To cut a long story short, that ACK was sent when the socket didn't belong to anybody. Instead of allowing packets that pertain to a socket that belongs to user x, allow packets that pertain to a connection that was initiated by a socket from user x.
The longer story.
To understand the issue, it helps to understand how wget and HTTP requests work in general.
In
wget http://cachefly.cachefly.net/10mb.test
wget establishes a TCP connection to cachefly.cachefly.net, and once established sends a request in the HTTP protocol that says: "Please send me the content of /10mb.test (GET /10mb.test HTTP/1.1) and by the way, could you please not close the connection after you're done (Connection: Keep-alive). The reason it does that is because in case the server replies with a redirection for a URL on the same IP address, it can reuse the connection.
Now the server can reply with either, "here comes the data you requested, beware it's 10MB large (Content-Length: 10485760), and yes OK, I'll leave the connection open". Or if it doesn't know the size of the data, "Here's the data, sorry I can't leave the connection open but I'll tell when you can stop downloading the data by closing my end of the connection".
In the URL above, we're in the first case.
So, as soon as wget has obtained the headers for the response, it knows its job is done once it has downloaded 10MB of data.
Basically, what wget does is read the data until 10MB have been received and exit. But at that point, there's more to be done. What about the server? It's been told to leave the connection open.
Before exiting, wget closes (close system call) the file descriptor for the socket. Upon, the close, the system finishes acknowledging the data sent by the server and sends a FIN to say: "I won't be sending any more data". At that point close returns and wget exits. There is no socket associated to the TCP connection anymore (at least not one owned by any user). However it's not finished yet. Upon receiving that FIN, the HTTP server sees end-of-file when reading the next request from the client. In HTTP, that means "no more request, I'll close my end". So it sends its FIN as well, to say, "I won't be sending anything either, that connection is going away".
Upon receiving that FIN, the client sends a "ACK". But, at that point, wget is long gone, so that ACK is not from any user. Which is why it is blocked by your firewall. Because the server doesn't receive the ACK, it's going to send the FIN over and over until it gives up and you'll see more dropped ACKs. That also means that by dropping those ACKs, you're needlessly using resources of the server (which needs to maintain a socket in the LAST-ACK state) for quite some time.
The behavior would have been different if the client had not requested "Keep-alive" or the server had not replied with "Keep-alive".
As already mentioned, if you're using the connection tracker, what you want to do is let every packet in the ESTABLISHED and RELATED states through and only worry about NEW packets.
If you allow NEW packets from user x but not packets from user y, then other packets for established connections by user x will go through, and because there can't be established connections by user y (since we're blocking the NEW packets that would establish the connection), there will not be any packet for user y connections going through.
You can't unless you patch the kernel. And you have better things to do.
The counters are incremented __nf_ct_refresh_acct if the parameter do_acct is nonzero. This function is called through two wrappers: nf_ct_refresh_acct, which increments the counters, and nf_ct_refresh, which doesn't. The choice of wrapper is made according to the protocol type: protocols that can track do, the ones that can't don't.
The amount of computation is tiny. It's just two synchronized additions. Modern processors have very deep instruction pipelines, which makes conditional instructions expensive: the processor tries to predict which branch will be taken to start executing the next few instructions before it's determined the result of the test, and if the prediction is wrong, a lot of work needs to be discarded. The additions do require synchronization between the CPUs, because all CPUs have to be updating the same counter; on typical multicore architecture, this means that the core has to lock the cache line containing the counters. If the feature was optional, the CPU would have to read the configuration value, which doesn't require exclusive access so is a little less expensive. Still it would be an extremely tiny gain, to be balanced by the slightly less tiny loss for the majority of users who want the counters. It's just not worth having an option to disable this feature.
Best Answer
OUTPUT is for packets that are emitted by the host. Their destination is usually another host, but can be the same host via the loopback interface, so not all packets that go through OUTPUT are in fact outgoing.
FORWARD is for packets that are neither emitted by the host nor directed to the host. They are the packets that the host is merely routing.
When you start digging into packet mangling and NAT, the full story is rather more complex.