Messages in this thread | | | Date | Wed, 11 Nov 2009 11:36:15 +0000 (GMT) | From | Hugh Dickins <> | Subject | Re: [PATCH 2/6] mm: mlocking in try_to_unmap_one |
| |
On Wed, 11 Nov 2009, KOSAKI Motohiro wrote:
Though it doesn't quite answer your question, I'll just reinsert the last paragraph of my description here...
> > try_to_unmap_file()'s TTU_MUNLOCK nonlinear handling was particularly > > amusing: once unravelled, it turns out to have been choosing between > > two different ways of doing the same nothing. Ah, no, one way was > > actually returning SWAP_FAIL when it meant to return SWAP_SUCCESS.
... > > @@ -1081,45 +1053,23 @@ static int try_to_unmap_file(struct page ... > > > > - if (list_empty(&mapping->i_mmap_nonlinear)) > > + /* We don't bother to try to find the munlocked page in nonlinears */ > > + if (MLOCK_PAGES && TTU_ACTION(flags) == TTU_MUNLOCK) > > goto out; > > I have dumb question. > Does this shortcut exiting code makes any behavior change?
Not dumb. My intention was to make no behaviour change with any of this patch; but in checking back before completing the description, I suddenly realized that that shortcut intentionally avoids the
if (max_nl_size == 0) { /* all nonlinears locked or reserved ? */ ret = SWAP_FAIL; goto out; }
(which doesn't show up in the patch: you'll have to look at rmap.c), which used to have the effect of try_to_munlock() returning SWAP_FAIL in the case when there were one or more VM_NONLINEAR vmas of the file, but none of them (and none of the covering linear vmas) VM_LOCKED.
That should have been a SWAP_SUCCESS case, or with my changes another SWAP_AGAIN, either of which would make munlock_vma_page() count_vm_event(UNEVICTABLE_PGMUNLOCKED); which would be correct; but the SWAP_FAIL meant that count was not incremented in this case.
Actually, I've double-fixed that, because I also changed munlock_vma_page() to increment the count whenever ret != SWAP_MLOCK; which seemed more appropriate, but would have been a no-op if try_to_munlock() only returned SWAP_SUCCESS or SWAP_AGAIN or SWAP_MLOCK as it claimed.
But I wasn't very inclined to boast of fixing that bug, since my testing didn't give confidence that those /proc/vmstat unevictable_pgs_*lock* counts are being properly maintained anyway - when I locked the same pages in two vmas then unlocked them in both, I ended up with mlocked bigger than munlocked (with or without my 2/6 patch); which I suspect is wrong, but rather off my present course towards KSM swapping...
Hugh
| |