* KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> [2010-03-30 13:51:59]: > On Tue, 30 Mar 2010 13:06:48 +0900 > Daisuke Nishimura <nishimura@xxxxxxxxxxxxxxxxx> wrote: > > > On Tue, 30 Mar 2010 12:11:19 +0900, KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote: > > > On Tue, 30 Mar 2010 11:49:03 +0900 > > > Daisuke Nishimura <nishimura@xxxxxxxxxxxxxxxxx> wrote: > > > > > > > On Tue, 30 Mar 2010 11:23:01 +0900, KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote: > > > > > SHARED mapped file cache is not moved by patch [1/2] ??? > > > > > It sounds strange. > > > > > > > > > hmm, I'm sorry I'm not so good at user applications, but is it usual to use > > > > VM_SHARED file caches(!tmpfs) ? > > > > And is it better for us to move them only when page_mapcount() == 1 ? > > > > > > > > > > Considering shared library which has only one user, moving MAP_SHARED makes sense. > > > Unfortunately, there are people who creates their own shared library just for > > > their private dlopen() etc. (shared library for private use...) > > > > > > So, I think moving MAP_SHARED files makes sense. > > > IIRC, the libraries are loaded with MAP_PRIVATE and MAP_SHARED is not set. > > Thank you for your explanations. > > I'll update my patches to allow to move MAP_SHARED(but page_mapcount() == 1) > > file caches, and resend. > > > > Hmm, considering again...current summary is following...right ? > > - If page is an anon, it's not moved if page_mapcount() > 2. > - If page is a page cache, it's not moved if page_mapcount() > 2. > - If page is a shmem, it's not moved regardless of mapcount. > - If pte is swap, it's not moved refcnt > 2. > > I think following is straightforward and simple. > > - If page is an anon or swap of anon, it's not moved if referer > 2. What is referer in this context? The cgroup refering to the page? > (i.e. inherited from it's parent) > - If page is file,shmem or swap of shmem, it's moved regardless of referer. > But pages only under "from" memcg can be moved. > > I doubt adding too much speciality to shmem is not good. > Yep, I tend to agree, but I need to take a closer look again at the patches. -- Three Cheers, Balbir -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>