[PATCH net] net: pin protocol module before socket allocation

[PATCH net] net: pin protocol module before socket allocation

[Chengfeng Ye]

sk_prot_alloc() reads prot->slab and starts allocating the socket
before it gets a reference to prot->owner. A protocol module can begin
unloading after its protocol was selected but before the reference is
taken, allowing this interleaving:

  CPU 0                                CPU 1
  slab = prot->slab
                                       proto_unregister(prot)
                                         kmem_cache_destroy(prot->slab)
  kmem_cache_alloc(slab, ...)

kmem_cache_alloc() then dereferences a freed struct kmem_cache and can
crash or corrupt memory. The kernel reported:

  Oops: general protection fault, probably for non-canonical address
  KASAN: maybe wild-memory-access in range
  RIP: kmem_cache_alloc_noprof+0x63/0x370
  Call Trace:
   sk_prot_alloc+0x74/0x2c0
   sk_alloc+0x2b/0x6c0
   inet_create+0x2cd/0xd40
   __sock_create+0x1c3/0x430
   __sys_socket+0x116/0x1d0

Take the module reference before reading prot->slab so module removal
cannot destroy the cache during allocation. Drop that reference after
freeing the allocation on either failure path; on success sk_prot_free()
continues to release it as before.

Fixes: a79af59efd20 ("[NET]: Fix module reference counts for loadable protocol modules")
Signed-off-by: Chengfeng Ye <nicoyip.dev@gmail.com>
---
 net/core/sock.c | 13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)

diff --git a/net/core/sock.c b/net/core/sock.c
index 1ad41904db25..59f4b15fd594 100644
--- a/net/core/sock.c
+++ b/net/core/sock.c
@@ -2240,33 +2240,34 @@ static struct sock *sk_prot_alloc(struct proto *prot, gfp_t priority,
 	struct sock *sk;
 	struct kmem_cache *slab;
 
+	if (!try_module_get(prot->owner))
+		return NULL;
+
 	slab = prot->slab;
 	if (slab != NULL) {
 		sk = kmem_cache_alloc(slab, priority & ~__GFP_ZERO);
 		if (!sk)
-			return sk;
+			goto out_module_put;
 		if (want_init_on_alloc(priority))
 			sk_prot_clear_nulls(sk, prot->obj_size);
 	} else
 		sk = kmalloc(prot->obj_size, priority);
 
-	if (sk != NULL) {
-		if (security_sk_alloc(sk, family, priority))
-			goto out_free;
-
-		if (!try_module_get(prot->owner))
-			goto out_free_sec;
-	}
+	if (!sk)
+		goto out_module_put;
+
+	if (security_sk_alloc(sk, family, priority))
+		goto out_free;
 
 	return sk;
 
-out_free_sec:
-	security_sk_free(sk);
 out_free:
 	if (slab != NULL)
 		kmem_cache_free(slab, sk);
 	else
 		kfree(sk);
+out_module_put:
+	module_put(prot->owner);
 	return NULL;
 }
 
-- 
2.43.0

Re: [PATCH net] net: pin protocol module before socket allocation

[Kuniyuki Iwashima]

On Sun, Aug 23, 2026 at 10:13 AM Chengfeng Ye <nicoyip.dev@gmail.com> wrote:
>
> sk_prot_alloc() reads prot->slab and starts allocating the socket
> before it gets a reference to prot->owner. A protocol module can begin
> unloading after its protocol was selected but before the reference is
> taken, allowing this interleaving:
>
>   CPU 0                                CPU 1
>   slab = prot->slab
>                                        proto_unregister(prot)
>                                          kmem_cache_destroy(prot->slab)
>   kmem_cache_alloc(slab, ...)
>
> kmem_cache_alloc() then dereferences a freed struct kmem_cache and can
> crash or corrupt memory. The kernel reported:
>
>   Oops: general protection fault, probably for non-canonical address
>   KASAN: maybe wild-memory-access in range
>   RIP: kmem_cache_alloc_noprof+0x63/0x370
>   Call Trace:
>    sk_prot_alloc+0x74/0x2c0
>    sk_alloc+0x2b/0x6c0
>    inet_create+0x2cd/0xd40

inet_create() rather looks broken.

inet_unregister_protosw() calls synchronize_net(), but the following
proto_unregister() might be done after rcu_read_unlock() and before
WARN_ON(!answer_prot->slab) in inet_create().

Which module did you use to trigger the bug ?


>    __sock_create+0x1c3/0x430
>    __sys_socket+0x116/0x1d0
>
> Take the module reference before reading prot->slab so module removal
> cannot destroy the cache during allocation. Drop that reference after
> freeing the allocation on either failure path; on success sk_prot_free()
> continues to release it as before.
>
> Fixes: a79af59efd20 ("[NET]: Fix module reference counts for loadable protocol modules")
> Signed-off-by: Chengfeng Ye <nicoyip.dev@gmail.com>
> ---
>  net/core/sock.c | 13 ++++++++-----
>  1 file changed, 8 insertions(+), 5 deletions(-)
>
> diff --git a/net/core/sock.c b/net/core/sock.c
> index 1ad41904db25..59f4b15fd594 100644
> --- a/net/core/sock.c
> +++ b/net/core/sock.c
> @@ -2240,33 +2240,34 @@ static struct sock *sk_prot_alloc(struct proto *prot, gfp_t priority,
>         struct sock *sk;
>         struct kmem_cache *slab;
>
> +       if (!try_module_get(prot->owner))
> +               return NULL;
> +
>         slab = prot->slab;
>         if (slab != NULL) {
>                 sk = kmem_cache_alloc(slab, priority & ~__GFP_ZERO);
>                 if (!sk)
> -                       return sk;
> +                       goto out_module_put;
>                 if (want_init_on_alloc(priority))
>                         sk_prot_clear_nulls(sk, prot->obj_size);
>         } else
>                 sk = kmalloc(prot->obj_size, priority);
>
> -       if (sk != NULL) {
> -               if (security_sk_alloc(sk, family, priority))
> -                       goto out_free;
> -
> -               if (!try_module_get(prot->owner))
> -                       goto out_free_sec;
> -       }
> +       if (!sk)
> +               goto out_module_put;
> +
> +       if (security_sk_alloc(sk, family, priority))
> +               goto out_free;
>
>         return sk;
>
> -out_free_sec:
> -       security_sk_free(sk);
>  out_free:
>         if (slab != NULL)
>                 kmem_cache_free(slab, sk);
>         else
>                 kfree(sk);
> +out_module_put:
> +       module_put(prot->owner);
>         return NULL;
>  }
>
> --
> 2.43.0

[PATCH net v2] net: pin protocol module before inet socket allocation

[Chengfeng Ye]

inet_create() and inet6_create() look up the protocol under
rcu_read_lock(), then drop RCU before using the resulting proto.
sk_alloc() uses GFP_KERNEL, so RCU cannot be held across it.

A loadable protocol can be unregistered in that window.
inet_unregister_protosw() waits with synchronize_net() only for
readers still in the RCU section. After rcu_read_unlock(), module
exit can run proto_unregister(), destroy prot->slab, and free the
module. inet_create() then uses a dangling proto pointer:

  CPU 0 inet_create                     CPU 1 l2tp_ip_exit
  rcu_read_lock()
  answer_prot = answer->prot
  rcu_read_unlock()                     inet_unregister_protosw()
                                          synchronize_net()
                                        proto_unregister()
                                          kmem_cache_destroy(slab)
  WARN_ON(!answer_prot->slab)
  sk_alloc() -> kmem_cache_alloc(stale)

This was reproduced with socket(AF_INET, SOCK_DGRAM, IPPROTO_L2TP)
racing delete_module("l2tp_ip"):

  Oops: general protection fault, probably for non-canonical address
  KASAN: maybe wild-memory-access in range
  RIP: kmem_cache_alloc_noprof+0x63/0x370
  Call Trace:
   sk_prot_alloc+0x74/0x2c0
   sk_alloc+0x2b/0x6c0
   inet_create+0x2cd/0xd40
   __sock_create+0x1c3/0x430
   __sys_socket+0x116/0x1d0

__sock_create() already converts family lookup into a module
reference before dropping RCU, but that pins inet, not the protocol
module. Pin answer_prot->owner while still under RCU, then allocate.
sk_prot_alloc() still takes the socket-lifetime reference; drop the
temporary one on every path after sk_alloc(). inet6_create() has the
same hole.

Fixes: a79af59efd20 ("[NET]: Fix module reference counts for loadable protocol modules")
Cc: stable@vger.kernel.org
Signed-off-by: Chengfeng Ye <nicoyip.dev@gmail.com>
---
v2: Pin answer_prot->owner in inet_create()/inet6_create() under
    RCU instead of reordering try_module_get() in sk_prot_alloc().
    sk_alloc() can sleep, so the lookup must be converted into a
    module reference before rcu_read_unlock(). Suggested by
    Kuniyuki Iwashima.

 net/ipv4/af_inet.c  | 10 ++++++++--
 net/ipv6/af_inet6.c | 10 ++++++++--
 2 files changed, 16 insertions(+), 4 deletions(-)

diff --git a/net/ipv4/af_inet.c b/net/ipv4/af_inet.c
index 32d006c1a8ee..88a18ac0e6e4 100644
--- a/net/ipv4/af_inet.c
+++ b/net/ipv4/af_inet.c
@@ -325,6 +325,10 @@ static int inet_create(struct net *net, struct socket *sock, int protocol,
 	sock->ops = answer->ops;
 	answer_prot = answer->prot;
 	answer_flags = answer->flags;
+	if (!try_module_get(answer_prot->owner)) {
+		err = -EPROTONOSUPPORT;
+		goto out_rcu_unlock;
+	}
 	rcu_read_unlock();
 
 	WARN_ON(!answer_prot->slab);
@@ -332,7 +336,7 @@ static int inet_create(struct net *net, struct socket *sock, int protocol,
 	err = -ENOMEM;
 	sk = sk_alloc(net, PF_INET, GFP_KERNEL, answer_prot, kern);
 	if (!sk)
-		goto out;
+		goto out_module_put;
 
 	err = 0;
 	if (INET_PROTOSW_REUSE & answer_flags)
@@ -398,6 +402,8 @@ static int inet_create(struct net *net, struct socket *sock, int protocol,
 		if (err)
 			goto out_sk_release;
 	}
+out_module_put:
+	module_put(answer_prot->owner);
 out:
 	return err;
 out_rcu_unlock:
@@ -406,7 +412,7 @@ static int inet_create(struct net *net, struct socket *sock, int protocol,
 out_sk_release:
 	sk_common_release(sk);
 	sock->sk = NULL;
-	goto out;
+	goto out_module_put;
 }
 
 
diff --git a/net/ipv6/af_inet6.c b/net/ipv6/af_inet6.c
index 282912a11999..82b390ce3e0c 100644
--- a/net/ipv6/af_inet6.c
+++ b/net/ipv6/af_inet6.c
@@ -170,6 +170,10 @@ static int inet6_create(struct net *net, struct socket *sock, int protocol,
 	sock->ops = answer->ops;
 	answer_prot = answer->prot;
 	answer_flags = answer->flags;
+	if (!try_module_get(answer_prot->owner)) {
+		err = -EPROTONOSUPPORT;
+		goto out_rcu_unlock;
+	}
 	rcu_read_unlock();
 
 	WARN_ON(!answer_prot->slab);
@@ -177,7 +181,7 @@ static int inet6_create(struct net *net, struct socket *sock, int protocol,
 	err = -ENOBUFS;
 	sk = sk_alloc(net, PF_INET6, GFP_KERNEL, answer_prot, kern);
 	if (!sk)
-		goto out;
+		goto out_module_put;
 
 	sock_init_data(sock, sk);
 
@@ -251,6 +255,8 @@ static int inet6_create(struct net *net, struct socket *sock, int protocol,
 		if (err)
 			goto out_sk_release;
 	}
+out_module_put:
+	module_put(answer_prot->owner);
 out:
 	return err;
 out_rcu_unlock:
@@ -259,7 +265,7 @@ static int inet6_create(struct net *net, struct socket *sock, int protocol,
 out_sk_release:
 	sk_common_release(sk);
 	sock->sk = NULL;
-	goto out;
+	goto out_module_put;
 }
 
 int __inet6_bind(struct sock *sk, struct sockaddr_unsized *uaddr, int addr_len,
-- 
2.43.0

Re: [PATCH net] net: pin protocol module before socket allocation

[Chengfeng Ye]

On Tue, Aug 25, 2026 at 2:22 AM Kuniyuki Iwashima <kuniyu@google.com> wrote:
> inet_create() rather looks broken.
>
> inet_unregister_protosw() calls synchronize_net(), but the following
> proto_unregister() might be done after rcu_read_unlock() and before
> WARN_ON(!answer_prot->slab) in inet_create().

Right, that is the root cause. As sk_alloc() uses GFP_KERNEL, it
cannot run under rcu_read_lock(). I just sent a v2 by moving the extra
module reference get/put to inet_create().

> Which module did you use to trigger the bug ?

It is l2tp_ip, via socket(AF_INET, SOCK_DGRAM, IPPROTO_L2TP) racing
delete_module("l2tp_ip"). I think inet6_create() also suffers from
this bug as it has a similar pattern, so v2 also fixes that path as
well.

Best regards,
Chengfeng