1TCP(7)                     Linux Programmer's Manual                    TCP(7)


6       tcp - TCP protocol


9       #include <sys/socket.h>
10       #include <netinet/in.h>
11       #include <netinet/tcp.h>
13       tcp_socket = socket(AF_INET, SOCK_STREAM, 0);


16       This  is  an  implementation  of  the  TCP protocol defined in RFC 793,
17       RFC 1122 and RFC 2001 with the NewReno and SACK  extensions.   It  pro‐
18       vides  a  reliable, stream-oriented, full-duplex connection between two
19       sockets on top of ip(7), for both v4 and v6 versions.   TCP  guarantees
20       that the data arrives in order and retransmits lost packets.  It gener‐
21       ates and checks a per-packet checksum  to  catch  transmission  errors.
22       TCP does not preserve record boundaries.
24       A  newly  created  TCP socket has no remote or local address and is not
25       fully specified.  To create an outgoing TCP connection  use  connect(2)
26       to establish a connection to another TCP socket.  To receive new incom‐
27       ing connections, first bind(2) the socket to a local address  and  port
28       and  then  call  listen(2)  to put the socket into the listening state.
29       After that a new socket for each incoming connection  can  be  accepted
30       using  accept(2).   A socket which has had accept(2) or connect(2) suc‐
31       cessfully called on it is fully specified and may transmit data.   Data
32       cannot be transmitted on listening or not yet connected sockets.
34       Linux supports RFC 1323 TCP high performance extensions.  These include
35       Protection Against Wrapped Sequence Numbers (PAWS), Window Scaling  and
36       Timestamps.   Window scaling allows the use of large (> 64 kB) TCP win‐
37       dows in order to support links with high latency or bandwidth.  To make
38       use of them, the send and receive buffer sizes must be increased.  They
39       can  be  set  globally   with   the   /proc/sys/net/ipv4/tcp_wmem   and
40       /proc/sys/net/ipv4/tcp_rmem  files,  or  on individual sockets by using
41       the SO_SNDBUF and SO_RCVBUF socket options with the setsockopt(2) call.
43       The maximum sizes for socket buffers declared  via  the  SO_SNDBUF  and
44       SO_RCVBUF    mechanisms    are   limited   by   the   values   in   the
45       /proc/sys/net/core/rmem_max  and   /proc/sys/net/core/wmem_max   files.
46       Note that TCP actually allocates twice the size of the buffer requested
47       in the setsockopt(2) call, and so a succeeding getsockopt(2) call  will
48       not  return  the  same size of buffer as requested in the setsockopt(2)
49       call.  TCP uses the extra space for administrative purposes and  inter‐
50       nal  kernel  structures,  and  the /proc file values reflect the larger
51       sizes compared to the actual TCP windows.  On  individual  connections,
52       the socket buffer size must be set prior to the listen(2) or connect(2)
53       calls in order to have it take effect.  See socket(7) for more informa‐
54       tion.
56       TCP  supports  urgent data.  Urgent data is used to signal the receiver
57       that some important message is part of the  data  stream  and  that  it
58       should  be  processed as soon as possible.  To send urgent data specify
59       the MSG_OOB option to send(2).  When urgent data is received, the  ker‐
60       nel sends a SIGURG signal to the process or process group that has been
61       set as the socket "owner" using the SIOCSPGRP or FIOSETOWN  ioctls  (or
62       the POSIX.1-specified fcntl(2) F_SETOWN operation).  When the SO_OOBIN‐
63       LINE socket option is enabled, urgent data is put into the normal  data
64       stream  (a program can test for its location using the SIOCATMARK ioctl
65       described below), otherwise it can be received only  when  the  MSG_OOB
66       flag is set for recv(2) or recvmsg(2).
68       When out-of-band data is present, select(2) indicates the file descrip‐
69       tor as having an exceptional condition and poll (2) indicates a POLLPRI
70       event.
72       Linux  2.4  introduced  a number of changes for improved throughput and
73       scaling, as well as enhanced functionality.   Some  of  these  features
74       include  support for zero-copy sendfile(2), Explicit Congestion Notifi‐
75       cation, new management of TIME_WAIT sockets, keep-alive socket  options
76       and support for Duplicate SACK extensions.
78   Address formats
79       TCP  is built on top of IP (see ip(7)).  The address formats defined by
80       ip(7) apply to TCP.  TCP supports  point-to-point  communication  only;
81       broadcasting and multicasting are not supported.
83   /proc interfaces
84       System-wide  TCP  parameter  settings  can  be accessed by files in the
85       directory /proc/sys/net/ipv4/.  In addition, most IP  /proc  interfaces
86       also  apply  to TCP; see ip(7).  Variables described as Boolean take an
87       integer value, with a nonzero value ("true") meaning  that  the  corre‐
88       sponding option is enabled, and a zero value ("false") meaning that the
89       option is disabled.
91       tcp_abc (Integer; default: 0; Linux 2.6.15 to Linux 3.8)
92              Control the Appropriate Byte Count (ABC), defined in  RFC  3465.
93              ABC  is  a  way  of increasing the congestion window (cwnd) more
94              slowly in response to partial acknowledgments.  Possible  values
95              are:
97              0  increase cwnd once per acknowledgment (no ABC)
99              1  increase cwnd once per acknowledgment of full sized segment
101              2  allow  increase  cwnd by two if acknowledgment is of two seg‐
102                 ments to compensate for delayed acknowledgments.
104       tcp_abort_on_overflow (Boolean; default: disabled; since Linux 2.4)
105              Enable resetting connections if the  listening  service  is  too
106              slow  and  unable  to keep up and accept them.  It means that if
107              overflow occurred due to a burst, the connection  will  recover.
108              Enable  this option only if you are really sure that the listen‐
109              ing  daemon  cannot  be  tuned  to  accept  connections  faster.
110              Enabling this option can harm the clients of your server.
112       tcp_adv_win_scale (integer; default: 2; since Linux 2.4)
113              Count   buffering   overhead  as  bytes/2^tcp_adv_win_scale,  if
114              tcp_adv_win_scale    is    greater    than    0;    or    bytes-
115              bytes/2^(-tcp_adv_win_scale),  if tcp_adv_win_scale is less than
116              or equal to zero.
118              The socket receive buffer space is shared between  the  applica‐
119              tion  and  kernel.   TCP maintains part of the buffer as the TCP
120              window, this is the size of the receive window advertised to the
121              other  end.   The rest of the space is used as the "application"
122              buffer, used to isolate the network from scheduling and applica‐
123              tion  latencies.   The  tcp_adv_win_scale  default  value  of  2
124              implies that the space used for the application  buffer  is  one
125              fourth that of the total.
127       tcp_allowed_congestion_control  (String; default: see text; since Linux
128       2.4.20)
129              Show/set the congestion control algorithm choices  available  to
130              unprivileged  processes  (see the description of the TCP_CONGES‐
131              TION socket option).  The items in the  list  are  separated  by
132              white  space and terminated by a newline character.  The list is
133              a subset of those  listed  in  tcp_available_congestion_control.
134              The  default value for this list is "reno" plus the default set‐
135              ting of tcp_congestion_control.
137       tcp_autocorking (Boolean; default: enabled; since Linux 3.14)
138              If this option is enabled, the kernel tries  to  coalesce  small
139              writes  (from consecutive write(2) and sendmsg(2) calls) as much
140              as possible, in order to decrease the total number of sent pack‐
141              ets.   Coalescing  is  done if at least one prior packet for the
142              flow is waiting  in  Qdisc  queues  or  device  transmit  queue.
143              Applications  can still use the TCP_CORK socket option to obtain
144              optimal behavior when they know how/when to uncork  their  sock‐
145              ets.
147       tcp_available_congestion_control   (String;   read-only;   since  Linux
148       2.4.20)
149              Show a list of the congestion-control algorithms that are regis‐
150              tered.   The  items in the list are separated by white space and
151              terminated by a newline character.  This list is a limiting  set
152              for  the  list  in tcp_allowed_congestion_control.  More conges‐
153              tion-control algorithms may be available  as  modules,  but  not
154              loaded.
156       tcp_app_win (integer; default: 31; since Linux 2.4)
157              This  variable  defines  how  many  bytes  of the TCP window are
158              reserved for buffering overhead.
160              A maximum of (window/2^tcp_app_win, mss) bytes in the window are
161              reserved  for the application buffer.  A value of 0 implies that
162              no amount is reserved.
164       tcp_base_mss (Integer; default: 512; since Linux 2.6.17)
165              The initial value of search_low to be used by the  packetization
166              layer  Path  MTU  discovery  (MTU  probing).   If MTU probing is
167              enabled, this is the initial MSS used by the connection.
169       tcp_bic (Boolean; default: disabled; Linux 2.4.27/2.6.6 to 2.6.13)
170              Enable BIC TCP  congestion  control  algorithm.   BIC-TCP  is  a
171              sender-side-only change that ensures a linear RTT fairness under
172              large windows while offering both scalability and  bounded  TCP-
173              friendliness.  The protocol combines two schemes called additive
174              increase and binary search increase.  When the congestion window
175              is  large, additive increase with a large increment ensures lin‐
176              ear RTT fairness as well as good scalability.  Under small  con‐
177              gestion  windows,  binary search increase provides TCP friendli‐
178              ness.
180       tcp_bic_low_window (integer; default: 14; Linux 2.4.27/2.6.6 to 2.6.13)
181              Set the threshold window (in packets) where BIC  TCP  starts  to
182              adjust  the  congestion  window.   Below  this threshold BIC TCP
183              behaves the same as the default TCP Reno.
185       tcp_bic_fast_convergence (Boolean; default: enabled; Linux 2.4.27/2.6.6
186       to 2.6.13)
187              Force  BIC  TCP to more quickly respond to changes in congestion
188              window.  Allows two flows sharing the same  connection  to  con‐
189              verge more rapidly.
191       tcp_congestion_control (String; default: see text; since Linux 2.4.13)
192              Set  the default congestion-control algorithm to be used for new
193              connections.  The algorithm  "reno"  is  always  available,  but
194              additional choices may be available depending on kernel configu‐
195              ration.  The default value for this file is set as part of  ker‐
196              nel configuration.
198       tcp_dma_copybreak (integer; default: 4096; since Linux 2.6.24)
199              Lower  limit, in bytes, of the size of socket reads that will be
200              offloaded to a DMA copy engine, if one is present in the  system
201              and the kernel was configured with the CONFIG_NET_DMA option.
203       tcp_dsack (Boolean; default: enabled; since Linux 2.4)
204              Enable RFC 2883 TCP Duplicate SACK support.
206       tcp_ecn (Integer; default: see below; since Linux 2.4)
207              Enable RFC 3168 Explicit Congestion Notification.
209              This file can have one of the following values:
211              0      Disable  ECN.  Neither initiate nor accept ECN.  This was
212                     the default up to and including Linux 2.6.30.
214              1      Enable ECN when requested  by  incoming  connections  and
215                     also request ECN on outgoing connection attempts.
217              2      Enable ECN when requested by incoming connections, but do
218                     not request ECN on outgoing connections.  This  value  is
219                     supported, and is the default, since Linux 2.6.31.
221              When   enabled,  connectivity  to  some  destinations  could  be
222              affected due to older, misbehaving middle boxes along the  path,
223              causing  connections  to be dropped.  However, to facilitate and
224              encourage deployment with option 1,  and  to  work  around  such
225              buggy  equipment,  the  tcp_ecn_fallback  option has been intro‐
226              duced.
228       tcp_ecn_fallback (Boolean; default: enabled; since Linux 4.1)
229              Enable RFC 3168, Section fallback.  When enabled,  out‐
230              going  ECN-setup  SYNs  that  time  out  within  the  normal SYN
231              retransmission timeout will be resent with CWR and ECE cleared.
233       tcp_fack (Boolean; default: enabled; since Linux 2.2)
234              Enable TCP Forward Acknowledgement support.
236       tcp_fin_timeout (integer; default: 60; since Linux 2.2)
237              This specifies how many seconds to wait for a final  FIN  packet
238              before the socket is forcibly closed.  This is strictly a viola‐
239              tion of the TCP specification, but required to  prevent  denial-
240              of-service attacks.  In Linux 2.2, the default value was 180.
242       tcp_frto (integer; default: see below; since Linux 2.4.21/2.6)
243              Enable F-RTO, an enhanced recovery algorithm for TCP retransmis‐
244              sion timeouts (RTOs).  It is particularly beneficial in wireless
245              environments  where packet loss is typically due to random radio
246              interference rather than intermediate  router  congestion.   See
247              RFC 4138 for more details.
249              This file can have one of the following values:
251              0  Disabled.   This  was  the  default up to and including Linux
252                 2.6.23.
254              1  The basic version F-RTO algorithm is enabled.
256              2  Enable SACK-enhanced F-RTO if flow uses SACK.  The basic ver‐
257                 sion can be used also when SACK is in use though in that case
258                 scenario(s) exists  where  F-RTO  interacts  badly  with  the
259                 packet  counting of the SACK-enabled TCP flow.  This value is
260                 the default since Linux 2.6.24.
262              Before Linux 2.6.22, this parameter was a  Boolean  value,  sup‐
263              porting just values 0 and 1 above.
265       tcp_frto_response (integer; default: 0; since Linux 2.6.22)
266              When  F-RTO  has  detected that a TCP retransmission timeout was
267              spurious (i.e., the timeout would have been avoided had TCP  set
268              a  longer  retransmission timeout), TCP has several options con‐
269              cerning what to do next.  Possible values are:
271              0  Rate halving  based;  a  smooth  and  conservative  response,
272                 results  in  halved  congestion  window (cwnd) and slow-start
273                 threshold (ssthresh) after one RTT.
275              1  Very conservative  response;  not  recommended  because  even
276                 though  being  valid,  it  interacts  poorly with the rest of
277                 Linux TCP; halves cwnd and ssthresh immediately.
279              2  Aggressive response; undoes congestion-control measures  that
280                 are  now known to be unnecessary (ignoring the possibility of
281                 a lost retransmission that would require TCP to be more  cau‐
282                 tious); cwnd and ssthresh are restored to the values prior to
283                 timeout.
285       tcp_keepalive_intvl (integer; default: 75; since Linux 2.4)
286              The number of seconds between TCP keep-alive probes.
288       tcp_keepalive_probes (integer; default: 9; since Linux 2.2)
289              The maximum number of TCP keep-alive probes to send before  giv‐
290              ing  up  and  killing  the connection if no response is obtained
291              from the other end.
293       tcp_keepalive_time (integer; default: 7200; since Linux 2.2)
294              The number of seconds a connection needs to be idle  before  TCP
295              begins sending out keep-alive probes.  Keep-alives are sent only
296              when the SO_KEEPALIVE socket option  is  enabled.   The  default
297              value  is  7200 seconds (2 hours).  An idle connection is termi‐
298              nated after approximately an additional 11 minutes (9 probes  an
299              interval of 75 seconds apart) when keep-alive is enabled.
301              Note that underlying connection tracking mechanisms and applica‐
302              tion timeouts may be much shorter.
304       tcp_low_latency (Boolean; default: disabled; since Linux 2.4.21/2.6)
305              If enabled, the TCP stack  makes  decisions  that  prefer  lower
306              latency as opposed to higher throughput.  It this option is dis‐
307              abled, then higher throughput is preferred.  An  example  of  an
308              application  where  this  default  should  be changed would be a
309              Beowulf compute cluster.
311       tcp_max_orphans (integer; default: see below; since Linux 2.4)
312              The maximum number of orphaned (not attached to  any  user  file
313              handle)  TCP sockets allowed in the system.  When this number is
314              exceeded, the orphaned connection is  reset  and  a  warning  is
315              printed.   This  limit  exists only to prevent simple denial-of-
316              service attacks.  Lowering this limit is not recommended.   Net‐
317              work  conditions  might  require  you  to increase the number of
318              orphans allowed, but note that each orphan can eat up to  ~64 kB
319              of  unswappable  memory.  The default initial value is set equal
320              to the  kernel  parameter  NR_FILE.   This  initial  default  is
321              adjusted depending on the memory in the system.
323       tcp_max_syn_backlog (integer; default: see below; since Linux 2.2)
324              The  maximum  number  of  queued  connection requests which have
325              still  not  received  an  acknowledgement  from  the  connecting
326              client.  If this number is exceeded, the kernel will begin drop‐
327              ping requests.  The default value of 256 is  increased  to  1024
328              when the memory present in the system is adequate or greater (>=
329              128 MB), and reduced to 128 for those systems with very low mem‐
330              ory (<= 32 MB).
332              Prior to Linux 2.6.20, it was recommended that if this needed to
333              be increased above 1024, the  size  of  the  SYNACK  hash  table
334              (TCP_SYNQ_HSIZE) in include/net/tcp.h should be modified to keep
336                  TCP_SYNQ_HSIZE * 16 <= tcp_max_syn_backlog
338              and the kernel should be recompiled.  In Linux 2.6.20, the fixed
339              sized TCP_SYNQ_HSIZE was removed in favor of dynamic sizing.
341       tcp_max_tw_buckets (integer; default: see below; since Linux 2.4)
342              The maximum number of sockets in TIME_WAIT state allowed in  the
343              system.  This limit exists only to prevent simple denial-of-ser‐
344              vice attacks.   The  default  value  of  NR_FILE*2  is  adjusted
345              depending  on  the  memory  in  the  system.   If this number is
346              exceeded, the socket is closed and a warning is printed.
348       tcp_moderate_rcvbuf   (Boolean;   default:   enabled;    since    Linux
349       2.4.17/2.6.7)
350              If  enabled, TCP performs receive buffer auto-tuning, attempting
351              to automatically size the buffer (no greater  than  tcp_rmem[2])
352              to match the size required by the path for full throughput.
354       tcp_mem (since Linux 2.4)
355              This  is  a  vector of 3 integers: [low, pressure, high].  These
356              bounds, measured in units of the system page size, are  used  by
357              TCP  to  track its memory usage.  The defaults are calculated at
358              boot time from the amount of available memory.   (TCP  can  only
359              use  low  memory  for  this,  which  is  limited  to  around 900
360              megabytes on 32-bit systems.  64-bit systems do not suffer  this
361              limitation.)
363              low       TCP  doesn't  regulate  its memory allocation when the
364                        number of pages it has  allocated  globally  is  below
365                        this number.
367              pressure  When  the  amount  of  memory allocated by TCP exceeds
368                        this number of pages, TCP moderates  its  memory  con‐
369                        sumption.   This  memory pressure state is exited once
370                        the number of pages  allocated  falls  below  the  low
371                        mark.
373              high      The  maximum  number of pages, globally, that TCP will
374                        allocate.   This  value  overrides  any  other  limits
375                        imposed by the kernel.
377       tcp_mtu_probing (integer; default: 0; since Linux 2.6.17)
378              This parameter controls TCP Packetization-Layer Path MTU Discov‐
379              ery.  The following values may be assigned to the file:
381              0  Disabled
383              1  Disabled by default, enabled when an ICMP black hole detected
385              2  Always enabled, use initial MSS of tcp_base_mss.
387       tcp_no_metrics_save (Boolean; default: disabled; since Linux 2.6.6)
388              By default, TCP saves various connection metrics  in  the  route
389              cache  when  the  connection  closes, so that connections estab‐
390              lished in the near future can use these to  set  initial  condi‐
391              tions.   Usually, this increases overall performance, but it may
392              sometimes cause performance degradation.  If tcp_no_metrics_save
393              is enabled, TCP will not cache metrics on closing connections.
395       tcp_orphan_retries (integer; default: 8; since Linux 2.4)
396              The  maximum number of attempts made to probe the other end of a
397              connection which has been closed by our end.
399       tcp_reordering (integer; default: 3; since Linux 2.4)
400              The maximum a packet can be reordered in  a  TCP  packet  stream
401              without  TCP assuming packet loss and going into slow start.  It
402              is not advisable to  change  this  number.   This  is  a  packet
403              reordering  detection  metric  designed  to minimize unnecessary
404              back off and retransmits provoked by reordering of packets on  a
405              connection.
407       tcp_retrans_collapse (Boolean; default: enabled; since Linux 2.2)
408              Try to send full-sized packets during retransmit.
410       tcp_retries1 (integer; default: 3; since Linux 2.2)
411              The  number  of times TCP will attempt to retransmit a packet on
412              an established connection normally, without the extra effort  of
413              getting the network layers involved.  Once we exceed this number
414              of retransmits, we first have the network layer update the route
415              if  possible before each new retransmit.  The default is the RFC
416              specified minimum of 3.
418       tcp_retries2 (integer; default: 15; since Linux 2.2)
419              The maximum number of times a TCP  packet  is  retransmitted  in
420              established  state  before  giving up.  The default value is 15,
421              which corresponds to a duration of approximately between  13  to
422              30  minutes,  depending  on  the  retransmission  timeout.   The
423              RFC 1122 specified minimum limit of  100  seconds  is  typically
424              deemed too short.
426       tcp_rfc1337 (Boolean; default: disabled; since Linux 2.2)
427              Enable TCP behavior conformant with RFC 1337.  When disabled, if
428              a RST is received in TIME_WAIT state, we close the socket  imme‐
429              diately without waiting for the end of the TIME_WAIT period.
431       tcp_rmem (since Linux 2.4)
432              This  is  a  vector  of  3 integers: [min, default, max].  These
433              parameters are used by TCP to  regulate  receive  buffer  sizes.
434              TCP  dynamically adjusts the size of the receive buffer from the
435              defaults listed below, in the range of these  values,  depending
436              on memory available in the system.
438              min       minimum  size  of  the receive buffer used by each TCP
439                        socket.  The default value is the  system  page  size.
440                        (On  Linux  2.4, the default value is 4 kB, lowered to
441                        PAGE_SIZE bytes in low-memory systems.)  This value is
442                        used  to  ensure that in memory pressure mode, alloca‐
443                        tions below this size will still succeed.  This is not
444                        used  to bound the size of the receive buffer declared
445                        using SO_RCVBUF on a socket.
447              default   the default size of  the  receive  buffer  for  a  TCP
448                        socket.   This  value  overwrites  the initial default
449                        buffer    size     from     the     generic     global
450                        net.core.rmem_default  defined for all protocols.  The
451                        default value is 87380 bytes.   (On  Linux  2.4,  this
452                        will  be  lowered to 43689 in low-memory systems.)  If
453                        larger receive buffer sizes are  desired,  this  value
454                        should  be  increased  (to  affect  all  sockets).  To
455                        employ  large  TCP  windows,   the   net.ipv4.tcp_win‐
456                        dow_scaling must be enabled (default).
458              max       the  maximum  size  of the receive buffer used by each
459                        TCP socket.  This value does not override  the  global
460                        net.core.rmem_max.  This is not used to limit the size
461                        of the receive buffer declared using  SO_RCVBUF  on  a
462                        socket.   The  default  value  is calculated using the
463                        formula
465                            max(87380, min(4 MB, tcp_mem[1]*PAGE_SIZE/128))
467                        (On Linux 2.4, the default is 87380*2  bytes,  lowered
468                        to 87380 in low-memory systems).
470       tcp_sack (Boolean; default: enabled; since Linux 2.2)
471              Enable RFC 2018 TCP Selective Acknowledgements.
473       tcp_slow_start_after_idle   (Boolean;  default:  enabled;  since  Linux
474       2.6.18)
475              If enabled, provide RFC 2861 behavior and time out  the  conges‐
476              tion  window after an idle period.  An idle period is defined as
477              the current RTO (retransmission timeout).  If disabled, the con‐
478              gestion window will not be timed out after an idle period.
480       tcp_stdurg (Boolean; default: disabled; since Linux 2.2)
481              If  this option is enabled, then use the RFC 1122 interpretation
482              of the TCP urgent-pointer field.  According to this  interpreta‐
483              tion, the urgent pointer points to the last byte of urgent data.
484              If this option is disabled, then use the  BSD-compatible  inter‐
485              pretation  of  the  urgent pointer: the urgent pointer points to
486              the first byte after the urgent data.  Enabling this option  may
487              lead to interoperability problems.
489       tcp_syn_retries (integer; default: 5; since Linux 2.2)
490              The  maximum number of times initial SYNs for an active TCP con‐
491              nection attempt will be retransmitted.  This value should not be
492              higher  than  255.  The default value is 5, which corresponds to
493              approximately 180 seconds.
495       tcp_synack_retries (integer; default: 5; since Linux 2.2)
496              The maximum number of times a SYN/ACK segment for a passive  TCP
497              connection  will  be  retransmitted.   This number should not be
498              higher than 255.
500       tcp_syncookies (Boolean; since Linux 2.2)
501              Enable TCP syncookies.  The kernel must be  compiled  with  CON‐
502              FIG_SYN_COOKIES.  Send out syncookies when the syn backlog queue
503              of a socket overflows.  The syncookies feature attempts to  pro‐
504              tect a socket from a SYN flood attack.  This should be used as a
505              last resort, if at all.  This is a violation of the  TCP  proto‐
506              col,  and  conflicts  with other areas of TCP such as TCP exten‐
507              sions.  It can cause problems for clients and relays.  It is not
508              recommended  as a tuning mechanism for heavily loaded servers to
509              help with overloaded or misconfigured  conditions.   For  recom‐
510              mended alternatives see tcp_max_syn_backlog, tcp_synack_retries,
511              and tcp_abort_on_overflow.
513       tcp_timestamps (integer; default: 1; since Linux 2.2)
514              Set to one of the following values to enable or disable RFC 1323
515              TCP timestamps:
517              0  Disable timestamps.
519              1  Enable timestamps as defined in RFC1323 and use random offset
520                 for each connection rather than only using the current time.
522              2  As for the value 1,  but  without  random  offsets.   Setting
523                 tcp_timestamps to this value is meaningful since Linux 4.10.
525       tcp_tso_win_divisor (integer; default: 3; since Linux 2.6.9)
526              This parameter controls what percentage of the congestion window
527              can be consumed by  a  single  TCP  Segmentation  Offload  (TSO)
528              frame.   The  setting  of  this  parameter is a tradeoff between
529              burstiness and building larger TSO frames.
531       tcp_tw_recycle (Boolean; default: disabled; Linux 2.4 to 4.11)
532              Enable fast  recycling  of  TIME_WAIT  sockets.   Enabling  this
533              option is not recommended as the remote IP may not use monotoni‐
534              cally increasing timestamps (devices behind  NAT,  devices  with
535              per-connection  timestamp offsets).  See RFC 1323 (PAWS) and RFC
536              6191.
538       tcp_tw_reuse (Boolean; default: disabled; since Linux 2.4.19/2.6)
539              Allow to reuse TIME_WAIT sockets for new connections when it  is
540              safe  from protocol viewpoint.  It should not be changed without
541              advice/request of technical experts.
543       tcp_vegas_cong_avoid (Boolean; default: disabled; Linux 2.2 to 2.6.13)
544              Enable TCP Vegas congestion avoidance algorithm.  TCP Vegas is a
545              sender-side-only  change  to  TCP  that anticipates the onset of
546              congestion by estimating the bandwidth.  TCP Vegas  adjusts  the
547              sending  rate  by  modifying  the  congestion window.  TCP Vegas
548              should provide less packet loss, but it is not as aggressive  as
549              TCP Reno.
551       tcp_westwood (Boolean; default: disabled; Linux 2.4.26/2.6.3 to 2.6.13)
552              Enable  TCP  Westwood+  congestion control algorithm.  TCP West‐
553              wood+ is a sender-side-only modification of the TCP Reno  proto‐
554              col  stack that optimizes the performance of TCP congestion con‐
555              trol.  It is based on end-to-end  bandwidth  estimation  to  set
556              congestion  window  and  slow start threshold after a congestion
557              episode.  Using this estimation, TCP Westwood+ adaptively sets a
558              slow  start  threshold  and a congestion window which takes into
559              account the bandwidth used at the  time  congestion  is  experi‐
560              enced.   TCP  Westwood+  significantly  increases  fairness with
561              respect to TCP Reno in wired networks and throughput over  wire‐
562              less links.
564       tcp_window_scaling (Boolean; default: enabled; since Linux 2.2)
565              Enable RFC 1323 TCP window scaling.  This feature allows the use
566              of a large window (> 64 kB) on  a  TCP  connection,  should  the
567              other  end support it.  Normally, the 16 bit window length field
568              in the TCP header limits the window size to less than 64 kB.  If
569              larger  windows  are desired, applications can increase the size
570              of their socket buffers and the window scaling  option  will  be
571              employed.  If tcp_window_scaling is disabled, TCP will not nego‐
572              tiate the use of window scaling with the other end  during  con‐
573              nection setup.
575       tcp_wmem (since Linux 2.4)
576              This  is  a  vector  of  3 integers: [min, default, max].  These
577              parameters are used by TCP to regulate send buffer  sizes.   TCP
578              dynamically adjusts the size of the send buffer from the default
579              values listed below, in the range of these values, depending  on
580              memory available.
582              min       Minimum  size  of  the  send  buffer  used by each TCP
583                        socket.  The default value is the  system  page  size.
584                        (On Linux 2.4, the default value is 4 kB.)  This value
585                        is used to ensure that in memory pressure mode,  allo‐
586                        cations  below  this size will still succeed.  This is
587                        not used to bound the size of the send buffer declared
588                        using SO_SNDBUF on a socket.
590              default   The  default size of the send buffer for a TCP socket.
591                        This value overwrites the initial default buffer  size
592                        from            the           generic           global
593                        /proc/sys/net/core/wmem_default defined for all proto‐
594                        cols.   The  default  value  is 16 kB.  If larger send
595                        buffer  sizes  are  desired,  this  value  should   be
596                        increased  (to  affect  all sockets).  To employ large
597                        TCP windows, the /proc/sys/net/ipv4/tcp_window_scaling
598                        must be set to a nonzero value (default).
600              max       The  maximum  size of the send buffer used by each TCP
601                        socket.  This value does not  override  the  value  in
602                        /proc/sys/net/core/wmem_max.   This  is  not  used  to
603                        limit the size  of  the  send  buffer  declared  using
604                        SO_SNDBUF  on  a  socket.  The default value is calcu‐
605                        lated using the formula
607                            max(65536, min(4 MB, tcp_mem[1]*PAGE_SIZE/128))
609                        (On Linux 2.4, the default value  is  128 kB,  lowered
610                        64 kB depending on low-memory systems.)
612       tcp_workaround_signed_windows  (Boolean; default: disabled; since Linux
613       2.6.26)
614              If enabled, assume that no receipt of  a  window-scaling  option
615              means  that  the remote TCP is broken and treats the window as a
616              signed quantity.  If disabled, assume that the remote TCP is not
617              broken  even  if  we do not receive a window scaling option from
618              it.
620   Socket options
621       To set or get a TCP socket option, call getsockopt(2) to read  or  set‐
622       sockopt(2)  to  write  the option with the option level argument set to
623       IPPROTO_TCP.  Unless otherwise noted, optval is a pointer  to  an  int.
624       In  addition,  most IPPROTO_IP socket options are valid on TCP sockets.
625       For more information see ip(7).
627       TCP_CONGESTION (since Linux 2.6.13)
628              The argument for this option is a string.   This  option  allows
629              the  caller  to  set  the TCP congestion control algorithm to be
630              used,  on  a  per-socket  basis.   Unprivileged  processes   are
631              restricted to choosing one of the algorithms in tcp_allowed_con‐
632              gestion_control   (described   above).    Privileged   processes
633              (CAP_NET_ADMIN) can choose from any of the available congestion-
634              control algorithms (see the description of tcp_available_conges‐
635              tion_control above).
637       TCP_CORK (since Linux 2.2)
638              If  set,  don't  send  out  partial  frames.  All queued partial
639              frames are sent when the option is cleared again.  This is  use‐
640              ful  for  prepending  headers before calling sendfile(2), or for
641              throughput optimization.  As currently implemented, there  is  a
642              200  millisecond  ceiling on the time for which output is corked
643              by TCP_CORK.  If this ceiling is reached, then  queued  data  is
644              automatically  transmitted.   This  option  can be combined with
645              TCP_NODELAY only since Linux 2.5.71.  This option should not  be
646              used in code intended to be portable.
648       TCP_DEFER_ACCEPT (since Linux 2.4)
649              Allow  a  listener  to be awakened only when data arrives on the
650              socket.  Takes an integer value (seconds), this  can  bound  the
651              maximum number of attempts TCP will make to complete the connec‐
652              tion.  This option should not be used in  code  intended  to  be
653              portable.
655       TCP_INFO (since Linux 2.4)
656              Used  to  collect  information  about  this  socket.  The kernel
657              returns   a   struct   tcp_info   as   defined   in   the   file
658              /usr/include/linux/tcp.h.   This  option  should  not be used in
659              code intended to be portable.
661       TCP_KEEPCNT (since Linux 2.4)
662              The maximum number of keepalive probes TCP  should  send  before
663              dropping the connection.  This option should not be used in code
664              intended to be portable.
666       TCP_KEEPIDLE (since Linux 2.4)
667              The time (in seconds) the connection needs to remain idle before
668              TCP  starts  sending  keepalive  probes,  if  the  socket option
669              SO_KEEPALIVE has been set on this socket.   This  option  should
670              not be used in code intended to be portable.
672       TCP_KEEPINTVL (since Linux 2.4)
673              The time (in seconds) between individual keepalive probes.  This
674              option should not be used in code intended to be portable.
676       TCP_LINGER2 (since Linux 2.4)
677              The lifetime of orphaned FIN_WAIT2 state sockets.   This  option
678              can  be  used  to  override  the system-wide setting in the file
679              /proc/sys/net/ipv4/tcp_fin_timeout for this socket.  This is not
680              to  be confused with the socket(7) level option SO_LINGER.  This
681              option should not be used in code intended to be portable.
683       TCP_MAXSEG
684              The maximum segment size for outgoing TCP packets.  In Linux 2.2
685              and  earlier,  and  in Linux 2.6.28 and later, if this option is
686              set before connection establishment, it  also  changes  the  MSS
687              value  announced to the other end in the initial packet.  Values
688              greater than the (eventual) interface MTU have no  effect.   TCP
689              will  also  impose its minimum and maximum bounds over the value
690              provided.
692       TCP_NODELAY
693              If set, disable the Nagle algorithm.  This means  that  segments
694              are  always  sent  as  soon as possible, even if there is only a
695              small amount of data.  When not  set,  data  is  buffered  until
696              there  is  a sufficient amount to send out, thereby avoiding the
697              frequent sending of small packets, which results  in  poor  uti‐
698              lization of the network.  This option is overridden by TCP_CORK;
699              however, setting this option forces an explicit flush of pending
700              output, even if TCP_CORK is currently set.
702       TCP_QUICKACK (since Linux 2.4.4)
703              Enable quickack mode if set or disable quickack mode if cleared.
704              In quickack mode, acks are sent immediately, rather than delayed
705              if  needed  in accordance to normal TCP operation.  This flag is
706              not permanent, it only enables a  switch  to  or  from  quickack
707              mode.   Subsequent operation of the TCP protocol will once again
708              enter/leave quickack mode depending on  internal  protocol  pro‐
709              cessing  and  factors such as delayed ack timeouts occurring and
710              data transfer.  This option should not be used in code  intended
711              to be portable.
713       TCP_SYNCNT (since Linux 2.4)
714              Set  the  number  of SYN retransmits that TCP should send before
715              aborting the attempt to connect.  It cannot  exceed  255.   This
716              option should not be used in code intended to be portable.
718       TCP_USER_TIMEOUT (since Linux 2.6.37)
719              This  option  takes  an  unsigned  int as an argument.  When the
720              value is greater than 0, it specifies the maximum amount of time
721              in  milliseconds that transmitted data may remain unacknowledged
722              before TCP will forcibly close the corresponding connection  and
723              return  ETIMEDOUT  to  the  application.  If the option value is
724              specified as 0, TCP will to use the system default.
726              Increasing user timeouts allows  a  TCP  connection  to  survive
727              extended  periods  without  end-to-end connectivity.  Decreasing
728              user timeouts allows applications to "fail fast", if so desired.
729              Otherwise,  failure  may  take up to 20 minutes with the current
730              system defaults in a normal WAN environment.
732              This option can be set during any state of a TCP connection, but
733              is effective only during the synchronized states of a connection
735              LAST-ACK).    Moreover,   when   used  with  the  TCP  keepalive
736              (SO_KEEPALIVE) option, TCP_USER_TIMEOUT will override  keepalive
737              to  determine  when to close a connection due to keepalive fail‐
738              ure.
740              The option has no effect on when TCP retransmits a  packet,  nor
741              when a keepalive probe is sent.
743              This  option,  like many others, will be inherited by the socket
744              returned by accept(2), if it was set on the listening socket.
746              Further details on the user timeout  feature  can  be  found  in
747              RFC 793 and RFC 5482 ("TCP User Timeout Option").
749       TCP_WINDOW_CLAMP (since Linux 2.4)
750              Bound the size of the advertised window to this value.  The ker‐
751              nel imposes a minimum size of  SOCK_MIN_RCVBUF/2.   This  option
752              should not be used in code intended to be portable.
754   Sockets API
755       TCP  provides  limited  support for out-of-band data, in the form of (a
756       single byte of) urgent data.  In Linux this  means  if  the  other  end
757       sends  newer out-of-band data the older urgent data is inserted as nor‐
758       mal data into the stream (even when SO_OOBINLINE  is  not  set).   This
759       differs from BSD-based stacks.
761       Linux  uses  the  BSD  compatible  interpretation of the urgent pointer
762       field by default.  This violates RFC 1122, but is required for interop‐
763       erability    with    other    stacks.     It   can   be   changed   via
764       /proc/sys/net/ipv4/tcp_stdurg.
766       It is possible to peek at out-of-band data using the  recv(2)  MSG_PEEK
767       flag.
769       Since  version  2.4,  Linux  supports the use of MSG_TRUNC in the flags
770       argument of recv(2) (and recvmsg(2)).  This flag  causes  the  received
771       bytes of data to be discarded, rather than passed back in a caller-sup‐
772       plied buffer.  Since Linux 2.4.4, MSG_TRUNC also has this  effect  when
773       used in conjunction with MSG_OOB to receive out-of-band data.
775   Ioctls
776       The  following ioctl(2) calls return information in value.  The correct
777       syntax is:
779              int value;
780              error = ioctl(tcp_socket, ioctl_type, &value);
782       ioctl_type is one of the following:
784       SIOCINQ
785              Returns the amount of queued unread data in the receive  buffer.
786              The socket must not be in LISTEN state, otherwise an error (EIN‐
787              VAL) is returned.   SIOCINQ  is  defined  in  <linux/sockios.h>.
788              Alternatively,  you  can use the synonymous FIONREAD, defined in
789              <sys/ioctl.h>.
791       SIOCATMARK
792              Returns true (i.e., value is nonzero) if the inbound data stream
793              is at the urgent mark.
795              If the SO_OOBINLINE socket option is set, and SIOCATMARK returns
796              true, then the next read from the socket will return the  urgent
797              data.  If the SO_OOBINLINE socket option is not set, and SIOCAT‐
798              MARK returns true, then the  next  read  from  the  socket  will
799              return the bytes following the urgent data (to actually read the
800              urgent data requires the recv(MSG_OOB) flag).
802              Note that a read never reads across  the  urgent  mark.   If  an
803              application  is  informed  of  the  presence  of urgent data via
804              select(2) (using the exceptfds argument) or through delivery  of
805              a SIGURG signal, then it can advance up to the mark using a loop
806              which repeatedly tests SIOCATMARK and performs a read  (request‐
807              ing any number of bytes) as long as SIOCATMARK returns false.
809       SIOCOUTQ
810              Returns the amount of unsent data in the socket send queue.  The
811              socket must not be in LISTEN state, otherwise an error  (EINVAL)
812              is  returned.  SIOCOUTQ is defined in <linux/sockios.h>.  Alter‐
813              natively, you  can  use  the  synonymous  TIOCOUTQ,  defined  in
814              <sys/ioctl.h>.
816   Error handling
817       When  a  network  error  occurs, TCP tries to resend the packet.  If it
818       doesn't succeed after some time, either ETIMEDOUT or the last  received
819       error on this connection is reported.
821       Some  applications  require  a quicker error notification.  This can be
822       enabled with the IPPROTO_IP level IP_RECVERR socket option.  When  this
823       option  is  enabled,  all incoming errors are immediately passed to the
824       user program.  Use this option with care — it makes TCP  less  tolerant
825       to routing changes and other normal network conditions.


829              Passed socket address type in sin_family was not AF_INET.
831       EPIPE  The  other  end closed the socket unexpectedly or a read is exe‐
832              cuted on a shut down socket.
834       ETIMEDOUT
835              The other end didn't acknowledge retransmitted data  after  some
836              time.
838       Any  errors  defined  for ip(7) or the generic socket layer may also be
839       returned for TCP.


842       Support for Explicit Congestion  Notification,  zero-copy  sendfile(2),
843       reordering  support and some SACK extensions (DSACK) were introduced in
844       2.4.  Support for forward acknowledgement (FACK), TIME_WAIT  recycling,
845       and per-connection keepalive socket options were introduced in 2.3.


848       Not all errors are documented.
849       IPv6 is not described.


852       accept(2),  bind(2),  connect(2), getsockopt(2), listen(2), recvmsg(2),
853       sendfile(2), sendmsg(2), socket(2), ip(7), socket(7)
855       RFC 793 for the TCP specification.
856       RFC 1122 for the TCP requirements and a description of the Nagle  algo‐
857       rithm.
858       RFC 1323 for TCP timestamp and window scaling options.
859       RFC 1337 for a description of TIME_WAIT assassination hazards.
860       RFC 3168 for a description of Explicit Congestion Notification.
861       RFC 2581 for TCP congestion control algorithms.
862       RFC 2018 and RFC 2883 for SACK and extensions to SACK.


865       This  page  is  part of release 4.15 of the Linux man-pages project.  A
866       description of the project, information about reporting bugs,  and  the
867       latest     version     of     this    page,    can    be    found    at
868       https://www.kernel.org/doc/man-pages/.
872Linux                             2017-09-15                            TCP(7)