Semaphore currently fails because their build VMs are based on Ubuntu 14.04 which ship util-linux (at with it, libmount) in version 2.20. A PPA I found would give us version 2.25, but not 2.26. Why the version bump, and which minimum libmount requirement does the new code have?

util-linux-2.26 is the version that introduced this monitor.

This is still WIP, so no reason to worry about Semaphore. If we end up merging it, we can always provide the old code as workaround.

@dvdhrm @zonque : If it helps, I can provide a PPA with util-linux 2.26 backported to 14.04. It would be a shame to lose the semaphore builds from now on. (I run my distcheck tests in wily, i. e. the current development series; that has a current kernel, util-linux, etc.)

@martinpitt : That would be great! But lets first figure out how to solve the libmount root-only restriction. If that requires util-linux changes, we have to wait for a new release, anyway.

@dvdhrm:

This is still not ready to be merged. It currently breaks systemd --user as libmount
requires you to be root to monitor 'utab'. Reason for this is
userspace_monitor_get_fd()" in libmount, which calls open(O_CREAT) on the
utab-lockfile (why?).

inotify_add_watch() requires that the file you want to monitor already exists, otherwise you need to monitor all directory for all IN_CREATE events. It's painful because you need to care about filenames on event etc.

open(O_CREAT) is not problem for non-root user if the file already exists. So the question is why on "systemd --user" the file /run/mount/utab.lock does not exist if there is already running pid1 systemd that already monitoring the file. (or do you mix old pid1 systemd and new systemd --user? :-)

@karelzak It's actually not that hard to handle situations where the dir doesn't exist. We recently added a similar fix to the sd-network stuff in 870395a. Please have a look!

@poettering I already had code to monitor all the directory, but it means that after each event you have to check filenames. The current implementation is simple and elegant, but unfortunately useless for non-root users :-(

I'll try to improve it...

@karelzak the commit i linked avoids having to check file names. It only looks for IN_ISDIR, that's all...

@poettering I see, nice solution. Thanks.

The libmount should be fixed now (util-linux/util-linux@7dc0f5c). I guess we can wait for util-linux v2.27 (Jul/Aug 2015) rather than try to backport etc...

I'll send an updated version of the patch later (it's necessary to use mnt_monitor_next_change() to verify that an event is really about utab change).

BTW, I think there is a bug in sd_network_monitor_flush(), the code assumes that inotify_add_watch() always allocates a new FD so it's safe to remove the old watch
by inotify_rm_watch(fd, e->wd). But this is not true, for already monitored filesystem objects inotify_add_watch() returns the same (already used) FD.

I guess that "mkdir /run/systemd/netif/FOO" in situation you monitor /run/systemd/netif/ will remove the watch of the directory at all. IMHO you need something like:

karelzak@625671b (the patch is absolutely no tested, just my guess when I read the code :-))

@karelzak hmm, makes sense... Though I figure we can simplify the patch a bit by simply returning the wd from monitor_add_inotify_watch() as return value, under the assumption it is strictly positive, and thus no ambiguity is created between returning an error and a wd.

@teg there's something to fix in sd-network!

Please, close. Replaced by new version #986.