netlink: make sure we allow at least one dump skb

Commit under Fixes tightened up the memory accounting for Netlink
sockets. Looks like the accounting is too strict for some existing
use cases, Marek reported issues with nl80211 / WiFi iw CLI.

To reduce number of iterations Netlink dumps try to allocate
messages based on the size of the buffer passed to previous
recvmsg() calls. If user space uses a larger buffer in recvmsg()
than sk_rcvbuf we will allocate an skb we won't be able to queue.

Make sure we always allow at least one skb to be queued.
Same workaround is already present in netlink_attachskb().
Alternative would be to cap the allocation size to
  rcvbuf - rmem_alloc
but as I said, the workaround is already present in other places.

Reported-by: Marek Szyprowski <[email protected]>
Link: https://lore.kernel.org/[email protected]
Fixes: ae8f160 ("netlink: Fix wraparounds of sk->sk_rmem_alloc.")
Tested-by: Marek Szyprowski <[email protected]>
Reviewed-by: Kuniyuki Iwashima <[email protected]>
Link: https://patch.msgid.link/[email protected]
Signed-off-by: Jakub Kicinski <[email protected]>

Author: kuba-moo

net/netlink/af_netlink.c

@@ -2258,11 +2258,11 @@ static int netlink_dump(struct sock *sk, bool lock_taken)
 	struct netlink_ext_ack extack = {};
 	struct netlink_callback *cb;
 	struct sk_buff *skb = NULL;
+	unsigned int rmem, rcvbuf;
 	size_t max_recvmsg_len;
 	struct module *module;
 	int err = -ENOBUFS;
 	int alloc_min_size;
-	unsigned int rmem;
 	int alloc_size;
 
 	if (!lock_taken)
@@ -2294,8 +2294,9 @@ static int netlink_dump(struct sock *sk, bool lock_taken)
 	if (!skb)
 		goto errout_skb;
 
+	rcvbuf = READ_ONCE(sk->sk_rcvbuf);
 	rmem = atomic_add_return(skb->truesize, &sk->sk_rmem_alloc);
-	if (rmem >= READ_ONCE(sk->sk_rcvbuf)) {
+	if (rmem != skb->truesize && rmem >= rcvbuf) {
 		atomic_sub(skb->truesize, &sk->sk_rmem_alloc);
 		goto errout_skb;
 	}