dillon at apollo.backplane.com
Mon Jan 2 08:46:32 PST 2006
:On 2006-01-02, Csaba Henk <csaba.henk at xxxxxxx> wrote:
:> I just posted two patches to submit@ which make nullfs work with the new
:> namecache API (yes, it's without vnode overlaying).
:Anchors for the future:
What a nice Christmas present! I was hoping someone would come along
and rewrite nullfs!
The unmount patch also looks good. I see you are able to avoid
calling vget by using nc_mount, which is possible now that the
unmount code passes around mount structures instead of vnodes. Great!
I will get these into HEAD today. It won't make it into 1.4 for
the release but it will make it into a future update.
There is one issue with regards to overlayed namecache entries.
The issue occurs when you try to do a rename() or other namespace
operation and you have multiple namecache records refering to the
same physical namespace. Since rename() locks the namespace by locking
the related namecache records, it becomes possible for a namespace
operation in the nullfs layer to race a namespace operation in the
underlying filesystem layer. The system probably wouldn't be too happy
when this happens.
But, as I said, this issue is more the responsibility of the namecache
locking layer then the filesystem layer. In order to make it work
properly we'll probably need to extend the namecache structure to
recognize the difference between an overlay and a base filesystem
entry and to have the overlay also lock the base filesystem entry
(if present). Plus there would need to be a deadlock detector. Not a
trivial problem, but at least one that can be encapsulated within the
<dillon at xxxxxxxxxxxxx>
More information about the Kernel