Re: Bug#2828: Horrible aliasing problem (was xosview)
- To: firstname.lastname@example.org
- Subject: Re: Bug#2828: Horrible aliasing problem (was xosview)
- From: email@example.com (Kai Henningsen)
- Date: 01 May 1996 10:25:00 +0200
- Message-id: <682Kpf$UcsB@khms.westfalen.de>
- In-reply-to: <m0uENu8-0002ZFC@chiark.chu.cam.ac.uk>
firstname.lastname@example.org (Ian Jackson) wrote on 30.04.96 in <m0uENu8-0002ZFC@chiark.chu.cam.ac.uk>:
> Aaaaaaarrgh ! The classic pointer-aliasing problem strikes
> filesystems ...
> Suggestions and/or sympathy welcome.
Just some random thoughts on the problem.
First, we can partition this in two sub-problems:
A. (At least one of) the directories is a symlink. The file isn't.
B. The file is a symlink.
(Of course, both possibilities go for both old and new file.)
B should be somewhat more easy; you know a file's a symlink when you
install it, and it ought not to change, and there's not much reason for
system admins to change it.
So A is the problem here.
Maybe you could only check the directory components, and make sure that
you check each component only once. That ought to have much better
performance than checking every file, but is much harder to do.
Second, an obvious (but, from earlier comments from you, not very
attractive) way would be md5sums on every file.
Third, in the case that triggered this, dpkg was removing a file it just
installed, wasn't it? This case could certainly be solved by keeping
device/inode numbers during the install only.
Fourth, this is probably the exact reason mirror doesn't handle symlink-
directory changes well.
Hmm. I tend to think the check-directory-components thing is the best bet