On Thu, Apr 22, 2010 at 11:14 PM, Mel Gorman <mel@xxxxxxxxx> wrote: > On Thu, Apr 22, 2010 at 07:51:53PM +0900, KAMEZAWA Hiroyuki wrote: >> On Thu, 22 Apr 2010 19:31:06 +0900 >> KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote: >> >> > On Thu, 22 Apr 2010 19:13:12 +0900 >> > Minchan Kim <minchan.kim@xxxxxxxxx> wrote: >> > >> > > On Thu, Apr 22, 2010 at 6:46 PM, KAMEZAWA Hiroyuki >> > > <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote: >> > >> > > > Hmm..in my test, the case was. >> > > > >> > > > Before try_to_unmap: >> > > > mapcount=1, SwapCache, remap_swapcache=1 >> > > > After remap >> > > > mapcount=0, SwapCache, rc=0. >> > > > >> > > > So, I think there may be some race in rmap_walk() and vma handling or >> > > > anon_vma handling. migration_entry isn't found by rmap_walk. >> > > > >> > > > Hmm..it seems this kind patch will be required for debug. >> > > >> >> Ok, here is my patch for _fix_. But still testing... >> Running well at least for 30 minutes, where I can see bug in 10minutes. >> But this patch is too naive. please think about something better fix. >> >> == >> From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> >> >> At adjust_vma(), vma's start address and pgoff is updated under >> write lock of mmap_sem. This means the vma's rmap information >> update is atoimic only under read lock of mmap_sem. >> >> >> Even if it's not atomic, in usual case, try_to_ummap() etc... >> just fails to decrease mapcount to be 0. no problem. >> >> But at page migration's rmap_walk(), it requires to know all >> migration_entry in page tables and recover mapcount. >> >> So, this race in vma's address is critical. When rmap_walk meet >> the race, rmap_walk will mistakenly get -EFAULT and don't call >> rmap_one(). This patch adds a lock for vma's rmap information. >> But, this is _very slow_. > > Ok wow. That is exceptionally well-spotted. This looks like a proper bug > that compaction exposes as opposed to a bug that compaction introduces. > >> We need something sophisitcated, light-weight update for this.. >> > > In the event the VMA is backed by a file, the mapping i_mmap_lock is taken for > the duration of the update and is taken elsewhere where the VMA information > is read such as rmap_walk_file() > > In the event the VMA is anon, vma_adjust currently talks no locks and your > patch introduces a new one but why not use the anon_vma lock here? Am I > missing something that requires the new lock? rmap_walk_anon doesn't hold vma's anon_vma->lock. It holds page->anon_vma->lock. -- Kind regards, Minchan Kim -- 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