aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
-rw-r--r--Documentation/networking/ip-sysctl.txt355
1 files changed, 177 insertions, 178 deletions
diff --git a/Documentation/networking/ip-sysctl.txt b/Documentation/networking/ip-sysctl.txt
index bbcc8deda172..a0f6842368c3 100644
--- a/Documentation/networking/ip-sysctl.txt
+++ b/Documentation/networking/ip-sysctl.txt
@@ -101,6 +101,11 @@ inet_peer_gc_maxtime - INTEGER
101 101
102TCP variables: 102TCP variables:
103 103
104somaxconn - INTEGER
105 Limit of socket listen() backlog, known in userspace as SOMAXCONN.
106 Defaults to 128. See also tcp_max_syn_backlog for additional tuning
107 for TCP sockets.
108
104tcp_abc - INTEGER 109tcp_abc - INTEGER
105 Controls Appropriate Byte Count (ABC) defined in RFC3465. 110 Controls Appropriate Byte Count (ABC) defined in RFC3465.
106 ABC is a way of increasing congestion window (cwnd) more slowly 111 ABC is a way of increasing congestion window (cwnd) more slowly
@@ -112,48 +117,51 @@ tcp_abc - INTEGER
112 of two segments to compensate for delayed acknowledgments. 117 of two segments to compensate for delayed acknowledgments.
113 Default: 0 (off) 118 Default: 0 (off)
114 119
115tcp_syn_retries - INTEGER 120tcp_abort_on_overflow - BOOLEAN
116 Number of times initial SYNs for an active TCP connection attempt 121 If listening service is too slow to accept new connections,
117 will be retransmitted. Should not be higher than 255. Default value 122 reset them. Default state is FALSE. It means that if overflow
118 is 5, which corresponds to ~180seconds. 123 occurred due to a burst, connection will recover. Enable this
124 option _only_ if you are really sure that listening daemon
125 cannot be tuned to accept connections faster. Enabling this
126 option can harm clients of your server.
119 127
120tcp_synack_retries - INTEGER 128tcp_adv_win_scale - INTEGER
121 Number of times SYNACKs for a passive TCP connection attempt will 129 Count buffering overhead as bytes/2^tcp_adv_win_scale
122 be retransmitted. Should not be higher than 255. Default value 130 (if tcp_adv_win_scale > 0) or bytes-bytes/2^(-tcp_adv_win_scale),
123 is 5, which corresponds to ~180seconds. 131 if it is <= 0.
132 Default: 2
124 133
125tcp_keepalive_time - INTEGER 134tcp_allowed_congestion_control - STRING
126 How often TCP sends out keepalive messages when keepalive is enabled. 135 Show/set the congestion control choices available to non-privileged
127 Default: 2hours. 136 processes. The list is a subset of those listed in
137 tcp_available_congestion_control.
138 Default is "reno" and the default setting (tcp_congestion_control).
128 139
129tcp_keepalive_probes - INTEGER 140tcp_app_win - INTEGER
130 How many keepalive probes TCP sends out, until it decides that the 141 Reserve max(window/2^tcp_app_win, mss) of window for application
131 connection is broken. Default value: 9. 142 buffer. Value 0 is special, it means that nothing is reserved.
143 Default: 31
132 144
133tcp_keepalive_intvl - INTEGER 145tcp_available_congestion_control - STRING
134 How frequently the probes are send out. Multiplied by 146 Shows the available congestion control choices that are registered.
135 tcp_keepalive_probes it is time to kill not responding connection, 147 More congestion control algorithms may be available as modules,
136 after probes started. Default value: 75sec i.e. connection 148 but not loaded.
137 will be aborted after ~11 minutes of retries.
138 149
139tcp_retries1 - INTEGER 150tcp_congestion_control - STRING
140 How many times to retry before deciding that something is wrong 151 Set the congestion control algorithm to be used for new
141 and it is necessary to report this suspicion to network layer. 152 connections. The algorithm "reno" is always available, but
142 Minimal RFC value is 3, it is default, which corresponds 153 additional choices may be available based on kernel configuration.
143 to ~3sec-8min depending on RTO. 154 Default is set as part of kernel configuration.
144 155
145tcp_retries2 - INTEGER 156tcp_dsack - BOOLEAN
146 How may times to retry before killing alive TCP connection. 157 Allows TCP to send "duplicate" SACKs.
147 RFC1122 says that the limit should be longer than 100 sec.
148 It is too small number. Default value 15 corresponds to ~13-30min
149 depending on RTO.
150 158
151tcp_orphan_retries - INTEGER 159tcp_ecn - BOOLEAN
152 How may times to retry before killing TCP connection, closed 160 Enable Explicit Congestion Notification in TCP.
153 by our side. Default value 7 corresponds to ~50sec-16min 161
154 depending on RTO. If you machine is loaded WEB server, 162tcp_fack - BOOLEAN
155 you should think about lowering this value, such sockets 163 Enable FACK congestion avoidance and fast retransmission.
156 may consume significant resources. Cf. tcp_max_orphans. 164 The value is not used, if tcp_sack is not enabled.
157 165
158tcp_fin_timeout - INTEGER 166tcp_fin_timeout - INTEGER
159 Time to hold socket in state FIN-WAIT-2, if it was closed 167 Time to hold socket in state FIN-WAIT-2, if it was closed
@@ -166,24 +174,33 @@ tcp_fin_timeout - INTEGER
166 because they eat maximum 1.5K of memory, but they tend 174 because they eat maximum 1.5K of memory, but they tend
167 to live longer. Cf. tcp_max_orphans. 175 to live longer. Cf. tcp_max_orphans.
168 176
169tcp_max_tw_buckets - INTEGER 177tcp_frto - BOOLEAN
170 Maximal number of timewait sockets held by system simultaneously. 178 Enables F-RTO, an enhanced recovery algorithm for TCP retransmission
171 If this number is exceeded time-wait socket is immediately destroyed 179 timeouts. It is particularly beneficial in wireless environments
172 and warning is printed. This limit exists only to prevent 180 where packet loss is typically due to random radio interference
173 simple DoS attacks, you _must_ not lower the limit artificially, 181 rather than intermediate router congestion.
174 but rather increase it (probably, after increasing installed memory),
175 if network conditions require more than default value.
176 182
177tcp_tw_recycle - BOOLEAN 183tcp_keepalive_time - INTEGER
178 Enable fast recycling TIME-WAIT sockets. Default value is 0. 184 How often TCP sends out keepalive messages when keepalive is enabled.
179 It should not be changed without advice/request of technical 185 Default: 2hours.
180 experts.
181 186
182tcp_tw_reuse - BOOLEAN 187tcp_keepalive_probes - INTEGER
183 Allow to reuse TIME-WAIT sockets for new connections when it is 188 How many keepalive probes TCP sends out, until it decides that the
184 safe from protocol viewpoint. Default value is 0. 189 connection is broken. Default value: 9.
185 It should not be changed without advice/request of technical 190
186 experts. 191tcp_keepalive_intvl - INTEGER
192 How frequently the probes are send out. Multiplied by
193 tcp_keepalive_probes it is time to kill not responding connection,
194 after probes started. Default value: 75sec i.e. connection
195 will be aborted after ~11 minutes of retries.
196
197tcp_low_latency - BOOLEAN
198 If set, the TCP stack makes decisions that prefer lower
199 latency as opposed to higher throughput. By default, this
200 option is not set meaning that higher throughput is preferred.
201 An example of an application where this default should be
202 changed would be a Beowulf compute cluster.
203 Default: 0
187 204
188tcp_max_orphans - INTEGER 205tcp_max_orphans - INTEGER
189 Maximal number of TCP sockets not attached to any user file handle, 206 Maximal number of TCP sockets not attached to any user file handle,
@@ -197,41 +214,6 @@ tcp_max_orphans - INTEGER
197 more aggressively. Let me to remind again: each orphan eats 214 more aggressively. Let me to remind again: each orphan eats
198 up to ~64K of unswappable memory. 215 up to ~64K of unswappable memory.
199 216
200tcp_abort_on_overflow - BOOLEAN
201 If listening service is too slow to accept new connections,
202 reset them. Default state is FALSE. It means that if overflow
203 occurred due to a burst, connection will recover. Enable this
204 option _only_ if you are really sure that listening daemon
205 cannot be tuned to accept connections faster. Enabling this
206 option can harm clients of your server.
207
208tcp_syncookies - BOOLEAN
209 Only valid when the kernel was compiled with CONFIG_SYNCOOKIES
210 Send out syncookies when the syn backlog queue of a socket
211 overflows. This is to prevent against the common 'syn flood attack'
212 Default: FALSE
213
214 Note, that syncookies is fallback facility.
215 It MUST NOT be used to help highly loaded servers to stand
216 against legal connection rate. If you see synflood warnings
217 in your logs, but investigation shows that they occur
218 because of overload with legal connections, you should tune
219 another parameters until this warning disappear.
220 See: tcp_max_syn_backlog, tcp_synack_retries, tcp_abort_on_overflow.
221
222 syncookies seriously violate TCP protocol, do not allow
223 to use TCP extensions, can result in serious degradation
224 of some services (f.e. SMTP relaying), visible not by you,
225 but your clients and relays, contacting you. While you see
226 synflood warnings in logs not being really flooded, your server
227 is seriously misconfigured.
228
229tcp_stdurg - BOOLEAN
230 Use the Host requirements interpretation of the TCP urg pointer field.
231 Most hosts use the older BSD interpretation, so if you turn this on
232 Linux might not communicate correctly with them.
233 Default: FALSE
234
235tcp_max_syn_backlog - INTEGER 217tcp_max_syn_backlog - INTEGER
236 Maximal number of remembered connection requests, which are 218 Maximal number of remembered connection requests, which are
237 still did not receive an acknowledgment from connecting client. 219 still did not receive an acknowledgment from connecting client.
@@ -239,24 +221,34 @@ tcp_max_syn_backlog - INTEGER
239 and 128 for low memory machines. If server suffers of overload, 221 and 128 for low memory machines. If server suffers of overload,
240 try to increase this number. 222 try to increase this number.
241 223
242tcp_window_scaling - BOOLEAN 224tcp_max_tw_buckets - INTEGER
243 Enable window scaling as defined in RFC1323. 225 Maximal number of timewait sockets held by system simultaneously.
226 If this number is exceeded time-wait socket is immediately destroyed
227 and warning is printed. This limit exists only to prevent
228 simple DoS attacks, you _must_ not lower the limit artificially,
229 but rather increase it (probably, after increasing installed memory),
230 if network conditions require more than default value.
244 231
245tcp_timestamps - BOOLEAN 232tcp_mem - vector of 3 INTEGERs: min, pressure, max
246 Enable timestamps as defined in RFC1323. 233 min: below this number of pages TCP is not bothered about its
234 memory appetite.
247 235
248tcp_sack - BOOLEAN 236 pressure: when amount of memory allocated by TCP exceeds this number
249 Enable select acknowledgments (SACKS). 237 of pages, TCP moderates its memory consumption and enters memory
238 pressure mode, which is exited when memory consumption falls
239 under "min".
250 240
251tcp_fack - BOOLEAN 241 max: number of pages allowed for queueing by all TCP sockets.
252 Enable FACK congestion avoidance and fast retransmission.
253 The value is not used, if tcp_sack is not enabled.
254 242
255tcp_dsack - BOOLEAN 243 Defaults are calculated at boot time from amount of available
256 Allows TCP to send "duplicate" SACKs. 244 memory.
257 245
258tcp_ecn - BOOLEAN 246tcp_orphan_retries - INTEGER
259 Enable Explicit Congestion Notification in TCP. 247 How may times to retry before killing TCP connection, closed
248 by our side. Default value 7 corresponds to ~50sec-16min
249 depending on RTO. If you machine is loaded WEB server,
250 you should think about lowering this value, such sockets
251 may consume significant resources. Cf. tcp_max_orphans.
260 252
261tcp_reordering - INTEGER 253tcp_reordering - INTEGER
262 Maximal reordering of packets in a TCP stream. 254 Maximal reordering of packets in a TCP stream.
@@ -267,20 +259,23 @@ tcp_retrans_collapse - BOOLEAN
267 On retransmit try to send bigger packets to work around bugs in 259 On retransmit try to send bigger packets to work around bugs in
268 certain TCP stacks. 260 certain TCP stacks.
269 261
270tcp_wmem - vector of 3 INTEGERs: min, default, max 262tcp_retries1 - INTEGER
271 min: Amount of memory reserved for send buffers for TCP socket. 263 How many times to retry before deciding that something is wrong
272 Each TCP socket has rights to use it due to fact of its birth. 264 and it is necessary to report this suspicion to network layer.
273 Default: 4K 265 Minimal RFC value is 3, it is default, which corresponds
266 to ~3sec-8min depending on RTO.
274 267
275 default: Amount of memory allowed for send buffers for TCP socket 268tcp_retries2 - INTEGER
276 by default. This value overrides net.core.wmem_default used 269 How may times to retry before killing alive TCP connection.
277 by other protocols, it is usually lower than net.core.wmem_default. 270 RFC1122 says that the limit should be longer than 100 sec.
278 Default: 16K 271 It is too small number. Default value 15 corresponds to ~13-30min
272 depending on RTO.
279 273
280 max: Maximal amount of memory allowed for automatically selected 274tcp_rfc1337 - BOOLEAN
281 send buffers for TCP socket. This value does not override 275 If set, the TCP stack behaves conforming to RFC1337. If unset,
282 net.core.wmem_max, "static" selection via SO_SNDBUF does not use this. 276 we are not conforming to RFC, but prevent TCP TIME_WAIT
283 Default: 128K 277 assassination.
278 Default: 0
284 279
285tcp_rmem - vector of 3 INTEGERs: min, default, max 280tcp_rmem - vector of 3 INTEGERs: min, default, max
286 min: Minimal size of receive buffer used by TCP sockets. 281 min: Minimal size of receive buffer used by TCP sockets.
@@ -299,79 +294,91 @@ tcp_rmem - vector of 3 INTEGERs: min, default, max
299 net.core.rmem_max, "static" selection via SO_RCVBUF does not use this. 294 net.core.rmem_max, "static" selection via SO_RCVBUF does not use this.
300 Default: 87380*2 bytes. 295 Default: 87380*2 bytes.
301 296
302tcp_mem - vector of 3 INTEGERs: min, pressure, max 297tcp_sack - BOOLEAN
303 min: below this number of pages TCP is not bothered about its 298 Enable select acknowledgments (SACKS).
304 memory appetite.
305 299
306 pressure: when amount of memory allocated by TCP exceeds this number 300tcp_slow_start_after_idle - BOOLEAN
307 of pages, TCP moderates its memory consumption and enters memory 301 If set, provide RFC2861 behavior and time out the congestion
308 pressure mode, which is exited when memory consumption falls 302 window after an idle period. An idle period is defined at
309 under "min". 303 the current RTO. If unset, the congestion window will not
304 be timed out after an idle period.
305 Default: 1
310 306
311 max: number of pages allowed for queueing by all TCP sockets. 307tcp_stdurg - BOOLEAN
308 Use the Host requirements interpretation of the TCP urg pointer field.
309 Most hosts use the older BSD interpretation, so if you turn this on
310 Linux might not communicate correctly with them.
311 Default: FALSE
312 312
313 Defaults are calculated at boot time from amount of available 313tcp_synack_retries - INTEGER
314 memory. 314 Number of times SYNACKs for a passive TCP connection attempt will
315 be retransmitted. Should not be higher than 255. Default value
316 is 5, which corresponds to ~180seconds.
315 317
316tcp_app_win - INTEGER 318tcp_syncookies - BOOLEAN
317 Reserve max(window/2^tcp_app_win, mss) of window for application 319 Only valid when the kernel was compiled with CONFIG_SYNCOOKIES
318 buffer. Value 0 is special, it means that nothing is reserved. 320 Send out syncookies when the syn backlog queue of a socket
319 Default: 31 321 overflows. This is to prevent against the common 'syn flood attack'
322 Default: FALSE
320 323
321tcp_adv_win_scale - INTEGER 324 Note, that syncookies is fallback facility.
322 Count buffering overhead as bytes/2^tcp_adv_win_scale 325 It MUST NOT be used to help highly loaded servers to stand
323 (if tcp_adv_win_scale > 0) or bytes-bytes/2^(-tcp_adv_win_scale), 326 against legal connection rate. If you see synflood warnings
324 if it is <= 0. 327 in your logs, but investigation shows that they occur
325 Default: 2 328 because of overload with legal connections, you should tune
329 another parameters until this warning disappear.
330 See: tcp_max_syn_backlog, tcp_synack_retries, tcp_abort_on_overflow.
326 331
327tcp_rfc1337 - BOOLEAN 332 syncookies seriously violate TCP protocol, do not allow
328 If set, the TCP stack behaves conforming to RFC1337. If unset, 333 to use TCP extensions, can result in serious degradation
329 we are not conforming to RFC, but prevent TCP TIME_WAIT 334 of some services (f.e. SMTP relaying), visible not by you,
330 assassination. 335 but your clients and relays, contacting you. While you see
331 Default: 0 336 synflood warnings in logs not being really flooded, your server
337 is seriously misconfigured.
332 338
333tcp_low_latency - BOOLEAN 339tcp_syn_retries - INTEGER
334 If set, the TCP stack makes decisions that prefer lower 340 Number of times initial SYNs for an active TCP connection attempt
335 latency as opposed to higher throughput. By default, this 341 will be retransmitted. Should not be higher than 255. Default value
336 option is not set meaning that higher throughput is preferred. 342 is 5, which corresponds to ~180seconds.
337 An example of an application where this default should be 343
338 changed would be a Beowulf compute cluster. 344tcp_timestamps - BOOLEAN
339 Default: 0 345 Enable timestamps as defined in RFC1323.
340 346
341tcp_tso_win_divisor - INTEGER 347tcp_tso_win_divisor - INTEGER
342 This allows control over what percentage of the congestion window 348 This allows control over what percentage of the congestion window
343 can be consumed by a single TSO frame. 349 can be consumed by a single TSO frame.
344 The setting of this parameter is a choice between burstiness and 350 The setting of this parameter is a choice between burstiness and
345 building larger TSO frames. 351 building larger TSO frames.
346 Default: 3 352 Default: 3
347 353
348tcp_frto - BOOLEAN 354tcp_tw_recycle - BOOLEAN
349 Enables F-RTO, an enhanced recovery algorithm for TCP retransmission 355 Enable fast recycling TIME-WAIT sockets. Default value is 0.
350 timeouts. It is particularly beneficial in wireless environments 356 It should not be changed without advice/request of technical
351 where packet loss is typically due to random radio interference 357 experts.
352 rather than intermediate router congestion.
353 358
354tcp_allowed_congestion_control - STRING 359tcp_tw_reuse - BOOLEAN
355 Show/set the congestion control choices available to non-privileged 360 Allow to reuse TIME-WAIT sockets for new connections when it is
356 processes. The list is a subset of those listed in 361 safe from protocol viewpoint. Default value is 0.
357 tcp_available_congestion_control. 362 It should not be changed without advice/request of technical
358 Default is "reno" and the default setting (tcp_congestion_control). 363 experts.
359 364
360tcp_available_congestion_control - STRING 365tcp_window_scaling - BOOLEAN
361 Shows the available congestion control choices that are registered. 366 Enable window scaling as defined in RFC1323.
362 More congestion control algorithms may be available as modules,
363 but not loaded.
364 367
365tcp_congestion_control - STRING 368tcp_wmem - vector of 3 INTEGERs: min, default, max
366 Set the congestion control algorithm to be used for new 369 min: Amount of memory reserved for send buffers for TCP socket.
367 connections. The algorithm "reno" is always available, but 370 Each TCP socket has rights to use it due to fact of its birth.
368 additional choices may be available based on kernel configuration. 371 Default: 4K
369 Default is set as part of kernel configuration.
370 372
371somaxconn - INTEGER 373 default: Amount of memory allowed for send buffers for TCP socket
372 Limit of socket listen() backlog, known in userspace as SOMAXCONN. 374 by default. This value overrides net.core.wmem_default used
373 Defaults to 128. See also tcp_max_syn_backlog for additional tuning 375 by other protocols, it is usually lower than net.core.wmem_default.
374 for TCP sockets. 376 Default: 16K
377
378 max: Maximal amount of memory allowed for automatically selected
379 send buffers for TCP socket. This value does not override
380 net.core.wmem_max, "static" selection via SO_SNDBUF does not use this.
381 Default: 128K
375 382
376tcp_workaround_signed_windows - BOOLEAN 383tcp_workaround_signed_windows - BOOLEAN
377 If set, assume no receipt of a window scaling option means the 384 If set, assume no receipt of a window scaling option means the
@@ -380,13 +387,6 @@ tcp_workaround_signed_windows - BOOLEAN
380 not receive a window scaling option from them. 387 not receive a window scaling option from them.
381 Default: 0 388 Default: 0
382 389
383tcp_slow_start_after_idle - BOOLEAN
384 If set, provide RFC2861 behavior and time out the congestion
385 window after an idle period. An idle period is defined at
386 the current RTO. If unset, the congestion window will not
387 be timed out after an idle period.
388 Default: 1
389
390CIPSOv4 Variables: 390CIPSOv4 Variables:
391 391
392cipso_cache_enable - BOOLEAN 392cipso_cache_enable - BOOLEAN
@@ -986,4 +986,3 @@ no_cong_thresh FIXME
986slot_timeout FIXME 986slot_timeout FIXME
987warn_noreply_time FIXME 987warn_noreply_time FIXME
988 988
989$Id: ip-sysctl.txt,v 1.20 2001/12/13 09:00:18 davem Exp $