Messages in this thread | | | Date | Tue, 8 Dec 2009 17:04:19 -0800 | From | Chris Wright <> | Subject | Re: [PATCH 2/9] ksm: let shared pages be swappable |
| |
* KAMEZAWA Hiroyuki (kamezawa.hiroyu@jp.fujitsu.com) wrote: > On Fri, 4 Dec 2009 09:16:40 -0800 > Chris Wright <chrisw@redhat.com> wrote: > > * KAMEZAWA Hiroyuki (kamezawa.hiroyu@jp.fujitsu.com) wrote: > > > Hmm, can't we use ZERO_PAGE we have now ? > > > If do so, > > > - no mapcount check > > > - never on LRU > > > - don't have to maintain shared information because ZERO_PAGE itself has > > > copy-on-write nature. > > > > It's a somewhat special case, but wouldn't it be useful to have a generic > > method to recognize this kind of sharing since it's a generic issue? > > I just remembered that why ZERO_PAGE was removed (in past). It was becasue > cache-line ping-pong at fork beacause of page->mapcount. And KSM introduces > zero-pages which have mapcount again. If no problems in realitsitc usage of > KVM, ignore me.
KVM is not exactly fork heavy (although it's not the only possible user of KSM). And the CoW path has fault + copy already.
Semi-related...it can make good sense to make the KSM trees per NUMA node. Would mean things like page of zeroes would collapse to number of NUMA nodes pages rather than a single page, but has the benefit of not adding remote access (although, probably more useful for text pages than zero pages).
thanks, -chris
| |