net/dccp/ackvec.c

   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/dccp.h>
  16 #include <linux/init.h>
  17 #include <linux/errno.h>
  18 #include <linux/kernel.h>
  19 #include <linux/skbuff.h>
  20 #include <linux/slab.h>
  21
  22 #include <net/sock.h>
  23
  24 static struct kmem_cache *dccp_ackvec_slab;
  25 static struct kmem_cache *dccp_ackvec_record_slab;
  26
  27 static struct dccp_ackvec_record *dccp_ackvec_record_new(void)
  28 {
  29         struct dccp_ackvec_record *avr =
  30                         kmem_cache_alloc(dccp_ackvec_record_slab, GFP_ATOMIC);
  31
  32         if (avr != NULL)
  33                 INIT_LIST_HEAD(&avr->dccpavr_node);
  34
  35         return avr;
  36 }
  37
  38 static void dccp_ackvec_record_delete(struct dccp_ackvec_record *avr)
  39 {
  40         if (unlikely(avr == NULL))
  41                 return;
  42         /* Check if deleting a linked record */
  43         WARN_ON(!list_empty(&avr->dccpavr_node));
  44         kmem_cache_free(dccp_ackvec_record_slab, avr);
  45 }
  46
  47 static void dccp_ackvec_insert_avr(struct dccp_ackvec *av,
  48                                    struct dccp_ackvec_record *avr)
  49 {
  50         /*
  51          * AVRs are sorted by seqno. Since we are sending them in order, we
  52          * just add the AVR at the head of the list.
  53          * -sorbo.
  54          */
  55         if (!list_empty(&av->dccpav_records)) {
  56                 const struct dccp_ackvec_record *head =
  57                                         list_entry(av->dccpav_records.next,
  58                                                    struct dccp_ackvec_record,
  59                                                    dccpavr_node);
  60                 BUG_ON(before48(avr->dccpavr_ack_seqno,
  61                                 head->dccpavr_ack_seqno));
  62         }
  63
  64         list_add(&avr->dccpavr_node, &av->dccpav_records);
  65 }
  66
  67 int dccp_insert_option_ackvec(struct sock *sk, struct sk_buff *skb)
  68 {
  69         struct dccp_sock *dp = dccp_sk(sk);
  70         struct dccp_ackvec *av = dp->dccps_hc_rx_ackvec;
  71         /* Figure out how many options do we need to represent the ackvec */
  72         const u16 nr_opts = (av->dccpav_vec_len +
  73                              DCCP_MAX_ACKVEC_OPT_LEN - 1) /
  74                             DCCP_MAX_ACKVEC_OPT_LEN;
  75         u16 len = av->dccpav_vec_len + 2 * nr_opts, i;
  76         u32 elapsed_time;
  77         const unsigned char *tail, *from;
  78         unsigned char *to;
  79         struct dccp_ackvec_record *avr;
  80         suseconds_t delta;
  81
  82         if (DCCP_SKB_CB(skb)->dccpd_opt_len + len > DCCP_MAX_OPT_LEN)
  83                 return -1;
  84
  85         delta = ktime_us_delta(ktime_get_real(), av->dccpav_time);
  86         elapsed_time = delta / 10;
  87
  88         if (elapsed_time != 0 &&
  89             dccp_insert_option_elapsed_time(sk, skb, elapsed_time))
  90                 return -1;
  91
  92         avr = dccp_ackvec_record_new();
  93         if (avr == NULL)
  94                 return -1;
  95
  96         DCCP_SKB_CB(skb)->dccpd_opt_len += len;
  97
  98         to   = skb_push(skb, len);
  99         len  = av->dccpav_vec_len;
 100         from = av->dccpav_buf + av->dccpav_buf_head;
 101         tail = av->dccpav_buf + DCCP_MAX_ACKVEC_LEN;
 102
 103         for (i = 0; i < nr_opts; ++i) {
 104                 int copylen = len;
 105
 106                 if (len > DCCP_MAX_ACKVEC_OPT_LEN)
 107                         copylen = DCCP_MAX_ACKVEC_OPT_LEN;
 108
 109                 *to++ = DCCPO_ACK_VECTOR_0;
 110                 *to++ = copylen + 2;
 111
 112                 /* Check if buf_head wraps */
 113                 if (from + copylen > tail) {
 114                         const u16 tailsize = tail - from;
 115
 116                         memcpy(to, from, tailsize);
 117                         to      += tailsize;
 118                         len     -= tailsize;
 119                         copylen -= tailsize;
 120                         from    = av->dccpav_buf;
 121                 }
 122
 123                 memcpy(to, from, copylen);
 124                 from += copylen;
 125                 to   += copylen;
 126                 len  -= copylen;
 127         }
 128
 129         /*
 130          *      From RFC 4340, A.2:
 131          *
 132          *      For each acknowledgement it sends, the HC-Receiver will add an
 133          *      acknowledgement record.  ack_seqno will equal the HC-Receiver
 134          *      sequence number it used for the ack packet; ack_ptr will equal
 135          *      buf_head; ack_ackno will equal buf_ackno; and ack_nonce will
 136          *      equal buf_nonce.
 137          */
 138         avr->dccpavr_ack_seqno = DCCP_SKB_CB(skb)->dccpd_seq;
 139         avr->dccpavr_ack_ptr   = av->dccpav_buf_head;
 140         avr->dccpavr_ack_ackno = av->dccpav_buf_ackno;
 141         avr->dccpavr_ack_nonce = av->dccpav_buf_nonce;
 142         avr->dccpavr_sent_len  = av->dccpav_vec_len;
 143
 144         dccp_ackvec_insert_avr(av, avr);
 145
 146         dccp_pr_debug("%s ACK Vector 0, len=%d, ack_seqno=%llu, "
 147                       "ack_ackno=%llu\n",
 148                       dccp_role(sk), avr->dccpavr_sent_len,
 149                       (unsigned long long)avr->dccpavr_ack_seqno,
 150                       (unsigned long long)avr->dccpavr_ack_ackno);
 151         return 0;
 152 }
 153
 154 struct dccp_ackvec *dccp_ackvec_alloc(const gfp_t priority)
 155 {
 156         struct dccp_ackvec *av = kmem_cache_alloc(dccp_ackvec_slab, priority);
 157
 158         if (av != NULL) {
 159                 av->dccpav_buf_head     = DCCP_MAX_ACKVEC_LEN - 1;
 160                 av->dccpav_buf_ackno    = UINT48_MAX + 1;
 161                 av->dccpav_buf_nonce = av->dccpav_buf_nonce = 0;
 162                 av->dccpav_time      = ktime_set(0, 0);
 163                 av->dccpav_vec_len      = 0;
 164                 INIT_LIST_HEAD(&av->dccpav_records);
 165         }
 166
 167         return av;
 168 }
 169
 170 void dccp_ackvec_free(struct dccp_ackvec *av)
 171 {
 172         if (unlikely(av == NULL))
 173                 return;
 174
 175         if (!list_empty(&av->dccpav_records)) {
 176                 struct dccp_ackvec_record *avr, *next;
 177
 178                 list_for_each_entry_safe(avr, next, &av->dccpav_records,
 179                                          dccpavr_node) {
 180                         list_del_init(&avr->dccpavr_node);
 181                         dccp_ackvec_record_delete(avr);
 182                 }
 183         }
 184
 185         kmem_cache_free(dccp_ackvec_slab, av);
 186 }
 187
 188 static inline u8 dccp_ackvec_state(const struct dccp_ackvec *av,
 189                                    const u32 index)
 190 {
 191         return av->dccpav_buf[index] & DCCP_ACKVEC_STATE_MASK;
 192 }
 193
 194 static inline u8 dccp_ackvec_len(const struct dccp_ackvec *av,
 195                                  const u32 index)
 196 {
 197         return av->dccpav_buf[index] & DCCP_ACKVEC_LEN_MASK;
 198 }
 199
 200 /*
 201  * If several packets are missing, the HC-Receiver may prefer to enter multiple
 202  * bytes with run length 0, rather than a single byte with a larger run length;
 203  * this simplifies table updates if one of the missing packets arrives.
 204  */
 205 static inline int dccp_ackvec_set_buf_head_state(struct dccp_ackvec *av,
 206                                                  const unsigned int packets,
 207                                                  const unsigned char state)
 208 {
 209         unsigned int gap;
 210         long new_head;
 211
 212         if (av->dccpav_vec_len + packets > DCCP_MAX_ACKVEC_LEN)
 213                 return -ENOBUFS;
 214
 215         gap      = packets - 1;
 216         new_head = av->dccpav_buf_head - packets;
 217
 218         if (new_head < 0) {
 219                 if (gap > 0) {
 220                         memset(av->dccpav_buf, DCCP_ACKVEC_STATE_NOT_RECEIVED,
 221                                gap + new_head + 1);
 222                         gap = -new_head;
 223                 }
 224                 new_head += DCCP_MAX_ACKVEC_LEN;
 225         }
 226
 227         av->dccpav_buf_head = new_head;
 228
 229         if (gap > 0)
 230                 memset(av->dccpav_buf + av->dccpav_buf_head + 1,
 231                        DCCP_ACKVEC_STATE_NOT_RECEIVED, gap);
 232
 233         av->dccpav_buf[av->dccpav_buf_head] = state;
 234         av->dccpav_vec_len += packets;
 235         return 0;
 236 }
 237
 238 /*
 239  * Implements the RFC 4340, Appendix A
 240  */
 241 int dccp_ackvec_add(struct dccp_ackvec *av, const struct sock *sk,
 242                     const u64 ackno, const u8 state)
 243 {
 244         /*
 245          * Check at the right places if the buffer is full, if it is, tell the
 246          * caller to start dropping packets till the HC-Sender acks our ACK
 247          * vectors, when we will free up space in dccpav_buf.
 248          *
 249          * We may well decide to do buffer compression, etc, but for now lets
 250          * just drop.
 251          *
 252          * From Appendix A.1.1 (`New Packets'):
 253          *
 254          *      Of course, the circular buffer may overflow, either when the
 255          *      HC-Sender is sending data at a very high rate, when the
 256          *      HC-Receiver's acknowledgements are not reaching the HC-Sender,
 257          *      or when the HC-Sender is forgetting to acknowledge those acks
 258          *      (so the HC-Receiver is unable to clean up old state). In this
 259          *      case, the HC-Receiver should either compress the buffer (by
 260          *      increasing run lengths when possible), transfer its state to
 261          *      a larger buffer, or, as a last resort, drop all received
 262          *      packets, without processing them whatsoever, until its buffer
 263          *      shrinks again.
 264          */
 265
 266         /* See if this is the first ackno being inserted */
 267         if (av->dccpav_vec_len == 0) {
 268                 av->dccpav_buf[av->dccpav_buf_head] = state;
 269                 av->dccpav_vec_len = 1;
 270         } else if (after48(ackno, av->dccpav_buf_ackno)) {
 271                 const u64 delta = dccp_delta_seqno(av->dccpav_buf_ackno,
 272                                                    ackno);
 273
 274                 /*
 275                  * Look if the state of this packet is the same as the
 276                  * previous ackno and if so if we can bump the head len.
 277                  */
 278                 if (delta == 1 &&
 279                     dccp_ackvec_state(av, av->dccpav_buf_head) == state &&
 280                     (dccp_ackvec_len(av, av->dccpav_buf_head) <
 281                      DCCP_ACKVEC_LEN_MASK))
 282                         av->dccpav_buf[av->dccpav_buf_head]++;
 283                 else if (dccp_ackvec_set_buf_head_state(av, delta, state))
 284                         return -ENOBUFS;
 285         } else {
 286                 /*
 287                  * A.1.2.  Old Packets
 288                  *
 289                  *      When a packet with Sequence Number S <= buf_ackno
 290                  *      arrives, the HC-Receiver will scan the table for
 291                  *      the byte corresponding to S. (Indexing structures
 292                  *      could reduce the complexity of this scan.)
 293                  */
 294                 u64 delta = dccp_delta_seqno(ackno, av->dccpav_buf_ackno);
 295                 u32 index = av->dccpav_buf_head;
 296
 297                 while (1) {
 298                         const u8 len = dccp_ackvec_len(av, index);
 299                         const u8 state = dccp_ackvec_state(av, index);
 300                         /*
 301                          * valid packets not yet in dccpav_buf have a reserved
 302                          * entry, with a len equal to 0.
 303                          */
 304                         if (state == DCCP_ACKVEC_STATE_NOT_RECEIVED &&
 305                             len == 0 && delta == 0) { /* Found our
 306                                                          reserved seat! */
 307                                 dccp_pr_debug("Found %llu reserved seat!\n",
 308                                               (unsigned long long)ackno);
 309                                 av->dccpav_buf[index] = state;
 310                                 goto out;
 311                         }
 312                         /* len == 0 means one packet */
 313                         if (delta < len + 1)
 314                                 goto out_duplicate;
 315
 316                         delta -= len + 1;
 317                         if (++index == DCCP_MAX_ACKVEC_LEN)
 318                                 index = 0;
 319                 }
 320         }
 321
 322         av->dccpav_buf_ackno = ackno;
 323         av->dccpav_time = ktime_get_real();
 324 out:
 325         return 0;
 326
 327 out_duplicate:
 328         /* Duplicate packet */
 329         dccp_pr_debug("Received a dup or already considered lost "
 330                       "packet: %llu\n", (unsigned long long)ackno);
 331         return -EILSEQ;
 332 }
 333
 334 #ifdef CONFIG_IP_DCCP_DEBUG
 335 void dccp_ackvector_print(const u64 ackno, const unsigned char *vector, int len)
 336 {
 337         dccp_pr_debug_cat("ACK vector len=%d, ackno=%llu |", len,
 338                          (unsigned long long)ackno);
 339
 340         while (len--) {
 341                 const u8 state = (*vector & DCCP_ACKVEC_STATE_MASK) >> 6;
 342                 const u8 rl = *vector & DCCP_ACKVEC_LEN_MASK;
 343
 344                 dccp_pr_debug_cat("%d,%d|", state, rl);
 345                 ++vector;
 346         }
 347
 348         dccp_pr_debug_cat("\n");
 349 }
 350
 351 void dccp_ackvec_print(const struct dccp_ackvec *av)
 352 {
 353         dccp_ackvector_print(av->dccpav_buf_ackno,
 354                              av->dccpav_buf + av->dccpav_buf_head,
 355                              av->dccpav_vec_len);
 356 }
 357 #endif
 358
 359 static void dccp_ackvec_throw_record(struct dccp_ackvec *av,
 360                                      struct dccp_ackvec_record *avr)
 361 {
 362         struct dccp_ackvec_record *next;
 363
 364         /* sort out vector length */
 365         if (av->dccpav_buf_head <= avr->dccpavr_ack_ptr)
 366                 av->dccpav_vec_len = avr->dccpavr_ack_ptr - av->dccpav_buf_head;
 367         else
 368                 av->dccpav_vec_len = DCCP_MAX_ACKVEC_LEN - 1
 369                                      - av->dccpav_buf_head
 370                                      + avr->dccpavr_ack_ptr;
 371
 372         /* free records */
 373         list_for_each_entry_safe_from(avr, next, &av->dccpav_records,
 374                                       dccpavr_node) {
 375                 list_del_init(&avr->dccpavr_node);
 376                 dccp_ackvec_record_delete(avr);
 377         }
 378 }
 379
 380 void dccp_ackvec_check_rcv_ackno(struct dccp_ackvec *av, struct sock *sk,
 381                                  const u64 ackno)
 382 {
 383         struct dccp_ackvec_record *avr;
 384
 385         /*
 386          * If we traverse backwards, it should be faster when we have large
 387          * windows. We will be receiving ACKs for stuff we sent a while back
 388          * -sorbo.
 389          */
 390         list_for_each_entry_reverse(avr, &av->dccpav_records, dccpavr_node) {
 391                 if (ackno == avr->dccpavr_ack_seqno) {
 392                         dccp_pr_debug("%s ACK packet 0, len=%d, ack_seqno=%llu, "
 393                                       "ack_ackno=%llu, ACKED!\n",
 394                                       dccp_role(sk), 1,
 395                                       (unsigned long long)avr->dccpavr_ack_seqno,
 396                                       (unsigned long long)avr->dccpavr_ack_ackno);
 397                         dccp_ackvec_throw_record(av, avr);
 398                         break;
 399                 } else if (avr->dccpavr_ack_seqno > ackno)
 400                         break; /* old news */
 401         }
 402 }
 403
 404 static void dccp_ackvec_check_rcv_ackvector(struct dccp_ackvec *av,
 405                                             struct sock *sk, u64 *ackno,
 406                                             const unsigned char len,
 407                                             const unsigned char *vector)
 408 {
 409         unsigned char i;
 410         struct dccp_ackvec_record *avr;
 411
 412         /* Check if we actually sent an ACK vector */
 413         if (list_empty(&av->dccpav_records))
 414                 return;
 415
 416         i = len;
 417         /*
 418          * XXX
 419          * I think it might be more efficient to work backwards. See comment on
 420          * rcv_ackno. -sorbo.
 421          */
 422         avr = list_entry(av->dccpav_records.next, struct dccp_ackvec_record,
 423                          dccpavr_node);
 424         while (i--) {
 425                 const u8 rl = *vector & DCCP_ACKVEC_LEN_MASK;
 426                 u64 ackno_end_rl;
 427
 428                 dccp_set_seqno(&ackno_end_rl, *ackno - rl);
 429
 430                 /*
 431                  * If our AVR sequence number is greater than the ack, go
 432                  * forward in the AVR list until it is not so.
 433                  */
 434                 list_for_each_entry_from(avr, &av->dccpav_records,
 435                                          dccpavr_node) {
 436                         if (!after48(avr->dccpavr_ack_seqno, *ackno))
 437                                 goto found;
 438                 }
 439                 /* End of the dccpav_records list, not found, exit */
 440                 break;
 441 found:
 442                 if (between48(avr->dccpavr_ack_seqno, ackno_end_rl, *ackno)) {
 443                         const u8 state = *vector & DCCP_ACKVEC_STATE_MASK;
 444                         if (state != DCCP_ACKVEC_STATE_NOT_RECEIVED) {
 445                                 dccp_pr_debug("%s ACK vector 0, len=%d, "
 446                                               "ack_seqno=%llu, ack_ackno=%llu, "
 447                                               "ACKED!\n",
 448                                               dccp_role(sk), len,
 449                                               (unsigned long long)
 450                                               avr->dccpavr_ack_seqno,
 451                                               (unsigned long long)
 452                                               avr->dccpavr_ack_ackno);
 453                                 dccp_ackvec_throw_record(av, avr);
 454                                 break;
 455                         }
 456                         /*
 457                          * If it wasn't received, continue scanning... we might
 458                          * find another one.
 459                          */
 460                 }
 461
 462                 dccp_set_seqno(ackno, ackno_end_rl - 1);
 463                 ++vector;
 464         }
 465 }
 466
 467 int dccp_ackvec_parse(struct sock *sk, const struct sk_buff *skb,
 468                       u64 *ackno, const u8 opt, const u8 *value, const u8 len)
 469 {
 470         if (len > DCCP_MAX_ACKVEC_OPT_LEN)
 471                 return -1;
 472
 473         /* dccp_ackvector_print(DCCP_SKB_CB(skb)->dccpd_ack_seq, value, len); */
 474         dccp_ackvec_check_rcv_ackvector(dccp_sk(sk)->dccps_hc_rx_ackvec, sk,
 475                                         ackno, len, value);
 476         return 0;
 477 }
 478
 479 int __init dccp_ackvec_init(void)
 480 {
 481         dccp_ackvec_slab = kmem_cache_create("dccp_ackvec",
 482                                              sizeof(struct dccp_ackvec), 0,
 483                                              SLAB_HWCACHE_ALIGN, NULL);
 484         if (dccp_ackvec_slab == NULL)
 485                 goto out_err;
 486
 487         dccp_ackvec_record_slab =
 488                         kmem_cache_create("dccp_ackvec_record",
 489                                           sizeof(struct dccp_ackvec_record),
 490                                           0, SLAB_HWCACHE_ALIGN, NULL);
 491         if (dccp_ackvec_record_slab == NULL)
 492                 goto out_destroy_slab;
 493
 494         return 0;
 495
 496 out_destroy_slab:
 497         kmem_cache_destroy(dccp_ackvec_slab);
 498         dccp_ackvec_slab = NULL;
 499 out_err:
 500         DCCP_CRIT("Unable to create Ack Vector slab cache");
 501         return -ENOBUFS;
 502 }
 503
 504 void dccp_ackvec_exit(void)
 505 {
 506         if (dccp_ackvec_slab != NULL) {
 507                 kmem_cache_destroy(dccp_ackvec_slab);
 508                 dccp_ackvec_slab = NULL;
 509         }
 510         if (dccp_ackvec_record_slab != NULL) {
 511                 kmem_cache_destroy(dccp_ackvec_record_slab);
 512                 dccp_ackvec_record_slab = NULL;
 513         }
 514 }