hash-based notify and dnotify?

[James Peach]

On Jan 30, 2007, at 1:06 PM, Volker Lendecke wrote:

> Hi!
>
> Right now I'm looking at the file change notify support in
> Samba3. I am at a point where I've got the Samba4
> implementation working so that it almost survives the same
> set of RAW-NOTIFY subtests that Samba4 does.
>
> Samba4 only contains an inotify backend to call out to Unix,
> Samba3 contains 3 ones: dnotify, FAM and a hash-based system
> that regularly scans directories.
>
> If you google for dnotify you will find that it is first not
> the cleanest interface, it can miss some updates and worst
> it does not tell us what has changed. It only tells us
> *something* has changed.
>
> The directory scanning interface is broken in similar ways,
> the worst thing is: It is terribly inefficient.
>
> FAM seems mostly unused these days

Only IRIX AFAIK.

> but has a moderately
> usable interface. You can select on a socket and ask the
> subsystem for what has changed.
>
> Tridge has implemented a scheme where the Samba-internal
> events are handled with correct Windows semantics, at least
> to the level where we right now have seen it in the tests.
> The backends I described help us in seeing Unix-triggered
> events that were caused outside of Samba.
>
> I would like to ask if we can drop the dnotify and
> hash-based interfaces and keep inotify as our main
> interface. This would leave the Linux 2.4 based and other
> unix systems without inotify behind for the unix-triggered
> events. The workaround for these would be to compile the SGI
> fam, which from what I've seen implements both dnotify and a
> hash-based system similar to what Samba3 has done for ages.

Sounds good to me. I'm not sure that FAM is maintained anymore, so it  
might not be as much of a fallback as you might like.

--
James Peach | jpeach at samba.org