[PATCH 3/3] balancenuma: fix page locking in migrating thp

[PATCH 3/3] balancenuma: fix page locking in migrating thp

[Hillf Danton]

If fail to migrate thp(due to unsuccessful isolation), and if the original
pmd entry is changed after reaquiring page table lock, it is unsafe to
release the page lock as page maybe unstable.

It is fixed by raising extra page count before trying migration.

Signed-off-by: Hillf Danton <dhillf@gmail.com>
---

--- a/mm/huge_memory.c	Fri Nov 23 22:07:38 2012
+++ b/mm/huge_memory.c	Fri Nov 23 22:55:32 2012
@@ -1060,11 +1060,13 @@ int do_huge_pmd_numa_page(struct mm_stru
 	}
 	spin_unlock(&mm->page_table_lock);

+	get_page(page);
 	/* Migrate the THP to the requested node */
 	migrated = migrate_misplaced_transhuge_page(mm, vma,
 				pmdp, pmd, addr,
 				page, target_nid);
 	if (migrated) {
+		put_page(page);
 		count_vm_numa_event(NUMA_THP_MIGRATE_SUCCESS);
 		current_nid = target_nid;
 	} else {
@@ -1072,8 +1074,10 @@ int do_huge_pmd_numa_page(struct mm_stru
 		spin_lock(&mm->page_table_lock);
 		if (unlikely(!pmd_same(pmd, *pmdp))) {
 			unlock_page(page);
+			put_page(page);
 			goto out_unlock;
 		}
+		put_page(page);
 		goto clear_pmdnuma;
 	}

--

Re: [PATCH 3/3] balancenuma: fix page locking in migrating thp

[Mel Gorman]

On Sat, Nov 24, 2012 at 12:27:39PM +0800, Hillf Danton wrote:
> If fail to migrate thp(due to unsuccessful isolation), and if the original
> pmd entry is changed after reaquiring page table lock, it is unsafe to
> release the page lock as page maybe unstable.
> 
> It is fixed by raising extra page count before trying migration.
> 

Less sure of this one but the locking here is a mess. A reference count
is already taken. Isolating the page takes another reference count so the
first one can be dropped without the page being freed underneath us. The
migration function takes care of unlocking the page on a successful
migration. On unsuccessful migration, the PMD is rechecked under the
lock before clearing the pmd_numa. I'm not sure where you are seeing
the instability but I haven't fired up my brain either (it's the weekend
:)). I'll take a fresh look at this Monday.

Thanks!

-- 
Mel Gorman
SUSE Labs