On 11/07/2014 06:38 AM, Randy Witt wrote:
On 11/06/2014 05:29 AM, Koen Kooi wrote:

Op 6 nov. 2014, om 12:59 heeft ChenQi <qi.c...@windriver.com> het volgende geschreven:



On 11/06/2014 06:33 PM, Koen Kooi wrote:
Op 6 nov. 2014, om 08:59 heeft ChenQi <qi.c...@windriver.com> het volgende geschreven:

On 11/06/2014 03:48 PM, Koen Kooi wrote:
Op 6 nov. 2014, om 08:32 heeft Chen Qi <qi.c...@windriver.com> het volgende geschreven:

In systemd, /etc/sysctl.conf is actually ignored by systemd-sysctl,
because this command only examine *.conf files under a bunch of directories
like /etc/sysctl.d or /usr/lib/sysctl.d.

The problem is we are used to configuring kernel parameters in /etc/sysctl.conf, so it would be really strange if the configuration in that file doesn't have any
effect.

This patch reference Fedora's solution to this problem, creating a symlink to
/etc/sysctl.conf under /etc/sysctl.d/.
Shouldn't this be done in procps instead?

Actually, the problem is not about `sysctl' command.
procps provides `sysctl', but busybox also provides this command.
It's very possible that on our generated image, procps is not installed but `sysctl' command is available. Both busybox's and procps's `sysctl' command takes /etc/sysctl.conf into consideration.
Right, but only procps installs that file.

As busybox provides `sysctl' utility, is it reasonable that it also provides a corresponding configuration file (/etc/sysctl.conf)? Should we make a patch for busybox?


Now, systemd provides a similar utility called `systemd-sysctl' which is executed at boot time via systemd-sysctl.service.

So our actually problem is that systemd-sysctl ignores /etc/sysctl.conf, which makes it somewhat strange, especially to users who are used to configuring parameters in sysctl.conf. And this patch solves this problem by adding a symlink under /etc/sysctl.d/.

That's why I think we should put this in systemd.
You're adding a symlink to a file which only exists if you install procps, which isn't in RDEPENDS.


As I said before, procps is *not* necessary for the sysctl mechanism to have effect.
(Think about systemd-based core-image-minimal image.)

Busybox provides `sysctl', systemd provides `systemd-sysctl'.
(It's an easy program, there might exist other packages that provide it too.)

/etc/sysctl.conf is a configuration file which is very likely to be modified or created by administrators to configure kernel parameters. (You can't expect administrators to all start learning systemd, trying to understand the gap and differences. In addition, they may have scripts that edit /etc/sysctl.conf to automate their work.)

The point of the symlink is to ensure that when users edit /etc/sysctl.conf (or create one), configurations in that file will have effect at boot time.

Just think about this problem from a standpoint of user experience.

You still haven't convinced me that shipping a broken symlink is a good idea. I'm pretty sure it's not even allowed in out commit guidelines.

I agree, even in Fedora's case, both /etc/sysctl.conf and /etc/sysctl.d/99-sysctl.conf come from the same package, initscripts-9.51-2.



Koen & Witt,

When I started out to solve this problem, I basically came up with three solutions. Please see details below. I chose to use solution 3 because I think it's better than the other two. Please tell me which one you like, or please tell me your solution.

Solution 1:
Make 'sysctl.conf' managed by ALTERNATIVES mechanism in OE. Busybox, procps and systemd all provide this file. And the symlink is still in systemd.
(My opinion for this one: over engineering)

Solution 2:
Make base-files provide sysctl.conf and /etc/sysctl.d/99-sysctl.conf (in case of systemd). procps no longer provides sysctl.conf. (My opinion for this one: sysctl.conf doesn't logically belong to base-files.)

Solution 3:
Let systemd provide a broken symlink '/etc/sysctl.d/99-sysctl.conf' and don't change anything else.

Which of the above three solutions do you like? Do you have any other suggestion?

//Chen Qi

--
_______________________________________________
Openembedded-core mailing list
Openembedded-core@lists.openembedded.org
http://lists.openembedded.org/mailman/listinfo/openembedded-core

Reply via email to