Commit | Line | Data |
---|---|---|
eda7acdd PK |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* Multipath TCP | |
3 | * | |
4 | * Copyright (c) 2017 - 2019, Intel Corporation. | |
5 | */ | |
6 | ||
c85adced GT |
7 | #define pr_fmt(fmt) "MPTCP: " fmt |
8 | ||
eda7acdd | 9 | #include <linux/kernel.h> |
a24d22b2 | 10 | #include <crypto/sha2.h> |
eda7acdd PK |
11 | #include <net/tcp.h> |
12 | #include <net/mptcp.h> | |
13 | #include "protocol.h" | |
a877de06 | 14 | #include "mib.h" |
eda7acdd | 15 | |
ed66bfb4 GT |
16 | #include <trace/events/mptcp.h> |
17 | ||
65492c5a PA |
18 | static bool mptcp_cap_flag_sha256(u8 flags) |
19 | { | |
20 | return (flags & MPTCP_CAP_FLAG_MASK) == MPTCP_CAP_HMAC_SHA256; | |
21 | } | |
22 | ||
cfde141e PA |
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) | |
eda7acdd | 26 | { |
eda7acdd | 27 | u8 subtype = *ptr >> 4; |
648ef4b8 | 28 | int expected_opsize; |
fe33d386 | 29 | u16 subopt; |
eda7acdd PK |
30 | u8 version; |
31 | u8 flags; | |
5c4a824d | 32 | u8 i; |
eda7acdd PK |
33 | |
34 | switch (subtype) { | |
35 | case MPTCPOPT_MP_CAPABLE: | |
cc7972ea CP |
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; | |
fe33d386 | 42 | subopt = OPTION_MPTCP_MPC_ACK; |
cc7972ea | 43 | } else { |
fe33d386 | 44 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK) { |
cc7972ea | 45 | expected_opsize = TCPOLEN_MPTCP_MPC_SYNACK; |
fe33d386 PA |
46 | subopt = OPTION_MPTCP_MPC_SYNACK; |
47 | } else { | |
cc7972ea | 48 | expected_opsize = TCPOLEN_MPTCP_MPC_SYN; |
fe33d386 PA |
49 | subopt = OPTION_MPTCP_MPC_SYN; |
50 | } | |
cc7972ea | 51 | } |
208e8f66 GT |
52 | |
53 | /* Cfr RFC 8684 Section 3.3.0: | |
54 | * If a checksum is present but its use had | |
55 | * not been negotiated in the MP_CAPABLE handshake, the receiver MUST | |
56 | * close the subflow with a RST, as it is not behaving as negotiated. | |
57 | * If a checksum is not present when its use has been negotiated, the | |
58 | * receiver MUST close the subflow with a RST, as it is considered | |
59 | * broken | |
60 | * We parse even option with mismatching csum presence, so that | |
61 | * later in subflow_data_ready we can trigger the reset. | |
62 | */ | |
63 | if (opsize != expected_opsize && | |
64 | (expected_opsize != TCPOLEN_MPTCP_MPC_ACK_DATA || | |
65 | opsize != TCPOLEN_MPTCP_MPC_ACK_DATA_CSUM)) | |
eda7acdd PK |
66 | break; |
67 | ||
cc7972ea | 68 | /* try to be gentle vs future versions on the initial syn */ |
eda7acdd | 69 | version = *ptr++ & MPTCP_VERSION_MASK; |
cc7972ea CP |
70 | if (opsize != TCPOLEN_MPTCP_MPC_SYN) { |
71 | if (version != MPTCP_SUPPORTED_VERSION) | |
72 | break; | |
73 | } else if (version < MPTCP_SUPPORTED_VERSION) { | |
eda7acdd | 74 | break; |
cc7972ea | 75 | } |
eda7acdd PK |
76 | |
77 | flags = *ptr++; | |
65492c5a | 78 | if (!mptcp_cap_flag_sha256(flags) || |
eda7acdd PK |
79 | (flags & MPTCP_CAP_EXTENSIBILITY)) |
80 | break; | |
81 | ||
82 | /* RFC 6824, Section 3.1: | |
83 | * "For the Checksum Required bit (labeled "A"), if either | |
84 | * host requires the use of checksums, checksums MUST be used. | |
85 | * In other words, the only way for checksums not to be used | |
86 | * is if both hosts in their SYNs set A=0." | |
eda7acdd PK |
87 | */ |
88 | if (flags & MPTCP_CAP_CHECKSUM_REQD) | |
74c7dfbe | 89 | mp_opt->suboptions |= OPTION_MPTCP_CSUMREQD; |
eda7acdd | 90 | |
a086aeba | 91 | mp_opt->deny_join_id0 = !!(flags & MPTCP_CAP_DENY_JOIN_ID0); |
df377be3 | 92 | |
fe33d386 | 93 | mp_opt->suboptions |= subopt; |
cc7972ea CP |
94 | if (opsize >= TCPOLEN_MPTCP_MPC_SYNACK) { |
95 | mp_opt->sndr_key = get_unaligned_be64(ptr); | |
96 | ptr += 8; | |
97 | } | |
98 | if (opsize >= TCPOLEN_MPTCP_MPC_ACK) { | |
eda7acdd PK |
99 | mp_opt->rcvr_key = get_unaligned_be64(ptr); |
100 | ptr += 8; | |
eda7acdd | 101 | } |
208e8f66 | 102 | if (opsize >= TCPOLEN_MPTCP_MPC_ACK_DATA) { |
cc7972ea CP |
103 | /* Section 3.1.: |
104 | * "the data parameters in a MP_CAPABLE are semantically | |
105 | * equivalent to those in a DSS option and can be used | |
106 | * interchangeably." | |
107 | */ | |
74c7dfbe | 108 | mp_opt->suboptions |= OPTION_MPTCP_DSS; |
cc7972ea CP |
109 | mp_opt->use_map = 1; |
110 | mp_opt->mpc_map = 1; | |
237ff253 | 111 | mp_opt->use_ack = 0; |
cc7972ea CP |
112 | mp_opt->data_len = get_unaligned_be16(ptr); |
113 | ptr += 2; | |
114 | } | |
208e8f66 | 115 | if (opsize == TCPOLEN_MPTCP_MPC_ACK_DATA_CSUM) { |
ba2c89e0 | 116 | mp_opt->csum = get_unaligned((__force __sum16 *)ptr); |
74c7dfbe | 117 | mp_opt->suboptions |= OPTION_MPTCP_CSUMREQD; |
208e8f66 GT |
118 | ptr += 2; |
119 | } | |
120 | pr_debug("MP_CAPABLE version=%x, flags=%x, optlen=%d sndr=%llu, rcvr=%llu len=%d csum=%u", | |
cc7972ea | 121 | version, flags, opsize, mp_opt->sndr_key, |
208e8f66 | 122 | mp_opt->rcvr_key, mp_opt->data_len, mp_opt->csum); |
eda7acdd PK |
123 | break; |
124 | ||
f296234c | 125 | case MPTCPOPT_MP_JOIN: |
f296234c | 126 | if (opsize == TCPOLEN_MPTCP_MPJ_SYN) { |
89e23277 | 127 | mp_opt->suboptions |= OPTION_MPTCP_MPJ_SYN; |
f296234c PK |
128 | mp_opt->backup = *ptr++ & MPTCPOPT_BACKUP; |
129 | mp_opt->join_id = *ptr++; | |
130 | mp_opt->token = get_unaligned_be32(ptr); | |
131 | ptr += 4; | |
132 | mp_opt->nonce = get_unaligned_be32(ptr); | |
133 | ptr += 4; | |
134 | pr_debug("MP_JOIN bkup=%u, id=%u, token=%u, nonce=%u", | |
135 | mp_opt->backup, mp_opt->join_id, | |
136 | mp_opt->token, mp_opt->nonce); | |
137 | } else if (opsize == TCPOLEN_MPTCP_MPJ_SYNACK) { | |
89e23277 | 138 | mp_opt->suboptions |= OPTION_MPTCP_MPJ_SYNACK; |
f296234c PK |
139 | mp_opt->backup = *ptr++ & MPTCPOPT_BACKUP; |
140 | mp_opt->join_id = *ptr++; | |
141 | mp_opt->thmac = get_unaligned_be64(ptr); | |
142 | ptr += 8; | |
143 | mp_opt->nonce = get_unaligned_be32(ptr); | |
144 | ptr += 4; | |
145 | pr_debug("MP_JOIN bkup=%u, id=%u, thmac=%llu, nonce=%u", | |
146 | mp_opt->backup, mp_opt->join_id, | |
147 | mp_opt->thmac, mp_opt->nonce); | |
148 | } else if (opsize == TCPOLEN_MPTCP_MPJ_ACK) { | |
89e23277 | 149 | mp_opt->suboptions |= OPTION_MPTCP_MPJ_ACK; |
f296234c PK |
150 | ptr += 2; |
151 | memcpy(mp_opt->hmac, ptr, MPTCPOPT_HMAC_LEN); | |
152 | pr_debug("MP_JOIN hmac"); | |
f296234c PK |
153 | } |
154 | break; | |
155 | ||
eda7acdd PK |
156 | case MPTCPOPT_DSS: |
157 | pr_debug("DSS"); | |
648ef4b8 MM |
158 | ptr++; |
159 | ||
cc7972ea CP |
160 | /* we must clear 'mpc_map' be able to detect MP_CAPABLE |
161 | * map vs DSS map in mptcp_incoming_options(), and reconstruct | |
162 | * map info accordingly | |
163 | */ | |
164 | mp_opt->mpc_map = 0; | |
648ef4b8 MM |
165 | flags = (*ptr++) & MPTCP_DSS_FLAG_MASK; |
166 | mp_opt->data_fin = (flags & MPTCP_DSS_DATA_FIN) != 0; | |
167 | mp_opt->dsn64 = (flags & MPTCP_DSS_DSN64) != 0; | |
168 | mp_opt->use_map = (flags & MPTCP_DSS_HAS_MAP) != 0; | |
169 | mp_opt->ack64 = (flags & MPTCP_DSS_ACK64) != 0; | |
170 | mp_opt->use_ack = (flags & MPTCP_DSS_HAS_ACK); | |
171 | ||
172 | pr_debug("data_fin=%d dsn64=%d use_map=%d ack64=%d use_ack=%d", | |
173 | mp_opt->data_fin, mp_opt->dsn64, | |
174 | mp_opt->use_map, mp_opt->ack64, | |
175 | mp_opt->use_ack); | |
176 | ||
177 | expected_opsize = TCPOLEN_MPTCP_DSS_BASE; | |
178 | ||
179 | if (mp_opt->use_ack) { | |
180 | if (mp_opt->ack64) | |
181 | expected_opsize += TCPOLEN_MPTCP_DSS_ACK64; | |
182 | else | |
183 | expected_opsize += TCPOLEN_MPTCP_DSS_ACK32; | |
184 | } | |
185 | ||
186 | if (mp_opt->use_map) { | |
187 | if (mp_opt->dsn64) | |
188 | expected_opsize += TCPOLEN_MPTCP_DSS_MAP64; | |
189 | else | |
190 | expected_opsize += TCPOLEN_MPTCP_DSS_MAP32; | |
191 | } | |
192 | ||
390b95a5 GT |
193 | /* Always parse any csum presence combination, we will enforce |
194 | * RFC 8684 Section 3.3.0 checks later in subflow_data_ready | |
648ef4b8 MM |
195 | */ |
196 | if (opsize != expected_opsize && | |
197 | opsize != expected_opsize + TCPOLEN_MPTCP_DSS_CHECKSUM) | |
198 | break; | |
199 | ||
74c7dfbe | 200 | mp_opt->suboptions |= OPTION_MPTCP_DSS; |
648ef4b8 MM |
201 | if (mp_opt->use_ack) { |
202 | if (mp_opt->ack64) { | |
203 | mp_opt->data_ack = get_unaligned_be64(ptr); | |
204 | ptr += 8; | |
205 | } else { | |
206 | mp_opt->data_ack = get_unaligned_be32(ptr); | |
207 | ptr += 4; | |
208 | } | |
209 | ||
210 | pr_debug("data_ack=%llu", mp_opt->data_ack); | |
211 | } | |
212 | ||
213 | if (mp_opt->use_map) { | |
214 | if (mp_opt->dsn64) { | |
215 | mp_opt->data_seq = get_unaligned_be64(ptr); | |
216 | ptr += 8; | |
217 | } else { | |
218 | mp_opt->data_seq = get_unaligned_be32(ptr); | |
219 | ptr += 4; | |
220 | } | |
221 | ||
222 | mp_opt->subflow_seq = get_unaligned_be32(ptr); | |
223 | ptr += 4; | |
224 | ||
225 | mp_opt->data_len = get_unaligned_be16(ptr); | |
226 | ptr += 2; | |
227 | ||
390b95a5 | 228 | if (opsize == expected_opsize + TCPOLEN_MPTCP_DSS_CHECKSUM) { |
74c7dfbe | 229 | mp_opt->suboptions |= OPTION_MPTCP_CSUMREQD; |
ba2c89e0 | 230 | mp_opt->csum = get_unaligned((__force __sum16 *)ptr); |
390b95a5 GT |
231 | ptr += 2; |
232 | } | |
233 | ||
234 | pr_debug("data_seq=%llu subflow_seq=%u data_len=%u csum=%d:%u", | |
648ef4b8 | 235 | mp_opt->data_seq, mp_opt->subflow_seq, |
74c7dfbe PA |
236 | mp_opt->data_len, !!(mp_opt->suboptions & OPTION_MPTCP_CSUMREQD), |
237 | mp_opt->csum); | |
648ef4b8 MM |
238 | } |
239 | ||
eda7acdd PK |
240 | break; |
241 | ||
3df523ab PK |
242 | case MPTCPOPT_ADD_ADDR: |
243 | mp_opt->echo = (*ptr++) & MPTCP_ADDR_ECHO; | |
244 | if (!mp_opt->echo) { | |
245 | if (opsize == TCPOLEN_MPTCP_ADD_ADDR || | |
246 | opsize == TCPOLEN_MPTCP_ADD_ADDR_PORT) | |
1b1a6ef5 | 247 | mp_opt->addr.family = AF_INET; |
3df523ab PK |
248 | #if IS_ENABLED(CONFIG_MPTCP_IPV6) |
249 | else if (opsize == TCPOLEN_MPTCP_ADD_ADDR6 || | |
250 | opsize == TCPOLEN_MPTCP_ADD_ADDR6_PORT) | |
1b1a6ef5 | 251 | mp_opt->addr.family = AF_INET6; |
3df523ab PK |
252 | #endif |
253 | else | |
254 | break; | |
255 | } else { | |
256 | if (opsize == TCPOLEN_MPTCP_ADD_ADDR_BASE || | |
257 | opsize == TCPOLEN_MPTCP_ADD_ADDR_BASE_PORT) | |
1b1a6ef5 | 258 | mp_opt->addr.family = AF_INET; |
3df523ab PK |
259 | #if IS_ENABLED(CONFIG_MPTCP_IPV6) |
260 | else if (opsize == TCPOLEN_MPTCP_ADD_ADDR6_BASE || | |
261 | opsize == TCPOLEN_MPTCP_ADD_ADDR6_BASE_PORT) | |
1b1a6ef5 | 262 | mp_opt->addr.family = AF_INET6; |
3df523ab PK |
263 | #endif |
264 | else | |
265 | break; | |
266 | } | |
267 | ||
74c7dfbe | 268 | mp_opt->suboptions |= OPTION_MPTCP_ADD_ADDR; |
f7dafee1 | 269 | mp_opt->addr.id = *ptr++; |
a086aeba PA |
270 | mp_opt->addr.port = 0; |
271 | mp_opt->ahmac = 0; | |
1b1a6ef5 | 272 | if (mp_opt->addr.family == AF_INET) { |
f7dafee1 | 273 | memcpy((u8 *)&mp_opt->addr.addr.s_addr, (u8 *)ptr, 4); |
3df523ab PK |
274 | ptr += 4; |
275 | if (opsize == TCPOLEN_MPTCP_ADD_ADDR_PORT || | |
276 | opsize == TCPOLEN_MPTCP_ADD_ADDR_BASE_PORT) { | |
f7dafee1 | 277 | mp_opt->addr.port = htons(get_unaligned_be16(ptr)); |
3df523ab PK |
278 | ptr += 2; |
279 | } | |
280 | } | |
281 | #if IS_ENABLED(CONFIG_MPTCP_IPV6) | |
282 | else { | |
f7dafee1 | 283 | memcpy(mp_opt->addr.addr6.s6_addr, (u8 *)ptr, 16); |
3df523ab PK |
284 | ptr += 16; |
285 | if (opsize == TCPOLEN_MPTCP_ADD_ADDR6_PORT || | |
286 | opsize == TCPOLEN_MPTCP_ADD_ADDR6_BASE_PORT) { | |
f7dafee1 | 287 | mp_opt->addr.port = htons(get_unaligned_be16(ptr)); |
3df523ab PK |
288 | ptr += 2; |
289 | } | |
290 | } | |
291 | #endif | |
292 | if (!mp_opt->echo) { | |
293 | mp_opt->ahmac = get_unaligned_be64(ptr); | |
294 | ptr += 8; | |
295 | } | |
90a4aea8 | 296 | pr_debug("ADD_ADDR%s: id=%d, ahmac=%llu, echo=%d, port=%d", |
1b1a6ef5 | 297 | (mp_opt->addr.family == AF_INET6) ? "6" : "", |
f7dafee1 | 298 | mp_opt->addr.id, mp_opt->ahmac, mp_opt->echo, ntohs(mp_opt->addr.port)); |
3df523ab PK |
299 | break; |
300 | ||
301 | case MPTCPOPT_RM_ADDR: | |
5c4a824d GT |
302 | if (opsize < TCPOLEN_MPTCP_RM_ADDR_BASE + 1 || |
303 | opsize > TCPOLEN_MPTCP_RM_ADDR_BASE + MPTCP_RM_IDS_MAX) | |
3df523ab PK |
304 | break; |
305 | ||
8e60eed6 GT |
306 | ptr++; |
307 | ||
74c7dfbe | 308 | mp_opt->suboptions |= OPTION_MPTCP_RM_ADDR; |
5c4a824d GT |
309 | mp_opt->rm_list.nr = opsize - TCPOLEN_MPTCP_RM_ADDR_BASE; |
310 | for (i = 0; i < mp_opt->rm_list.nr; i++) | |
311 | mp_opt->rm_list.ids[i] = *ptr++; | |
312 | pr_debug("RM_ADDR: rm_list_nr=%d", mp_opt->rm_list.nr); | |
3df523ab PK |
313 | break; |
314 | ||
40453a5c GT |
315 | case MPTCPOPT_MP_PRIO: |
316 | if (opsize != TCPOLEN_MPTCP_PRIO) | |
317 | break; | |
318 | ||
74c7dfbe | 319 | mp_opt->suboptions |= OPTION_MPTCP_PRIO; |
40453a5c GT |
320 | mp_opt->backup = *ptr++ & MPTCP_PRIO_BKUP; |
321 | pr_debug("MP_PRIO: prio=%d", mp_opt->backup); | |
322 | break; | |
323 | ||
50c504a2 FW |
324 | case MPTCPOPT_MP_FASTCLOSE: |
325 | if (opsize != TCPOLEN_MPTCP_FASTCLOSE) | |
326 | break; | |
327 | ||
328 | ptr += 2; | |
329 | mp_opt->rcvr_key = get_unaligned_be64(ptr); | |
330 | ptr += 8; | |
74c7dfbe | 331 | mp_opt->suboptions |= OPTION_MPTCP_FASTCLOSE; |
1e75629c | 332 | pr_debug("MP_FASTCLOSE: recv_key=%llu", mp_opt->rcvr_key); |
50c504a2 FW |
333 | break; |
334 | ||
dc87efdb FW |
335 | case MPTCPOPT_RST: |
336 | if (opsize != TCPOLEN_MPTCP_RST) | |
337 | break; | |
338 | ||
339 | if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_RST)) | |
340 | break; | |
74c7dfbe PA |
341 | |
342 | mp_opt->suboptions |= OPTION_MPTCP_RST; | |
dc87efdb FW |
343 | flags = *ptr++; |
344 | mp_opt->reset_transient = flags & MPTCP_RST_TRANSIENT; | |
345 | mp_opt->reset_reason = *ptr; | |
9ddd1cac GT |
346 | pr_debug("MP_RST: transient=%u reason=%u", |
347 | mp_opt->reset_transient, mp_opt->reset_reason); | |
dc87efdb FW |
348 | break; |
349 | ||
5580d41b GT |
350 | case MPTCPOPT_MP_FAIL: |
351 | if (opsize != TCPOLEN_MPTCP_FAIL) | |
352 | break; | |
353 | ||
354 | ptr += 2; | |
74c7dfbe | 355 | mp_opt->suboptions |= OPTION_MPTCP_FAIL; |
5580d41b GT |
356 | mp_opt->fail_seq = get_unaligned_be64(ptr); |
357 | pr_debug("MP_FAIL: data_seq=%llu", mp_opt->fail_seq); | |
358 | break; | |
359 | ||
eda7acdd PK |
360 | default: |
361 | break; | |
362 | } | |
363 | } | |
364 | ||
0799e21b | 365 | void mptcp_get_options(const struct sk_buff *skb, |
cfde141e | 366 | struct mptcp_options_received *mp_opt) |
cec37a6e | 367 | { |
cec37a6e | 368 | const struct tcphdr *th = tcp_hdr(skb); |
cfde141e PA |
369 | const unsigned char *ptr; |
370 | int length; | |
cec37a6e | 371 | |
cfde141e | 372 | /* initialize option status */ |
74c7dfbe | 373 | mp_opt->suboptions = 0; |
cfde141e PA |
374 | |
375 | length = (th->doff * 4) - sizeof(struct tcphdr); | |
cec37a6e PK |
376 | ptr = (const unsigned char *)(th + 1); |
377 | ||
378 | while (length > 0) { | |
379 | int opcode = *ptr++; | |
380 | int opsize; | |
381 | ||
382 | switch (opcode) { | |
383 | case TCPOPT_EOL: | |
384 | return; | |
385 | case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ | |
386 | length--; | |
387 | continue; | |
388 | default: | |
07718be2 MM |
389 | if (length < 2) |
390 | return; | |
cec37a6e PK |
391 | opsize = *ptr++; |
392 | if (opsize < 2) /* "silly options" */ | |
393 | return; | |
394 | if (opsize > length) | |
395 | return; /* don't parse partial options */ | |
396 | if (opcode == TCPOPT_MPTCP) | |
cfde141e | 397 | mptcp_parse_option(skb, ptr, opsize, mp_opt); |
cec37a6e PK |
398 | ptr += opsize - 2; |
399 | length -= opsize; | |
400 | } | |
401 | } | |
402 | } | |
403 | ||
cc7972ea CP |
404 | bool mptcp_syn_options(struct sock *sk, const struct sk_buff *skb, |
405 | unsigned int *size, struct mptcp_out_options *opts) | |
cec37a6e PK |
406 | { |
407 | struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); | |
408 | ||
cc7972ea CP |
409 | /* we will use snd_isn to detect first pkt [re]transmission |
410 | * in mptcp_established_options_mp() | |
411 | */ | |
412 | subflow->snd_isn = TCP_SKB_CB(skb)->end_seq; | |
cec37a6e | 413 | if (subflow->request_mptcp) { |
cec37a6e | 414 | opts->suboptions = OPTION_MPTCP_MPC_SYN; |
06fe1719 | 415 | opts->csum_reqd = mptcp_is_checksum_enabled(sock_net(sk)); |
bab6b88e | 416 | opts->allow_join_id0 = mptcp_allow_join_id0(sock_net(sk)); |
cec37a6e PK |
417 | *size = TCPOLEN_MPTCP_MPC_SYN; |
418 | return true; | |
ec3edaa7 PK |
419 | } else if (subflow->request_join) { |
420 | pr_debug("remote_token=%u, nonce=%u", subflow->remote_token, | |
421 | subflow->local_nonce); | |
422 | opts->suboptions = OPTION_MPTCP_MPJ_SYN; | |
423 | opts->join_id = subflow->local_id; | |
424 | opts->token = subflow->remote_token; | |
425 | opts->nonce = subflow->local_nonce; | |
426 | opts->backup = subflow->request_bkup; | |
427 | *size = TCPOLEN_MPTCP_MPJ_SYN; | |
428 | return true; | |
cec37a6e PK |
429 | } |
430 | return false; | |
431 | } | |
432 | ||
ec3edaa7 PK |
433 | static void clear_3rdack_retransmission(struct sock *sk) |
434 | { | |
435 | struct inet_connection_sock *icsk = inet_csk(sk); | |
436 | ||
437 | sk_stop_timer(sk, &icsk->icsk_delack_timer); | |
438 | icsk->icsk_ack.timeout = 0; | |
439 | icsk->icsk_ack.ato = 0; | |
440 | icsk->icsk_ack.pending &= ~(ICSK_ACK_SCHED | ICSK_ACK_TIMER); | |
441 | } | |
442 | ||
cc7972ea | 443 | static bool mptcp_established_options_mp(struct sock *sk, struct sk_buff *skb, |
d87903b6 | 444 | bool snd_data_fin_enable, |
cc7972ea | 445 | unsigned int *size, |
6d0060f6 | 446 | struct mptcp_out_options *opts) |
cec37a6e PK |
447 | { |
448 | struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); | |
06fe1719 | 449 | struct mptcp_sock *msk = mptcp_sk(subflow->conn); |
cc7972ea CP |
450 | struct mptcp_ext *mpext; |
451 | unsigned int data_len; | |
c94b1f96 | 452 | u8 len; |
cc7972ea | 453 | |
ec3edaa7 PK |
454 | /* When skb is not available, we better over-estimate the emitted |
455 | * options len. A full DSS option (28 bytes) is longer than | |
456 | * TCPOLEN_MPTCP_MPC_ACK_DATA(22) or TCPOLEN_MPTCP_MPJ_ACK(24), so | |
457 | * tell the caller to defer the estimate to | |
458 | * mptcp_established_options_dss(), which will reserve enough space. | |
459 | */ | |
460 | if (!skb) | |
461 | return false; | |
cc7972ea | 462 | |
d87903b6 PA |
463 | /* MPC/MPJ needed only on 3rd ack packet, DATA_FIN and TCP shutdown take precedence */ |
464 | if (subflow->fully_established || snd_data_fin_enable || | |
465 | subflow->snd_isn != TCP_SKB_CB(skb)->seq || | |
466 | sk->sk_state != TCP_ESTABLISHED) | |
ec3edaa7 PK |
467 | return false; |
468 | ||
469 | if (subflow->mp_capable) { | |
cc7972ea CP |
470 | mpext = mptcp_get_ext(skb); |
471 | data_len = mpext ? mpext->data_len : 0; | |
cec37a6e | 472 | |
f7cc8890 | 473 | /* we will check ops->data_len in mptcp_write_options() to |
cc7972ea CP |
474 | * discriminate between TCPOLEN_MPTCP_MPC_ACK_DATA and |
475 | * TCPOLEN_MPTCP_MPC_ACK | |
476 | */ | |
f7cc8890 | 477 | opts->data_len = data_len; |
cec37a6e PK |
478 | opts->suboptions = OPTION_MPTCP_MPC_ACK; |
479 | opts->sndr_key = subflow->local_key; | |
480 | opts->rcvr_key = subflow->remote_key; | |
06fe1719 | 481 | opts->csum_reqd = READ_ONCE(msk->csum_enabled); |
bab6b88e | 482 | opts->allow_join_id0 = mptcp_allow_join_id0(sock_net(sk)); |
cc7972ea CP |
483 | |
484 | /* Section 3.1. | |
485 | * The MP_CAPABLE option is carried on the SYN, SYN/ACK, and ACK | |
486 | * packets that start the first subflow of an MPTCP connection, | |
487 | * as well as the first packet that carries data | |
488 | */ | |
c94b1f96 GT |
489 | if (data_len > 0) { |
490 | len = TCPOLEN_MPTCP_MPC_ACK_DATA; | |
491 | if (opts->csum_reqd) { | |
c5b39e26 | 492 | /* we need to propagate more info to csum the pseudo hdr */ |
f7cc8890 DC |
493 | opts->data_seq = mpext->data_seq; |
494 | opts->subflow_seq = mpext->subflow_seq; | |
495 | opts->csum = mpext->csum; | |
c94b1f96 GT |
496 | len += TCPOLEN_MPTCP_DSS_CHECKSUM; |
497 | } | |
498 | *size = ALIGN(len, 4); | |
499 | } else { | |
cc7972ea | 500 | *size = TCPOLEN_MPTCP_MPC_ACK; |
c94b1f96 | 501 | } |
cc7972ea CP |
502 | |
503 | pr_debug("subflow=%p, local_key=%llu, remote_key=%llu map_len=%d", | |
504 | subflow, subflow->local_key, subflow->remote_key, | |
505 | data_len); | |
506 | ||
cec37a6e | 507 | return true; |
ec3edaa7 PK |
508 | } else if (subflow->mp_join) { |
509 | opts->suboptions = OPTION_MPTCP_MPJ_ACK; | |
510 | memcpy(opts->hmac, subflow->hmac, MPTCPOPT_HMAC_LEN); | |
511 | *size = TCPOLEN_MPTCP_MPJ_ACK; | |
512 | pr_debug("subflow=%p", subflow); | |
513 | ||
bcd97734 PA |
514 | /* we can use the full delegate action helper only from BH context |
515 | * If we are in process context - sk is flushing the backlog at | |
516 | * socket lock release time - just set the appropriate flag, will | |
517 | * be handled by the release callback | |
518 | */ | |
519 | if (sock_owned_by_user(sk)) | |
520 | set_bit(MPTCP_DELEGATE_ACK, &subflow->delegated_status); | |
521 | else | |
522 | mptcp_subflow_delegate(subflow, MPTCP_DELEGATE_ACK); | |
ec3edaa7 | 523 | return true; |
cec37a6e PK |
524 | } |
525 | return false; | |
526 | } | |
527 | ||
6d0060f6 | 528 | static void mptcp_write_data_fin(struct mptcp_subflow_context *subflow, |
9c29e361 | 529 | struct sk_buff *skb, struct mptcp_ext *ext) |
6d0060f6 | 530 | { |
017512a0 PA |
531 | /* The write_seq value has already been incremented, so the actual |
532 | * sequence number for the DATA_FIN is one less. | |
533 | */ | |
534 | u64 data_fin_tx_seq = READ_ONCE(mptcp_sk(subflow->conn)->write_seq) - 1; | |
7279da61 | 535 | |
9c29e361 | 536 | if (!ext->use_map || !skb->len) { |
6d0060f6 MM |
537 | /* RFC6824 requires a DSS mapping with specific values |
538 | * if DATA_FIN is set but no data payload is mapped | |
539 | */ | |
6d37a0b8 | 540 | ext->data_fin = 1; |
6d0060f6 MM |
541 | ext->use_map = 1; |
542 | ext->dsn64 = 1; | |
017512a0 | 543 | ext->data_seq = data_fin_tx_seq; |
6d0060f6 MM |
544 | ext->subflow_seq = 0; |
545 | ext->data_len = 1; | |
7279da61 | 546 | } else if (ext->data_seq + ext->data_len == data_fin_tx_seq) { |
6d37a0b8 MM |
547 | /* If there's an existing DSS mapping and it is the |
548 | * final mapping, DATA_FIN consumes 1 additional byte of | |
549 | * mapping space. | |
6d0060f6 | 550 | */ |
6d37a0b8 | 551 | ext->data_fin = 1; |
6d0060f6 MM |
552 | ext->data_len++; |
553 | } | |
554 | } | |
555 | ||
556 | static bool mptcp_established_options_dss(struct sock *sk, struct sk_buff *skb, | |
d87903b6 | 557 | bool snd_data_fin_enable, |
6d0060f6 | 558 | unsigned int *size, |
6d0060f6 MM |
559 | struct mptcp_out_options *opts) |
560 | { | |
561 | struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); | |
7279da61 | 562 | struct mptcp_sock *msk = mptcp_sk(subflow->conn); |
6d0060f6 MM |
563 | unsigned int dss_size = 0; |
564 | struct mptcp_ext *mpext; | |
6d0060f6 | 565 | unsigned int ack_size; |
d22f4988 | 566 | bool ret = false; |
d87903b6 | 567 | u64 ack_seq; |
6d0060f6 | 568 | |
c5b39e26 | 569 | opts->csum_reqd = READ_ONCE(msk->csum_enabled); |
0bac966a | 570 | mpext = skb ? mptcp_get_ext(skb) : NULL; |
6d0060f6 | 571 | |
7279da61 | 572 | if (!skb || (mpext && mpext->use_map) || snd_data_fin_enable) { |
c5b39e26 | 573 | unsigned int map_size = TCPOLEN_MPTCP_DSS_BASE + TCPOLEN_MPTCP_DSS_MAP64; |
6d0060f6 | 574 | |
c5b39e26 GT |
575 | if (mpext) { |
576 | if (opts->csum_reqd) | |
577 | map_size += TCPOLEN_MPTCP_DSS_CHECKSUM; | |
6d0060f6 | 578 | |
6d0060f6 | 579 | opts->ext_copy = *mpext; |
c5b39e26 | 580 | } |
6d0060f6 | 581 | |
c5b39e26 | 582 | dss_size = map_size; |
7279da61 | 583 | if (skb && snd_data_fin_enable) |
9c29e361 | 584 | mptcp_write_data_fin(subflow, skb, &opts->ext_copy); |
1bff1e43 | 585 | opts->suboptions = OPTION_MPTCP_DSS; |
d22f4988 CP |
586 | ret = true; |
587 | } | |
588 | ||
2398e399 PA |
589 | /* passive sockets msk will set the 'can_ack' after accept(), even |
590 | * if the first subflow may have the already the remote key handy | |
591 | */ | |
d22f4988 | 592 | opts->ext_copy.use_ack = 0; |
dc093db5 | 593 | if (!READ_ONCE(msk->can_ack)) { |
d22f4988 CP |
594 | *size = ALIGN(dss_size, 4); |
595 | return ret; | |
6d0060f6 MM |
596 | } |
597 | ||
e3859603 | 598 | ack_seq = READ_ONCE(msk->ack_seq); |
37198e93 | 599 | if (READ_ONCE(msk->use_64bit_ack)) { |
a0c1d0ea | 600 | ack_size = TCPOLEN_MPTCP_DSS_ACK64; |
e3859603 | 601 | opts->ext_copy.data_ack = ack_seq; |
a0c1d0ea CP |
602 | opts->ext_copy.ack64 = 1; |
603 | } else { | |
604 | ack_size = TCPOLEN_MPTCP_DSS_ACK32; | |
e3859603 | 605 | opts->ext_copy.data_ack32 = (uint32_t)ack_seq; |
a0c1d0ea CP |
606 | opts->ext_copy.ack64 = 0; |
607 | } | |
608 | opts->ext_copy.use_ack = 1; | |
1bff1e43 | 609 | opts->suboptions = OPTION_MPTCP_DSS; |
ea4ca586 | 610 | WRITE_ONCE(msk->old_wspace, __mptcp_space((struct sock *)msk)); |
6d0060f6 MM |
611 | |
612 | /* Add kind/length/subtype/flag overhead if mapping is not populated */ | |
613 | if (dss_size == 0) | |
614 | ack_size += TCPOLEN_MPTCP_DSS_BASE; | |
615 | ||
616 | dss_size += ack_size; | |
617 | ||
6d0060f6 MM |
618 | *size = ALIGN(dss_size, 4); |
619 | return true; | |
620 | } | |
621 | ||
761c124e GT |
622 | static u64 add_addr_generate_hmac(u64 key1, u64 key2, |
623 | struct mptcp_addr_info *addr) | |
3df523ab | 624 | { |
761c124e | 625 | u16 port = ntohs(addr->port); |
bd697222 | 626 | u8 hmac[SHA256_DIGEST_SIZE]; |
3df523ab | 627 | u8 msg[19]; |
761c124e | 628 | int i = 0; |
3df523ab | 629 | |
761c124e GT |
630 | msg[i++] = addr->id; |
631 | if (addr->family == AF_INET) { | |
632 | memcpy(&msg[i], &addr->addr.s_addr, 4); | |
633 | i += 4; | |
634 | } | |
635 | #if IS_ENABLED(CONFIG_MPTCP_IPV6) | |
636 | else if (addr->family == AF_INET6) { | |
637 | memcpy(&msg[i], &addr->addr6.s6_addr, 16); | |
638 | i += 16; | |
639 | } | |
640 | #endif | |
641 | msg[i++] = port >> 8; | |
642 | msg[i++] = port & 0xFF; | |
3df523ab | 643 | |
761c124e | 644 | mptcp_crypto_hmac_sha(key1, key2, msg, i, hmac); |
3df523ab | 645 | |
bd697222 | 646 | return get_unaligned_be64(&hmac[SHA256_DIGEST_SIZE - sizeof(u64)]); |
3df523ab | 647 | } |
3df523ab | 648 | |
84dfe367 | 649 | static bool mptcp_established_options_add_addr(struct sock *sk, struct sk_buff *skb, |
f643b803 GT |
650 | unsigned int *size, |
651 | unsigned int remaining, | |
652 | struct mptcp_out_options *opts) | |
3df523ab PK |
653 | { |
654 | struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); | |
655 | struct mptcp_sock *msk = mptcp_sk(subflow->conn); | |
84dfe367 GT |
656 | bool drop_other_suboptions = false; |
657 | unsigned int opt_size = *size; | |
6a6c05a8 | 658 | bool echo; |
1b1c7a0e | 659 | int len; |
3df523ab | 660 | |
1f5e9e2f YL |
661 | /* add addr will strip the existing options, be sure to avoid breaking |
662 | * MPC/MPJ handshakes | |
663 | */ | |
f643b803 | 664 | if (!mptcp_pm_should_add_signal(msk) || |
1f5e9e2f YL |
665 | (opts->suboptions & (OPTION_MPTCP_MPJ_ACK | OPTION_MPTCP_MPC_ACK)) || |
666 | !mptcp_pm_add_addr_signal(msk, skb, opt_size, remaining, &opts->addr, | |
af7939f3 | 667 | &echo, &drop_other_suboptions)) |
1b1c7a0e PK |
668 | return false; |
669 | ||
1f5e9e2f YL |
670 | if (drop_other_suboptions) |
671 | remaining += opt_size; | |
af7939f3 | 672 | len = mptcp_add_addr_len(opts->addr.family, echo, !!opts->addr.port); |
1b1c7a0e PK |
673 | if (remaining < len) |
674 | return false; | |
3df523ab | 675 | |
1b1c7a0e | 676 | *size = len; |
1f5e9e2f YL |
677 | if (drop_other_suboptions) { |
678 | pr_debug("drop other suboptions"); | |
679 | opts->suboptions = 0; | |
1bff1e43 PA |
680 | |
681 | /* note that e.g. DSS could have written into the memory | |
682 | * aliased by ahmac, we must reset the field here | |
683 | * to avoid appending the hmac even for ADD_ADDR echo | |
684 | * options | |
685 | */ | |
686 | opts->ahmac = 0; | |
84dfe367 | 687 | *size -= opt_size; |
1f5e9e2f | 688 | } |
fef6b7ec | 689 | opts->suboptions |= OPTION_MPTCP_ADD_ADDR; |
761c124e | 690 | if (!echo) { |
45b1a122 | 691 | MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_ADDADDRTX); |
761c124e GT |
692 | opts->ahmac = add_addr_generate_hmac(msk->local_key, |
693 | msk->remote_key, | |
694 | &opts->addr); | |
45b1a122 PA |
695 | } else { |
696 | MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_ECHOADDTX); | |
3df523ab | 697 | } |
4a2777a8 | 698 | pr_debug("addr_id=%d, ahmac=%llu, echo=%d, port=%d", |
30f60bae | 699 | opts->addr.id, opts->ahmac, echo, ntohs(opts->addr.port)); |
3df523ab PK |
700 | |
701 | return true; | |
702 | } | |
703 | ||
5cb104ae GT |
704 | static bool mptcp_established_options_rm_addr(struct sock *sk, |
705 | unsigned int *size, | |
706 | unsigned int remaining, | |
707 | struct mptcp_out_options *opts) | |
708 | { | |
709 | struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); | |
710 | struct mptcp_sock *msk = mptcp_sk(subflow->conn); | |
6445e17a GT |
711 | struct mptcp_rm_list rm_list; |
712 | int i, len; | |
5cb104ae GT |
713 | |
714 | if (!mptcp_pm_should_rm_signal(msk) || | |
6445e17a | 715 | !(mptcp_pm_rm_addr_signal(msk, remaining, &rm_list))) |
5cb104ae GT |
716 | return false; |
717 | ||
6445e17a GT |
718 | len = mptcp_rm_addr_len(&rm_list); |
719 | if (len < 0) | |
720 | return false; | |
721 | if (remaining < len) | |
5cb104ae GT |
722 | return false; |
723 | ||
6445e17a | 724 | *size = len; |
5cb104ae | 725 | opts->suboptions |= OPTION_MPTCP_RM_ADDR; |
6445e17a | 726 | opts->rm_list = rm_list; |
5cb104ae | 727 | |
6445e17a GT |
728 | for (i = 0; i < opts->rm_list.nr; i++) |
729 | pr_debug("rm_list_ids[%d]=%d", i, opts->rm_list.ids[i]); | |
45b1a122 | 730 | MPTCP_ADD_STATS(sock_net(sk), MPTCP_MIB_RMADDRTX, opts->rm_list.nr); |
5cb104ae GT |
731 | return true; |
732 | } | |
733 | ||
06706542 GT |
734 | static bool mptcp_established_options_mp_prio(struct sock *sk, |
735 | unsigned int *size, | |
736 | unsigned int remaining, | |
737 | struct mptcp_out_options *opts) | |
738 | { | |
739 | struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); | |
740 | ||
1bff1e43 PA |
741 | /* can't send MP_PRIO with MPC, as they share the same option space: |
742 | * 'backup'. Also it makes no sense at all | |
743 | */ | |
13ac17a3 | 744 | if (!subflow->send_mp_prio || (opts->suboptions & OPTIONS_MPTCP_MPC)) |
06706542 GT |
745 | return false; |
746 | ||
ec99a470 DC |
747 | /* account for the trailing 'nop' option */ |
748 | if (remaining < TCPOLEN_MPTCP_PRIO_ALIGN) | |
06706542 GT |
749 | return false; |
750 | ||
ec99a470 | 751 | *size = TCPOLEN_MPTCP_PRIO_ALIGN; |
06706542 GT |
752 | opts->suboptions |= OPTION_MPTCP_PRIO; |
753 | opts->backup = subflow->request_bkup; | |
754 | ||
755 | pr_debug("prio=%d", opts->backup); | |
756 | ||
757 | return true; | |
758 | } | |
759 | ||
c25aeb4e | 760 | static noinline bool mptcp_established_options_rst(struct sock *sk, struct sk_buff *skb, |
dc87efdb FW |
761 | unsigned int *size, |
762 | unsigned int remaining, | |
763 | struct mptcp_out_options *opts) | |
764 | { | |
765 | const struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); | |
766 | ||
767 | if (remaining < TCPOLEN_MPTCP_RST) | |
c25aeb4e | 768 | return false; |
dc87efdb FW |
769 | |
770 | *size = TCPOLEN_MPTCP_RST; | |
771 | opts->suboptions |= OPTION_MPTCP_RST; | |
772 | opts->reset_transient = subflow->reset_transient; | |
773 | opts->reset_reason = subflow->reset_reason; | |
0c1f78a4 | 774 | MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPRSTTX); |
c25aeb4e GT |
775 | |
776 | return true; | |
777 | } | |
778 | ||
f284c0c7 PA |
779 | static bool mptcp_established_options_fastclose(struct sock *sk, |
780 | unsigned int *size, | |
781 | unsigned int remaining, | |
782 | struct mptcp_out_options *opts) | |
783 | { | |
784 | struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); | |
785 | struct mptcp_sock *msk = mptcp_sk(subflow->conn); | |
786 | ||
787 | if (likely(!subflow->send_fastclose)) | |
788 | return false; | |
789 | ||
790 | if (remaining < TCPOLEN_MPTCP_FASTCLOSE) | |
791 | return false; | |
792 | ||
793 | *size = TCPOLEN_MPTCP_FASTCLOSE; | |
794 | opts->suboptions |= OPTION_MPTCP_FASTCLOSE; | |
795 | opts->rcvr_key = msk->remote_key; | |
796 | ||
797 | pr_debug("FASTCLOSE key=%llu", opts->rcvr_key); | |
0c1f78a4 | 798 | MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPFASTCLOSETX); |
f284c0c7 PA |
799 | return true; |
800 | } | |
801 | ||
c25aeb4e GT |
802 | static bool mptcp_established_options_mp_fail(struct sock *sk, |
803 | unsigned int *size, | |
804 | unsigned int remaining, | |
805 | struct mptcp_out_options *opts) | |
806 | { | |
807 | struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); | |
808 | ||
809 | if (likely(!subflow->send_mp_fail)) | |
810 | return false; | |
811 | ||
812 | if (remaining < TCPOLEN_MPTCP_FAIL) | |
813 | return false; | |
814 | ||
815 | *size = TCPOLEN_MPTCP_FAIL; | |
816 | opts->suboptions |= OPTION_MPTCP_FAIL; | |
817 | opts->fail_seq = subflow->map_seq; | |
818 | ||
819 | pr_debug("MP_FAIL fail_seq=%llu", opts->fail_seq); | |
0c1f78a4 | 820 | MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPFAILTX); |
c25aeb4e GT |
821 | |
822 | return true; | |
dc87efdb FW |
823 | } |
824 | ||
6d0060f6 MM |
825 | bool mptcp_established_options(struct sock *sk, struct sk_buff *skb, |
826 | unsigned int *size, unsigned int remaining, | |
827 | struct mptcp_out_options *opts) | |
828 | { | |
d87903b6 PA |
829 | struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); |
830 | struct mptcp_sock *msk = mptcp_sk(subflow->conn); | |
6d0060f6 | 831 | unsigned int opt_size = 0; |
d87903b6 | 832 | bool snd_data_fin; |
6d0060f6 MM |
833 | bool ret = false; |
834 | ||
3df523ab PK |
835 | opts->suboptions = 0; |
836 | ||
1e39e5a3 | 837 | if (unlikely(__mptcp_check_fallback(msk) && !mptcp_check_infinite_map(skb))) |
e1ff9e82 DC |
838 | return false; |
839 | ||
dc87efdb | 840 | if (unlikely(skb && TCP_SKB_CB(skb)->tcp_flags & TCPHDR_RST)) { |
f284c0c7 PA |
841 | if (mptcp_established_options_fastclose(sk, &opt_size, remaining, opts) || |
842 | mptcp_established_options_mp_fail(sk, &opt_size, remaining, opts)) { | |
c25aeb4e GT |
843 | *size += opt_size; |
844 | remaining -= opt_size; | |
845 | } | |
f284c0c7 | 846 | /* MP_RST can be used with MP_FASTCLOSE and MP_FAIL if there is room */ |
c25aeb4e GT |
847 | if (mptcp_established_options_rst(sk, skb, &opt_size, remaining, opts)) { |
848 | *size += opt_size; | |
849 | remaining -= opt_size; | |
850 | } | |
dc87efdb FW |
851 | return true; |
852 | } | |
d5824847 | 853 | |
d87903b6 | 854 | snd_data_fin = mptcp_data_fin_enabled(msk); |
ce395d0e | 855 | if (mptcp_established_options_mp(sk, skb, snd_data_fin, &opt_size, opts)) |
6d0060f6 | 856 | ret = true; |
ce395d0e | 857 | else if (mptcp_established_options_dss(sk, skb, snd_data_fin, &opt_size, opts)) { |
04fac2ca MB |
858 | unsigned int mp_fail_size; |
859 | ||
6d0060f6 | 860 | ret = true; |
04fac2ca MB |
861 | if (mptcp_established_options_mp_fail(sk, &mp_fail_size, |
862 | remaining - opt_size, opts)) { | |
863 | *size += opt_size + mp_fail_size; | |
864 | remaining -= opt_size - mp_fail_size; | |
c25aeb4e GT |
865 | return true; |
866 | } | |
867 | } | |
6d0060f6 MM |
868 | |
869 | /* we reserved enough space for the above options, and exceeding the | |
870 | * TCP option space would be fatal | |
871 | */ | |
872 | if (WARN_ON_ONCE(opt_size > remaining)) | |
873 | return false; | |
874 | ||
875 | *size += opt_size; | |
876 | remaining -= opt_size; | |
84dfe367 | 877 | if (mptcp_established_options_add_addr(sk, skb, &opt_size, remaining, opts)) { |
3df523ab PK |
878 | *size += opt_size; |
879 | remaining -= opt_size; | |
880 | ret = true; | |
5cb104ae GT |
881 | } else if (mptcp_established_options_rm_addr(sk, &opt_size, remaining, opts)) { |
882 | *size += opt_size; | |
883 | remaining -= opt_size; | |
884 | ret = true; | |
3df523ab | 885 | } |
6d0060f6 | 886 | |
06706542 GT |
887 | if (mptcp_established_options_mp_prio(sk, &opt_size, remaining, opts)) { |
888 | *size += opt_size; | |
889 | remaining -= opt_size; | |
890 | ret = true; | |
891 | } | |
892 | ||
6d0060f6 MM |
893 | return ret; |
894 | } | |
895 | ||
cec37a6e PK |
896 | bool mptcp_synack_options(const struct request_sock *req, unsigned int *size, |
897 | struct mptcp_out_options *opts) | |
898 | { | |
899 | struct mptcp_subflow_request_sock *subflow_req = mptcp_subflow_rsk(req); | |
900 | ||
901 | if (subflow_req->mp_capable) { | |
902 | opts->suboptions = OPTION_MPTCP_MPC_SYNACK; | |
903 | opts->sndr_key = subflow_req->local_key; | |
06fe1719 | 904 | opts->csum_reqd = subflow_req->csum_reqd; |
bab6b88e | 905 | opts->allow_join_id0 = subflow_req->allow_join_id0; |
cec37a6e PK |
906 | *size = TCPOLEN_MPTCP_MPC_SYNACK; |
907 | pr_debug("subflow_req=%p, local_key=%llu", | |
908 | subflow_req, subflow_req->local_key); | |
909 | return true; | |
f296234c PK |
910 | } else if (subflow_req->mp_join) { |
911 | opts->suboptions = OPTION_MPTCP_MPJ_SYNACK; | |
912 | opts->backup = subflow_req->backup; | |
913 | opts->join_id = subflow_req->local_id; | |
914 | opts->thmac = subflow_req->thmac; | |
915 | opts->nonce = subflow_req->local_nonce; | |
916 | pr_debug("req=%p, bkup=%u, id=%u, thmac=%llu, nonce=%u", | |
917 | subflow_req, opts->backup, opts->join_id, | |
918 | opts->thmac, opts->nonce); | |
919 | *size = TCPOLEN_MPTCP_MPJ_SYNACK; | |
920 | return true; | |
cec37a6e PK |
921 | } |
922 | return false; | |
923 | } | |
924 | ||
d5824847 | 925 | static bool check_fully_established(struct mptcp_sock *msk, struct sock *ssk, |
f296234c | 926 | struct mptcp_subflow_context *subflow, |
0be534f5 PA |
927 | struct sk_buff *skb, |
928 | struct mptcp_options_received *mp_opt) | |
d22f4988 CP |
929 | { |
930 | /* here we can process OoO, in-window pkts, only in-sequence 4th ack | |
f296234c | 931 | * will make the subflow fully established |
d22f4988 | 932 | */ |
f296234c PK |
933 | if (likely(subflow->fully_established)) { |
934 | /* on passive sockets, check for 3rd ack retransmission | |
935 | * note that msk is always set by subflow_syn_recv_sock() | |
936 | * for mp_join subflows | |
937 | */ | |
938 | if (TCP_SKB_CB(skb)->seq == subflow->ssn_offset + 1 && | |
939 | TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq && | |
74c7dfbe | 940 | subflow->mp_join && (mp_opt->suboptions & OPTIONS_MPTCP_MPJ) && |
70c708e8 | 941 | !subflow->request_join) |
d5824847 | 942 | tcp_send_ack(ssk); |
dfc8d060 | 943 | goto check_notify; |
f296234c PK |
944 | } |
945 | ||
d5824847 PA |
946 | /* we must process OoO packets before the first subflow is fully |
947 | * established. OoO packets are instead a protocol violation | |
948 | * for MP_JOIN subflows as the peer must not send any data | |
949 | * before receiving the forth ack - cfr. RFC 8684 section 3.2. | |
f296234c | 950 | */ |
d5824847 PA |
951 | if (TCP_SKB_CB(skb)->seq != subflow->ssn_offset + 1) { |
952 | if (subflow->mp_join) | |
953 | goto reset; | |
dfc8d060 DS |
954 | if (subflow->is_mptfo && mp_opt->suboptions & OPTION_MPTCP_MPC_ACK) |
955 | goto set_fully_established; | |
f296234c | 956 | return subflow->mp_capable; |
d5824847 | 957 | } |
d22f4988 | 958 | |
b3ea6b27 PA |
959 | if (subflow->remote_key_valid && |
960 | (((mp_opt->suboptions & OPTION_MPTCP_DSS) && mp_opt->use_ack) || | |
961 | ((mp_opt->suboptions & OPTION_MPTCP_ADD_ADDR) && !mp_opt->echo))) { | |
f296234c | 962 | /* subflows are fully established as soon as we get any |
67b12f79 | 963 | * additional ack, including ADD_ADDR. |
f296234c | 964 | */ |
e4a0fa47 | 965 | goto set_fully_established; |
f296234c | 966 | } |
d22f4988 | 967 | |
d22f4988 | 968 | /* If the first established packet does not contain MP_CAPABLE + data |
d5824847 PA |
969 | * then fallback to TCP. Fallback scenarios requires a reset for |
970 | * MP_JOIN subflows. | |
d22f4988 | 971 | */ |
74c7dfbe | 972 | if (!(mp_opt->suboptions & OPTIONS_MPTCP_MPC)) { |
d5824847 PA |
973 | if (subflow->mp_join) |
974 | goto reset; | |
d22f4988 | 975 | subflow->mp_capable = 0; |
e1ff9e82 | 976 | pr_fallback(msk); |
d51991e2 | 977 | mptcp_do_fallback(ssk); |
d22f4988 CP |
978 | return false; |
979 | } | |
f296234c | 980 | |
df377be3 GT |
981 | if (mp_opt->deny_join_id0) |
982 | WRITE_ONCE(msk->pm.remote_deny_join_id0, true); | |
983 | ||
d6085fe1 PA |
984 | if (unlikely(!READ_ONCE(msk->pm.server_side))) |
985 | pr_warn_once("bogus mpc option on established client sk"); | |
e4a0fa47 | 986 | |
5b49c41a | 987 | set_fully_established: |
e4a0fa47 PA |
988 | mptcp_data_lock((struct sock *)msk); |
989 | __mptcp_subflow_fully_established(msk, subflow, mp_opt); | |
990 | mptcp_data_unlock((struct sock *)msk); | |
f296234c | 991 | |
dfc8d060 | 992 | check_notify: |
5b950ff4 PA |
993 | /* if the subflow is not already linked into the conn_list, we can't |
994 | * notify the PM: this subflow is still on the listener queue | |
995 | * and the PM possibly acquiring the subflow lock could race with | |
996 | * the listener close | |
997 | */ | |
998 | if (likely(subflow->pm_notified) || list_empty(&subflow->node)) | |
f296234c PK |
999 | return true; |
1000 | ||
1001 | subflow->pm_notified = 1; | |
ec3edaa7 | 1002 | if (subflow->mp_join) { |
d5824847 | 1003 | clear_3rdack_retransmission(ssk); |
62535200 | 1004 | mptcp_pm_subflow_established(msk); |
ec3edaa7 | 1005 | } else { |
7a486c44 | 1006 | mptcp_pm_fully_established(msk, ssk); |
ec3edaa7 | 1007 | } |
d22f4988 | 1008 | return true; |
d5824847 PA |
1009 | |
1010 | reset: | |
1011 | mptcp_subflow_reset(ssk); | |
1012 | return false; | |
d22f4988 CP |
1013 | } |
1014 | ||
1502328f | 1015 | u64 __mptcp_expand_seq(u64 old_seq, u64 cur_seq) |
cc9d2566 | 1016 | { |
1502328f PA |
1017 | u32 old_seq32, cur_seq32; |
1018 | ||
1019 | old_seq32 = (u32)old_seq; | |
1020 | cur_seq32 = (u32)cur_seq; | |
1021 | cur_seq = (old_seq & GENMASK_ULL(63, 32)) + cur_seq32; | |
1022 | if (unlikely(cur_seq32 < old_seq32 && before(old_seq32, cur_seq32))) | |
1023 | return cur_seq + (1LL << 32); | |
1024 | ||
1025 | /* reverse wrap could happen, too */ | |
1026 | if (unlikely(cur_seq32 > old_seq32 && after(old_seq32, cur_seq32))) | |
1027 | return cur_seq - (1LL << 32); | |
1028 | return cur_seq; | |
cc9d2566 PA |
1029 | } |
1030 | ||
38967f42 PA |
1031 | static void __mptcp_snd_una_update(struct mptcp_sock *msk, u64 new_snd_una) |
1032 | { | |
1033 | msk->bytes_acked += new_snd_una - msk->snd_una; | |
1034 | msk->snd_una = new_snd_una; | |
1035 | } | |
1036 | ||
6f8a612a | 1037 | static void ack_update_msk(struct mptcp_sock *msk, |
6e628cd3 | 1038 | struct sock *ssk, |
6f8a612a | 1039 | struct mptcp_options_received *mp_opt) |
cc9d2566 | 1040 | { |
7439d687 | 1041 | u64 new_wnd_end, new_snd_una, snd_nxt = READ_ONCE(msk->snd_nxt); |
6f8a612a | 1042 | struct sock *sk = (struct sock *)msk; |
7439d687 PA |
1043 | u64 old_snd_una; |
1044 | ||
1045 | mptcp_data_lock(sk); | |
cc9d2566 PA |
1046 | |
1047 | /* avoid ack expansion on update conflict, to reduce the risk of | |
1048 | * wrongly expanding to a future ack sequence number, which is way | |
1049 | * more dangerous than missing an ack | |
1050 | */ | |
7439d687 | 1051 | old_snd_una = msk->snd_una; |
1502328f | 1052 | new_snd_una = mptcp_expand_seq(old_snd_una, mp_opt->data_ack, mp_opt->ack64); |
cc9d2566 | 1053 | |
0d199e43 FW |
1054 | /* ACK for data not even sent yet? Ignore.*/ |
1055 | if (unlikely(after64(new_snd_una, snd_nxt))) | |
1056 | new_snd_una = old_snd_una; | |
cc9d2566 | 1057 | |
6f8a612a FW |
1058 | new_wnd_end = new_snd_una + tcp_sk(ssk)->snd_wnd; |
1059 | ||
219d0499 | 1060 | if (after64(new_wnd_end, msk->wnd_end)) |
7439d687 | 1061 | msk->wnd_end = new_wnd_end; |
219d0499 PA |
1062 | |
1063 | /* this assumes mptcp_incoming_options() is invoked after tcp_ack() */ | |
d09d818e | 1064 | if (after64(msk->wnd_end, READ_ONCE(msk->snd_nxt))) |
219d0499 | 1065 | __mptcp_check_push(sk, ssk); |
6f8a612a | 1066 | |
7439d687 | 1067 | if (after64(new_snd_una, old_snd_una)) { |
38967f42 | 1068 | __mptcp_snd_una_update(msk, new_snd_una); |
7439d687 | 1069 | __mptcp_data_acked(sk); |
cc9d2566 | 1070 | } |
7439d687 | 1071 | mptcp_data_unlock(sk); |
ed66bfb4 GT |
1072 | |
1073 | trace_ack_update_msk(mp_opt->data_ack, | |
1074 | old_snd_una, new_snd_una, | |
1075 | new_wnd_end, msk->wnd_end); | |
cc9d2566 PA |
1076 | } |
1077 | ||
1a49b2c2 | 1078 | bool mptcp_update_rcv_data_fin(struct mptcp_sock *msk, u64 data_fin_seq, bool use_64bit) |
3721b9b6 MM |
1079 | { |
1080 | /* Skip if DATA_FIN was already received. | |
1081 | * If updating simultaneously with the recvmsg loop, values | |
1082 | * should match. If they mismatch, the peer is misbehaving and | |
1083 | * we will prefer the most recent information. | |
1084 | */ | |
781bf13d | 1085 | if (READ_ONCE(msk->rcv_data_fin)) |
3721b9b6 MM |
1086 | return false; |
1087 | ||
1a49b2c2 | 1088 | WRITE_ONCE(msk->rcv_data_fin_seq, |
1502328f | 1089 | mptcp_expand_seq(READ_ONCE(msk->ack_seq), data_fin_seq, use_64bit)); |
3721b9b6 MM |
1090 | WRITE_ONCE(msk->rcv_data_fin, 1); |
1091 | ||
1092 | return true; | |
1093 | } | |
1094 | ||
1b1c7a0e PK |
1095 | static bool add_addr_hmac_valid(struct mptcp_sock *msk, |
1096 | struct mptcp_options_received *mp_opt) | |
1097 | { | |
1098 | u64 hmac = 0; | |
1099 | ||
1100 | if (mp_opt->echo) | |
1101 | return true; | |
1102 | ||
761c124e GT |
1103 | hmac = add_addr_generate_hmac(msk->remote_key, |
1104 | msk->local_key, | |
1105 | &mp_opt->addr); | |
1b1c7a0e PK |
1106 | |
1107 | pr_debug("msk=%p, ahmac=%llu, mp_opt->ahmac=%llu\n", | |
742e2f36 | 1108 | msk, hmac, mp_opt->ahmac); |
1b1c7a0e PK |
1109 | |
1110 | return hmac == mp_opt->ahmac; | |
1111 | } | |
1112 | ||
6787b7e3 JW |
1113 | /* Return false if a subflow has been reset, else return true */ |
1114 | bool mptcp_incoming_options(struct sock *sk, struct sk_buff *skb) | |
648ef4b8 | 1115 | { |
d22f4988 | 1116 | struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(sk); |
1b1c7a0e | 1117 | struct mptcp_sock *msk = mptcp_sk(subflow->conn); |
cfde141e | 1118 | struct mptcp_options_received mp_opt; |
648ef4b8 MM |
1119 | struct mptcp_ext *mpext; |
1120 | ||
6e628cd3 PA |
1121 | if (__mptcp_check_fallback(msk)) { |
1122 | /* Keep it simple and unconditionally trigger send data cleanup and | |
1123 | * pending queue spooling. We will need to acquire the data lock | |
1124 | * for more accurate checks, and once the lock is acquired, such | |
1125 | * helpers are cheap. | |
1126 | */ | |
1127 | mptcp_data_lock(subflow->conn); | |
219d0499 PA |
1128 | if (sk_stream_memory_free(sk)) |
1129 | __mptcp_check_push(subflow->conn, sk); | |
c026d33b PA |
1130 | |
1131 | /* on fallback we just need to ignore the msk-level snd_una, as | |
1132 | * this is really plain TCP | |
1133 | */ | |
38967f42 | 1134 | __mptcp_snd_una_update(msk, READ_ONCE(msk->snd_nxt)); |
c026d33b | 1135 | |
6e628cd3 PA |
1136 | __mptcp_data_acked(subflow->conn); |
1137 | mptcp_data_unlock(subflow->conn); | |
6787b7e3 | 1138 | return true; |
6e628cd3 | 1139 | } |
e1ff9e82 | 1140 | |
0799e21b | 1141 | mptcp_get_options(skb, &mp_opt); |
6787b7e3 JW |
1142 | |
1143 | /* The subflow can be in close state only if check_fully_established() | |
1144 | * just sent a reset. If so, tell the caller to ignore the current packet. | |
1145 | */ | |
cfde141e | 1146 | if (!check_fully_established(msk, sk, subflow, skb, &mp_opt)) |
6787b7e3 | 1147 | return sk->sk_state != TCP_CLOSE; |
648ef4b8 | 1148 | |
f6c2ef59 PA |
1149 | if (unlikely(mp_opt.suboptions != OPTION_MPTCP_DSS)) { |
1150 | if ((mp_opt.suboptions & OPTION_MPTCP_FASTCLOSE) && | |
1151 | msk->local_key == mp_opt.rcvr_key) { | |
1152 | WRITE_ONCE(msk->rcv_fastclose, true); | |
1153 | mptcp_schedule_work((struct sock *)msk); | |
1e75629c | 1154 | MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPFASTCLOSERX); |
f6c2ef59 | 1155 | } |
50c504a2 | 1156 | |
f6c2ef59 PA |
1157 | if ((mp_opt.suboptions & OPTION_MPTCP_ADD_ADDR) && |
1158 | add_addr_hmac_valid(msk, &mp_opt)) { | |
1159 | if (!mp_opt.echo) { | |
d1ace2d9 | 1160 | mptcp_pm_add_addr_received(sk, &mp_opt.addr); |
f6c2ef59 PA |
1161 | MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_ADDADDR); |
1162 | } else { | |
1163 | mptcp_pm_add_addr_echoed(msk, &mp_opt.addr); | |
1164 | mptcp_pm_del_add_timer(msk, &mp_opt.addr, true); | |
1165 | MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_ECHOADD); | |
1166 | } | |
1167 | ||
1168 | if (mp_opt.addr.port) | |
1169 | MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_PORTADD); | |
a877de06 | 1170 | } |
2fbdd9ea | 1171 | |
f6c2ef59 PA |
1172 | if (mp_opt.suboptions & OPTION_MPTCP_RM_ADDR) |
1173 | mptcp_pm_rm_addr_received(msk, &mp_opt.rm_list); | |
1b1c7a0e | 1174 | |
f6c2ef59 PA |
1175 | if (mp_opt.suboptions & OPTION_MPTCP_PRIO) { |
1176 | mptcp_pm_mp_prio_received(sk, mp_opt.backup); | |
1177 | MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPPRIORX); | |
1178 | } | |
d0876b22 | 1179 | |
f6c2ef59 PA |
1180 | if (mp_opt.suboptions & OPTION_MPTCP_FAIL) { |
1181 | mptcp_pm_mp_fail_received(sk, mp_opt.fail_seq); | |
1182 | MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPFAILRX); | |
1183 | } | |
40453a5c | 1184 | |
f6c2ef59 PA |
1185 | if (mp_opt.suboptions & OPTION_MPTCP_RST) { |
1186 | subflow->reset_seen = 1; | |
1187 | subflow->reset_reason = mp_opt.reset_reason; | |
1188 | subflow->reset_transient = mp_opt.reset_transient; | |
e40dd439 | 1189 | MPTCP_INC_STATS(sock_net(sk), MPTCP_MIB_MPRSTRX); |
f6c2ef59 | 1190 | } |
5580d41b | 1191 | |
f6c2ef59 PA |
1192 | if (!(mp_opt.suboptions & OPTION_MPTCP_DSS)) |
1193 | return true; | |
dc87efdb FW |
1194 | } |
1195 | ||
cc9d2566 PA |
1196 | /* we can't wait for recvmsg() to update the ack_seq, otherwise |
1197 | * monodirectional flows will stuck | |
1198 | */ | |
cfde141e | 1199 | if (mp_opt.use_ack) |
6f8a612a | 1200 | ack_update_msk(msk, sk, &mp_opt); |
cc9d2566 | 1201 | |
06827b34 MM |
1202 | /* Zero-data-length packets are dropped by the caller and not |
1203 | * propagated to the MPTCP layer, so the skb extension does not | |
1204 | * need to be allocated or populated. DATA_FIN information, if | |
1205 | * present, needs to be updated here before the skb is freed. | |
43b54c6e MM |
1206 | */ |
1207 | if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) { | |
1208 | if (mp_opt.data_fin && mp_opt.data_len == 1 && | |
a5cb752b PA |
1209 | mptcp_update_rcv_data_fin(msk, mp_opt.data_seq, mp_opt.dsn64)) |
1210 | mptcp_schedule_work((struct sock *)msk); | |
06827b34 | 1211 | |
6787b7e3 | 1212 | return true; |
43b54c6e MM |
1213 | } |
1214 | ||
648ef4b8 MM |
1215 | mpext = skb_ext_add(skb, SKB_EXT_MPTCP); |
1216 | if (!mpext) | |
6787b7e3 | 1217 | return true; |
648ef4b8 MM |
1218 | |
1219 | memset(mpext, 0, sizeof(*mpext)); | |
1220 | ||
f6c2ef59 | 1221 | if (likely(mp_opt.use_map)) { |
cfde141e | 1222 | if (mp_opt.mpc_map) { |
cc7972ea CP |
1223 | /* this is an MP_CAPABLE carrying MPTCP data |
1224 | * we know this map the first chunk of data | |
1225 | */ | |
1226 | mptcp_crypto_key_sha(subflow->remote_key, NULL, | |
1227 | &mpext->data_seq); | |
1228 | mpext->data_seq++; | |
1229 | mpext->subflow_seq = 1; | |
1230 | mpext->dsn64 = 1; | |
1231 | mpext->mpc_map = 1; | |
a77895db | 1232 | mpext->data_fin = 0; |
cc7972ea | 1233 | } else { |
cfde141e PA |
1234 | mpext->data_seq = mp_opt.data_seq; |
1235 | mpext->subflow_seq = mp_opt.subflow_seq; | |
1236 | mpext->dsn64 = mp_opt.dsn64; | |
1237 | mpext->data_fin = mp_opt.data_fin; | |
cc7972ea | 1238 | } |
cfde141e | 1239 | mpext->data_len = mp_opt.data_len; |
648ef4b8 | 1240 | mpext->use_map = 1; |
74c7dfbe | 1241 | mpext->csum_reqd = !!(mp_opt.suboptions & OPTION_MPTCP_CSUMREQD); |
208e8f66 GT |
1242 | |
1243 | if (mpext->csum_reqd) | |
1244 | mpext->csum = mp_opt.csum; | |
648ef4b8 | 1245 | } |
6787b7e3 JW |
1246 | |
1247 | return true; | |
648ef4b8 MM |
1248 | } |
1249 | ||
f3589be0 | 1250 | static void mptcp_set_rwin(struct tcp_sock *tp, struct tcphdr *th) |
fa3fe2b1 FW |
1251 | { |
1252 | const struct sock *ssk = (const struct sock *)tp; | |
f3589be0 PA |
1253 | struct mptcp_subflow_context *subflow; |
1254 | u64 ack_seq, rcv_wnd_old, rcv_wnd_new; | |
fa3fe2b1 | 1255 | struct mptcp_sock *msk; |
f3589be0 PA |
1256 | u32 new_win; |
1257 | u64 win; | |
fa3fe2b1 FW |
1258 | |
1259 | subflow = mptcp_subflow_ctx(ssk); | |
1260 | msk = mptcp_sk(subflow->conn); | |
1261 | ||
f3589be0 PA |
1262 | ack_seq = READ_ONCE(msk->ack_seq); |
1263 | rcv_wnd_new = ack_seq + tp->rcv_wnd; | |
1264 | ||
1265 | rcv_wnd_old = atomic64_read(&msk->rcv_wnd_sent); | |
1266 | if (after64(rcv_wnd_new, rcv_wnd_old)) { | |
1267 | u64 rcv_wnd; | |
fa3fe2b1 | 1268 | |
f3589be0 PA |
1269 | for (;;) { |
1270 | rcv_wnd = atomic64_cmpxchg(&msk->rcv_wnd_sent, rcv_wnd_old, rcv_wnd_new); | |
1271 | ||
1272 | if (rcv_wnd == rcv_wnd_old) | |
1273 | break; | |
6bec0411 PA |
1274 | |
1275 | rcv_wnd_old = rcv_wnd; | |
1276 | if (before64(rcv_wnd_new, rcv_wnd_old)) { | |
38acb626 | 1277 | MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_RCVWNDCONFLICTUPDATE); |
f3589be0 | 1278 | goto raise_win; |
38acb626 PA |
1279 | } |
1280 | MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_RCVWNDCONFLICT); | |
f3589be0 PA |
1281 | } |
1282 | return; | |
1283 | } | |
1284 | ||
1285 | if (rcv_wnd_new != rcv_wnd_old) { | |
1286 | raise_win: | |
1287 | win = rcv_wnd_old - ack_seq; | |
1288 | tp->rcv_wnd = min_t(u64, win, U32_MAX); | |
1289 | new_win = tp->rcv_wnd; | |
1290 | ||
1291 | /* Make sure we do not exceed the maximum possible | |
1292 | * scaled window. | |
1293 | */ | |
1294 | if (unlikely(th->syn)) | |
1295 | new_win = min(new_win, 65535U) << tp->rx_opt.rcv_wscale; | |
1296 | if (!tp->rx_opt.rcv_wscale && | |
0f1e4d06 | 1297 | READ_ONCE(sock_net(ssk)->ipv4.sysctl_tcp_workaround_signed_windows)) |
f3589be0 PA |
1298 | new_win = min(new_win, MAX_TCP_WINDOW); |
1299 | else | |
1300 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
1301 | ||
1302 | /* RFC1323 scaling applied */ | |
1303 | new_win >>= tp->rx_opt.rcv_wscale; | |
1304 | th->window = htons(new_win); | |
38acb626 | 1305 | MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_RCVWNDSHARED); |
f3589be0 | 1306 | } |
fa3fe2b1 FW |
1307 | } |
1308 | ||
ba2c89e0 | 1309 | __sum16 __mptcp_make_csum(u64 data_seq, u32 subflow_seq, u16 data_len, __wsum sum) |
c94b1f96 GT |
1310 | { |
1311 | struct csum_pseudo_header header; | |
1312 | __wsum csum; | |
1313 | ||
1314 | /* cfr RFC 8684 3.3.1.: | |
1315 | * the data sequence number used in the pseudo-header is | |
1316 | * always the 64-bit value, irrespective of what length is used in the | |
1317 | * DSS option itself. | |
1318 | */ | |
f7cc8890 DC |
1319 | header.data_seq = cpu_to_be64(data_seq); |
1320 | header.subflow_seq = htonl(subflow_seq); | |
1321 | header.data_len = htons(data_len); | |
c94b1f96 GT |
1322 | header.csum = 0; |
1323 | ||
c312ee21 | 1324 | csum = csum_partial(&header, sizeof(header), sum); |
ba2c89e0 | 1325 | return csum_fold(csum); |
c94b1f96 GT |
1326 | } |
1327 | ||
ba2c89e0 | 1328 | static __sum16 mptcp_make_csum(const struct mptcp_ext *mpext) |
f7cc8890 DC |
1329 | { |
1330 | return __mptcp_make_csum(mpext->data_seq, mpext->subflow_seq, mpext->data_len, | |
c312ee21 | 1331 | ~csum_unfold(mpext->csum)); |
f7cc8890 DC |
1332 | } |
1333 | ||
ba2c89e0 PA |
1334 | static void put_len_csum(u16 len, __sum16 csum, void *data) |
1335 | { | |
1336 | __sum16 *sumptr = data + 2; | |
1337 | __be16 *ptr = data; | |
1338 | ||
1339 | put_unaligned_be16(len, ptr); | |
1340 | ||
1341 | put_unaligned(csum, sumptr); | |
1342 | } | |
1343 | ||
ea66758c | 1344 | void mptcp_write_options(struct tcphdr *th, __be32 *ptr, struct tcp_sock *tp, |
fa3fe2b1 | 1345 | struct mptcp_out_options *opts) |
eda7acdd | 1346 | { |
d7889cfa GT |
1347 | const struct sock *ssk = (const struct sock *)tp; |
1348 | struct mptcp_subflow_context *subflow; | |
c25aeb4e | 1349 | |
8cca39e2 MB |
1350 | /* Which options can be used together? |
1351 | * | |
1352 | * X: mutually exclusive | |
1353 | * O: often used together | |
1354 | * C: can be used together in some cases | |
1355 | * P: could be used together but we prefer not to (optimisations) | |
1356 | * | |
1357 | * Opt: | MPC | MPJ | DSS | ADD | RM | PRIO | FAIL | FC | | |
1358 | * ------|------|------|------|------|------|------|------|------| | |
1359 | * MPC |------|------|------|------|------|------|------|------| | |
1360 | * MPJ | X |------|------|------|------|------|------|------| | |
1361 | * DSS | X | X |------|------|------|------|------|------| | |
1362 | * ADD | X | X | P |------|------|------|------|------| | |
1363 | * RM | C | C | C | P |------|------|------|------| | |
1364 | * PRIO | X | C | C | C | C |------|------|------| | |
1365 | * FAIL | X | X | C | X | X | X |------|------| | |
1366 | * FC | X | X | X | X | X | X | X |------| | |
1367 | * RST | X | X | X | X | X | X | O | O | | |
1368 | * ------|------|------|------|------|------|------|------|------| | |
1369 | * | |
1370 | * The same applies in mptcp_established_options() function. | |
1bff1e43 PA |
1371 | */ |
1372 | if (likely(OPTION_MPTCP_DSS & opts->suboptions)) { | |
1373 | struct mptcp_ext *mpext = &opts->ext_copy; | |
1374 | u8 len = TCPOLEN_MPTCP_DSS_BASE; | |
1375 | u8 flags = 0; | |
1376 | ||
1377 | if (mpext->use_ack) { | |
1378 | flags = MPTCP_DSS_HAS_ACK; | |
1379 | if (mpext->ack64) { | |
1380 | len += TCPOLEN_MPTCP_DSS_ACK64; | |
1381 | flags |= MPTCP_DSS_ACK64; | |
1382 | } else { | |
1383 | len += TCPOLEN_MPTCP_DSS_ACK32; | |
1384 | } | |
1385 | } | |
1386 | ||
1387 | if (mpext->use_map) { | |
1388 | len += TCPOLEN_MPTCP_DSS_MAP64; | |
1389 | ||
1390 | /* Use only 64-bit mapping flags for now, add | |
1391 | * support for optional 32-bit mappings later. | |
1392 | */ | |
1393 | flags |= MPTCP_DSS_HAS_MAP | MPTCP_DSS_DSN64; | |
1394 | if (mpext->data_fin) | |
1395 | flags |= MPTCP_DSS_DATA_FIN; | |
1396 | ||
1397 | if (opts->csum_reqd) | |
1398 | len += TCPOLEN_MPTCP_DSS_CHECKSUM; | |
1399 | } | |
1400 | ||
1401 | *ptr++ = mptcp_option(MPTCPOPT_DSS, len, 0, flags); | |
1402 | ||
1403 | if (mpext->use_ack) { | |
1404 | if (mpext->ack64) { | |
1405 | put_unaligned_be64(mpext->data_ack, ptr); | |
1406 | ptr += 2; | |
1407 | } else { | |
1408 | put_unaligned_be32(mpext->data_ack32, ptr); | |
1409 | ptr += 1; | |
1410 | } | |
1411 | } | |
1412 | ||
1413 | if (mpext->use_map) { | |
1414 | put_unaligned_be64(mpext->data_seq, ptr); | |
1415 | ptr += 2; | |
1416 | put_unaligned_be32(mpext->subflow_seq, ptr); | |
1417 | ptr += 1; | |
1418 | if (opts->csum_reqd) { | |
1e39e5a3 GT |
1419 | /* data_len == 0 is reserved for the infinite mapping, |
1420 | * the checksum will also be set to 0. | |
1421 | */ | |
ba2c89e0 | 1422 | put_len_csum(mpext->data_len, |
d7e6f583 | 1423 | (mpext->data_len ? mptcp_make_csum(mpext) : 0), |
ba2c89e0 | 1424 | ptr); |
1bff1e43 PA |
1425 | } else { |
1426 | put_unaligned_be32(mpext->data_len << 16 | | |
1427 | TCPOPT_NOP << 8 | TCPOPT_NOP, ptr); | |
1428 | } | |
110b6d1f | 1429 | ptr += 1; |
1bff1e43 | 1430 | } |
902c8f86 MB |
1431 | |
1432 | /* We might need to add MP_FAIL options in rare cases */ | |
1433 | if (unlikely(OPTION_MPTCP_FAIL & opts->suboptions)) | |
1434 | goto mp_fail; | |
13ac17a3 | 1435 | } else if (OPTIONS_MPTCP_MPC & opts->suboptions) { |
06fe1719 | 1436 | u8 len, flag = MPTCP_CAP_HMAC_SHA256; |
eda7acdd | 1437 | |
c94b1f96 | 1438 | if (OPTION_MPTCP_MPC_SYN & opts->suboptions) { |
eda7acdd | 1439 | len = TCPOLEN_MPTCP_MPC_SYN; |
c94b1f96 | 1440 | } else if (OPTION_MPTCP_MPC_SYNACK & opts->suboptions) { |
cec37a6e | 1441 | len = TCPOLEN_MPTCP_MPC_SYNACK; |
f7cc8890 | 1442 | } else if (opts->data_len) { |
cc7972ea | 1443 | len = TCPOLEN_MPTCP_MPC_ACK_DATA; |
c94b1f96 GT |
1444 | if (opts->csum_reqd) |
1445 | len += TCPOLEN_MPTCP_DSS_CHECKSUM; | |
1446 | } else { | |
eda7acdd | 1447 | len = TCPOLEN_MPTCP_MPC_ACK; |
c94b1f96 | 1448 | } |
eda7acdd | 1449 | |
06fe1719 GT |
1450 | if (opts->csum_reqd) |
1451 | flag |= MPTCP_CAP_CHECKSUM_REQD; | |
1452 | ||
bab6b88e GT |
1453 | if (!opts->allow_join_id0) |
1454 | flag |= MPTCP_CAP_DENY_JOIN_ID0; | |
1455 | ||
3df523ab PK |
1456 | *ptr++ = mptcp_option(MPTCPOPT_MP_CAPABLE, len, |
1457 | MPTCP_SUPPORTED_VERSION, | |
06fe1719 | 1458 | flag); |
cc7972ea CP |
1459 | |
1460 | if (!((OPTION_MPTCP_MPC_SYNACK | OPTION_MPTCP_MPC_ACK) & | |
1461 | opts->suboptions)) | |
1462 | goto mp_capable_done; | |
1463 | ||
eda7acdd PK |
1464 | put_unaligned_be64(opts->sndr_key, ptr); |
1465 | ptr += 2; | |
cc7972ea CP |
1466 | if (!((OPTION_MPTCP_MPC_ACK) & opts->suboptions)) |
1467 | goto mp_capable_done; | |
1468 | ||
1469 | put_unaligned_be64(opts->rcvr_key, ptr); | |
1470 | ptr += 2; | |
f7cc8890 | 1471 | if (!opts->data_len) |
cc7972ea CP |
1472 | goto mp_capable_done; |
1473 | ||
c94b1f96 | 1474 | if (opts->csum_reqd) { |
ba2c89e0 PA |
1475 | put_len_csum(opts->data_len, |
1476 | __mptcp_make_csum(opts->data_seq, | |
1477 | opts->subflow_seq, | |
1478 | opts->data_len, | |
1479 | ~csum_unfold(opts->csum)), | |
1480 | ptr); | |
c94b1f96 | 1481 | } else { |
f7cc8890 | 1482 | put_unaligned_be32(opts->data_len << 16 | |
c94b1f96 GT |
1483 | TCPOPT_NOP << 8 | TCPOPT_NOP, ptr); |
1484 | } | |
cc7972ea | 1485 | ptr += 1; |
6d0060f6 | 1486 | |
1bff1e43 PA |
1487 | /* MPC is additionally mutually exclusive with MP_PRIO */ |
1488 | goto mp_capable_done; | |
71b077e4 PA |
1489 | } else if (OPTIONS_MPTCP_MPJ & opts->suboptions) { |
1490 | if (OPTION_MPTCP_MPJ_SYN & opts->suboptions) { | |
1491 | *ptr++ = mptcp_option(MPTCPOPT_MP_JOIN, | |
1492 | TCPOLEN_MPTCP_MPJ_SYN, | |
1493 | opts->backup, opts->join_id); | |
1494 | put_unaligned_be32(opts->token, ptr); | |
1495 | ptr += 1; | |
1496 | put_unaligned_be32(opts->nonce, ptr); | |
1497 | ptr += 1; | |
1498 | } else if (OPTION_MPTCP_MPJ_SYNACK & opts->suboptions) { | |
1499 | *ptr++ = mptcp_option(MPTCPOPT_MP_JOIN, | |
1500 | TCPOLEN_MPTCP_MPJ_SYNACK, | |
1501 | opts->backup, opts->join_id); | |
1502 | put_unaligned_be64(opts->thmac, ptr); | |
1503 | ptr += 2; | |
1504 | put_unaligned_be32(opts->nonce, ptr); | |
1505 | ptr += 1; | |
1506 | } else { | |
1507 | *ptr++ = mptcp_option(MPTCPOPT_MP_JOIN, | |
1508 | TCPOLEN_MPTCP_MPJ_ACK, 0, 0); | |
1509 | memcpy(ptr, opts->hmac, MPTCPOPT_HMAC_LEN); | |
1510 | ptr += 5; | |
1511 | } | |
1bff1e43 | 1512 | } else if (OPTION_MPTCP_ADD_ADDR & opts->suboptions) { |
6eb3d1e3 GT |
1513 | u8 len = TCPOLEN_MPTCP_ADD_ADDR_BASE; |
1514 | u8 echo = MPTCP_ADDR_ECHO; | |
1515 | ||
e1ef6832 | 1516 | #if IS_ENABLED(CONFIG_MPTCP_IPV6) |
fef6b7ec | 1517 | if (opts->addr.family == AF_INET6) |
e1ef6832 GT |
1518 | len = TCPOLEN_MPTCP_ADD_ADDR6_BASE; |
1519 | #endif | |
1520 | ||
30f60bae | 1521 | if (opts->addr.port) |
22fb85ff GT |
1522 | len += TCPOLEN_MPTCP_PORT_LEN; |
1523 | ||
3df523ab | 1524 | if (opts->ahmac) { |
6eb3d1e3 GT |
1525 | len += sizeof(opts->ahmac); |
1526 | echo = 0; | |
3df523ab | 1527 | } |
3df523ab | 1528 | |
6eb3d1e3 | 1529 | *ptr++ = mptcp_option(MPTCPOPT_ADD_ADDR, |
30f60bae | 1530 | len, echo, opts->addr.id); |
fef6b7ec | 1531 | if (opts->addr.family == AF_INET) { |
30f60bae | 1532 | memcpy((u8 *)ptr, (u8 *)&opts->addr.addr.s_addr, 4); |
e1ef6832 | 1533 | ptr += 1; |
3df523ab | 1534 | } |
3df523ab | 1535 | #if IS_ENABLED(CONFIG_MPTCP_IPV6) |
fef6b7ec | 1536 | else if (opts->addr.family == AF_INET6) { |
30f60bae | 1537 | memcpy((u8 *)ptr, opts->addr.addr6.s6_addr, 16); |
e1ef6832 | 1538 | ptr += 4; |
3df523ab | 1539 | } |
3df523ab PK |
1540 | #endif |
1541 | ||
30f60bae | 1542 | if (!opts->addr.port) { |
22fb85ff GT |
1543 | if (opts->ahmac) { |
1544 | put_unaligned_be64(opts->ahmac, ptr); | |
1545 | ptr += 2; | |
1546 | } | |
1547 | } else { | |
30f60bae GT |
1548 | u16 port = ntohs(opts->addr.port); |
1549 | ||
22fb85ff GT |
1550 | if (opts->ahmac) { |
1551 | u8 *bptr = (u8 *)ptr; | |
1552 | ||
30f60bae | 1553 | put_unaligned_be16(port, bptr); |
22fb85ff GT |
1554 | bptr += 2; |
1555 | put_unaligned_be64(opts->ahmac, bptr); | |
1556 | bptr += 8; | |
1557 | put_unaligned_be16(TCPOPT_NOP << 8 | | |
1558 | TCPOPT_NOP, bptr); | |
1559 | ||
1560 | ptr += 3; | |
1561 | } else { | |
30f60bae | 1562 | put_unaligned_be32(port << 16 | |
22fb85ff GT |
1563 | TCPOPT_NOP << 8 | |
1564 | TCPOPT_NOP, ptr); | |
1565 | ptr += 1; | |
1566 | } | |
3df523ab | 1567 | } |
f284c0c7 PA |
1568 | } else if (unlikely(OPTION_MPTCP_FASTCLOSE & opts->suboptions)) { |
1569 | /* FASTCLOSE is mutually exclusive with others except RST */ | |
1570 | *ptr++ = mptcp_option(MPTCPOPT_MP_FASTCLOSE, | |
1571 | TCPOLEN_MPTCP_FASTCLOSE, | |
1572 | 0, 0); | |
1573 | put_unaligned_be64(opts->rcvr_key, ptr); | |
1574 | ptr += 2; | |
1575 | ||
902c8f86 MB |
1576 | if (OPTION_MPTCP_RST & opts->suboptions) |
1577 | goto mp_rst; | |
1578 | return; | |
1579 | } else if (unlikely(OPTION_MPTCP_FAIL & opts->suboptions)) { | |
1580 | mp_fail: | |
1581 | /* MP_FAIL is mutually exclusive with others except RST */ | |
1582 | subflow = mptcp_subflow_ctx(ssk); | |
1583 | subflow->send_mp_fail = 0; | |
1584 | ||
1585 | *ptr++ = mptcp_option(MPTCPOPT_MP_FAIL, | |
1586 | TCPOLEN_MPTCP_FAIL, | |
1587 | 0, 0); | |
1588 | put_unaligned_be64(opts->fail_seq, ptr); | |
1589 | ptr += 2; | |
1590 | ||
f284c0c7 PA |
1591 | if (OPTION_MPTCP_RST & opts->suboptions) |
1592 | goto mp_rst; | |
1593 | return; | |
1594 | } else if (unlikely(OPTION_MPTCP_RST & opts->suboptions)) { | |
1595 | mp_rst: | |
1596 | *ptr++ = mptcp_option(MPTCPOPT_RST, | |
1597 | TCPOLEN_MPTCP_RST, | |
1598 | opts->reset_transient, | |
1599 | opts->reset_reason); | |
1600 | return; | |
3df523ab | 1601 | } |
3df523ab | 1602 | |
1bff1e43 | 1603 | if (OPTION_MPTCP_PRIO & opts->suboptions) { |
1bff1e43 PA |
1604 | subflow = mptcp_subflow_ctx(ssk); |
1605 | subflow->send_mp_prio = 0; | |
1606 | ||
1607 | *ptr++ = mptcp_option(MPTCPOPT_MP_PRIO, | |
1608 | TCPOLEN_MPTCP_PRIO, | |
1609 | opts->backup, TCPOPT_NOP); | |
c21b50d5 | 1610 | |
3c976f4c | 1611 | MPTCP_INC_STATS(sock_net(ssk), MPTCP_MIB_MPPRIOTX); |
1bff1e43 PA |
1612 | } |
1613 | ||
1614 | mp_capable_done: | |
3df523ab | 1615 | if (OPTION_MPTCP_RM_ADDR & opts->suboptions) { |
6445e17a GT |
1616 | u8 i = 1; |
1617 | ||
3df523ab | 1618 | *ptr++ = mptcp_option(MPTCPOPT_RM_ADDR, |
6445e17a GT |
1619 | TCPOLEN_MPTCP_RM_ADDR_BASE + opts->rm_list.nr, |
1620 | 0, opts->rm_list.ids[0]); | |
1621 | ||
1622 | while (i < opts->rm_list.nr) { | |
1623 | u8 id1, id2, id3, id4; | |
1624 | ||
1625 | id1 = opts->rm_list.ids[i]; | |
1626 | id2 = i + 1 < opts->rm_list.nr ? opts->rm_list.ids[i + 1] : TCPOPT_NOP; | |
1627 | id3 = i + 2 < opts->rm_list.nr ? opts->rm_list.ids[i + 2] : TCPOPT_NOP; | |
1628 | id4 = i + 3 < opts->rm_list.nr ? opts->rm_list.ids[i + 3] : TCPOPT_NOP; | |
1629 | put_unaligned_be32(id1 << 24 | id2 << 16 | id3 << 8 | id4, ptr); | |
1630 | ptr += 1; | |
1631 | i += 4; | |
1632 | } | |
3df523ab PK |
1633 | } |
1634 | ||
fa3fe2b1 | 1635 | if (tp) |
f3589be0 | 1636 | mptcp_set_rwin(tp, th); |
eda7acdd | 1637 | } |
dc87efdb FW |
1638 | |
1639 | __be32 mptcp_get_reset_option(const struct sk_buff *skb) | |
1640 | { | |
1641 | const struct mptcp_ext *ext = mptcp_get_ext(skb); | |
1642 | u8 flags, reason; | |
1643 | ||
1644 | if (ext) { | |
1645 | flags = ext->reset_transient; | |
1646 | reason = ext->reset_reason; | |
1647 | ||
1648 | return mptcp_option(MPTCPOPT_RST, TCPOLEN_MPTCP_RST, | |
1649 | flags, reason); | |
1650 | } | |
1651 | ||
1652 | return htonl(0u); | |
1653 | } | |
1654 | EXPORT_SYMBOL_GPL(mptcp_get_reset_option); |