xfrm: Use spin_lock_bh() in xfrm_input()

[PATCH] xfrm: Use spin_lock_bh() in xfrm_input()

Posted by Zhang Yiqun 2 years ago

This patch is to change spin_lock() into spin_lock_bh(), which can
disable bottem half in calling. If we leave this as spin_lock(),
it may stuck in a deadlock, because the callback in bottem half in
crypto driver will also call xfrm_input() again.

Signed-off-by: Zhang Yiqun <zhangyiqun@phytium.com.cn>
---
 net/xfrm/xfrm_input.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/net/xfrm/xfrm_input.c b/net/xfrm/xfrm_input.c
index bd4ce21d76d7..f4cd46d73b1e 100644
--- a/net/xfrm/xfrm_input.c
+++ b/net/xfrm/xfrm_input.c
@@ -581,7 +581,7 @@ int xfrm_input(struct sk_buff *skb, int nexthdr, __be32 spi, int encap_type)
 		}
 
 lock:
-		spin_lock(&x->lock);
+		spin_lock_bh(&x->lock);
 
 		if (unlikely(x->km.state != XFRM_STATE_VALID)) {
 			if (x->km.state == XFRM_STATE_ACQ)
@@ -607,7 +607,7 @@ int xfrm_input(struct sk_buff *skb, int nexthdr, __be32 spi, int encap_type)
 			goto drop_unlock;
 		}
 
-		spin_unlock(&x->lock);
+		spin_unlock_bh(&x->lock);
 
 		if (xfrm_tunnel_check(skb, x, family)) {
 			XFRM_INC_STATS(net, LINUX_MIB_XFRMINSTATEMODEERROR);
-- 
2.17.1

Re: [PATCH] xfrm: Use spin_lock_bh() in xfrm_input()

Posted by Eric Dumazet 2 years ago

On Mon, Dec 18, 2023 at 9:43 AM Zhang Yiqun <zhangyiqun@phytium.com.cn> wrote:
>
> This patch is to change spin_lock() into spin_lock_bh(), which can
> disable bottem half in calling. If we leave this as spin_lock(),
> it may stuck in a deadlock, because the callback in bottem half in
> crypto driver will also call xfrm_input() again.
>
> Signed-off-by: Zhang Yiqun <zhangyiqun@phytium.com.cn>

When was the bug added ?
We need a FIxes: tag.

Also a stack trace to show the deadlock (or lockdep complaint ) would
be needed as well.

> ---
>  net/xfrm/xfrm_input.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/net/xfrm/xfrm_input.c b/net/xfrm/xfrm_input.c
> index bd4ce21d76d7..f4cd46d73b1e 100644
> --- a/net/xfrm/xfrm_input.c
> +++ b/net/xfrm/xfrm_input.c
> @@ -581,7 +581,7 @@ int xfrm_input(struct sk_buff *skb, int nexthdr, __be32 spi, int encap_type)
>                 }
>
>  lock:
> -               spin_lock(&x->lock);
> +               spin_lock_bh(&x->lock);
>
>                 if (unlikely(x->km.state != XFRM_STATE_VALID)) {
>                         if (x->km.state == XFRM_STATE_ACQ)
> @@ -607,7 +607,7 @@ int xfrm_input(struct sk_buff *skb, int nexthdr, __be32 spi, int encap_type)
>                         goto drop_unlock;
>                 }
>
> -               spin_unlock(&x->lock);
> +               spin_unlock_bh(&x->lock);
>
>                 if (xfrm_tunnel_check(skb, x, family)) {
>                         XFRM_INC_STATS(net, LINUX_MIB_XFRMINSTATEMODEERROR);

This patch is not correct anyway.
There are five places in xfrm_input() where x->lock is either locked
or unlocked.

Please tell us how this was tested.
Thanks.

Re: [PATCH] xfrm: Use spin_lock_bh() in xfrm_input()

Posted by Steffen Klassert 2 years ago

On Mon, Dec 18, 2023 at 12:07:50PM +0100, Eric Dumazet wrote:
> On Mon, Dec 18, 2023 at 9:43 AM Zhang Yiqun <zhangyiqun@phytium.com.cn> wrote:
> >
> > This patch is to change spin_lock() into spin_lock_bh(), which can
> > disable bottem half in calling. If we leave this as spin_lock(),
> > it may stuck in a deadlock, because the callback in bottem half in
> > crypto driver will also call xfrm_input() again.
> >
> > Signed-off-by: Zhang Yiqun <zhangyiqun@phytium.com.cn>
> 
> When was the bug added ?
> We need a FIxes: tag.

This looks more like a 'crypto driver' bug. xfrm_input() runs in
the RX path and therefore expects to run with BHs off.