mm/rmap: ensure the validity of mapping vma which referenced an anon page
Subsystems (2)
PATCH 1/1 mm/rmap: ensure the validity of mapping vma which referenced an anon page
From: chenqiwu <[email protected]>
When finding all the mapping vmas for an anon page by anon_vma, there is a panic risk that one mapping vma or its vm_mm has been released by someone. Like the following crash during kswapd reclaiming pages:
Unable to handle kernel NULL pointer dereference at virtual address
00000048
PC is at page_vma_mapped_walk+0x54/0x16c
LR is at page_referenced_one+0x44/0x140
[......]
CPU: 1 PID: 161 Comm: kswapd0 Call trace: [<ffffff9a080832d0>] el1_da+0x24/0x3c [<ffffff9a0823ea18>] page_vma_mapped_walk+0x54/0x16c [<ffffff9a0823fd8c>] page_referenced_one+0x44/0x140 [<ffffff9a08240ef0>] rmap_walk_anon+0x124/0x168 [<ffffff9a0823fcfc>] page_referenced+0x144/0x190 [<ffffff9a08213e6c>] shrink_active_list+0x25c/0x478 [<ffffff9a082126c8>] kswapd+0x7b0/0x9c8 [<ffffff9a080daffc>] kthread+0x154/0x18c [<ffffff9a0808563c>] ret_from_fork+0x10/0x18
The PC is pointed to the following code line:
bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw)
{
struct mm_struct *mm = pvmw->vma->vm_mm;
......
pgd = pgd_offset(mm, pvmw->address); //PC
......
}Because the current pvmw->vma->vm_mm is a kernel NULL pointer, which causing crash when pgd_offset() dereferences the mm pointer.
This patch fixes the problem by ensuring that both the mapping vma and its vm_mm are valid. If not, we just continue to traverse the anon_vma->rb_root to avoid the potential junk pointer dereference.
--- mm/rmap.c | 3 +++ 1 file changed, 3 insertions(+)
mm/rmap.c +3 -0
@@ -1863,6 +1863,9 @@ static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,if (rwc->invalid_vma && rwc->invalid_vma (vma, rwc->arg))continue ; +if (!vma && !vma->vm_mm) +continue ; +if (!rwc->rmap_one (page, vma, address, rwc->arg))break ;if (rwc->done && rwc->done (page))
Signature · 1 line
1.9.1
Quoted 5 lines · On Wed, Mar 11, 2020 at 03:45:13PM +0800, [email protected] wrote:
From: chenqiwu <[email protected]>
When finding all the mapping vmas for an anon page by
anon_vma, there is a panic risk that one mapping vma or itsvm_mmhas been released by someone.
What? Who would be able to release the VMA or the mm_struct? We hold anon_vma lock. It doesn't make sense to me. Something else is broken.
Quoted 45 lines
Like the following crash during kswapd reclaiming pages:
Unable to handle kernel NULL pointer dereference at virtual address00000048
PC is atpage_vma_mapped_walk+0x54/0x16c
LR is atpage_referenced_one+0x44/0x140
[......]CPU: 1 PID: 161 Comm: kswapd0 Call trace: [<ffffff9a080832d0>] el1_da+0x24/0x3c [<ffffff9a0823ea18>] page_vma_mapped_walk+0x54/0x16c [<ffffff9a0823fd8c>] page_referenced_one+0x44/0x140 [<ffffff9a08240ef0>] rmap_walk_anon+0x124/0x168 [<ffffff9a0823fcfc>] page_referenced+0x144/0x190 [<ffffff9a08213e6c>] shrink_active_list+0x25c/0x478 [<ffffff9a082126c8>] kswapd+0x7b0/0x9c8 [<ffffff9a080daffc>] kthread+0x154/0x18c [<ffffff9a0808563c>] ret_from_fork+0x10/0x18The PC is pointed to the following code line:
boolpage_vma_mapped_walk(struct page_vma_mapped_walk *pvmw){ struct mm_struct *mm = pvmw->vma->vm_mm; ...... pgd = pgd_offset(mm, pvmw->address); //PC ...... }Because the current
pvmw->vma->vm_mmis a kernel NULL pointer, which causing crash whenpgd_offset()dereferences the mm pointer.This patch fixes the problem by ensuring that both the mapping vma and its
vm_mmare valid. If not, we just continue to traverse theanon_vma->rb_rootto avoid the potential junk pointer dereference.Signed-off-by chenqiwu <[email protected]>--- mm/rmap.c | 3 +++ 1 file changed, 3 insertions(+)mm/rmap.c +0
-0
mm/rmap.c static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc, · line 1868
Quoted 6 lines
mm/rmap.c +3
-0@@ -1863,6 +1863,9 @@ static void rmap_walk_anon(struct page *page, struct rmap_walk_control *rwc,if (rwc->invalid_vma && rwc->invalid_vma (vma, rwc->arg))continue ; +if (!vma && !vma->vm_mm) +continue ; +
Even if the premis of the patch is true, 'vma' got used twice in the loop iteration before you check if it's non-NULL.
Quoted 7 lines
mm/rmap.c +0
-0if (!rwc->rmap_one (page, vma, address, rwc->arg))break ;if (rwc->done && rwc->done (page))Signature · 1 line
1.9.1
Signature · 1 line
Kirill A. Shutemov
Quoted 10 lines · Kirill A. Shutemov
Quoted 5 lines · On Wed, Mar 11, 2020 at 03:45:13PM +0800, [email protected] wrote:
From: chenqiwu <[email protected]>
When finding all the mapping vmas for an anon page by
anon_vma, there is a panic risk that one mapping vma or itsvm_mmhas been released by someone.What? Who would be able to release the VMA or the
mm_struct? We holdanon_vmalock. It doesn't make sense to me. Something else is broken.
There is only log (no kdump) in my hand. It's hard to say who released the VMA's mm_struct before we hold the anon_vma lock.