[PATCH] ext4: Fix data races when using cached status extents

[PATCH] ext4: Fix data races when using cached status extents

[Jan Kara]

When using cached extent stored in extent status tree in tree->cache_es
another process holding ei->i_es_lock for reading can be racing with us
setting new value of tree->cache_es. If the compiler would decide to
refetch tree->cache_es at an unfortunate moment, it could result in a
bogus in_range() check. Fix the possible race by using READ_ONCE() when
using tree->cache_es only under ei->i_es_lock for reading.

Reported-by: syzbot+4a03518df1e31b537066@syzkaller.appspotmail.com
Link: https://lore.kernel.org/all/000000000000d3b33905fa0fd4a6@google.com
Suggested-by: Dmitry Vyukov <dvyukov@google.com>
Signed-off-by: Jan Kara <jack@suse.cz>
---
 fs/ext4/extents_status.c | 28 ++++++++++++----------------
 1 file changed, 12 insertions(+), 16 deletions(-)

diff --git a/fs/ext4/extents_status.c b/fs/ext4/extents_status.c
index 7bc221038c6c..ca2cb926894f 100644
--- a/fs/ext4/extents_status.c
+++ b/fs/ext4/extents_status.c
@@ -267,14 +267,12 @@ static void __es_find_extent_range(struct inode *inode,
 
 	/* see if the extent has been cached */
 	es->es_lblk = es->es_len = es->es_pblk = 0;
-	if (tree->cache_es) {
-		es1 = tree->cache_es;
-		if (in_range(lblk, es1->es_lblk, es1->es_len)) {
-			es_debug("%u cached by [%u/%u) %llu %x\n",
-				 lblk, es1->es_lblk, es1->es_len,
-				 ext4_es_pblock(es1), ext4_es_status(es1));
-			goto out;
-		}
+	es1 = READ_ONCE(tree->cache_es);
+	if (es1 && in_range(lblk, es1->es_lblk, es1->es_len)) {
+		es_debug("%u cached by [%u/%u) %llu %x\n",
+			 lblk, es1->es_lblk, es1->es_len,
+			 ext4_es_pblock(es1), ext4_es_status(es1));
+		goto out;
 	}
 
 	es1 = __es_tree_search(&tree->root, lblk);
@@ -931,14 +929,12 @@ int ext4_es_lookup_extent(struct inode *inode, ext4_lblk_t lblk,
 
 	/* find extent in cache firstly */
 	es->es_lblk = es->es_len = es->es_pblk = 0;
-	if (tree->cache_es) {
-		es1 = tree->cache_es;
-		if (in_range(lblk, es1->es_lblk, es1->es_len)) {
-			es_debug("%u cached by [%u/%u)\n",
-				 lblk, es1->es_lblk, es1->es_len);
-			found = 1;
-			goto out;
-		}
+	es1 = READ_ONCE(tree->cache_es);
+	if (es1 && in_range(lblk, es1->es_lblk, es1->es_len)) {
+		es_debug("%u cached by [%u/%u)\n",
+			 lblk, es1->es_lblk, es1->es_len);
+		found = 1;
+		goto out;
 	}
 
 	node = tree->root.rb_node;
-- 
2.35.3

Re: [PATCH] ext4: Fix data races when using cached status extents

[Theodore Ts'o]

On Wed, May 03, 2023 at 08:21:28PM +0200, Jan Kara wrote:
> When using cached extent stored in extent status tree in tree->cache_es
> another process holding ei->i_es_lock for reading can be racing with us
> setting new value of tree->cache_es. If the compiler would decide to
> refetch tree->cache_es at an unfortunate moment, it could result in a
> bogus in_range() check. Fix the possible race by using READ_ONCE() when
> using tree->cache_es only under ei->i_es_lock for reading.
> 
> Reported-by: syzbot+4a03518df1e31b537066@syzkaller.appspotmail.com
> Link: https://lore.kernel.org/all/000000000000d3b33905fa0fd4a6@google.com
> Suggested-by: Dmitry Vyukov <dvyukov@google.com>
> Signed-off-by: Jan Kara <jack@suse.cz>

Don't we also need a WRITE_ONCE here?

diff --git a/fs/ext4/extents_status.c b/fs/ext4/extents_status.c
index 7bc221038c6c..4694582cf255 100644
--- a/fs/ext4/extents_status.c
+++ b/fs/ext4/extents_status.c
@@ -293,7 +293,7 @@ static void __es_find_extent_range(struct inode *inode,
 	}
 
 	if (es1 && matching_fn(es1)) {
-		tree->cache_es = es1;
+		WRITE_ONCE(tree->cache_es, es1);
 		es->es_lblk = es1->es_lblk;
 		es->es_len = es1->es_len;
 		es->es_pblk = es1->es_pblk;

					- Ted

Re: [PATCH] ext4: Fix data races when using cached status extents

[Jan Kara]

On Wed 03-05-23 22:41:57, Theodore Ts'o wrote:
> On Wed, May 03, 2023 at 08:21:28PM +0200, Jan Kara wrote:
> > When using cached extent stored in extent status tree in tree->cache_es
> > another process holding ei->i_es_lock for reading can be racing with us
> > setting new value of tree->cache_es. If the compiler would decide to
> > refetch tree->cache_es at an unfortunate moment, it could result in a
> > bogus in_range() check. Fix the possible race by using READ_ONCE() when
> > using tree->cache_es only under ei->i_es_lock for reading.
> > 
> > Reported-by: syzbot+4a03518df1e31b537066@syzkaller.appspotmail.com
> > Link: https://lore.kernel.org/all/000000000000d3b33905fa0fd4a6@google.com
> > Suggested-by: Dmitry Vyukov <dvyukov@google.com>
> > Signed-off-by: Jan Kara <jack@suse.cz>
> 
> Don't we also need a WRITE_ONCE here?

Right, we should do that as well. I'll update the patch.

								Honza

> 
> diff --git a/fs/ext4/extents_status.c b/fs/ext4/extents_status.c
> index 7bc221038c6c..4694582cf255 100644
> --- a/fs/ext4/extents_status.c
> +++ b/fs/ext4/extents_status.c
> @@ -293,7 +293,7 @@ static void __es_find_extent_range(struct inode *inode,
>  	}
>  
>  	if (es1 && matching_fn(es1)) {
> -		tree->cache_es = es1;
> +		WRITE_ONCE(tree->cache_es, es1);
>  		es->es_lblk = es1->es_lblk;
>  		es->es_len = es1->es_len;
>  		es->es_pblk = es1->es_pblk;
> 
> 					- Ted
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR