On 26.11.18 02:44, Wei Yang wrote: > On Fri, Nov 23, 2018 at 09:46:52AM +0100, David Hildenbrand wrote: >> On 23.11.18 09:42, Michal Hocko wrote: >>> On Thu 22-11-18 16:26:40, David Hildenbrand wrote: >>>> On 22.11.18 11:12, Wei Yang wrote: >>>>> During online_pages phase, pgdat->nr_zones will be updated in case this >>>>> zone is empty. >>>>> >>>>> Currently the online_pages phase is protected by the global lock >>>>> mem_hotplug_begin(), which ensures there is no contention during the >>>>> update of nr_zones. But this global lock introduces scalability issues. >>>>> >>>>> This patch is a preparation for removing the global lock during >>>>> online_pages phase. Also this patch changes the documentation of >>>>> node_size_lock to include the protectioin of nr_zones. >>>> >>>> I looked into locking recently, and there is more to it. >>>> >>>> Please read: >>>> >>>> commit dee6da22efac451d361f5224a60be2796d847b51 >>>> Author: David Hildenbrand <david@xxxxxxxxxx> >>>> Date: Tue Oct 30 15:10:44 2018 -0700 >>>> >>>> memory-hotplug.rst: add some details about locking internals >>>> >>>> Let's document the magic a bit, especially why device_hotplug_lock is >>>> required when adding/removing memory and how it all play together with >>>> requests to online/offline memory from user space. >>>> >>>> Short summary: Onlining/offlining of memory requires the device_hotplug_lock >>>> as of now. >>> >>> Well, I would tend to disagree here. You might be describing the current >>> state of art but the device_hotplug_lock doesn't make much sense for the >>> memory hotplug in principle. There is absolutely nothing in the core MM >> >> There are collisions with CPU hotplug that require this lock (when nodes >> come and go as far as I remember). And there is the problematic lock >> inversion that can happen when adding/remving memory. This all has to be >> sorted out, we'll have to see if we really need it for >> onlining/offlining, though, however ... >> > > Seems I get a little understanding on this part. > > There are two hotplug: > * CPU hotplug > * Memory hotplug. > > There are two phase for Memory hotplug: > * physical add/remove > * logical online/offline > > All of them are protected by device_hotplug_lock now, so we need to be > careful to release this in any case. Is my understanding correct? Yes, e.g. the acpi driver always held the device_hotplug_lock (due to possible problems with concurrent cpu/memory hot(un)plug). Onlining offlining of devices (including cpu/memory) from user space always held the device_hotplug_lock. So this part was executed sequentially for a long time. I recently made sure that any adding/removing/onlining/offlining correctly grabs the device_hotplug_lock AND the mem_hotplug_lock in any case (because it was inconsistent and broken), so it is all executed sequentially. So when getting rid of mem_hotplug_lock we only have to care about all users that don't take the device_hotplug_lock. > >>> that would require this lock. The current state just uses a BKL in some >>> sense and we really want to get rid of that longterm. This patch is a tiny >>> step in that direction and I suspect many more will need to come on the >>> way. We really want to end up with a clear scope of each lock being >>> taken. A project for a brave soul... >> >> ... for now I don't consider "optimize for parallel >> onlining/offlining/adding/removing of memory and cpus" really necessary. >> What is necessary indeed is to not slowdown the whole system just >> because some memory is coming/going. Therefore I agree, this patch is a >> step into the right direction. >> > > Agree. > > The target is to accelerate the hot-plug without slow down the normal > process. Indeed! -- Thanks, David / dhildenb
