[autofs] Abusing read_link for fun and profit

[H. Peter Anvin]

Well, the recent discussion gave me an idea for how one may be able to 
handle ghosting (and possibly direct mounts, although I'd still prefer 
to do it via dentry traps) cleanly: by putting a read_link method on the 
autofs leaf directories and push a lot of the lookup there.  That means 
lstat() will return the "ghost", as if it had been a symlink, and thus 
GUI tools and ls -l will be happy, but unlike a real symlink, tools like 
ls -l will not then try to look at the "real" directory behind it.

Thus all is good.

I was a bit uncertain if this could actually be done without causing 
infinite recursion, but some initial experiments seems to make it work.

With this, I might be able to do without dentry traps for autofs 5, 
although I would still prefer to have them, for the sheer reason that it 
might otherwise require a plain preposterous number of superblocks 
(mounts.)  The reason for this is that any time an automounted 
filesystem is on top of another mounted filesystem, as in the case of 
the following key:

foo	/ foosrv:/export/root	/bar foosrv:/export/bar

... it is functionally equivalent to foo/bar being a direct mount in 
order for everything to be done cleanly.  Similarly, "scaffolding" 
directories could be done using a ramfs, but now we're up to a total 
number of two superblocks per submount point (one for the "direct 
mount", one for the mount itself), plus one for the ramfs.  The ramfs 
could be eliminated by making the autofs more ramfs-like, which is 
pretty much a no-brainer anyway... the current autofs replicates way 
more code than it needs to.

This is way interesting...

	-hpa

_______________________________________________
autofs mailing list
[EMAIL PROTECTED]
http://linux.kernel.org/mailman/listinfo/autofs