not use just one uuid?

Also think that locking will be heavily repeated (at least on a single
node), so it is needed to make lock as fast as possible --- without any
string operations, hashing, searching, associated spinlocks.

I'd propose to change the interface to

lockspace_handle = dm_cluster_lock_init(char *uuid)
dm_cluster_lock_by_str(void *lockspace_handle, ...)

--- assume that "uuid" already uniquely identifies the snapshot --- and
that second uuid is unnecessary. You can then get rid of "find_dcl"
function and its complexities (spinlock, list walking, strlen, strcmp) --- so that they won't be executed on the fast path --- and lockspace_handle
could be a single pointer to struct dm_cluster_lock.

Mikulas


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]