Coping with the TCP TIME-WAIT state on busy Linux servers

TL;DR: Do not enable net.ipv4.tcp_tw_recycle.
The Linux kernel documentation is not very helpful about what net.ipv4.tcp_tw_recycle does:

Enable fast recycling TIME-WAIT sockets. Default value is 0. It should not be changed without advice/request of technical experts.

Its sibling, net.ipv4.tcp_tw_reuse is a little bit more documented but the language is about the same:

Allow to reuse TIME-WAIT sockets for new connections when it is safe from protocol viewpoint. Default value is 0. It should not be changed without advice/request of technical experts.

The mere result of this lack of documentation is that we find numerous tuning guides advising to set both these settings to 1 to reduce the number of entries in the TIME-WAIT state. However, as stated by tcp(7) manual page, the net.ipv4.tcp_tw_recycle option is quite problematic for public-facing servers as it won’t handle connections from two different computers behind the same NAT device, which is a problem hard to detect and waiting to bite you:

Enable fast recycling of TIME-WAIT sockets. Enabling this option is not recommended since this causes problems when working with NAT (Network Address Translation).

I will provide here a more detailed explanation in the hope to teach people who are wrong on the Internet.

 

 Duty Calls — Someone is wrong on the Internet

As a sidenote, despite the use of ipv4 in its name, the net.ipv4.tcp_tw_recycle control also applies to IPv6. Also, keep in mind we are looking at the TCP stack of Linux. This is completely unrelated to Netfilter connection tracking which may be tweaked in other ways

About TIME-WAIT state

Let’s rewind a bit and have a close look at this TIME-WAIT state. What is it? See the TCP state diagram below

TCP state diagram

Only the end closing the connection first will reach the TIME-WAIT state. The other end will follow a path which usually permits to quickly get rid of the connection.
You can have a look at the current state of connections with ss -tan:

$ ss -tan | head -5
LISTEN     0  511             *:80              *:*     
SYN-RECV   0  0     192.0.2.145:80    203.0.113.5:35449
SYN-RECV   0  0     192.0.2.145:80   203.0.113.27:53599
ESTAB      0  0     192.0.2.145:80   203.0.113.27:33605
TIME-WAIT  0  0     192.0.2.145:80   203.0.113.47:50685 

Purpose

There are two purposes for the TIME-WAIT state:

• The most known one is to prevent delayed segments from one connection being accepted by a later connection relying on the same quadruplet (source address, source port, destination address, destination port). The sequence number also needs to be in a certain range to be accepted. This narrows a bit the problem but it still exists, especially on fast connections with large receive windows. RFC 1337 explains in details what happens when the TIME-WAIT state is deficient. Here is an example of what could be avoided if the TIME-WAIT state wasn’t shortened:

Due to a shortened TIME-WAIT state, a delayed TCP segment has been accepted in an unrelated connection.

• The other purpose is to ensure the remote end has closed the connection. When the last ACK is lost, the remote end stays in the LAST-ACK state⁴. Without the TIME-WAIT state, a connection could be reopened while the remote end still thinks the previous connection is valid. When it receives a SYN segment (and the sequence number matches), it will answer with a RST as it is not expecting such a segment. The new connection will be aborted with an error:

If the remote end stays in LAST-ACK state because the last ACK was lost, opening a new connection with the same quadruplet will not work.

RFC 793 requires the TIME-WAIT state to last twice the time of the MSL. On Linux, this duration is not tunable and is defined in include/net/tcp.h as one minute:

#define TCP_TIMEWAIT_LEN (60*HZ) /* how long to wait to destroy TIME-WAIT
                                  * state, about 60 seconds     */

There have been propositions to turn this into a tunable value but it has been refused on the ground the TIME-WAIT state is a good thing.

Problems

Now, let’s see why this state can be annoying on a server handling a lot of connections. There are three aspects of the problem:

• the slot taken in the connection table preventing new connections of the same kind,
• the memory occupied by the socket structure in the kernel, and
• the additional CPU usage.

The result of ss -tan state time-wait | wc -l is not a problem per se!

Connection table slot

A connection in the TIME-WAIT state is kept for one minute in the connection table. This means, another connection with the same quadruplet (source address, source port, destination address, destination port) cannot exist.
For a web server, the destination address and the destination port are likely to be constant. If your web server is behind a L7 load-balancer, the source address will also be constant. On Linux, the client port is by default allocated in a port range of about 30,000 ports (this can be changed by tuning net.ipv4.ip_local_port_range). This means that only 30,000 connections can be established between the web server and the load-balancer every minute, so about 500 connections per second.
If the TIME-WAIT sockets are on the client side, such a situation is easy to detect. The call to connect() will return EADDRNOTAVAIL and the application will log some error message about that. On the server side, this is more complex as there is no log and no counter to rely on. In doubt, you should just try to come with something sensible to list the number of used quadruplets:

$ ss -tan 'sport = :80' | awk '{print $(NF)" "$(NF-1)}' | \
>     sed 's/:[^ ]*//g' | sort | uniq -c
    696 10.24.2.30 10.33.1.64
   1881 10.24.2.30 10.33.1.65
   5314 10.24.2.30 10.33.1.66
   5293 10.24.2.30 10.33.1.67
   3387 10.24.2.30 10.33.1.68
   2663 10.24.2.30 10.33.1.69
   1129 10.24.2.30 10.33.1.70
  10536 10.24.2.30 10.33.1.73 

The solution is more quadruplets⁵. This can be done in several ways (in the order of difficulty to setup):

• use more client ports by setting net.ipv4.ip_local_port_range to a wider range,
• use more server ports by asking the web server to listen to several additional ports (81, 82, 83, …),
• use more client IP by configuring additional IP on the load balancer and use them in a round-robin fashion,
• use more server IP by configuring additional IP on the web server⁶.

Of course, a last solution is to tweak net.ipv4.tcp_tw_reuse and net.ipv4.tcp_tw_recycle. Don’t do that yet, we will cover those settings later.

Memory

With many connections to handle, leaving a socket open for one additional minute may cost your server some memory. For example, if you want to handle about 10,000 new connections per second, you will have about 600,000 sockets in the TIME-WAIT state. How much memory does it represent? Not that much!
First, from the application point of view, a TIME-WAIT socket does not consume any memory: the socket has been closed. In the kernel, a TIME-WAIT socket is present in three structures (for three different purposes):

1. A hash table of connections, named the “TCP established hash table” (despite containing connections in other states) is used to locate an existing connection, for example when receiving a new segment.
   Each bucket of this hash table contains both a list of connections in the TIME-WAIT state and a list of regular active connections. The size of the hash table depends on the system memory and is printed at boot:
   

   $ dmesg | grep "TCP established hash table"
   [    0.169348] TCP established hash table entries: 65536 (order: 8, 1048576 bytes)
   

   It is possible to override it by specifying the number of entries on the kernel command line with the thash_entries parameter.
   
   
   
   Each element of the list of connections in the TIME-WAIT state is a struct tcp_timewait_sock, while the type for other states is struct tcp_sock:
   

   struct tcp_timewait_sock {
       struct inet_timewait_sock tw_sk;
       u32    tw_rcv_nxt;
       u32    tw_snd_nxt;
       u32    tw_rcv_wnd;
       u32    tw_ts_offset;
       u32    tw_ts_recent;
       long   tw_ts_recent_stamp;
   };
   
   struct inet_timewait_sock {
       struct sock_common  __tw_common;
   
       int                     tw_timeout;
       volatile unsigned char  tw_substate;
       unsigned char           tw_rcv_wscale;
       __be16 tw_sport;
       unsigned int tw_ipv6only     : 1,
                    tw_transparent  : 1,
                    tw_pad          : 6,
                    tw_tos          : 8,
                    tw_ipv6_offset  : 16;
       unsigned long            tw_ttd;
       struct inet_bind_bucket *tw_tb;
       struct hlist_node        tw_death_node;
   };
   

2. A set of lists of connections, called the “death row”, is used to expire the connections in the TIME-WAIT state. They are ordered by how much time left before expiration.
   
   
   
   It uses the same memory space as for the entries in the hash table of connections. This is the struct hlist_node tw_death_node member of struct inet_timewait_sock
   
3. A hash table of bound ports, holding the locally bound ports and the associated parameters, is used to determine if it is safe to listen to a given port or to find a free port in the case of dynamic bind. The size of this hash table is the same as the size of the hash table of connections:
   

   $ dmesg | grep "TCP bind hash table"
   [    0.169962] TCP bind hash table entries: 65536 (order: 8, 1048576 bytes)
   

   Each element is a struct inet_bind_socket. There is one element for each locally bound port. A TIME-WAIT connection to a web server is locally bound to the port 80 and shares the same entry as its sibling TIME-WAIT connections. On the other hand, a connection to a remote service is locally bound to some random port and does not share its entry.
   

So, we are only concerned by the space occupied by struct tcp_timewait_sock and struct inet_bind_socket. There is one struct tcp_timewait_sock for each connection in the TIME-WAIT state, inbound or outbound. There is one dedicated struct inet_bind_socket for each outbound connection and none for an inbound connection.
A struct tcp_timewait_sock is only 168 bytes while a struct inet_bind_socket is 48 bytes:

$ sudo apt-get install linux-image-$(uname -r)-dbg
[...]
$ gdb /usr/lib/debug/boot/vmlinux-$(uname -r)
(gdb) print sizeof(struct tcp_timewait_sock)
 $1 = 168
(gdb) print sizeof(struct tcp_sock)
 $2 = 1776
(gdb) print sizeof(struct inet_bind_bucket)
 $3 = 48

So, if you have about 40,000 inbound connections in the TIME-WAIT state, it should eat less than 10MB of memory. If you have about 40,000 outbound connections in the TIME-WAIT state, you need to account for 2.5MB of additional memory. Let’s check that by looking at the output of slabtop. Here is the result on a server with about 50,000 connections in the TIME-WAIT state, 45,000 of which are outbound connections:

$ sudo slabtop -o | grep -E '(^  OBJS|tw_sock_TCP|tcp_bind_bucket)'
  OBJS ACTIVE  USE OBJ SIZE  SLABS OBJ/SLAB CACHE SIZE NAME                   
 50955  49725  97%    0.25K   3397       15     13588K tw_sock_TCP            
 44840  36556  81%    0.06K    760       59      3040K tcp_bind_bucket

There is nothing to change here: the memory used by TIME-WAIT connections is really small. If your server need to handle thousands of new connections per second, you need far more memory to be able to efficiently push data to clients. The overhead of TIME-WAIT connections is negligible.

CPU

On the CPU side, searching for a free local port can be a bit expensive. The work is done by the inet_csk_get_port() function which uses a lock and iterate on locally bound ports until a free port is found. A large number of entries in this hash table is usually not a problem if you have a lot of outbound connections in the TIME-WAIT state (like ephemeral connections to a memcached server): the connections usually share the same profile, the function will quickly find a free port as it iterates on them sequentially.

Other solutions

If you still think you have a problem with TIME-WAIT connections after reading the previous section, there are three additional solutions to solve them:

• disable socket lingering,
• net.ipv4.tcp_tw_reuse, and
• net.ipv4.tcp_tw_recycle.

Socket lingering

When close() is called, any remaining data in the kernel buffers will be sent in the background and the socket will eventually transition to the TIME-WAIT state. The application can continue to work immediatly and assume that all data will eventually be safely delivered.
However, an application can choose to disable this behaviour, known as socket lingering. There are two flavors:

1. In the first one, any remaining data will be discarded and instead of closing the connection with the normal four-packet connection termination sequence, the connection will be closed with a RST (and therefore, the peer will detect an error) and will be immediatly destroyed. No TIME-WAIT state in this case.
   
2. With the second flavor, if there is any data still remaining in the socket send buffer, the process will sleep when calling close() until either all the data is sent and acknowledged by the peer or the configured linger timer expires. It is possible for a process to not sleep by setting the socket as non-blocking. In this case, the same process happens in the background. It permits the remaining data to be sent during a configured timeout but if the data is succesfully sent, the normal close sequence is run and you get a TIME-WAIT state. And on the other case, you’ll get the connection close with a RST and the remaining data is discarded.
   

In both cases, disabling socket lingering is not a one-size-fits-all solution. It may be used by some applications like HAProxy or Nginx when it is safe to use from the upper protocol point of view. There are good reasons to not disable it unconditionnaly.

net.ipv4.tcp_tw_reuse

The TIME-WAIT state prevents delayed segments to be accepted in an unrelated connection. However, on certain conditions, it is possible to assume a new connection’s segment cannot be misinterpreted with an old connection’s segment.
RFC 1323 presents a set of TCP extensions to improve performance over high-bandwidth paths. Among other things, it defines a new TCP option carrying two four-byte timestamp fields. The first one is the current value of the timestamp clock of the TCP sending the option while the second one is the most recent timestamp received from the remote host.
By enabling net.ipv4.tcp_tw_reuse, Linux will reuse an existing connection in the TIME-WAIT state for a new outgoing connection if the new timestamp is strictly bigger than the most recent timestamp recorded for the previous connection: an outgoing connection in the TIME-WAIT state can be reused after just one second.
How is it safe? The first purpose of the TIME-WAIT state was to avoid duplicate segments to be accepted in an unrelated connection. Thanks to the use of timestamps, such a duplicate segments will come with an outdated timestamp and therefore be discarded.
The second purpose was to ensure the remote end is not in the LAST-ACK state because of the lost of the last ACK. The remote end will retransmit the FIN segment until:

1. it gives up (and tear down the connection), or
2. it receives the ACK it is waiting (and tear down the connection), or
3. it receives a RST (and tear down the connection).

If the FIN segments are received in a timely manner, the local end socket will still be in the TIME-WAIT state and the expected ACK segments will be sent.
Once a new connection replaces the TIME-WAIT entry, the SYN segment of the new connection is ignored (thanks to the timestamps) and won’t be answered by a RST but only by a retransmission of the FIN segment. The FIN segment will then be answered with a RST (because the local connection is in the SYN-SENT state) which will allow the transition out of the LAST-ACK state. The initial SYN segment will eventually be resent (after one second) because there was no answer and the connection will be established without apparent error, except a slight delay:

If the remote end stays in LAST-ACK state because the last ACK was lost, the remote connection will be reset when the local end transition to the SYN-SENT state.

It should be noted that when a connection is reused, the TWRecycled counter is increased (despite its name).

net.ipv4.tcp_tw_recycle

This mechanism also relies on the timestamp option but affects both incoming and outgoing connections which is handy when the server usually closes the connection first
The TIME-WAIT state is scheduled to expire sooner: it will be removed after the retransmission timeout (RTO) interval which is computed from the RTT and its variance. You can spot the appropriate values for a living connection with the ss command:

$ ss --info  sport = :2112 dport = :4057
State      Recv-Q Send-Q    Local Address:Port        Peer Address:Port   
ESTAB      0      1831936   10.47.0.113:2112          10.65.1.42:4057    
         cubic wscale:7,7 rto:564 rtt:352.5/4 ato:40 cwnd:386 ssthresh:200 send 4.5Mbps rcv_space:5792

To keep the same guarantees the TIME-WAIT state was providing, while reducing the expiration timer, when a connection enters the TIME-WAIT state, the latest timestamp is remembered in a dedicated structure containing various metrics for previous known destinations. Then, Linux will drop any segment from the remote host whose timestamp is not strictly bigger than the latest recorded timestamp, unless the TIME-WAIT state would have expired:

if (tmp_opt.saw_tstamp &&
    tcp_death_row.sysctl_tw_recycle &&
    (dst = inet_csk_route_req(sk, &fl4, req, want_cookie)) != NULL &&
    fl4.daddr == saddr &&
    (peer = rt_get_peer((struct rtable *)dst, fl4.daddr)) != NULL) {
        inet_peer_refcheck(peer);
        if ((u32)get_seconds() - peer->tcp_ts_stamp < TCP_PAWS_MSL &&
            (s32)(peer->tcp_ts - req->ts_recent) >
                                        TCP_PAWS_WINDOW) {
                NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
                goto drop_and_release;
        }
}

When the remote host is in fact a NAT device, the condition on timestamps will forbid allof the hosts except one behind the NAT device to connect during one minute because they do not share the same timestamp clock. In doubt, this is far better to disable this option since it leads to difficult to detect and difficult to diagnose problems.
The LAST-ACK state is handled in the exact same way as for net.ipv4.tcp_tw_recycle.

Summary

The universal solution is to increase the number of possible quadruplets by using, for example, more server ports. This will allow you to not exhaust the possible connections with TIME-WAIT entries.
On the server side, do not enable net.ipv4.tcp_tw_recycle unless you are pretty sure you will never have NAT devices in the mix. Enabling net.ipv4.tcp_tw_reuse is useless for incoming connections.
On the client side, enabling net.ipv4.tcp_tw_reuse is another almost-safe solution. Enabling net.ipv4.tcp_tw_recycle in addition to net.ipv4.tcp_tw_reuse is mostly useless.
Moreover, when designing protocols, don’t let clients close first. Clients won’t have to deal with the TIME-WAIT state pushing the responsability to servers which are better suited to handle this.
And a final quote by W. Richard Stevens, in Unix Network Programming:

The TIME_WAIT state is our friend and is there to help us (i.e., to let old duplicate segments expire in the network). Instead of trying to avoid the state, we should understand it.

---