Skip to contentSkip to conversation

net: socket: set sock->sk to NULL after calling proto_ops::release()

PATCH In review Eric Biggers 5 participants 8 messages Feb 21, 2019 – Feb 25, 2019

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 net: socket: set sock->sk to NULL after calling proto_ops::release()

From: Eric Biggers <[email protected]>

Commit 9060cb719e61 ("net: crypto set sk to NULL when af_alg_release.") fixed a use-after-free in sockfs_setattr() when an AF_ALG socket is closed concurrently with fchownat(). However, it ignored that many other proto_ops::release() methods don't set sock->sk to NULL and therefore allow the same use-after-free:

- base_sock_release
- bnep_sock_release
- cmtp_sock_release
- data_sock_release
- dn_release
- hci_sock_release
- hidp_sock_release
- iucv_sock_release
- l2cap_sock_release
- llcp_sock_release
- llc_ui_release
- rawsock_release
- rfcomm_sock_release
- sco_sock_release
- svc_release
- vcc_release
- x25_release

Rather than fixing all these and relying on every socket type to get this right forever, just make __sock_release() set sock->sk to NULL itself after calling proto_ops::release().

Reproducer that produces the KASAN splat when any of these socket types are configured into the kernel:

    #include <pthread.h>
    #include <stdlib.h>
    #include <sys/socket.h>
    #include <unistd.h>

    pthread_t t;
    volatile int fd;

    void *close_thread(void *arg)
    {
        for (;;) {
            usleep(rand() % 100);
            close(fd);
        }
    }

    int main()
    {
        pthread_create(&t, NULL, close_thread, NULL);
        for (;;) {
            fd = socket(rand() % 50, rand() % 11, 0);
            fchownat(fd, "", 1000, 1000, 0x1000);
            close(fd);
        }
    }
Fixes 86741ec25462 ("net: core: Add a UID field to struct sock.")
Cc <[email protected]> # v4.10+

---

NOTE: I am not an expert in the networking code, so please carefully check that I haven't missed some reason why this simple fix won't do.

 net/socket.c | 1 +
 1 file changed, 1 insertion(+)
net/socket.c +1 -0
index d80d87a395ea..320f51b22b19 100644
@@ -577,6 +577,7 @@ static void __sock_release(struct socket *sock, struct inode *inode)
 		if (inode)
 			inode_lock(inode);
 		sock->ops->release(sock);
+		sock->sk = NULL;
 		if (inode)
 			inode_unlock(inode);
 		sock->ops = NULL;
Signature · 1 line
2.21.0.rc0.258.g878e2cd30e-goog
#1 lore raw
Quoted 3 lines · Eric Biggers

Rather than fixing all these and relying on every socket type to get this right forever, just make __sock_release() set sock->sk to NULL itself after calling proto_ops::release().

	Is there any case when we would want sock->sk non-NULL when

sock->sk->sk_socket is NULL?

IOW, why not make sock_orphan() take care of that, at the same
time we dissociate sock from socket? After all, on the other end of things we have both sock_graft() and sock_init_data() set both sock->sk and sk->sk_socket...

	Looking at the assignments to sock->sk, I see

* atalk_release(), bcm_release(), raw_release(), pfkey_release(), pppol2tp_release(), packet_release(), smc_release(), xsk_release() - directly after sock_orphan()
* rxrpc_release() - directly *before* sock_orphan()

* ax25_release() - somewhat after sock_orphan(); no idea if anything done in between needs it still set. Doesn't looks like there could be - readers of sock->sk in there are all in methods that can't overlap with ->release().
* kcm_release() - similar.
* rds_release() - similar.
* tipc_release() - similar.
* unix_accept() - similar, and there I'm fairly sure it could be done right after sock_orphan()
* netrom_release() - similar. BTW, why the hell does nr_accept() check for NULL sock->sk, while other methods (callable for exact same sockets) assume that it's non-NULL? AFAICS, the check in nr_accept() is pointless - nr_create() can leave sock->sk NULL only if it fails, in which case it's going to be immediately hit by ->release() and freed by iput() in sock_release().
* rose_release() - similar, complete with ->accept() oddity.

* caif_release() - done *before* sock_orphan(), with other bit of the latter duplicated just prior. NFI why; it messes with debugfs, so there's a whole kettle of copulating slugs involved ;-/

* ieee802154_sock_release(), inet_release(), pn_socket_release() - done before calliing into protocol's ->close(), which ends up calling sock_orphan(), either directly or via sk_common_release(). I suspect that we would be find with zeroing sock->sk delayed until sock_orphan() in all of those, but that needs more RTFS than I want to go into at the moment.

* netlink_release() - somewhat after sock_orphan(); not sure, calls of ->netlink_unbind() done inbetween might want it still set.

* qrtr_release() - lacks sock_orphan(), might be broken

* vsock_releae() - NFI.

* vcc_create() - clears sock->sk in the very beginning; AFAICS that's pointless.

#2 lore raw
In reply to Eric Biggers
Quoted 8 lines · Eric Biggers

From: Eric Biggers <[email protected]>

Commit 9060cb719e61 ("net: crypto set sk to NULL when af_alg_release.") fixed a use-after-free in sockfs_setattr() when an AF_ALG socket is closed concurrently with fchownat(). However, it ignored that many other proto_ops::release() methods don't set sock->sk to NULL and therefore allow the same use-after-free:

I fail to see how setting a pointer to NULL can avoid races.

We lack some kind of protection, rcu or something, if another thread can change sock->sk at anytime while sockfs_setattr() is used.

sockfs_setattr()
...

     if (sock->sk)

// even if sock->sk was not NULL for the if (...).

// it can be NULL right now, compiler could read sock->sk a second time and catch a NULL.

        sock->sk->sk_uid = iattr->ia_uid;
Quoted 75 lines

- base_sock_release
- bnep_sock_release
- cmtp_sock_release
- data_sock_release
- dn_release
- hci_sock_release
- hidp_sock_release
- iucv_sock_release
- l2cap_sock_release
- llcp_sock_release
- llc_ui_release
- rawsock_release
- rfcomm_sock_release
- sco_sock_release
- svc_release
- vcc_release
- x25_release

Rather than fixing all these and relying on every socket type to get this right forever, just make __sock_release() set sock->sk to NULL itself after calling proto_ops::release().

Reproducer that produces the KASAN splat when any of these socket types are configured into the kernel:

    #include <pthread.h>
    #include <stdlib.h>
    #include <sys/socket.h>
    #include <unistd.h>

    pthread_t t;
    volatile int fd;

    void *close_thread(void *arg)
    {
        for (;;) {
            usleep(rand() % 100);
            close(fd);
        }
    }

    int main()
    {
        pthread_create(&t, NULL, close_thread, NULL);
        for (;;) {
            fd = socket(rand() % 50, rand() % 11, 0);
            fchownat(fd, "", 1000, 1000, 0x1000);
            close(fd);
        }
    }
Fixes 86741ec25462 ("net: core: Add a UID field to struct sock.")
Cc <[email protected]> # v4.10+

---

NOTE: I am not an expert in the networking code, so please carefully check that I haven't missed some reason why this simple fix won't do.

 net/socket.c | 1 +
 1 file changed, 1 insertion(+)
net/socket.c +1 -0
index d80d87a395ea..320f51b22b19 100644
@@ -577,6 +577,7 @@ static void __sock_release(struct socket *sock, struct inode *inode)
 		if (inode)
 			inode_lock(inode);
 		sock->ops->release(sock);
+		sock->sk = NULL;
 		if (inode)
 			inode_unlock(inode);
 		sock->ops = NULL;
#3 lore raw

Hi Eric,

Quoted 20 lines · Eric Dumazet
Quoted 8 lines · Eric Biggers

From: Eric Biggers <[email protected]>

Commit 9060cb719e61 ("net: crypto set sk to NULL when af_alg_release.") fixed a use-after-free in sockfs_setattr() when an AF_ALG socket is closed concurrently with fchownat(). However, it ignored that many other proto_ops::release() methods don't set sock->sk to NULL and therefore allow the same use-after-free:

I fail to see how setting a pointer to NULL can avoid races.

We lack some kind of protection, rcu or something, if another thread can change sock->sk at anytime while sockfs_setattr() is used.

sockfs_setattr()
...

Quoted 9 lines
     if (sock->sk)

// even if sock->sk was not NULL for the if (...).

// it can be NULL right now, compiler could read sock->sk a second time and catch a NULL.

        sock->sk->sk_uid = iattr->ia_uid;

->setattr() is called under inode_lock(), which __sock_release() also takes. So the uses of sock->sk are serialized. See commit 6d8c50dcb029 ("socket: close race condition between sock_close() and sockfs_setattr()").

The issue now is that if ->setattr() happens *after* __sock_release() (which is possible if fchownat() gets the reference to the file's 'struct path', then the file is close()d by another thread, then fchownat() continues), it will see stale sock->sk because for many socket types it wasn't set to NULL earlier.

- Eric

Acked-by on v1 1/1
#6 lore raw
In reply to Eric Biggers
Quoted 60 lines · Eric Biggers

From: Eric Biggers <[email protected]>

Commit 9060cb719e61 ("net: crypto set sk to NULL when af_alg_release.") fixed a use-after-free in sockfs_setattr() when an AF_ALG socket is closed concurrently with fchownat(). However, it ignored that many other proto_ops::release() methods don't set sock->sk to NULL and therefore allow the same use-after-free:

- base_sock_release
- bnep_sock_release
- cmtp_sock_release
- data_sock_release
- dn_release
- hci_sock_release
- hidp_sock_release
- iucv_sock_release
- l2cap_sock_release
- llcp_sock_release
- llc_ui_release
- rawsock_release
- rfcomm_sock_release
- sco_sock_release
- svc_release
- vcc_release
- x25_release

Rather than fixing all these and relying on every socket type to get this right forever, just make __sock_release() set sock->sk to NULL itself after calling proto_ops::release().

Reproducer that produces the KASAN splat when any of these socket types are configured into the kernel:

    #include <pthread.h>
    #include <stdlib.h>
    #include <sys/socket.h>
    #include <unistd.h>

    pthread_t t;
    volatile int fd;

    void *close_thread(void *arg)
    {
        for (;;) {
            usleep(rand() % 100);
            close(fd);
        }
    }

    int main()
    {
        pthread_create(&t, NULL, close_thread, NULL);
        for (;;) {
            fd = socket(rand() % 50, rand() % 11, 0);
            fchownat(fd, "", 1000, 1000, 0x1000);
            close(fd);
        }
    }
Quoted 3 lines
Fixes 86741ec25462 ("net: core: Add a UID field to struct sock.")
Cc <[email protected]> # v4.10+
#4 lore raw
In reply to Eric Biggers
Quoted 3 lines · Eric Biggers

->setattr() is called under inode_lock(), which __sock_release() also takes. So the uses of sock->sk are serialized. See commit 6d8c50dcb029 ("socket: close race condition between sock_close() and sockfs_setattr()").

Oh right, we added another inode_lock()/inode_unlock() for sock_close()

Quoted 8 lines

The issue now is that if ->setattr() happens *after* __sock_release() (which is possible if fchownat() gets the reference to the file's 'struct path', then the file is close()d by another thread, then fchownat() continues), it will see stale sock->sk because for many socket types it wasn't set to NULL earlier.

- Eric

#5 lore raw
Quoted 9 lines · Eric Dumazet
Quoted 3 lines · Eric Biggers

->setattr() is called under inode_lock(), which __sock_release() also takes. So the uses of sock->sk are serialized. See commit 6d8c50dcb029 ("socket: close race condition between sock_close() and sockfs_setattr()").

Oh right, we added another inode_lock()/inode_unlock() for sock_close()

An interesting question is whether anything else will be confused by

	sock->sk && sock->sk->sk_socket != sock

I'd still like to figure out if we could simply make sock_orphan() do something like

	if (likely(sk->sk_socket))
		sk->sk_socket->sk = NULL;

just before sk_set_socket(sk, NULL);

That would make for much easier rules; the question is whether anything relies upon the windows when linkage between socket and sock is not symmetrical...

maintainer
#7 lore raw
In reply to Eric Biggers

From: Eric Biggers <[email protected]>
Date: Thu, 21 Feb 2019 14:13:56 -0800

Quoted 7 lines

From: Eric Biggers <[email protected]>

Commit 9060cb719e61 ("net: crypto set sk to NULL when af_alg_release.") fixed a use-after-free in sockfs_setattr() when an AF_ALG socket is closed concurrently with fchownat(). However, it ignored that many other proto_ops::release() methods don't set sock->sk to NULL and therefore allow the same use-after-free:

...

Quoted 6 lines

Rather than fixing all these and relying on every socket type to get this right forever, just make __sock_release() set sock->sk to NULL itself after calling proto_ops::release().

Reproducer that produces the KASAN splat when any of these socket types are configured into the kernel:

...

Quoted 2 lines
Fixes 86741ec25462 ("net: core: Add a UID field to struct sock.")

Applied and queued up for -stable, thanks Eric.