Can somebody explain this changing `du` output?

Question

Just tripped across a weird case of du. Putting arguments one way gives:

$ du -s htdocs/.git/ feuerkraft/.git/ media/.git/
40564   htdocs/.git/
40740   feuerkraft/.git/
204 media/.git/

Putting them in a different order changes the size of the last directory:

$ du -s media/.git/ htdocs/.git/ feuerkraft/.git/ 
40696   media/.git/
40564   htdocs/.git/
248 feuerkraft/.git/

Those should all be regular directories:

$ ls -l */.git -d -i
4891918 drwxrwxr-x 1 ingo ingo 138 Mar 28 08:41 feuerkraft/.git
4892896 drwxrwxr-x 1 ingo ingo 138 Mar 28 08:44 htdocs/.git
4893874 drwxrwxr-x 1 ingo ingo 138 Mar 28 08:42 media/.git

Doing it manually gives:

$ for i in htdocs/.git/ feuerkraft/.git/ media/.git/; do du -s "${i}"; done
40564   htdocs/.git/
40740   feuerkraft/.git/
40696   media/.git/

When I copy the directory with cp -axv the problem stays, if I copy the directory with cp -rxv it disappears.

Digging deeper:

$ du -s feuerkraft/.git/objects/pack/pack-26b69bdec4b3fd894bf7984d23e1938b522e6d53.pack media/.git/objects/pack/pack-26b69bdec4b3fd894bf7984d23e1938b522e6d53.pack htdocs/.git/objects/pack/pack-31a80589fbb132827d19b718c07aba1aa6bb9c02.pack  | wc -l
2
$ ls feuerkraft/.git/objects/pack/pack-26b69bdec4b3fd894bf7984d23e1938b522e6d53.pack media/.git/objects/pack/pack-26b69bdec4b3fd894bf7984d23e1938b522e6d53.pack htdocs/.git/objects/pack/pack-31a80589fbb132827d19b718c07aba1aa6bb9c02.pack  | wc -l 
3

du list only 2 files if three are given, ls lists 3.

Filesystem is btrfs. Is this a bug or something with compression/copy-on-write or whatever?

I think it is a bug, because there is no reason why the file size should be different(unless they are being modified at the same time). You should file it to concerned developers. — Registered User, Mar 28 '15 at 08:05

Grumbel · Accepted Answer · 2015-03-28T13:15:50.957

This explains it:

$ ls -l feuerkraft/.git/objects/pack/pack-26b69bdec4b3fd894bf7984d23e1938b522e6d53.pack media/.git/objects/pack/pack-26b69bdec4b3fd894bf7984d23e1938b522e6d53.pack htdocs/.git/objects/pack/pack-31a80589fbb132827d19b718c07aba1aa6bb9c02.pack -i
13150428 -r--r--r-- 2 ingo ingo 41191354 Mar 15 14:14 feuerkraft/.git/objects/pack/pack-26b69bdec4b3fd894bf7984d23e1938b522e6d53.pack
13151244 -r--r--r-- 1 ingo ingo 41197513 Mar 28 08:44 htdocs/.git/objects/pack/pack-31a80589fbb132827d19b718c07aba1aa6bb9c02.pack
13150428 -r--r--r-- 2 ingo ingo 41191354 Mar 15 14:14 media/.git/objects/pack/pack-26b69bdec4b3fd894bf7984d23e1938b522e6d53.pack

The feuerkraft and media files point to the same inode, i.e. hard links. This is a feature of git clone and documented in the manpage:

$ man git clone
...
   --no-hardlinks
       Force the cloning process from a repository on a local filesystem to copy
       the files under the .git/objects directory instead of using hardlinks.
       This may be desirable if you are trying to make a back-up of your
       repository.

This also explains why git sets the pack files to read-only, always wondered about that. du also has an option to count hard links double:

$ man du
...
   -l, --count-links
          count sizes many times if hard linked

And that gives the expected results:

$ du -ls media/.git/ htdocs/.git/ feuerkraft/.git/
40696   media/.git/
40564   htdocs/.git/
40740   feuerkraft/.git/

Can somebody explain this changing `du` output?

1 Answers1