On Tue, 29 May 2001, Greywolf wrote:

> Would I be putting it more simply to the lesser technical than yourselves
> (such as, oh, me :-) to say that the layers still need that struct lock *
> drilling down to the base layer so that locking on a node will actually
> work, i.e. all layers ATOP the base point down to the lock at the base
> level, while the struct lock * at the base is a real lock on the node,
> hence any references on any other layer will actually perform the lock op
> [if they're even permitted to do that]?
> 
> Or is that not even close?

I think that's about it, except that "need" is a bit strong. With what we
have now, there's a struct lock * in the vnode. Upper layers will look at
the struct lock * in the lower node, and put it in their vnode (so a layer
on a layer will work right).

When VOP_LOCK() gets called on a layered node, if the struct lock * is
non-NULL, the layer just goes ahead and calls the lock manager using that
lock.

If the struct lock * is NULL, the layer calls VOP_LOCK on the underlying
vnode. The clustered file system Chuck spoke of would export a NULL.

So right now, the struct lock * is a speed improvement - we don't have to
recursivly call down a stack of nodes to be able to lock & unlock a node.

Chuck's changes would still work - we'd just always recursivly call down.
Life would just be a bit slower.

Take care,

Bill