git.oblomov.eu Git - linux-2.6/blob - net/dccp/ccids/lib/packet_history.c

   1 /*
   2  *  net/dccp/packet_history.c
   3  *
   4  *  Copyright (c) 2007   The University of Aberdeen, Scotland, UK
   5  *  Copyright (c) 2005-7 The University of Waikato, Hamilton, New Zealand.
   6  *
   7  *  An implementation of the DCCP protocol
   8  *
   9  *  This code has been developed by the University of Waikato WAND
  10  *  research group. For further information please see http://www.wand.net.nz/
  11  *  or e-mail Ian McDonald - ian.mcdonald@jandi.co.nz
  12  *
  13  *  This code also uses code from Lulea University, rereleased as GPL by its
  14  *  authors:
  15  *  Copyright (c) 2003 Nils-Erik Mattsson, Joacim Haggmark, Magnus Erixzon
  16  *
  17  *  Changes to meet Linux coding standards, to make it meet latest ccid3 draft
  18  *  and to make it work as a loadable module in the DCCP stack written by
  19  *  Arnaldo Carvalho de Melo <acme@conectiva.com.br>.
  20  *
  21  *  Copyright (c) 2005 Arnaldo Carvalho de Melo <acme@conectiva.com.br>
  22  *
  23  *  This program is free software; you can redistribute it and/or modify
  24  *  it under the terms of the GNU General Public License as published by
  25  *  the Free Software Foundation; either version 2 of the License, or
  26  *  (at your option) any later version.
  27  *
  28  *  This program is distributed in the hope that it will be useful,
  29  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  30  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  31  *  GNU General Public License for more details.
  32  *
  33  *  You should have received a copy of the GNU General Public License
  34  *  along with this program; if not, write to the Free Software
  35  *  Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  36  */
  37
  38 #include <linux/string.h>
  39 #include <linux/slab.h>
  40 #include "packet_history.h"
  41 #include "../../dccp.h"
  42
  43 /*
  44  * Transmitter History Routines
  45  */
  46 static struct kmem_cache *tfrc_tx_hist_slab;
  47
  48 int __init tfrc_tx_packet_history_init(void)
  49 {
  50         tfrc_tx_hist_slab = kmem_cache_create("tfrc_tx_hist",
  51                                               sizeof(struct tfrc_tx_hist_entry),
  52                                               0, SLAB_HWCACHE_ALIGN, NULL);
  53         return tfrc_tx_hist_slab == NULL ? -ENOBUFS : 0;
  54 }
  55
  56 void tfrc_tx_packet_history_exit(void)
  57 {
  58         if (tfrc_tx_hist_slab != NULL) {
  59                 kmem_cache_destroy(tfrc_tx_hist_slab);
  60                 tfrc_tx_hist_slab = NULL;
  61         }
  62 }
  63
  64 int tfrc_tx_hist_add(struct tfrc_tx_hist_entry **headp, u64 seqno)
  65 {
  66         struct tfrc_tx_hist_entry *entry = kmem_cache_alloc(tfrc_tx_hist_slab, gfp_any());
  67
  68         if (entry == NULL)
  69                 return -ENOBUFS;
  70         entry->seqno = seqno;
  71         entry->stamp = ktime_get_real();
  72         entry->next  = *headp;
  73         *headp       = entry;
  74         return 0;
  75 }
  76 EXPORT_SYMBOL_GPL(tfrc_tx_hist_add);
  77
  78 void tfrc_tx_hist_purge(struct tfrc_tx_hist_entry **headp)
  79 {
  80         struct tfrc_tx_hist_entry *head = *headp;
  81
  82         while (head != NULL) {
  83                 struct tfrc_tx_hist_entry *next = head->next;
  84
  85                 kmem_cache_free(tfrc_tx_hist_slab, head);
  86                 head = next;
  87         }
  88
  89         *headp = NULL;
  90 }
  91 EXPORT_SYMBOL_GPL(tfrc_tx_hist_purge);
  92
  93 /*
  94  *      Receiver History Routines
  95  */
  96 static struct kmem_cache *tfrc_rx_hist_slab;
  97
  98 int __init tfrc_rx_packet_history_init(void)
  99 {
 100         tfrc_rx_hist_slab = kmem_cache_create("tfrc_rxh_cache",
 101                                               sizeof(struct tfrc_rx_hist_entry),
 102                                               0, SLAB_HWCACHE_ALIGN, NULL);
 103         return tfrc_rx_hist_slab == NULL ? -ENOBUFS : 0;
 104 }
 105
 106 void tfrc_rx_packet_history_exit(void)
 107 {
 108         if (tfrc_rx_hist_slab != NULL) {
 109                 kmem_cache_destroy(tfrc_rx_hist_slab);
 110                 tfrc_rx_hist_slab = NULL;
 111         }
 112 }
 113
 114 static inline void tfrc_rx_hist_entry_from_skb(struct tfrc_rx_hist_entry *entry,
 115                                                const struct sk_buff *skb,
 116                                                const u64 ndp)
 117 {
 118         const struct dccp_hdr *dh = dccp_hdr(skb);
 119
 120         entry->tfrchrx_seqno = DCCP_SKB_CB(skb)->dccpd_seq;
 121         entry->tfrchrx_ccval = dh->dccph_ccval;
 122         entry->tfrchrx_type  = dh->dccph_type;
 123         entry->tfrchrx_ndp   = ndp;
 124         entry->tfrchrx_tstamp = ktime_get_real();
 125 }
 126
 127 void tfrc_rx_hist_add_packet(struct tfrc_rx_hist *h,
 128                              const struct sk_buff *skb,
 129                              const u64 ndp)
 130 {
 131         struct tfrc_rx_hist_entry *entry = tfrc_rx_hist_last_rcv(h);
 132
 133         tfrc_rx_hist_entry_from_skb(entry, skb, ndp);
 134 }
 135 EXPORT_SYMBOL_GPL(tfrc_rx_hist_add_packet);
 136
 137 /* has the packet contained in skb been seen before? */
 138 int tfrc_rx_hist_duplicate(struct tfrc_rx_hist *h, struct sk_buff *skb)
 139 {
 140         const u64 seq = DCCP_SKB_CB(skb)->dccpd_seq;
 141         int i;
 142
 143         if (dccp_delta_seqno(tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno, seq) <= 0)
 144                 return 1;
 145
 146         for (i = 1; i <= h->loss_count; i++)
 147                 if (tfrc_rx_hist_entry(h, i)->tfrchrx_seqno == seq)
 148                         return 1;
 149
 150         return 0;
 151 }
 152 EXPORT_SYMBOL_GPL(tfrc_rx_hist_duplicate);
 153
 154 static void tfrc_rx_hist_swap(struct tfrc_rx_hist *h, const u8 a, const u8 b)
 155 {
 156         const u8 idx_a = tfrc_rx_hist_index(h, a),
 157                  idx_b = tfrc_rx_hist_index(h, b);
 158         struct tfrc_rx_hist_entry *tmp = h->ring[idx_a];
 159
 160         h->ring[idx_a] = h->ring[idx_b];
 161         h->ring[idx_b] = tmp;
 162 }
 163
 164 /*
 165  * Private helper functions for loss detection.
 166  *
 167  * In the descriptions, `Si' refers to the sequence number of entry number i,
 168  * whose NDP count is `Ni' (lower case is used for variables).
 169  * Note: All __xxx_loss functions expect that a test against duplicates has been
 170  *       performed already: the seqno of the skb must not be less than the seqno
 171  *       of loss_prev; and it must not equal that of any valid history entry.
 172  */
 173 static void __do_track_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u64 n1)
 174 {
 175         u64 s0 = tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno,
 176             s1 = DCCP_SKB_CB(skb)->dccpd_seq;
 177
 178         if (!dccp_loss_free(s0, s1, n1)) {      /* gap between S0 and S1 */
 179                 h->loss_count = 1;
 180                 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 1), skb, n1);
 181         }
 182 }
 183
 184 static void __one_after_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u32 n2)
 185 {
 186         u64 s0 = tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno,
 187             s1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_seqno,
 188             s2 = DCCP_SKB_CB(skb)->dccpd_seq;
 189
 190         if (likely(dccp_delta_seqno(s1, s2) > 0)) {     /* S1  <  S2 */
 191                 h->loss_count = 2;
 192                 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 2), skb, n2);
 193                 return;
 194         }
 195
 196         /* S0  <  S2  <  S1 */
 197
 198         if (dccp_loss_free(s0, s2, n2)) {
 199                 u64 n1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_ndp;
 200
 201                 if (dccp_loss_free(s2, s1, n1)) {
 202                         /* hole is filled: S0, S2, and S1 are consecutive */
 203                         h->loss_count = 0;
 204                         h->loss_start = tfrc_rx_hist_index(h, 1);
 205                 } else
 206                         /* gap between S2 and S1: just update loss_prev */
 207                         tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_loss_prev(h), skb, n2);
 208
 209         } else {        /* gap between S0 and S2 */
 210                 /*
 211                  * Reorder history to insert S2 between S0 and S1
 212                  */
 213                 tfrc_rx_hist_swap(h, 0, 3);
 214                 h->loss_start = tfrc_rx_hist_index(h, 3);
 215                 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 1), skb, n2);
 216                 h->loss_count = 2;
 217         }
 218 }
 219
 220 /* return 1 if a new loss event has been identified */
 221 static int __two_after_loss(struct tfrc_rx_hist *h, struct sk_buff *skb, u32 n3)
 222 {
 223         u64 s0 = tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno,
 224             s1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_seqno,
 225             s2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_seqno,
 226             s3 = DCCP_SKB_CB(skb)->dccpd_seq;
 227
 228         if (likely(dccp_delta_seqno(s2, s3) > 0)) {     /* S2  <  S3 */
 229                 h->loss_count = 3;
 230                 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 3), skb, n3);
 231                 return 1;
 232         }
 233
 234         /* S3  <  S2 */
 235
 236         if (dccp_delta_seqno(s1, s3) > 0) {             /* S1  <  S3  <  S2 */
 237                 /*
 238                  * Reorder history to insert S3 between S1 and S2
 239                  */
 240                 tfrc_rx_hist_swap(h, 2, 3);
 241                 tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 2), skb, n3);
 242                 h->loss_count = 3;
 243                 return 1;
 244         }
 245
 246         /* S0  <  S3  <  S1 */
 247
 248         if (dccp_loss_free(s0, s3, n3)) {
 249                 u64 n1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_ndp;
 250
 251                 if (dccp_loss_free(s3, s1, n1)) {
 252                         /* hole between S0 and S1 filled by S3 */
 253                         u64 n2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_ndp;
 254
 255                         if (dccp_loss_free(s1, s2, n2)) {
 256                                 /* entire hole filled by S0, S3, S1, S2 */
 257                                 h->loss_start = tfrc_rx_hist_index(h, 2);
 258                                 h->loss_count = 0;
 259                         } else {
 260                                 /* gap remains between S1 and S2 */
 261                                 h->loss_start = tfrc_rx_hist_index(h, 1);
 262                                 h->loss_count = 1;
 263                         }
 264
 265                 } else /* gap exists between S3 and S1, loss_count stays at 2 */
 266                         tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_loss_prev(h), skb, n3);
 267
 268                 return 0;
 269         }
 270
 271         /*
 272          * The remaining case:  S0  <  S3  <  S1  <  S2;  gap between S0 and S3
 273          * Reorder history to insert S3 between S0 and S1.
 274          */
 275         tfrc_rx_hist_swap(h, 0, 3);
 276         h->loss_start = tfrc_rx_hist_index(h, 3);
 277         tfrc_rx_hist_entry_from_skb(tfrc_rx_hist_entry(h, 1), skb, n3);
 278         h->loss_count = 3;
 279
 280         return 1;
 281 }
 282
 283 /* recycle RX history records to continue loss detection if necessary */
 284 static void __three_after_loss(struct tfrc_rx_hist *h)
 285 {
 286         /*
 287          * At this stage we know already that there is a gap between S0 and S1
 288          * (since S0 was the highest sequence number received before detecting
 289          * the loss). To recycle the loss record, it is thus only necessary to
 290          * check for other possible gaps between S1/S2 and between S2/S3.
 291          */
 292         u64 s1 = tfrc_rx_hist_entry(h, 1)->tfrchrx_seqno,
 293             s2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_seqno,
 294             s3 = tfrc_rx_hist_entry(h, 3)->tfrchrx_seqno;
 295         u64 n2 = tfrc_rx_hist_entry(h, 2)->tfrchrx_ndp,
 296             n3 = tfrc_rx_hist_entry(h, 3)->tfrchrx_ndp;
 297
 298         if (dccp_loss_free(s1, s2, n2)) {
 299
 300                 if (dccp_loss_free(s2, s3, n3)) {
 301                         /* no gap between S2 and S3: entire hole is filled */
 302                         h->loss_start = tfrc_rx_hist_index(h, 3);
 303                         h->loss_count = 0;
 304                 } else {
 305                         /* gap between S2 and S3 */
 306                         h->loss_start = tfrc_rx_hist_index(h, 2);
 307                         h->loss_count = 1;
 308                 }
 309
 310         } else {        /* gap between S1 and S2 */
 311                 h->loss_start = tfrc_rx_hist_index(h, 1);
 312                 h->loss_count = 2;
 313         }
 314 }
 315
 316 /**
 317  *  tfrc_rx_handle_loss  -  Loss detection and further processing
 318  *  @h:             The non-empty RX history object
 319  *  @lh:            Loss Intervals database to update
 320  *  @skb:           Currently received packet
 321  *  @ndp:           The NDP count belonging to @skb
 322  *  @calc_first_li: Caller-dependent computation of first loss interval in @lh
 323  *  @sk:            Used by @calc_first_li (see tfrc_lh_interval_add)
 324  *  Chooses action according to pending loss, updates LI database when a new
 325  *  loss was detected, and does required post-processing. Returns 1 when caller
 326  *  should send feedback, 0 otherwise.
 327  *  Since it also takes care of reordering during loss detection and updates the
 328  *  records accordingly, the caller should not perform any more RX history
 329  *  operations when loss_count is greater than 0 after calling this function.
 330  */
 331 int tfrc_rx_handle_loss(struct tfrc_rx_hist *h,
 332                         struct tfrc_loss_hist *lh,
 333                         struct sk_buff *skb, const u64 ndp,
 334                         u32 (*calc_first_li)(struct sock *), struct sock *sk)
 335 {
 336         int is_new_loss = 0;
 337
 338         if (tfrc_rx_hist_duplicate(h, skb))
 339                 return 0;
 340
 341         if (h->loss_count == 0) {
 342                 __do_track_loss(h, skb, ndp);
 343         } else if (h->loss_count == 1) {
 344                 __one_after_loss(h, skb, ndp);
 345         } else if (h->loss_count != 2) {
 346                 DCCP_BUG("invalid loss_count %d", h->loss_count);
 347         } else if (__two_after_loss(h, skb, ndp)) {
 348                 /*
 349                  * Update Loss Interval database and recycle RX records
 350                  */
 351                 is_new_loss = tfrc_lh_interval_add(lh, h, calc_first_li, sk);
 352                 __three_after_loss(h);
 353         }
 354         return is_new_loss;
 355 }
 356 EXPORT_SYMBOL_GPL(tfrc_rx_handle_loss);
 357
 358 void tfrc_rx_hist_purge(struct tfrc_rx_hist *h)
 359 {
 360         int i;
 361
 362         for (i = 0; i <= TFRC_NDUPACK; ++i)
 363                 if (h->ring[i] != NULL) {
 364                         kmem_cache_free(tfrc_rx_hist_slab, h->ring[i]);
 365                         h->ring[i] = NULL;
 366                 }
 367 }
 368 EXPORT_SYMBOL_GPL(tfrc_rx_hist_purge);
 369
 370 static int tfrc_rx_hist_alloc(struct tfrc_rx_hist *h)
 371 {
 372         int i;
 373
 374         memset(h, 0, sizeof(*h));
 375
 376         for (i = 0; i <= TFRC_NDUPACK; i++) {
 377                 h->ring[i] = kmem_cache_alloc(tfrc_rx_hist_slab, GFP_ATOMIC);
 378                 if (h->ring[i] == NULL) {
 379                         tfrc_rx_hist_purge(h);
 380                         return -ENOBUFS;
 381                 }
 382         }
 383         return 0;
 384 }
 385
 386 int tfrc_rx_hist_init(struct tfrc_rx_hist *h, struct sock *sk)
 387 {
 388         if (tfrc_rx_hist_alloc(h))
 389                 return -ENOBUFS;
 390         /*
 391          * Initialise first entry with GSR to start loss detection as early as
 392          * possible. Code using this must not use any other fields. The entry
 393          * will be overwritten once the CCID updates its received packets.
 394          */
 395         tfrc_rx_hist_loss_prev(h)->tfrchrx_seqno = dccp_sk(sk)->dccps_gsr;
 396         return 0;
 397 }
 398 EXPORT_SYMBOL_GPL(tfrc_rx_hist_init);
 399
 400 /**
 401  * tfrc_rx_hist_rtt_last_s - reference entry to compute RTT samples against
 402  */
 403 static inline struct tfrc_rx_hist_entry *
 404                         tfrc_rx_hist_rtt_last_s(const struct tfrc_rx_hist *h)
 405 {
 406         return h->ring[0];
 407 }
 408
 409 /**
 410  * tfrc_rx_hist_rtt_prev_s: previously suitable (wrt rtt_last_s) RTT-sampling entry
 411  */
 412 static inline struct tfrc_rx_hist_entry *
 413                         tfrc_rx_hist_rtt_prev_s(const struct tfrc_rx_hist *h)
 414 {
 415         return h->ring[h->rtt_sample_prev];
 416 }
 417
 418 /**
 419  * tfrc_rx_hist_sample_rtt  -  Sample RTT from timestamp / CCVal
 420  * Based on ideas presented in RFC 4342, 8.1. Returns 0 if it was not able
 421  * to compute a sample with given data - calling function should check this.
 422  */
 423 u32 tfrc_rx_hist_sample_rtt(struct tfrc_rx_hist *h, const struct sk_buff *skb)
 424 {
 425         u32 sample = 0,
 426             delta_v = SUB16(dccp_hdr(skb)->dccph_ccval,
 427                             tfrc_rx_hist_rtt_last_s(h)->tfrchrx_ccval);
 428
 429         if (delta_v < 1 || delta_v > 4) {       /* unsuitable CCVal delta */
 430                 if (h->rtt_sample_prev == 2) {  /* previous candidate stored */
 431                         sample = SUB16(tfrc_rx_hist_rtt_prev_s(h)->tfrchrx_ccval,
 432                                        tfrc_rx_hist_rtt_last_s(h)->tfrchrx_ccval);
 433                         if (sample)
 434                                 sample = 4 / sample *
 435                                          ktime_us_delta(tfrc_rx_hist_rtt_prev_s(h)->tfrchrx_tstamp,
 436                                                         tfrc_rx_hist_rtt_last_s(h)->tfrchrx_tstamp);
 437                         else    /*
 438                                  * FIXME: This condition is in principle not
 439                                  * possible but occurs when CCID is used for
 440                                  * two-way data traffic. I have tried to trace
 441                                  * it, but the cause does not seem to be here.
 442                                  */
 443                                 DCCP_BUG("please report to dccp@vger.kernel.org"
 444                                          " => prev = %u, last = %u",
 445                                          tfrc_rx_hist_rtt_prev_s(h)->tfrchrx_ccval,
 446                                          tfrc_rx_hist_rtt_last_s(h)->tfrchrx_ccval);
 447                 } else if (delta_v < 1) {
 448                         h->rtt_sample_prev = 1;
 449                         goto keep_ref_for_next_time;
 450                 }
 451
 452         } else if (delta_v == 4) /* optimal match */
 453                 sample = ktime_to_us(net_timedelta(tfrc_rx_hist_rtt_last_s(h)->tfrchrx_tstamp));
 454         else {                   /* suboptimal match */
 455                 h->rtt_sample_prev = 2;
 456                 goto keep_ref_for_next_time;
 457         }
 458
 459         if (unlikely(sample > DCCP_SANE_RTT_MAX)) {
 460                 DCCP_WARN("RTT sample %u too large, using max\n", sample);
 461                 sample = DCCP_SANE_RTT_MAX;
 462         }
 463
 464         h->rtt_sample_prev = 0;        /* use current entry as next reference */
 465 keep_ref_for_next_time:
 466
 467         return sample;
 468 }
 469 EXPORT_SYMBOL_GPL(tfrc_rx_hist_sample_rtt);