Hello, it's too late since I sent previos patch. https://lkml.org/lkml/2015/6/3/37 This patch is alomost new compared to previos approach. I think this is more simple, clear and easy to review. One thing I should notice is that I have tested this patch and couldn't find any critical problem so I rebased patchset onto recent mmotm(ie, mmotm-2015-10-15-15-20) to send formal patchset. Unfortunately, I start to see sudden discarding of the page we shouldn't do. IOW, application's valid anonymous page was disappeared suddenly. When I look through THP changes, I think we could lose dirty bit of pte between freeze_page and unfreeze_page when we mark it as migration entry and restore it. So, I added below simple code without enough considering and cannot see the problem any more. I hope it's good hint to find right fix this problem. diff --git a/mm/huge_memory.c b/mm/huge_memory.c index d5ea516ffb54..e881c04f5950 100644 --- a/mm/huge_memory.c +++ b/mm/huge_memory.c @@ -3138,6 +3138,9 @@ static void unfreeze_page_vma(struct vm_area_struct *vma, struct page *page, if (is_write_migration_entry(swp_entry)) entry = maybe_mkwrite(entry, vma); + if (PageDirty(page)) + SetPageDirty(page); + flush_dcache_page(page); set_pte_at(vma->vm_mm, address, pte + i, entry); Although it fixes abvove problem, I can encounter below another bug in several hours. BUG: Bad rss-counter state mm:ffff88007fc28000 idx:1 val:439 BUG: Bad rss-counter state mm:ffff88007fc28000 idx:2 val:73 Or BUG: Bad rss-counter state mm:ffff88007fc28000 idx:1 val:512 It seems we are zapping THP page without decreasing MM_ANONPAGES and MM_SWAPENTS. Of course, it could be a bug of MADV_FREE and recent changes of THP reveals it. What I can say is I couldn't see any problem until mmotm-2015-10-06-16-30 so I guess there is some conflict with THP-refcount redesign of Kirill or it makes to reveal MADV_FREE's hidden bug. I will hunt it down but I hope Kirill might catch it up earlier than me. Major thing with this patch is two things. 1. Work with MADV_FREE on PG_dirty page. So far, MADV_FREE doesn't work with page which is not in swap cache but has PG_dirty(ex, swapped-in page). Details are in [3/5]. 2. Make MADV_FREE discard path simple Current logic for discarding hinted page is really mess so [4/5] makes it simple and clean. 3. Fix with KSM page A process can have KSM page which is no dirty bit in page table entry and no PG_dirty in page->flags so VM could discard it wrongly. [5/5] fixes it. Minchan Kim (5): [1/5] mm: MADV_FREE trivial clean up [2/5] mm: skip huge zero page in MADV_FREE [3/5] mm: clear PG_dirty to mark page freeable [4/5] mm: simplify reclaim path for MADV_FREE [5/5] mm: mark stable page dirty in KSM include/linux/rmap.h | 6 +---- mm/huge_memory.c | 9 ++++---- mm/ksm.c | 12 ++++++++++ mm/madvise.c | 29 +++++++++++------------- mm/rmap.c | 46 +++++++------------------------------ mm/swap_state.c | 5 ++-- mm/vmscan.c | 64 ++++++++++++++++------------------------------------ 7 files changed, 60 insertions(+), 111 deletions(-) -- 1.9.1 -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>