strange interaction between open and cwd

[Baz Walter]

On 05/05/10 00:44, Nobody wrote:
> On Tue, 04 May 2010 14:36:06 +0100, Baz Walter wrote:
>
>> this will work so long as the file is in a part of the filesystem that can
>> be traversed from the current directory to the root. what i'm not sure
>> about is whether it's possible to cross filesystem boundaries using this
>> kind of technique.
>
> At least on Linux, the kernel "fixes" the links at mount points, i.e.
> within the root directory of a mounted filesystem, ".." refers to
> the directory containing the mount point on the parent filesystem, while
> the mount point refers to the root directory of the mounted filesystem.
>
> This also appears to work correctly for bind mounts (mounting an arbitrary
> directory to another directory, which results in a directory hierarchy
> appearing at multiple locations within the filesystem), i.e. ".." refers
> to the appropriate directory for each "instance".
>
> OTOH, the algorithm can fail if a directory is moved (whether by rename()
> or remounting) between the stat("..") and the listdir().

i think the algorithm also can't guarantee the intended result when 
crossing filesystem boundaries. IIUC, a stat() call on the root 
directory of a mounted filesystem will give the same inode number as its 
parent. so if several filesystems are mounted in the same parent 
directory, there is no way to tell which of them is the "right" one.