Messages in this thread | | | Subject | Re: CPU Hotplug rework | From | Peter Zijlstra <> | Date | Mon, 26 Mar 2012 15:38:18 +0200 |
| |
On Mon, 2012-03-26 at 09:09 -0400, Steven Rostedt wrote: > On Mon, 2012-03-26 at 10:02 +0200, Peter Zijlstra wrote: > > On Mon, 2012-03-26 at 11:11 +1030, Rusty Russell wrote: > > > Obviously, having callbacks hanging around until the CPU comes back is > > > not viable, nor is blocking preempt during the callbacks. Calling > > > get_online_cpus() is too heavy. > > > > -rt has patches (albeit somewhat ugly) to make get_online_cpus() a br > > style rw-lock. > > Unfortunately, -rt still has broken cpu hotplug.
This is still the issue where we take the hotplug lock for write, but then rely on other threads to complete -- say CPU_DOWN_PREPARE notifiers doing flush_workqueue(), and the worker kthread needing to acquire the hotplug lock for read, right?
Yes, that is bothersome, but doesn't hinder the work required to get get_online_cpus() usably fast, right?
| |