I want to use the functionality of umount -l, then remove the underlying device as safely as possible afterward.

  • I can't use umount --force to unmount the filesystem as it has become invisible to new processes.
  • I can't use lsof for an accurate list of the open files as the filesystem has become invisible to new processes.
  • If I use lsof before umount -l, there is a race contition of a new file being opened in between the two invocations.

I'm testing a work-around: sync && blockdev --setro /dev/<device>

The manual for blockdev --setro only says:

Set read-only.

Is the man page missing something? This seems to create a file on a --setro device:

# mount /dev/loop0 mountpoint/
# blockdev --setro /dev/loop0
# echo test > mountpoint/f
# sync
# umount mountpoint
# mount /dev/loop0 mountpoint/
mount: /tmp/mountpoint: WARNING: device write-protected, mounted read-only.
# cat mountpoint/f


$ uname -a
Linux svelte 4.9.39-1-MANJARO #1 SMP PREEMPT Fri Jul 21 08:25:24 UTC 2017 x86_64 GNU/Linux
$ blockdev --version
blockdev from util-linux 2.30

I guess blockdev --setro works similar to chmod: It affects only future openings of the object.

But I can offer a work-around for your detection problem:

  1. The /proc/$PID/cwd value is changed to / after the lazy unmount.
  2. The paths of open files of the process shown in /proc/$PID/fd are moved up to /, e.g. /mnt/tmp/output becomes /output.

So you can first filter all processes with cmd /. There may be false positives among them but this is very fast. The next step (not necessarily complete but probably faster) is to check /proc/$PID/fd of all these processes for files which do not exist at their shown paths.

The complete but probably not so fast check is to run stat for all files in /proc/$PID/fd. It shows the original device. So you might check this value before the umount to make things easier.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.