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

TL;DR

Do not enable net.ipv4.tcp_tw_recycle—it doesn’t even exist anymore since Linux 4.12. Most of the time, TIME-WAIT sockets are harmless. Otherwise, jump to the summary for the recommended solutions.

The Linux kernel documentation is not very helpful about what net.ipv4.tcp_tw_recycle and net.ipv4.tcp_tw_reuse do. This lack of documentation opens the path to 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 the 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 wrking with NAT (Network Address Translation).

I will provide here a more detailed explanation of how to properly handle the TIME-WAIT state. 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.¹

• About the TIME-WAIT state
  • Purpose
  • Problems
    • Connection table slot
    • Memory
    • CPU
• Other solutions
  • Socket lingering
  • net.ipv4.tcp_tw_reuse
  • net.ipv4.tcp_tw_recycle
• Summary

About the 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:²

Only the end closing the connection first will reach the TIME-WAIT state. The other end will follow a path which usually permits it to quickly get rid of the connection.

You can have a look at the current state of connections with ss -tan:

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 detail what happens when the TIME-WAIT state is deficient.³ Here is an example of what could be avoided if the TIME-WAIT state wasn’t shortened:

• 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 state.⁴ 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:

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:

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 used 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 an 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:

The solution is more quadruplets.⁵ 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;⁶ or
• use more server IP by configuring additional IP on the web server.⁷

The last solution is to tweak net.ipv4.tcp_tw_reuse and net.ipv4.tcp_tw_recycle. Don’t do that yet, we will cover these 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)

   1 2	$ 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:⁸

   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; };

   1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26

   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 is 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.⁹

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 a 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)

   1 2	$ 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.

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:

If you have about 40,000 inbound connections in the TIME-WAIT state, it should eat less than 10 MiB of memory. If you have about 40,000 outbound connections in the TIME-WAIT state, you need to account for 2.5 MiB 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:

There is nothing to change here: the memory used by TIME-WAIT connections is really small. If your server needs 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 iterates 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 immediately and assume that all data will eventually be safely delivered.

However, an application can choose to disable this behavior, 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 immediately destroyed. No TIME-WAIT state in this case.
2. With the second flavor, any data is 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 successfully 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 unconditionally.

net.ipv4.tcp_tw_reuse

The TIME-WAIT state prevents delayed segments to be accepted in an unrelated connection. However, in 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 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 loss 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 on time, 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:

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. This is handy when the server usually closes the connection first.¹⁰

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:

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:

When the remote host is a NAT device, the condition on timestamps will forbid all 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.

Update (2017-09)

Starting from Linux 4.10 (commit 95a22caee396), Linux will randomize timestamp offsets for each connection, making this option completely broken, with or without NAT. It has been completely removed from Linux 4.12.

The LAST-ACK state is handled in the same way as for net.ipv4.tcp_tw_reuse.

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, never enable net.ipv4.tcp_tw_recycle due to its side effects. 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 responsibility 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.

参考链接

---

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