On Mon, Jul 29, 2019 at 10:35:15AM +0200, Michal Hocko wrote: > On Mon 29-07-19 17:20:52, Minchan Kim wrote: > > On Mon, Jul 29, 2019 at 09:45:23AM +0200, Michal Hocko wrote: > > > On Mon 29-07-19 16:10:37, Minchan Kim wrote: > > > > In our testing(carmera recording), Miguel and Wei found unmap_page_range > > > > takes above 6ms with preemption disabled easily. When I see that, the > > > > reason is it holds page table spinlock during entire 512 page operation > > > > in a PMD. 6.2ms is never trivial for user experince if RT task couldn't > > > > run in the time because it could make frame drop or glitch audio problem. > > > > > > Where is the time spent during the tear down? 512 pages doesn't sound > > > like a lot to tear down. Is it the TLB flushing? > > > > Miguel confirmed there is no such big latency without mark_page_accessed > > in zap_pte_range so I guess it's the contention of LRU lock as well as > > heavy activate_page overhead which is not trivial, either. > > Please give us more details ideally with some numbers. I had a time to benchmark it via adding some trace_printk hooks between pte_offset_map_lock and pte_unmap_unlock in zap_pte_range. The testing device is 2018 premium mobile device. I can get 2ms delay rather easily to release 2M(ie, 512 pages) when the task runs on little core even though it doesn't have any IPI and LRU lock contention. It's already too heavy. If I remove activate_page, 35-40% overhead of zap_pte_range is gone so most of overhead(about 0.7ms) comes from activate_page via mark_page_accessed. Thus, if there are LRU contention, that 0.7ms could accumulate up to several ms.
