On (06/18/15 12:39), Minchan Kim wrote: [..] > > ah, I see. > > it doesn't hold the lock `until all the pages are done`. it holds it > > as long as zs_can_compact() returns > 0. hm, I'm not entirely sure that > > this patch set has increased the locking time (in average). > > I see your point. Sorry for the consusing. > My point is not average but max time. I bet your patch will increase > it and it will affect others who want to allocate zspage in parallel on > another CPU. makes sense. [..] > > > Yes, it's not easy and I believe a few artificial testing are not enough > > > to prove no regression but we don't have any choice. > > > Actually, I think this patchset does make sense. Although it might have > > > a problem on situation heavy memory pressure by lacking of fragment space, > > > > > > I tested exactly this scenario yesterday (and sent an email). We leave `no holes' > > in classes only in ~1.35% of cases. so, no, this argument is not valid. we preserve > > fragmentation. > > Thanks, Sergey. > > I want to test by myself to simulate worst case scenario to make to use up > reserved memory by zram. For it, please fix below first and resubmit, please. > > 1. doesn't hold lock until class compation is done. > It could prevent another allocation on another CPU. > I want to make worst case scenario and it needs it. > > 2. No touch ZS_ALMOST_FULL waterline. It can put more zspages > in ZS_ALMOST_FULL list so it couldn't be selected by migration > source. > > With new patchset, I want to watch min(free_pages of the system), > zram.max_used_pages, testing time and so on. > > Really sorry for bothering you, Sergey but I think it's important > feature on zram so I want to be careful because risk management is > my role. ok. will take a day or two to gather new numbers. -ss -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>