1 // SPDX-License-Identifier: GPL-2.0
2 /* Multipath TCP
3 *
4 * Copyright (c) 2017 - 2019, Intel Corporation.
5 */
6
7 #define pr_fmt(fmt) "MPTCP: " fmt
8
9 #include <linux/kernel.h>
10 #include <crypto/sha2.h>
11 #include <net/tcp.h>
12 #include <net/mptcp.h>
13 #include "protocol.h"
14 #include "mib.h"
15
16 #include <trace/events/mptcp.h>
17
mptcp_cap_flag_sha256(u8 flags)18 static bool mptcp_cap_flag_sha256(u8 flags)
19 {
20 return (flags & MPTCP_CAP_FLAG_MASK) == MPTCP_CAP_HMAC_SHA256;
21 }
22
mptcp_parse_option(const struct sk_buff * skb,const unsigned char * ptr,int opsize,struct mptcp_options_received * mp_opt)23 static void mptcp_parse_option(const struct sk_buff *skb,
24 const unsigned char *ptr, int opsize,
25 struct mptcp_options_received *mp_opt)
26 {
27 u8 subtype = *ptr >> 4;
28 int expected_opsize;
29 u16 subopt;
30 u8 version;
31 u8 flags;
32 u8 i;
33
34 switch (subtype) {
35 case MPTCPOPT_MP_CAPABLE:
36 /* strict size checking */
37 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
38 if (skb->len > tcp_hdr(skb)->doff << 2)
39 expected_opsize = TCPOLEN_MPTCP_MPC_ACK_DATA;
40 else
41 expected_opsize = TCPOLEN_MPTCP_MPC_ACK;
42 subopt = OPTION_MPTCP_MPC_ACK;
43 } else {
44 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK) {
45 expected_opsize = TCPOLEN_MPTCP_MPC_SYNACK;
46 subopt = OPTION_MPTCP_MPC_SYNACK;
47 } else {
48 expected_opsize = TCPOLEN_MPTCP_MPC_SYN;
49 subopt = OPTION_MPTCP_MPC_SYN;
50 }
51 }
52
53 /* Only the MPC + ACK can be used with a RM_ADDR */
54 if (subopt == OPTION_MPTCP_MPC_ACK) {
55 if ((mp_opt->suboptions & ~OPTION_MPTCP_RM_ADDR) != 0)
56 break;
57 } else if (mp_opt->suboptions != 0) {
58 break;
59 }
60
61 /* Cfr RFC 8684 Section 3.3.0:
62 * If a checksum is present but its use had
63 * not been negotiated in the MP_CAPABLE handshake, the receiver MUST
64 * close the subflow with a RST, as it is not behaving as negotiated.
65 * If a checksum is not present when its use has been negotiated, the
66 * receiver MUST close the subflow with a RST, as it is considered
67 * broken
68 * We parse even option with mismatching csum presence, so that
69 * later in subflow_data_ready we can trigger the reset.
70 */
71 if (opsize != expected_opsize &&
72 (expected_opsize != TCPOLEN_MPTCP_MPC_ACK_DATA ||
73 opsize != TCPOLEN_MPTCP_MPC_ACK_DATA_CSUM))
74 break;
75
76 /* try to be gentle vs future versions on the initial syn */
77 version = *ptr++ & MPTCP_VERSION_MASK;
78 if (opsize != TCPOLEN_MPTCP_MPC_SYN) {
79 if (version != MPTCP_SUPPORTED_VERSION)
80 break;
81 } else if (version < MPTCP_SUPPORTED_VERSION) {
82 break;
83 }
84
85 flags = *ptr++;
86 if (!mptcp_cap_flag_sha256(flags) ||
87 (flags & MPTCP_CAP_EXTENSIBILITY))
88 break;
89
90 /* RFC 6824, Section 3.1:
91 * "For the Checksum Required bit (labeled "A"), if either
92 * host requires the use of checksums, checksums MUST be used.
93 * In other words, the only way for checksums not to be used
94 * is if both hosts in their SYNs set A=0."
95 */
96 if ((flags & MPTCP_CAP_CHECKSUM_REQD) &&
97 opsize < TCPOLEN_MPTCP_MPC_ACK_DATA)
98 mp_opt->suboptions |= OPTION_MPTCP_CSUMREQD;
99
100 mp_opt->deny_join_id0 = !!(flags & MPTCP_CAP_DENY_JOIN_ID0);
101
102 mp_opt->suboptions |= subopt;
103 if (opsize >= TCPOLEN_MPTCP_MPC_SYNACK) {
104 mp_opt->sndr_key = get_unaligned_be64(ptr);
105 ptr += 8;
106 }
107 if (opsize >= TCPOLEN_MPTCP_MPC_ACK) {
108 mp_opt->rcvr_key = get_unaligned_be64(ptr);
109 ptr += 8;
110 }
111 if (opsize >= TCPOLEN_MPTCP_MPC_ACK_DATA) {
112 /* Section 3.1.:
113 * "the data parameters in a MP_CAPABLE are semantically
114 * equivalent to those in a DSS option and can be used
115 * interchangeably."
116 */
117 mp_opt->suboptions |= OPTION_MPTCP_DSS;
118 mp_opt->use_map = 1;
119 mp_opt->mpc_map = 1;
120 mp_opt->data_len = get_unaligned_be16(ptr);
121 ptr += 2;
122 }
123 if (opsize == TCPOLEN_MPTCP_MPC_ACK_DATA_CSUM) {
124 mp_opt->csum = get_unaligned((__force __sum16 *)ptr);
125 mp_opt->suboptions |= OPTION_MPTCP_CSUMREQD;
126 ptr += 2;
127 }
128 pr_debug("MP_CAPABLE version=%x, flags=%x, optlen=%d sndr=%llu, rcvr=%llu len=%d csum=%u\n",
129 version, flags, opsize, mp_opt->sndr_key,
130 mp_opt->rcvr_key, mp_opt->data_len, mp_opt->csum);
131 break;
132
133 case MPTCPOPT_MP_JOIN:
134 /* Can be used with a restricted number of other options */
135 if ((mp_opt->suboptions & ~(OPTION_MPTCP_RM_ADDR |
136 OPTION_MPTCP_PRIO)) != 0)
137 break;
138
139 if (opsize == TCPOLEN_MPTCP_MPJ_SYN) {
140 mp_opt->suboptions |= OPTION_MPTCP_MPJ_SYN;
141 mp_opt->backup = *ptr++ & MPTCPOPT_BACKUP;
142 mp_opt->join_id = *ptr++;
143 mp_opt->token = get_unaligned_be32(ptr);
144 ptr += 4;
145 mp_opt->nonce = get_unaligned_be32(ptr);
146 ptr += 4;
147 pr_debug("MP_JOIN bkup=%u, id=%u, token=%u, nonce=%u\n",
148 mp_opt->backup, mp_opt->join_id,
149 mp_opt->token, mp_opt->nonce);
150 } else if (opsize == TCPOLEN_MPTCP_MPJ_SYNACK) {
151 mp_opt->suboptions |= OPTION_MPTCP_MPJ_SYNACK;
152 mp_opt->backup = *ptr++ & MPTCPOPT_BACKUP;
153 mp_opt->join_id = *ptr++;
154 mp_opt->thmac = get_unaligned_be64(ptr);
155 ptr += 8;
156 mp_opt->nonce = get_unaligned_be32(ptr);
157 ptr += 4;
158 pr_debug("MP_JOIN bkup=%u, id=%u, thmac=%llu, nonce=%u\n",
159 mp_opt->backup, mp_opt->join_id,
160 mp_opt->thmac, mp_opt->nonce);
161 } else if (opsize == TCPOLEN_MPTCP_MPJ_ACK) {
162 mp_opt->suboptions |= OPTION_MPTCP_MPJ_ACK;
163 ptr += 2;
164 memcpy(mp_opt->hmac, ptr, MPTCPOPT_HMAC_LEN);
165 pr_debug("MP_JOIN hmac\n");
166 }
167 break;
168
169 case MPTCPOPT_DSS:
170 /* Can be used with a restricted number of other options */
171 if ((mp_opt->suboptions & ~(OPTION_MPTCP_ADD_ADDR |
172 OPTION_MPTCP_RM_ADDR |
173 OPTION_MPTCP_PRIO |
174 OPTION_MPTCP_FASTCLOSE |
175 OPTION_MPTCP_FAIL)) != 0)
176 break;
177
178 pr_debug("DSS\n");
179 ptr++;
180
181 flags = (*ptr++) & MPTCP_DSS_FLAG_MASK;
182 mp_opt->dsn64 = (flags & MPTCP_DSS_DSN64) != 0;
183 mp_opt->use_map = (flags & MPTCP_DSS_HAS_MAP) != 0;
184 mp_opt->ack64 = (flags & MPTCP_DSS_ACK64) != 0;
185 mp_opt->use_ack = (flags & MPTCP_DSS_HAS_ACK);
186
187 expected_opsize = TCPOLEN_MPTCP_DSS_BASE;
188
189 if (mp_opt->use_ack) {
190 if (mp_opt->ack64)
191 expected_opsize += TCPOLEN_MPTCP_DSS_ACK64;
192 else
193 expected_opsize += TCPOLEN_MPTCP_DSS_ACK32;
194 }
195
196 if (mp_opt->use_map) {
197 mp_opt->data_fin = (flags & MPTCP_DSS_DATA_FIN) != 0;
198 if (mp_opt->dsn64)
199 expected_opsize += TCPOLEN_MPTCP_DSS_MAP64;
200 else
201 expected_opsize += TCPOLEN_MPTCP_DSS_MAP32;
202 }
203
204 pr_debug("data_fin=%d dsn64=%d use_map=%d ack64=%d use_ack=%d\n",
205 mp_opt->data_fin, mp_opt->dsn64,
206 mp_opt->use_map, mp_opt->ack64,
207 mp_opt->use_ack);
208
209 /* Always parse any csum presence combination, we will enforce
210 * RFC 8684 Section 3.3.0 checks later in subflow_data_ready
211 */
212 if (opsize != expected_opsize &&
213 opsize != expected_opsize + TCPOLEN_MPTCP_DSS_CHECKSUM) {
214 mp_opt->dsn64 = 0;
215 mp_opt->use_map = 0;
216 mp_opt->ack64 = 0;
217 mp_opt->use_ack = 0;
218 mp_opt->data_fin = 0;
219 break;
220 }
221
222 mp_opt->suboptions |= OPTION_MPTCP_DSS;
223 if (mp_opt->use_ack) {
224 if (mp_opt->ack64) {
225 mp_opt->data_ack = get_unaligned_be64(ptr);
226 ptr += 8;
227 } else {
228 mp_opt->data_ack = get_unaligned_be32(ptr);
229 ptr += 4;
230 }
231
232 pr_debug("data_ack=%llu\n", mp_opt->data_ack);
233 }
234
235 if (mp_opt->use_map) {
236 if (mp_opt->dsn64) {
237 mp_opt->data_seq = get_unaligned_be64(ptr);
238 ptr += 8;
239 } else {
240 mp_opt->data_seq = get_unaligned_be32(ptr);
241 ptr += 4;
242 }
243
244 mp_opt->subflow_seq = get_unaligned_be32(ptr);
245 ptr += 4;
246
247 mp_opt->data_len = get_unaligned_be16(ptr);
248 ptr += 2;
249
250 if (opsize == expected_opsize + TCPOLEN_MPTCP_DSS_CHECKSUM) {
251 mp_opt->suboptions |= OPTION_MPTCP_CSUMREQD;
252 mp_opt->csum = get_unaligned((__force __sum16 *)ptr);
253 ptr += 2;
254 }
255
256 pr_debug("data_seq=%llu subflow_seq=%u data_len=%u csum=%d:%u\n",
257 mp_opt->data_seq, mp_opt->subflow_seq,
258 mp_opt->data_len, !!(mp_opt->suboptions & OPTION_MPTCP_CSUMREQD),
259 mp_opt->csum);
260 }
261
262 break;
263
264 case MPTCPOPT_ADD_ADDR:
265 /* Can be used with a restricted number of other options */
266 if ((mp_opt->suboptions & ~(OPTIONS_MPTCP_DSS |
267 OPTION_MPTCP_RM_ADDR |
268 OPTION_MPTCP_PRIO)) != 0)
269 break;
270
271 mp_opt->echo = (*ptr++) & MPTCP_ADDR_ECHO;
272 if (!mp_opt->echo) {
273 if (opsize == TCPOLEN_MPTCP_ADD_ADDR ||
274 opsize == TCPOLEN_MPTCP_ADD_ADDR_PORT)
275 mp_opt->addr.family = AF_INET;
276 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
277 else if (opsize == TCPOLEN_MPTCP_ADD_ADDR6 ||
278 opsize == TCPOLEN_MPTCP_ADD_ADDR6_PORT)
279 mp_opt->addr.family = AF_INET6;
280 #endif
281 else
282 break;
283 } else {
284 if (opsize == TCPOLEN_MPTCP_ADD_ADDR_BASE ||
285 opsize == TCPOLEN_MPTCP_ADD_ADDR_BASE_PORT)
286 mp_opt->addr.family = AF_INET;
287 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
288 else if (opsize == TCPOLEN_MPTCP_ADD_ADDR6_BASE ||
289 opsize == TCPOLEN_MPTCP_ADD_ADDR6_BASE_PORT)
290 mp_opt->addr.family = AF_INET6;
291 #endif
292 else
293 break;
294 }
295
296 mp_opt->suboptions |= OPTION_MPTCP_ADD_ADDR;
297 mp_opt->addr.id = *ptr++;
298 mp_opt->addr.port = 0;
299 mp_opt->ahmac = 0;
300 if (mp_opt->addr.family == AF_INET) {
301 memcpy((u8 *)&mp_opt->addr.addr.s_addr, (u8 *)ptr, 4);
302 ptr += 4;
303 if (opsize == TCPOLEN_MPTCP_ADD_ADDR_PORT ||
304 opsize == TCPOLEN_MPTCP_ADD_ADDR_BASE_PORT) {
305 mp_opt->addr.port = htons(get_unaligned_be16(ptr));
306 ptr += 2;
307 }
308 }
309 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
310 else {
311 memcpy(mp_opt->addr.addr6.s6_addr, (u8 *)ptr, 16);
312 ptr += 16;
313 if (opsize == TCPOLEN_MPTCP_ADD_ADDR6_PORT ||
314 opsize == TCPOLEN_MPTCP_ADD_ADDR6_BASE_PORT) {
315 mp_opt->addr.port = htons(get_unaligned_be16(ptr));
316 ptr += 2;
317 }
318 }
319 #endif
320 if (!mp_opt->echo) {
321 mp_opt->ahmac = get_unaligned_be64(ptr);
322 ptr += 8;
323 }
324 pr_debug("ADD_ADDR%s: id=%d, ahmac=%llu, echo=%d, port=%d\n",
325 (mp_opt->addr.family == AF_INET6) ? "6" : "",
326 mp_opt->addr.id, mp_opt->ahmac, mp_opt->echo, ntohs(mp_opt->addr.port));
327 break;
328
329 case MPTCPOPT_RM_ADDR:
330 /* Can be used with a restricted number of other options */
331 if ((mp_opt->suboptions & ~(OPTION_MPTCP_MPC_ACK |
332 OPTIONS_MPTCP_MPJ |
333 OPTIONS_MPTCP_DSS |
334 OPTION_MPTCP_ADD_ADDR |
335 OPTION_MPTCP_PRIO)) != 0)
336 break;
337
338 if (opsize < TCPOLEN_MPTCP_RM_ADDR_BASE + 1 ||
339 opsize > TCPOLEN_MPTCP_RM_ADDR_BASE + MPTCP_RM_IDS_MAX)
340 break;
341
342 ptr++;
343
344 mp_opt->suboptions |= OPTION_MPTCP_RM_ADDR;
345 mp_opt->rm_list.nr = opsize - TCPOLEN_MPTCP_RM_ADDR_BASE;
346 for (i = 0; i < mp_opt->rm_list.nr; i++)
347 mp_opt->rm_list.ids[i] = *ptr++;
348 pr_debug("RM_ADDR: rm_list_nr=%d\n", mp_opt->rm_list.nr);
349 break;
350
351 case MPTCPOPT_MP_PRIO:
352 /* Can be used with a restricted number of other options */
353 if ((mp_opt->suboptions & ~(OPTIONS_MPTCP_MPJ |
354 OPTIONS_MPTCP_DSS |
355 OPTION_MPTCP_ADD_ADDR |
356 OPTION_MPTCP_RM_ADDR)) != 0)
357 break;
358
359 if (opsize != TCPOLEN_MPTCP_PRIO)
360 break;
361
362 mp_opt->suboptions |= OPTION_MPTCP_PRIO;
363 mp_opt->backup = *ptr++ & MPTCP_PRIO_BKUP;
364 pr_debug("MP_PRIO: prio=%d\n", mp_opt->backup);
365 break;
366
367 case MPTCPOPT_MP_FASTCLOSE:
368 /* Can be used with a restricted number of other options */
369 if ((mp_opt->suboptions & ~(OPTIONS_MPTCP_DSS |
370 OPTION_MPTCP_RST)) != 0)
371 break;
372
373 if (opsize != TCPOLEN_MPTCP_FASTCLOSE)
374 break;
375
376 ptr += 2;
377 mp_opt->rcvr_key = get_unaligned_be64(ptr);
378 ptr += 8;
379 mp_opt->suboptions |= OPTION_MPTCP_FASTCLOSE;
380 pr_debug("MP_FASTCLOSE: recv_key=%llu\n", mp_opt->rcvr_key);
381 break;
382
383 case MPTCPOPT_RST:
384 /* Can be used with a restricted number of other options */
385 if ((mp_opt->suboptions & ~(OPTION_MPTCP_FAIL |
386 OPTION_MPTCP_FASTCLOSE)) != 0)
387 break;
388
389 if (opsize != TCPOLEN_MPTCP_RST)
390 break;
391
392 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_RST))
393 break;
394
395 mp_opt->suboptions |= OPTION_MPTCP_RST;
396 flags = *ptr++;
397 mp_opt->reset_transient = flags & MPTCP_RST_TRANSIENT;
398 mp_opt->reset_reason = *ptr;
399 pr_debug("MP_RST: transient=%u reason=%u\n",
400 mp_opt->reset_transient, mp_opt->reset_reason);
401 break;
402
403 case MPTCPOPT_MP_FAIL:
404 /* Can be used with a restricted number of other options */
405 if ((mp_opt->suboptions & ~(OPTIONS_MPTCP_DSS |
406 OPTION_MPTCP_RST)) != 0)
407 break;
408
409 if (opsize != TCPOLEN_MPTCP_FAIL)
410 break;
411
412 ptr += 2;
413 mp_opt->suboptions |= OPTION_MPTCP_FAIL;
414 mp_opt->fail_seq = get_unaligned_be64(ptr);
415 pr_debug("MP_FAIL: data_seq=%llu\n", mp_opt->fail_seq);
416 break;
417
418 default:
419 break;
420 }
421 }
422
mptcp_get_options(const struct sk_buff * skb,struct mptcp_options_received * mp_opt)423 void mptcp_get_options(const struct sk_buff *skb,
424 struct mptcp_options_received *mp_opt)
425 {
426 const struct tcphdr *th = tcp_hdr(skb);
427 const unsigned char *ptr;
428 int length;
429
430 /* Ensure that casting the whole status to u32 is efficient and safe */
431 BUILD_BUG_ON(sizeof_field(struct mptcp_options_received, status) != sizeof(u32));
432 BUILD_BUG_ON(!IS_ALIGNED(offsetof(struct mptcp_options_received, status),
433 sizeof(u32)));
434 *(u32 *)&mp_opt->status = 0;
435
436 length = (th->doff * 4) - sizeof(struct tcphdr);
437 ptr = (const unsigned char *)(th + 1);
438
439 while (length > 0) {
440 int opcode = *ptr++;
441 int opsize;
442
443 switch (opcode) {
444 case TCPOPT_EOL:
445 return;
446 case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */
447 length--;
448 continue;
449 default:
450 if (length < 2)
451 return;
452 opsize = *ptr++;
453 if (opsize < 2) /* "silly options" */
454 return;
455 if (opsize > length)
456 return; /* don't parse partial options */
457 if (opcode == TCPOPT_MPTCP)
458 mptcp_parse_option(skb, ptr, opsize, mp_opt);
459 ptr += opsize - 2;
460 length -= opsize;
461 }
462 }
463 }
464
mptcp_syn_options(struct sock * sk,const struct sk_buff * skb,unsigned int * size,struct mptcp_out_options * opts)465 bool mptcp_syn_options(struct sock *sk, const struct sk_buff *skb,
466 unsigned int *size, struct mptcp_out_options *opts)
467 {
468 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
469
470 /* we will use snd_isn to detect first pkt [re]transmission
471 * in mptcp_established_options_mp()
472 */
473 subflow->snd_isn = TCP_SKB_CB(skb)->end_seq;
474 if (subflow->request_mptcp) {
475 if (unlikely(subflow_simultaneous_connect(sk))) {
476 WARN_ON_ONCE(!mptcp_try_fallback(sk, MPTCP_MIB_SIMULTCONNFALLBACK));
477
478 /* Ensure mptcp_finish_connect() will not process the
479 * MPC handshake.
480 */
481 subflow->request_mptcp = 0;
482 return false;
483 }
484
485 opts->suboptions = OPTION_MPTCP_MPC_SYN;
486 opts->csum_reqd = mptcp_is_checksum_enabled(sock_net(sk));
487 opts->allow_join_id0 = mptcp_allow_join_id0(sock_net(sk));
488 *size = TCPOLEN_MPTCP_MPC_SYN;
489 return true;
490 } else if (subflow->request_join) {
491 pr_debug("remote_token=%u, nonce=%u\n", subflow->remote_token,
492 subflow->local_nonce);
493 opts->suboptions = OPTION_MPTCP_MPJ_SYN;
494 opts->join_id = subflow->local_id;
495 opts->token = subflow->remote_token;
496 opts->nonce = subflow->local_nonce;
497 opts->backup = subflow->request_bkup;
498 *size = TCPOLEN_MPTCP_MPJ_SYN;
499 return true;
500 }
501 return false;
502 }
503
clear_3rdack_retransmission(struct sock * sk)504 static void clear_3rdack_retransmission(struct sock *sk)
505 {
506 struct inet_connection_sock *icsk = inet_csk(sk);
507
508 sk_stop_timer(sk, &icsk->icsk_delack_timer);
509 icsk->icsk_ack.ato = 0;
510 icsk->icsk_ack.pending &= ~(ICSK_ACK_SCHED | ICSK_ACK_TIMER);
511 }
512
mptcp_established_options_mp(struct sock * sk,struct sk_buff * skb,bool snd_data_fin_enable,int * size,struct mptcp_out_options * opts)513 static bool mptcp_established_options_mp(struct sock *sk, struct sk_buff *skb,
514 bool snd_data_fin_enable, int *size,
515 struct mptcp_out_options *opts)
516 {
517 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
518 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
519 struct mptcp_ext *mpext;
520 unsigned int data_len;
521 u8 len;
522
523 /* When skb is not available, we better over-estimate the emitted
524 * options len. A full DSS option (28 bytes) is longer than
525 * TCPOLEN_MPTCP_MPC_ACK_DATA(22) or TCPOLEN_MPTCP_MPJ_ACK(24), so
526 * tell the caller to defer the estimate to
527 * mptcp_established_options_dss(), which will reserve enough space.
528 */
529 if (!skb)
530 return false;
531
532 /* MPC/MPJ needed only on 3rd ack packet, DATA_FIN and TCP shutdown take precedence */
533 if (subflow->fully_established || snd_data_fin_enable ||
534 subflow->snd_isn != TCP_SKB_CB(skb)->seq ||
535 sk->sk_state != TCP_ESTABLISHED)
536 return false;
537
538 if (subflow->mp_capable) {
539 mpext = mptcp_get_ext(skb);
540 data_len = mpext ? mpext->data_len : 0;
541
542 /* we will check ops->data_len in mptcp_write_options() to
543 * discriminate between TCPOLEN_MPTCP_MPC_ACK_DATA and
544 * TCPOLEN_MPTCP_MPC_ACK
545 */
546 opts->data_len = data_len;
547 opts->suboptions = OPTION_MPTCP_MPC_ACK;
548 opts->sndr_key = subflow->local_key;
549 opts->rcvr_key = subflow->remote_key;
550 opts->csum_reqd = READ_ONCE(msk->csum_enabled);
551 opts->allow_join_id0 = mptcp_allow_join_id0(sock_net(sk));
552
553 /* Section 3.1.
554 * The MP_CAPABLE option is carried on the SYN, SYN/ACK, and ACK
555 * packets that start the first subflow of an MPTCP connection,
556 * as well as the first packet that carries data
557 */
558 if (data_len > 0) {
559 len = TCPOLEN_MPTCP_MPC_ACK_DATA;
560 if (opts->csum_reqd) {
561 /* we need to propagate more info to csum the pseudo hdr */
562 opts->data_seq = mpext->data_seq;
563 opts->subflow_seq = mpext->subflow_seq;
564 opts->csum = mpext->csum;
565 len += TCPOLEN_MPTCP_DSS_CHECKSUM;
566 }
567 *size = ALIGN(len, 4);
568 } else {
569 *size = TCPOLEN_MPTCP_MPC_ACK;
570 }
571
572 pr_debug("subflow=%p, local_key=%llu, remote_key=%llu map_len=%d\n",
573 subflow, subflow->local_key, subflow->remote_key,
574 data_len);
575
576 return true;
577 } else if (subflow->mp_join) {
578 opts->suboptions = OPTION_MPTCP_MPJ_ACK;
579 memcpy(opts->hmac, subflow->hmac, MPTCPOPT_HMAC_LEN);
580 *size = TCPOLEN_MPTCP_MPJ_ACK;
581 pr_debug("subflow=%p\n", subflow);
582
583 /* we can use the full delegate action helper only from BH context
584 * If we are in process context - sk is flushing the backlog at
585 * socket lock release time - just set the appropriate flag, will
586 * be handled by the release callback
587 */
588 if (sock_owned_by_user(sk))
589 set_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status);
590 else
591 mptcp_subflow_delegate(subflow, MPTCP_DELEGATE_ACK);
592 return true;
593 }
594 return false;
595 }
596
mptcp_write_data_fin(struct mptcp_subflow_context * subflow,struct sk_buff * skb,struct mptcp_ext * ext)597 static void mptcp_write_data_fin(struct mptcp_subflow_context *subflow,
598 struct sk_buff *skb, struct mptcp_ext *ext)
599 {
600 /* The write_seq value has already been incremented, so the actual
601 * sequence number for the DATA_FIN is one less.
602 */
603 u64 data_fin_tx_seq = READ_ONCE(mptcp_sk(subflow->conn)->write_seq) - 1;
604
605 if (!ext->use_map || !skb->len) {
606 /* RFC6824 requires a DSS mapping with specific values
607 * if DATA_FIN is set but no data payload is mapped
608 */
609 ext->data_fin = 1;
610 ext->use_map = 1;
611 ext->dsn64 = 1;
612 ext->data_seq = data_fin_tx_seq;
613 ext->subflow_seq = 0;
614 ext->data_len = 1;
615 ext->csum = 0;
616 } else if (ext->data_seq + ext->data_len == data_fin_tx_seq) {
617 /* If there's an existing DSS mapping and it is the
618 * final mapping, DATA_FIN consumes 1 additional byte of
619 * mapping space.
620 */
621 ext->data_fin = 1;
622 ext->data_len++;
623 }
624 }
625
mptcp_established_options_dss(struct sock * sk,struct sk_buff * skb,bool snd_data_fin_enable,int * size,struct mptcp_out_options * opts)626 static bool mptcp_established_options_dss(struct sock *sk, struct sk_buff *skb,
627 bool snd_data_fin_enable, int *size,
628 struct mptcp_out_options *opts)
629 {
630 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
631 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
632 struct tcp_sock *tp = tcp_sk(sk);
633 unsigned int dss_size = 0;
634 struct mptcp_ext *mpext;
635 unsigned int ack_size;
636 bool ret = false;
637
638 /* Zero `use_ack` and `use_map` flags with one shot. */
639 memset(&opts->ext_copy.flags, 0, sizeof(opts->ext_copy.flags));
640 opts->csum_reqd = READ_ONCE(msk->csum_enabled);
641 mpext = skb ? mptcp_get_ext(skb) : NULL;
642
643 if (!skb || (mpext && mpext->use_map) || snd_data_fin_enable) {
644 unsigned int map_size = TCPOLEN_MPTCP_DSS_BASE + TCPOLEN_MPTCP_DSS_MAP64;
645
646 if (mpext) {
647 if (opts->csum_reqd)
648 map_size += TCPOLEN_MPTCP_DSS_CHECKSUM;
649
650 opts->ext_copy = *mpext;
651 }
652
653 dss_size = map_size;
654 if (skb && snd_data_fin_enable)
655 mptcp_write_data_fin(subflow, skb, &opts->ext_copy);
656 opts->suboptions = OPTION_MPTCP_DSS;
657 ret = true;
658 }
659
660 /* passive sockets msk will set the 'can_ack' after accept(), even
661 * if the first subflow may have the already the remote key handy
662 */
663 if (!READ_ONCE(msk->can_ack)) {
664 *size = ALIGN(dss_size, 4);
665 return ret;
666 }
667
668 if (READ_ONCE(msk->use_64bit_ack)) {
669 ack_size = TCPOLEN_MPTCP_DSS_ACK64;
670 opts->ext_copy.ack64 = 1;
671 } else {
672 ack_size = TCPOLEN_MPTCP_DSS_ACK32;
673 opts->ext_copy.ack64 = 0;
674 }
675 opts->ext_copy.use_ack = 1;
676 opts->suboptions = OPTION_MPTCP_DSS;
677
678 /* Add kind/length/subtype/flag overhead if mapping is not populated */
679 if (dss_size == 0)
680 ack_size += TCPOLEN_MPTCP_DSS_BASE;
681
682 /* The caller is __tcp_transmit_skb(), and will compute the new rcv
683 * wnd soon: ensure that the window can shrink.
684 */
685 if (skb)
686 tp->rcv_wnd = tp->rcv_nxt - tp->rcv_wup;
687
688 dss_size += ack_size;
689
690 *size = ALIGN(dss_size, 4);
691 return true;
692 }
693
add_addr_generate_hmac(u64 key1,u64 key2,struct mptcp_addr_info * addr)694 static u64 add_addr_generate_hmac(u64 key1, u64 key2,
695 struct mptcp_addr_info *addr)
696 {
697 u16 port = ntohs(addr->port);
698 u8 hmac[SHA256_DIGEST_SIZE];
699 u8 msg[19];
700 int i = 0;
701
702 msg[i++] = addr->id;
703 if (addr->family == AF_INET) {
704 memcpy(&msg[i], &addr->addr.s_addr, 4);
705 i += 4;
706 }
707 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
708 else if (addr->family == AF_INET6) {
709 memcpy(&msg[i], &addr->addr6.s6_addr, 16);
710 i += 16;
711 }
712 #endif
713 msg[i++] = port >> 8;
714 msg[i++] = port & 0xFF;
715
716 mptcp_crypto_hmac_sha(key1, key2, msg, i, hmac);
717
718 return get_unaligned_be64(&hmac[SHA256_DIGEST_SIZE - sizeof(u64)]);
719 }
720
mptcp_established_options_add_addr(struct sock * sk,struct sk_buff * skb,int * size,unsigned int remaining,bool has_ts,struct mptcp_out_options * opts)721 static bool mptcp_established_options_add_addr(struct sock *sk,
722 struct sk_buff *skb, int *size,
723 unsigned int remaining,
724 bool has_ts,
725 struct mptcp_out_options *opts)
726 {
727 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
728 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
729 struct mptcp_addr_info addr;
730 bool drop_ts = has_ts;
731 bool echo;
732
733 /* add addr will strip the existing options, be sure to avoid breaking
734 * MPC/MPJ handshakes
735 */
736 if (!mptcp_pm_should_add_signal(msk) ||
737 (opts->suboptions & (OPTION_MPTCP_MPJ_ACK | OPTION_MPTCP_MPC_ACK)) ||
738 !skb || !skb_is_tcp_pure_ack(skb) ||
739 !mptcp_pm_add_addr_signal(msk, size, remaining, &addr, &echo,
740 &drop_ts))
741 return false;
742
743 pr_debug("drop other suboptions\n");
744 opts->suboptions = OPTION_MPTCP_ADD_ADDR;
745 opts->drop_ts = drop_ts;
746 opts->addr = addr;
747 if (!echo) {
748 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_ADDADDRTX);
749 opts->ahmac = add_addr_generate_hmac(READ_ONCE(msk->local_key),
750 READ_ONCE(msk->remote_key),
751 &opts->addr);
752 } else {
753 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_ECHOADDTX);
754 opts->ahmac = 0;
755 }
756 pr_debug("addr_id=%d, ahmac=%llu, echo=%d, port=%d\n",
757 opts->addr.id, opts->ahmac, echo, ntohs(opts->addr.port));
758
759 return true;
760 }
761
mptcp_established_options_rm_addr(struct sock * sk,int * size,unsigned int remaining,struct mptcp_out_options * opts)762 static bool mptcp_established_options_rm_addr(struct sock *sk, int *size,
763 unsigned int remaining,
764 struct mptcp_out_options *opts)
765 {
766 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
767 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
768 struct mptcp_rm_list rm_list;
769 int i;
770
771 if (!mptcp_pm_should_rm_signal(msk) ||
772 !(mptcp_pm_rm_addr_signal(msk, remaining, &rm_list, size)))
773 return false;
774
775 opts->suboptions |= OPTION_MPTCP_RM_ADDR;
776 opts->rm_list = rm_list;
777
778 for (i = 0; i < opts->rm_list.nr; i++)
779 pr_debug("rm_list_ids[%d]=%d\n", i, opts->rm_list.ids[i]);
780 MPTCP_ADD_STATS(sock_net(sk), MPTCP_MIB_RMADDRTX, opts->rm_list.nr);
781 return true;
782 }
783
mptcp_established_options_mp_prio(struct sock * sk,int * size,unsigned int remaining,struct mptcp_out_options * opts)784 static bool mptcp_established_options_mp_prio(struct sock *sk, int *size,
785 unsigned int remaining,
786 struct mptcp_out_options *opts)
787 {
788 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
789
790 /* can't send MP_PRIO with MPC, as they share the same option space:
791 * 'backup'. Also it makes no sense at all
792 */
793 if (!subflow->send_mp_prio || (opts->suboptions & OPTIONS_MPTCP_MPC))
794 return false;
795
796 /* account for the trailing 'nop' option */
797 if (remaining < TCPOLEN_MPTCP_PRIO_ALIGN)
798 return false;
799
800 *size = TCPOLEN_MPTCP_PRIO_ALIGN;
801 opts->suboptions |= OPTION_MPTCP_PRIO;
802 opts->backup = subflow->request_bkup;
803
804 pr_debug("prio=%d\n", opts->backup);
805
806 return true;
807 }
808
mptcp_established_options_rst(struct sock * sk,int * size,unsigned int remaining,struct mptcp_out_options * opts)809 static noinline bool mptcp_established_options_rst(struct sock *sk,
810 int *size,
811 unsigned int remaining,
812 struct mptcp_out_options *opts)
813 {
814 const struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
815
816 if (remaining < TCPOLEN_MPTCP_RST)
817 return false;
818
819 *size = TCPOLEN_MPTCP_RST;
820 opts->suboptions |= OPTION_MPTCP_RST;
821 opts->reset_transient = subflow->reset_transient;
822 opts->reset_reason = subflow->reset_reason;
823 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPRSTTX);
824
825 return true;
826 }
827
mptcp_established_options_fastclose(struct sock * sk,int * size,unsigned int remaining,struct mptcp_out_options * opts)828 static bool mptcp_established_options_fastclose(struct sock *sk, int *size,
829 unsigned int remaining,
830 struct mptcp_out_options *opts)
831 {
832 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
833 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
834
835 if (likely(!subflow->send_fastclose))
836 return false;
837
838 if (remaining < TCPOLEN_MPTCP_FASTCLOSE)
839 return false;
840
841 *size = TCPOLEN_MPTCP_FASTCLOSE;
842 opts->suboptions |= OPTION_MPTCP_FASTCLOSE;
843 opts->rcvr_key = READ_ONCE(msk->remote_key);
844
845 pr_debug("FASTCLOSE key=%llu\n", opts->rcvr_key);
846 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPFASTCLOSETX);
847 return true;
848 }
849
mptcp_established_options_mp_fail(struct sock * sk,int * size,unsigned int remaining,struct mptcp_out_options * opts)850 static bool mptcp_established_options_mp_fail(struct sock *sk, int *size,
851 unsigned int remaining,
852 struct mptcp_out_options *opts)
853 {
854 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
855
856 if (likely(!subflow->send_mp_fail))
857 return false;
858
859 if (remaining < TCPOLEN_MPTCP_FAIL)
860 return false;
861
862 *size = TCPOLEN_MPTCP_FAIL;
863 opts->suboptions |= OPTION_MPTCP_FAIL;
864 opts->fail_seq = subflow->map_seq;
865
866 pr_debug("MP_FAIL fail_seq=%llu\n", opts->fail_seq);
867 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPFAILTX);
868
869 return true;
870 }
871
mptcp_established_options(struct sock * sk,struct sk_buff * skb,unsigned int remaining,bool has_ts,struct mptcp_out_options * opts)872 int mptcp_established_options(struct sock *sk, struct sk_buff *skb,
873 unsigned int remaining, bool has_ts,
874 struct mptcp_out_options *opts)
875 {
876 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
877 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
878 int total_size = 0;
879 bool snd_data_fin;
880 bool ret = false;
881 int opt_size = 0;
882
883 opts->suboptions = 0;
884
885 /* Force later mptcp_write_options(), but do not use any actual
886 * option space.
887 */
888 if (unlikely(__mptcp_check_fallback(msk) && !mptcp_check_infinite_map(skb)))
889 return 0;
890
891 if (unlikely(skb && TCP_SKB_CB(skb)->tcp_flags & TCPHDR_RST)) {
892 if (mptcp_established_options_fastclose(sk, &opt_size, remaining, opts) ||
893 mptcp_established_options_mp_fail(sk, &opt_size, remaining, opts)) {
894 total_size += opt_size;
895 remaining -= opt_size;
896 }
897 /* MP_RST can be used with MP_FASTCLOSE and MP_FAIL if there is room */
898 if (mptcp_established_options_rst(sk, &opt_size, remaining, opts)) {
899 total_size += opt_size;
900 remaining -= opt_size;
901 }
902 return total_size;
903 }
904
905 snd_data_fin = mptcp_data_fin_enabled(msk);
906 if (mptcp_established_options_mp(sk, skb, snd_data_fin, &opt_size, opts))
907 ret = true;
908 else if (mptcp_established_options_dss(sk, skb, snd_data_fin, &opt_size, opts)) {
909 int mp_fail_size;
910
911 ret = true;
912 if (mptcp_established_options_mp_fail(sk, &mp_fail_size,
913 remaining - opt_size, opts)) {
914 total_size += opt_size + mp_fail_size;
915 remaining -= opt_size - mp_fail_size;
916 return total_size;
917 }
918 }
919
920 /* we reserved enough space for the above options, and exceeding the
921 * TCP option space would be fatal
922 */
923 if (WARN_ON_ONCE(opt_size > remaining))
924 return -1;
925
926 total_size += opt_size;
927 remaining -= opt_size;
928 if (mptcp_established_options_add_addr(sk, skb, &opt_size, remaining,
929 has_ts, opts)) {
930 total_size += opt_size;
931 remaining -= opt_size;
932 ret = true;
933 } else if (mptcp_established_options_rm_addr(sk, &opt_size, remaining, opts)) {
934 total_size += opt_size;
935 remaining -= opt_size;
936 ret = true;
937 }
938
939 if (mptcp_established_options_mp_prio(sk, &opt_size, remaining, opts)) {
940 total_size += opt_size;
941 remaining -= opt_size;
942 ret = true;
943 }
944
945 return ret ? total_size : -1;
946 }
947
mptcp_synack_options(const struct request_sock * req,unsigned int * size,struct mptcp_out_options * opts)948 bool mptcp_synack_options(const struct request_sock *req, unsigned int *size,
949 struct mptcp_out_options *opts)
950 {
951 struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req);
952
953 if (subflow_req->mp_capable) {
954 opts->suboptions = OPTION_MPTCP_MPC_SYNACK;
955 opts->sndr_key = subflow_req->local_key;
956 opts->csum_reqd = subflow_req->csum_reqd;
957 opts->allow_join_id0 = subflow_req->allow_join_id0;
958 *size = TCPOLEN_MPTCP_MPC_SYNACK;
959 pr_debug("subflow_req=%p, local_key=%llu\n",
960 subflow_req, subflow_req->local_key);
961 return true;
962 } else if (subflow_req->mp_join) {
963 opts->suboptions = OPTION_MPTCP_MPJ_SYNACK;
964 opts->backup = subflow_req->request_bkup;
965 opts->join_id = subflow_req->local_id;
966 opts->thmac = subflow_req->thmac;
967 opts->nonce = subflow_req->local_nonce;
968 pr_debug("req=%p, bkup=%u, id=%u, thmac=%llu, nonce=%u\n",
969 subflow_req, opts->backup, opts->join_id,
970 opts->thmac, opts->nonce);
971 *size = TCPOLEN_MPTCP_MPJ_SYNACK;
972 return true;
973 }
974 return false;
975 }
976
check_fully_established(struct mptcp_sock * msk,struct sock * ssk,struct mptcp_subflow_context * subflow,struct sk_buff * skb,struct mptcp_options_received * mp_opt)977 static bool check_fully_established(struct mptcp_sock *msk, struct sock *ssk,
978 struct mptcp_subflow_context *subflow,
979 struct sk_buff *skb,
980 struct mptcp_options_received *mp_opt)
981 {
982 /* here we can process OoO, in-window pkts, only in-sequence 4th ack
983 * will make the subflow fully established
984 */
985 if (likely(subflow->fully_established)) {
986 /* on passive sockets, check for 3rd ack retransmission
987 * note that msk is always set by subflow_syn_recv_sock()
988 * for mp_join subflows
989 */
990 if (TCP_SKB_CB(skb)->seq == subflow->ssn_offset + 1 &&
991 TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq &&
992 subflow->mp_join && (mp_opt->suboptions & OPTIONS_MPTCP_MPJ) &&
993 !subflow->request_join)
994 tcp_send_ack(ssk);
995 goto check_notify;
996 }
997
998 /* we must process OoO packets before the first subflow is fully
999 * established. OoO packets are instead a protocol violation
1000 * for MP_JOIN subflows as the peer must not send any data
1001 * before receiving the forth ack - cfr. RFC 8684 section 3.2.
1002 */
1003 if (TCP_SKB_CB(skb)->seq != subflow->ssn_offset + 1) {
1004 if (subflow->mp_join)
1005 goto reset;
1006 if (subflow->is_mptfo && mp_opt->suboptions & OPTION_MPTCP_MPC_ACK)
1007 goto set_fully_established;
1008 return subflow->mp_capable;
1009 }
1010
1011 if (subflow->remote_key_valid &&
1012 (((mp_opt->suboptions & OPTION_MPTCP_DSS) && mp_opt->use_ack) ||
1013 ((mp_opt->suboptions & OPTION_MPTCP_ADD_ADDR) &&
1014 (!mp_opt->echo || subflow->mp_join)))) {
1015 /* subflows are fully established as soon as we get any
1016 * additional ack, including ADD_ADDR.
1017 */
1018 goto set_fully_established;
1019 }
1020
1021 /* If the first established packet does not contain MP_CAPABLE + data
1022 * then fallback to TCP. Fallback scenarios requires a reset for
1023 * MP_JOIN subflows.
1024 */
1025 if (!(mp_opt->suboptions & OPTIONS_MPTCP_MPC)) {
1026 if (subflow->mp_join)
1027 goto reset;
1028 subflow->mp_capable = 0;
1029 if (!mptcp_try_fallback(ssk, MPTCP_MIB_MPCAPABLEDATAFALLBACK)) {
1030 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_FALLBACKFAILED);
1031 goto reset;
1032 }
1033 return false;
1034 }
1035
1036 if (unlikely(!READ_ONCE(msk->pm.server_side)))
1037 pr_warn_once("bogus mpc option on established client sk");
1038
1039 set_fully_established:
1040 if (mp_opt->deny_join_id0)
1041 WRITE_ONCE(msk->pm.remote_deny_join_id0, true);
1042
1043 mptcp_data_lock((struct sock *)msk);
1044 __mptcp_subflow_fully_established(msk, subflow, mp_opt);
1045 mptcp_data_unlock((struct sock *)msk);
1046
1047 check_notify:
1048 /* if the subflow is not already linked into the conn_list, we can't
1049 * notify the PM: this subflow is still on the listener queue
1050 * and the PM possibly acquiring the subflow lock could race with
1051 * the listener close
1052 */
1053 if (likely(subflow->pm_notified) || list_empty(&subflow->node))
1054 return true;
1055
1056 subflow->pm_notified = 1;
1057 if (subflow->mp_join) {
1058 clear_3rdack_retransmission(ssk);
1059 mptcp_pm_subflow_established(msk);
1060 } else {
1061 mptcp_pm_fully_established(msk, ssk);
1062 }
1063 return true;
1064
1065 reset:
1066 mptcp_subflow_reset(ssk);
1067 return false;
1068 }
1069
__mptcp_expand_seq(u64 old_seq,u64 cur_seq)1070 u64 __mptcp_expand_seq(u64 old_seq, u64 cur_seq)
1071 {
1072 u32 old_seq32, cur_seq32;
1073
1074 old_seq32 = (u32)old_seq;
1075 cur_seq32 = (u32)cur_seq;
1076 cur_seq = (old_seq & GENMASK_ULL(63, 32)) + cur_seq32;
1077 if (unlikely(cur_seq32 < old_seq32 && before(old_seq32, cur_seq32)))
1078 return cur_seq + (1LL << 32);
1079
1080 /* reverse wrap could happen, too */
1081 if (unlikely(cur_seq32 > old_seq32 && after(old_seq32, cur_seq32)))
1082 return cur_seq - (1LL << 32);
1083 return cur_seq;
1084 }
1085
__mptcp_snd_una_update(struct mptcp_sock * msk,u64 new_snd_una)1086 static void __mptcp_snd_una_update(struct mptcp_sock *msk, u64 new_snd_una)
1087 {
1088 msk->bytes_acked += new_snd_una - msk->snd_una;
1089 WRITE_ONCE(msk->snd_una, new_snd_una);
1090 }
1091
rwin_update(struct mptcp_sock * msk,struct sock * ssk,struct sk_buff * skb)1092 static void rwin_update(struct mptcp_sock *msk, struct sock *ssk,
1093 struct sk_buff *skb)
1094 {
1095 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
1096 struct tcp_sock *tp = tcp_sk(ssk);
1097 u64 mptcp_rcv_wnd;
1098
1099 /* Avoid touching extra cachelines if TCP is going to accept this
1100 * skb without filling the TCP-level window even with a possibly
1101 * outdated mptcp-level rwin.
1102 */
1103 if (!skb->len || skb->len < tcp_receive_window(tp))
1104 return;
1105
1106 mptcp_rcv_wnd = atomic64_read(&msk->rcv_wnd_sent);
1107 if (!after64(mptcp_rcv_wnd, subflow->rcv_wnd_sent))
1108 return;
1109
1110 /* Some other subflow grew the mptcp-level rwin since rcv_wup,
1111 * resync.
1112 */
1113 tp->rcv_wnd += mptcp_rcv_wnd - subflow->rcv_wnd_sent;
1114 tcp_update_max_rcv_wnd_seq(tp);
1115 subflow->rcv_wnd_sent = mptcp_rcv_wnd;
1116 }
1117
ack_update_msk(struct mptcp_sock * msk,struct sock * ssk,struct mptcp_options_received * mp_opt)1118 static void ack_update_msk(struct mptcp_sock *msk,
1119 struct sock *ssk,
1120 struct mptcp_options_received *mp_opt)
1121 {
1122 u64 new_wnd_end, new_snd_una, snd_nxt = READ_ONCE(msk->snd_nxt);
1123 struct sock *sk = (struct sock *)msk;
1124 u64 old_snd_una;
1125
1126 mptcp_data_lock(sk);
1127
1128 /* avoid ack expansion on update conflict, to reduce the risk of
1129 * wrongly expanding to a future ack sequence number, which is way
1130 * more dangerous than missing an ack
1131 */
1132 old_snd_una = msk->snd_una;
1133 new_snd_una = mptcp_expand_seq(old_snd_una, mp_opt->data_ack, mp_opt->ack64);
1134
1135 /* ACK for data not even sent yet? Ignore.*/
1136 if (unlikely(after64(new_snd_una, snd_nxt)))
1137 new_snd_una = old_snd_una;
1138
1139 new_wnd_end = new_snd_una + tcp_sk(ssk)->snd_wnd;
1140
1141 if (after64(new_wnd_end, msk->wnd_end))
1142 WRITE_ONCE(msk->wnd_end, new_wnd_end);
1143
1144 /* this assumes mptcp_incoming_options() is invoked after tcp_ack() */
1145 if (after64(msk->wnd_end, snd_nxt))
1146 __mptcp_check_push(sk, ssk);
1147
1148 if (after64(new_snd_una, old_snd_una)) {
1149 __mptcp_snd_una_update(msk, new_snd_una);
1150 __mptcp_data_acked(sk);
1151 }
1152 msk->last_ack_recv = tcp_jiffies32;
1153 mptcp_data_unlock(sk);
1154
1155 trace_ack_update_msk(mp_opt->data_ack,
1156 old_snd_una, new_snd_una,
1157 new_wnd_end, READ_ONCE(msk->wnd_end));
1158 }
1159
mptcp_update_rcv_data_fin(struct mptcp_sock * msk,u64 data_fin_seq,bool use_64bit)1160 bool mptcp_update_rcv_data_fin(struct mptcp_sock *msk, u64 data_fin_seq, bool use_64bit)
1161 {
1162 /* Skip if DATA_FIN was already received.
1163 * If updating simultaneously with the recvmsg loop, values
1164 * should match. If they mismatch, the peer is misbehaving and
1165 * we will prefer the most recent information.
1166 */
1167 if (READ_ONCE(msk->rcv_data_fin))
1168 return false;
1169
1170 WRITE_ONCE(msk->rcv_data_fin_seq,
1171 mptcp_expand_seq(READ_ONCE(msk->ack_seq), data_fin_seq, use_64bit));
1172 WRITE_ONCE(msk->rcv_data_fin, 1);
1173
1174 return true;
1175 }
1176
add_addr_hmac_valid(struct mptcp_sock * msk,struct mptcp_options_received * mp_opt)1177 static bool add_addr_hmac_valid(struct mptcp_sock *msk,
1178 struct mptcp_options_received *mp_opt)
1179 {
1180 u64 hmac = 0;
1181
1182 if (mp_opt->echo)
1183 return true;
1184
1185 hmac = add_addr_generate_hmac(READ_ONCE(msk->remote_key),
1186 READ_ONCE(msk->local_key),
1187 &mp_opt->addr);
1188
1189 pr_debug("msk=%p, ahmac=%llu, mp_opt->ahmac=%llu\n",
1190 msk, hmac, mp_opt->ahmac);
1191
1192 return hmac == mp_opt->ahmac;
1193 }
1194
mptcp_over_limit(struct sock * sk,struct sock * ssk,const struct sk_buff * skb)1195 static bool mptcp_over_limit(struct sock *sk, struct sock *ssk,
1196 const struct sk_buff *skb)
1197 {
1198 struct mptcp_sock *msk = mptcp_sk(sk);
1199 u32 rcvbuf = READ_ONCE(sk->sk_rcvbuf);
1200
1201 if (likely((u32)sk_rmem_alloc_get(sk) <= rcvbuf &&
1202 READ_ONCE(msk->backlog_len) <= rcvbuf))
1203 return false;
1204
1205 /* Avoid silently dropping pure acks, fin, rst or already-acked segm. */
1206 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq ||
1207 TCP_SKB_CB(skb)->tcp_flags & (TCPHDR_FIN | TCPHDR_RST) ||
1208 !after(TCP_SKB_CB(skb)->end_seq, tcp_sk(ssk)->rcv_nxt))
1209 return false;
1210
1211 /* Dropped due to memory constraints, schedule an ack. */
1212 inet_csk(ssk)->icsk_ack.pending |= ICSK_ACK_NOMEM | ICSK_ACK_NOW;
1213 inet_csk_schedule_ack(ssk);
1214
1215 /* Plain TCP (fallback) and skb is dropped before the TCP recv queue. */
1216 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRCVQDROP);
1217
1218 return true;
1219 }
1220
1221 /* Return false when the caller must drop the packet, i.e. in case of error,
1222 * subflow has been reset, or over memory limits.
1223 */
mptcp_incoming_options(struct sock * sk,struct sk_buff * skb)1224 bool mptcp_incoming_options(struct sock *sk, struct sk_buff *skb)
1225 {
1226 struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk);
1227 struct mptcp_sock *msk = mptcp_sk(subflow->conn);
1228 struct mptcp_options_received mp_opt;
1229 struct mptcp_ext *mpext;
1230
1231 if (__mptcp_check_fallback(msk)) {
1232 /* Keep it simple and unconditionally trigger send data cleanup and
1233 * pending queue spooling. We will need to acquire the data lock
1234 * for more accurate checks, and once the lock is acquired, such
1235 * helpers are cheap.
1236 */
1237 mptcp_data_lock(subflow->conn);
1238 if (sk_stream_memory_free(sk))
1239 __mptcp_check_push(subflow->conn, sk);
1240
1241 /* on fallback we just need to ignore the msk-level snd_una, as
1242 * this is really plain TCP
1243 */
1244 __mptcp_snd_una_update(msk, READ_ONCE(msk->snd_nxt));
1245
1246 __mptcp_data_acked(subflow->conn);
1247 mptcp_data_unlock(subflow->conn);
1248 return !mptcp_over_limit(subflow->conn, sk, skb);
1249 }
1250
1251 mptcp_get_options(skb, &mp_opt);
1252
1253 /* The subflow can be in close state only if check_fully_established()
1254 * just sent a reset. If so, tell the caller to ignore the current packet.
1255 */
1256 if (!check_fully_established(msk, sk, subflow, skb, &mp_opt))
1257 return sk->sk_state != TCP_CLOSE;
1258
1259 if (unlikely(mp_opt.suboptions != OPTION_MPTCP_DSS)) {
1260 if ((mp_opt.suboptions & OPTION_MPTCP_FASTCLOSE) &&
1261 READ_ONCE(msk->local_key) == mp_opt.rcvr_key) {
1262 WRITE_ONCE(msk->rcv_fastclose, true);
1263 mptcp_schedule_work((struct sock *)msk);
1264 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPFASTCLOSERX);
1265 }
1266
1267 if ((mp_opt.suboptions & OPTION_MPTCP_ADD_ADDR) &&
1268 add_addr_hmac_valid(msk, &mp_opt)) {
1269 if (!mp_opt.echo) {
1270 mptcp_pm_add_addr_received(sk, &mp_opt.addr);
1271 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_ADDADDR);
1272 } else {
1273 mptcp_pm_add_addr_echoed(msk, &mp_opt.addr);
1274 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_ECHOADD);
1275 }
1276
1277 if (mp_opt.addr.port)
1278 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_PORTADD);
1279 }
1280
1281 if (mp_opt.suboptions & OPTION_MPTCP_RM_ADDR)
1282 mptcp_pm_rm_addr_received(msk, &mp_opt.rm_list);
1283
1284 if (mp_opt.suboptions & OPTION_MPTCP_PRIO) {
1285 mptcp_pm_mp_prio_received(sk, mp_opt.backup);
1286 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPPRIORX);
1287 }
1288
1289 if (mp_opt.suboptions & OPTION_MPTCP_FAIL) {
1290 mptcp_pm_mp_fail_received(sk, mp_opt.fail_seq);
1291 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPFAILRX);
1292 }
1293
1294 if (mp_opt.suboptions & OPTION_MPTCP_RST) {
1295 subflow->reset_seen = 1;
1296 subflow->reset_reason = mp_opt.reset_reason;
1297 subflow->reset_transient = mp_opt.reset_transient;
1298 MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPRSTRX);
1299 }
1300
1301 if (!(mp_opt.suboptions & OPTION_MPTCP_DSS))
1302 return true;
1303 }
1304
1305 /* we can't wait for recvmsg() to update the ack_seq, otherwise
1306 * monodirectional flows will stuck
1307 */
1308 if (mp_opt.use_ack)
1309 ack_update_msk(msk, sk, &mp_opt);
1310 rwin_update(msk, sk, skb);
1311
1312 /* Zero-data-length packets are dropped by the caller and not
1313 * propagated to the MPTCP layer, so the skb extension does not
1314 * need to be allocated or populated. DATA_FIN information, if
1315 * present, needs to be updated here before the skb is freed.
1316 */
1317 if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) {
1318 if (mp_opt.data_fin && mp_opt.data_len == 1 &&
1319 mptcp_update_rcv_data_fin(msk, mp_opt.data_seq, mp_opt.dsn64))
1320 mptcp_schedule_work((struct sock *)msk);
1321
1322 return true;
1323 }
1324
1325 mpext = skb_ext_add(skb, SKB_EXT_MPTCP);
1326 if (!mpext)
1327 return false;
1328
1329 memset(mpext, 0, sizeof(*mpext));
1330
1331 if (likely(mp_opt.use_map)) {
1332 if (mp_opt.mpc_map) {
1333 /* this is an MP_CAPABLE carrying MPTCP data
1334 * we know this map the first chunk of data
1335 */
1336 mptcp_crypto_key_sha(subflow->remote_key, NULL,
1337 &mpext->data_seq);
1338 mpext->data_seq++;
1339 mpext->subflow_seq = 1;
1340 mpext->dsn64 = 1;
1341 mpext->mpc_map = 1;
1342 mpext->data_fin = 0;
1343 } else {
1344 mpext->data_seq = mp_opt.data_seq;
1345 mpext->subflow_seq = mp_opt.subflow_seq;
1346 mpext->dsn64 = mp_opt.dsn64;
1347 mpext->data_fin = mp_opt.data_fin;
1348 }
1349 mpext->data_len = mp_opt.data_len;
1350 mpext->use_map = 1;
1351 mpext->csum_reqd = !!(mp_opt.suboptions & OPTION_MPTCP_CSUMREQD);
1352
1353 if (mpext->csum_reqd)
1354 mpext->csum = mp_opt.csum;
1355 }
1356
1357 return true;
1358 }
1359
mptcp_set_rwin(struct mptcp_sock * msk,struct tcp_sock * tp,struct tcphdr * th,u64 ack_seq)1360 static u64 mptcp_set_rwin(struct mptcp_sock *msk, struct tcp_sock *tp,
1361 struct tcphdr *th, u64 ack_seq)
1362 {
1363 const struct sock *ssk = (const struct sock *)tp;
1364 u64 rcv_wnd_old, rcv_wnd_new;
1365 u32 new_win;
1366 u64 win;
1367
1368 rcv_wnd_new = ack_seq + tp->rcv_wnd;
1369
1370 rcv_wnd_old = atomic64_read(&msk->rcv_wnd_sent);
1371 if (after64(rcv_wnd_new, rcv_wnd_old)) {
1372 u64 rcv_wnd;
1373
1374 for (;;) {
1375 rcv_wnd = atomic64_cmpxchg(&msk->rcv_wnd_sent, rcv_wnd_old, rcv_wnd_new);
1376
1377 if (rcv_wnd == rcv_wnd_old)
1378 break;
1379
1380 rcv_wnd_old = rcv_wnd;
1381 if (before64(rcv_wnd_new, rcv_wnd_old)) {
1382 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_RCVWNDCONFLICTUPDATE);
1383 goto raise_win;
1384 }
1385 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_RCVWNDCONFLICT);
1386 }
1387 goto update_wspace;
1388 }
1389
1390 if (rcv_wnd_new != rcv_wnd_old) {
1391 raise_win:
1392 /* The msk-level rcv wnd is after the tcp level one,
1393 * sync the latter.
1394 */
1395 rcv_wnd_new = rcv_wnd_old;
1396 win = rcv_wnd_old - ack_seq;
1397 new_win = min_t(u64, win, U32_MAX);
1398 tp->rcv_wnd = new_win;
1399 tcp_update_max_rcv_wnd_seq(tp);
1400
1401 /* Make sure we do not exceed the maximum possible
1402 * scaled window.
1403 */
1404 if (unlikely(th->syn))
1405 new_win = min(new_win, 65535U) << tp->rx_opt.rcv_wscale;
1406 if (!tp->rx_opt.rcv_wscale &&
1407 READ_ONCE(sock_net(ssk)->ipv4.sysctl_tcp_workaround_signed_windows))
1408 new_win = min(new_win, MAX_TCP_WINDOW);
1409 else
1410 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
1411
1412 /* RFC1323 scaling applied */
1413 new_win >>= tp->rx_opt.rcv_wscale;
1414 th->window = htons(new_win);
1415 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_RCVWNDSHARED);
1416 }
1417
1418 update_wspace:
1419 WRITE_ONCE(msk->old_wspace, tp->rcv_wnd);
1420 return rcv_wnd_new;
1421 }
1422
mptcp_track_rwin(struct tcp_sock * tp)1423 static void mptcp_track_rwin(struct tcp_sock *tp)
1424 {
1425 const struct sock *ssk = (const struct sock *)tp;
1426 struct mptcp_subflow_context *subflow;
1427 struct mptcp_sock *msk;
1428
1429 if (!ssk)
1430 return;
1431
1432 subflow = mptcp_subflow_ctx(ssk);
1433 msk = mptcp_sk(subflow->conn);
1434 WRITE_ONCE(msk->old_wspace, tp->rcv_wnd);
1435 }
1436
__mptcp_make_csum(u64 data_seq,u32 subflow_seq,u16 data_len,__wsum sum)1437 __sum16 __mptcp_make_csum(u64 data_seq, u32 subflow_seq, u16 data_len, __wsum sum)
1438 {
1439 struct csum_pseudo_header header;
1440 __wsum csum;
1441
1442 /* cfr RFC 8684 3.3.1.:
1443 * the data sequence number used in the pseudo-header is
1444 * always the 64-bit value, irrespective of what length is used in the
1445 * DSS option itself.
1446 */
1447 header.data_seq = cpu_to_be64(data_seq);
1448 header.subflow_seq = htonl(subflow_seq);
1449 header.data_len = htons(data_len);
1450 header.csum = 0;
1451
1452 csum = csum_partial(&header, sizeof(header), sum);
1453 return csum_fold(csum);
1454 }
1455
mptcp_make_csum(const struct mptcp_ext * mpext)1456 static __sum16 mptcp_make_csum(const struct mptcp_ext *mpext)
1457 {
1458 return __mptcp_make_csum(mpext->data_seq, mpext->subflow_seq, mpext->data_len,
1459 ~csum_unfold(mpext->csum));
1460 }
1461
put_len_csum(u16 len,__sum16 csum,void * data)1462 static void put_len_csum(u16 len, __sum16 csum, void *data)
1463 {
1464 __sum16 *sumptr = data + 2;
1465 __be16 *ptr = data;
1466
1467 put_unaligned_be16(len, ptr);
1468
1469 put_unaligned(csum, sumptr);
1470 }
1471
mptcp_write_options(struct tcphdr * th,__be32 * ptr,struct tcp_sock * tp,struct mptcp_out_options * opts)1472 void mptcp_write_options(struct tcphdr *th, __be32 *ptr, struct tcp_sock *tp,
1473 struct mptcp_out_options *opts)
1474 {
1475 const struct sock *ssk = (const struct sock *)tp;
1476 struct mptcp_subflow_context *subflow;
1477
1478 /* Which options can be used together?
1479 *
1480 * X: mutually exclusive
1481 * O: often used together
1482 * C: can be used together in some cases
1483 * P: could be used together but we prefer not to (optimisations)
1484 *
1485 * Opt: | MPC | MPJ | DSS | ADD | RM | PRIO | FAIL | FC |
1486 * ------|------|------|------|------|------|------|------|------|
1487 * MPC |------|------|------|------|------|------|------|------|
1488 * MPJ | X |------|------|------|------|------|------|------|
1489 * DSS | X | X |------|------|------|------|------|------|
1490 * ADD | X | X | P |------|------|------|------|------|
1491 * RM | C | C | C | P |------|------|------|------|
1492 * PRIO | X | C | C | C | C |------|------|------|
1493 * FAIL | X | X | C | X | X | X |------|------|
1494 * FC | X | X | P | X | X | X | X |------|
1495 * RST | X | X | X | X | X | X | O | O |
1496 * ------|------|------|------|------|------|------|------|------|
1497 *
1498 * The same applies in mptcp_established_options() function.
1499 */
1500 if (likely(OPTION_MPTCP_DSS & opts->suboptions)) {
1501 struct mptcp_ext *mpext = &opts->ext_copy;
1502 u8 len = TCPOLEN_MPTCP_DSS_BASE;
1503 u8 flags = 0;
1504
1505 if (mpext->use_ack) {
1506 flags = MPTCP_DSS_HAS_ACK;
1507 if (mpext->ack64) {
1508 len += TCPOLEN_MPTCP_DSS_ACK64;
1509 flags |= MPTCP_DSS_ACK64;
1510 } else {
1511 len += TCPOLEN_MPTCP_DSS_ACK32;
1512 }
1513 }
1514
1515 if (mpext->use_map) {
1516 len += TCPOLEN_MPTCP_DSS_MAP64;
1517
1518 /* Use only 64-bit mapping flags for now, add
1519 * support for optional 32-bit mappings later.
1520 */
1521 flags |= MPTCP_DSS_HAS_MAP | MPTCP_DSS_DSN64;
1522 if (mpext->data_fin)
1523 flags |= MPTCP_DSS_DATA_FIN;
1524
1525 if (opts->csum_reqd)
1526 len += TCPOLEN_MPTCP_DSS_CHECKSUM;
1527 }
1528
1529 *ptr++ = mptcp_option(MPTCPOPT_DSS, len, 0, flags);
1530
1531 if (mpext->use_ack) {
1532 struct mptcp_sock *msk;
1533 u64 ack_seq;
1534
1535 /* DSS option is set only by mptcp_established_options,
1536 * the caller is __tcp_transmit_skb() and ssk is always
1537 * not NULL.
1538 */
1539 subflow = mptcp_subflow_ctx(ssk);
1540 msk = mptcp_sk(subflow->conn);
1541 ack_seq = READ_ONCE(msk->ack_seq);
1542 if (mpext->ack64) {
1543 put_unaligned_be64(ack_seq, ptr);
1544 ptr += 2;
1545 } else {
1546 put_unaligned_be32(ack_seq, ptr);
1547 ptr += 1;
1548 }
1549 subflow->rcv_wnd_sent = mptcp_set_rwin(msk, tp, th,
1550 ack_seq);
1551 }
1552
1553 if (mpext->use_map) {
1554 put_unaligned_be64(mpext->data_seq, ptr);
1555 ptr += 2;
1556 put_unaligned_be32(mpext->subflow_seq, ptr);
1557 ptr += 1;
1558 if (opts->csum_reqd) {
1559 /* data_len == 0 is reserved for the infinite mapping,
1560 * the checksum will also be set to 0.
1561 */
1562 put_len_csum(mpext->data_len,
1563 (mpext->data_len ? mptcp_make_csum(mpext) : 0),
1564 ptr);
1565 } else {
1566 put_unaligned_be32(mpext->data_len << 16 |
1567 TCPOPT_NOP << 8 | TCPOPT_NOP, ptr);
1568 }
1569 ptr += 1;
1570 }
1571
1572 /* We might need to add MP_FAIL options in rare cases */
1573 if (unlikely(OPTION_MPTCP_FAIL & opts->suboptions))
1574 goto mp_fail;
1575 } else if (OPTIONS_MPTCP_MPC & opts->suboptions) {
1576 u8 len, flag = MPTCP_CAP_HMAC_SHA256;
1577
1578 if (OPTION_MPTCP_MPC_SYN & opts->suboptions) {
1579 len = TCPOLEN_MPTCP_MPC_SYN;
1580 } else if (OPTION_MPTCP_MPC_SYNACK & opts->suboptions) {
1581 len = TCPOLEN_MPTCP_MPC_SYNACK;
1582 } else if (opts->data_len) {
1583 len = TCPOLEN_MPTCP_MPC_ACK_DATA;
1584 if (opts->csum_reqd)
1585 len += TCPOLEN_MPTCP_DSS_CHECKSUM;
1586 } else {
1587 len = TCPOLEN_MPTCP_MPC_ACK;
1588 }
1589
1590 if (opts->csum_reqd)
1591 flag |= MPTCP_CAP_CHECKSUM_REQD;
1592
1593 if (!opts->allow_join_id0)
1594 flag |= MPTCP_CAP_DENY_JOIN_ID0;
1595
1596 *ptr++ = mptcp_option(MPTCPOPT_MP_CAPABLE, len,
1597 MPTCP_SUPPORTED_VERSION,
1598 flag);
1599
1600 if (!((OPTION_MPTCP_MPC_SYNACK | OPTION_MPTCP_MPC_ACK) &
1601 opts->suboptions))
1602 goto mp_capable_done;
1603
1604 put_unaligned_be64(opts->sndr_key, ptr);
1605 ptr += 2;
1606 if (!((OPTION_MPTCP_MPC_ACK) & opts->suboptions))
1607 goto mp_capable_done;
1608
1609 put_unaligned_be64(opts->rcvr_key, ptr);
1610 ptr += 2;
1611 if (!opts->data_len)
1612 goto mp_capable_done;
1613
1614 if (opts->csum_reqd) {
1615 put_len_csum(opts->data_len,
1616 __mptcp_make_csum(opts->data_seq,
1617 opts->subflow_seq,
1618 opts->data_len,
1619 ~csum_unfold(opts->csum)),
1620 ptr);
1621 } else {
1622 put_unaligned_be32(opts->data_len << 16 |
1623 TCPOPT_NOP << 8 | TCPOPT_NOP, ptr);
1624 }
1625 ptr += 1;
1626
1627 /* MPC is additionally mutually exclusive with MP_PRIO */
1628 goto mp_capable_done;
1629 } else if (OPTIONS_MPTCP_MPJ & opts->suboptions) {
1630 if (OPTION_MPTCP_MPJ_SYN & opts->suboptions) {
1631 *ptr++ = mptcp_option(MPTCPOPT_MP_JOIN,
1632 TCPOLEN_MPTCP_MPJ_SYN,
1633 opts->backup, opts->join_id);
1634 put_unaligned_be32(opts->token, ptr);
1635 ptr += 1;
1636 put_unaligned_be32(opts->nonce, ptr);
1637 ptr += 1;
1638 } else if (OPTION_MPTCP_MPJ_SYNACK & opts->suboptions) {
1639 *ptr++ = mptcp_option(MPTCPOPT_MP_JOIN,
1640 TCPOLEN_MPTCP_MPJ_SYNACK,
1641 opts->backup, opts->join_id);
1642 put_unaligned_be64(opts->thmac, ptr);
1643 ptr += 2;
1644 put_unaligned_be32(opts->nonce, ptr);
1645 ptr += 1;
1646 } else {
1647 *ptr++ = mptcp_option(MPTCPOPT_MP_JOIN,
1648 TCPOLEN_MPTCP_MPJ_ACK, 0, 0);
1649 memcpy(ptr, opts->hmac, MPTCPOPT_HMAC_LEN);
1650 ptr += 5;
1651 }
1652 } else if (OPTION_MPTCP_ADD_ADDR & opts->suboptions) {
1653 u8 len = TCPOLEN_MPTCP_ADD_ADDR_BASE;
1654 u8 echo = MPTCP_ADDR_ECHO;
1655
1656 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1657 if (opts->addr.family == AF_INET6)
1658 len = TCPOLEN_MPTCP_ADD_ADDR6_BASE;
1659 #endif
1660
1661 if (opts->addr.port)
1662 len += TCPOLEN_MPTCP_PORT_LEN;
1663
1664 if (opts->ahmac) {
1665 len += sizeof(opts->ahmac);
1666 echo = 0;
1667 }
1668
1669 *ptr++ = mptcp_option(MPTCPOPT_ADD_ADDR,
1670 len, echo, opts->addr.id);
1671 if (opts->addr.family == AF_INET) {
1672 memcpy((u8 *)ptr, (u8 *)&opts->addr.addr.s_addr, 4);
1673 ptr += 1;
1674 }
1675 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
1676 else if (opts->addr.family == AF_INET6) {
1677 memcpy((u8 *)ptr, opts->addr.addr6.s6_addr, 16);
1678 ptr += 4;
1679 }
1680 #endif
1681
1682 if (!opts->addr.port) {
1683 if (opts->ahmac) {
1684 put_unaligned_be64(opts->ahmac, ptr);
1685 ptr += 2;
1686 }
1687 } else {
1688 u16 port = ntohs(opts->addr.port);
1689
1690 if (opts->ahmac) {
1691 u8 *bptr = (u8 *)ptr;
1692
1693 put_unaligned_be16(port, bptr);
1694 bptr += 2;
1695 put_unaligned_be64(opts->ahmac, bptr);
1696 bptr += 8;
1697 put_unaligned_be16(TCPOPT_NOP << 8 |
1698 TCPOPT_NOP, bptr);
1699
1700 ptr += 3;
1701 } else {
1702 put_unaligned_be32(port << 16 |
1703 TCPOPT_NOP << 8 |
1704 TCPOPT_NOP, ptr);
1705 ptr += 1;
1706 }
1707 }
1708 } else if (unlikely(OPTION_MPTCP_FASTCLOSE & opts->suboptions)) {
1709 /* FASTCLOSE is mutually exclusive with others except RST */
1710 *ptr++ = mptcp_option(MPTCPOPT_MP_FASTCLOSE,
1711 TCPOLEN_MPTCP_FASTCLOSE,
1712 0, 0);
1713 put_unaligned_be64(opts->rcvr_key, ptr);
1714 ptr += 2;
1715
1716 if (OPTION_MPTCP_RST & opts->suboptions)
1717 goto mp_rst;
1718 return;
1719 } else if (unlikely(OPTION_MPTCP_FAIL & opts->suboptions)) {
1720 mp_fail:
1721 /* MP_FAIL is mutually exclusive with others except RST */
1722 subflow = mptcp_subflow_ctx(ssk);
1723 subflow->send_mp_fail = 0;
1724
1725 *ptr++ = mptcp_option(MPTCPOPT_MP_FAIL,
1726 TCPOLEN_MPTCP_FAIL,
1727 0, 0);
1728 put_unaligned_be64(opts->fail_seq, ptr);
1729 ptr += 2;
1730
1731 if (OPTION_MPTCP_RST & opts->suboptions)
1732 goto mp_rst;
1733 return;
1734 } else if (unlikely(OPTION_MPTCP_RST & opts->suboptions)) {
1735 mp_rst:
1736 *ptr++ = mptcp_option(MPTCPOPT_RST,
1737 TCPOLEN_MPTCP_RST,
1738 opts->reset_transient,
1739 opts->reset_reason);
1740 return;
1741 } else if (unlikely(!opts->suboptions)) {
1742 /* Fallback to TCP */
1743 mptcp_track_rwin(tp);
1744 return;
1745 }
1746
1747 if (OPTION_MPTCP_PRIO & opts->suboptions) {
1748 subflow = mptcp_subflow_ctx(ssk);
1749 subflow->send_mp_prio = 0;
1750
1751 *ptr++ = mptcp_option(MPTCPOPT_MP_PRIO,
1752 TCPOLEN_MPTCP_PRIO,
1753 opts->backup, TCPOPT_NOP);
1754
1755 MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPPRIOTX);
1756 }
1757
1758 mp_capable_done:
1759 if (OPTION_MPTCP_RM_ADDR & opts->suboptions) {
1760 u8 i = 1;
1761
1762 *ptr++ = mptcp_option(MPTCPOPT_RM_ADDR,
1763 TCPOLEN_MPTCP_RM_ADDR_BASE + opts->rm_list.nr,
1764 0, opts->rm_list.ids[0]);
1765
1766 while (i < opts->rm_list.nr) {
1767 u8 id1, id2, id3, id4;
1768
1769 id1 = opts->rm_list.ids[i];
1770 id2 = i + 1 < opts->rm_list.nr ? opts->rm_list.ids[i + 1] : TCPOPT_NOP;
1771 id3 = i + 2 < opts->rm_list.nr ? opts->rm_list.ids[i + 2] : TCPOPT_NOP;
1772 id4 = i + 3 < opts->rm_list.nr ? opts->rm_list.ids[i + 3] : TCPOPT_NOP;
1773 put_unaligned_be32(id1 << 24 | id2 << 16 | id3 << 8 | id4, ptr);
1774 ptr += 1;
1775 i += 4;
1776 }
1777 }
1778 }
1779
mptcp_get_reset_option(const struct sk_buff * skb)1780 __be32 mptcp_get_reset_option(const struct sk_buff *skb)
1781 {
1782 const struct mptcp_ext *ext = mptcp_get_ext(skb);
1783 u8 flags, reason;
1784
1785 if (ext) {
1786 flags = ext->reset_transient;
1787 reason = ext->reset_reason;
1788
1789 return mptcp_option(MPTCPOPT_RST, TCPOLEN_MPTCP_RST,
1790 flags, reason);
1791 }
1792
1793 return htonl(0u);
1794 }
1795 EXPORT_SYMBOL_GPL(mptcp_get_reset_option);
1796