lkml.org 
[lkml]   [2015]   [Oct]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [PATCH 0/5] MADV_FREE refactoring and fix KSM page
On Mon, Oct 19, 2015 at 07:01:50PM +0900, Minchan Kim wrote:
> On Mon, Oct 19, 2015 at 03:31:42PM +0900, Minchan Kim wrote:
> > 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);
>
> The condition of PageDirty was typo. I didn't add the condition.
> Just added.
>
> SetPageDirty(page);

I reviewed THP refcount redesign patch and It seems below patch fixes
MADV_FREE problem. It works well for hours.

From 104a0940b4c0f97e61de9fee0fd602926ff28312 Mon Sep 17 00:00:00 2001
From: Minchan Kim <minchan@kernel.org>
Date: Tue, 20 Oct 2015 16:00:52 +0900
Subject: [PATCH] mm: mark head page dirty in split_huge_page

In thp split in old THP refcount, we mappped all of pages
(ie, head + tails) to pte_mkdirty and mark PG_flags to every
tail pages.

But with THP refcount redesign, we can lose dirty bit in page table
and PG_dirty for head page if we want to free the THP page using
migration_entry.

It ends up discarding head page by madvise_free suddenly.
This patch fixes it by mark the head page PG_dirty when VM splits
the THP page.

Signed-off-by: Minchan Kim <minchan@kernel.org>
---
mm/huge_memory.c | 1 +
1 file changed, 1 insertion(+)

diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index adccfb48ce57..7fbbd42554a1 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -3258,6 +3258,7 @@ static void __split_huge_page(struct page *page, struct list_head *list)
atomic_sub(tail_mapcount, &head->_count);

ClearPageCompound(head);
+ SetPageDirty(head);
spin_unlock_irq(&zone->lru_lock);

unfreeze_page(page_anon_vma(head), head);
--
1.9.1


\
 
 \ /
  Last update: 2015-10-20 09:41    [W:0.115 / U:0.076 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site