11553 Want pluggable TCP congestion control algorithms
Portions contributed by: Cody Peter Mello <cody.mello@joyent.com>
Reviewed by: Dan McDonald <danmcd@joyent.com>
Reviewed by: Robert Mustacchi <robert.mustacchi@joyent.com>
1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2019 Joyent, Inc.
24 * Copyright (c) 2013, OmniTI Computer Consulting, Inc. All rights reserved.
25 * Copyright (c) 2013, 2016 by Delphix. All rights reserved.
26 */
27
28 #ifndef _INET_TCP_IMPL_H
29 #define _INET_TCP_IMPL_H
30
31 /*
32 * TCP implementation private declarations. These interfaces are
33 * used to build the IP module and are not meant to be accessed
34 * by any modules except IP itself. They are undocumented and are
35 * subject to change without notice.
36 */
37
38 #ifdef __cplusplus
39 extern "C" {
40 #endif
41
42 #ifdef _KERNEL
43
44 #include <sys/cpuvar.h>
45 #include <sys/clock_impl.h> /* For LBOLT_FASTPATH{,64} */
46 #include <inet/optcom.h>
47 #include <inet/tcp.h>
48 #include <inet/tunables.h>
49
50 #define TCP_MOD_ID 5105
51
52 extern struct qinit tcp_sock_winit;
53 extern struct qinit tcp_winit;
54
55 extern sock_downcalls_t sock_tcp_downcalls;
56
57 /*
58 * Note that by default, the _snd_lowat_fraction tunable controls the value of
59 * the transmit low water mark. TCP_XMIT_LOWATER (and thus the _xmit_lowat
60 * property) is only used if the administrator has disabled _snd_lowat_fraction
61 * by setting it to 0.
62 */
63 #define TCP_XMIT_LOWATER 4096
64 #define TCP_XMIT_HIWATER 49152
65 #define TCP_RECV_LOWATER 2048
66 #define TCP_RECV_HIWATER 128000
67
68 /*
69 * Bind hash list size and has function. It has to be a power of 2 for
70 * hashing.
71 */
72 #define TCP_BIND_FANOUT_SIZE 1024
73 #define TCP_BIND_HASH(lport) (ntohs(lport) & (TCP_BIND_FANOUT_SIZE - 1))
74
75 /*
76 * This implementation follows the 4.3BSD interpretation of the urgent
77 * pointer and not RFC 1122. Switching to RFC 1122 behavior would cause
78 * incompatible changes in protocols like telnet and rlogin.
79 */
80 #define TCP_OLD_URP_INTERPRETATION 1
81
82 /* TCP option length */
83 #define TCPOPT_NOP_LEN 1
84 #define TCPOPT_MAXSEG_LEN 4
85 #define TCPOPT_WS_LEN 3
86 #define TCPOPT_REAL_WS_LEN (TCPOPT_WS_LEN+1)
87 #define TCPOPT_TSTAMP_LEN 10
88 #define TCPOPT_REAL_TS_LEN (TCPOPT_TSTAMP_LEN+2)
89 #define TCPOPT_SACK_OK_LEN 2
90 #define TCPOPT_REAL_SACK_OK_LEN (TCPOPT_SACK_OK_LEN+2)
91 #define TCPOPT_REAL_SACK_LEN 4
92 #define TCPOPT_MAX_SACK_LEN 36
93 #define TCPOPT_HEADER_LEN 2
94
95 /* Round up the value to the nearest mss. */
96 #define MSS_ROUNDUP(value, mss) ((((value) - 1) / (mss) + 1) * (mss))
97
98 /*
99 * Was this tcp created via socket() interface?
100 */
101 #define TCP_IS_SOCKET(tcp) ((tcp)->tcp_issocket)
102
103 /*
104 * Is this tcp not attached to any upper client?
105 */
106 #define TCP_IS_DETACHED(tcp) ((tcp)->tcp_detached)
107
108 /* TCP timers related data structures. Refer to tcp_timers.c. */
109 typedef struct tcp_timer_s {
110 conn_t *connp;
111 void (*tcpt_proc)(void *);
112 callout_id_t tcpt_tid;
113 } tcp_timer_t;
114
115 extern kmem_cache_t *tcp_timercache;
116
117 /*
118 * Macro for starting various timers. Retransmission timer has its own macro,
119 * TCP_TIMER_RESTART(). tim is in millisec.
120 */
121 #define TCP_TIMER(tcp, f, tim) \
122 tcp_timeout(tcp->tcp_connp, f, tim)
123 #define TCP_TIMER_CANCEL(tcp, id) \
124 tcp_timeout_cancel(tcp->tcp_connp, id)
125
126 /*
127 * To restart the TCP retransmission timer. intvl is in millisec.
128 */
129 #define TCP_TIMER_RESTART(tcp, intvl) { \
130 if ((tcp)->tcp_timer_tid != 0) \
131 (void) TCP_TIMER_CANCEL((tcp), (tcp)->tcp_timer_tid); \
132 (tcp)->tcp_timer_tid = TCP_TIMER((tcp), tcp_timer, (intvl)); \
133 }
134
135
136 /*
137 * Maximum TIME_WAIT timeout. It is defined here (instead of tcp_tunables.c)
138 * so that other parameters can be derived from it.
139 */
140 #define TCP_TIME_WAIT_MAX (10 * MINUTES)
141
142 /*
143 * TCP_TIME_WAIT_DELAY governs how often the time_wait_collector runs.
144 * Running it every 5 seconds seems to yield a reasonable balance between
145 * cleanup liveliness and system load.
146 */
147 #define TCP_TIME_WAIT_DELAY (5 * SECONDS)
148
149 #define TCP_TIME_WAIT_BUCKETS ((TCP_TIME_WAIT_MAX / TCP_TIME_WAIT_DELAY) + 1)
150
151 /*
152 * For scalability, we must not run a timer for every TCP connection
153 * in TIME_WAIT state. To see why, consider (for time wait interval of
154 * 1 minutes):
155 * 10,000 connections/sec * 60 seconds/time wait = 600,000 active conn's
156 *
157 * Since TIME_WAIT expiration occurs on a per-squeue basis, handling
158 * connections from all netstacks on the system, a simple queue is inadequate
159 * for pending entries. This is because tcp_time_wait_interval may differ
160 * between connections, causing tail insertion to violate expiration order.
161 *
162 * Instead of performing expensive sorting or unnecessary list traversal to
163 * counteract interval variance between netstacks, a timing wheel structure is
164 * used. The duration covered by each bucket in the wheel is determined by the
165 * TCP_TIME_WAIT_DELAY (5 seconds). The number of buckets in the wheel is
166 * determined by dividing the maximum TIME_WAIT interval (10 minutes) by
167 * TCP_TIME_WAIT_DELAY, with one added bucket for rollover protection.
168 * (Yielding 121 buckets with the current parameters) When items are inserted
169 * into the set of buckets, they are indexed by using their expiration time
170 * divided by the bucket size, modulo the number of buckets. This means that
171 * when each bucket is processed, all items within should have expired within
172 * the last TCP_TIME_WAIT_DELAY interval.
173 *
174 * Since bucket timer schedules are rounded to the nearest TCP_TIME_WAIT_DELAY
175 * interval to ensure all connections in the pending bucket will be expired, a
176 * per-squeue offset is used when doing TIME_WAIT scheduling. This offset is
177 * between 0 and the TCP_TIME_WAIT_DELAY and is designed to avoid scheduling
178 * all of the tcp_time_wait_collector threads to run in lock-step. The offset
179 * is fixed while there are any connections present in the buckets.
180 *
181 * When a tcp_t enters TIME_WAIT state, a timer is started (timeout is
182 * tcps_time_wait_interval). When the tcp_t is detached (upper layer closes
183 * the end point), it is scheduled to be cleaned up by the squeue-driving
184 * tcp_time_wait_collector (also using tcps_time_wait_interval). This means
185 * that the TIME_WAIT state can be extended (up to doubled) if the tcp_t
186 * doesn't become detached for a long time.
187 *
188 * The list manipulations (including tcp_time_wait_next/prev)
189 * are protected by the tcp_time_wait_lock. The content of the
190 * detached TIME_WAIT connections is protected by the normal perimeters.
191 *
192 * These connection lists are per squeue and squeues are shared across the
193 * tcp_stack_t instances. Things in a tcp_time_wait_bucket remain associated
194 * with the tcp_stack_t and conn_netstack. Any tcp_t connections stored in the
195 * tcp_free_list are disassociated and have NULL tcp_tcps and conn_netstack
196 * pointers.
197 */
198 typedef struct tcp_squeue_priv_s {
199 kmutex_t tcp_time_wait_lock;
200 boolean_t tcp_time_wait_collector_active;
201 callout_id_t tcp_time_wait_tid;
202 uint64_t tcp_time_wait_cnt;
203 int64_t tcp_time_wait_schedule;
204 int64_t tcp_time_wait_offset;
205 tcp_t *tcp_time_wait_bucket[TCP_TIME_WAIT_BUCKETS];
206 tcp_t *tcp_free_list;
207 uint_t tcp_free_list_cnt;
208 } tcp_squeue_priv_t;
209
210 /*
211 * Parameters for TCP Initial Send Sequence number (ISS) generation. When
212 * tcp_strong_iss is set to 1, which is the default, the ISS is calculated
213 * by adding three components: a time component which grows by 1 every 4096
214 * nanoseconds (versus every 4 microseconds suggested by RFC 793, page 27);
215 * a per-connection component which grows by 125000 for every new connection;
216 * and an "extra" component that grows by a random amount centered
217 * approximately on 64000. This causes the ISS generator to cycle every
218 * 4.89 hours if no TCP connections are made, and faster if connections are
219 * made.
220 *
221 * When tcp_strong_iss is set to 0, ISS is calculated by adding two
222 * components: a time component which grows by 250000 every second; and
223 * a per-connection component which grows by 125000 for every new connections.
224 *
225 * A third method, when tcp_strong_iss is set to 2, for generating ISS is
226 * prescribed by Steve Bellovin. This involves adding time, the 125000 per
227 * connection, and a one-way hash (MD5) of the connection ID <sport, dport,
228 * src, dst>, a "truly" random (per RFC 1750) number, and a console-entered
229 * password.
230 */
231 #define ISS_INCR 250000
232 #define ISS_NSEC_SHT 12
233
234 /* Macros for timestamp comparisons */
235 #define TSTMP_GEQ(a, b) ((int32_t)((a)-(b)) >= 0)
236 #define TSTMP_LT(a, b) ((int32_t)((a)-(b)) < 0)
237
238 /*
239 * Initialize cwnd according to RFC 3390. def_max_init_cwnd is
240 * either tcp_slow_start_initial or tcp_slow_start_after idle
241 * depending on the caller. If the upper layer has not used the
242 * TCP_INIT_CWND option to change the initial cwnd, tcp_init_cwnd
243 * should be 0 and we use the formula in RFC 3390 to set tcp_cwnd.
244 * If the upper layer has changed set the tcp_init_cwnd, just use
245 * it to calculate the tcp_cwnd.
246 *
247 * "An Argument for Increasing TCP's Initial Congestion Window"
248 * ACM SIGCOMM Computer Communications Review, vol. 40 (2010), pp. 27-33
249 * -- Nandita Dukkipati, Tiziana Refice, Yuchung Cheng,
250 * Hsiao-keng Jerry Chu, Tom Herbert, Amit Agarwal,
251 * Arvind Jain, Natalia Sutin
252 *
253 * "Based on the results from our experiments, we believe the
254 * initial congestion window should be at least ten segments
255 * and the same be investigated for standardization by the IETF."
256 *
257 * As such, the def_max_init_cwnd argument with which this macro is
258 * invoked is either the tcps_slow_start_initial or
259 * tcps_slow_start_after_idle which both default to 0 and will respect
260 * RFC 3390 exactly. If the tunables are explicitly set by the operator,
261 * then the initial congestion window should be set as the operator
262 * demands, within reason. We shall arbitrarily define reason as a
263 * maximum of 16 (same as used by the TCP_INIT_CWND setsockopt).
264 */
265
266 /* Maximum TCP initial cwin (start/restart). */
267 #define TCP_MAX_INIT_CWND 16
268
269 #define TCP_SET_INIT_CWND(tcp, mss, def_max_init_cwnd) \
270 { \
271 if ((tcp)->tcp_init_cwnd == 0) { \
272 if (def_max_init_cwnd == 0) { \
273 (tcp)->tcp_cwnd = MIN(4 * (mss), \
274 MAX(2 * (mss), 4380 / (mss) * (mss))); \
275 } else { \
276 (tcp)->tcp_cwnd = MIN(TCP_MAX_INIT_CWND * (mss),\
277 def_max_init_cwnd * (mss)); \
278 } \
279 } else { \
280 (tcp)->tcp_cwnd = (tcp)->tcp_init_cwnd * (mss); \
281 } \
282 tcp->tcp_cwnd_cnt = 0; \
283 }
284
285 /*
286 * Set ECN capable transport (ECT) code point in IP header.
287 *
288 * Note that there are 2 ECT code points '01' and '10', which are called
289 * ECT(1) and ECT(0) respectively. Here we follow the original ECT code
290 * point ECT(0) for TCP as described in RFC 2481.
291 */
292 #define TCP_SET_ECT(tcp, iph) \
293 if ((tcp)->tcp_connp->conn_ipversion == IPV4_VERSION) { \
294 /* We need to clear the code point first. */ \
295 ((ipha_t *)(iph))->ipha_type_of_service &= 0xFC; \
296 ((ipha_t *)(iph))->ipha_type_of_service |= IPH_ECN_ECT0; \
297 } else { \
298 ((ip6_t *)(iph))->ip6_vcf &= htonl(0xFFCFFFFF); \
299 ((ip6_t *)(iph))->ip6_vcf |= htonl(IPH_ECN_ECT0 << 20); \
300 }
301
302 /*
303 * TCP options struct returned from tcp_parse_options.
304 */
305 typedef struct tcp_opt_s {
306 uint32_t tcp_opt_mss;
307 uint32_t tcp_opt_wscale;
308 uint32_t tcp_opt_ts_val;
309 uint32_t tcp_opt_ts_ecr;
310 tcp_t *tcp;
311 } tcp_opt_t;
312
313 /*
314 * Flags returned from tcp_parse_options.
315 */
316 #define TCP_OPT_MSS_PRESENT 1
317 #define TCP_OPT_WSCALE_PRESENT 2
318 #define TCP_OPT_TSTAMP_PRESENT 4
319 #define TCP_OPT_SACK_OK_PRESENT 8
320 #define TCP_OPT_SACK_PRESENT 16
321
322 /*
323 * Write-side flow-control is implemented via the per instance STREAMS
324 * write-side Q by explicitly setting QFULL to stop the flow of mblk_t(s)
325 * and clearing QFULL and calling qbackenable() to restart the flow based
326 * on the number of TCP unsent bytes (i.e. those not on the wire waiting
327 * for a remote ACK).
328 *
329 * This is different than a standard STREAMS kmod which when using the
330 * STREAMS Q the framework would automatictly flow-control based on the
331 * defined hiwat/lowat values as mblk_t's are enqueued/dequeued.
332 *
333 * As of FireEngine TCP write-side flow-control needs to take into account
334 * both the unsent tcp_xmit list bytes but also any squeue_t enqueued bytes
335 * (i.e. from tcp_wput() -> tcp_output()).
336 *
337 * This is accomplished by adding a new tcp_t fields, tcp_squeue_bytes, to
338 * count the number of bytes enqueued by tcp_wput() and the number of bytes
339 * dequeued and processed by tcp_output().
340 *
341 * So, the total number of bytes unsent is (squeue_bytes + unsent) with all
342 * flow-control uses of unsent replaced with the macro TCP_UNSENT_BYTES.
343 */
344 extern void tcp_clrqfull(tcp_t *);
345 extern void tcp_setqfull(tcp_t *);
346
347 #define TCP_UNSENT_BYTES(tcp) \
348 ((tcp)->tcp_squeue_bytes + (tcp)->tcp_unsent)
349
350 /*
351 * Linked list struct to store listener connection limit configuration per
352 * IP stack. The list is stored at tcps_listener_conf in tcp_stack_t.
353 *
354 * tl_port: the listener port of this limit configuration
355 * tl_ratio: the maximum amount of memory consumed by all concurrent TCP
356 * connections created by a listener does not exceed 1/tl_ratio
357 * of the total system memory. Note that this is only an
358 * approximation.
359 * tl_link: linked list struct
360 */
361 typedef struct tcp_listener_s {
362 in_port_t tl_port;
363 uint32_t tl_ratio;
364 list_node_t tl_link;
365 } tcp_listener_t;
366
367 /*
368 * If there is a limit set on the number of connections allowed per each
369 * listener, the following struct is used to store that counter. It keeps
370 * the number of TCP connection created by a listener. Note that this needs
371 * to be separated from the listener since the listener can go away before
372 * all the connections are gone.
373 *
374 * When the struct is allocated, tlc_cnt is set to 1. When a new connection
375 * is created by the listener, tlc_cnt is incremented by 1. When a connection
376 * created by the listener goes away, tlc_count is decremented by 1. When the
377 * listener itself goes away, tlc_cnt is decremented by one. The last
378 * connection (or the listener) which decrements tlc_cnt to zero frees the
379 * struct.
380 *
381 * tlc_max is the maximum number of concurrent TCP connections created from a
382 * listner. It is calculated when the tcp_listen_cnt_t is allocated.
383 *
384 * tlc_report_time stores the time when cmn_err() is called to report that the
385 * max has been exceeeded. Report is done at most once every
386 * TCP_TLC_REPORT_INTERVAL mins for a listener.
387 *
388 * tlc_drop stores the number of connection attempt dropped because the
389 * limit has reached.
390 */
391 typedef struct tcp_listen_cnt_s {
392 uint32_t tlc_max;
393 uint32_t tlc_cnt;
394 int64_t tlc_report_time;
395 uint32_t tlc_drop;
396 } tcp_listen_cnt_t;
397
398 #define TCP_TLC_REPORT_INTERVAL (30 * MINUTES)
399
400 #define TCP_DECR_LISTEN_CNT(tcp) \
401 { \
402 ASSERT((tcp)->tcp_listen_cnt->tlc_cnt > 0); \
403 if (atomic_dec_32_nv(&(tcp)->tcp_listen_cnt->tlc_cnt) == 0) \
404 kmem_free((tcp)->tcp_listen_cnt, sizeof (tcp_listen_cnt_t)); \
405 (tcp)->tcp_listen_cnt = NULL; \
406 }
407
408 /* Increment and decrement the number of connections in tcp_stack_t. */
409 #define TCPS_CONN_INC(tcps) \
410 atomic_inc_64( \
411 (uint64_t *)&(tcps)->tcps_sc[CPU->cpu_seqid]->tcp_sc_conn_cnt)
412
413 #define TCPS_CONN_DEC(tcps) \
414 atomic_dec_64( \
415 (uint64_t *)&(tcps)->tcps_sc[CPU->cpu_seqid]->tcp_sc_conn_cnt)
416
417 /*
418 * When the system is under memory pressure, stack variable tcps_reclaim is
419 * true, we shorten the connection timeout abort interval to tcp_early_abort
420 * seconds. Defined in tcp.c.
421 */
422 extern uint32_t tcp_early_abort;
423
424 /*
425 * To reach to an eager in Q0 which can be dropped due to an incoming
426 * new SYN request when Q0 is full, a new doubly linked list is
427 * introduced. This list allows to select an eager from Q0 in O(1) time.
428 * This is needed to avoid spending too much time walking through the
429 * long list of eagers in Q0 when tcp_drop_q0() is called. Each member of
430 * this new list has to be a member of Q0.
431 * This list is headed by listener's tcp_t. When the list is empty,
432 * both the pointers - tcp_eager_next_drop_q0 and tcp_eager_prev_drop_q0,
433 * of listener's tcp_t point to listener's tcp_t itself.
434 *
435 * Given an eager in Q0 and a listener, MAKE_DROPPABLE() puts the eager
436 * in the list. MAKE_UNDROPPABLE() takes the eager out of the list.
437 * These macros do not affect the eager's membership to Q0.
438 */
439 #define MAKE_DROPPABLE(listener, eager) \
440 if ((eager)->tcp_eager_next_drop_q0 == NULL) { \
441 (listener)->tcp_eager_next_drop_q0->tcp_eager_prev_drop_q0\
442 = (eager); \
443 (eager)->tcp_eager_prev_drop_q0 = (listener); \
444 (eager)->tcp_eager_next_drop_q0 = \
445 (listener)->tcp_eager_next_drop_q0; \
446 (listener)->tcp_eager_next_drop_q0 = (eager); \
447 }
448
449 #define MAKE_UNDROPPABLE(eager) \
450 if ((eager)->tcp_eager_next_drop_q0 != NULL) { \
451 (eager)->tcp_eager_next_drop_q0->tcp_eager_prev_drop_q0 \
452 = (eager)->tcp_eager_prev_drop_q0; \
453 (eager)->tcp_eager_prev_drop_q0->tcp_eager_next_drop_q0 \
454 = (eager)->tcp_eager_next_drop_q0; \
455 (eager)->tcp_eager_prev_drop_q0 = NULL; \
456 (eager)->tcp_eager_next_drop_q0 = NULL; \
457 }
458
459 /*
460 * The format argument to pass to tcp_display().
461 * DISP_PORT_ONLY means that the returned string has only port info.
462 * DISP_ADDR_AND_PORT means that the returned string also contains the
463 * remote and local IP address.
464 */
465 #define DISP_PORT_ONLY 1
466 #define DISP_ADDR_AND_PORT 2
467
468 #define IP_ADDR_CACHE_SIZE 2048
469 #define IP_ADDR_CACHE_HASH(faddr) \
470 (ntohl(faddr) & (IP_ADDR_CACHE_SIZE -1))
471
472 /*
473 * TCP reassembly macros. We hide starting and ending sequence numbers in
474 * b_next and b_prev of messages on the reassembly queue. The messages are
475 * chained using b_cont. These macros are used in tcp_reass() so we don't
476 * have to see the ugly casts and assignments.
477 */
478 #define TCP_REASS_SEQ(mp) ((uint32_t)(uintptr_t)((mp)->b_next))
479 #define TCP_REASS_SET_SEQ(mp, u) ((mp)->b_next = \
480 (mblk_t *)(uintptr_t)(u))
481 #define TCP_REASS_END(mp) ((uint32_t)(uintptr_t)((mp)->b_prev))
482 #define TCP_REASS_SET_END(mp, u) ((mp)->b_prev = \
483 (mblk_t *)(uintptr_t)(u))
484
485 #define tcps_time_wait_interval tcps_propinfo_tbl[0].prop_cur_uval
486 #define tcps_conn_req_max_q tcps_propinfo_tbl[1].prop_cur_uval
487 #define tcps_conn_req_max_q0 tcps_propinfo_tbl[2].prop_cur_uval
488 #define tcps_conn_req_min tcps_propinfo_tbl[3].prop_cur_uval
489 #define tcps_conn_grace_period tcps_propinfo_tbl[4].prop_cur_uval
490 #define tcps_cwnd_max_ tcps_propinfo_tbl[5].prop_cur_uval
491 #define tcps_dbg tcps_propinfo_tbl[6].prop_cur_uval
492 #define tcps_smallest_nonpriv_port tcps_propinfo_tbl[7].prop_cur_uval
493 #define tcps_ip_abort_cinterval tcps_propinfo_tbl[8].prop_cur_uval
494 #define tcps_ip_abort_linterval tcps_propinfo_tbl[9].prop_cur_uval
495 #define tcps_ip_abort_interval tcps_propinfo_tbl[10].prop_cur_uval
496 #define tcps_ip_notify_cinterval tcps_propinfo_tbl[11].prop_cur_uval
497 #define tcps_ip_notify_interval tcps_propinfo_tbl[12].prop_cur_uval
498 #define tcps_ipv4_ttl tcps_propinfo_tbl[13].prop_cur_uval
499 #define tcps_keepalive_interval_high tcps_propinfo_tbl[14].prop_max_uval
500 #define tcps_keepalive_interval tcps_propinfo_tbl[14].prop_cur_uval
501 #define tcps_keepalive_interval_low tcps_propinfo_tbl[14].prop_min_uval
502 #define tcps_maxpsz_multiplier tcps_propinfo_tbl[15].prop_cur_uval
503 #define tcps_mss_def_ipv4 tcps_propinfo_tbl[16].prop_cur_uval
504 #define tcps_mss_max_ipv4 tcps_propinfo_tbl[17].prop_cur_uval
505 #define tcps_mss_min tcps_propinfo_tbl[18].prop_cur_uval
506 #define tcps_naglim_def tcps_propinfo_tbl[19].prop_cur_uval
507 #define tcps_rexmit_interval_initial_high \
508 tcps_propinfo_tbl[20].prop_max_uval
509 #define tcps_rexmit_interval_initial tcps_propinfo_tbl[20].prop_cur_uval
510 #define tcps_rexmit_interval_initial_low \
511 tcps_propinfo_tbl[20].prop_min_uval
512 #define tcps_rexmit_interval_max_high tcps_propinfo_tbl[21].prop_max_uval
513 #define tcps_rexmit_interval_max tcps_propinfo_tbl[21].prop_cur_uval
514 #define tcps_rexmit_interval_max_low tcps_propinfo_tbl[21].prop_min_uval
515 #define tcps_rexmit_interval_min_high tcps_propinfo_tbl[22].prop_max_uval
516 #define tcps_rexmit_interval_min tcps_propinfo_tbl[22].prop_cur_uval
517 #define tcps_rexmit_interval_min_low tcps_propinfo_tbl[22].prop_min_uval
518 #define tcps_deferred_ack_interval tcps_propinfo_tbl[23].prop_cur_uval
519 #define tcps_snd_lowat_fraction tcps_propinfo_tbl[24].prop_cur_uval
520 #define tcps_dupack_fast_retransmit tcps_propinfo_tbl[25].prop_cur_uval
521 #define tcps_ignore_path_mtu tcps_propinfo_tbl[26].prop_cur_bval
522 #define tcps_smallest_anon_port tcps_propinfo_tbl[27].prop_cur_uval
523 #define tcps_largest_anon_port tcps_propinfo_tbl[28].prop_cur_uval
524 #define tcps_xmit_hiwat tcps_propinfo_tbl[29].prop_cur_uval
525 #define tcps_xmit_lowat tcps_propinfo_tbl[30].prop_cur_uval
526 #define tcps_recv_hiwat tcps_propinfo_tbl[31].prop_cur_uval
527 #define tcps_recv_hiwat_minmss tcps_propinfo_tbl[32].prop_cur_uval
528 #define tcps_fin_wait_2_flush_interval_high \
529 tcps_propinfo_tbl[33].prop_max_uval
530 #define tcps_fin_wait_2_flush_interval tcps_propinfo_tbl[33].prop_cur_uval
531 #define tcps_fin_wait_2_flush_interval_low \
532 tcps_propinfo_tbl[33].prop_min_uval
533 #define tcps_max_buf tcps_propinfo_tbl[34].prop_cur_uval
534 #define tcps_strong_iss tcps_propinfo_tbl[35].prop_cur_uval
535 #define tcps_rtt_updates tcps_propinfo_tbl[36].prop_cur_uval
536 #define tcps_wscale_always tcps_propinfo_tbl[37].prop_cur_bval
537 #define tcps_tstamp_always tcps_propinfo_tbl[38].prop_cur_bval
538 #define tcps_tstamp_if_wscale tcps_propinfo_tbl[39].prop_cur_bval
539 #define tcps_rexmit_interval_extra tcps_propinfo_tbl[40].prop_cur_uval
540 #define tcps_deferred_acks_max tcps_propinfo_tbl[41].prop_cur_uval
541 #define tcps_slow_start_after_idle tcps_propinfo_tbl[42].prop_cur_uval
542 #define tcps_slow_start_initial tcps_propinfo_tbl[43].prop_cur_uval
543 #define tcps_sack_permitted tcps_propinfo_tbl[44].prop_cur_uval
544 #define tcps_ipv6_hoplimit tcps_propinfo_tbl[45].prop_cur_uval
545 #define tcps_mss_def_ipv6 tcps_propinfo_tbl[46].prop_cur_uval
546 #define tcps_mss_max_ipv6 tcps_propinfo_tbl[47].prop_cur_uval
547 #define tcps_rev_src_routes tcps_propinfo_tbl[48].prop_cur_bval
548 #define tcps_local_dack_interval tcps_propinfo_tbl[49].prop_cur_uval
549 #define tcps_local_dacks_max tcps_propinfo_tbl[50].prop_cur_uval
550 #define tcps_ecn_permitted tcps_propinfo_tbl[51].prop_cur_uval
551 #define tcps_rst_sent_rate_enabled tcps_propinfo_tbl[52].prop_cur_bval
552 #define tcps_rst_sent_rate tcps_propinfo_tbl[53].prop_cur_uval
553 #define tcps_push_timer_interval tcps_propinfo_tbl[54].prop_cur_uval
554 #define tcps_use_smss_as_mss_opt tcps_propinfo_tbl[55].prop_cur_bval
555 #define tcps_keepalive_abort_interval_high \
556 tcps_propinfo_tbl[56].prop_max_uval
557 #define tcps_keepalive_abort_interval \
558 tcps_propinfo_tbl[56].prop_cur_uval
559 #define tcps_keepalive_abort_interval_low \
560 tcps_propinfo_tbl[56].prop_min_uval
561 #define tcps_wroff_xtra tcps_propinfo_tbl[57].prop_cur_uval
562 #define tcps_dev_flow_ctl tcps_propinfo_tbl[58].prop_cur_bval
563 #define tcps_reass_timeout tcps_propinfo_tbl[59].prop_cur_uval
564 #define tcps_iss_incr tcps_propinfo_tbl[65].prop_cur_uval
565
566
567 /*
568 * As defined in RFC 6298, the RTO is the average estimates (SRTT) plus a
569 * multiple of the deviation estimates (K * RTTVAR):
570 *
571 * RTO = SRTT + max(G, K * RTTVAR)
572 *
573 * K is defined in the RFC as 4, and G is the clock granularity. We constrain
574 * the minimum mean deviation to TCP_SD_MIN when processing new RTTs, so this
575 * becomes:
576 *
577 * RTO = SRTT + 4 * RTTVAR
578 *
579 * In practice, however, we make several additions to it. As we use a finer
580 * grained clock than BSD and update RTO for every ACK, we add in another 1/4 of
581 * RTT to the deviation of RTO to accommodate burstiness of 1/4 of window size:
582 *
583 * RTO = SRTT + (SRTT / 4) + 4 * RTTVAR
584 *
585 * Since tcp_rtt_sa is 8 times the SRTT, and tcp_rtt_sd is 4 times the RTTVAR,
586 * this becomes:
587 *
588 * RTO = (tcp_rtt_sa / 8) + ((tcp_rtt_sa / 8) / 4) + tcp_rtt_sd
589 * RTO = (tcp_rtt_sa / 2^3) + (tcp_rtt_sa / 2^5) + tcp_rtt_sd
590 * RTO = (tcp_rtt_sa >> 3) + (tcp_rtt_sa >> 5) + tcp_rtt_sd
591 *
592 * The "tcp_rexmit_interval_extra" and "tcp_conn_grace_period" tunables are
593 * used to help account for extreme environments where the algorithm fails to
594 * work; by default they should be 0. (The latter tunable is only used for
595 * calculating the intial RTO, and so is optionally passed in as "extra".) We
596 * add them here:
597 *
598 * RTO = (tcp_rtt_sa >> 3) + (tcp_rtt_sa >> 5) + tcp_rtt_sd +
599 * tcps_rexmit_interval_extra + tcps_conn_grace_period
600 *
601 * We then pin the RTO within our configured boundaries (sections 2.4 and 2.5
602 * of RFC 6298).
603 */
604 static __GNU_INLINE clock_t
605 tcp_calculate_rto(tcp_t *tcp, tcp_stack_t *tcps, uint32_t extra)
606 {
607 clock_t rto;
608
609 rto = NSEC2MSEC((tcp->tcp_rtt_sa >> 3) + (tcp->tcp_rtt_sa >> 5) +
610 tcp->tcp_rtt_sd) + tcps->tcps_rexmit_interval_extra + extra;
611
612 if (rto < tcp->tcp_rto_min) {
613 rto = tcp->tcp_rto_min;
614 } else if (rto > tcp->tcp_rto_max) {
615 rto = tcp->tcp_rto_max;
616 }
617
618 return (rto);
619 }
620
621 extern struct qinit tcp_rinitv4, tcp_rinitv6;
622 extern boolean_t do_tcp_fusion;
623
624 /*
625 * Object to represent database of options to search passed to
626 * {sock,tpi}optcom_req() interface routine to take care of option
627 * management and associated methods.
628 */
629 extern optdb_obj_t tcp_opt_obj;
630 extern uint_t tcp_max_optsize;
631
632 extern int tcp_squeue_flag;
633
634 extern uint_t tcp_free_list_max_cnt;
635
636 /*
637 * Functions in tcp.c.
638 */
639 extern void tcp_acceptor_hash_insert(t_uscalar_t, tcp_t *);
640 extern tcp_t *tcp_acceptor_hash_lookup(t_uscalar_t, tcp_stack_t *);
641 extern void tcp_acceptor_hash_remove(tcp_t *);
642 extern mblk_t *tcp_ack_mp(tcp_t *);
643 extern int tcp_build_hdrs(tcp_t *);
644 extern void tcp_cleanup(tcp_t *);
645 extern int tcp_clean_death(tcp_t *, int);
646 extern void tcp_clean_death_wrapper(void *, mblk_t *, void *,
647 ip_recv_attr_t *);
648 extern void tcp_close_common(conn_t *, int);
649 extern void tcp_close_detached(tcp_t *);
650 extern void tcp_close_mpp(mblk_t **);
651 extern void tcp_closei_local(tcp_t *);
652 extern sock_lower_handle_t tcp_create(int, int, int, sock_downcalls_t **,
653 uint_t *, int *, int, cred_t *);
654 extern conn_t *tcp_create_common(cred_t *, boolean_t, boolean_t, int *);
655 extern void tcp_disconnect(tcp_t *, mblk_t *);
656 extern char *tcp_display(tcp_t *, char *, char);
657 extern int tcp_do_bind(conn_t *, struct sockaddr *, socklen_t, cred_t *,
658 boolean_t);
659 extern int tcp_do_connect(conn_t *, const struct sockaddr *, socklen_t,
660 cred_t *, pid_t);
661 extern int tcp_do_listen(conn_t *, struct sockaddr *, socklen_t, int,
662 cred_t *, boolean_t);
663 extern int tcp_do_unbind(conn_t *);
664 extern boolean_t tcp_eager_blowoff(tcp_t *, t_scalar_t);
665 extern void tcp_eager_cleanup(tcp_t *, boolean_t);
666 extern void tcp_eager_kill(void *, mblk_t *, void *, ip_recv_attr_t *);
667 extern void tcp_eager_unlink(tcp_t *);
668 extern void tcp_init_values(tcp_t *, tcp_t *);
669 extern void tcp_ipsec_cleanup(tcp_t *);
670 extern int tcp_maxpsz_set(tcp_t *, boolean_t);
671 extern void tcp_mss_set(tcp_t *, uint32_t);
672 extern void tcp_reinput(conn_t *, mblk_t *, ip_recv_attr_t *, ip_stack_t *);
673 extern int tcp_rsrv(queue_t *);
674 extern uint_t tcp_rwnd_reopen(tcp_t *);
675 extern int tcp_rwnd_set(tcp_t *, uint32_t);
676 extern int tcp_set_destination(tcp_t *);
677 extern void tcp_set_ws_value(tcp_t *);
678 extern void tcp_stop_lingering(tcp_t *);
679 extern void tcp_update_pmtu(tcp_t *, boolean_t);
680 extern mblk_t *tcp_zcopy_backoff(tcp_t *, mblk_t *, boolean_t);
681 extern boolean_t tcp_zcopy_check(tcp_t *);
682 extern void tcp_zcopy_notify(tcp_t *);
683 extern void tcp_get_proto_props(tcp_t *, struct sock_proto_props *);
684
685 /*
686 * Bind related functions in tcp_bind.c
687 */
688 extern int tcp_bind_check(conn_t *, struct sockaddr *, socklen_t,
689 cred_t *, boolean_t);
690 extern void tcp_bind_hash_insert(tf_t *, tcp_t *, int);
691 extern void tcp_bind_hash_remove(tcp_t *);
692 extern in_port_t tcp_bindi(tcp_t *, in_port_t, const in6_addr_t *,
693 int, boolean_t, boolean_t, boolean_t);
694 extern in_port_t tcp_update_next_port(in_port_t, const tcp_t *,
695 boolean_t);
696
697 /*
698 * Fusion related functions in tcp_fusion.c.
699 */
700 extern void tcp_fuse(tcp_t *, uchar_t *, tcpha_t *);
701 extern void tcp_unfuse(tcp_t *);
702 extern boolean_t tcp_fuse_output(tcp_t *, mblk_t *, uint32_t);
703 extern void tcp_fuse_output_urg(tcp_t *, mblk_t *);
704 extern boolean_t tcp_fuse_rcv_drain(queue_t *, tcp_t *, mblk_t **);
705 extern size_t tcp_fuse_set_rcv_hiwat(tcp_t *, size_t);
706 extern int tcp_fuse_maxpsz(tcp_t *);
707 extern void tcp_fuse_backenable(tcp_t *);
708 extern void tcp_iss_key_init(uint8_t *, int, tcp_stack_t *);
709
710 /*
711 * Output related functions in tcp_output.c.
712 */
713 extern void tcp_close_output(void *, mblk_t *, void *, ip_recv_attr_t *);
714 extern void tcp_output(void *, mblk_t *, void *, ip_recv_attr_t *);
715 extern void tcp_output_urgent(void *, mblk_t *, void *, ip_recv_attr_t *);
716 extern void tcp_rexmit_after_error(tcp_t *);
717 extern void tcp_sack_rexmit(tcp_t *, uint_t *);
718 extern void tcp_send_data(tcp_t *, mblk_t *);
719 extern void tcp_send_synack(void *, mblk_t *, void *, ip_recv_attr_t *);
720 extern void tcp_shutdown_output(void *, mblk_t *, void *, ip_recv_attr_t *);
721 extern void tcp_ss_rexmit(tcp_t *);
722 extern void tcp_update_xmit_tail(tcp_t *, uint32_t);
723 extern int tcp_wput(queue_t *, mblk_t *);
724 extern void tcp_wput_data(tcp_t *, mblk_t *, boolean_t);
725 extern int tcp_wput_sock(queue_t *, mblk_t *);
726 extern int tcp_wput_fallback(queue_t *, mblk_t *);
727 extern void tcp_xmit_ctl(char *, tcp_t *, uint32_t, uint32_t, int);
728 extern void tcp_xmit_listeners_reset(mblk_t *, ip_recv_attr_t *,
729 ip_stack_t *i, conn_t *);
730 extern mblk_t *tcp_xmit_mp(tcp_t *, mblk_t *, int32_t, int32_t *,
731 mblk_t **, uint32_t, boolean_t, uint32_t *, boolean_t);
732
733 /*
734 * Input related functions in tcp_input.c.
735 */
736 extern void tcp_icmp_input(void *, mblk_t *, void *, ip_recv_attr_t *);
737 extern void tcp_input_data(void *, mblk_t *, void *, ip_recv_attr_t *);
738 extern void tcp_input_listener_unbound(void *, mblk_t *, void *,
739 ip_recv_attr_t *);
740 extern boolean_t tcp_paws_check(tcp_t *, const tcp_opt_t *);
741 extern int tcp_parse_options(tcpha_t *, tcp_opt_t *);
742 extern uint_t tcp_rcv_drain(tcp_t *);
743 extern void tcp_rcv_enqueue(tcp_t *, mblk_t *, uint_t, cred_t *);
744 extern boolean_t tcp_verifyicmp(conn_t *, void *, icmph_t *, icmp6_t *,
745 ip_recv_attr_t *);
746
747 /*
748 * Kernel socket related functions in tcp_socket.c.
749 */
750 extern int tcp_fallback(sock_lower_handle_t, queue_t *, boolean_t,
751 so_proto_quiesced_cb_t, sock_quiesce_arg_t *);
752 extern boolean_t tcp_newconn_notify(tcp_t *, ip_recv_attr_t *);
753
754 /*
755 * Timer related functions in tcp_timers.c.
756 */
757 extern void tcp_ack_timer(void *);
758 extern void tcp_close_linger_timeout(void *);
759 extern void tcp_keepalive_timer(void *);
760 extern void tcp_push_timer(void *);
761 extern void tcp_reass_timer(void *);
762 extern mblk_t *tcp_timermp_alloc(int);
763 extern void tcp_timermp_free(tcp_t *);
764 extern timeout_id_t tcp_timeout(conn_t *, void (*)(void *), hrtime_t);
765 extern clock_t tcp_timeout_cancel(conn_t *, timeout_id_t);
766 extern void tcp_timer(void *arg);
767 extern void tcp_timers_stop(tcp_t *);
768
769 /*
770 * TCP TPI related functions in tcp_tpi.c.
771 */
772 extern void tcp_addr_req(tcp_t *, mblk_t *);
773 extern void tcp_capability_req(tcp_t *, mblk_t *);
774 extern boolean_t tcp_conn_con(tcp_t *, uchar_t *, mblk_t *,
775 mblk_t **, ip_recv_attr_t *);
776 extern void tcp_err_ack(tcp_t *, mblk_t *, int, int);
777 extern void tcp_err_ack_prim(tcp_t *, mblk_t *, int, int, int);
778 extern void tcp_info_req(tcp_t *, mblk_t *);
779 extern void tcp_send_conn_ind(void *, mblk_t *, void *);
780 extern void tcp_send_pending(void *, mblk_t *, void *, ip_recv_attr_t *);
781 extern int tcp_tpi_accept(queue_t *, mblk_t *);
782 extern void tcp_tpi_bind(tcp_t *, mblk_t *);
783 extern int tcp_tpi_close(queue_t *, int, cred_t *);
784 extern int tcp_tpi_close_accept(queue_t *, int, cred_t *);
785 extern void tcp_tpi_connect(tcp_t *, mblk_t *);
786 extern int tcp_tpi_opt_get(queue_t *, t_scalar_t, t_scalar_t, uchar_t *);
787 extern int tcp_tpi_opt_set(queue_t *, uint_t, int, int, uint_t, uchar_t *,
788 uint_t *, uchar_t *, void *, cred_t *);
789 extern void tcp_tpi_unbind(tcp_t *, mblk_t *);
790 extern void tcp_tli_accept(tcp_t *, mblk_t *);
791 extern void tcp_use_pure_tpi(tcp_t *);
792 extern void tcp_do_capability_ack(tcp_t *, struct T_capability_ack *,
793 t_uscalar_t);
794
795 /*
796 * TCP option processing related functions in tcp_opt_data.c
797 */
798 extern int tcp_opt_get(conn_t *, int, int, uchar_t *);
799 extern int tcp_opt_set(conn_t *, uint_t, int, int, uint_t, uchar_t *,
800 uint_t *, uchar_t *, void *, cred_t *);
801
802 /*
803 * TCP time wait processing related functions in tcp_time_wait.c.
804 */
805 extern void tcp_time_wait_append(tcp_t *);
806 extern void tcp_time_wait_collector(void *);
807 extern boolean_t tcp_time_wait_remove(tcp_t *, tcp_squeue_priv_t *);
808 extern void tcp_time_wait_processing(tcp_t *, mblk_t *, uint32_t,
809 uint32_t, int, tcpha_t *, ip_recv_attr_t *);
810
811 /*
812 * Misc functions in tcp_misc.c.
813 */
814 extern uint32_t tcp_find_listener_conf(tcp_stack_t *, in_port_t);
815 extern void tcp_ioctl_abort_conn(queue_t *, mblk_t *);
816 extern void tcp_listener_conf_cleanup(tcp_stack_t *);
817 extern void tcp_stack_cpu_add(tcp_stack_t *, processorid_t);
818
819 #endif /* _KERNEL */
820
821 #ifdef __cplusplus
822 }
823 #endif
824
825 #endif /* _INET_TCP_IMPL_H */
--- EOF ---