bcachefs: add check for btree id against max in try read node

[PATCH v1] bcachefs: add check for btree id against max in try read node

Posted by Piotr Zalewski 1 year, 4 months ago

Add check for read node's btree_id against BTREE_ID_NR_MAX in
try_read_btree_node to prevent triggering EBUG_ON condition in
bch2_btree_id_root[1].

[1] https://syzkaller.appspot.com/bug?extid=cf7b2215b5d70600ec00

Reported-by: syzbot+cf7b2215b5d70600ec00@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=cf7b2215b5d70600ec00
Fixes: 4409b8081d16 ("bcachefs: Repair pass for scanning for btree nodes")
Signed-off-by: Piotr Zalewski <pZ010001011111@proton.me>
---
 fs/bcachefs/btree_node_scan.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/fs/bcachefs/btree_node_scan.c b/fs/bcachefs/btree_node_scan.c
index b28c649c6838..a24b4004e8f0 100644
--- a/fs/bcachefs/btree_node_scan.c
+++ b/fs/bcachefs/btree_node_scan.c
@@ -171,6 +171,9 @@ static void try_read_btree_node(struct find_btree_nodes *f, struct bch_dev *ca,
 	if (BTREE_NODE_LEVEL(bn) >= BTREE_MAX_DEPTH)
 		return;
 
+	if (BTREE_NODE_ID(bn) >= BTREE_ID_NR_MAX)
+		return;
+
 	rcu_read_lock();
 	struct found_btree_node n = {
 		.btree_id	= BTREE_NODE_ID(bn),
-- 
2.46.2

Re: [PATCH v1] bcachefs: add check for btree id against max in try read node

Posted by Piotr Zalewski 1 year, 4 months ago

> Add check for read node's btree_id against BTREE_ID_NR_MAX in
> try_read_btree_node to prevent triggering EBUG_ON condition in
> bch2_btree_id_root[1].

There seems to be some kind of (pro)regression regarding this issue. Now it
triggers a deadlock. I also found related bug detected by syzbot[2] for
which someone supplied the same patch recently and it triggered the same
deadlock. I will look into it (perhaps simple check isn't a proper fix 
here).

> [1] https://syzkaller.appspot.com/bug?extid=cf7b2215b5d70600ec00

[2] https://syzkaller.appspot.com/bug?extid=9f41e4b255897d99d4e9

> Reported-by: syzbot+cf7b2215b5d70600ec00@syzkaller.appspotmail.com
> Closes: https://syzkaller.appspot.com/bug?extid=cf7b2215b5d70600ec00
> Fixes: 4409b8081d16 ("bcachefs: Repair pass for scanning for btree nodes")
> Signed-off-by: Piotr Zalewski pZ010001011111@proton.me
> 
> ---
> fs/bcachefs/btree_node_scan.c | 3 +++
> 1 file changed, 3 insertions(+)
> 
> diff --git a/fs/bcachefs/btree_node_scan.c b/fs/bcachefs/btree_node_scan.c
> index b28c649c6838..a24b4004e8f0 100644
> --- a/fs/bcachefs/btree_node_scan.c
> +++ b/fs/bcachefs/btree_node_scan.c
> @@ -171,6 +171,9 @@ static void try_read_btree_node(struct find_btree_nodes *f, struct bch_dev *ca,
> if (BTREE_NODE_LEVEL(bn) >= BTREE_MAX_DEPTH)
> 
> return;
> 
> + if (BTREE_NODE_ID(bn) >= BTREE_ID_NR_MAX)
> 
> + return;
> +
> rcu_read_lock();
> struct found_btree_node n = {
> .btree_id = BTREE_NODE_ID(bn),
> --
> 2.46.2
>

Re: [PATCH v1] bcachefs: add check for btree id against max in try read node

Posted by Kent Overstreet 1 year, 4 months ago

On Mon, Oct 07, 2024 at 02:53:13AM GMT, Piotr Zalewski wrote:
> > Add check for read node's btree_id against BTREE_ID_NR_MAX in
> > try_read_btree_node to prevent triggering EBUG_ON condition in
> > bch2_btree_id_root[1].
> 
> There seems to be some kind of (pro)regression regarding this issue. Now it
> triggers a deadlock. I also found related bug detected by syzbot[2] for
> which someone supplied the same patch recently and it triggered the same
> deadlock. I will look into it (perhaps simple check isn't a proper fix 
> here).
> 
> > [1] https://syzkaller.appspot.com/bug?extid=cf7b2215b5d70600ec00
> 
> [2] https://syzkaller.appspot.com/bug?extid=9f41e4b255897d99d4e9

Your patch looks correct, [2] shouldn't be possible with it - and it
appears to only be reported on upstream commits. I'll apply your patch.

Re: [PATCH v1] bcachefs: add check for btree id against max in try read node

Posted by Piotr Zalewski 1 year, 4 months ago

On Monday, October 7th, 2024 at 5:05 AM, Kent Overstreet <kent.overstreet@linux.dev> wrote:

> On Mon, Oct 07, 2024 at 02:53:13AM GMT, Piotr Zalewski wrote:
> 
> > > Add check for read node's btree_id against BTREE_ID_NR_MAX in
> > > try_read_btree_node to prevent triggering EBUG_ON condition in
> > > bch2_btree_id_root[1].
> > 
> > There seems to be some kind of (pro)regression regarding this issue. Now it
> > triggers a deadlock. I also found related bug detected by syzbot[2] for
> > which someone supplied the same patch recently and it triggered the same
> > deadlock. I will look into it (perhaps simple check isn't a proper fix
> > here).
> > 
> > > [1] https://syzkaller.appspot.com/bug?extid=cf7b2215b5d70600ec00
> > 
> > [2] https://syzkaller.appspot.com/bug?extid=9f41e4b255897d99d4e9
> 
> 
> Your patch looks correct, [2] shouldn't be possible with it - and it
> appears to only be reported on upstream commits. I'll apply your patch.

Hi Kent,

Today I sat down to the other bug found by syzbot and retested it on
master, it triggers the same deadlock warning[3]. It seems like in
bch2_replicas_entry_validate mutex doesn't have to be always acquired? 
Also, I found the issue for specifically that warning[4]. The "count" is
high for the issue, is there some patch for that already pending? I 
couldn't find any.

[3] https://syzkaller.appspot.com/bug?extid=a1b59c8e1a3f022fd301
[4] https://syzkaller.appspot.com/bug?extid=4d24267b490e2b68a5fa

Best regards, Piotr Zalewski