I'm not sure to understand your first sentence. Resolving the last component is already the existing behavior, but maybe not the intended one.

Anyway, I agree that the path should be resolved without its last component.

I wrote a new patch for this. I hope that this one will not break anything.

Rémy.


2013/4/2 Pádraig Brady <P@draigbrady.com>
On 04/01/2013 09:10 PM, Rémy Lefevre wrote:
> Rémy.
>
>
> 2013/4/1 Pádraig Brady <P@draigbrady.com <mailto:P@draigbrady.com>>
>
>     On 04/01/2013 03:40 PM, Rémy Lefevre wrote:
>     > Overwriting relative symlink leads to undesirable behavior. Consider the
>     > following example:
>     >
>     > # Create some directories
>     > mkdir test
>     > mkdir test/folder1
>     > mkdir test/folder2
>     >
>     > #Create some files
>     > touch test/folder1/file1
>     > touch test/folder1/file2
>     >
>     > #Create a relative symlink in folder2 to file1
>     > ln -sr test/folder1/file1 test/folder2/link
>     >
>     > #Check the link
>     > ls -l test/folder2/link
>     > # Correctly output a link to ../folder1/file1
>     >
>     > #Overwrite the symlink to point to file2
>     > ln -sfr test/folder1/file2 test/folder2/link
>     >
>     > #Check the link
>     > ls -l test/folder2/link
>     > # Wrongly output a link to file2 instead of ../folder1/file2
>     >
>     >
>     > This undesirable behavior is due to a dereferencing of the target when the
>     > relative path is computed. Passing CAN_NOLINKS flag to
>     > canonicalize_filename_mode solves the problem.
>
>     Doing that though breaks `ln -sr target1 target2 dir` where dir is a symlink.
>     Also if /some/other/component of the path is a symlink, you probably want
>     that resolved? You might even want the final component of the link name
>     resolved in some cases. So perhaps the approach here is to only disable
>     dereferencing when -n is set, and even then only for the last_component()?
>
>     In other words, `ln -nsf ...` means update the specified link name no matter what it is.
>     -n used only be significant when the link name was to a directory,
>     but with -r it's also significant if linking outside its containing directory.
>
>     I'll sort out a patch later.
>
> You are right. It breaks any path composed of link. Sorry for this bad patch.
>
> But could you provide me an example where the final component of the link should be resolved ? Does it make sense as the link will be overwritten? I must be missing something.

Yes resolving the last component would be a departure from existing behavior.
So it there is no need to conditionalize this on -n, and we just need
to resolve the path without the last_component() and then tack that on.

thanks,
Pádraig.