Re: [PATCH] mm: do not access page->mapping directly on page_endio

On Wed, Feb 22, 2017 at 03:53:16PM +0100, Michal Hocko wrote:

On Wed 22-02-17 23:35:17, Minchan Kim wrote:

quoted

On Wed, Feb 22, 2017 at 01:11:00PM +0100, Michal Hocko wrote:

quoted

On Wed 22-02-17 14:39:24, Minchan Kim wrote:

quoted

With rw_page, page_endio is used for completing IO on a page
and it propagates write error to the address space if the IO
fails. The problem is it accesses page->mapping directly which
might be okay for file-backed pages but it shouldn't for
anonymous page. Otherwise, it can corrupt one of field from
anon_vma under us and system goes panic randomly.

I was about to say that anonymous pages shouldn't hit that path because
the end_swap_bio_write doesn call page_endio. But then I've noticed that

No. For driver to support rw_page, every swap_writepage calls rw_page.

swap_writepage
  bdev_writepage
    ops->rw_page

Ohh, you are right, I have missed this option. I was looking at the
normal swapout path which uses bio.

quoted

quoted

zram does call this function. On a closer look, though, it doesn't seem
to call it with err != 0 so it cannot hit this path. So I am wondering
whether this actually fixes anything. Why it has been marked for stable?

Look at other drivers to support rw_page, not zram, esp, brd.
They can be used for swap device and then can hit the case.

In fact, I encountered the BUG during zram development(i.e., it doesn't
land to upstream) and it was really hard to figure it out because it made
random crash, sometime mmap_sem lockdep, sometime other places where
places never related to zram/zsmalloc, sometime not reproducible.

When I consider how that bug is subtle and people do fast-swap test as brd,
it's worth to add stable mark, I think.

Sure, could you add this to the changelog. Along with Fixes tag? I
suspect it is dd6bd0d9c7db ("swap: use bdev_read_page() /
bdev_write_page()") which has introduced this but I didn't look too
close. The patch is trivially correct.

Sure. Thanks for the review.

Andrew, Could you change description with this?

From 9efb87a873db67a9e6ebf44fdabf7d05fe4b4e21 Mon Sep 17 00:00:00 2001

From: Minchan Kim <minchan@kernel.org>
Date: Fri, 4 Nov 2016 09:12:39 +0900
Subject: [PATCH v2] mm: do not access page->mapping directly on page_endio

With rw_page, page_endio is used for completing IO on a page
and it propagates write error to the address space if the IO
fails. The problem is it accesses page->mapping directly which
might be okay for file-backed pages but it shouldn't for
anonymous page. Otherwise, it can corrupt one of field from
anon_vma under us and system goes panic randomly.

swap_writepage
  bdev_writepage
    ops->rw_page

I encountered the BUG during developing new zram feature and
it was really hard to figure it out because it made random
crash, somtime mmap_sem lockdep, sometime other places where
places never related to zram/zsmalloc, and not reproducible
with some configuration.

When I consider how that bug is subtle and people do fast-swap
test with brd, it's worth to add stable mark, I think.

Fixes: dd6bd0d9c7db ("swap: use bdev_read_page() / bdev_write_page()")
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: <redacted>
Signed-off-by: Minchan Kim <minchan@kernel.org>
---
* from v1
  * add more detailed description with Fix tag - Michal

 mm/filemap.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/mm/filemap.c b/mm/filemap.c
index 2ba46f410c7c..1944c631e3e6 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c

@@ -1008,9 +1008,12 @@ void page_endio(struct page *page, bool is_write, int err)
 		unlock_page(page);
 	} else {
 		if (err) {
+			struct address_space *mapping;
+
 			SetPageError(page);
-			if (page->mapping)
-				mapping_set_error(page->mapping, err);
+			mapping = page_mapping(page);
+			if (mapping)
+				mapping_set_error(mapping, err);
 		}
 		end_page_writeback(page);
 	}

-- 
2.7.4