TCP(4)

💖 tcp(4) — Internet Transmission Control Protocol 💖

okay so real talk: this page is from the Kernel Interfaces Manual, section 4 (TCP(4)), and it is literally the entire personality of TCP. bestie it is LONG, so grab a snack 🍓 and let's get into it.

---

✨ NAME ✨

tcp — Internet Transmission Control Protocol

🌸 SYNOPSIS 🌸

#include <sys/types.h>
#include <sys/socket.h>
#include <netinet/in.h>
#include <netinet/tcp.h>

int
socket(AF_INET, SOCK_STREAM, 0);

🦋 DESCRIPTION 🦋

The TCP protocol gives you reliable, flow-controlled, two-way transmission of data — like a group chat where nobody's messages get dropped and everyone takes turns 🫶. It's a byte-stream protocol used to support the SOCK_STREAM abstraction.

TCP uses the standard Internet address format and, on top of that, hands you a per-host collection of “port addresses”. So every address is made of an Internet address (which specifies the host and the network) plus a specific TCP port on the host that identifies the peer entity 💅.

Sockets using the TCP protocol are either “active” or “passive”. Active sockets initiate connections to passive sockets. By default, TCP sockets are created active — so to make a passive socket you have to call listen(2) after binding the socket with bind(2). Only passive sockets may use accept(2) to accept incoming connections. Only active sockets may use connect(2) to initiate connections. that's the whole rule and it's not negotiable, iykyk ⭐

Passive sockets may “underspecify” their location so they can match incoming connection requests from multiple networks — this technique is called “wildcard addressing” and it's how a single server serves clients on multiple networks 🌈. To create a socket that listens on all networks, you must bind the Internet address INADDR_ANY. You can still specify the TCP port at that point; if you don't, the system will assign one. Once a connection has been established, the socket's address is fixed by the peer entity's location. The address assigned to the socket is the address associated with the network interface that packets are being transmitted and received through. Normally, that address corresponds to the peer entity's network.

TCP supports a bunch of socket options that you set with setsockopt(2) and test with getsockopt(2) — here's the full lineup, no skips 👑

🎀 the socket options 🎀

🌟 TCP_INFO

Info about a socket's underlying TCP session can be retrieved by passing the read-only option TCP_INFO to getsockopt(2). It accepts a single argument: a pointer to an instance of struct tcp_info.

bestie this API is subject to change 😭 — consult the source to figure out which fields are actually filled out by this option right now. The FreeBSD specific additions include send window size, receive window size, and bandwidth-controlled window space.

🌟 TCP_CCALGOOPT

Set or query congestion control algorithm specific parameters. See mod_cc(4) for the details 💜.

🌟 TCP_CONGESTION

Select or query the congestion control algorithm that TCP will use for the connection. See mod_cc(4) for the details 💜.

🌟 TCP_FASTOPEN

Enable or disable TCP Fast Open (TFO) 🚀. Heads up: to use this option, the kernel must be built with the TCP_RFC7413 option.

This option can be set on the socket either before or after listen(2) is invoked. Clearing this option on a listen socket after it has been set has no effect on existing TFO connections or on TFO connections in progress — it only prevents new TFO connections from being established. no takebacks, basically 💅

For passively-created sockets, you can query the TCP_FASTOPEN socket option to find out whether the connection was established using TFO. Note that connections established via a TFO SYN, but which fall back to using a non-TFO SYN|ACK, will still have the TCP_FASTOPEN socket option set. sneaky but true 🙈

In addition to the facilities defined in RFC7413, this implementation supports a pre-shared key (PSK) mode of operation, where the TFO server requires the client to be in possession of a shared secret in order for the client to successfully open TFO connections with the server. This is useful, for example, in environments where TFO servers are exposed to both internal and external clients and you only want to allow TFO connections from internal clients 🎀

In the PSK mode of operation, the server generates and sends TFO cookies to requesting clients as usual. However, when validating cookies received in TFO SYNs from clients, the server requires the client-supplied cookie to equal:

SipHash24(key=16-byte-psk, msg=cookie-sent-to-client)

Multiple concurrent valid pre-shared keys are supported, so time-based rolling PSK invalidation policies can be implemented in the system ✨. The default number of concurrent pre-shared keys is 2.

This can be adjusted with the TCP_RFC7413_MAX_PSKS kernel option.

🌟 TCP_FUNCTION_BLK

Select or query the set of functions that TCP will use for this connection. This lets a user select an alternate TCP stack 👑. The alternate TCP stack must already be loaded in the kernel. To list the available TCP stacks, see functions_available in the “FIB support” — TCP sockets are FIB-aware. They inherit the FIB of the process which created the socket, or that of the listening socket for sockets created by accept(2). In particular, the FIB is not inherited from that of the interface where the initiating SYN packet was received. When an incoming connection request arrives to a listening socket, the initial handshake also occurs in the FIB of the listening socket, not that of the received packet.

By default, a TCP listening socket can accept connections originating from any FIB. If the net.inet.tcp.bind_all_fibs tunable is set to 0, a listening socket will only accept connections originating from the listening socket's own FIB. Connection requests from other FIBs will be treated as though there is no listening socket for the destination address and port 💀. In this mode, multiple listening sockets owned by the same user can listen on the same address and port so long as they belong to different FIBs — similar to the behavior of the SO_REUSEPORT socket option. If the tunable is set to 0, all sockets added to a load-balancing group created with the SO_REUSEPORT_LB socket option must belong to the same FIB. “MIB (sysctl) Variables” section further down. To list the default TCP stack, see functions_default in the “MIB (sysctl) Variables” section.

🌟 TCP_KEEPINIT

This setsockopt(2) option accepts a per-socket timeout argument of u_int in seconds, for new, non-established TCP connections. For the global default in milliseconds, see keepinit in the “MIB (sysctl) Variables” section further down ⏳.

🌟 TCP_KEEPIDLE

This setsockopt(2) option accepts an argument of u_int for the amount of time, in seconds, that the connection must be idle before keepalive probes (if enabled) are sent for the connection of this socket. If set on a listening socket, the value is inherited by the newly created socket upon accept(2). For the global default in milliseconds, see keepidle in the “MIB (sysctl) Variables” section further down ⏳.

🌟 TCP_KEEPINTVL

This setsockopt(2) option accepts an argument of u_int to set the per-socket interval, in seconds, between keepalive probes sent to a peer. If set on a listening socket, the value is inherited by the newly created socket upon accept(2). For the global default in milliseconds, see keepintvl in the “MIB (sysctl) Variables” section further down ⏳.

🌟 TCP_KEEPCNT

This setsockopt(2) option accepts an argument of u_int and lets you tune, per-socket, the number of probes sent with no response before the connection gets dropped 💀. If set on a listening socket, the value is inherited by the newly created socket upon accept(2). For the global default, see keepcnt in the “MIB (sysctl) Variables” section further down.

🌟 TCP_NODELAY

Under most circumstances, TCP sends data the moment it's presented; but when outstanding data has not yet been acknowledged, it gathers small amounts of output to be sent in a single packet once an acknowledgement is received 🧚. For a small number of clients — think window systems sending a stream of mouse events that receive no replies — this packetization can cause significant delays. The boolean option TCP_NODELAY defeats this algorithm. periodt. 🔥

🌟 TCP_MAXSEG

By default, a sender-TCP and a receiver-TCP negotiate among themselves to determine the maximum segment size to be used for each connection 🤝. The TCP_MAXSEG option allows the user to determine the result of this negotiation, and to reduce it if desired.

---

(more coming, bestie — this page is not done with us yet 💖)

• TCP_MAXUNACKTIME — okay so this one takes a u_int argument and sets a per-socket timer, in seconds, for how long the connection is allowed to make no progress before things get spicy 🔥. "Progress" here means at least 1 byte gets acknowledged inside that window. If the connection just sits there giving nothing, the TCP stack will straight-up terminate it with a reset 💀. Heads up though: the default value is zero, and zero means no progress checks at all — the vibe is "we trust you."
• TCP_NOOPT — bestie, TCP is a talker. It usually sends a bunch of options in every packet, all matching the various TCP extensions this implementation supports. This boolean option is your mute button 🎀: it disables TCP option use on a per-connection basis.
• TCP_NOPUSH — so by convention, the sender-TCP sets the "push" bit and starts transmitting immediately (if it's permitted to) at the end of every single user call to write(2) or writev(2). When you set this option to a non-zero value, TCP will hold everything back — it delays sending any data at all until either the socket is closed or the internal send buffer is full. Pure "wait for it… wait for it…" energy ngl 🦋.
• TCP_MD5SIG — this enables MD5 digests (also known as TCP-MD5) on writes to the socket you picked. Outgoing traffic gets digested, and digests on incoming traffic get verified. The moment this option is enabled on a socket, all inbound and outgoing TCP segments must be signed with MD5 digests — no exceptions, periodt.

One super common use for this in a FreeBSD router deployment is letting those routers interwork with Cisco equipment at peering points 🤝. Support for this feature conforms to RFC 2385.

For this option to actually function correctly, the admin has to add a tcp-md5 key entry to the system's security associations database (SADB) using the setkey(8) utility. And jsyk, this entry can only be specified on a per-host basis at this time.

If an SADB entry can't be found for the destination, the system does not send any outgoing segments and drops any inbound segments 😭. However — during connection negotiation, a non-signed segment will be accepted if an SADB entry does not exist between hosts. And when a non-signed segment is accepted, the established connection is not protected with MD5 digests. So like, it goes through, but unprotected. Choose wisely 💅.

• TCP_STATS — manage collection of connection level statistics using the stats(3) framework. Each dropped segment is taken into account in the TCP protocol statistics 📊 — nothing gets to quietly disappear, the receipts are kept.
• TCP_TXTLS_ENABLE — enables in-kernel Transport Layer Security (TLS) for data written to this socket ✨. See ktls(4) for more details.
• TCP_TXTLS_MODE — the integer argument here can be used to get or set the current TLS transmit mode of a socket. See ktls(4) for more details.
• TCP_RXTLS_ENABLE — enables in-kernel TLS for data read from this socket 🌸. See ktls(4) for more details.
• TCP_REUSPORT_LB_NUMA — changes NUMA affinity filtering for an established TCP listen socket. This option takes a single integer argument which specifies the NUMA domain to filter on for this listen socket. The argument can also have the following special values:
  • TCP_REUSPORT_LB_NUMA_NODOM — remove NUMA filtering for this listen socket 🙅‍♀️.
  • TCP_REUSPORT_LB_NUMA_CURDOM — filter traffic associated with the domain where the calling thread is currently executing. This is typically used after a process or thread inherits a listen socket from its parent, and sets its CPU affinity to a particular core 🎯.
• TCP_REMOTE_UDP_ENCAPS_PORT — set and get the remote UDP encapsulation port. It can only be set on a closed TCP socket ⚠️ (closed only, those are the rules).

---

Okay so the option level for the setsockopt(2) call is the protocol number for TCP — you can grab that from getprotobyname(3), or just use IPPROTO_TCP. All options are declared in <netinet/tcp.h> 📚.

Options at the IP transport level may be used with TCP too; see ip(4). And incoming connection requests that are source-routed are noted, with the reverse source route used in responding 👀.

The default congestion control algorithm for TCP is cc_cubic(4) 💙. Other congestion control algorithms can be made available using the mod_cc(4) framework.

🌈 MIB (sysctl) Variables 🌈

The TCP protocol implements a number of variables in the net.inet.tcp branch of the sysctl(3) MIB, which can also be read or modified with sysctl(8) 👑.

• ack_war_timewindow, ack_war_cnt — the challenge ACK throttling algorithm defined in RFC 5961 limits the number of challenge ACKs sent per TCP connection to ack_war_cnt during the time interval specified in milliseconds by ack_war_timewindow. Setting ack_war_timewindow or ack_war_cnt to zero disables challenge ACK throttling 🛑.
• always_keepalive — assume that SO_KEEPALIVE is set on all TCP connections, so the kernel will periodically send a packet to the remote host to verify the connection is still up 🫶.
• blackhole — if enabled, disable sending of RST when a connection is attempted to a port where there is no socket accepting connections. See blackhole(4).
• blackhole_local — see blackhole(4).
• cc — a number of variables for congestion control are under the net.inet.tcp.cc node. See mod_cc(4).
• cc.newreno — variables for NewReno congestion control are under the net.inet.tcp.cc.newreno node. See cc_newreno(4).
• delacktime — maximum amount of time, in milliseconds, before a delayed ACK is sent ⏱️.
• delayed_ack — delay ACK to try and piggyback it onto a data packet or another ACK. Efficiency queen behavior 💅.
• do_prr — perform SACK loss recovery using the Proportional Rate Reduction (PRR) algorithm described in RFC6937. This improves the effectiveness of retransmissions — particularly in environments with ACK thinning or burst loss events — since the chances of running out of the ACK clock are reduced, preventing lengthy and performance reducing RTO based loss recovery (default is true) ✅.
• do_tcpdrain — flush packets in the TCP reassembly queue if the system is low on mbufs 🚿.
• drop_synfin — drop TCP packets with both SYN and FIN set. Like, pick a lane bestie 😭.
• ecn.enable — enable support for TCP Explicit Congestion Notification (ECN). ECN lets a TCP sender reduce the transmission rate in order to avoid packet drops 🌱.
  • 0 — disable ECN.
  • 1 — allow incoming connections to request ECN. Outgoing connections will request ECN.
  • 2 — allow incoming connections to request ECN. Outgoing connections will not request ECN. (default)
  • 3 — negotiate on incoming connection for Accurate ECN, ECN, or no ECN. Outgoing connections will request Accurate ECN and fall back to ECN depending on the capabilities of the server.
  • 4 — negotiate on incoming connection for Accurate ECN, ECN, or no ECN. Outgoing connections will not request ECN.
• ecn.maxretries — number of retries (SYN or SYN/ACK retransmits) before disabling ECN on a specific connection. This is needed to help with connection establishment when a broken firewall is in the network path 🧱.
• fast_finwait2_recycle — recycle TCP FIN_WAIT_2 connections faster when the socket is marked as SBS_CANTRCVMORE (no user process has the socket open, data received on the socket cannot be read). The timeout used here is finwait2_timeout ♻️.
• fastopen.acceptany — when non-zero, all client-supplied TFO cookies will be considered to be valid. The default is 0. (So by default, no, it does not just believe everyone 🙈.)
• fastopen.autokey — when this and net.inet.tcp.fastopen.server_enable are non-zero, a new key will be automatically generated after this specified seconds. The default is 120 🔑.
• fastopen.ccache_bucket_limit — the maximum number of entries in a client cookie cache bucket. The default value can be tuned with the TCP_FASTOPEN_CCACHE_BUCKET_LIMIT_DEFAULT kernel option or by setting net.inet.tcp.fastopen_ccache_bucket_limit in the loader(8) 🧺.
• fastopen.ccache_buckets — how many buckets the client cookie cache gets. Read-only, bestie 🙈 you can't just write to it. Tune it with the TCP_FASTOPEN_CCACHE_BUCKETS_DEFAULT kernel option, or by setting fastopen.ccache_buckets in the loader(8).
• fastopen.ccache_list — prints out the client cookie cache so you can 👀 vibe check it. Read-only.
• fastopen.client_enable — when it's zero, no new active (that's client-side) TFO connections can be created, periodt. On the transition from enabled → disabled, the client cookie cache gets cleared and disabled. Important: the enabled → disabled transition does not touch any active TFO connections already in progress — it only stops new ones from being established. Default is 0.
• fastopen.keylen — the key length, in bytes. Read-only.
• fastopen.maxkeys — the maximum number of keys supported. Read-only.
• fastopen.maxpsks — the maximum number of pre-shared keys supported. Read-only.
• fastopen.numkeys — the current number of keys installed. Read-only.
• fastopen.numpsks — the current number of pre-shared keys installed. Read-only.
• fastopen.path_disable_time — so when a failure happens while trying to create a new active (i.e., client) TFO connection, new active connections on that same path — path being determined by the tuple {client_ip, server_ip, server_port} — get forced to be non-TFO for this many seconds 😭. One thing though: the path disable mechanism relies on state stored in client cookie cache entries, so it's possible the disable time for a given path gets reduced if the matching client cookie cache entry gets reused due to resource pressure before the disable period has even elapsed. The default is TCP_FASTOPEN_PATH_DISABLE_TIME_DEFAULT.
• fastopen.psk_enable — when it's non-zero, pre-shared key (PSK) mode is enabled for all TFO servers ✨. On the enabled → disabled transition, all installed pre-shared keys are removed. The default is 0.
• fastopen.server_enable — when zero, no new passive (that's server-side) TFO connections can be created. On the enabled → disabled transition, all installed keys and pre-shared keys are removed 🫣. On the disabled → enabled transition, if fastopen.autokey is non-zero and there are no keys installed, a new key gets generated immediately. And same deal as the client one: the enabled → disabled transition does not affect any passive TFO connections already in progress — it only prevents new ones from being established. The default is 0.
• fastopen.setkey — installs a new key, by writing net.inet.tcp.fastopen.keylen bytes to this sysctl.
• fastopen.setpsk — installs a new pre-shared key, by writing net.inet.tcp.fastopen.keylen bytes to this sysctl.
• finwait2_timeout — the timeout to use for fast recycling of TCP FIN_WAIT_2 connections (fast_finwait2_recycle). Defaults to 60 seconds.
• functions_available — the list of available TCP function blocks (aka TCP stacks) 👑
• functions_default — the default TCP function block (TCP stack).
• hostcache — okay so the TCP host cache is genuinely iconic: it caches connection details and metrics so future connections between the same hosts perform better 🔥. When a TCP connection completes, a host caches info for that connection for some defined period of time. There are a bunch of hostcache variables hanging out under this node — peep hostcache.enable.
• hostcache.bucketlimit — the maximum number of entries for the same hash. Defaults to 30.
• hostcache.cachelimit — overall entry limit for the hostcache. Defaults to hashsize * bucketlimit.
• hostcache.count — the current number of entries in the host cache.
• hostcache.enable — enable/disable the host cache:
  • 0 — disable the host cache.
  • 1 — enable the host cache. (default)
• hostcache.expire — time in seconds for how long an entry should be kept in the host cache since it was last accessed. Defaults to 3600 (1 hour).
• hostcache.hashsize — the size of the TCP hostcache hashtable. Heads up: this number has to be a power of two, or it gets rejected 🚫. Defaults to 512.
• hostcache.histo — provides a histogram of the hostcache hash utilization. Cute little bar chart moment 📊.
• hostcache.list — provides a complete list of all the current entries in the host cache.
• hostcache.prune — time in seconds between pruning expired host cache entries. Defaults to 300 (5 minutes).
• hostcache.purge — expire all entries on the next pruning of host cache entries. Any non-zero setting gets reset to zero once the purge is running.
  • 0 — do not purge all entries when pruning the host cache (default).
  • 1 — purge all entries when doing the next pruning.
  • 2 — purge all entries and also reseed the hash salt.
• hostcache.purgenow — immediately purges all entries once set to any value. Setting it to 2 will also reseed the hash salt.
• icmp_may_rst — certain ICMP unreachable messages may abort connections in SYN-SENT state.
• initcwnd_segments — enables the ability to specify the initial congestion window in number of segments. The default value is 10, as suggested by RFC 6928. Changing the value on the fly would not affect connections using a congestion window from the hostcache. ⚠️ Caution: this regulates the burst of packets allowed to be sent in the first RTT. The value should be relative to the link capacity — start with small values for lower-capacity links. Large bursts can cause buffer overruns and packet drops if routers have small buffers or the link is experiencing congestion. fr fr be careful with this one.
• insecure_rst — use the criteria defined in RFC793 instead of RFC5961 for accepting RST segments. Default is false.
• insecure_syn — use the criteria defined in RFC793 instead of RFC5961 for accepting SYN segments. Default is false.
• insecure_ack — use the criteria defined in RFC793 for validating SEG.ACK. Default is false.
• isn_reseed_interval — the interval (in seconds) specifying how often the secret data used in RFC 1948 initial sequence number calculations should be reseeded. By default this variable is set to zero, indicating that no reseeding will occur. And ngl, reseeding shouldn't be necessary, and it will break TIME_WAIT recycling for a few minutes 💀.
• keepcnt — number of keepalive probes sent, with no response, before a connection is dropped. The default is 8 packets.
• keepidle — amount of time, in milliseconds, that the connection must be idle before sending keepalive probes (if enabled). The default is 7200000 msec (7.2M msec, 2 hours).
• keepinit — timeout, in milliseconds, for new, non-established TCP connections. The default is 75000 msec (75K msec, 75 sec).
• keepintvl — the interval, in milliseconds, between keepalive probes sent to remote machines when no response is received on a keepidle probe. The default is 75000 msec (75K msec, 75 sec).
• log_in_vain — logs any connection attempts to ports where there's no socket accepting connections. The value of 1 limits the logging to SYN (connection establishment) packets only. A value of 2 results in any TCP packets to closed ports being logged. Any value not listed above disables the logging (default is 0, i.e., logging is disabled).
• minmss — minimum TCP Maximum Segment Size; used to prevent a denial of service attack from an unreasonably low MSS 🛡️.
• msl — the Maximum Segment Lifetime, in milliseconds, for a packet.
• msl_local — the Maximum Segment Lifetime, in milliseconds, for a packet when both endpoints are local. msl_local is only used if nolocaltimewait, which is deprecated, is zero 🙃.
• mssdflt — the default value used for the TCP Maximum Segment Size ("MSS") for IPv4 when no advice to the contrary is received from MSS negotiation.
• newcwv — turns on the New Congestion Window Validation mechanism, exactly as described in RFC 7661. It gently reduces the congestion window during periods when TCP is application limited and the network bandwidth isn't being used completely. That lowkey prevents self-inflicted packet losses once the app starts sending data at a higher speed. ✨
• nolocaltimewait — suppresses creation of TCP TIME_WAIT states for connections where both endpoints are local. The default is 0. > ⚠️ bestie, nolocaltimewait is deprecated and will be removed in FreeBSD 16. msl_local can be used instead. 🫶
• path_mtu_discovery — enables Path MTU Discovery. 🛰️
• pcbcount — number of active protocol control blocks (read-only). 👑
• perconn_stats_enable — controls the default collection of statistics for all connections using the stats(3) framework. 0 disables it, 1 enables it, and 2 enables random sampling across log id connection groups, with all connections in a group receiving the same setting. 🎀
• perconn_stats_sample_rates — a CSV list of template_spec=percent key-value pairs that controls the per template sampling rates when stats(3) sampling is enabled. 🧮
• persmax — maximum persistence interval, msec. ⏳
• persmin — minimum persistence interval, msec. ⏳
• pmtud_blackhole_detection — enables automatic path MTU blackhole detection. In case of retransmits of MSS sized segments, the OS lowers the MSS to check if it's an MTU problem. If the current MSS is greater than the configured value to try (net.inet.tcp.pmtud_blackhole_mss and net.inet.tcp.v6pmtud_blackhole_mss), it gets set to this value; otherwise, the MSS gets set to the default values (net.inet.tcp.mssdflt and net.inet.tcp.v6mssdflt). Settings:
  • 0 — disable path MTU blackhole detection.
  • 1 — enable path MTU blackhole detection for IPv4 and IPv6.
  • 2 — enable path MTU blackhole detection only for IPv4.
  • 3 — enable path MTU blackhole detection only for IPv6. 🕳️
• pmtud_blackhole_mss — MSS to try for IPv4 if PMTU blackhole detection is turned on. 🕳️
• reass.cursegments — the current total number of segments present in all reassembly queues. 📦
• reass.maxqueuelen — the maximum number of segments allowed in each reassembly queue. By default, the system chooses a limit based on each TCP connection's receive buffer size and maximum segment size (MSS). The actual limit applied to a session's reassembly queue will be the lower of the system-calculated automatic limit and the user-specified reass.maxqueuelen limit. 🎚️
• reass.maxsegments — the maximum limit on the total number of segments across all reassembly queues. The limit can be adjusted as a tunable. 🔧
• recvbuf_auto — enables automatic receive buffer sizing as a connection progresses. 🌱
• recvbuf_max — maximum size of the automatic receive buffer. 📈
• recvspace — initial TCP receive window (buffer size). 🪟
• retries — maximum number of consecutive timer based retransmits sent after a data segment is lost (default and maximum is 12). 🔁
• rexmit_drop_options — drops TCP options from the third and later retransmitted SYN segments of a connection. ✂️
• rexmit_initial, rexmit_min, rexmit_slop, rexmit_max — adjust the retransmit timer calculation for TCP. A new connection starts with the timer set to rexmit_initial. The rexmit_slop is typically added to the raw calculation to take into account occasional variances that the SRTT (smoothed round-trip time) can't accommodate, while the minimum specifies an absolute minimum. A number of TCP RFCs suggest a 1 second minimum, but those RFCs tend to focus on streaming behavior and fail to deal with the fact that a 1 second minimum has severe detrimental effects over lossy interactive connections — such as a 802.11b wireless link — and over very fast but lossy connections for those cases not covered by the fast retransmit code. For this reason, we use 200ms of slop and a near-0 minimum, which gives us an effective minimum of 200ms (similar to Linux). The initial value is used before an RTT measurement has been performed. rexmit_min and rexmit_max set minimum and maximum timer values that a connection may have. ⏱️
• rfc1323 — implements the window scaling and timestamp options of RFC 1323/RFC 7323 (default is 1). Settings:
  • 0 — disable window scaling and timestamp option.
  • 1 — enable window scaling and timestamp option.
  • 2 — enable only window scaling.
  • 3 — enable only timestamp option. 🪟
• rfc3042 — enables the Limited Transmit algorithm as described in RFC 3042. It helps avoid timeouts on lossy links and also when the congestion window is small, as happens on short transfers. 🚦
• rfc3390 — enables support for RFC 3390, which allows for a variable-sized starting congestion window on new connections, depending on the maximum segment size. This helps throughput in general, but particularly affects short transfers and high-bandwidth large propagation-delay connections. 🚀
• sack.enable — enables support for RFC 2018, the TCP Selective Acknowledgment option, which allows the receiver to inform the sender about all successfully arrived segments, allowing the sender to retransmit the missing segments only. 🧩
• sack.globalholes — global number of TCP SACK holes currently allocated. 🕳️
• sack.globalmaxholes — maximum number of SACK holes per system, across all connections. Defaults to 65536. 🏆
• sack.lrd — enables Lost Retransmission Detection for SACK-enabled sessions, enabled by default. Under severe congestion, a retransmission can be lost, which then leads to a mandatory Retransmission Timeout (RTO), followed by slow-start. LRD will try to resend the repeatedly lost packet, preventing the time-consuming RTO and performance reducing slow-start or purge of the SACK scoreboard. 🕵️
• sack.maxholes — maximum number of SACK holes per connection. Defaults to 128. 🎯
• sack.revised — enables three updated mechanisms from RFC6675 (default is true). It calculates the bytes in flight using the algorithm described in RFC 6675, and is also an improvement when Proportional Rate Reduction is enabled. Next, Rescue Retransmission helps timely loss recovery when the trailing segments of a transmission are lost while no additional data is ready to be sent; in case a partial ACK without a SACK block is received during SACK loss recovery, the trailing segment is immediately resent rather than waiting for a Retransmission timeout. Finally, SACK loss recovery is also engaged once two segments plus one byte are SACKed — even if no traditional duplicate ACKs were observed. > ⚠️ sack.revised is deprecated and will be removed in FreeBSD 16. sack.enable will always follow RFC6675. 💀
• sendbuf_auto — enables automatic send buffer sizing. 📤
• sendbuf_auto_lowat — modifies the threshold for auto send buffer growth to account for SO_SNDLOWAT. 🎚️
• sendbuf_inc — incrementor step size of the automatic send buffer. ➕
• sendbuf_max — maximum size of the automatic send buffer. 📊
• sendspace — initial TCP send window (buffer size). 🪟
• syncache — variables under the net.inet.tcp.syncache node are documented in syncache(4). 🗂️
• syncookies — determines whether or not SYN cookies should be generated for outbound SYN-ACK packets. SYN cookies are a great help during SYN flood attacks, and they're enabled by default. (See syncookies(4).) 🍪
• syncookies_only — see syncookies(4). 👀
• tcbhashsize — size of the TCP control-block hash table (read-only). This is tuned using the kernel option TCBHASHSIZE or by setting net.inet.tcp.tcbhashsize in loader(8). 🔑
• tolerate_missing_ts — bestie, this one's a whole mood. It lets you tolerate missing timestamps (RFC 1323/RFC 7323) on TCP segments belonging to connections where TCP timestamp support was negotiated. Why would you ever need this? Because as of June 2021, several TCP stacks are known to straight-up violate RFC 7323 — including modern, widely deployed ones 😭. So the default is 1, i.e. missing timestamps are tolerated. Periodt.
• ts_offset_per_conn — when initializing the TCP timestamps, use a per-connection offset instead of a per-host-pair offset. Default? Per-connection offsets, as recommended in RFC 7323. That's the slay option, obvi.
• tso — enable TCP Segmentation Offload. Iconic little performance moment ✨
• udp_tunneling_overhead — the overhead taken into account when using UDP encapsulation. Since MSS clamping by middleboxes will most likely not work (rude), values larger than 8 (the size of the UDP header) are also supported. Supported values are between 8 and 1024. The default is 8.
• udp_tunneling_port — the local UDP encapsulation port. A value of 0 indicates UDP encapsulation is disabled — and yep, the default is 0. So it's off until you say otherwise 🫶
• v6mssdflt — the default value used for the TCP Maximum Segment Size (“MSS”) for IPv6 when no advice to the contrary is received from MSS negotiation.
• v6pmtud_blackhole_mss — MSS to try for IPv6 if PMTU blackhole detection is turned on. See pmtud_blackhole_detection 👀

---

💀 ERRORS 💀

A socket operation may fail with one of these errors returned — and ngl, some of these are giving "your wifi died at 3am" energy:

• [EISCONN] — when trying to establish a connection on a socket which already has one. Girl, it's already connected 🥴
• [ENOBUFS] or [ENOMEM] — when the system runs out of memory for an internal data structure.
• [ETIMEDOUT] — when a connection was dropped due to excessive retransmissions. It tried so hard and got so far 😭
• [ECONNRESET] — when the remote peer forces the connection to be closed. Rude but ok.
• [ECONNREFUSED] — when the remote peer actively refuses connection establishment (usually because no process is listening to the port). Nobody's home, bestie 🏚️
• [EADDRINUSE] — when an attempt is made to create a socket with a port which has already been allocated. Taken. Claimed. Off the market.
• [EADDRNOTAVAIL] — when an attempt is made to create a socket with a network address for which no network interface exists. That address literally does not exist ✨
• [EAFNOSUPPORT] — when an attempt is made to bind or connect a socket to a multicast address.
• [EINVAL] — when trying to change TCP function blocks at an invalid point in the session.
• [ENOENT] — when trying to use a TCP function block that is not available.

---

🌈 SEE ALSO 🌈

Go read these, they're all lowkey iconic in their own way:

getsockopt(2), setfib(2), socket(2), stats(3), sysctl(3), blackhole(4), inet(4), intro(4), ip(4), ktls(4), mod_cc(4), siftr(4), syncache(4), tcp_bbr(4), tcp_rack(4), setkey(8), sysctl(8), tcp_functions(9)

And the legendary literature 💅:

• V. Jacobson, B. Braden, and D. Borman, TCP Extensions for High Performance, RFC 1323.
• D. Borman, B. Braden, V. Jacobson, and R. Scheffenegger, TCP Extensions for High Performance, RFC 7323.
• A. Heffernan, Protection of BGP Sessions via the TCP MD5 Signature Option, RFC 2385.
• K. Ramakrishnan, S. Floyd, and D. Black, The Addition of Explicit Congestion Notification (ECN) to IP, RFC 3168.
• A. Ramaiah, R. Stewart, and M. Dalal, Improving TCP's Robustness to Blind In-Window Attacks, RFC 5961.

---

🦋 HISTORY 🦋

The TCP protocol appeared in 4.2BSD — like, ancient history, the origin story 🫶. The RFC 1323 extensions for window scaling and timestamps were added in 4.4BSD. And the TCP_INFO option was introduced in Linux 2.6 and is subject to change — so like, don't get too attached, it's giving "situationship" 💀.

---

and that's the whole page, bestie!! you ate that, fr fr 💖✨ this little beauty ships with Debian, and the page you just read is dated September 5, 2025 (yes, TCP(4) — the one and only 👑). go forth and configure responsibly 🌸🎀