Hi linux-fsdevel,
When you go into /proc/<pid>/fd and type 'ls -l', you see the numbered
file descriptors as symlinks to the real files, devices, etc. Opening
the file descriptors allows you to access the contents, even if the
object has already been deleted.
In several virtualized environments (e.g. kubernetes/docker, chroot,
etc), you can still access the contents, but the target of the symlink
(output of readlink) is shown as an absolute path from the *virtual*
root directory. This may not be the same environment the caller is in,
resulting in (apparently) broken links.
One solution would be to show the path relative to /proc/<pid>/root
instead, turning
"6 -> /some/directory/foo" (broken link)
into
"6 -> ../root/some/directory/foo" (relative link using process' actual
root directory),
or possibly
"6 -> /proc/1234/root/some/directory/foo" (absolute link using actual
root directory).
Pros:
-----
+ This fixes some (maybe all?) of the erroneously broken links in
/proc/<pid>/fd
+ This would work for all processes, virtualized or not.
+ It can't break existing code (any further), unless existing code
depends on erroneously broken links.
+ It's short.
Cons:
-----
- Maybe there is existing code that assumes absolute links instead of
relative links?
Implementation details:
-----------------------
It looks like adding some code to do_proc_readlink
(https://github.com/torvalds/linux/blob/690e2aba7beb1ef06352803bea41a68a3c695015/fs/proc/base.c
) might do the trick; d_path() already appends ' (deleted)' to the link
if necessary, prepending '../root' to paths that start with '/' should
work the same way.
Question:
---------
I'd like to write a patch for this, but after reading some of the
patches in linux-fsdevel, it looks like there are lots of
implementation details I could get wrong through ignorance of kernel
internals. What would you recommend for a good example to copy?
Conversely, if this is a 5-minute job for someone here, would anyone
like to take it on?
Thanks!
--
Charles Howes <vger.kernel.org@xxxxxxxxxx>
