[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Bug#580814: Parallellizing the boot in Debian Squeeze - ready for wider testing

On 05/11/2010 01:09 PM, Julian Andres Klode wrote:
On Tue, May 11, 2010 at 12:49:46PM +0200, Julien Cristau wrote:
On Tue, May 11, 2010 at 12:37:56 +0200, Julian Andres Klode wrote:

On Sun, May 09, 2010 at 05:25:16PM +0200, Tollef Fog Heen wrote:

Is it really a good idea to have init depend on such an option?
(It's also disabled in all defconfigs.)

Here's what I heard from Kay Sievers:

Apr 30 16:13:27<juliank>        kay: Is there a way to get it working without this option?
Apr 30 16:13:58<kay>    no, no chance
Apr 30 16:14:18<kay>    it's one of the building blocks to track/babysit processes
Apr 30 16:17:24<kay>    it will not be "debug" some day, but always required, yes
Apr 30 16:17:57<kay>    there is otherwise no way to reliably kill a service, you need a simple "container" to kill
Apr 30 16:18:17<kay>    otherwise processes can fork faster than you can kill them
Apr 30 16:18:38<kay>    cgroups provide a race free way to kill an entire group of processes
Apr 30 16:19:13<kay>    and they can also tell you that all processes of a service died ->  empty group
Apr 30 16:20:30<juliank>        kay: Does CONFIG_CGROUP_DEBUG have any impact on performance?
Apr 30 16:20:48<kay>    it should not be noticeable
Apr 30 16:21:03<kay>    only if you actually use them, and even then they are very cheap
Apr 30 16:21:20<kay>    it's not much more than a tag sticked to a process
Apr 30 16:21:38<kay>    and a way to handle the tagged things then ...

This does mean that when you use something like screen, the tty it was connected to is from then on unusable, right? As the cgroup that contains the screen process also contains the getty and it doesn't kill one without the other as that is in no way reliable :-)



Reply to: