>> As for dmapi - they probably want it at the VFS layer so other FS can
>> implement it, rather than just tunneling through VFS with an ioctl to
>> XFS specific API.
Jeremy, you're over-simplifying and picking the low-hanging fruit.
Having a working implementation, no matter how objectionable this one may be
at the moment, does serve to help other people see the issues involved in
making DMAPI work; it does help people see the types of problems that DMAPI
has to solve. Everytime someone says it can be done entirely at the Linux VFS
level, and that filesystems shouldn't have to be modified to be aware of it, I
wonder if they've really studied the issues.
>From: Christoph Hellwig <hch@xxxxxxxxxxxxx>
>DMAPI is broken by design and can't work reliably with the Linux VFS
>design unless you take special care (see the mountpoint option for
>it).
The fact that the mount event must include the name of the mountpoint is
clearly questionable. Unfortunately, Linux is probably the first time anyone
who does dmapi has run into an OS that allows a filesystem to be mounted on
multiple mountpoints at the same time--it was easy for the people designing
the dmapi spec to not anticipate this complication.
On the other hand, I will grant that the mount event is one piece that can be
done entirely in the Linux VFS layer--originally, that's where I had it.
Dean
|