Thread (22 messages) 22 messages, 7 authors, 2022-08-05

Re: [PATCH v4] net: fix refcount bug in sk_psock_get (2)

From: Jakub Kicinski <kuba@kernel.org>
Date: 2022-08-03 15:14:27
Also in: bpf, linux-kernel-mentees, lkml
Subsystem: bpf [general] (safe dynamic programs and tools), the rest · Maintainers: Alexei Starovoitov, Daniel Borkmann, Andrii Nakryiko, Eduard Zingerman, Kumar Kartikeya Dwivedi, Linus Torvalds

On Wed,  3 Aug 2022 20:41:22 +0800 Hawkins Jiawei wrote:
-/* Pointer stored in sk_user_data might not be suitable for copying
- * when cloning the socket. For instance, it can point to a reference
- * counted object. sk_user_data bottom bit is set if pointer must not
- * be copied.
+/* flag bits in sk_user_data
+ *
+ * SK_USER_DATA_NOCOPY - Pointer stored in sk_user_data might
+ * not be suitable for copying when cloning the socket.
+ * For instance, it can point to a reference counted object.
+ * sk_user_data bottom bit is set if pointer must not be copied.
+ *
+ * SK_USER_DATA_BPF    - Managed by BPF
I'd use this opportunity to add more info here, BPF is too general.
Maybe "Pointer is used by a BPF reuseport array"? Martin, WDYT?
+ * SK_USER_DATA_PSOCK  - Mark whether pointer stored in sk_user_data points
+ * to psock type. This bit should be set when sk_user_data is
+ * assigned to a psock object.
+/**
+ * rcu_dereference_sk_user_data_psock - return psock if sk_user_data
+ * points to the psock type(SK_USER_DATA_PSOCK flag is set), otherwise
+ * return NULL
+ *
+ * @sk: socket
+ */
+static inline
+struct sk_psock *rcu_dereference_sk_user_data_psock(const struct sock *sk)
nit: the return type more commonly goes on the same line as "static
inline"
+{
+	uintptr_t __tmp = (uintptr_t)rcu_dereference(__sk_user_data((sk)));
+
+	if (__tmp & SK_USER_DATA_PSOCK)
+		return (struct sk_psock *)(__tmp & SK_USER_DATA_PTRMASK);
+
+	return NULL;
+}
As a follow up we can probably generalize this into
 __rcu_dereference_sk_user_data_cond(sk, bit)

and make the psock just call that:

static inline struct sk_psock *
rcu_dereference_sk_user_data_psock(const struct sock *sk)
{
	return __rcu_dereference_sk_user_data_cond(sk, SK_USER_DATA_PSOCK);
}

then reuseport can also benefit, maybe:
diff --git a/kernel/bpf/reuseport_array.c b/kernel/bpf/reuseport_array.c
index e2618fb5870e..ad5c447a690c 100644
--- a/kernel/bpf/reuseport_array.c
+++ b/kernel/bpf/reuseport_array.c
@@ -21,14 +21,11 @@ static struct reuseport_array *reuseport_array(struct bpf_map *map)
 /* The caller must hold the reuseport_lock */
 void bpf_sk_reuseport_detach(struct sock *sk)
 {
-	uintptr_t sk_user_data;
+	struct sock __rcu **socks;
 
 	write_lock_bh(&sk->sk_callback_lock);
-	sk_user_data = (uintptr_t)sk->sk_user_data;
-	if (sk_user_data & SK_USER_DATA_BPF) {
-		struct sock __rcu **socks;
-
-		socks = (void *)(sk_user_data & SK_USER_DATA_PTRMASK);
+	socks = __rcu_dereference_sk_user_data_cond(sk, SK_USER_DATA_BPF);
+	if (socks) {
 		WRITE_ONCE(sk->sk_user_data, NULL);
 		/*
 		 * Do not move this NULL assignment outside of

But that must be a separate patch, not part of this fix.
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help