[Unionfs] UNIONFS_MMAP (was: rw nfs branch ...)

[Shaya Potter]

On Tue, 2006-03-07 at 15:24 -0500, Erez Zadok wrote:
> In message <200603071424.07490.hpj at urpla.net>, Hans-Peter Jansen writes:
> 
> > > About /proc soft link problem, I guess Shaya Potter's mmap patch is
> > > the answer. But I didn't try it and it seems to be needed some
> > > fixing, for example lseek at readdir.
> > 
> > Yeah, Shaya's mmap support patch sounds very promising, and I'm going to 
> > try it as soon as my time window will allow more experiments. But 
> > pressure is high currently to deploy the stuff..
> 
> We've begun the long journey to get full mmap support working.  Just in the
> past day or so we've applied Shaya's patch, slightly modified to the CVS
> tree.  It is NOT on by default: you'll have to compile unionfs with
> UNIONFS_MMAP.

thanks for taking the patch.

> We have done very little testing on it; it's highly experimental, but looks
> promising as a starting point.  So I preferred to have at least _something_
> available for people to try, and hopefully help us enhance and fix it (while
> we're continuing to stabilize the rest of the code in preparation for a
> first release).

the main thing I didn't like was the fsync() after copyup.  It was
necessary at one point, but I'm unsure its necessary now.  Need to do
some more tests, which I don't directly have time right now (but might
happen do to other things I need to test).

another comment, I tried to the "mapping" flip trick, but from my
experience, that really wont work with regular linux file systems, as it
depends on the buffer's being consistent (for instance for ext2/3),
which is not the case if copyup happens.

> So, consider this a request for testing from those of you who feel
> comfortable enough with the unionfs code, to try out UNIONFS_MMAP, find
> what's missing/broken, and give us feedback (or patches :-)

people should try doing this test.

map a file write/shared, modify it, unmount the fs (no need for copyup
to be involved, I seem to recall this being a problem w/ wrapfs as well,
believe I filed a bug a while ago).

see if the changes persist, I remember that being a problem.