> It seems that using the directory's mode bits etc. for a symbolic
> link's is leading to problems.

Yup.  Too much stuff assumes that any filesystem object has a full
complement of attributes, like owner and mode bits.  I can understand a
desire to make symlinks non-inode entities; it sounds like a good
objective to me.  But the way it's being done is causing trouble.  IMO,
for what it's worth, as long as they really are filesystem entities
they should be treated as such, and if they aren't, distinguished
"impossible" values should be returned, rather than grabbing other
values from somewhere more or less random.

> Has anyone considered using the link's target's mode bits if the
> target exists and the directory's mode bits if there is no target?

Quite aside from whether this would actually fix any of the problems, I
really really dislike having to namei() the target in order to lstat()
the link.  In particular, a symlink pointing onto a dead NFS server
should still be lstat()table, seems to me.

					der Mouse

			    mouse@collatz.mcrcim.mcgill.edu