From mboxrd@z Thu Jan 1 00:00:00 1970 From: Christoph Lameter Subject: Re: [PATCH v2 3/5] change number_of_cpusets to an atomic Date: Tue, 24 Apr 2012 13:27:09 -0500 (CDT) Message-ID: References: <1335209867-1831-1-git-send-email-glommer@parallels.com> <1335209867-1831-4-git-send-email-glommer@parallels.com> <4F96D1A8.6040604@parallels.com> <4F96D50D.4020804@parallels.com> Mime-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII Cc: Tejun Heo , netdev-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, cgroups-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, Li Zefan , kamezawa.hiroyu-+CUm20s59erQFUHtdCDX3A@public.gmane.org, David Miller , devel-GEFAQzZX7r8dnm+yROfE0A@public.gmane.org To: Glauber Costa Return-path: In-Reply-To: <4F96D50D.4020804-bzQdu9zFT3WakBO8gow8eQ@public.gmane.org> Sender: cgroups-owner-u79uwXL29TY76Z2rM5mHXA@public.gmane.org List-Id: netdev.vger.kernel.org On Tue, 24 Apr 2012, Glauber Costa wrote: > > > It doesn't seem to be the case here. > > > > How did you figure that? number_of_cpusets was introduced exactly because > > the functions are used in places where we do not pay the cost of calling > > __cpuset_node_allowed_soft/hardwall. Have a look at these. They may take > > locks etc etc in critical allocation paths > I am not arguing that. > > You want to avoid the cost of processing a function, that's fair. > (Note that by "function call cost" I don't mean the cost of processing a > function, but the cost of a (potentially empty) function call.) > The real question is: Are you okay with the cost of a branch + a global > variable (which is almost read only) fetch? No and that is why the static branching comes in. It takes away the global read of the number_of_cpusets variable in the critical paths. > The test of a global variable can - and do as of right now - avoid all the > expensive operations like locking, sleeping, etc, and if you don't need to > squeeze every nanosecond you can, they are often simpler - and therefore > better - than static branching. Better than static branching? This is in critical VM functions and reducing the cache footprint there is good for everyone. > Just to mention one point I am coming across these days - that initiated all > this: static patching holds the cpu_hotplug.lock. So it can't be called if you > hold any lock that has been already held under the cpu_hotplug.lock. This will > probably mean any lock the cpuset cgroup needs to take, because it is called - > and to do a lot of things - from the cpu hotplug handler, that holds the > cpu_hotplug.lock. Transitions from one to two cpusets are rare and are only done when a cpuset is created in the /dev/cpuset hierachy). You could move the code modification outside of locks or defer action into an event thread if there are locks in the way.