rsync - Apr 2012 - Bug#666024: rsync --link-dest can incorrectly hardlink together destination files

[please Cc: 666024-forwarded at bugs.debian.org on any replies, thanks]

Please see the attached Debian bug report, which includes a helpful
bug-reproducing shell script; I've confirmed this still happens with
3.0.9.

Paul Slootman


On Tue 27 Mar 2012, Ian Jackson wrote:
> Package: rsync
> Version: 3.0.7-2
> 
> With rsync --link-dest, if two different source files (not hardlinked
> to each other) are both candidates for being hardlinked from files of
> the same name in the link-dest directory, both of them may be so
> hardlinked.  The effect is that on the destination the two files
> become hardlined to each other even though they are not hardlinked
> together in the source.
> 
> I discovered this because my backup system has a stage which checks
> that the generated backup directory is truly semantically identical to
> the source volume.  When --link-dest is used in a backup mechanism the
> effect is that if files were previously hardlinks on the source
> volume, but something makes them no longer hardlinks by breaking the
> link and copying the contents and attributes of the file, this change
> is not reflected in the backups.
> 
> The bug can be reproduced with the attached shell script.
> Run the script, and observe these lines in the output:
>  1390154    4 -rw-rw-r--   4 ian      ian             9 Mar 27 22:32
./dst/b
>  1390154    4 -rw-rw-r--   4 ian      ian             9 Mar 27 22:32
./dst/a
>  ...
>  1390155    4 -rw-rw-r--   1 ian      ian             9 Mar 27 22:32
./src/b
>  1390152    4 -rw-rw-r--   1 ian      ian             9 Mar 27 22:32
./src/a
> You can see that the inode numbers of src/a and src/b are different,
> whereas the inode numbers of dst/a and dst/b are the same.
> 
> Thanks,
> Ian.
> 
> #!/bin/sh
> set -e
> rm -rf d
> mkdir d
> cd d
> set -x
> 
> mkdir src
> echo contents >src/a
> ln src/a src/b
> 
> rsync -aH src/. linkdest
> 
> rm src/b
> cp -a src/a src/b
> 
> rsync -aH --link-dest=`pwd`/linkdest/. src/. dst
> 
> find -ls

On Tue 27 Mar 2012, Ian Jackson wrote:
> 
> I wrote:
> > With rsync --link-dest, if two different source files (not hardlinked
> > to each other) are both candidates for being hardlinked from files of
> > the same name in the link-dest directory, both of them may be so
> > hardlinked.  The effect is that on the destination the two files
> > become hardlined to each other even though they are not hardlinked
> > together in the source.
> 
> That's garbled.  I meant:
> 
> With rsync --link-dest, if two different source files (not hardlinked
> to each other) have identical contents and are both candidates for
> being hardlinked from files of their respective names in the link-dest
> directory, and the link-dest files are hardlinks to each other, rsync
> may wrongly make the two destination files hardlinks to each other.
> 
> Hopefully the test case script made this clear.
> 
> Ian.