On Wed 12-09-18 20:12:30, Arun KS wrote: > On 2018-09-12 18:47, Michal Hocko wrote: > > On Wed 12-09-18 22:57:43, Balbir Singh wrote: > > > On Wed, Sep 12, 2018 at 12:38:53PM +0200, Michal Hocko wrote: > > > > On Wed 12-09-18 14:56:45, Arun KS wrote: > > > > > When free pages are done with pageblock_order, time spend on > > > > > coalescing pages by buddy allocator can be reduced. With > > > > > section size of 256MB, hot add latency of a single section > > > > > shows improvement from 50-60 ms to less than 1 ms, hence > > > > > improving the hot add latency by 60%. > > > > > > > > Where does the improvement come from? You are still doing the same > > > > amount of work except that the number of callbacks is lower. Is this the > > > > real source of 60% improvement? > > > > > > > > > > It looks like only the first page of the pageblock is initialized, is > > > some of the cost amortized in terms of doing one initialization for > > > the page with order (order) and then relying on split_page and helpers > > > to do the rest? Of course the number of callbacks reduce by a > > > significant > > > number as well. > > > > Ohh, I have missed that part. Now when re-reading I can see the reason > > for the perf improvement. It is most likely the higher order free which > > ends up being much cheaper. This part makes some sense. > > > > How much is this feasible is another question. Do not forget we have > > those external providers of the online callback and those would need to > > be updated as well. > Sure Michal, I ll look into this. > > > > > Btw. the normal memmap init code path does the same per-page free as > > well. If we really want to speed the hotplug path then I guess the init > > one would see a bigger improvement and those two should be in sync. > Thanks for pointers, Will look further. I haven't looked closer and I will be travelling next week so just hint. Have a look at the nobootmem and how it frees pages to the page allocator in __free_pages_boot_core. Seems exactly what you want and it also answers your question about reference counting. -- Michal Hocko SUSE Labs
