[PATCH v2] skbuff: fix a data race in skb_queue_len()

[PATCH v2] skbuff: fix a data race in skb_queue_len()

[Qian Cai]

sk_buff.qlen can be accessed concurrently as noticed by KCSAN,

 BUG: KCSAN: data-race in __skb_try_recv_from_queue / unix_dgram_sendmsg

 read to 0xffff8a1b1d8a81c0 of 4 bytes by task 5371 on cpu 96:
  unix_dgram_sendmsg+0x9a9/0xb70 include/linux/skbuff.h:1821
				 net/unix/af_unix.c:1761
  ____sys_sendmsg+0x33e/0x370
  ___sys_sendmsg+0xa6/0xf0
  __sys_sendmsg+0x69/0xf0
  __x64_sys_sendmsg+0x51/0x70
  do_syscall_64+0x91/0xb47
  entry_SYSCALL_64_after_hwframe+0x49/0xbe

 write to 0xffff8a1b1d8a81c0 of 4 bytes by task 1 on cpu 99:
  __skb_try_recv_from_queue+0x327/0x410 include/linux/skbuff.h:2029
  __skb_try_recv_datagram+0xbe/0x220
  unix_dgram_recvmsg+0xee/0x850
  ____sys_recvmsg+0x1fb/0x210
  ___sys_recvmsg+0xa2/0xf0
  __sys_recvmsg+0x66/0xf0
  __x64_sys_recvmsg+0x51/0x70
  do_syscall_64+0x91/0xb47
  entry_SYSCALL_64_after_hwframe+0x49/0xbe

Since only the read is operating as lockless, it could introduce a logic
bug in unix_recvq_full() due to the load tearing. Fix it by adding
a lockless variant of skb_queue_len() and unix_recvq_full() where
READ_ONCE() is on the read while WRITE_ONCE() is on the write similar to
the commit d7d16a89350a ("net: add skb_queue_empty_lockless()").

Signed-off-by: Qian Cai <cai@lca.pw>
---

v2: add lockless variant helpers and WRITE_ONCE().

 include/linux/skbuff.h | 14 +++++++++++++-
 net/unix/af_unix.c     |  9 ++++++++-
 2 files changed, 21 insertions(+), 2 deletions(-)

diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
index 3d13a4b717e9..de5eade20e52 100644
--- a/include/linux/skbuff.h
+++ b/include/linux/skbuff.h
@@ -1822,6 +1822,18 @@ static inline __u32 skb_queue_len(const struct sk_buff_head *list_)
 }
 
 /**
+ *	skb_queue_len	- get queue length
+ *	@list_: list to measure
+ *
+ *	Return the length of an &sk_buff queue.
+ *	This variant can be used in lockless contexts.
+ */
+static inline __u32 skb_queue_len_lockless(const struct sk_buff_head *list_)
+{
+	return READ_ONCE(list_->qlen);
+}
+
+/**
  *	__skb_queue_head_init - initialize non-spinlock portions of sk_buff_head
  *	@list: queue to initialize
  *
@@ -2026,7 +2038,7 @@ static inline void __skb_unlink(struct sk_buff *skb, struct sk_buff_head *list)
 {
 	struct sk_buff *next, *prev;
 
-	list->qlen--;
+	WRITE_ONCE(list->qlen, list->qlen - 1);
 	next	   = skb->next;
 	prev	   = skb->prev;
 	skb->next  = skb->prev = NULL;
diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c
index 321af97c7bbe..349e7fbfbc67 100644
--- a/net/unix/af_unix.c
+++ b/net/unix/af_unix.c
@@ -194,6 +194,12 @@ static inline int unix_recvq_full(struct sock const *sk)
 	return skb_queue_len(&sk->sk_receive_queue) > sk->sk_max_ack_backlog;
 }
 
+static inline int unix_recvq_full_lockless(struct sock const *sk)
+{
+	return skb_queue_len_lockless(&sk->sk_receive_queue) >
+		sk->sk_max_ack_backlog;
+}
+
 struct sock *unix_peer_get(struct sock *s)
 {
 	struct sock *peer;
@@ -1758,7 +1764,8 @@ static int unix_dgram_sendmsg(struct socket *sock, struct msghdr *msg,
 	 * - unix_peer(sk) == sk by time of get but disconnected before lock
 	 */
 	if (other != sk &&
-	    unlikely(unix_peer(other) != sk && unix_recvq_full(other))) {
+	    unlikely(unix_peer(other) != sk &&
+	    unix_recvq_full_lockless(other))) {
 		if (timeo) {
 			timeo = unix_wait_for_peer(other, timeo);
 
-- 
1.8.3.1

Re: [PATCH v2] skbuff: fix a data race in skb_queue_len()

[Eric Dumazet]

On 2/4/20 8:15 AM, Qian Cai wrote:
> sk_buff.qlen can be accessed concurrently as noticed by KCSAN,
> 
> 
> Since only the read is operating as lockless, it could introduce a logic
> bug in unix_recvq_full() due to the load tearing. Fix it by adding
> a lockless variant of skb_queue_len() and unix_recvq_full() where
> READ_ONCE() is on the read while WRITE_ONCE() is on the write similar to
> the commit d7d16a89350a ("net: add skb_queue_empty_lockless()").
> 
> Signed-off-by: Qian Cai <cai@lca.pw>
> ---
> 
> v2: add lockless variant helpers and WRITE_ONCE().
> 
>  include/linux/skbuff.h | 14 +++++++++++++-
>  net/unix/af_unix.c     |  9 ++++++++-
>  2 files changed, 21 insertions(+), 2 deletions(-)
> 
> diff --git a/include/linux/skbuff.h b/include/linux/skbuff.h
> index 3d13a4b717e9..de5eade20e52 100644
> --- a/include/linux/skbuff.h
> +++ b/include/linux/skbuff.h
> @@ -1822,6 +1822,18 @@ static inline __u32 skb_queue_len(const struct sk_buff_head *list_)
>  }
>  
>  /**
> + *	skb_queue_len	- get queue length

Please fix to use the exact name.

> + *	@list_: list to measure
> + *
> + *	Return the length of an &sk_buff queue.
> + *	This variant can be used in lockless contexts.
> + */
> +static inline __u32 skb_queue_len_lockless(const struct sk_buff_head *list_)
> +{
> +	return READ_ONCE(list_->qlen);
> +}
> +
> +/**
>   *	__skb_queue_head_init - initialize non-spinlock portions of sk_buff_head
>   *	@list: queue to initialize
>   *
> @@ -2026,7 +2038,7 @@ static inline void __skb_unlink(struct sk_buff *skb, struct sk_buff_head *list)
>  {
>  	struct sk_buff *next, *prev;
>  
> -	list->qlen--;
> +	WRITE_ONCE(list->qlen, list->qlen - 1);
>  	next	   = skb->next;
>  	prev	   = skb->prev;
>  	skb->next  = skb->prev = NULL;
> diff --git a/net/unix/af_unix.c b/net/unix/af_unix.c
> index 321af97c7bbe..349e7fbfbc67 100644
> --- a/net/unix/af_unix.c
> +++ b/net/unix/af_unix.c
> @@ -194,6 +194,12 @@ static inline int unix_recvq_full(struct sock const *sk)
>  	return skb_queue_len(&sk->sk_receive_queue) > sk->sk_max_ack_backlog;
>  }
>  
> +static inline int unix_recvq_full_lockless(struct sock const *sk)

The const attribute is misplaced. It should be :

static inline bool unix_recvq_full_lockless(const struct sock *sk)

> +{
> +	return skb_queue_len_lockless(&sk->sk_receive_queue) >
> +		sk->sk_max_ack_backlog;

You probably also need a READ_ONCE() for sk->sk_max_ack_backlog

It is a matter of time before syzbot finds how to trigger the race.

Since you added a nice unix_recvq_full_lockless() helper, lets make it right.

Thanks.