Re: Symlink indirection

Re: Symlink indirection

[junkio]

"AW" == Andrew Walrond <andrew@walrond.org> gives an example of
a/{x,y,z}, b/{y,z}, c/z mounted on d/. in that order, later
mounts covering the earlier ones.

AW> echo "d/w" > d/w would create a new file in directory a.

Personally I'd rather expect this to happen in c/.  Imagine a/
being on read-only medium like CD-ROM containing bunch of source
files, b/ to hold patched source, and c/ to hold binaries
resulting from compilation.  That is,

    rm -fr a b c d
    mkdir a b c d
    mount /cdrom a
    mount --bind a d
    mount --bind --overlay b d
    (cd b && bzip2 -d <../patch-2.9.91.bz2 | patch -p1)
    mount --bind --overlay c d
    (cd c && make mrproper && cat ../.config >.config &&
    make oldconfig && make dep && make bzImage)

Back to your example; what do you wish to happen when we do
this?

    $ mv d/z d/zz && test -f d/z && cat d/z

Here we rename d/z (which is really c/z) to zz.  Does this
reveal z that used to be hidden by that, namely b/z, and "cat
d/z" now shows "b/z"?

Or just like the case of creating a new file, does the union
"remember" the fact that the directory "d" should not contain
"z" anymore, and "test -f d/z" fails?

Re: Symlink indirection

[Andrew Walrond]

junkio@cox.net wrote:
> "AW" == Andrew Walrond <andrew@walrond.org> gives an example of
> a/{x,y,z}, b/{y,z}, c/z mounted on d/. in that order, later
> mounts covering the earlier ones.
> 
> AW> echo "d/w" > d/w would create a new file in directory a.
> 
> Personally I'd rather expect this to happen in c/.  Imagine a/
> being on read-only medium like CD-ROM containing bunch of source
> files, b/ to hold patched source, and c/ to hold binaries
> resulting from compilation.  That is,
> 
>     rm -fr a b c d
>     mkdir a b c d
>     mount /cdrom a
>     mount --bind a d
>     mount --bind --overlay b d
>     (cd b && bzip2 -d <../patch-2.9.91.bz2 | patch -p1)
>     mount --bind --overlay c d
>     (cd c && make mrproper && cat ../.config >.config &&
>     make oldconfig && make dep && make bzImage)

A nice example.

Lets ditch --overlay and replace it with:

--transparent
	just like --overlay
--read-transparent
	a bit like a pane of glass. You can see stuff behind it (read) but 
throw a tomato at it and it sticks to the glass (write)

Then, simplifying your example a bit, we can mount the source cd then 
mount a directory --read-transparent over the top to hold both our 
patched source and compiled binaries.

Or if you want to keep patched-source and binaries seperate,

mount /cdrom stuff
mount --bind --read-transparent patched-src stuff

cd stuff
patch the src

mount --bind --read-transparent binaries stuff

compile your code

> 
> Back to your example; what do you wish to happen when we do
> this?
> 
>     $ mv d/z d/zz && test -f d/z && cat d/z
> 
> Here we rename d/z (which is really c/z) to zz.  Does this
> reveal z that used to be hidden by that, namely b/z, and "cat
> d/z" now shows "b/z"?

Yes - exactly

> 
> Or just like the case of creating a new file, does the union
> "remember" the fact that the directory "d" should not contain
> "z" anymore, and "test -f d/z" fails?
> 

No. Thats not necessary.

Re: Symlink indirection

[John Bradford]

> > "AW" == Andrew Walrond <andrew@walrond.org> gives an example of
> > a/{x,y,z}, b/{y,z}, c/z mounted on d/. in that order, later
> > mounts covering the earlier ones.
> > 
> > AW> echo "d/w" > d/w would create a new file in directory a.

I disagree.  It should create it in directory d, even though that is
the mount point.

A union mount should include files from another directory, but writes
should go to the actual named directory.

> > Back to your example; what do you wish to happen when we do
> > this?
> > 
> >     $ mv d/z d/zz && test -f d/z && cat d/z
> > 
> > Here we rename d/z (which is really c/z) to zz.  Does this
> > reveal z that used to be hidden by that, namely b/z, and "cat
> > d/z" now shows "b/z"?
> 
> Yes - exactly

Union mounts should be read only.

If read-write union mounts are needed, I don't think that we should
implement them significantly differently to the way they work in BSD.

John.

Union mounts

[John Bradford]

> I disagree.  It should create it in directory d, even though that is
> the mount point.
> 
> A union mount should include files from another directory, but writes
> should go to the actual named directory.
> 
> Union mounts should be read only.
> 
> If read-write union mounts are needed, I don't think that we should
> implement them significantly differently to the way they work in BSD.

That wasn't very well explained, what I mean is this:

Example:

# cd /
# mkdir foo
# mount -o union /dev/hda2 /foo
# echo foobar > foo/bar
# umount /dev/hda2
# cat foo/bar
foobar

That's what I would consider to be the most useful way to implement
union mounts - the contents of /dev/hda2 would be accessible,
read-only, at /foo/bar, with files that already exist in /foo/bar
replacing files that would otherwise be visible from /dev/hda2.

Writes would go to the directory foo, which is just an ordinary
subdirectory of the root filesystem.

This is completely different to the mount_union behavior in BSD, where
writes go to the most recently added union mount.

However, it might be best to implement things the BSD way for
compatibility reasons, but I'm not sure how widespread the use of
mount_union is.  It's probably not widely used.

John.