[PATCH v2] tar: regression with hardlinked symlinks

[Harald van Dijk]

ping

FWIW, this problem actually came up with an archive containing a fsroot 
including a busybox installation again, but this time, the archive was 
created after a deduplication step had hardlinked identical files and 
symlinks.

On 4/1/18 1:37 PM, Harald van Dijk wrote:
> When the creation of symlinks got delayed until after all other files 
> had been extracted, that didn't take into account that the archive may 
> contain hardlinks to those symlinks. The extraction of those hardlinks 
> did not get delayed so would fail.
> 
> This patch generalises the symlink_placeholders approach to a 
> link_placeholders which allows delaying the extraction of hardlinks too. 
> It changes the order in which delayed links are added to ensure that a 
> hardlink to an earlier symlink will create the symlink first.
> 
> It delays all hardlinks without regards to the target. It may be 
> desirable to instead tweak it to determine whether the link target is a 
> delayed symlink, but that is a cost/benefits tradeoff. As hardlinks are 
> relatively rare, I guess the benefit of such an optimisation would be 
> small, but I am not really qualified to make that determination.
> 
> Test case:
> 
>    mkdir dir
>    >dir/a
>    ln -s ../dir/a dir/b
>    ln dir/b dir/c
>    mkdir new
>    tar cf - dir/* | tar -C new -xf -
> 
> busybox ad4e9613:
> 
> tar: can't create hardlink 'dir/c' to 'dir/b': No such file or directory
> 
> busybox ad4e9613 + patch:
> 
> No error output. The archive is successfully extracted.
> 
> The security hole that the delayed extraction of symlinks was meant to 
> prevent should not be re-introduced by the delayed creation of 
> hardlinks: the parent components of archive entries still get created 
> first, so if an archive contains a file "exploit", a symlink "parent -> 
> ..", and a hardlink "parent/exploit -> exploit", the presence of that 
> hardlink causes "parent" to be immediately created as a regular 
> directory. The symlink creation "parent -> .." will therefore fail. This 
> could use careful review though.
> 
> Cheers,
> Harald van Dijk