Re: [PATCH 4/5] bpf: sockmap, sockhash: return file descriptors from privileged lookup

Jakub Sitnicki wrote:

On Tue, Mar 10, 2020 at 06:47 PM CET, Lorenz Bauer wrote:

quoted

Allow callers with CAP_NET_ADMIN to retrieve file descriptors from a
sockmap and sockhash. O_CLOEXEC is enforced on all fds.

Without this, it's difficult to resize or otherwise rebuild existing
sockmap or sockhashes.

Suggested-by: Jakub Sitnicki <jakub@cloudflare.com>
Signed-off-by: Lorenz Bauer <redacted>
---
 net/core/sock_map.c | 19 +++++++++++++++++++
 1 file changed, 19 insertions(+)

diff --git a/net/core/sock_map.c b/net/core/sock_map.c
index 03e04426cd21..3228936aa31e 100644
--- a/net/core/sock_map.c
+++ b/net/core/sock_map.c

@@ -347,12 +347,31 @@ static void *sock_map_lookup(struct bpf_map *map, void *key)
 static int __sock_map_copy_value(struct bpf_map *map, struct sock *sk,
 				 void *value)
 {
+	struct file *file;
+	int fd;
+
 	switch (map->value_size) {
 	case sizeof(u64):
 		sock_gen_cookie(sk);
 		*(u64 *)value = atomic64_read(&sk->sk_cookie);
 		return 0;

+	case sizeof(u32):
+		if (!capable(CAP_NET_ADMIN))
+			return -EPERM;
+
+		fd = get_unused_fd_flags(O_CLOEXEC);
+		if (unlikely(fd < 0))
+			return fd;
+
+		read_lock_bh(&sk->sk_callback_lock);
+		file = get_file(sk->sk_socket->file);

I think this deserves a second look.

We don't lock the sock, so what if tcp_close orphans it before we enter
this critical section? Looks like sk->sk_socket might be NULL.

I'd find a test that tries to trigger the race helpful, like:

  thread A: loop in lookup FD from map
  thread B: loop in insert FD into map, close FD

Agreed, this was essentially my question above as well.

When the psock is created we call sock_hold() and will only do a sock_put()
after an rcu grace period when its removed. So at least if you have the
sock here it should have a sk_refcnt. (Note the user data is set to NULL
so if you do reference psock you need to check its non-null.)

Is that enough to ensure sk_socket? Seems not to me, tcp_close for example
will still happen and call sock_orphan(sk) based on my admittddly quick
look.

Further, even if you do check sk->sk_socket is non-null what does it mean
to return a file with a socket that is closed, deleted from the sock_map
and psock removed? At this point is it just a dangling reference?

Still a bit confused as well what would or should happen when the sock is closed
after you have the file reference? I could probably dig up what exactly
would happen but I think we need it in the commiit message so we understand
it. I also didn't dig up the details here but if the receiver of the
fd crashes or otherwise disappears this hopefully all get cleaned up?

quoted

+		read_unlock_bh(&sk->sk_callback_lock);
+
+		fd_install(fd, file);
+		*(u32 *)value = fd;
+		return 0;
+
 	default:
 		return -ENOSPC;
 	}