blob: a68d5269abd8d6b08a170153c6eff5be8a75ef92 [file] [log] [blame]
Dave Barach68b0fb02017-02-28 15:15:56 -05001/*
Florin Coras222e1f412019-02-16 20:47:32 -08002 * Copyright (c) 2016-2019 Cisco and/or its affiliates.
Dave Barach68b0fb02017-02-28 15:15:56 -05003 * Licensed under the Apache License, Version 2.0 (the "License");
4 * you may not use this file except in compliance with the License.
5 * You may obtain a copy of the License at:
6 *
7 * http://www.apache.org/licenses/LICENSE-2.0
8 *
9 * Unless required by applicable law or agreed to in writing, software
10 * distributed under the License is distributed on an "AS IS" BASIS,
11 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 * See the License for the specific language governing permissions and
13 * limitations under the License.
14 */
15
16#include <vnet/tcp/tcp.h>
Florin Coras999840c2020-03-18 20:31:34 +000017#include <vnet/tcp/tcp_inlines.h>
Florin Corasb2215d62017-08-01 16:56:58 -070018#include <math.h>
Dave Barach68b0fb02017-02-28 15:15:56 -050019
Dave Barach2c25a622017-06-26 11:35:07 -040020typedef enum _tcp_output_next
Dave Barach68b0fb02017-02-28 15:15:56 -050021{
22 TCP_OUTPUT_NEXT_DROP,
Dave Barach2c25a622017-06-26 11:35:07 -040023 TCP_OUTPUT_NEXT_IP_LOOKUP,
Florin Corasf9d05682018-04-26 08:26:52 -070024 TCP_OUTPUT_NEXT_IP_REWRITE,
25 TCP_OUTPUT_NEXT_IP_ARP,
Dave Barach68b0fb02017-02-28 15:15:56 -050026 TCP_OUTPUT_N_NEXT
27} tcp_output_next_t;
28
29#define foreach_tcp4_output_next \
30 _ (DROP, "error-drop") \
Florin Corasf9d05682018-04-26 08:26:52 -070031 _ (IP_LOOKUP, "ip4-lookup") \
32 _ (IP_REWRITE, "ip4-rewrite") \
33 _ (IP_ARP, "ip4-arp")
Dave Barach68b0fb02017-02-28 15:15:56 -050034
35#define foreach_tcp6_output_next \
36 _ (DROP, "error-drop") \
Florin Corasf9d05682018-04-26 08:26:52 -070037 _ (IP_LOOKUP, "ip6-lookup") \
38 _ (IP_REWRITE, "ip6-rewrite") \
39 _ (IP_ARP, "ip6-discover-neighbor")
Dave Barach68b0fb02017-02-28 15:15:56 -050040
41static char *tcp_error_strings[] = {
42#define tcp_error(n,s) s,
43#include <vnet/tcp/tcp_error.def>
44#undef tcp_error
45};
46
47typedef struct
48{
Clement Durand6cf260c2017-04-13 13:27:04 +020049 tcp_header_t tcp_header;
50 tcp_connection_t tcp_connection;
Dave Barach68b0fb02017-02-28 15:15:56 -050051} tcp_tx_trace_t;
52
Filip Tehlare275bed2019-03-06 00:06:56 -080053static u8 *
Dave Barach68b0fb02017-02-28 15:15:56 -050054format_tcp_tx_trace (u8 * s, va_list * args)
55{
56 CLIB_UNUSED (vlib_main_t * vm) = va_arg (*args, vlib_main_t *);
57 CLIB_UNUSED (vlib_node_t * node) = va_arg (*args, vlib_node_t *);
Clement Durand6cf260c2017-04-13 13:27:04 +020058 tcp_tx_trace_t *t = va_arg (*args, tcp_tx_trace_t *);
Florin Coras30928f82020-01-27 19:21:28 -080059 tcp_connection_t *tc = &t->tcp_connection;
Christophe Fontained3c008d2017-10-02 18:10:54 +020060 u32 indent = format_get_indent (s);
Dave Barach68b0fb02017-02-28 15:15:56 -050061
Florin Coras30928f82020-01-27 19:21:28 -080062 s = format (s, "%U state %U\n%U%U", format_tcp_connection_id, tc,
63 format_tcp_state, tc->state, format_white_space, indent,
64 format_tcp_header, &t->tcp_header, 128);
Dave Barach68b0fb02017-02-28 15:15:56 -050065
66 return s;
67}
68
Filip Tehlare275bed2019-03-06 00:06:56 -080069#ifndef CLIB_MARCH_VARIANT
Dave Barach68b0fb02017-02-28 15:15:56 -050070static u8
Florin Coras4eeeaaf2017-09-05 14:03:37 -040071tcp_window_compute_scale (u32 window)
Dave Barach68b0fb02017-02-28 15:15:56 -050072{
73 u8 wnd_scale = 0;
Florin Coras4eeeaaf2017-09-05 14:03:37 -040074 while (wnd_scale < TCP_MAX_WND_SCALE && (window >> wnd_scale) > TCP_WND_MAX)
Dave Barach68b0fb02017-02-28 15:15:56 -050075 wnd_scale++;
76 return wnd_scale;
77}
78
79/**
Florin Coras6534b7a2017-07-18 05:38:03 -040080 * TCP's initial window
Florin Corase04c2992017-03-01 08:17:34 -080081 */
82always_inline u32
83tcp_initial_wnd_unscaled (tcp_connection_t * tc)
84{
Florin Coras6534b7a2017-07-18 05:38:03 -040085 /* RFC 6928 recommends the value lower. However at the time our connections
86 * are initialized, fifos may not be allocated. Therefore, advertise the
87 * smallest possible unscaled window size and update once fifos are
88 * assigned to the session.
89 */
90 /*
91 tcp_update_rcv_mss (tc);
92 TCP_IW_N_SEGMENTS * tc->mss;
93 */
Florin Coras9094b5c2019-08-12 14:17:47 -070094 return tcp_cfg.min_rx_fifo;
Florin Corase04c2992017-03-01 08:17:34 -080095}
96
97/**
Dave Barach68b0fb02017-02-28 15:15:56 -050098 * Compute initial window and scale factor. As per RFC1323, window field in
99 * SYN and SYN-ACK segments is never scaled.
100 */
101u32
102tcp_initial_window_to_advertise (tcp_connection_t * tc)
103{
Florin Corase80b5912018-12-12 19:25:43 -0800104 /* Compute rcv wscale only if peer advertised support for it */
105 if (tc->state != TCP_STATE_SYN_RCVD || tcp_opts_wscale (&tc->rcv_opts))
Florin Coras9094b5c2019-08-12 14:17:47 -0700106 tc->rcv_wscale = tcp_window_compute_scale (tcp_cfg.max_rx_fifo);
Florin Corase80b5912018-12-12 19:25:43 -0800107
Florin Corase04c2992017-03-01 08:17:34 -0800108 tc->rcv_wnd = tcp_initial_wnd_unscaled (tc);
Dave Barach68b0fb02017-02-28 15:15:56 -0500109
110 return clib_min (tc->rcv_wnd, TCP_WND_MAX);
111}
112
Florin Corasf20fd1a2019-03-28 13:21:19 -0700113static inline void
Florin Coras6792ec02017-03-13 03:49:51 -0700114tcp_update_rcv_wnd (tcp_connection_t * tc)
115{
Florin Corasf20fd1a2019-03-28 13:21:19 -0700116 u32 available_space, wnd;
Florin Coras6792ec02017-03-13 03:49:51 -0700117 i32 observed_wnd;
Florin Corasf20fd1a2019-03-28 13:21:19 -0700118
Florin Corase04c2992017-03-01 08:17:34 -0800119 /*
120 * Figure out how much space we have available
121 */
Florin Corasd2aab832018-05-22 11:39:59 -0700122 available_space = transport_max_rx_enqueue (&tc->connection);
Florin Corasf20fd1a2019-03-28 13:21:19 -0700123 if (PREDICT_FALSE (available_space < tc->rcv_opts.mss))
Florin Coras182d2192019-07-24 10:27:20 -0700124 {
125 tc->rcv_wnd = 0;
126 return;
127 }
Dave Barach68b0fb02017-02-28 15:15:56 -0500128
Florin Corase04c2992017-03-01 08:17:34 -0800129 /*
130 * Use the above and what we know about what we've previously advertised
131 * to compute the new window
132 */
Florin Coras6792ec02017-03-13 03:49:51 -0700133 observed_wnd = (i32) tc->rcv_wnd - (tc->rcv_nxt - tc->rcv_las);
Florin Corase04c2992017-03-01 08:17:34 -0800134
135 /* Bad. Thou shalt not shrink */
Florin Corasf20fd1a2019-03-28 13:21:19 -0700136 if (PREDICT_FALSE ((i32) available_space < observed_wnd))
Florin Corase04c2992017-03-01 08:17:34 -0800137 {
Florin Corasf20fd1a2019-03-28 13:21:19 -0700138 wnd = clib_max (observed_wnd, 0);
Florin Corasa436a422019-08-20 07:09:31 -0700139 TCP_EVT (TCP_EVT_RCV_WND_SHRUNK, tc, observed_wnd, available_space);
Florin Corase04c2992017-03-01 08:17:34 -0800140 }
Florin Coras6792ec02017-03-13 03:49:51 -0700141 else
Florin Corase04c2992017-03-01 08:17:34 -0800142 {
Florin Coras6792ec02017-03-13 03:49:51 -0700143 wnd = available_space;
Florin Corase04c2992017-03-01 08:17:34 -0800144 }
145
Florin Coras3e350af2017-03-30 02:54:28 -0700146 /* Make sure we have a multiple of rcv_wscale */
147 if (wnd && tc->rcv_wscale)
148 {
Florin Coras679b78f2019-03-06 21:16:26 -0800149 wnd &= ~((1 << tc->rcv_wscale) - 1);
Florin Coras3e350af2017-03-30 02:54:28 -0700150 if (wnd == 0)
151 wnd = 1 << tc->rcv_wscale;
152 }
Florin Coras6792ec02017-03-13 03:49:51 -0700153
154 tc->rcv_wnd = clib_min (wnd, TCP_WND_MAX << tc->rcv_wscale);
Dave Barach68b0fb02017-02-28 15:15:56 -0500155}
156
157/**
Florin Coras0dbd5172018-06-25 16:19:34 -0700158 * Compute and return window to advertise, scaled as per RFC1323
159 */
Florin Coras47596832019-03-12 18:58:54 -0700160static inline u32
Florin Coras0dbd5172018-06-25 16:19:34 -0700161tcp_window_to_advertise (tcp_connection_t * tc, tcp_state_t state)
162{
163 if (state < TCP_STATE_ESTABLISHED)
164 return tcp_initial_window_to_advertise (tc);
165
166 tcp_update_rcv_wnd (tc);
Florin Coras0dbd5172018-06-25 16:19:34 -0700167 return tc->rcv_wnd >> tc->rcv_wscale;
168}
169
Florin Coras0dbd5172018-06-25 16:19:34 -0700170static int
Florin Corascedcf602019-08-27 12:15:43 -0700171tcp_make_syn_options (tcp_connection_t * tc, tcp_options_t * opts)
Dave Barach68b0fb02017-02-28 15:15:56 -0500172{
173 u8 len = 0;
174
175 opts->flags |= TCP_OPTS_FLAG_MSS;
Florin Corascedcf602019-08-27 12:15:43 -0700176 opts->mss = tc->mss;
Dave Barach68b0fb02017-02-28 15:15:56 -0500177 len += TCP_OPTION_LEN_MSS;
178
179 opts->flags |= TCP_OPTS_FLAG_WSCALE;
Florin Corascedcf602019-08-27 12:15:43 -0700180 opts->wscale = tc->rcv_wscale;
Dave Barach68b0fb02017-02-28 15:15:56 -0500181 len += TCP_OPTION_LEN_WINDOW_SCALE;
182
183 opts->flags |= TCP_OPTS_FLAG_TSTAMP;
184 opts->tsval = tcp_time_now ();
185 opts->tsecr = 0;
186 len += TCP_OPTION_LEN_TIMESTAMP;
187
Florin Coras93992a92017-05-24 18:03:56 -0700188 if (TCP_USE_SACKS)
189 {
190 opts->flags |= TCP_OPTS_FLAG_SACK_PERMITTED;
191 len += TCP_OPTION_LEN_SACK_PERMITTED;
192 }
Dave Barach68b0fb02017-02-28 15:15:56 -0500193
194 /* Align to needed boundary */
195 len += (TCP_OPTS_ALIGN - len % TCP_OPTS_ALIGN) % TCP_OPTS_ALIGN;
196 return len;
197}
198
Florin Coras0dbd5172018-06-25 16:19:34 -0700199static int
Dave Barach68b0fb02017-02-28 15:15:56 -0500200tcp_make_synack_options (tcp_connection_t * tc, tcp_options_t * opts)
201{
202 u8 len = 0;
203
204 opts->flags |= TCP_OPTS_FLAG_MSS;
Florin Corasc8343412017-05-04 14:25:50 -0700205 opts->mss = tc->mss;
Dave Barach68b0fb02017-02-28 15:15:56 -0500206 len += TCP_OPTION_LEN_MSS;
207
Florin Coras93992a92017-05-24 18:03:56 -0700208 if (tcp_opts_wscale (&tc->rcv_opts))
Dave Barach68b0fb02017-02-28 15:15:56 -0500209 {
210 opts->flags |= TCP_OPTS_FLAG_WSCALE;
211 opts->wscale = tc->rcv_wscale;
212 len += TCP_OPTION_LEN_WINDOW_SCALE;
213 }
214
Florin Coras93992a92017-05-24 18:03:56 -0700215 if (tcp_opts_tstamp (&tc->rcv_opts))
Dave Barach68b0fb02017-02-28 15:15:56 -0500216 {
217 opts->flags |= TCP_OPTS_FLAG_TSTAMP;
218 opts->tsval = tcp_time_now ();
219 opts->tsecr = tc->tsval_recent;
220 len += TCP_OPTION_LEN_TIMESTAMP;
221 }
222
Florin Coras93992a92017-05-24 18:03:56 -0700223 if (tcp_opts_sack_permitted (&tc->rcv_opts))
Dave Barach68b0fb02017-02-28 15:15:56 -0500224 {
225 opts->flags |= TCP_OPTS_FLAG_SACK_PERMITTED;
226 len += TCP_OPTION_LEN_SACK_PERMITTED;
227 }
228
229 /* Align to needed boundary */
230 len += (TCP_OPTS_ALIGN - len % TCP_OPTS_ALIGN) % TCP_OPTS_ALIGN;
231 return len;
232}
233
Florin Coras0dbd5172018-06-25 16:19:34 -0700234static int
Dave Barach68b0fb02017-02-28 15:15:56 -0500235tcp_make_established_options (tcp_connection_t * tc, tcp_options_t * opts)
236{
237 u8 len = 0;
238
239 opts->flags = 0;
240
Florin Coras93992a92017-05-24 18:03:56 -0700241 if (tcp_opts_tstamp (&tc->rcv_opts))
Dave Barach68b0fb02017-02-28 15:15:56 -0500242 {
243 opts->flags |= TCP_OPTS_FLAG_TSTAMP;
Vladimir Kropylev1cfcb782019-07-02 11:25:26 +0300244 opts->tsval = tcp_tstamp (tc);
Dave Barach68b0fb02017-02-28 15:15:56 -0500245 opts->tsecr = tc->tsval_recent;
246 len += TCP_OPTION_LEN_TIMESTAMP;
247 }
Florin Coras93992a92017-05-24 18:03:56 -0700248 if (tcp_opts_sack_permitted (&tc->rcv_opts))
Dave Barach68b0fb02017-02-28 15:15:56 -0500249 {
250 if (vec_len (tc->snd_sacks))
251 {
252 opts->flags |= TCP_OPTS_FLAG_SACK;
Florin Corase5b17912019-02-21 16:46:24 -0800253 if (tc->snd_sack_pos >= vec_len (tc->snd_sacks))
254 tc->snd_sack_pos = 0;
255 opts->sacks = &tc->snd_sacks[tc->snd_sack_pos];
256 opts->n_sack_blocks = vec_len (tc->snd_sacks) - tc->snd_sack_pos;
257 opts->n_sack_blocks = clib_min (opts->n_sack_blocks,
Florin Corasc28764f2017-04-26 00:08:42 -0700258 TCP_OPTS_MAX_SACK_BLOCKS);
Florin Corase5b17912019-02-21 16:46:24 -0800259 tc->snd_sack_pos += opts->n_sack_blocks;
Dave Barach68b0fb02017-02-28 15:15:56 -0500260 len += 2 + TCP_OPTION_LEN_SACK_BLOCK * opts->n_sack_blocks;
261 }
262 }
263
264 /* Align to needed boundary */
265 len += (TCP_OPTS_ALIGN - len % TCP_OPTS_ALIGN) % TCP_OPTS_ALIGN;
266 return len;
267}
268
269always_inline int
270tcp_make_options (tcp_connection_t * tc, tcp_options_t * opts,
271 tcp_state_t state)
272{
273 switch (state)
274 {
275 case TCP_STATE_ESTABLISHED:
Florin Coras25579b42018-06-06 17:55:02 -0700276 case TCP_STATE_CLOSE_WAIT:
Florin Coras54ddf432018-12-21 13:54:09 -0800277 case TCP_STATE_FIN_WAIT_1:
278 case TCP_STATE_LAST_ACK:
279 case TCP_STATE_CLOSING:
280 case TCP_STATE_FIN_WAIT_2:
281 case TCP_STATE_TIME_WAIT:
282 case TCP_STATE_CLOSED:
Dave Barach68b0fb02017-02-28 15:15:56 -0500283 return tcp_make_established_options (tc, opts);
284 case TCP_STATE_SYN_RCVD:
285 return tcp_make_synack_options (tc, opts);
286 case TCP_STATE_SYN_SENT:
Florin Corascedcf602019-08-27 12:15:43 -0700287 return tcp_make_syn_options (tc, opts);
Dave Barach68b0fb02017-02-28 15:15:56 -0500288 default:
Florin Coras371ca502018-02-21 12:07:41 -0800289 clib_warning ("State not handled! %d", state);
Dave Barach68b0fb02017-02-28 15:15:56 -0500290 return 0;
291 }
292}
293
Florin Corasc8343412017-05-04 14:25:50 -0700294/**
Florin Corasb26743d2018-06-26 09:31:04 -0700295 * Update burst send vars
296 *
297 * - Updates snd_mss to reflect the effective segment size that we can send
298 * by taking into account all TCP options, including SACKs.
299 * - Cache 'on the wire' options for reuse
300 * - Updates receive window which can be reused for a burst.
301 *
302 * This should *only* be called when doing bursts
Florin Corasc8343412017-05-04 14:25:50 -0700303 */
304void
Florin Corasb26743d2018-06-26 09:31:04 -0700305tcp_update_burst_snd_vars (tcp_connection_t * tc)
Florin Corasc8343412017-05-04 14:25:50 -0700306{
Florin Corasb26743d2018-06-26 09:31:04 -0700307 tcp_main_t *tm = &tcp_main;
308
Florin Corasc8343412017-05-04 14:25:50 -0700309 /* Compute options to be used for connection. These may be reused when
310 * sending data or to compute the effective mss (snd_mss) */
Florin Corasb26743d2018-06-26 09:31:04 -0700311 tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts,
312 TCP_STATE_ESTABLISHED);
Florin Corasc8343412017-05-04 14:25:50 -0700313
314 /* XXX check if MTU has been updated */
Florin Coras93992a92017-05-24 18:03:56 -0700315 tc->snd_mss = clib_min (tc->mss, tc->rcv_opts.mss) - tc->snd_opts_len;
Florin Corasdb84e572017-05-09 18:54:52 -0700316 ASSERT (tc->snd_mss > 0);
Florin Corasb26743d2018-06-26 09:31:04 -0700317
318 tcp_options_write (tm->wrk_ctx[tc->c_thread_index].cached_opts,
319 &tc->snd_opts);
320
321 tcp_update_rcv_wnd (tc);
Florin Coras52814732019-06-12 15:38:19 -0700322
Florin Corasbbcfaac2019-10-10 13:52:04 -0700323 if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
Florin Corasd6ae4bf2019-10-12 18:10:20 -0700324 tcp_bt_check_app_limited (tc);
Florin Corasca2831a2019-07-04 17:05:59 -0700325
326 if (tc->snd_una == tc->snd_nxt)
Florin Corasc31dc312019-10-06 14:06:14 -0700327 {
328 tcp_cc_event (tc, TCP_CC_EVT_START_TX);
Florin Coras11e9e352019-11-13 19:09:47 -0800329 tcp_connection_tx_pacer_reset (tc, tc->cwnd, TRANSPORT_PACER_MIN_BURST);
Florin Corasc31dc312019-10-06 14:06:14 -0700330 }
Florin Corasc8343412017-05-04 14:25:50 -0700331}
332
Filip Tehlare275bed2019-03-06 00:06:56 -0800333#endif /* CLIB_MARCH_VARIANT */
Florin Corasc8343412017-05-04 14:25:50 -0700334
Florin Coras0dbd5172018-06-25 16:19:34 -0700335static void *
Dave Barach68b0fb02017-02-28 15:15:56 -0500336tcp_reuse_buffer (vlib_main_t * vm, vlib_buffer_t * b)
337{
Florin Corasb2215d62017-08-01 16:56:58 -0700338 if (b->flags & VLIB_BUFFER_NEXT_PRESENT)
339 vlib_buffer_free_one (vm, b->next_buffer);
Florin Coras4eeeaaf2017-09-05 14:03:37 -0400340 /* Zero all flags but free list index and trace flag */
341 b->flags &= VLIB_BUFFER_NEXT_PRESENT - 1;
Florin Coras1f152cd2017-08-18 19:28:03 -0700342 b->current_data = 0;
343 b->current_length = 0;
344 b->total_length_not_including_first_buffer = 0;
345 vnet_buffer (b)->tcp.flags = 0;
Florin Corasb2215d62017-08-01 16:56:58 -0700346
Dave Barach68b0fb02017-02-28 15:15:56 -0500347 /* Leave enough space for headers */
Florin Coras1ee78302019-02-05 15:51:15 -0800348 return vlib_buffer_make_headroom (b, TRANSPORT_MAX_HDRS_LEN);
Florin Coras1f152cd2017-08-18 19:28:03 -0700349}
350
Filip Tehlare275bed2019-03-06 00:06:56 -0800351#ifndef CLIB_MARCH_VARIANT
Florin Coras0dbd5172018-06-25 16:19:34 -0700352static void *
Florin Coras1f152cd2017-08-18 19:28:03 -0700353tcp_init_buffer (vlib_main_t * vm, vlib_buffer_t * b)
354{
355 ASSERT ((b->flags & VLIB_BUFFER_NEXT_PRESENT) == 0);
Florin Coras4eeeaaf2017-09-05 14:03:37 -0400356 b->flags |= VNET_BUFFER_F_LOCALLY_ORIGINATED;
Florin Coras1f152cd2017-08-18 19:28:03 -0700357 b->total_length_not_including_first_buffer = 0;
Florin Coras24793e32018-05-09 11:34:25 -0700358 b->current_data = 0;
Florin Corasd79b41e2017-03-04 05:37:52 -0800359 vnet_buffer (b)->tcp.flags = 0;
Florin Corasf988e692017-11-27 04:34:14 -0500360 VLIB_BUFFER_TRACE_TRAJECTORY_INIT (b);
Florin Coras1f152cd2017-08-18 19:28:03 -0700361 /* Leave enough space for headers */
Florin Coras1ee78302019-02-05 15:51:15 -0800362 return vlib_buffer_make_headroom (b, TRANSPORT_MAX_HDRS_LEN);
Dave Barach68b0fb02017-02-28 15:15:56 -0500363}
364
Srikanth A02833ff2019-10-02 17:48:58 -0700365
366/* Compute TCP checksum in software when offloading is disabled for a connection */
367u16
368ip6_tcp_compute_checksum_custom (vlib_main_t * vm, vlib_buffer_t * p0,
369 ip46_address_t * src, ip46_address_t * dst)
370{
371 ip_csum_t sum0;
372 u16 payload_length_host_byte_order;
373 u32 i;
374
375 /* Initialize checksum with ip header. */
376 sum0 = clib_host_to_net_u16 (vlib_buffer_length_in_chain (vm, p0)) +
377 clib_host_to_net_u16 (IP_PROTOCOL_TCP);
378 payload_length_host_byte_order = vlib_buffer_length_in_chain (vm, p0);
379
380 for (i = 0; i < ARRAY_LEN (src->ip6.as_uword); i++)
381 {
382 sum0 = ip_csum_with_carry
383 (sum0, clib_mem_unaligned (&src->ip6.as_uword[i], uword));
384 sum0 = ip_csum_with_carry
385 (sum0, clib_mem_unaligned (&dst->ip6.as_uword[i], uword));
386 }
387
388 return ip_calculate_l4_checksum (vm, p0, sum0,
389 payload_length_host_byte_order, NULL, 0,
390 NULL);
391}
392
393u16
394ip4_tcp_compute_checksum_custom (vlib_main_t * vm, vlib_buffer_t * p0,
395 ip46_address_t * src, ip46_address_t * dst)
396{
397 ip_csum_t sum0;
398 u32 payload_length_host_byte_order;
399
400 payload_length_host_byte_order = vlib_buffer_length_in_chain (vm, p0);
401 sum0 =
402 clib_host_to_net_u32 (payload_length_host_byte_order +
403 (IP_PROTOCOL_TCP << 16));
404
405 sum0 = ip_csum_with_carry (sum0, clib_mem_unaligned (&src->ip4, u32));
406 sum0 = ip_csum_with_carry (sum0, clib_mem_unaligned (&dst->ip4, u32));
407
408 return ip_calculate_l4_checksum (vm, p0, sum0,
409 payload_length_host_byte_order, NULL, 0,
410 NULL);
411}
412
413static inline u16
414tcp_compute_checksum (tcp_connection_t * tc, vlib_buffer_t * b)
415{
416 u16 checksum = 0;
Florin Corasbbcfaac2019-10-10 13:52:04 -0700417 if (PREDICT_FALSE (tc->cfg_flags & TCP_CFG_F_NO_CSUM_OFFLOAD))
Srikanth A02833ff2019-10-02 17:48:58 -0700418 {
419 tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
420 vlib_main_t *vm = wrk->vm;
421
422 if (tc->c_is_ip4)
423 checksum = ip4_tcp_compute_checksum_custom
424 (vm, b, &tc->c_lcl_ip, &tc->c_rmt_ip);
425 else
426 checksum = ip6_tcp_compute_checksum_custom
427 (vm, b, &tc->c_lcl_ip, &tc->c_rmt_ip);
428 }
429 else
430 {
431 b->flags |= VNET_BUFFER_F_OFFLOAD_TCP_CKSUM;
432 }
433 return checksum;
434}
435
Dave Barach68b0fb02017-02-28 15:15:56 -0500436/**
437 * Prepare ACK
438 */
Florin Corasbdf7fd62019-01-31 17:31:01 -0800439static inline void
Dave Barach68b0fb02017-02-28 15:15:56 -0500440tcp_make_ack_i (tcp_connection_t * tc, vlib_buffer_t * b, tcp_state_t state,
441 u8 flags)
442{
443 tcp_options_t _snd_opts, *snd_opts = &_snd_opts;
444 u8 tcp_opts_len, tcp_hdr_opts_len;
445 tcp_header_t *th;
446 u16 wnd;
447
448 wnd = tcp_window_to_advertise (tc, state);
449
450 /* Make and write options */
451 tcp_opts_len = tcp_make_established_options (tc, snd_opts);
452 tcp_hdr_opts_len = tcp_opts_len + sizeof (tcp_header_t);
453
454 th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, tc->snd_nxt,
455 tc->rcv_nxt, tcp_hdr_opts_len, flags, wnd);
456
457 tcp_options_write ((u8 *) (th + 1), snd_opts);
Srikanth A02833ff2019-10-02 17:48:58 -0700458
459 th->checksum = tcp_compute_checksum (tc, b);
460
Dave Barach68b0fb02017-02-28 15:15:56 -0500461 vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
Vladimir Kropylev398afbd2019-06-25 00:06:52 +0300462
463 if (wnd == 0)
464 tcp_zero_rwnd_sent_on (tc);
465 else
466 tcp_zero_rwnd_sent_off (tc);
Dave Barach68b0fb02017-02-28 15:15:56 -0500467}
468
469/**
470 * Convert buffer to ACK
471 */
Florin Corasbdf7fd62019-01-31 17:31:01 -0800472static inline void
Dave Barach68b0fb02017-02-28 15:15:56 -0500473tcp_make_ack (tcp_connection_t * tc, vlib_buffer_t * b)
474{
Dave Barach68b0fb02017-02-28 15:15:56 -0500475 tcp_make_ack_i (tc, b, TCP_STATE_ESTABLISHED, TCP_FLAG_ACK);
Florin Corasa436a422019-08-20 07:09:31 -0700476 TCP_EVT (TCP_EVT_ACK_SENT, tc);
Florin Coras3e350af2017-03-30 02:54:28 -0700477 tc->rcv_las = tc->rcv_nxt;
Dave Barach68b0fb02017-02-28 15:15:56 -0500478}
479
480/**
481 * Convert buffer to FIN-ACK
482 */
Florin Coras999840c2020-03-18 20:31:34 +0000483static void
Florin Corasd79b41e2017-03-04 05:37:52 -0800484tcp_make_fin (tcp_connection_t * tc, vlib_buffer_t * b)
Dave Barach68b0fb02017-02-28 15:15:56 -0500485{
Florin Corasbdf7fd62019-01-31 17:31:01 -0800486 tcp_make_ack_i (tc, b, TCP_STATE_ESTABLISHED, TCP_FLAG_FIN | TCP_FLAG_ACK);
Florin Coras4eeeaaf2017-09-05 14:03:37 -0400487}
Dave Barach68b0fb02017-02-28 15:15:56 -0500488
Florin Coras4eeeaaf2017-09-05 14:03:37 -0400489/**
490 * Convert buffer to SYN
491 */
492void
493tcp_make_syn (tcp_connection_t * tc, vlib_buffer_t * b)
494{
495 u8 tcp_hdr_opts_len, tcp_opts_len;
496 tcp_header_t *th;
497 u16 initial_wnd;
498 tcp_options_t snd_opts;
499
500 initial_wnd = tcp_initial_window_to_advertise (tc);
501
502 /* Make and write options */
Dave Barachb7b92992018-10-17 10:38:51 -0400503 clib_memset (&snd_opts, 0, sizeof (snd_opts));
Florin Corascedcf602019-08-27 12:15:43 -0700504 tcp_opts_len = tcp_make_syn_options (tc, &snd_opts);
Florin Coras4eeeaaf2017-09-05 14:03:37 -0400505 tcp_hdr_opts_len = tcp_opts_len + sizeof (tcp_header_t);
506
507 th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, tc->iss,
508 tc->rcv_nxt, tcp_hdr_opts_len, TCP_FLAG_SYN,
509 initial_wnd);
510 vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
511 tcp_options_write ((u8 *) (th + 1), &snd_opts);
Srikanth A02833ff2019-10-02 17:48:58 -0700512 th->checksum = tcp_compute_checksum (tc, b);
Dave Barach68b0fb02017-02-28 15:15:56 -0500513}
514
515/**
516 * Convert buffer to SYN-ACK
517 */
Florin Coras999840c2020-03-18 20:31:34 +0000518static void
Dave Barach68b0fb02017-02-28 15:15:56 -0500519tcp_make_synack (tcp_connection_t * tc, vlib_buffer_t * b)
520{
Dave Barach68b0fb02017-02-28 15:15:56 -0500521 tcp_options_t _snd_opts, *snd_opts = &_snd_opts;
522 u8 tcp_opts_len, tcp_hdr_opts_len;
523 tcp_header_t *th;
524 u16 initial_wnd;
Dave Barach68b0fb02017-02-28 15:15:56 -0500525
Dave Barachb7b92992018-10-17 10:38:51 -0400526 clib_memset (snd_opts, 0, sizeof (*snd_opts));
Dave Barach68b0fb02017-02-28 15:15:56 -0500527 initial_wnd = tcp_initial_window_to_advertise (tc);
Dave Barach68b0fb02017-02-28 15:15:56 -0500528 tcp_opts_len = tcp_make_synack_options (tc, snd_opts);
529 tcp_hdr_opts_len = tcp_opts_len + sizeof (tcp_header_t);
530
531 th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, tc->iss,
532 tc->rcv_nxt, tcp_hdr_opts_len,
533 TCP_FLAG_SYN | TCP_FLAG_ACK, initial_wnd);
Dave Barach68b0fb02017-02-28 15:15:56 -0500534 tcp_options_write ((u8 *) (th + 1), snd_opts);
535
536 vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
Srikanth A02833ff2019-10-02 17:48:58 -0700537 th->checksum = tcp_compute_checksum (tc, b);
Dave Barach68b0fb02017-02-28 15:15:56 -0500538}
539
Florin Coras5484daa2020-03-27 23:55:06 +0000540static void
541tcp_enqueue_to_ip_lookup (tcp_worker_ctx_t * wrk, vlib_buffer_t * b, u32 bi,
542 u8 is_ip4, u32 fib_index)
Dave Barach68b0fb02017-02-28 15:15:56 -0500543{
Florin Coras5484daa2020-03-27 23:55:06 +0000544 tcp_main_t *tm = &tcp_main;
Florin Corasbe72ae62018-11-01 11:23:03 -0700545 vlib_main_t *vm = wrk->vm;
Dave Barach68b0fb02017-02-28 15:15:56 -0500546
Damjan Marion213b5aa2017-07-13 21:19:27 +0200547 b->flags |= VNET_BUFFER_F_LOCALLY_ORIGINATED;
Dave Barach68b0fb02017-02-28 15:15:56 -0500548 b->error = 0;
549
Florin Coras56b39f62018-03-27 17:29:32 -0700550 vnet_buffer (b)->sw_if_index[VLIB_TX] = fib_index;
551 vnet_buffer (b)->sw_if_index[VLIB_RX] = 0;
Dave Barach68b0fb02017-02-28 15:15:56 -0500552
Florin Corasf988e692017-11-27 04:34:14 -0500553 tcp_trajectory_add_start (b, 1);
Dave Barach68b0fb02017-02-28 15:15:56 -0500554
Florin Coras5484daa2020-03-27 23:55:06 +0000555 session_add_pending_tx_buffer (vm->thread_index, bi,
556 tm->ipl_next_node[!is_ip4]);
Florin Coras9d063042017-09-14 03:08:00 -0400557
Florin Coras5484daa2020-03-27 23:55:06 +0000558 if (vm->thread_index == 0 && vlib_num_workers ())
559 session_queue_run_on_main_thread (wrk->vm);
Dave Barach68b0fb02017-02-28 15:15:56 -0500560}
561
Florin Coras0dbd5172018-06-25 16:19:34 -0700562static void
Florin Corasbe72ae62018-11-01 11:23:03 -0700563tcp_enqueue_to_output (tcp_worker_ctx_t * wrk, vlib_buffer_t * b, u32 bi,
564 u8 is_ip4)
Florin Coras1f152cd2017-08-18 19:28:03 -0700565{
Florin Coras2a7ea2e2019-10-16 22:06:08 -0700566 b->flags |= VNET_BUFFER_F_LOCALLY_ORIGINATED;
567 b->error = 0;
568
Florin Coras5484daa2020-03-27 23:55:06 +0000569 session_add_pending_tx_buffer (wrk->vm->thread_index, bi,
570 wrk->tco_next_node[!is_ip4]);
Florin Coras1f152cd2017-08-18 19:28:03 -0700571}
572
Filip Tehlare275bed2019-03-06 00:06:56 -0800573#endif /* CLIB_MARCH_VARIANT */
Florin Coras1f152cd2017-08-18 19:28:03 -0700574
Florin Coras0dbd5172018-06-25 16:19:34 -0700575static int
Florin Coras360336f2020-02-13 18:46:18 +0000576tcp_make_reset_in_place (vlib_main_t * vm, vlib_buffer_t * b, u8 is_ip4)
Dave Barach68b0fb02017-02-28 15:15:56 -0500577{
Dave Barach68b0fb02017-02-28 15:15:56 -0500578 ip4_header_t *ih4;
579 ip6_header_t *ih6;
Florin Coras360336f2020-02-13 18:46:18 +0000580 tcp_header_t *th;
581 ip4_address_t src_ip4, dst_ip4;
582 ip6_address_t src_ip6, dst_ip6;
Florin Corasdc629cd2017-05-09 00:52:37 -0700583 u16 src_port, dst_port;
Florin Coras360336f2020-02-13 18:46:18 +0000584 u32 tmp, len, seq, ack;
Florin Corasdc629cd2017-05-09 00:52:37 -0700585 u8 flags;
Dave Barach68b0fb02017-02-28 15:15:56 -0500586
587 /* Find IP and TCP headers */
Florin Coras360336f2020-02-13 18:46:18 +0000588 th = tcp_buffer_hdr (b);
Florin Corasdc629cd2017-05-09 00:52:37 -0700589
590 /* Save src and dst ip */
Dave Barach68b0fb02017-02-28 15:15:56 -0500591 if (is_ip4)
592 {
Florin Coras360336f2020-02-13 18:46:18 +0000593 ih4 = vlib_buffer_get_current (b);
Florin Corasdc629cd2017-05-09 00:52:37 -0700594 ASSERT ((ih4->ip_version_and_header_length & 0xF0) == 0x40);
Florin Coras360336f2020-02-13 18:46:18 +0000595 src_ip4.as_u32 = ih4->src_address.as_u32;
596 dst_ip4.as_u32 = ih4->dst_address.as_u32;
Dave Barach68b0fb02017-02-28 15:15:56 -0500597 }
598 else
599 {
Florin Coras360336f2020-02-13 18:46:18 +0000600 ih6 = vlib_buffer_get_current (b);
Dave Barach68b0fb02017-02-28 15:15:56 -0500601 ASSERT ((ih6->ip_version_traffic_class_and_flow_label & 0xF0) == 0x60);
Florin Coras360336f2020-02-13 18:46:18 +0000602 clib_memcpy_fast (&src_ip6, &ih6->src_address, sizeof (ip6_address_t));
603 clib_memcpy_fast (&dst_ip6, &ih6->dst_address, sizeof (ip6_address_t));
Dave Barach68b0fb02017-02-28 15:15:56 -0500604 }
605
Florin Coras360336f2020-02-13 18:46:18 +0000606 src_port = th->src_port;
607 dst_port = th->dst_port;
608 flags = TCP_FLAG_RST;
Florin Corasdc629cd2017-05-09 00:52:37 -0700609
Florin Coras360336f2020-02-13 18:46:18 +0000610 /*
611 * RFC 793. If the ACK bit is off, sequence number zero is used,
612 * <SEQ=0><ACK=SEG.SEQ+SEG.LEN><CTL=RST,ACK>
613 * If the ACK bit is on,
614 * <SEQ=SEG.ACK><CTL=RST>
615 */
616 if (tcp_ack (th))
Dave Barach68b0fb02017-02-28 15:15:56 -0500617 {
Florin Coras360336f2020-02-13 18:46:18 +0000618 seq = th->ack_number;
Florin Corasdc629cd2017-05-09 00:52:37 -0700619 ack = 0;
Dave Barach68b0fb02017-02-28 15:15:56 -0500620 }
Florin Coras360336f2020-02-13 18:46:18 +0000621 else
622 {
623 flags |= TCP_FLAG_ACK;
624 tmp = clib_net_to_host_u32 (th->seq_number);
625 len = vnet_buffer (b)->tcp.data_len + tcp_is_syn (th) + tcp_is_fin (th);
626 ack = clib_host_to_net_u32 (tmp + len);
627 seq = 0;
628 }
Dave Barach68b0fb02017-02-28 15:15:56 -0500629
Florin Coras360336f2020-02-13 18:46:18 +0000630 tcp_reuse_buffer (vm, b);
631 tcp_trajectory_add_start (b, 4);
632 th = vlib_buffer_push_tcp_net_order (b, dst_port, src_port, seq, ack,
633 sizeof (tcp_header_t), flags, 0);
Dave Barach68b0fb02017-02-28 15:15:56 -0500634
Dave Barach68b0fb02017-02-28 15:15:56 -0500635 if (is_ip4)
636 {
Florin Coras360336f2020-02-13 18:46:18 +0000637 ih4 = vlib_buffer_push_ip4 (vm, b, &dst_ip4, &src_ip4,
Florin Corasfdbc3822017-07-27 00:34:12 -0700638 IP_PROTOCOL_TCP, 1);
Florin Coras360336f2020-02-13 18:46:18 +0000639 th->checksum = ip4_tcp_udp_compute_checksum (vm, b, ih4);
Dave Barach68b0fb02017-02-28 15:15:56 -0500640 }
641 else
642 {
643 int bogus = ~0;
Florin Coras360336f2020-02-13 18:46:18 +0000644 ih6 = vlib_buffer_push_ip6 (vm, b, &dst_ip6, &src_ip6, IP_PROTOCOL_TCP);
645 th->checksum = ip6_tcp_udp_icmp_compute_checksum (vm, b, ih6, &bogus);
Dave Barach68b0fb02017-02-28 15:15:56 -0500646 ASSERT (!bogus);
647 }
648
649 return 0;
650}
651
Filip Tehlare275bed2019-03-06 00:06:56 -0800652#ifndef CLIB_MARCH_VARIANT
Dave Barach68b0fb02017-02-28 15:15:56 -0500653/**
654 * Send reset without reusing existing buffer
Florin Coras1f152cd2017-08-18 19:28:03 -0700655 *
656 * It extracts connection info out of original packet
Dave Barach68b0fb02017-02-28 15:15:56 -0500657 */
658void
Florin Corasd4c49be2019-02-07 00:15:53 -0800659tcp_send_reset_w_pkt (tcp_connection_t * tc, vlib_buffer_t * pkt,
660 u32 thread_index, u8 is_ip4)
Dave Barach68b0fb02017-02-28 15:15:56 -0500661{
Florin Corasd4c49be2019-02-07 00:15:53 -0800662 tcp_worker_ctx_t *wrk = tcp_get_worker (thread_index);
Florin Corasbe72ae62018-11-01 11:23:03 -0700663 vlib_main_t *vm = wrk->vm;
Dave Barach68b0fb02017-02-28 15:15:56 -0500664 vlib_buffer_t *b;
Florin Coras56b39f62018-03-27 17:29:32 -0700665 u32 bi, sw_if_index, fib_index;
Dave Barach68b0fb02017-02-28 15:15:56 -0500666 u8 tcp_hdr_len, flags = 0;
667 tcp_header_t *th, *pkt_th;
668 u32 seq, ack;
669 ip4_header_t *ih4, *pkt_ih4;
670 ip6_header_t *ih6, *pkt_ih6;
Florin Coras56b39f62018-03-27 17:29:32 -0700671 fib_protocol_t fib_proto;
Dave Barach68b0fb02017-02-28 15:15:56 -0500672
Florin Corasbdf7fd62019-01-31 17:31:01 -0800673 if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
Florin Coras66b11312017-07-31 17:18:03 -0700674 return;
675
Dave Barach68b0fb02017-02-28 15:15:56 -0500676 b = vlib_get_buffer (vm, bi);
Florin Coras56b39f62018-03-27 17:29:32 -0700677 sw_if_index = vnet_buffer (pkt)->sw_if_index[VLIB_RX];
678 fib_proto = is_ip4 ? FIB_PROTOCOL_IP4 : FIB_PROTOCOL_IP6;
679 fib_index = fib_table_get_index_for_sw_if_index (fib_proto, sw_if_index);
Florin Coras1f152cd2017-08-18 19:28:03 -0700680 tcp_init_buffer (vm, b);
Dave Barach68b0fb02017-02-28 15:15:56 -0500681
682 /* Make and write options */
683 tcp_hdr_len = sizeof (tcp_header_t);
684
685 if (is_ip4)
686 {
687 pkt_ih4 = vlib_buffer_get_current (pkt);
688 pkt_th = ip4_next_header (pkt_ih4);
689 }
690 else
691 {
692 pkt_ih6 = vlib_buffer_get_current (pkt);
693 pkt_th = ip6_next_header (pkt_ih6);
694 }
695
696 if (tcp_ack (pkt_th))
697 {
698 flags = TCP_FLAG_RST;
699 seq = pkt_th->ack_number;
Florin Coras776f3d82018-11-02 08:23:58 -0700700 ack = (tc->state >= TCP_STATE_SYN_RCVD) ? tc->rcv_nxt : 0;
Dave Barach68b0fb02017-02-28 15:15:56 -0500701 }
702 else
703 {
704 flags = TCP_FLAG_RST | TCP_FLAG_ACK;
705 seq = 0;
706 ack = clib_host_to_net_u32 (vnet_buffer (pkt)->tcp.seq_end);
707 }
708
709 th = vlib_buffer_push_tcp_net_order (b, pkt_th->dst_port, pkt_th->src_port,
710 seq, ack, tcp_hdr_len, flags, 0);
711
712 /* Swap src and dst ip */
713 if (is_ip4)
714 {
715 ASSERT ((pkt_ih4->ip_version_and_header_length & 0xF0) == 0x40);
716 ih4 = vlib_buffer_push_ip4 (vm, b, &pkt_ih4->dst_address,
Srikanth A02833ff2019-10-02 17:48:58 -0700717 &pkt_ih4->src_address, IP_PROTOCOL_TCP,
Florin Corasbbcfaac2019-10-10 13:52:04 -0700718 tcp_csum_offload (tc));
Dave Barach68b0fb02017-02-28 15:15:56 -0500719 th->checksum = ip4_tcp_udp_compute_checksum (vm, b, ih4);
720 }
721 else
722 {
723 int bogus = ~0;
Dave Barach68b0fb02017-02-28 15:15:56 -0500724 ASSERT ((pkt_ih6->ip_version_traffic_class_and_flow_label & 0xF0) ==
725 0x60);
Tarun Gupta2089c692019-11-04 16:35:59 -0800726 ih6 = vlib_buffer_push_ip6_custom (vm, b, &pkt_ih6->dst_address,
727 &pkt_ih6->src_address,
728 IP_PROTOCOL_TCP,
729 tc->ipv6_flow_label);
Dave Barach68b0fb02017-02-28 15:15:56 -0500730 th->checksum = ip6_tcp_udp_icmp_compute_checksum (vm, b, ih6, &bogus);
731 ASSERT (!bogus);
732 }
733
Florin Coras5484daa2020-03-27 23:55:06 +0000734 tcp_enqueue_to_ip_lookup (wrk, b, bi, is_ip4, fib_index);
Florin Corasa436a422019-08-20 07:09:31 -0700735 TCP_EVT (TCP_EVT_RST_SENT, tc);
Florin Coras1f421012019-07-26 10:18:51 -0700736 vlib_node_increment_counter (vm, tcp_node_index (output, tc->c_is_ip4),
737 TCP_ERROR_RST_SENT, 1);
Dave Barach68b0fb02017-02-28 15:15:56 -0500738}
739
Florin Coras1f152cd2017-08-18 19:28:03 -0700740/**
741 * Build and set reset packet for connection
742 */
743void
744tcp_send_reset (tcp_connection_t * tc)
745{
Florin Corasbe72ae62018-11-01 11:23:03 -0700746 tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
747 vlib_main_t *vm = wrk->vm;
Florin Coras1f152cd2017-08-18 19:28:03 -0700748 vlib_buffer_t *b;
749 u32 bi;
750 tcp_header_t *th;
751 u16 tcp_hdr_opts_len, advertise_wnd, opts_write_len;
752 u8 flags;
753
Florin Corasbdf7fd62019-01-31 17:31:01 -0800754 if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
Florin Coras1f152cd2017-08-18 19:28:03 -0700755 return;
756 b = vlib_get_buffer (vm, bi);
757 tcp_init_buffer (vm, b);
758
759 tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
760 tcp_hdr_opts_len = tc->snd_opts_len + sizeof (tcp_header_t);
Florin Coras4e783b92020-03-23 23:24:19 +0000761 advertise_wnd = tc->rcv_wnd >> tc->rcv_wscale;
Florin Coras1f152cd2017-08-18 19:28:03 -0700762 flags = TCP_FLAG_RST;
763 th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, tc->snd_nxt,
764 tc->rcv_nxt, tcp_hdr_opts_len, flags,
765 advertise_wnd);
766 opts_write_len = tcp_options_write ((u8 *) (th + 1), &tc->snd_opts);
Srikanth A02833ff2019-10-02 17:48:58 -0700767 th->checksum = tcp_compute_checksum (tc, b);
Florin Coras1f152cd2017-08-18 19:28:03 -0700768 ASSERT (opts_write_len == tc->snd_opts_len);
769 vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
Florin Corasdf36f492019-08-18 18:09:28 -0700770 tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
Florin Corasa436a422019-08-20 07:09:31 -0700771 TCP_EVT (TCP_EVT_RST_SENT, tc);
Florin Coras1f421012019-07-26 10:18:51 -0700772 vlib_node_increment_counter (vm, tcp_node_index (output, tc->c_is_ip4),
773 TCP_ERROR_RST_SENT, 1);
Florin Coras1f152cd2017-08-18 19:28:03 -0700774}
775
Florin Coras0dbd5172018-06-25 16:19:34 -0700776static void
Florin Corasbe72ae62018-11-01 11:23:03 -0700777tcp_push_ip_hdr (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
778 vlib_buffer_t * b)
Dave Barach68b0fb02017-02-28 15:15:56 -0500779{
Dave Barach68b0fb02017-02-28 15:15:56 -0500780 if (tc->c_is_ip4)
781 {
Florin Corasf4ce6ba2019-11-20 18:34:58 -0800782 vlib_buffer_push_ip4 (wrk->vm, b, &tc->c_lcl_ip4, &tc->c_rmt_ip4,
783 IP_PROTOCOL_TCP, tcp_csum_offload (tc));
Dave Barach68b0fb02017-02-28 15:15:56 -0500784 }
785 else
786 {
Florin Corasf4ce6ba2019-11-20 18:34:58 -0800787 vlib_buffer_push_ip6_custom (wrk->vm, b, &tc->c_lcl_ip6, &tc->c_rmt_ip6,
788 IP_PROTOCOL_TCP, tc->ipv6_flow_label);
Dave Barach68b0fb02017-02-28 15:15:56 -0500789 }
790}
791
792/**
793 * Send SYN
794 *
795 * Builds a SYN packet for a half-open connection and sends it to ipx_lookup.
796 * The packet is not forwarded through tcpx_output to avoid doing lookups
797 * in the half_open pool.
798 */
799void
800tcp_send_syn (tcp_connection_t * tc)
801{
Florin Corasbe72ae62018-11-01 11:23:03 -0700802 tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
803 vlib_main_t *vm = wrk->vm;
Dave Barach68b0fb02017-02-28 15:15:56 -0500804 vlib_buffer_t *b;
805 u32 bi;
Dave Barach68b0fb02017-02-28 15:15:56 -0500806
Florin Corasf988e692017-11-27 04:34:14 -0500807 /*
808 * Setup retransmit and establish timers before requesting buffer
809 * such that we can return if we've ran out.
810 */
Florin Coras0765d972020-03-18 21:26:41 +0000811 tcp_timer_update (&wrk->timer_wheel, tc, TCP_TIMER_RETRANSMIT_SYN,
Florin Corasf988e692017-11-27 04:34:14 -0500812 tc->rto * TCP_TO_TIMER_TICK);
813
Florin Corasbdf7fd62019-01-31 17:31:01 -0800814 if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
Florin Coras222e1f412019-02-16 20:47:32 -0800815 {
Florin Coras0765d972020-03-18 21:26:41 +0000816 tcp_timer_update (&wrk->timer_wheel, tc, TCP_TIMER_RETRANSMIT_SYN, 1);
Florin Coras222e1f412019-02-16 20:47:32 -0800817 return;
818 }
Florin Coras66b11312017-07-31 17:18:03 -0700819
Dave Barach68b0fb02017-02-28 15:15:56 -0500820 b = vlib_get_buffer (vm, bi);
Florin Coras1f152cd2017-08-18 19:28:03 -0700821 tcp_init_buffer (vm, b);
Florin Coras4eeeaaf2017-09-05 14:03:37 -0400822 tcp_make_syn (tc, b);
Dave Barach68b0fb02017-02-28 15:15:56 -0500823
824 /* Measure RTT with this */
Florin Corasefefc6b2018-11-07 12:49:19 -0800825 tc->rtt_ts = tcp_time_now_us (vlib_num_workers ()? 1 : 0);
Dave Barach68b0fb02017-02-28 15:15:56 -0500826 tc->rtt_seq = tc->snd_nxt;
Dave Barach68b0fb02017-02-28 15:15:56 -0500827 tc->rto_boff = 0;
828
Florin Corasbe72ae62018-11-01 11:23:03 -0700829 tcp_push_ip_hdr (wrk, tc, b);
830 tcp_enqueue_to_ip_lookup (wrk, b, bi, tc->c_is_ip4, tc->c_fib_index);
Florin Corasa436a422019-08-20 07:09:31 -0700831 TCP_EVT (TCP_EVT_SYN_SENT, tc);
Dave Barach68b0fb02017-02-28 15:15:56 -0500832}
833
Florin Coras7ac053b2018-11-05 15:57:21 -0800834void
835tcp_send_synack (tcp_connection_t * tc)
836{
837 tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
838 vlib_main_t *vm = wrk->vm;
839 vlib_buffer_t *b;
840 u32 bi;
841
Florin Coras0765d972020-03-18 21:26:41 +0000842 tcp_retransmit_timer_force_update (&wrk->timer_wheel, tc);
Florin Coras222e1f412019-02-16 20:47:32 -0800843
Florin Corasbdf7fd62019-01-31 17:31:01 -0800844 if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
Florin Coras222e1f412019-02-16 20:47:32 -0800845 {
Florin Coras0765d972020-03-18 21:26:41 +0000846 tcp_timer_update (&wrk->timer_wheel, tc, TCP_TIMER_RETRANSMIT, 1);
Florin Coras222e1f412019-02-16 20:47:32 -0800847 return;
848 }
Florin Coras7ac053b2018-11-05 15:57:21 -0800849
Florin Corasdf084782018-12-06 17:41:10 -0800850 tc->rtt_ts = tcp_time_now_us (tc->c_thread_index);
Florin Coras7ac053b2018-11-05 15:57:21 -0800851 b = vlib_get_buffer (vm, bi);
Florin Corasbdf7fd62019-01-31 17:31:01 -0800852 tcp_init_buffer (vm, b);
Florin Coras7ac053b2018-11-05 15:57:21 -0800853 tcp_make_synack (tc, b);
854 tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
Florin Corasa436a422019-08-20 07:09:31 -0700855 TCP_EVT (TCP_EVT_SYNACK_SENT, tc);
Florin Coras7ac053b2018-11-05 15:57:21 -0800856}
857
Florin Coras66b11312017-07-31 17:18:03 -0700858/**
Dave Barach68b0fb02017-02-28 15:15:56 -0500859 * Send FIN
860 */
861void
862tcp_send_fin (tcp_connection_t * tc)
863{
Florin Corasbe72ae62018-11-01 11:23:03 -0700864 tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
865 vlib_main_t *vm = wrk->vm;
Florin Coras9d063042017-09-14 03:08:00 -0400866 vlib_buffer_t *b;
867 u32 bi;
868 u8 fin_snt = 0;
869
Florin Corasc977e7c2018-10-16 20:30:31 -0700870 fin_snt = tc->flags & TCP_CONN_FINSNT;
871 if (fin_snt)
Florin Coras47596832019-03-12 18:58:54 -0700872 tc->snd_nxt -= 1;
Florin Corasc977e7c2018-10-16 20:30:31 -0700873
Florin Corasbdf7fd62019-01-31 17:31:01 -0800874 if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
Florin Coraseb97e5f2018-10-15 21:35:42 -0700875 {
876 /* Out of buffers so program fin retransmit ASAP */
Florin Coras0765d972020-03-18 21:26:41 +0000877 tcp_timer_update (&wrk->timer_wheel, tc, TCP_TIMER_RETRANSMIT, 1);
Florin Coras85a3ddd2018-12-24 16:54:34 -0800878 if (fin_snt)
Florin Coras47596832019-03-12 18:58:54 -0700879 tc->snd_nxt += 1;
Florin Coras222e1f412019-02-16 20:47:32 -0800880 else
881 /* Make sure retransmit retries a fin not data */
882 tc->flags |= TCP_CONN_FINSNT;
Florin Coras85a3ddd2018-12-24 16:54:34 -0800883 return;
Florin Coraseb97e5f2018-10-15 21:35:42 -0700884 }
885
Florin Corascb711a42019-10-16 19:28:17 -0700886 /* If we have non-dupacks programmed, no need to send them */
887 if ((tc->flags & TCP_CONN_SNDACK) && !tc->pending_dupacks)
888 tc->flags &= ~TCP_CONN_SNDACK;
889
Florin Coras0765d972020-03-18 21:26:41 +0000890 tcp_retransmit_timer_force_update (&wrk->timer_wheel, tc);
Dave Barach68b0fb02017-02-28 15:15:56 -0500891 b = vlib_get_buffer (vm, bi);
Florin Coras24793e32018-05-09 11:34:25 -0700892 tcp_init_buffer (vm, b);
Florin Corasd79b41e2017-03-04 05:37:52 -0800893 tcp_make_fin (tc, b);
Florin Coras2a7ea2e2019-10-16 22:06:08 -0700894 tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
Florin Corasa436a422019-08-20 07:09:31 -0700895 TCP_EVT (TCP_EVT_FIN_SENT, tc);
Florin Coras47596832019-03-12 18:58:54 -0700896 /* Account for the FIN */
897 tc->snd_nxt += 1;
Florin Coras9d063042017-09-14 03:08:00 -0400898 if (!fin_snt)
Florin Coras4eeeaaf2017-09-05 14:03:37 -0400899 {
900 tc->flags |= TCP_CONN_FINSNT;
901 tc->flags &= ~TCP_CONN_FINPNDG;
Florin Coras47596832019-03-12 18:58:54 -0700902 tc->snd_una_max = seq_max (tc->snd_una_max, tc->snd_nxt);
Florin Corasa096f2d2017-09-28 23:49:42 -0400903 }
Dave Barach68b0fb02017-02-28 15:15:56 -0500904}
905
Dave Barach68b0fb02017-02-28 15:15:56 -0500906/**
Florin Corasfd247442019-03-12 16:56:26 -0700907 * Push TCP header and update connection variables. Should only be called
908 * for segments with data, not for 'control' packets.
Dave Barach68b0fb02017-02-28 15:15:56 -0500909 */
Florin Coras0dbd5172018-06-25 16:19:34 -0700910always_inline void
Florin Coras47596832019-03-12 18:58:54 -0700911tcp_push_hdr_i (tcp_connection_t * tc, vlib_buffer_t * b, u32 snd_nxt,
912 u8 compute_opts, u8 maybe_burst, u8 update_snd_nxt)
Dave Barach68b0fb02017-02-28 15:15:56 -0500913{
Florin Corasfd247442019-03-12 16:56:26 -0700914 u8 tcp_hdr_opts_len, flags = TCP_FLAG_ACK;
Dave Barach68b0fb02017-02-28 15:15:56 -0500915 u32 advertise_wnd, data_len;
Florin Corasb26743d2018-06-26 09:31:04 -0700916 tcp_main_t *tm = &tcp_main;
Dave Barach68b0fb02017-02-28 15:15:56 -0500917 tcp_header_t *th;
918
Florin Corasb26743d2018-06-26 09:31:04 -0700919 data_len = b->current_length;
920 if (PREDICT_FALSE (b->flags & VLIB_BUFFER_NEXT_PRESENT))
921 data_len += b->total_length_not_including_first_buffer;
922
Dave Barach68b0fb02017-02-28 15:15:56 -0500923 vnet_buffer (b)->tcp.flags = 0;
Florin Corasb26743d2018-06-26 09:31:04 -0700924 vnet_buffer (b)->tcp.connection_index = tc->c_c_index;
Dave Barach68b0fb02017-02-28 15:15:56 -0500925
Florin Corasc8343412017-05-04 14:25:50 -0700926 if (compute_opts)
927 tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
928
Florin Corasc8343412017-05-04 14:25:50 -0700929 tcp_hdr_opts_len = tc->snd_opts_len + sizeof (tcp_header_t);
Florin Corasb26743d2018-06-26 09:31:04 -0700930
931 if (maybe_burst)
932 advertise_wnd = tc->rcv_wnd >> tc->rcv_wscale;
933 else
Florin Coras47596832019-03-12 18:58:54 -0700934 advertise_wnd = tcp_window_to_advertise (tc, TCP_STATE_ESTABLISHED);
Florin Corasb26743d2018-06-26 09:31:04 -0700935
Florin Coras42ceddb2018-12-12 10:56:01 -0800936 if (PREDICT_FALSE (tc->flags & TCP_CONN_PSH_PENDING))
937 {
Florin Coras47596832019-03-12 18:58:54 -0700938 if (seq_geq (tc->psh_seq, snd_nxt)
939 && seq_lt (tc->psh_seq, snd_nxt + data_len))
Florin Coras42ceddb2018-12-12 10:56:01 -0800940 flags |= TCP_FLAG_PSH;
941 }
Florin Coras47596832019-03-12 18:58:54 -0700942 th = vlib_buffer_push_tcp (b, tc->c_lcl_port, tc->c_rmt_port, snd_nxt,
Dave Barach68b0fb02017-02-28 15:15:56 -0500943 tc->rcv_nxt, tcp_hdr_opts_len, flags,
944 advertise_wnd);
Dave Barach68b0fb02017-02-28 15:15:56 -0500945
Florin Corasb26743d2018-06-26 09:31:04 -0700946 if (maybe_burst)
947 {
Dave Barach178cf492018-11-13 16:34:13 -0500948 clib_memcpy_fast ((u8 *) (th + 1),
949 tm->wrk_ctx[tc->c_thread_index].cached_opts,
950 tc->snd_opts_len);
Florin Corasb26743d2018-06-26 09:31:04 -0700951 }
952 else
953 {
954 u8 len = tcp_options_write ((u8 *) (th + 1), &tc->snd_opts);
955 ASSERT (len == tc->snd_opts_len);
956 }
Dave Barach68b0fb02017-02-28 15:15:56 -0500957
Florin Coras93992a92017-05-24 18:03:56 -0700958 /*
959 * Update connection variables
960 */
961
Florin Coras47596832019-03-12 18:58:54 -0700962 if (update_snd_nxt)
963 tc->snd_nxt += data_len;
Florin Corasc28764f2017-04-26 00:08:42 -0700964 tc->rcv_las = tc->rcv_nxt;
Florin Coras3e350af2017-03-30 02:54:28 -0700965
Florin Corasedfe0ee2019-07-29 18:13:25 -0700966 tc->bytes_out += data_len;
967 tc->data_segs_out += 1;
968
Srikanth A02833ff2019-10-02 17:48:58 -0700969 th->checksum = tcp_compute_checksum (tc, b);
970
Florin Corasa436a422019-08-20 07:09:31 -0700971 TCP_EVT (TCP_EVT_PKTIZE, tc);
Dave Barach68b0fb02017-02-28 15:15:56 -0500972}
973
Florin Corasd6ae4bf2019-10-12 18:10:20 -0700974always_inline u32
975tcp_buffer_len (vlib_buffer_t * b)
976{
977 u32 data_len = b->current_length;
978 if (PREDICT_FALSE (b->flags & VLIB_BUFFER_NEXT_PRESENT))
979 data_len += b->total_length_not_including_first_buffer;
980 return data_len;
981}
982
Florin Coras0dbd5172018-06-25 16:19:34 -0700983u32
Florin Coras14ed6df2019-03-06 21:13:42 -0800984tcp_session_push_header (transport_connection_t * tconn, vlib_buffer_t * b)
Florin Coras0dbd5172018-06-25 16:19:34 -0700985{
Florin Coras14ed6df2019-03-06 21:13:42 -0800986 tcp_connection_t *tc = (tcp_connection_t *) tconn;
Florin Coras52814732019-06-12 15:38:19 -0700987
Florin Corasd6ae4bf2019-10-12 18:10:20 -0700988 if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
989 tcp_bt_track_tx (tc, tcp_buffer_len (b));
Florin Coras52814732019-06-12 15:38:19 -0700990
Florin Coras47596832019-03-12 18:58:54 -0700991 tcp_push_hdr_i (tc, b, tc->snd_nxt, /* compute opts */ 0, /* burst */ 1,
992 /* update_snd_nxt */ 1);
Florin Coras52814732019-06-12 15:38:19 -0700993
Florin Coras47596832019-03-12 18:58:54 -0700994 tc->snd_una_max = seq_max (tc->snd_nxt, tc->snd_una_max);
Florin Coras0dbd5172018-06-25 16:19:34 -0700995 tcp_validate_txf_size (tc, tc->snd_una_max - tc->snd_una);
996 /* If not tracking an ACK, start tracking */
997 if (tc->rtt_ts == 0 && !tcp_in_cong_recovery (tc))
998 {
Florin Corasd67f1122018-05-21 17:47:40 -0700999 tc->rtt_ts = tcp_time_now_us (tc->c_thread_index);
Florin Coras0dbd5172018-06-25 16:19:34 -07001000 tc->rtt_seq = tc->snd_nxt;
1001 }
1002 if (PREDICT_FALSE (!tcp_timer_is_active (tc, TCP_TIMER_RETRANSMIT)))
1003 {
Florin Coras0765d972020-03-18 21:26:41 +00001004 tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
1005 tcp_retransmit_timer_set (&wrk->timer_wheel, tc);
Florin Coras0dbd5172018-06-25 16:19:34 -07001006 tc->rto_boff = 0;
1007 }
1008 tcp_trajectory_add_start (b, 3);
1009 return 0;
1010}
1011
Dave Barach68b0fb02017-02-28 15:15:56 -05001012void
Florin Coras6792ec02017-03-13 03:49:51 -07001013tcp_send_ack (tcp_connection_t * tc)
Dave Barach68b0fb02017-02-28 15:15:56 -05001014{
Florin Corasbe72ae62018-11-01 11:23:03 -07001015 tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
1016 vlib_main_t *vm = wrk->vm;
Dave Barach68b0fb02017-02-28 15:15:56 -05001017 vlib_buffer_t *b;
1018 u32 bi;
1019
Florin Corasbdf7fd62019-01-31 17:31:01 -08001020 if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
Florin Coras222e1f412019-02-16 20:47:32 -08001021 {
1022 tcp_update_rcv_wnd (tc);
1023 return;
1024 }
Dave Barach68b0fb02017-02-28 15:15:56 -05001025 b = vlib_get_buffer (vm, bi);
Florin Coras24793e32018-05-09 11:34:25 -07001026 tcp_init_buffer (vm, b);
Dave Barach68b0fb02017-02-28 15:15:56 -05001027 tcp_make_ack (tc, b);
Florin Corasbe72ae62018-11-01 11:23:03 -07001028 tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
Dave Barach68b0fb02017-02-28 15:15:56 -05001029}
1030
Florin Coras7ac053b2018-11-05 15:57:21 -08001031void
Florin Coras26dd6de2019-07-23 23:54:47 -07001032tcp_program_ack (tcp_connection_t * tc)
Florin Coras7ac053b2018-11-05 15:57:21 -08001033{
1034 if (!(tc->flags & TCP_CONN_SNDACK))
1035 {
Florin Coras26dd6de2019-07-23 23:54:47 -07001036 session_add_self_custom_tx_evt (&tc->connection, 1);
Florin Coras7ac053b2018-11-05 15:57:21 -08001037 tc->flags |= TCP_CONN_SNDACK;
1038 }
1039}
1040
1041void
Florin Coras26dd6de2019-07-23 23:54:47 -07001042tcp_program_dupack (tcp_connection_t * tc)
Florin Coras7ac053b2018-11-05 15:57:21 -08001043{
1044 if (!(tc->flags & TCP_CONN_SNDACK))
1045 {
Florin Coras26dd6de2019-07-23 23:54:47 -07001046 session_add_self_custom_tx_evt (&tc->connection, 1);
Florin Coras7ac053b2018-11-05 15:57:21 -08001047 tc->flags |= TCP_CONN_SNDACK;
1048 }
1049 if (tc->pending_dupacks < 255)
1050 tc->pending_dupacks += 1;
1051}
1052
1053void
Florin Coras36ebcff2019-09-12 18:36:44 -07001054tcp_program_retransmit (tcp_connection_t * tc)
Florin Coras7ac053b2018-11-05 15:57:21 -08001055{
Florin Coras36ebcff2019-09-12 18:36:44 -07001056 if (!(tc->flags & TCP_CONN_RXT_PENDING))
Florin Coras7ac053b2018-11-05 15:57:21 -08001057 {
Florin Coras26dd6de2019-07-23 23:54:47 -07001058 session_add_self_custom_tx_evt (&tc->connection, 0);
Florin Coras36ebcff2019-09-12 18:36:44 -07001059 tc->flags |= TCP_CONN_RXT_PENDING;
Florin Coras7ac053b2018-11-05 15:57:21 -08001060 }
Florin Coras7ac053b2018-11-05 15:57:21 -08001061}
1062
Florin Corasb2215d62017-08-01 16:56:58 -07001063/**
1064 * Delayed ack timer handler
1065 *
1066 * Sends delayed ACK when timer expires
1067 */
Florin Coras6792ec02017-03-13 03:49:51 -07001068void
Florin Corasaa388692020-02-14 23:41:25 +00001069tcp_timer_delack_handler (tcp_connection_t * tc)
Florin Coras6792ec02017-03-13 03:49:51 -07001070{
Florin Coras6792ec02017-03-13 03:49:51 -07001071 tcp_send_ack (tc);
1072}
1073
Florin Corasb2215d62017-08-01 16:56:58 -07001074/**
Florin Coras017dc452019-08-30 11:06:35 -07001075 * Send window update ack
1076 *
1077 * Ensures that it will be sent only once, after a zero rwnd has been
1078 * advertised in a previous ack, and only if rwnd has grown beyond a
1079 * configurable value.
Vladimir Kropylev398afbd2019-06-25 00:06:52 +03001080 */
1081void
1082tcp_send_window_update_ack (tcp_connection_t * tc)
1083{
Vladimir Kropylev398afbd2019-06-25 00:06:52 +03001084 if (tcp_zero_rwnd_sent (tc))
1085 {
Florin Coras017dc452019-08-30 11:06:35 -07001086 tcp_update_rcv_wnd (tc);
1087 if (tc->rcv_wnd >= tcp_cfg.rwnd_min_update_ack * tc->snd_mss)
Vladimir Kropylev398afbd2019-06-25 00:06:52 +03001088 {
1089 tcp_zero_rwnd_sent_off (tc);
Florin Coras26dd6de2019-07-23 23:54:47 -07001090 tcp_program_ack (tc);
Vladimir Kropylev398afbd2019-06-25 00:06:52 +03001091 }
1092 }
1093}
1094
1095/**
Florin Coras36ee9f12018-11-02 12:52:10 -07001096 * Allocate a new buffer and build a new tcp segment
Dave Barach68b0fb02017-02-28 15:15:56 -05001097 *
Florin Coras36ee9f12018-11-02 12:52:10 -07001098 * @param wrk tcp worker
1099 * @param tc connection for which the segment will be allocated
1100 * @param offset offset of the first byte in the tx fifo
1101 * @param max_deq_byte segment size
1102 * @param[out] b pointer to buffer allocated
1103 *
1104 * @return the number of bytes in the segment or 0 if buffer cannot be
1105 * allocated or no data available
Florin Coras93992a92017-05-24 18:03:56 -07001106 */
Florin Coras36ee9f12018-11-02 12:52:10 -07001107static int
1108tcp_prepare_segment (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
1109 u32 offset, u32 max_deq_bytes, vlib_buffer_t ** b)
Dave Barach68b0fb02017-02-28 15:15:56 -05001110{
Florin Corasbe72ae62018-11-01 11:23:03 -07001111 u32 bytes_per_buffer = vnet_get_tcp_main ()->bytes_per_buffer;
1112 vlib_main_t *vm = wrk->vm;
Florin Corasbdf7fd62019-01-31 17:31:01 -08001113 u32 bi, seg_size;
Florin Coras93992a92017-05-24 18:03:56 -07001114 int n_bytes = 0;
Florin Corase87216f2017-08-17 16:59:22 -07001115 u8 *data;
Dave Barach68b0fb02017-02-28 15:15:56 -05001116
Florin Coras1ee78302019-02-05 15:51:15 -08001117 seg_size = max_deq_bytes + TRANSPORT_MAX_HDRS_LEN;
Florin Coras1f152cd2017-08-18 19:28:03 -07001118
Florin Corasb2215d62017-08-01 16:56:58 -07001119 /*
1120 * Prepare options
1121 */
Florin Corasc8343412017-05-04 14:25:50 -07001122 tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
1123
Florin Corasb2215d62017-08-01 16:56:58 -07001124 /*
1125 * Allocate and fill in buffer(s)
1126 */
Dave Barach68b0fb02017-02-28 15:15:56 -05001127
Florin Corasb2215d62017-08-01 16:56:58 -07001128 /* Easy case, buffer size greater than mss */
Florin Corasbe72ae62018-11-01 11:23:03 -07001129 if (PREDICT_TRUE (seg_size <= bytes_per_buffer))
Florin Corasb2215d62017-08-01 16:56:58 -07001130 {
Florin Corasbdf7fd62019-01-31 17:31:01 -08001131 if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
Florin Coras24793e32018-05-09 11:34:25 -07001132 return 0;
1133 *b = vlib_get_buffer (vm, bi);
1134 data = tcp_init_buffer (vm, *b);
Florin Coras31c99552019-03-01 13:00:58 -08001135 n_bytes = session_tx_fifo_peek_bytes (&tc->connection, data, offset,
1136 max_deq_bytes);
Florin Corasb2215d62017-08-01 16:56:58 -07001137 ASSERT (n_bytes == max_deq_bytes);
1138 b[0]->current_length = n_bytes;
Florin Coras47596832019-03-12 18:58:54 -07001139 tcp_push_hdr_i (tc, *b, tc->snd_una + offset, /* compute opts */ 0,
1140 /* burst */ 0, /* update_snd_nxt */ 0);
Florin Corasb2215d62017-08-01 16:56:58 -07001141 }
1142 /* Split mss into multiple buffers */
1143 else
1144 {
Florin Corasbdf7fd62019-01-31 17:31:01 -08001145 u32 chain_bi = ~0, n_bufs_per_seg, n_bufs;
1146 u16 n_peeked, len_to_deq;
Florin Corasb2215d62017-08-01 16:56:58 -07001147 vlib_buffer_t *chain_b, *prev_b;
Florin Corasb2215d62017-08-01 16:56:58 -07001148 int i;
1149
Florin Corasb2215d62017-08-01 16:56:58 -07001150 /* Make sure we have enough buffers */
Florin Corasbe72ae62018-11-01 11:23:03 -07001151 n_bufs_per_seg = ceil ((double) seg_size / bytes_per_buffer);
Florin Corasbdf7fd62019-01-31 17:31:01 -08001152 vec_validate_aligned (wrk->tx_buffers, n_bufs_per_seg - 1,
1153 CLIB_CACHE_LINE_BYTES);
1154 n_bufs = vlib_buffer_alloc (vm, wrk->tx_buffers, n_bufs_per_seg);
1155 if (PREDICT_FALSE (n_bufs != n_bufs_per_seg))
Florin Corasb2215d62017-08-01 16:56:58 -07001156 {
Florin Corasbdf7fd62019-01-31 17:31:01 -08001157 if (n_bufs)
1158 vlib_buffer_free (vm, wrk->tx_buffers, n_bufs);
1159 return 0;
Florin Corasb2215d62017-08-01 16:56:58 -07001160 }
1161
Florin Corasbdf7fd62019-01-31 17:31:01 -08001162 *b = vlib_get_buffer (vm, wrk->tx_buffers[--n_bufs]);
Florin Coras24793e32018-05-09 11:34:25 -07001163 data = tcp_init_buffer (vm, *b);
Florin Coras31c99552019-03-01 13:00:58 -08001164 n_bytes = session_tx_fifo_peek_bytes (&tc->connection, data, offset,
1165 bytes_per_buffer -
1166 TRANSPORT_MAX_HDRS_LEN);
Florin Corasb2215d62017-08-01 16:56:58 -07001167 b[0]->current_length = n_bytes;
1168 b[0]->flags |= VLIB_BUFFER_TOTAL_LENGTH_VALID;
1169 b[0]->total_length_not_including_first_buffer = 0;
Florin Corasb2215d62017-08-01 16:56:58 -07001170 max_deq_bytes -= n_bytes;
1171
1172 chain_b = *b;
1173 for (i = 1; i < n_bufs_per_seg; i++)
1174 {
1175 prev_b = chain_b;
Florin Corasbe72ae62018-11-01 11:23:03 -07001176 len_to_deq = clib_min (max_deq_bytes, bytes_per_buffer);
Florin Corasbdf7fd62019-01-31 17:31:01 -08001177 chain_bi = wrk->tx_buffers[--n_bufs];
Florin Corasb2215d62017-08-01 16:56:58 -07001178 chain_b = vlib_get_buffer (vm, chain_bi);
1179 chain_b->current_data = 0;
Florin Corase87216f2017-08-17 16:59:22 -07001180 data = vlib_buffer_get_current (chain_b);
Florin Coras31c99552019-03-01 13:00:58 -08001181 n_peeked = session_tx_fifo_peek_bytes (&tc->connection, data,
1182 offset + n_bytes,
1183 len_to_deq);
Florin Corasb2215d62017-08-01 16:56:58 -07001184 ASSERT (n_peeked == len_to_deq);
Florin Coras1f152cd2017-08-18 19:28:03 -07001185 n_bytes += n_peeked;
Florin Corasb2215d62017-08-01 16:56:58 -07001186 chain_b->current_length = n_peeked;
Florin Coras1f152cd2017-08-18 19:28:03 -07001187 chain_b->next_buffer = 0;
Florin Corasb2215d62017-08-01 16:56:58 -07001188
1189 /* update previous buffer */
1190 prev_b->next_buffer = chain_bi;
1191 prev_b->flags |= VLIB_BUFFER_NEXT_PRESENT;
1192
Florin Corasb2215d62017-08-01 16:56:58 -07001193 max_deq_bytes -= n_peeked;
Florin Coras1f152cd2017-08-18 19:28:03 -07001194 b[0]->total_length_not_including_first_buffer += n_peeked;
Florin Corasb2215d62017-08-01 16:56:58 -07001195 }
Florin Coras1f152cd2017-08-18 19:28:03 -07001196
Florin Coras47596832019-03-12 18:58:54 -07001197 tcp_push_hdr_i (tc, *b, tc->snd_una + offset, /* compute opts */ 0,
1198 /* burst */ 0, /* update_snd_nxt */ 0);
Florin Corasbdf7fd62019-01-31 17:31:01 -08001199
1200 if (PREDICT_FALSE (n_bufs))
1201 {
1202 clib_warning ("not all buffers consumed");
1203 vlib_buffer_free (vm, wrk->tx_buffers, n_bufs);
1204 }
Florin Corasb2215d62017-08-01 16:56:58 -07001205 }
1206
Florin Coras93992a92017-05-24 18:03:56 -07001207 ASSERT (n_bytes > 0);
Florin Corasbe72ae62018-11-01 11:23:03 -07001208 ASSERT (((*b)->current_data + (*b)->current_length) <= bytes_per_buffer);
Florin Corasbb292f42017-05-19 09:49:19 -07001209
Florin Coras36ee9f12018-11-02 12:52:10 -07001210 return n_bytes;
1211}
1212
1213/**
1214 * Build a retransmit segment
1215 *
1216 * @return the number of bytes in the segment or 0 if there's nothing to
1217 * retransmit
1218 */
1219static u32
1220tcp_prepare_retransmit_segment (tcp_worker_ctx_t * wrk,
1221 tcp_connection_t * tc, u32 offset,
1222 u32 max_deq_bytes, vlib_buffer_t ** b)
1223{
1224 u32 start, available_bytes;
1225 int n_bytes = 0;
1226
1227 ASSERT (tc->state >= TCP_STATE_ESTABLISHED);
1228 ASSERT (max_deq_bytes != 0);
1229
1230 /*
1231 * Make sure we can retransmit something
1232 */
Florin Coras31c99552019-03-01 13:00:58 -08001233 available_bytes = transport_max_tx_dequeue (&tc->connection);
Florin Coras36ee9f12018-11-02 12:52:10 -07001234 ASSERT (available_bytes >= offset);
1235 available_bytes -= offset;
1236 if (!available_bytes)
1237 return 0;
1238
1239 max_deq_bytes = clib_min (tc->snd_mss, max_deq_bytes);
1240 max_deq_bytes = clib_min (available_bytes, max_deq_bytes);
1241
Florin Coras36ee9f12018-11-02 12:52:10 -07001242 start = tc->snd_una + offset;
Florin Coras81cb8e42019-10-22 19:44:45 -07001243 ASSERT (seq_leq (start + max_deq_bytes, tc->snd_nxt));
1244
Florin Coras36ee9f12018-11-02 12:52:10 -07001245 n_bytes = tcp_prepare_segment (wrk, tc, offset, max_deq_bytes, b);
1246 if (!n_bytes)
1247 return 0;
1248
Florin Coras36ebcff2019-09-12 18:36:44 -07001249 tc->snd_rxt_bytes += n_bytes;
1250
Florin Corasbbcfaac2019-10-10 13:52:04 -07001251 if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
Florin Coras36ebcff2019-09-12 18:36:44 -07001252 tcp_bt_track_rxt (tc, start, start + n_bytes);
Dave Barach68b0fb02017-02-28 15:15:56 -05001253
Florin Corasedfe0ee2019-07-29 18:13:25 -07001254 tc->bytes_retrans += n_bytes;
1255 tc->segs_retrans += 1;
Florin Coras0765d972020-03-18 21:26:41 +00001256 tcp_worker_stats_inc (wrk, rxt_segs, 1);
Florin Corasa436a422019-08-20 07:09:31 -07001257 TCP_EVT (TCP_EVT_CC_RTX, tc, offset, n_bytes);
Florin Coras36ebcff2019-09-12 18:36:44 -07001258
Dave Barach68b0fb02017-02-28 15:15:56 -05001259 return n_bytes;
1260}
1261
Florin Coras36ebcff2019-09-12 18:36:44 -07001262static void
1263tcp_check_sack_reneging (tcp_connection_t * tc)
1264{
1265 sack_scoreboard_t *sb = &tc->sack_sb;
1266 sack_scoreboard_hole_t *hole;
1267
1268 hole = scoreboard_first_hole (sb);
1269 if (!sb->is_reneging && (!hole || hole->start == tc->snd_una))
1270 return;
1271
1272 scoreboard_clear_reneging (sb, tc->snd_una, tc->snd_nxt);
1273}
1274
Florin Coras6792ec02017-03-13 03:49:51 -07001275/**
1276 * Reset congestion control, switch cwnd to loss window and try again.
1277 */
1278static void
Florin Corasa3c32652019-07-03 17:47:22 -07001279tcp_cc_init_rxt_timeout (tcp_connection_t * tc)
Florin Coras6792ec02017-03-13 03:49:51 -07001280{
Florin Corasa436a422019-08-20 07:09:31 -07001281 TCP_EVT (TCP_EVT_CC_EVT, tc, 6);
Florin Coras36ebcff2019-09-12 18:36:44 -07001282
Florin Coras93992a92017-05-24 18:03:56 -07001283 tc->prev_ssthresh = tc->ssthresh;
1284 tc->prev_cwnd = tc->cwnd;
1285
Florin Coras36ebcff2019-09-12 18:36:44 -07001286 /* If we entrered loss without fast recovery, notify cc algo of the
1287 * congestion event such that it can update ssthresh and its state */
1288 if (!tcp_in_fastrecovery (tc))
1289 tcp_cc_congestion (tc);
Florin Corasa3c32652019-07-03 17:47:22 -07001290
Florin Coras36ebcff2019-09-12 18:36:44 -07001291 /* Let cc algo decide loss cwnd and ssthresh post unrecovered loss */
Florin Corasa3c32652019-07-03 17:47:22 -07001292 tcp_cc_loss (tc);
Florin Coras6792ec02017-03-13 03:49:51 -07001293
Florin Corasf1762d62017-09-24 19:43:08 -04001294 tc->rtt_ts = 0;
Florin Corasd2aab832018-05-22 11:39:59 -07001295 tc->cwnd_acc_bytes = 0;
Florin Corasedfe0ee2019-07-29 18:13:25 -07001296 tc->tr_occurences += 1;
Florin Coras3af90fc2017-05-03 21:09:42 -07001297 tcp_recovery_on (tc);
Florin Coras6792ec02017-03-13 03:49:51 -07001298}
1299
Florin Coras75c48c12019-08-02 15:17:21 -07001300void
Florin Corasaa388692020-02-14 23:41:25 +00001301tcp_timer_retransmit_handler (tcp_connection_t * tc)
Dave Barach68b0fb02017-02-28 15:15:56 -05001302{
Florin Corasaa388692020-02-14 23:41:25 +00001303 tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
Florin Corasbe72ae62018-11-01 11:23:03 -07001304 vlib_main_t *vm = wrk->vm;
Florin Corasb2215d62017-08-01 16:56:58 -07001305 vlib_buffer_t *b = 0;
Florin Coras3af90fc2017-05-03 21:09:42 -07001306 u32 bi, n_bytes;
Dave Barach68b0fb02017-02-28 15:15:56 -05001307
Florin Coras0765d972020-03-18 21:26:41 +00001308 tcp_worker_stats_inc (wrk, tr_events, 1);
Florin Coras75c48c12019-08-02 15:17:21 -07001309
Florin Corasaa388692020-02-14 23:41:25 +00001310 /* Should be handled by a different handler */
1311 if (PREDICT_FALSE (tc->state == TCP_STATE_SYN_SENT))
Florin Coras75c48c12019-08-02 15:17:21 -07001312 return;
1313
Florin Coras75c48c12019-08-02 15:17:21 -07001314 /* Wait-close and retransmit could pop at the same time */
1315 if (tc->state == TCP_STATE_CLOSED)
1316 return;
Dave Barach68b0fb02017-02-28 15:15:56 -05001317
Florin Corase04c2992017-03-01 08:17:34 -08001318 if (tc->state >= TCP_STATE_ESTABLISHED)
Dave Barach68b0fb02017-02-28 15:15:56 -05001319 {
Florin Corasa436a422019-08-20 07:09:31 -07001320 TCP_EVT (TCP_EVT_CC_EVT, tc, 2);
Florin Corase55a6d72018-10-31 23:09:22 -07001321
Florin Coras93992a92017-05-24 18:03:56 -07001322 /* Lost FIN, retransmit and return */
Florin Coras222e1f412019-02-16 20:47:32 -08001323 if (tc->flags & TCP_CONN_FINSNT)
Florin Coras93992a92017-05-24 18:03:56 -07001324 {
1325 tcp_send_fin (tc);
Florin Corasf988e692017-11-27 04:34:14 -05001326 tc->rto_boff += 1;
1327 tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
Florin Coras93992a92017-05-24 18:03:56 -07001328 return;
1329 }
1330
Florin Coras3ec66b02018-08-23 16:27:05 -07001331 /* Shouldn't be here. This condition is tricky because it has to take
1332 * into account boff > 0 due to persist timeout. */
Florin Coras47596832019-03-12 18:58:54 -07001333 if ((tc->rto_boff == 0 && tc->snd_una == tc->snd_nxt)
Florin Coras3ec66b02018-08-23 16:27:05 -07001334 || (tc->rto_boff > 0 && seq_geq (tc->snd_una, tc->snd_congestion)
1335 && !tcp_flight_size (tc)))
Florin Corasa096f2d2017-09-28 23:49:42 -04001336 {
Florin Coras3ec66b02018-08-23 16:27:05 -07001337 ASSERT (!tcp_in_recovery (tc));
1338 tc->rto_boff = 0;
Florin Corasa096f2d2017-09-28 23:49:42 -04001339 return;
1340 }
1341
Florin Coras3ec66b02018-08-23 16:27:05 -07001342 /* We're not in recovery so make sure rto_boff is 0. Can be non 0 due
1343 * to persist timer timeout */
Florin Coras4eeeaaf2017-09-05 14:03:37 -04001344 if (!tcp_in_recovery (tc) && tc->rto_boff > 0)
1345 {
1346 tc->rto_boff = 0;
1347 tcp_update_rto (tc);
1348 }
1349
Florin Coras548f7572019-06-07 15:31:06 -07001350 /* Peer is dead or network connectivity is lost. Close connection.
1351 * RFC 1122 section 4.2.3.5 recommends a value of at least 100s. For
1352 * a min rto of 0.2s we need to retry about 8 times. */
1353 if (tc->rto_boff >= TCP_RTO_BOFF_MAX)
1354 {
1355 tcp_send_reset (tc);
1356 tcp_connection_set_state (tc, TCP_STATE_CLOSED);
1357 session_transport_closing_notify (&tc->connection);
Florin Coras52be6742019-11-14 23:32:08 -08001358 session_transport_closed_notify (&tc->connection);
Florin Coras548f7572019-06-07 15:31:06 -07001359 tcp_connection_timers_reset (tc);
Florin Coras7a3a8662020-02-22 02:27:21 +00001360 tcp_program_cleanup (wrk, tc);
Florin Coras0765d972020-03-18 21:26:41 +00001361 tcp_worker_stats_inc (wrk, tr_abort, 1);
Florin Coras548f7572019-06-07 15:31:06 -07001362 return;
1363 }
1364
Florin Coras36ebcff2019-09-12 18:36:44 -07001365 if (tcp_opts_sack_permitted (&tc->rcv_opts))
1366 tcp_check_sack_reneging (tc);
Florin Coras4eeeaaf2017-09-05 14:03:37 -04001367
Florin Coras36ebcff2019-09-12 18:36:44 -07001368 /* Update send congestion to make sure that rxt has data to send */
1369 tc->snd_congestion = tc->snd_nxt;
Florin Corasa3c32652019-07-03 17:47:22 -07001370
Florin Coras36ebcff2019-09-12 18:36:44 -07001371 /* Send the first unacked segment. If we're short on buffers, return
1372 * as soon as possible */
Florin Coras81cb8e42019-10-22 19:44:45 -07001373 n_bytes = clib_min (tc->snd_mss, tc->snd_nxt - tc->snd_una);
1374 n_bytes = tcp_prepare_retransmit_segment (wrk, tc, 0, n_bytes, &b);
Florin Coras85a3ddd2018-12-24 16:54:34 -08001375 if (!n_bytes)
Florin Corase04c2992017-03-01 08:17:34 -08001376 {
Florin Coras0765d972020-03-18 21:26:41 +00001377 tcp_timer_update (&wrk->timer_wheel, tc, TCP_TIMER_RETRANSMIT, 1);
Florin Corase04c2992017-03-01 08:17:34 -08001378 return;
1379 }
Dave Barach68b0fb02017-02-28 15:15:56 -05001380
Dave Barachb7f1faa2017-08-29 11:43:37 -04001381 bi = vlib_get_buffer_index (vm, b);
Florin Corasbe72ae62018-11-01 11:23:03 -07001382 tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
Florin Coras75c48c12019-08-02 15:17:21 -07001383
1384 tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
Florin Coras0765d972020-03-18 21:26:41 +00001385 tcp_retransmit_timer_force_update (&wrk->timer_wheel, tc);
Florin Coras36ebcff2019-09-12 18:36:44 -07001386
1387 tc->rto_boff += 1;
1388 if (tc->rto_boff == 1)
1389 {
1390 tcp_cc_init_rxt_timeout (tc);
1391 /* Record timestamp. Eifel detection algorithm RFC3522 */
1392 tc->snd_rxt_ts = tcp_tstamp (tc);
1393 }
1394
1395 if (tcp_opts_sack_permitted (&tc->rcv_opts))
Florin Corasbe237bf2019-09-27 08:16:40 -07001396 scoreboard_init_rxt (&tc->sack_sb, tc->snd_una + n_bytes);
Florin Coras36ebcff2019-09-12 18:36:44 -07001397
1398 tcp_program_retransmit (tc);
Florin Coras93992a92017-05-24 18:03:56 -07001399 }
Florin Coras4eeeaaf2017-09-05 14:03:37 -04001400 /* Retransmit SYN-ACK */
1401 else if (tc->state == TCP_STATE_SYN_RCVD)
1402 {
Florin Corasa436a422019-08-20 07:09:31 -07001403 TCP_EVT (TCP_EVT_CC_EVT, tc, 2);
Florin Corase55a6d72018-10-31 23:09:22 -07001404
Florin Coras4eeeaaf2017-09-05 14:03:37 -04001405 tc->rtt_ts = 0;
1406
Florin Coras75c48c12019-08-02 15:17:21 -07001407 /* Passive open establish timeout */
1408 if (tc->rto > TCP_ESTABLISH_TIME >> 1)
1409 {
1410 tcp_connection_set_state (tc, TCP_STATE_CLOSED);
1411 tcp_connection_timers_reset (tc);
Florin Coras7a3a8662020-02-22 02:27:21 +00001412 tcp_program_cleanup (wrk, tc);
Florin Coras0765d972020-03-18 21:26:41 +00001413 tcp_worker_stats_inc (wrk, tr_abort, 1);
Florin Coras75c48c12019-08-02 15:17:21 -07001414 return;
1415 }
Florin Coras222e1f412019-02-16 20:47:32 -08001416
Florin Corasbdf7fd62019-01-31 17:31:01 -08001417 if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
Florin Corasf988e692017-11-27 04:34:14 -05001418 {
Florin Coras0765d972020-03-18 21:26:41 +00001419 tcp_timer_update (&wrk->timer_wheel, tc, TCP_TIMER_RETRANSMIT, 1);
Florin Corasf988e692017-11-27 04:34:14 -05001420 return;
1421 }
Florin Coras4eeeaaf2017-09-05 14:03:37 -04001422
Florin Coras75c48c12019-08-02 15:17:21 -07001423 tc->rto_boff += 1;
1424 if (tc->rto_boff > TCP_RTO_SYN_RETRIES)
1425 tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
1426
Florin Coras0765d972020-03-18 21:26:41 +00001427 tcp_retransmit_timer_force_update (&wrk->timer_wheel, tc);
Florin Coras75c48c12019-08-02 15:17:21 -07001428
Florin Coras4eeeaaf2017-09-05 14:03:37 -04001429 b = vlib_get_buffer (vm, bi);
Florin Coras24793e32018-05-09 11:34:25 -07001430 tcp_init_buffer (vm, b);
Florin Coras4eeeaaf2017-09-05 14:03:37 -04001431 tcp_make_synack (tc, b);
Florin Corasa436a422019-08-20 07:09:31 -07001432 TCP_EVT (TCP_EVT_SYN_RXT, tc, 1);
Florin Coras4eeeaaf2017-09-05 14:03:37 -04001433
1434 /* Retransmit timer already updated, just enqueue to output */
Florin Corasbe72ae62018-11-01 11:23:03 -07001435 tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
Dave Barach68b0fb02017-02-28 15:15:56 -05001436 }
Florin Coras93992a92017-05-24 18:03:56 -07001437 else
1438 {
1439 ASSERT (tc->state == TCP_STATE_CLOSED);
Florin Coras93992a92017-05-24 18:03:56 -07001440 return;
1441 }
Dave Barach68b0fb02017-02-28 15:15:56 -05001442}
1443
Florin Coras75c48c12019-08-02 15:17:21 -07001444/**
1445 * SYN retransmit timer handler. Active open only.
1446 */
Dave Barach68b0fb02017-02-28 15:15:56 -05001447void
Florin Corasaa388692020-02-14 23:41:25 +00001448tcp_timer_retransmit_syn_handler (tcp_connection_t * tc)
Dave Barach68b0fb02017-02-28 15:15:56 -05001449{
Florin Corasaa388692020-02-14 23:41:25 +00001450 tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
Florin Coras75c48c12019-08-02 15:17:21 -07001451 vlib_main_t *vm = wrk->vm;
Florin Coras75c48c12019-08-02 15:17:21 -07001452 vlib_buffer_t *b = 0;
1453 u32 bi;
Dave Barach68b0fb02017-02-28 15:15:56 -05001454
Florin Coras75c48c12019-08-02 15:17:21 -07001455 /* Note: the connection may have transitioned to ESTABLISHED... */
Florin Corasaa388692020-02-14 23:41:25 +00001456 if (PREDICT_FALSE (tc->state != TCP_STATE_SYN_SENT))
Florin Coras75c48c12019-08-02 15:17:21 -07001457 return;
1458
Florin Coras75c48c12019-08-02 15:17:21 -07001459 /* Half-open connection actually moved to established but we were
1460 * waiting for syn retransmit to pop to call cleanup from the right
1461 * thread. */
1462 if (tc->flags & TCP_CONN_HALF_OPEN_DONE)
1463 {
1464 if (tcp_half_open_connection_cleanup (tc))
1465 TCP_DBG ("could not remove half-open connection");
1466 return;
1467 }
1468
Florin Corasa436a422019-08-20 07:09:31 -07001469 TCP_EVT (TCP_EVT_CC_EVT, tc, 2);
Florin Coras75c48c12019-08-02 15:17:21 -07001470 tc->rtt_ts = 0;
1471
1472 /* Active open establish timeout */
1473 if (tc->rto >= TCP_ESTABLISH_TIME >> 1)
1474 {
1475 session_stream_connect_notify (&tc->connection, 1 /* fail */ );
1476 tcp_connection_cleanup (tc);
1477 return;
1478 }
1479
1480 if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
1481 {
Florin Coras0765d972020-03-18 21:26:41 +00001482 tcp_timer_update (&wrk->timer_wheel, tc, TCP_TIMER_RETRANSMIT_SYN, 1);
Florin Coras75c48c12019-08-02 15:17:21 -07001483 return;
1484 }
1485
1486 /* Try without increasing RTO a number of times. If this fails,
1487 * start growing RTO exponentially */
1488 tc->rto_boff += 1;
1489 if (tc->rto_boff > TCP_RTO_SYN_RETRIES)
1490 tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
1491
1492 b = vlib_get_buffer (vm, bi);
1493 tcp_init_buffer (vm, b);
1494 tcp_make_syn (tc, b);
1495
Florin Corasa436a422019-08-20 07:09:31 -07001496 TCP_EVT (TCP_EVT_SYN_RXT, tc, 0);
Florin Coras75c48c12019-08-02 15:17:21 -07001497
1498 /* This goes straight to ipx_lookup */
1499 tcp_push_ip_hdr (wrk, tc, b);
1500 tcp_enqueue_to_ip_lookup (wrk, b, bi, tc->c_is_ip4, tc->c_fib_index);
1501
Florin Coras0765d972020-03-18 21:26:41 +00001502 tcp_timer_update (&wrk->timer_wheel, tc, TCP_TIMER_RETRANSMIT_SYN,
Florin Coras75c48c12019-08-02 15:17:21 -07001503 tc->rto * TCP_TO_TIMER_TICK);
Dave Barach68b0fb02017-02-28 15:15:56 -05001504}
1505
1506/**
Florin Coras3e350af2017-03-30 02:54:28 -07001507 * Got 0 snd_wnd from peer, try to do something about it.
1508 *
1509 */
1510void
Florin Corasaa388692020-02-14 23:41:25 +00001511tcp_timer_persist_handler (tcp_connection_t * tc)
Florin Coras3e350af2017-03-30 02:54:28 -07001512{
Florin Corasaa388692020-02-14 23:41:25 +00001513 tcp_worker_ctx_t *wrk = tcp_get_worker (tc->c_thread_index);
Florin Corasbe72ae62018-11-01 11:23:03 -07001514 u32 bi, max_snd_bytes, available_bytes, offset;
1515 tcp_main_t *tm = vnet_get_tcp_main ();
1516 vlib_main_t *vm = wrk->vm;
Florin Coras3e350af2017-03-30 02:54:28 -07001517 vlib_buffer_t *b;
Florin Coras93992a92017-05-24 18:03:56 -07001518 int n_bytes = 0;
Florin Coras1f152cd2017-08-18 19:28:03 -07001519 u8 *data;
Florin Coras3e350af2017-03-30 02:54:28 -07001520
Florin Coras3e350af2017-03-30 02:54:28 -07001521 /* Problem already solved or worse */
Florin Coras7e74bf32019-03-06 16:51:58 -08001522 if (tc->state == TCP_STATE_CLOSED || tc->snd_wnd > tc->snd_mss
1523 || (tc->flags & TCP_CONN_FINSNT))
Florin Coras70f879d2020-03-13 17:54:42 +00001524 goto update_scheduler;
Florin Coras3e350af2017-03-30 02:54:28 -07001525
Florin Coras31c99552019-03-01 13:00:58 -08001526 available_bytes = transport_max_tx_dequeue (&tc->connection);
Florin Coras47596832019-03-12 18:58:54 -07001527 offset = tc->snd_nxt - tc->snd_una;
Florin Coras50958952017-08-29 14:50:13 -07001528
1529 /* Reprogram persist if no new bytes available to send. We may have data
1530 * next time */
1531 if (!available_bytes)
1532 {
Florin Coras0765d972020-03-18 21:26:41 +00001533 tcp_persist_timer_set (&wrk->timer_wheel, tc);
Florin Coras50958952017-08-29 14:50:13 -07001534 return;
1535 }
1536
1537 if (available_bytes <= offset)
Florin Coras70f879d2020-03-13 17:54:42 +00001538 goto update_scheduler;
Florin Coras50958952017-08-29 14:50:13 -07001539
Florin Coras3e350af2017-03-30 02:54:28 -07001540 /* Increment RTO backoff */
1541 tc->rto_boff += 1;
1542 tc->rto = clib_min (tc->rto << 1, TCP_RTO_MAX);
1543
Florin Corasb2215d62017-08-01 16:56:58 -07001544 /*
1545 * Try to force the first unsent segment (or buffer)
1546 */
Florin Corasbdf7fd62019-01-31 17:31:01 -08001547 if (PREDICT_FALSE (!vlib_buffer_alloc (vm, &bi, 1)))
Florin Coras3ec66b02018-08-23 16:27:05 -07001548 {
Florin Coras0765d972020-03-18 21:26:41 +00001549 tcp_persist_timer_set (&wrk->timer_wheel, tc);
Florin Coras3ec66b02018-08-23 16:27:05 -07001550 return;
1551 }
Florin Coras70f879d2020-03-13 17:54:42 +00001552
Florin Coras3e350af2017-03-30 02:54:28 -07001553 b = vlib_get_buffer (vm, bi);
Florin Coras1f152cd2017-08-18 19:28:03 -07001554 data = tcp_init_buffer (vm, b);
Florin Coras93992a92017-05-24 18:03:56 -07001555
Florin Coras1f152cd2017-08-18 19:28:03 -07001556 tcp_validate_txf_size (tc, offset);
Florin Corasc8343412017-05-04 14:25:50 -07001557 tc->snd_opts_len = tcp_make_options (tc, &tc->snd_opts, tc->state);
Florin Coras7e74bf32019-03-06 16:51:58 -08001558 max_snd_bytes = clib_min (tc->snd_mss,
1559 tm->bytes_per_buffer - TRANSPORT_MAX_HDRS_LEN);
1560 n_bytes = session_tx_fifo_peek_bytes (&tc->connection, data, offset,
1561 max_snd_bytes);
Florin Coras3e350af2017-03-30 02:54:28 -07001562 b->current_length = n_bytes;
Florin Coras84275e92017-09-26 12:30:40 -04001563 ASSERT (n_bytes != 0 && (tcp_timer_is_active (tc, TCP_TIMER_RETRANSMIT)
1564 || tc->snd_nxt == tc->snd_una_max
1565 || tc->rto_boff > 1));
Florin Coras93992a92017-05-24 18:03:56 -07001566
Florin Corasbbcfaac2019-10-10 13:52:04 -07001567 if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
Florin Corasdd60b1b2019-10-07 17:19:09 -07001568 {
1569 tcp_bt_check_app_limited (tc);
Florin Corasd6ae4bf2019-10-12 18:10:20 -07001570 tcp_bt_track_tx (tc, n_bytes);
Florin Corasdd60b1b2019-10-07 17:19:09 -07001571 }
1572
Florin Coras47596832019-03-12 18:58:54 -07001573 tcp_push_hdr_i (tc, b, tc->snd_nxt, /* compute opts */ 0,
1574 /* burst */ 0, /* update_snd_nxt */ 1);
1575 tc->snd_una_max = seq_max (tc->snd_nxt, tc->snd_una_max);
Florin Coras8b20bf52018-06-14 14:55:50 -07001576 tcp_validate_txf_size (tc, tc->snd_una_max - tc->snd_una);
Florin Corasbe72ae62018-11-01 11:23:03 -07001577 tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
Florin Coras3e350af2017-03-30 02:54:28 -07001578
Florin Coras1f152cd2017-08-18 19:28:03 -07001579 /* Just sent new data, enable retransmit */
Florin Coras0765d972020-03-18 21:26:41 +00001580 tcp_retransmit_timer_update (&wrk->timer_wheel, tc);
Florin Coras70f879d2020-03-13 17:54:42 +00001581
Florin Coras6080e0d2020-03-13 20:39:43 +00001582 return;
1583
Florin Coras70f879d2020-03-13 17:54:42 +00001584update_scheduler:
1585
Florin Coras6080e0d2020-03-13 20:39:43 +00001586 if (tcp_is_descheduled (tc))
Florin Coras70f879d2020-03-13 17:54:42 +00001587 transport_connection_reschedule (&tc->connection);
Florin Coras3e350af2017-03-30 02:54:28 -07001588}
1589
1590/**
Florin Coras6792ec02017-03-13 03:49:51 -07001591 * Retransmit first unacked segment
1592 */
Florin Corasbf4d5ce2018-10-19 16:26:24 -07001593int
Florin Corasbe72ae62018-11-01 11:23:03 -07001594tcp_retransmit_first_unacked (tcp_worker_ctx_t * wrk, tcp_connection_t * tc)
Dave Barach68b0fb02017-02-28 15:15:56 -05001595{
Florin Corasbe72ae62018-11-01 11:23:03 -07001596 vlib_main_t *vm = wrk->vm;
1597 vlib_buffer_t *b;
Florin Coras47596832019-03-12 18:58:54 -07001598 u32 bi, n_bytes;
Dave Barach68b0fb02017-02-28 15:15:56 -05001599
Florin Corasa436a422019-08-20 07:09:31 -07001600 TCP_EVT (TCP_EVT_CC_EVT, tc, 1);
Florin Corasbf4d5ce2018-10-19 16:26:24 -07001601
Florin Corasbe72ae62018-11-01 11:23:03 -07001602 n_bytes = tcp_prepare_retransmit_segment (wrk, tc, 0, tc->snd_mss, &b);
Florin Corasb2215d62017-08-01 16:56:58 -07001603 if (!n_bytes)
Florin Corasbf4d5ce2018-10-19 16:26:24 -07001604 return -1;
1605
Florin Corasb2215d62017-08-01 16:56:58 -07001606 bi = vlib_get_buffer_index (vm, b);
Florin Corasbe72ae62018-11-01 11:23:03 -07001607 tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
Florin Corasbf4d5ce2018-10-19 16:26:24 -07001608
1609 return 0;
Florin Coras6792ec02017-03-13 03:49:51 -07001610}
1611
Florin Coras36ee9f12018-11-02 12:52:10 -07001612static int
Florin Coras36ebcff2019-09-12 18:36:44 -07001613tcp_transmit_unsent (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
1614 u32 burst_size)
Florin Coras36ee9f12018-11-02 12:52:10 -07001615{
Florin Coras635acbf2019-07-31 14:11:05 -07001616 u32 offset, n_segs = 0, n_written, bi, available_wnd;
Florin Coras36ee9f12018-11-02 12:52:10 -07001617 vlib_main_t *vm = wrk->vm;
1618 vlib_buffer_t *b = 0;
1619
Florin Coras47596832019-03-12 18:58:54 -07001620 offset = tc->snd_nxt - tc->snd_una;
Florin Coras635acbf2019-07-31 14:11:05 -07001621 available_wnd = tc->snd_wnd - offset;
1622 burst_size = clib_min (burst_size, available_wnd / tc->snd_mss);
1623
Florin Corasd6ae4bf2019-10-12 18:10:20 -07001624 if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
1625 tcp_bt_check_app_limited (tc);
1626
Florin Coras36ee9f12018-11-02 12:52:10 -07001627 while (n_segs < burst_size)
1628 {
1629 n_written = tcp_prepare_segment (wrk, tc, offset, tc->snd_mss, &b);
1630 if (!n_written)
1631 goto done;
1632
1633 bi = vlib_get_buffer_index (vm, b);
1634 tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1635 offset += n_written;
1636 n_segs += 1;
Florin Coras47596832019-03-12 18:58:54 -07001637
Florin Corasd6ae4bf2019-10-12 18:10:20 -07001638 if (tc->cfg_flags & TCP_CFG_F_RATE_SAMPLE)
1639 tcp_bt_track_tx (tc, n_written);
1640
Florin Coras47596832019-03-12 18:58:54 -07001641 tc->snd_nxt += n_written;
1642 tc->snd_una_max = seq_max (tc->snd_nxt, tc->snd_una_max);
Florin Coras36ee9f12018-11-02 12:52:10 -07001643 }
1644
1645done:
1646 return n_segs;
1647}
1648
Florin Coras36ebcff2019-09-12 18:36:44 -07001649/**
1650 * Estimate send space using proportional rate reduction (RFC6937)
1651 */
Florin Corasbe237bf2019-09-27 08:16:40 -07001652int
Florin Coras36ebcff2019-09-12 18:36:44 -07001653tcp_fastrecovery_prr_snd_space (tcp_connection_t * tc)
1654{
1655 u32 pipe, prr_out;
1656 int space;
1657
1658 pipe = tcp_flight_size (tc);
1659 prr_out = tc->snd_rxt_bytes + (tc->snd_nxt - tc->snd_congestion);
1660
1661 if (pipe > tc->ssthresh)
1662 {
1663 space = ((int) tc->prr_delivered * ((f64) tc->ssthresh / tc->prev_cwnd))
1664 - prr_out;
1665 }
1666 else
1667 {
Florin Corasbe237bf2019-09-27 08:16:40 -07001668 int limit;
1669 limit = clib_max ((int) (tc->prr_delivered - prr_out), 0) + tc->snd_mss;
Florin Coras36ebcff2019-09-12 18:36:44 -07001670 space = clib_min (tc->ssthresh - pipe, limit);
1671 }
1672 space = clib_max (space, prr_out ? 0 : tc->snd_mss);
1673 return space;
1674}
1675
Florin Corasbe237bf2019-09-27 08:16:40 -07001676static inline u8
1677tcp_retransmit_should_retry_head (tcp_connection_t * tc,
1678 sack_scoreboard_t * sb)
1679{
1680 u32 tx_adv_sack = sb->high_sacked - tc->snd_congestion;
1681 f64 rr = (f64) tc->ssthresh / tc->prev_cwnd;
1682
Florin Corasb3dce892019-10-30 09:22:14 -07001683 if (tcp_fastrecovery_first (tc))
1684 return 1;
1685
Florin Corasbe237bf2019-09-27 08:16:40 -07001686 return (tx_adv_sack > (tc->snd_una - tc->prr_start) * rr);
1687}
1688
Florin Corascb711a42019-10-16 19:28:17 -07001689static inline u8
1690tcp_max_tx_deq (tcp_connection_t * tc)
1691{
1692 return (transport_max_tx_dequeue (&tc->connection)
1693 - (tc->snd_nxt - tc->snd_una));
1694}
1695
Florin Coras36ee9f12018-11-02 12:52:10 -07001696#define scoreboard_rescue_rxt_valid(_sb, _tc) \
1697 (seq_geq (_sb->rescue_rxt, _tc->snd_una) \
1698 && seq_leq (_sb->rescue_rxt, _tc->snd_congestion))
1699
Florin Coras6792ec02017-03-13 03:49:51 -07001700/**
Florin Coras36ebcff2019-09-12 18:36:44 -07001701 * Do retransmit with SACKs
Florin Coras6792ec02017-03-13 03:49:51 -07001702 */
Florin Coras36ebcff2019-09-12 18:36:44 -07001703static int
1704tcp_retransmit_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
1705 u32 burst_size)
Dave Barach68b0fb02017-02-28 15:15:56 -05001706{
Florin Coras36ebcff2019-09-12 18:36:44 -07001707 u32 n_written = 0, offset, max_bytes, n_segs = 0;
Florin Coras11e9e352019-11-13 19:09:47 -08001708 u8 snd_limited = 0, can_rescue = 0;
1709 u32 bi, max_deq, burst_bytes;
Florin Coras93992a92017-05-24 18:03:56 -07001710 sack_scoreboard_hole_t *hole;
Florin Coras36ee9f12018-11-02 12:52:10 -07001711 vlib_main_t *vm = wrk->vm;
1712 vlib_buffer_t *b = 0;
Florin Coras93992a92017-05-24 18:03:56 -07001713 sack_scoreboard_t *sb;
Florin Coras93992a92017-05-24 18:03:56 -07001714 int snd_space;
Dave Barach68b0fb02017-02-28 15:15:56 -05001715
Florin Coras36ebcff2019-09-12 18:36:44 -07001716 ASSERT (tcp_in_cong_recovery (tc));
Dave Barach68b0fb02017-02-28 15:15:56 -05001717
Florin Corasa8e71c82019-10-22 19:01:39 -07001718 burst_bytes = transport_connection_tx_pacer_burst (&tc->connection);
Florin Corasc31dc312019-10-06 14:06:14 -07001719 burst_size = clib_min (burst_size, burst_bytes / tc->snd_mss);
1720 if (!burst_size)
1721 {
1722 tcp_program_retransmit (tc);
1723 return 0;
1724 }
1725
Florin Coras36ebcff2019-09-12 18:36:44 -07001726 if (tcp_in_recovery (tc))
1727 snd_space = tcp_available_cc_snd_space (tc);
1728 else
1729 snd_space = tcp_fastrecovery_prr_snd_space (tc);
1730
Florin Corasca1c8f32018-05-23 21:01:30 -07001731 if (snd_space < tc->snd_mss)
Florin Coras11e9e352019-11-13 19:09:47 -08001732 goto done;
Florin Corasc31dc312019-10-06 14:06:14 -07001733
Florin Coras36ee9f12018-11-02 12:52:10 -07001734 sb = &tc->sack_sb;
Florin Corasbe237bf2019-09-27 08:16:40 -07001735
1736 /* Check if snd_una is a lost retransmit */
Florin Coras8a8b05c2019-10-16 10:07:39 -07001737 if (pool_elts (sb->holes)
1738 && seq_gt (sb->high_sacked, tc->snd_congestion)
Florin Corasbe237bf2019-09-27 08:16:40 -07001739 && tc->rxt_head != tc->snd_una
1740 && tcp_retransmit_should_retry_head (tc, sb))
1741 {
Florin Corasbf1f8b72019-11-04 14:39:33 -08001742 max_bytes = clib_min (tc->snd_mss, tc->snd_congestion - tc->snd_una);
1743 n_written = tcp_prepare_retransmit_segment (wrk, tc, 0, max_bytes, &b);
Florin Corasbe237bf2019-09-27 08:16:40 -07001744 if (!n_written)
1745 {
1746 tcp_program_retransmit (tc);
1747 goto done;
1748 }
1749 bi = vlib_get_buffer_index (vm, b);
1750 tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
1751 n_segs = 1;
1752
1753 tc->rxt_head = tc->snd_una;
1754 tc->rxt_delivered += n_written;
1755 tc->prr_delivered += n_written;
1756 ASSERT (tc->rxt_delivered <= tc->snd_rxt_bytes);
1757 }
1758
Florin Corasb3dce892019-10-30 09:22:14 -07001759 tcp_fastrecovery_first_off (tc);
1760
Florin Corasbe237bf2019-09-27 08:16:40 -07001761 TCP_EVT (TCP_EVT_CC_EVT, tc, 0);
Florin Coras36ee9f12018-11-02 12:52:10 -07001762 hole = scoreboard_get_hole (sb, sb->cur_rxt_hole);
1763
Florin Coras31c99552019-03-01 13:00:58 -08001764 max_deq = transport_max_tx_dequeue (&tc->connection);
Florin Coras47596832019-03-12 18:58:54 -07001765 max_deq -= tc->snd_nxt - tc->snd_una;
Florin Coras36ee9f12018-11-02 12:52:10 -07001766
Florin Corasbf4d5ce2018-10-19 16:26:24 -07001767 while (snd_space > 0 && n_segs < burst_size)
Florin Coras93992a92017-05-24 18:03:56 -07001768 {
Florin Coras81cb8e42019-10-22 19:44:45 -07001769 hole = scoreboard_next_rxt_hole (sb, hole, max_deq != 0, &can_rescue,
Florin Coras36ee9f12018-11-02 12:52:10 -07001770 &snd_limited);
Florin Coras93992a92017-05-24 18:03:56 -07001771 if (!hole)
1772 {
Florin Coras36ebcff2019-09-12 18:36:44 -07001773 /* We are out of lost holes to retransmit so send some new data. */
Florin Corasbe237bf2019-09-27 08:16:40 -07001774 if (max_deq > tc->snd_mss)
Florin Corasbf4d5ce2018-10-19 16:26:24 -07001775 {
Florin Corasc31dc312019-10-06 14:06:14 -07001776 u32 n_segs_new;
1777 int av_wnd;
1778
Florin Coras3eba7f12019-11-15 17:56:48 -08001779 /* Make sure we don't exceed available window and leave space
1780 * for one more packet, to avoid zero window acks */
Florin Corasc31dc312019-10-06 14:06:14 -07001781 av_wnd = (int) tc->snd_wnd - (tc->snd_nxt - tc->snd_una);
Florin Coras3eba7f12019-11-15 17:56:48 -08001782 av_wnd = clib_max (av_wnd - tc->snd_mss, 0);
Florin Corasc31dc312019-10-06 14:06:14 -07001783 snd_space = clib_min (snd_space, av_wnd);
Florin Coras36ee9f12018-11-02 12:52:10 -07001784 snd_space = clib_min (max_deq, snd_space);
1785 burst_size = clib_min (burst_size - n_segs,
1786 snd_space / tc->snd_mss);
Florin Coras36ebcff2019-09-12 18:36:44 -07001787 burst_size = clib_min (burst_size, TCP_RXT_MAX_BURST);
1788 n_segs_new = tcp_transmit_unsent (wrk, tc, burst_size);
1789 if (max_deq > n_segs_new * tc->snd_mss)
1790 tcp_program_retransmit (tc);
1791
1792 n_segs += n_segs_new;
Florin Coras36ee9f12018-11-02 12:52:10 -07001793 goto done;
Florin Corasbf4d5ce2018-10-19 16:26:24 -07001794 }
Florin Coras93992a92017-05-24 18:03:56 -07001795
Florin Coras36ebcff2019-09-12 18:36:44 -07001796 if (tcp_in_recovery (tc) || !can_rescue
1797 || scoreboard_rescue_rxt_valid (sb, tc))
Florin Coras36ee9f12018-11-02 12:52:10 -07001798 break;
1799
Florin Coras93992a92017-05-24 18:03:56 -07001800 /* If rescue rxt undefined or less than snd_una then one segment of
1801 * up to SMSS octets that MUST include the highest outstanding
1802 * unSACKed sequence number SHOULD be returned, and RescueRxt set to
1803 * RecoveryPoint. HighRxt MUST NOT be updated.
1804 */
Florin Coras3eba7f12019-11-15 17:56:48 -08001805 hole = scoreboard_last_hole (sb);
1806 max_bytes = clib_min (tc->snd_mss, hole->end - hole->start);
Florin Coras1f152cd2017-08-18 19:28:03 -07001807 max_bytes = clib_min (max_bytes, snd_space);
Florin Coras3eba7f12019-11-15 17:56:48 -08001808 offset = hole->end - tc->snd_una - max_bytes;
Florin Corasbe72ae62018-11-01 11:23:03 -07001809 n_written = tcp_prepare_retransmit_segment (wrk, tc, offset,
1810 max_bytes, &b);
Florin Coras24793e32018-05-09 11:34:25 -07001811 if (!n_written)
1812 goto done;
1813
Florin Coras3eba7f12019-11-15 17:56:48 -08001814 sb->rescue_rxt = tc->snd_congestion;
Florin Corasb2215d62017-08-01 16:56:58 -07001815 bi = vlib_get_buffer_index (vm, b);
Florin Corasbe72ae62018-11-01 11:23:03 -07001816 tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
Florin Corasbf4d5ce2018-10-19 16:26:24 -07001817 n_segs += 1;
Florin Coras93992a92017-05-24 18:03:56 -07001818 break;
1819 }
1820
Florin Coras1f152cd2017-08-18 19:28:03 -07001821 max_bytes = clib_min (hole->end - sb->high_rxt, snd_space);
1822 max_bytes = snd_limited ? clib_min (max_bytes, tc->snd_mss) : max_bytes;
1823 if (max_bytes == 0)
1824 break;
Florin Coras36ee9f12018-11-02 12:52:10 -07001825
Florin Coras93992a92017-05-24 18:03:56 -07001826 offset = sb->high_rxt - tc->snd_una;
Florin Corasbe72ae62018-11-01 11:23:03 -07001827 n_written = tcp_prepare_retransmit_segment (wrk, tc, offset, max_bytes,
1828 &b);
Florin Coras36ee9f12018-11-02 12:52:10 -07001829 ASSERT (n_written <= snd_space);
Florin Coras93992a92017-05-24 18:03:56 -07001830
1831 /* Nothing left to retransmit */
1832 if (n_written == 0)
Florin Corasb2215d62017-08-01 16:56:58 -07001833 break;
Florin Coras93992a92017-05-24 18:03:56 -07001834
Florin Corasb2215d62017-08-01 16:56:58 -07001835 bi = vlib_get_buffer_index (vm, b);
Florin Corasbe72ae62018-11-01 11:23:03 -07001836 tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
Florin Coras36ee9f12018-11-02 12:52:10 -07001837
1838 sb->high_rxt += n_written;
Florin Coras81cb8e42019-10-22 19:44:45 -07001839 ASSERT (seq_leq (sb->high_rxt, tc->snd_nxt));
1840
Florin Coras93992a92017-05-24 18:03:56 -07001841 snd_space -= n_written;
Florin Corasbf4d5ce2018-10-19 16:26:24 -07001842 n_segs += 1;
Florin Coras93992a92017-05-24 18:03:56 -07001843 }
1844
Florin Corasbf4d5ce2018-10-19 16:26:24 -07001845 if (hole)
Florin Coras36ebcff2019-09-12 18:36:44 -07001846 tcp_program_retransmit (tc);
Florin Corasbf4d5ce2018-10-19 16:26:24 -07001847
Florin Coras24793e32018-05-09 11:34:25 -07001848done:
Florin Coras36ebcff2019-09-12 18:36:44 -07001849
Florin Coras11e9e352019-11-13 19:09:47 -08001850 transport_connection_tx_pacer_reset_bucket (&tc->connection, 0);
Florin Corasbf4d5ce2018-10-19 16:26:24 -07001851 return n_segs;
Florin Coras93992a92017-05-24 18:03:56 -07001852}
1853
1854/**
1855 * Fast retransmit without SACK info
1856 */
Florin Coras36ebcff2019-09-12 18:36:44 -07001857static int
1858tcp_retransmit_no_sack (tcp_worker_ctx_t * wrk, tcp_connection_t * tc,
1859 u32 burst_size)
Florin Coras93992a92017-05-24 18:03:56 -07001860{
Florin Corasbf1f8b72019-11-04 14:39:33 -08001861 u32 n_written = 0, offset = 0, bi, max_deq, n_segs_now, max_bytes;
Florin Corasc31dc312019-10-06 14:06:14 -07001862 u32 burst_bytes, sent_bytes;
Florin Corasbe72ae62018-11-01 11:23:03 -07001863 vlib_main_t *vm = wrk->vm;
Florin Corasbf4d5ce2018-10-19 16:26:24 -07001864 int snd_space, n_segs = 0;
Florin Corasc31dc312019-10-06 14:06:14 -07001865 u8 cc_limited = 0;
Florin Coras93992a92017-05-24 18:03:56 -07001866 vlib_buffer_t *b;
1867
Florin Corasbf1f8b72019-11-04 14:39:33 -08001868 ASSERT (tcp_in_cong_recovery (tc));
Florin Corasa436a422019-08-20 07:09:31 -07001869 TCP_EVT (TCP_EVT_CC_EVT, tc, 0);
Florin Coras6792ec02017-03-13 03:49:51 -07001870
Florin Corasa8e71c82019-10-22 19:01:39 -07001871 burst_bytes = transport_connection_tx_pacer_burst (&tc->connection);
Florin Corasc31dc312019-10-06 14:06:14 -07001872 burst_size = clib_min (burst_size, burst_bytes / tc->snd_mss);
1873 if (!burst_size)
1874 {
1875 tcp_program_retransmit (tc);
1876 return 0;
1877 }
1878
Florin Coras87a9bf82019-06-12 11:26:35 -07001879 snd_space = tcp_available_cc_snd_space (tc);
Florin Corasc31dc312019-10-06 14:06:14 -07001880 cc_limited = snd_space < burst_bytes;
Florin Coras87a9bf82019-06-12 11:26:35 -07001881
Florin Coras36ee9f12018-11-02 12:52:10 -07001882 if (!tcp_fastrecovery_first (tc))
1883 goto send_unsent;
1884
1885 /* RFC 6582: [If a partial ack], retransmit the first unacknowledged
1886 * segment. */
Florin Corasbf4d5ce2018-10-19 16:26:24 -07001887 while (snd_space > 0 && n_segs < burst_size)
Dave Barach68b0fb02017-02-28 15:15:56 -05001888 {
Florin Corasbf1f8b72019-11-04 14:39:33 -08001889 max_bytes = clib_min (tc->snd_mss,
1890 tc->snd_congestion - tc->snd_una - offset);
1891 if (!max_bytes)
1892 break;
1893 n_written = tcp_prepare_retransmit_segment (wrk, tc, offset, max_bytes,
1894 &b);
Dave Barach68b0fb02017-02-28 15:15:56 -05001895
1896 /* Nothing left to retransmit */
Florin Coras6792ec02017-03-13 03:49:51 -07001897 if (n_written == 0)
Florin Corasb2215d62017-08-01 16:56:58 -07001898 break;
Dave Barach68b0fb02017-02-28 15:15:56 -05001899
Florin Corasb2215d62017-08-01 16:56:58 -07001900 bi = vlib_get_buffer_index (vm, b);
Florin Corasbe72ae62018-11-01 11:23:03 -07001901 tcp_enqueue_to_output (wrk, b, bi, tc->c_is_ip4);
Florin Coras6792ec02017-03-13 03:49:51 -07001902 snd_space -= n_written;
Florin Coras36ee9f12018-11-02 12:52:10 -07001903 offset += n_written;
Florin Corasbf4d5ce2018-10-19 16:26:24 -07001904 n_segs += 1;
Dave Barach68b0fb02017-02-28 15:15:56 -05001905 }
1906
Florin Coras36ee9f12018-11-02 12:52:10 -07001907 if (n_segs == burst_size)
1908 goto done;
Florin Corasbf4d5ce2018-10-19 16:26:24 -07001909
Florin Coras36ee9f12018-11-02 12:52:10 -07001910send_unsent:
1911
1912 /* RFC 6582: Send a new segment if permitted by the new value of cwnd. */
Florin Coras9ece3c02018-11-05 11:06:53 -08001913 if (snd_space < tc->snd_mss || tc->snd_mss == 0)
Florin Coras36ee9f12018-11-02 12:52:10 -07001914 goto done;
1915
Florin Coras31c99552019-03-01 13:00:58 -08001916 max_deq = transport_max_tx_dequeue (&tc->connection);
Florin Coras47596832019-03-12 18:58:54 -07001917 max_deq -= tc->snd_nxt - tc->snd_una;
Florin Coras36ee9f12018-11-02 12:52:10 -07001918 if (max_deq)
1919 {
1920 snd_space = clib_min (max_deq, snd_space);
1921 burst_size = clib_min (burst_size - n_segs, snd_space / tc->snd_mss);
Florin Coras36ebcff2019-09-12 18:36:44 -07001922 n_segs_now = tcp_transmit_unsent (wrk, tc, burst_size);
Florin Corasbf1f8b72019-11-04 14:39:33 -08001923 if (n_segs_now && max_deq > n_segs_now * tc->snd_mss)
Florin Coras36ebcff2019-09-12 18:36:44 -07001924 tcp_program_retransmit (tc);
Florin Coras36ee9f12018-11-02 12:52:10 -07001925 n_segs += n_segs_now;
1926 }
1927
Florin Coras36ee9f12018-11-02 12:52:10 -07001928done:
1929 tcp_fastrecovery_first_off (tc);
Florin Coras93992a92017-05-24 18:03:56 -07001930
Florin Corasc31dc312019-10-06 14:06:14 -07001931 sent_bytes = clib_min (n_segs * tc->snd_mss, burst_bytes);
1932 sent_bytes = cc_limited ? burst_bytes : sent_bytes;
1933 transport_connection_tx_pacer_update_bytes (&tc->connection, sent_bytes);
1934
1935 return n_segs;
Dave Barach68b0fb02017-02-28 15:15:56 -05001936}
Florin Coras26dd6de2019-07-23 23:54:47 -07001937
1938static int
1939tcp_send_acks (tcp_connection_t * tc, u32 max_burst_size)
1940{
1941 int j, n_acks;
1942
1943 if (!tc->pending_dupacks)
1944 {
Florin Corascb711a42019-10-16 19:28:17 -07001945 if (tcp_in_cong_recovery (tc) || !tcp_max_tx_deq (tc)
1946 || tc->state != TCP_STATE_ESTABLISHED)
1947 {
1948 tcp_send_ack (tc);
1949 return 1;
1950 }
1951 return 0;
Florin Coras26dd6de2019-07-23 23:54:47 -07001952 }
1953
1954 /* If we're supposed to send dupacks but have no ooo data
1955 * send only one ack */
1956 if (!vec_len (tc->snd_sacks))
1957 {
1958 tcp_send_ack (tc);
Florin Coras7fd59cc2020-03-12 15:50:57 +00001959 tc->dupacks_out += 1;
Florin Corascb711a42019-10-16 19:28:17 -07001960 tc->pending_dupacks = 0;
Florin Coras26dd6de2019-07-23 23:54:47 -07001961 return 1;
1962 }
1963
1964 /* Start with first sack block */
1965 tc->snd_sack_pos = 0;
1966
1967 /* Generate enough dupacks to cover all sack blocks. Do not generate
1968 * more sacks than the number of packets received. But do generate at
1969 * least 3, i.e., the number needed to signal congestion, if needed. */
1970 n_acks = vec_len (tc->snd_sacks) / TCP_OPTS_MAX_SACK_BLOCKS;
1971 n_acks = clib_min (n_acks, tc->pending_dupacks);
1972 n_acks = clib_max (n_acks, clib_min (tc->pending_dupacks, 3));
1973 for (j = 0; j < clib_min (n_acks, max_burst_size); j++)
1974 tcp_send_ack (tc);
1975
1976 if (n_acks < max_burst_size)
1977 {
1978 tc->pending_dupacks = 0;
1979 tc->snd_sack_pos = 0;
Florin Corasedfe0ee2019-07-29 18:13:25 -07001980 tc->dupacks_out += n_acks;
Florin Coras26dd6de2019-07-23 23:54:47 -07001981 return n_acks;
1982 }
1983 else
1984 {
1985 TCP_DBG ("constrained by burst size");
1986 tc->pending_dupacks = n_acks - max_burst_size;
Florin Corasedfe0ee2019-07-29 18:13:25 -07001987 tc->dupacks_out += max_burst_size;
Florin Coras26dd6de2019-07-23 23:54:47 -07001988 tcp_program_dupack (tc);
1989 return max_burst_size;
1990 }
1991}
1992
1993static int
Florin Coras36ebcff2019-09-12 18:36:44 -07001994tcp_do_retransmit (tcp_connection_t * tc, u32 max_burst_size)
Florin Coras26dd6de2019-07-23 23:54:47 -07001995{
Florin Coras26dd6de2019-07-23 23:54:47 -07001996 tcp_worker_ctx_t *wrk;
Florin Corasc31dc312019-10-06 14:06:14 -07001997 u32 n_segs;
Florin Coras26dd6de2019-07-23 23:54:47 -07001998
Florin Corasfd4c3fe2019-11-07 12:33:12 -08001999 if (PREDICT_FALSE (tc->state == TCP_STATE_CLOSED))
2000 return 0;
2001
Florin Coras26dd6de2019-07-23 23:54:47 -07002002 wrk = tcp_get_worker (tc->c_thread_index);
Florin Coras26dd6de2019-07-23 23:54:47 -07002003
Florin Corasc31dc312019-10-06 14:06:14 -07002004 if (tcp_opts_sack_permitted (&tc->rcv_opts))
2005 n_segs = tcp_retransmit_sack (wrk, tc, max_burst_size);
2006 else
2007 n_segs = tcp_retransmit_no_sack (wrk, tc, max_burst_size);
2008
Florin Coras26dd6de2019-07-23 23:54:47 -07002009 return n_segs;
2010}
2011
2012int
2013tcp_session_custom_tx (void *conn, u32 max_burst_size)
2014{
2015 tcp_connection_t *tc = (tcp_connection_t *) conn;
2016 u32 n_segs = 0;
2017
Florin Coras36ebcff2019-09-12 18:36:44 -07002018 if (tcp_in_cong_recovery (tc) && (tc->flags & TCP_CONN_RXT_PENDING))
Florin Coras26dd6de2019-07-23 23:54:47 -07002019 {
Florin Coras36ebcff2019-09-12 18:36:44 -07002020 tc->flags &= ~TCP_CONN_RXT_PENDING;
2021 n_segs = tcp_do_retransmit (tc, max_burst_size);
Florin Coras26dd6de2019-07-23 23:54:47 -07002022 max_burst_size -= n_segs;
2023 }
2024
2025 if (!(tc->flags & TCP_CONN_SNDACK))
2026 return n_segs;
2027
2028 tc->flags &= ~TCP_CONN_SNDACK;
2029
2030 /* We have retransmitted packets and no dupack */
2031 if (n_segs && !tc->pending_dupacks)
2032 return n_segs;
2033
2034 if (!max_burst_size)
2035 {
2036 tcp_program_ack (tc);
2037 return max_burst_size;
2038 }
2039
2040 n_segs += tcp_send_acks (tc, max_burst_size);
2041
2042 return n_segs;
2043}
Filip Tehlare275bed2019-03-06 00:06:56 -08002044#endif /* CLIB_MARCH_VARIANT */
Dave Barach68b0fb02017-02-28 15:15:56 -05002045
Florin Corasf9d05682018-04-26 08:26:52 -07002046static void
2047tcp_output_handle_link_local (tcp_connection_t * tc0, vlib_buffer_t * b0,
Florin Coras8b20bf52018-06-14 14:55:50 -07002048 u16 * next0, u32 * error0)
Florin Corasf9d05682018-04-26 08:26:52 -07002049{
2050 ip_adjacency_t *adj;
2051 adj_index_t ai;
2052
Florin Coras1c8ff632018-05-17 13:28:34 -07002053 /* Not thread safe but as long as the connection exists the adj should
2054 * not be removed */
Florin Corasf9d05682018-04-26 08:26:52 -07002055 ai = adj_nbr_find (FIB_PROTOCOL_IP6, VNET_LINK_IP6, &tc0->c_rmt_ip,
2056 tc0->sw_if_index);
2057 if (ai == ADJ_INDEX_INVALID)
2058 {
Florin Corasf9d05682018-04-26 08:26:52 -07002059 vnet_buffer (b0)->sw_if_index[VLIB_TX] = ~0;
2060 *next0 = TCP_OUTPUT_NEXT_DROP;
2061 *error0 = TCP_ERROR_LINK_LOCAL_RW;
2062 return;
2063 }
2064
2065 adj = adj_get (ai);
Florin Coras1c8ff632018-05-17 13:28:34 -07002066 if (PREDICT_TRUE (adj->lookup_next_index == IP_LOOKUP_NEXT_REWRITE))
Florin Corasf9d05682018-04-26 08:26:52 -07002067 *next0 = TCP_OUTPUT_NEXT_IP_REWRITE;
2068 else if (adj->lookup_next_index == IP_LOOKUP_NEXT_ARP)
2069 *next0 = TCP_OUTPUT_NEXT_IP_ARP;
Florin Coras1c8ff632018-05-17 13:28:34 -07002070 else
2071 {
2072 *next0 = TCP_OUTPUT_NEXT_DROP;
2073 *error0 = TCP_ERROR_LINK_LOCAL_RW;
2074 }
2075 vnet_buffer (b0)->ip.adj_index[VLIB_TX] = ai;
Florin Corasf9d05682018-04-26 08:26:52 -07002076}
2077
Florin Coras8b20bf52018-06-14 14:55:50 -07002078static void
2079tcp46_output_trace_frame (vlib_main_t * vm, vlib_node_runtime_t * node,
2080 u32 * to_next, u32 n_bufs)
Dave Barach68b0fb02017-02-28 15:15:56 -05002081{
Florin Coras8b20bf52018-06-14 14:55:50 -07002082 tcp_connection_t *tc;
2083 tcp_tx_trace_t *t;
2084 vlib_buffer_t *b;
2085 tcp_header_t *th;
2086 int i;
Dave Barach68b0fb02017-02-28 15:15:56 -05002087
Florin Coras30928f82020-01-27 19:21:28 -08002088 for (i = 0; i < n_bufs; i++)
Dave Barach68b0fb02017-02-28 15:15:56 -05002089 {
Florin Coras8b20bf52018-06-14 14:55:50 -07002090 b = vlib_get_buffer (vm, to_next[i]);
Florin Coras30928f82020-01-27 19:21:28 -08002091 if (!(b->flags & VLIB_BUFFER_IS_TRACED))
2092 continue;
Florin Coras8b20bf52018-06-14 14:55:50 -07002093 th = vlib_buffer_get_current (b);
2094 tc = tcp_connection_get (vnet_buffer (b)->tcp.connection_index,
2095 vm->thread_index);
2096 t = vlib_add_trace (vm, node, b, sizeof (*t));
Dave Barach178cf492018-11-13 16:34:13 -05002097 clib_memcpy_fast (&t->tcp_header, th, sizeof (t->tcp_header));
2098 clib_memcpy_fast (&t->tcp_connection, tc, sizeof (t->tcp_connection));
Florin Coras8b20bf52018-06-14 14:55:50 -07002099 }
2100}
Dave Barach68b0fb02017-02-28 15:15:56 -05002101
Florin Coras0dbd5172018-06-25 16:19:34 -07002102always_inline void
Florin Coras8b20bf52018-06-14 14:55:50 -07002103tcp_output_push_ip (vlib_main_t * vm, vlib_buffer_t * b0,
2104 tcp_connection_t * tc0, u8 is_ip4)
2105{
Florin Corasf4ce6ba2019-11-20 18:34:58 -08002106 TCP_EVT (TCP_EVT_OUTPUT, tc0,
2107 ((tcp_header_t *) vlib_buffer_get_current (b0))->flags,
2108 b0->current_length);
Srikanth A02833ff2019-10-02 17:48:58 -07002109
Florin Coras8b20bf52018-06-14 14:55:50 -07002110 if (is_ip4)
Florin Corasf4ce6ba2019-11-20 18:34:58 -08002111 vlib_buffer_push_ip4 (vm, b0, &tc0->c_lcl_ip4, &tc0->c_rmt_ip4,
2112 IP_PROTOCOL_TCP, tcp_csum_offload (tc0));
Florin Coras8b20bf52018-06-14 14:55:50 -07002113 else
Florin Corasf4ce6ba2019-11-20 18:34:58 -08002114 vlib_buffer_push_ip6_custom (vm, b0, &tc0->c_lcl_ip6, &tc0->c_rmt_ip6,
2115 IP_PROTOCOL_TCP, tc0->ipv6_flow_label);
Florin Coras8b20bf52018-06-14 14:55:50 -07002116}
Dave Barach68b0fb02017-02-28 15:15:56 -05002117
Florin Coras0dbd5172018-06-25 16:19:34 -07002118always_inline void
Simon Zhang1146ff42019-09-02 22:54:00 +08002119tcp_check_if_gso (tcp_connection_t * tc, vlib_buffer_t * b)
2120{
Florin Corasbbcfaac2019-10-10 13:52:04 -07002121 if (PREDICT_TRUE (!(tc->cfg_flags & TCP_CFG_F_TSO)))
Simon Zhang1146ff42019-09-02 22:54:00 +08002122 return;
Florin Corasbbcfaac2019-10-10 13:52:04 -07002123
Simon Zhang8a047ed2019-09-24 21:16:56 +08002124 u16 data_len = b->current_length - sizeof (tcp_header_t) - tc->snd_opts_len;
Simon Zhang1146ff42019-09-02 22:54:00 +08002125
Simon Zhang8a047ed2019-09-24 21:16:56 +08002126 if (PREDICT_FALSE (b->flags & VLIB_BUFFER_TOTAL_LENGTH_VALID))
2127 data_len += b->total_length_not_including_first_buffer;
2128
2129 if (PREDICT_TRUE (data_len <= tc->snd_mss))
2130 return;
2131 else
Simon Zhang1146ff42019-09-02 22:54:00 +08002132 {
2133 ASSERT ((b->flags & VNET_BUFFER_F_L3_HDR_OFFSET_VALID) != 0);
2134 ASSERT ((b->flags & VNET_BUFFER_F_L4_HDR_OFFSET_VALID) != 0);
2135 b->flags |= VNET_BUFFER_F_GSO;
2136 vnet_buffer2 (b)->gso_l4_hdr_sz =
2137 sizeof (tcp_header_t) + tc->snd_opts_len;
2138 vnet_buffer2 (b)->gso_size = tc->snd_mss;
2139 }
2140}
2141
2142always_inline void
Florin Coras8b20bf52018-06-14 14:55:50 -07002143tcp_output_handle_packet (tcp_connection_t * tc0, vlib_buffer_t * b0,
Florin Corasdf36f492019-08-18 18:09:28 -07002144 vlib_node_runtime_t * error_node, u16 * next0,
2145 u8 is_ip4)
Florin Coras8b20bf52018-06-14 14:55:50 -07002146{
Florin Coras540a8da2019-06-18 08:55:14 -07002147 /* If next_index is not drop use it */
Florin Coras07beade2019-06-20 12:18:31 -07002148 if (tc0->next_node_index)
2149 {
2150 *next0 = tc0->next_node_index;
2151 vnet_buffer (b0)->tcp.next_node_opaque = tc0->next_node_opaque;
2152 }
Florin Coras78dae002019-08-30 10:00:30 -07002153 else
2154 {
2155 *next0 = TCP_OUTPUT_NEXT_IP_LOOKUP;
2156 }
Florin Coras540a8da2019-06-18 08:55:14 -07002157
Florin Coras8b20bf52018-06-14 14:55:50 -07002158 vnet_buffer (b0)->sw_if_index[VLIB_TX] = tc0->c_fib_index;
2159 vnet_buffer (b0)->sw_if_index[VLIB_RX] = 0;
2160
2161 if (!is_ip4)
2162 {
Florin Corasdf36f492019-08-18 18:09:28 -07002163 u32 error0 = 0;
2164
Florin Coras8b20bf52018-06-14 14:55:50 -07002165 if (PREDICT_FALSE (ip6_address_is_link_local_unicast (&tc0->c_rmt_ip6)))
Florin Corasdf36f492019-08-18 18:09:28 -07002166 tcp_output_handle_link_local (tc0, b0, next0, &error0);
2167
2168 if (PREDICT_FALSE (error0))
2169 {
2170 b0->error = error_node->errors[error0];
2171 return;
2172 }
Florin Coras8b20bf52018-06-14 14:55:50 -07002173 }
2174
Florin Corasedfe0ee2019-07-29 18:13:25 -07002175 tc0->segs_out += 1;
Florin Coras8b20bf52018-06-14 14:55:50 -07002176}
2177
2178always_inline uword
2179tcp46_output_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
2180 vlib_frame_t * frame, int is_ip4)
2181{
2182 u32 n_left_from, *from, thread_index = vm->thread_index;
2183 vlib_buffer_t *bufs[VLIB_FRAME_SIZE], **b;
2184 u16 nexts[VLIB_FRAME_SIZE], *next;
Florin Corasdf36f492019-08-18 18:09:28 -07002185 vlib_node_runtime_t *error_node;
2186
2187 error_node = vlib_node_get_runtime (vm, tcp_node_index (output, is_ip4));
Florin Coras8b20bf52018-06-14 14:55:50 -07002188
2189 from = vlib_frame_vector_args (frame);
2190 n_left_from = frame->n_vectors;
Florin Corasbe72ae62018-11-01 11:23:03 -07002191 tcp_set_time_now (tcp_get_worker (thread_index));
Florin Coras8b20bf52018-06-14 14:55:50 -07002192
2193 if (PREDICT_FALSE (node->flags & VLIB_NODE_FLAG_TRACE))
2194 tcp46_output_trace_frame (vm, node, from, n_left_from);
2195
2196 vlib_get_buffers (vm, from, bufs, n_left_from);
2197 b = bufs;
2198 next = nexts;
2199
2200 while (n_left_from >= 4)
2201 {
Florin Coras8b20bf52018-06-14 14:55:50 -07002202 tcp_connection_t *tc0, *tc1;
2203
2204 {
2205 vlib_prefetch_buffer_header (b[2], STORE);
2206 CLIB_PREFETCH (b[2]->data, 2 * CLIB_CACHE_LINE_BYTES, STORE);
2207
2208 vlib_prefetch_buffer_header (b[3], STORE);
2209 CLIB_PREFETCH (b[3]->data, 2 * CLIB_CACHE_LINE_BYTES, STORE);
2210 }
2211
Florin Coras8b20bf52018-06-14 14:55:50 -07002212 tc0 = tcp_connection_get (vnet_buffer (b[0])->tcp.connection_index,
2213 thread_index);
2214 tc1 = tcp_connection_get (vnet_buffer (b[1])->tcp.connection_index,
2215 thread_index);
2216
Florin Coras78dae002019-08-30 10:00:30 -07002217 if (PREDICT_TRUE (!tc0 + !tc1 == 0))
2218 {
2219 tcp_output_push_ip (vm, b[0], tc0, is_ip4);
2220 tcp_output_push_ip (vm, b[1], tc1, is_ip4);
Florin Coras8b20bf52018-06-14 14:55:50 -07002221
Simon Zhang1146ff42019-09-02 22:54:00 +08002222 tcp_check_if_gso (tc0, b[0]);
2223 tcp_check_if_gso (tc1, b[1]);
2224
Florin Coras78dae002019-08-30 10:00:30 -07002225 tcp_output_handle_packet (tc0, b[0], error_node, &next[0], is_ip4);
2226 tcp_output_handle_packet (tc1, b[1], error_node, &next[1], is_ip4);
2227 }
2228 else
2229 {
2230 if (tc0 != 0)
2231 {
2232 tcp_output_push_ip (vm, b[0], tc0, is_ip4);
Simon Zhang1146ff42019-09-02 22:54:00 +08002233 tcp_check_if_gso (tc0, b[0]);
Florin Coras78dae002019-08-30 10:00:30 -07002234 tcp_output_handle_packet (tc0, b[0], error_node, &next[0],
2235 is_ip4);
2236 }
2237 else
2238 {
2239 b[0]->error = error_node->errors[TCP_ERROR_INVALID_CONNECTION];
2240 next[0] = TCP_OUTPUT_NEXT_DROP;
2241 }
2242 if (tc1 != 0)
2243 {
2244 tcp_output_push_ip (vm, b[1], tc1, is_ip4);
Simon Zhang1146ff42019-09-02 22:54:00 +08002245 tcp_check_if_gso (tc1, b[1]);
Florin Coras78dae002019-08-30 10:00:30 -07002246 tcp_output_handle_packet (tc1, b[1], error_node, &next[1],
2247 is_ip4);
2248 }
2249 else
2250 {
2251 b[1]->error = error_node->errors[TCP_ERROR_INVALID_CONNECTION];
2252 next[1] = TCP_OUTPUT_NEXT_DROP;
2253 }
2254 }
Florin Coras8b20bf52018-06-14 14:55:50 -07002255
2256 b += 2;
2257 next += 2;
2258 n_left_from -= 2;
2259 }
2260 while (n_left_from > 0)
2261 {
Florin Coras8b20bf52018-06-14 14:55:50 -07002262 tcp_connection_t *tc0;
2263
2264 if (n_left_from > 1)
2265 {
Florin Coras91ca4622018-06-30 11:27:59 -07002266 vlib_prefetch_buffer_header (b[1], STORE);
2267 CLIB_PREFETCH (b[1]->data, 2 * CLIB_CACHE_LINE_BYTES, STORE);
Florin Coras8b20bf52018-06-14 14:55:50 -07002268 }
2269
Florin Coras8b20bf52018-06-14 14:55:50 -07002270 tc0 = tcp_connection_get (vnet_buffer (b[0])->tcp.connection_index,
2271 thread_index);
2272
Florin Coras78dae002019-08-30 10:00:30 -07002273 if (PREDICT_TRUE (tc0 != 0))
2274 {
2275 tcp_output_push_ip (vm, b[0], tc0, is_ip4);
Simon Zhang1146ff42019-09-02 22:54:00 +08002276 tcp_check_if_gso (tc0, b[0]);
Florin Coras78dae002019-08-30 10:00:30 -07002277 tcp_output_handle_packet (tc0, b[0], error_node, &next[0], is_ip4);
2278 }
2279 else
2280 {
2281 b[0]->error = error_node->errors[TCP_ERROR_INVALID_CONNECTION];
2282 next[0] = TCP_OUTPUT_NEXT_DROP;
2283 }
Florin Coras8b20bf52018-06-14 14:55:50 -07002284
2285 b += 1;
2286 next += 1;
2287 n_left_from -= 1;
2288 }
2289
2290 vlib_buffer_enqueue_to_next (vm, node, from, nexts, frame->n_vectors);
Florin Corasdf36f492019-08-18 18:09:28 -07002291 vlib_node_increment_counter (vm, tcp_node_index (output, is_ip4),
2292 TCP_ERROR_PKTS_SENT, frame->n_vectors);
Florin Coras8b20bf52018-06-14 14:55:50 -07002293 return frame->n_vectors;
Dave Barach68b0fb02017-02-28 15:15:56 -05002294}
2295
Filip Tehlare275bed2019-03-06 00:06:56 -08002296VLIB_NODE_FN (tcp4_output_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
2297 vlib_frame_t * from_frame)
Dave Barach68b0fb02017-02-28 15:15:56 -05002298{
2299 return tcp46_output_inline (vm, node, from_frame, 1 /* is_ip4 */ );
2300}
2301
Filip Tehlare275bed2019-03-06 00:06:56 -08002302VLIB_NODE_FN (tcp6_output_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
2303 vlib_frame_t * from_frame)
Dave Barach68b0fb02017-02-28 15:15:56 -05002304{
2305 return tcp46_output_inline (vm, node, from_frame, 0 /* is_ip4 */ );
2306}
2307
Florin Corase69f4952017-03-07 10:06:24 -08002308/* *INDENT-OFF* */
Dave Barach68b0fb02017-02-28 15:15:56 -05002309VLIB_REGISTER_NODE (tcp4_output_node) =
2310{
Dave Barachf970d2f2018-11-26 18:34:33 -05002311 .name = "tcp4-output",
2312 /* Takes a vector of packets. */
2313 .vector_size = sizeof (u32),
2314 .n_errors = TCP_N_ERROR,
Dave Barach7fff3d22018-11-27 16:52:59 -05002315 .protocol_hint = VLIB_NODE_PROTO_HINT_TCP,
Dave Barachf970d2f2018-11-26 18:34:33 -05002316 .error_strings = tcp_error_strings,
2317 .n_next_nodes = TCP_OUTPUT_N_NEXT,
2318 .next_nodes = {
Dave Barach68b0fb02017-02-28 15:15:56 -05002319#define _(s,n) [TCP_OUTPUT_NEXT_##s] = n,
2320 foreach_tcp4_output_next
2321#undef _
Dave Barachf970d2f2018-11-26 18:34:33 -05002322 },
2323 .format_buffer = format_tcp_header,
2324 .format_trace = format_tcp_tx_trace,
Florin Corase69f4952017-03-07 10:06:24 -08002325};
2326/* *INDENT-ON* */
Dave Barach68b0fb02017-02-28 15:15:56 -05002327
Florin Corase69f4952017-03-07 10:06:24 -08002328/* *INDENT-OFF* */
Dave Barach68b0fb02017-02-28 15:15:56 -05002329VLIB_REGISTER_NODE (tcp6_output_node) =
2330{
Florin Corase69f4952017-03-07 10:06:24 -08002331 .name = "tcp6-output",
Dave Barach68b0fb02017-02-28 15:15:56 -05002332 /* Takes a vector of packets. */
Florin Corase69f4952017-03-07 10:06:24 -08002333 .vector_size = sizeof (u32),
2334 .n_errors = TCP_N_ERROR,
Dave Barach7fff3d22018-11-27 16:52:59 -05002335 .protocol_hint = VLIB_NODE_PROTO_HINT_TCP,
Florin Corase69f4952017-03-07 10:06:24 -08002336 .error_strings = tcp_error_strings,
2337 .n_next_nodes = TCP_OUTPUT_N_NEXT,
2338 .next_nodes = {
Dave Barach68b0fb02017-02-28 15:15:56 -05002339#define _(s,n) [TCP_OUTPUT_NEXT_##s] = n,
2340 foreach_tcp6_output_next
2341#undef _
Florin Corase69f4952017-03-07 10:06:24 -08002342 },
2343 .format_buffer = format_tcp_header,
2344 .format_trace = format_tcp_tx_trace,
2345};
2346/* *INDENT-ON* */
Dave Barach68b0fb02017-02-28 15:15:56 -05002347
Dave Barach68b0fb02017-02-28 15:15:56 -05002348typedef enum _tcp_reset_next
2349{
2350 TCP_RESET_NEXT_DROP,
2351 TCP_RESET_NEXT_IP_LOOKUP,
2352 TCP_RESET_N_NEXT
2353} tcp_reset_next_t;
2354
2355#define foreach_tcp4_reset_next \
2356 _(DROP, "error-drop") \
2357 _(IP_LOOKUP, "ip4-lookup")
2358
2359#define foreach_tcp6_reset_next \
2360 _(DROP, "error-drop") \
2361 _(IP_LOOKUP, "ip6-lookup")
2362
2363static uword
2364tcp46_send_reset_inline (vlib_main_t * vm, vlib_node_runtime_t * node,
2365 vlib_frame_t * from_frame, u8 is_ip4)
2366{
Florin Coras360336f2020-02-13 18:46:18 +00002367 u32 error0 = TCP_ERROR_RST_SENT, next0 = TCP_RESET_NEXT_IP_LOOKUP;
Dave Barach68b0fb02017-02-28 15:15:56 -05002368 u32 n_left_from, next_index, *from, *to_next;
Dave Barach68b0fb02017-02-28 15:15:56 -05002369
2370 from = vlib_frame_vector_args (from_frame);
2371 n_left_from = from_frame->n_vectors;
2372
2373 next_index = node->cached_next_index;
2374
2375 while (n_left_from > 0)
2376 {
2377 u32 n_left_to_next;
2378
2379 vlib_get_next_frame (vm, node, next_index, to_next, n_left_to_next);
2380
2381 while (n_left_from > 0 && n_left_to_next > 0)
2382 {
Dave Barach68b0fb02017-02-28 15:15:56 -05002383 vlib_buffer_t *b0;
Clement Durand6cf260c2017-04-13 13:27:04 +02002384 tcp_tx_trace_t *t0;
2385 tcp_header_t *th0;
Florin Coras360336f2020-02-13 18:46:18 +00002386 u32 bi0;
Dave Barach68b0fb02017-02-28 15:15:56 -05002387
2388 bi0 = from[0];
2389 to_next[0] = bi0;
2390 from += 1;
2391 to_next += 1;
2392 n_left_from -= 1;
2393 n_left_to_next -= 1;
2394
2395 b0 = vlib_get_buffer (vm, bi0);
Florin Coras360336f2020-02-13 18:46:18 +00002396 tcp_make_reset_in_place (vm, b0, is_ip4);
Dave Barach68b0fb02017-02-28 15:15:56 -05002397
2398 /* Prepare to send to IP lookup */
Florin Corasf988e692017-11-27 04:34:14 -05002399 vnet_buffer (b0)->sw_if_index[VLIB_TX] = ~0;
Dave Barach68b0fb02017-02-28 15:15:56 -05002400
Florin Corase69f4952017-03-07 10:06:24 -08002401 b0->error = node->errors[error0];
Damjan Marion213b5aa2017-07-13 21:19:27 +02002402 b0->flags |= VNET_BUFFER_F_LOCALLY_ORIGINATED;
Dave Barach68b0fb02017-02-28 15:15:56 -05002403 if (PREDICT_FALSE (b0->flags & VLIB_BUFFER_IS_TRACED))
2404 {
Clement Durand6cf260c2017-04-13 13:27:04 +02002405 th0 = vlib_buffer_get_current (b0);
2406 if (is_ip4)
2407 th0 = ip4_next_header ((ip4_header_t *) th0);
2408 else
2409 th0 = ip6_next_header ((ip6_header_t *) th0);
Clement Durand6cf260c2017-04-13 13:27:04 +02002410 t0 = vlib_add_trace (vm, node, b0, sizeof (*t0));
Dave Barach178cf492018-11-13 16:34:13 -05002411 clib_memcpy_fast (&t0->tcp_header, th0,
2412 sizeof (t0->tcp_header));
Dave Barach68b0fb02017-02-28 15:15:56 -05002413 }
2414
2415 vlib_validate_buffer_enqueue_x1 (vm, node, next_index, to_next,
2416 n_left_to_next, bi0, next0);
2417 }
2418 vlib_put_next_frame (vm, node, next_index, n_left_to_next);
2419 }
2420 return from_frame->n_vectors;
2421}
2422
Filip Tehlare275bed2019-03-06 00:06:56 -08002423VLIB_NODE_FN (tcp4_reset_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
2424 vlib_frame_t * from_frame)
Dave Barach68b0fb02017-02-28 15:15:56 -05002425{
2426 return tcp46_send_reset_inline (vm, node, from_frame, 1);
2427}
2428
Filip Tehlare275bed2019-03-06 00:06:56 -08002429VLIB_NODE_FN (tcp6_reset_node) (vlib_main_t * vm, vlib_node_runtime_t * node,
2430 vlib_frame_t * from_frame)
Dave Barach68b0fb02017-02-28 15:15:56 -05002431{
2432 return tcp46_send_reset_inline (vm, node, from_frame, 0);
2433}
2434
2435/* *INDENT-OFF* */
2436VLIB_REGISTER_NODE (tcp4_reset_node) = {
Dave Barach68b0fb02017-02-28 15:15:56 -05002437 .name = "tcp4-reset",
2438 .vector_size = sizeof (u32),
2439 .n_errors = TCP_N_ERROR,
2440 .error_strings = tcp_error_strings,
2441 .n_next_nodes = TCP_RESET_N_NEXT,
2442 .next_nodes = {
2443#define _(s,n) [TCP_RESET_NEXT_##s] = n,
2444 foreach_tcp4_reset_next
2445#undef _
2446 },
Clement Durand6cf260c2017-04-13 13:27:04 +02002447 .format_trace = format_tcp_tx_trace,
Dave Barach68b0fb02017-02-28 15:15:56 -05002448};
2449/* *INDENT-ON* */
2450
2451/* *INDENT-OFF* */
2452VLIB_REGISTER_NODE (tcp6_reset_node) = {
Dave Barach68b0fb02017-02-28 15:15:56 -05002453 .name = "tcp6-reset",
2454 .vector_size = sizeof (u32),
2455 .n_errors = TCP_N_ERROR,
2456 .error_strings = tcp_error_strings,
2457 .n_next_nodes = TCP_RESET_N_NEXT,
2458 .next_nodes = {
2459#define _(s,n) [TCP_RESET_NEXT_##s] = n,
2460 foreach_tcp6_reset_next
2461#undef _
2462 },
Clement Durand6cf260c2017-04-13 13:27:04 +02002463 .format_trace = format_tcp_tx_trace,
Dave Barach68b0fb02017-02-28 15:15:56 -05002464};
2465/* *INDENT-ON* */
2466
Dave Barach68b0fb02017-02-28 15:15:56 -05002467/*
2468 * fd.io coding-style-patch-verification: ON
2469 *
2470 * Local Variables:
2471 * eval: (c-set-style "gnu")
2472 * End:
2473 */