On Thu, Apr 02, 2020 at 02:50:28PM +0900, Joonsoo Kim wrote: > 2020년 3월 24일 (화) 오후 3:25, Joonsoo Kim <js1304@xxxxxxxxx>님이 작성: > > The problem looks not that easy. Hmm... > > > > In current code, there is a large time gap between the shadow entries > > are poped up and the page is charged to the memcg, especially, > > for readahead-ed pages. We cannot maintain the shadow entries of > > the readahead-ed pages until the pages are charged. > > > > My plan to solve this problem is propagating the charged mm to > > __read_swap_cache_async(), like as file cache, charging when > > the page is added on to the swap cache and calling workingset_refault() > > there. Charging will only occur if: > > > > 1. faulted page > > 2. readahead-ed page with the shadow entry for the same memcg > > > > Also, readahead only happens when shadow entry's memcg is the same > > with the charged memcg. If not the same, it's mostly not ours so > > readahead isn't needed. > > > > Please let me know how you think of the feasibility of this idea. > > Hello, Johannes. > > Could you let me know your opinion about the idea above? > In fact, since your reply is delayed, I completed the solution about the > above idea. If you want, I will submit it first. Then, we could discuss > the solution more easily. It's probably easiest if you send out your implementation and we discuss it over the code. Thanks!
