From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jacob Pan Subject: Re: [PATCH, v3 2/2] cgroups: introduce timer slack subsystem Date: Fri, 4 Feb 2011 09:27:55 -0800 Message-ID: <20110204092755.0bbfaa7a__29724.2486537018$1296840659$gmane$org@putvin> References: <1296679656-31163-1-git-send-email-kirill@shutemov.name> <1296679656-31163-3-git-send-email-kirill@shutemov.name> <20110202145605.6c9006fa@putvin> <20110203092229.GB1083@shutemov.name> <20110203095117.76a568fc@putvin> <20110203115743.1ef9a4c2@putvin> <20110204133439.GA7181@shutemov.name> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: <20110204133439.GA7181-oKw7cIdHH8eLwutG50LtGA@public.gmane.org> List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Sender: containers-bounces-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org Errors-To: containers-bounces-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org To: "Kirill A. Shutemov" Cc: containers-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org, linux-kernel-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, Paul Menage , "Paul E. McKenney" , Arjan van de Ven List-Id: containers.vger.kernel.org On Fri, 4 Feb 2011 15:34:39 +0200 "Kirill A. Shutemov" wrote: > On Thu, Feb 03, 2011 at 11:57:43AM -0800, Jacob Pan wrote: > > On Thu, 3 Feb 2011 10:12:51 -0800 > > Paul Menage wrote: > > > > > On Thu, Feb 3, 2011 at 9:51 AM, Jacob Pan > > > wrote: > > > > > > > > I think this logic defeats the purpose of having timer_slack > > > > subsystem in the first place. IMHO, the original intention was > > > > to have grouping effect of tasks in the cgroup. > > > > > > You can get the semantics you want by just setting min_slack_ns = > > > max_slack_ns. > > > > > true. it will just make set fail when min = max. it is awkward and > > counter intuitive when you want to change the group timer_slack. you > > will have to move both min and max to clamp the value, where set > > function can not be used. > > Interface is very similar to /sys/devices/system/cpu/cpuX/cpufreq. > I think it's sane. If you want some extention, you can do it with > userspace helper. > I don't disagree the current interface is usable. Just less intuitive. The situation is different for cpufreq, where you don't the situation of adding new entries to be adjusted in the existing max-min range. > > In addition, when a parent changes min = max, I don't see the > > current code enforce new settings on the children. Am i missing > > something? > > I've missed it. I'll fix. > > > In my use case, i want to put some apps into a managed group where > > relaxed slack value is used, but when time comes to move the app > > out of that cgroup, we would like to resore the original timer > > slack. I see having a current value per cgroup can be useful if we > > let timer code pick whether to use task slack value or the cgroup > > slack value. Or we have to cache the old value per task > > What's mean "original timer slack" if you are free to move a task > between a lot of cgroups and process itself free to change it anytime? > I need to manage tasks by a management software instead of letting the task change timer_slack by itself. The goal is to make management transparent and no modifications to the existing apps. Therefore, it is desirable to automatically enforce timer_slack when the apps are in the cgroup while automatically restore it when it is no longer under cgroup management. So the "original timer slack" can be the default 50us or whatever value chosen by the task itself. But the app itself should not care or even be aware of which cgroup it is in. So here are two optoins i can think of 1. add a new variable called cg_timer_slack_ns to struct task_struct{} cg_timer_slack_ns will be set by cgroup timer_slack subsystem, then we can retain the original per task value in timer_slack_ns. timer code will pick max(cg_timer_slack_ns, timer_slack_ns) if cg_timer_slack_ns is set. 2. leave task_struct unchanged, add a current_timer_slack to the cgroup. timer_slack cgroup does not modify per task timer_slack_ns. similar to option #1, let timer code pick the timer_slack to use based on whether the task is in timer_slack cgroup. Any thoughts? Thanks, Jacob