1 /* 2 * net/dccp/ackvec.c 3 * 4 * An implementation of the DCCP protocol 5 * Copyright (c) 2005 Arnaldo Carvalho de Melo <acme@ghostprotocols.net> 6 * 7 * This program is free software; you can redistribute it and/or modify it 8 * under the terms of the GNU General Public License as published by the 9 * Free Software Foundation; version 2 of the License; 10 */ 11 12 #include "ackvec.h" 13 #include "dccp.h" 14 15 #include <linux/init.h> 16 #include <linux/errno.h> 17 #include <linux/kernel.h> 18 #include <linux/skbuff.h> 19 #include <linux/slab.h> 20 21 #include <net/sock.h> 22 23 static struct kmem_cache *dccp_ackvec_slab; 24 static struct kmem_cache *dccp_ackvec_record_slab; 25 26 static struct dccp_ackvec_record *dccp_ackvec_record_new(void) 27 { 28 struct dccp_ackvec_record *avr = 29 kmem_cache_alloc(dccp_ackvec_record_slab, GFP_ATOMIC); 30 31 if (avr != NULL) 32 INIT_LIST_HEAD(&avr->avr_node); 33 34 return avr; 35 } 36 37 static void dccp_ackvec_record_delete(struct dccp_ackvec_record *avr) 38 { 39 if (unlikely(avr == NULL)) 40 return; 41 /* Check if deleting a linked record */ 42 WARN_ON(!list_empty(&avr->avr_node)); 43 kmem_cache_free(dccp_ackvec_record_slab, avr); 44 } 45 46 static void dccp_ackvec_insert_avr(struct dccp_ackvec *av, 47 struct dccp_ackvec_record *avr) 48 { 49 /* 50 * AVRs are sorted by seqno. Since we are sending them in order, we 51 * just add the AVR at the head of the list. 52 * -sorbo. 53 */ 54 if (!list_empty(&av->av_records)) { 55 const struct dccp_ackvec_record *head = 56 list_entry(av->av_records.next, 57 struct dccp_ackvec_record, 58 avr_node); 59 BUG_ON(before48(avr->avr_ack_seqno, head->avr_ack_seqno)); 60 } 61 62 list_add(&avr->avr_node, &av->av_records); 63 } 64 65 int dccp_insert_option_ackvec(struct sock *sk, struct sk_buff *skb) 66 { 67 struct dccp_sock *dp = dccp_sk(sk); 68 struct dccp_ackvec *av = dp->dccps_hc_rx_ackvec; 69 /* Figure out how many options do we need to represent the ackvec */ 70 const u8 nr_opts = DIV_ROUND_UP(av->av_vec_len, DCCP_SINGLE_OPT_MAXLEN); 71 u16 len = av->av_vec_len + 2 * nr_opts, i; 72 u32 elapsed_time; 73 const unsigned char *tail, *from; 74 unsigned char *to; 75 struct dccp_ackvec_record *avr; 76 suseconds_t delta; 77 78 if (DCCP_SKB_CB(skb)->dccpd_opt_len + len > DCCP_MAX_OPT_LEN) 79 return -1; 80 81 delta = ktime_us_delta(ktime_get_real(), av->av_time); 82 elapsed_time = delta / 10; 83 84 if (elapsed_time != 0 && 85 dccp_insert_option_elapsed_time(sk, skb, elapsed_time)) 86 return -1; 87 88 avr = dccp_ackvec_record_new(); 89 if (avr == NULL) 90 return -1; 91 92 DCCP_SKB_CB(skb)->dccpd_opt_len += len; 93 94 to = skb_push(skb, len); 95 len = av->av_vec_len; 96 from = av->av_buf + av->av_buf_head; 97 tail = av->av_buf + DCCP_MAX_ACKVEC_LEN; 98 99 for (i = 0; i < nr_opts; ++i) { 100 int copylen = len; 101 102 if (len > DCCP_SINGLE_OPT_MAXLEN) 103 copylen = DCCP_SINGLE_OPT_MAXLEN; 104 105 *to++ = DCCPO_ACK_VECTOR_0; 106 *to++ = copylen + 2; 107 108 /* Check if buf_head wraps */ 109 if (from + copylen > tail) { 110 const u16 tailsize = tail - from; 111 112 memcpy(to, from, tailsize); 113 to += tailsize; 114 len -= tailsize; 115 copylen -= tailsize; 116 from = av->av_buf; 117 } 118 119 memcpy(to, from, copylen); 120 from += copylen; 121 to += copylen; 122 len -= copylen; 123 } 124 125 /* 126 * From RFC 4340, A.2: 127 * 128 * For each acknowledgement it sends, the HC-Receiver will add an 129 * acknowledgement record. ack_seqno will equal the HC-Receiver 130 * sequence number it used for the ack packet; ack_ptr will equal 131 * buf_head; ack_ackno will equal buf_ackno; and ack_nonce will 132 * equal buf_nonce. 133 */ 134 avr->avr_ack_seqno = DCCP_SKB_CB(skb)->dccpd_seq; 135 avr->avr_ack_ptr = av->av_buf_head; 136 avr->avr_ack_ackno = av->av_buf_ackno; 137 avr->avr_ack_nonce = av->av_buf_nonce; 138 avr->avr_sent_len = av->av_vec_len; 139 140 dccp_ackvec_insert_avr(av, avr); 141 142 dccp_pr_debug("%s ACK Vector 0, len=%d, ack_seqno=%llu, " 143 "ack_ackno=%llu\n", 144 dccp_role(sk), avr->avr_sent_len, 145 (unsigned long long)avr->avr_ack_seqno, 146 (unsigned long long)avr->avr_ack_ackno); 147 return 0; 148 } 149 150 struct dccp_ackvec *dccp_ackvec_alloc(const gfp_t priority) 151 { 152 struct dccp_ackvec *av = kmem_cache_alloc(dccp_ackvec_slab, priority); 153 154 if (av != NULL) { 155 av->av_buf_head = DCCP_MAX_ACKVEC_LEN - 1; 156 av->av_buf_ackno = UINT48_MAX + 1; 157 av->av_buf_nonce = 0; 158 av->av_time = ktime_set(0, 0); 159 av->av_vec_len = 0; 160 INIT_LIST_HEAD(&av->av_records); 161 } 162 163 return av; 164 } 165 166 void dccp_ackvec_free(struct dccp_ackvec *av) 167 { 168 if (unlikely(av == NULL)) 169 return; 170 171 if (!list_empty(&av->av_records)) { 172 struct dccp_ackvec_record *avr, *next; 173 174 list_for_each_entry_safe(avr, next, &av->av_records, avr_node) { 175 list_del_init(&avr->avr_node); 176 dccp_ackvec_record_delete(avr); 177 } 178 } 179 180 kmem_cache_free(dccp_ackvec_slab, av); 181 } 182 183 static inline u8 dccp_ackvec_state(const struct dccp_ackvec *av, 184 const u32 index) 185 { 186 return av->av_buf[index] & DCCP_ACKVEC_STATE_MASK; 187 } 188 189 static inline u8 dccp_ackvec_len(const struct dccp_ackvec *av, 190 const u32 index) 191 { 192 return av->av_buf[index] & DCCP_ACKVEC_LEN_MASK; 193 } 194 195 /* 196 * If several packets are missing, the HC-Receiver may prefer to enter multiple 197 * bytes with run length 0, rather than a single byte with a larger run length; 198 * this simplifies table updates if one of the missing packets arrives. 199 */ 200 static inline int dccp_ackvec_set_buf_head_state(struct dccp_ackvec *av, 201 const unsigned int packets, 202 const unsigned char state) 203 { 204 unsigned int gap; 205 long new_head; 206 207 if (av->av_vec_len + packets > DCCP_MAX_ACKVEC_LEN) 208 return -ENOBUFS; 209 210 gap = packets - 1; 211 new_head = av->av_buf_head - packets; 212 213 if (new_head < 0) { 214 if (gap > 0) { 215 memset(av->av_buf, DCCP_ACKVEC_STATE_NOT_RECEIVED, 216 gap + new_head + 1); 217 gap = -new_head; 218 } 219 new_head += DCCP_MAX_ACKVEC_LEN; 220 } 221 222 av->av_buf_head = new_head; 223 224 if (gap > 0) 225 memset(av->av_buf + av->av_buf_head + 1, 226 DCCP_ACKVEC_STATE_NOT_RECEIVED, gap); 227 228 av->av_buf[av->av_buf_head] = state; 229 av->av_vec_len += packets; 230 return 0; 231 } 232 233 /* 234 * Implements the RFC 4340, Appendix A 235 */ 236 int dccp_ackvec_add(struct dccp_ackvec *av, const struct sock *sk, 237 const u64 ackno, const u8 state) 238 { 239 /* 240 * Check at the right places if the buffer is full, if it is, tell the 241 * caller to start dropping packets till the HC-Sender acks our ACK 242 * vectors, when we will free up space in av_buf. 243 * 244 * We may well decide to do buffer compression, etc, but for now lets 245 * just drop. 246 * 247 * From Appendix A.1.1 (`New Packets'): 248 * 249 * Of course, the circular buffer may overflow, either when the 250 * HC-Sender is sending data at a very high rate, when the 251 * HC-Receiver's acknowledgements are not reaching the HC-Sender, 252 * or when the HC-Sender is forgetting to acknowledge those acks 253 * (so the HC-Receiver is unable to clean up old state). In this 254 * case, the HC-Receiver should either compress the buffer (by 255 * increasing run lengths when possible), transfer its state to 256 * a larger buffer, or, as a last resort, drop all received 257 * packets, without processing them whatsoever, until its buffer 258 * shrinks again. 259 */ 260 261 /* See if this is the first ackno being inserted */ 262 if (av->av_vec_len == 0) { 263 av->av_buf[av->av_buf_head] = state; 264 av->av_vec_len = 1; 265 } else if (after48(ackno, av->av_buf_ackno)) { 266 const u64 delta = dccp_delta_seqno(av->av_buf_ackno, ackno); 267 268 /* 269 * Look if the state of this packet is the same as the 270 * previous ackno and if so if we can bump the head len. 271 */ 272 if (delta == 1 && 273 dccp_ackvec_state(av, av->av_buf_head) == state && 274 dccp_ackvec_len(av, av->av_buf_head) < DCCP_ACKVEC_LEN_MASK) 275 av->av_buf[av->av_buf_head]++; 276 else if (dccp_ackvec_set_buf_head_state(av, delta, state)) 277 return -ENOBUFS; 278 } else { 279 /* 280 * A.1.2. Old Packets 281 * 282 * When a packet with Sequence Number S <= buf_ackno 283 * arrives, the HC-Receiver will scan the table for 284 * the byte corresponding to S. (Indexing structures 285 * could reduce the complexity of this scan.) 286 */ 287 u64 delta = dccp_delta_seqno(ackno, av->av_buf_ackno); 288 u32 index = av->av_buf_head; 289 290 while (1) { 291 const u8 len = dccp_ackvec_len(av, index); 292 const u8 av_state = dccp_ackvec_state(av, index); 293 /* 294 * valid packets not yet in av_buf have a reserved 295 * entry, with a len equal to 0. 296 */ 297 if (av_state == DCCP_ACKVEC_STATE_NOT_RECEIVED && 298 len == 0 && delta == 0) { /* Found our 299 reserved seat! */ 300 dccp_pr_debug("Found %llu reserved seat!\n", 301 (unsigned long long)ackno); 302 av->av_buf[index] = state; 303 goto out; 304 } 305 /* len == 0 means one packet */ 306 if (delta < len + 1) 307 goto out_duplicate; 308 309 delta -= len + 1; 310 if (++index == DCCP_MAX_ACKVEC_LEN) 311 index = 0; 312 } 313 } 314 315 av->av_buf_ackno = ackno; 316 av->av_time = ktime_get_real(); 317 out: 318 return 0; 319 320 out_duplicate: 321 /* Duplicate packet */ 322 dccp_pr_debug("Received a dup or already considered lost " 323 "packet: %llu\n", (unsigned long long)ackno); 324 return -EILSEQ; 325 } 326 327 static void dccp_ackvec_throw_record(struct dccp_ackvec *av, 328 struct dccp_ackvec_record *avr) 329 { 330 struct dccp_ackvec_record *next; 331 332 /* sort out vector length */ 333 if (av->av_buf_head <= avr->avr_ack_ptr) 334 av->av_vec_len = avr->avr_ack_ptr - av->av_buf_head; 335 else 336 av->av_vec_len = DCCP_MAX_ACKVEC_LEN - 1 - 337 av->av_buf_head + avr->avr_ack_ptr; 338 339 /* free records */ 340 list_for_each_entry_safe_from(avr, next, &av->av_records, avr_node) { 341 list_del_init(&avr->avr_node); 342 dccp_ackvec_record_delete(avr); 343 } 344 } 345 346 void dccp_ackvec_check_rcv_ackno(struct dccp_ackvec *av, struct sock *sk, 347 const u64 ackno) 348 { 349 struct dccp_ackvec_record *avr; 350 351 /* 352 * If we traverse backwards, it should be faster when we have large 353 * windows. We will be receiving ACKs for stuff we sent a while back 354 * -sorbo. 355 */ 356 list_for_each_entry_reverse(avr, &av->av_records, avr_node) { 357 if (ackno == avr->avr_ack_seqno) { 358 dccp_pr_debug("%s ACK packet 0, len=%d, ack_seqno=%llu, " 359 "ack_ackno=%llu, ACKED!\n", 360 dccp_role(sk), 1, 361 (unsigned long long)avr->avr_ack_seqno, 362 (unsigned long long)avr->avr_ack_ackno); 363 dccp_ackvec_throw_record(av, avr); 364 break; 365 } else if (avr->avr_ack_seqno > ackno) 366 break; /* old news */ 367 } 368 } 369 370 static void dccp_ackvec_check_rcv_ackvector(struct dccp_ackvec *av, 371 struct sock *sk, u64 *ackno, 372 const unsigned char len, 373 const unsigned char *vector) 374 { 375 unsigned char i; 376 struct dccp_ackvec_record *avr; 377 378 /* Check if we actually sent an ACK vector */ 379 if (list_empty(&av->av_records)) 380 return; 381 382 i = len; 383 /* 384 * XXX 385 * I think it might be more efficient to work backwards. See comment on 386 * rcv_ackno. -sorbo. 387 */ 388 avr = list_entry(av->av_records.next, struct dccp_ackvec_record, avr_node); 389 while (i--) { 390 const u8 rl = *vector & DCCP_ACKVEC_LEN_MASK; 391 u64 ackno_end_rl; 392 393 dccp_set_seqno(&ackno_end_rl, *ackno - rl); 394 395 /* 396 * If our AVR sequence number is greater than the ack, go 397 * forward in the AVR list until it is not so. 398 */ 399 list_for_each_entry_from(avr, &av->av_records, avr_node) { 400 if (!after48(avr->avr_ack_seqno, *ackno)) 401 goto found; 402 } 403 /* End of the av_records list, not found, exit */ 404 break; 405 found: 406 if (between48(avr->avr_ack_seqno, ackno_end_rl, *ackno)) { 407 const u8 state = *vector & DCCP_ACKVEC_STATE_MASK; 408 if (state != DCCP_ACKVEC_STATE_NOT_RECEIVED) { 409 dccp_pr_debug("%s ACK vector 0, len=%d, " 410 "ack_seqno=%llu, ack_ackno=%llu, " 411 "ACKED!\n", 412 dccp_role(sk), len, 413 (unsigned long long) 414 avr->avr_ack_seqno, 415 (unsigned long long) 416 avr->avr_ack_ackno); 417 dccp_ackvec_throw_record(av, avr); 418 break; 419 } 420 /* 421 * If it wasn't received, continue scanning... we might 422 * find another one. 423 */ 424 } 425 426 dccp_set_seqno(ackno, ackno_end_rl - 1); 427 ++vector; 428 } 429 } 430 431 int dccp_ackvec_parse(struct sock *sk, const struct sk_buff *skb, 432 u64 *ackno, const u8 opt, const u8 *value, const u8 len) 433 { 434 if (len > DCCP_SINGLE_OPT_MAXLEN) 435 return -1; 436 437 /* dccp_ackvector_print(DCCP_SKB_CB(skb)->dccpd_ack_seq, value, len); */ 438 dccp_ackvec_check_rcv_ackvector(dccp_sk(sk)->dccps_hc_rx_ackvec, sk, 439 ackno, len, value); 440 return 0; 441 } 442 443 int __init dccp_ackvec_init(void) 444 { 445 dccp_ackvec_slab = kmem_cache_create("dccp_ackvec", 446 sizeof(struct dccp_ackvec), 0, 447 SLAB_HWCACHE_ALIGN, NULL); 448 if (dccp_ackvec_slab == NULL) 449 goto out_err; 450 451 dccp_ackvec_record_slab = 452 kmem_cache_create("dccp_ackvec_record", 453 sizeof(struct dccp_ackvec_record), 454 0, SLAB_HWCACHE_ALIGN, NULL); 455 if (dccp_ackvec_record_slab == NULL) 456 goto out_destroy_slab; 457 458 return 0; 459 460 out_destroy_slab: 461 kmem_cache_destroy(dccp_ackvec_slab); 462 dccp_ackvec_slab = NULL; 463 out_err: 464 DCCP_CRIT("Unable to create Ack Vector slab cache"); 465 return -ENOBUFS; 466 } 467 468 void dccp_ackvec_exit(void) 469 { 470 if (dccp_ackvec_slab != NULL) { 471 kmem_cache_destroy(dccp_ackvec_slab); 472 dccp_ackvec_slab = NULL; 473 } 474 if (dccp_ackvec_record_slab != NULL) { 475 kmem_cache_destroy(dccp_ackvec_record_slab); 476 dccp_ackvec_record_slab = NULL; 477 } 478 } 479