On Thu, 22 Apr 2010 10:28:20 +0100 Mel Gorman <mel@xxxxxxxxx> wrote: > On Wed, Apr 21, 2010 at 10:46:45AM -0500, Christoph Lameter wrote: > > On Wed, 21 Apr 2010, Mel Gorman wrote: > > > > > > > 2. Is the BUG_ON check in > > > > > include/linux/swapops.h#migration_entry_to_page() now wrong? (I > > > > > think yes, but I'm not sure and I'm having trouble verifying it) > > > > > > > > The bug check ensures that migration entries only occur when the page > > > > is locked. This patch changes that behavior. This is going too oops > > > > therefore in unmap_and_move() when you try to remove the migration_ptes > > > > from an unlocked page. > > > > > > > > > > It's not unmap_and_move() that the problem is occurring on but during a > > > page fault - presumably in do_swap_page but I'm not 100% certain. > > > > remove_migration_pte() calls migration_entry_to_page(). So it must do that > > only if the page is still locked. > > > > Correct, but the other call path is > > do_swap_page > -> migration_entry_wait > -> migration_entry_to_page > > with migration_entry_wait expecting the page to be locked. There is a dangling > migration PTEs coming from somewhere. I thought it was from unmapped swapcache > first, but that cannot be the case. There is a race somewhere. > > > You need to ensure that the page is not unlocked in move_to_new_page() if > > the migration ptes are kept. > > > > move_to_new_page() only unlocks the new page not the original page. So that is safe. > > > > And it seems that the old page is also unlocked in unmap_and_move() only > > after the migration_ptes have been removed? So we are fine after all...? > > > > You'd think but migration PTEs are being left behind in some circumstance. I > thought it was due to this series, but it's unlikely. It's more a case that > compaction heavily exercises migration. > > We can clean up the old migration PTEs though when they are encountered > like in the following patch for example? I'll continue investigating why > this dangling migration pte exists as closing that race would be a > better fix. > > ==== CUT HERE ==== > mm,migration: Remove dangling migration ptes pointing to unlocked pages > > Due to some yet-to-be-identified race, it is possible for migration PTEs > to be left behind, When later paged-in, a BUG is triggered that assumes > that all migration PTEs are point to a page currently being migrated and > so must be locked. > > Rather than calling BUG, this patch notes the existance of dangling migration > PTEs in migration_entry_wait() and cleans them up. > I use similar patch for debugging. In my patch, this when this function founds dangling migration entry, return error code and do_swap_page() returns VM_FAULT_SIGBUS. 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. -Kame -- 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>