net/batman-adv/tp_meter.c

   1 /* Copyright (C) 2012-2017  B.A.T.M.A.N. contributors:
   2  *
   3  * Edo Monticelli, Antonio Quartulli
   4  *
   5  * This program is free software; you can redistribute it and/or
   6  * modify it under the terms of version 2 of the GNU General Public
   7  * License as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it will be useful, but
  10  * WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  12  * General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, see <http://www.gnu.org/licenses/>.
  16  */
  17
  18 #include "tp_meter.h"
  19 #include "main.h"
  20
  21 #include <linux/atomic.h>
  22 #include <linux/bug.h>
  23 #include <linux/byteorder/generic.h>
  24 #include <linux/cache.h>
  25 #include <linux/compiler.h>
  26 #include <linux/err.h>
  27 #include <linux/etherdevice.h>
  28 #include <linux/fs.h>
  29 #include <linux/if_ether.h>
  30 #include <linux/jiffies.h>
  31 #include <linux/kernel.h>
  32 #include <linux/kref.h>
  33 #include <linux/kthread.h>
  34 #include <linux/list.h>
  35 #include <linux/netdevice.h>
  36 #include <linux/param.h>
  37 #include <linux/printk.h>
  38 #include <linux/random.h>
  39 #include <linux/rculist.h>
  40 #include <linux/rcupdate.h>
  41 #include <linux/sched.h>
  42 #include <linux/skbuff.h>
  43 #include <linux/slab.h>
  44 #include <linux/spinlock.h>
  45 #include <linux/stddef.h>
  46 #include <linux/string.h>
  47 #include <linux/timer.h>
  48 #include <linux/wait.h>
  49 #include <linux/workqueue.h>
  50 #include <uapi/linux/batman_adv.h>
  51
  52 #include "hard-interface.h"
  53 #include "log.h"
  54 #include "netlink.h"
  55 #include "originator.h"
  56 #include "packet.h"
  57 #include "send.h"
  58
  59 /**
  60  * BATADV_TP_DEF_TEST_LENGTH - Default test length if not specified by the user
  61  *  in milliseconds
  62  */
  63 #define BATADV_TP_DEF_TEST_LENGTH 10000
  64
  65 /**
  66  * BATADV_TP_AWND - Advertised window by the receiver (in bytes)
  67  */
  68 #define BATADV_TP_AWND 0x20000000
  69
  70 /**
  71  * BATADV_TP_RECV_TIMEOUT - Receiver activity timeout. If the receiver does not
  72  *  get anything for such amount of milliseconds, the connection is killed
  73  */
  74 #define BATADV_TP_RECV_TIMEOUT 1000
  75
  76 /**
  77  * BATADV_TP_MAX_RTO - Maximum sender timeout. If the sender RTO gets beyond
  78  * such amound of milliseconds, the receiver is considered unreachable and the
  79  * connection is killed
  80  */
  81 #define BATADV_TP_MAX_RTO 30000
  82
  83 /**
  84  * BATADV_TP_FIRST_SEQ - First seqno of each session. The number is rather high
  85  *  in order to immediately trigger a wrap around (test purposes)
  86  */
  87 #define BATADV_TP_FIRST_SEQ ((u32)-1 - 2000)
  88
  89 /**
  90  * BATADV_TP_PLEN - length of the payload (data after the batadv_unicast header)
  91  *  to simulate
  92  */
  93 #define BATADV_TP_PLEN (BATADV_TP_PACKET_LEN - ETH_HLEN - \
  94                         sizeof(struct batadv_unicast_packet))
  95
  96 static u8 batadv_tp_prerandom[4096] __read_mostly;
  97
  98 /**
  99  * batadv_tp_session_cookie - generate session cookie based on session ids
 100  * @session: TP session identifier
 101  * @icmp_uid: icmp pseudo uid of the tp session
 102  *
 103  * Return: 32 bit tp_meter session cookie
 104  */
 105 static u32 batadv_tp_session_cookie(const u8 session[2], u8 icmp_uid)
 106 {
 107         u32 cookie;
 108
 109         cookie = icmp_uid << 16;
 110         cookie |= session[0] << 8;
 111         cookie |= session[1];
 112
 113         return cookie;
 114 }
 115
 116 /**
 117  * batadv_tp_cwnd - compute the new cwnd size
 118  * @base: base cwnd size value
 119  * @increment: the value to add to base to get the new size
 120  * @min: minumim cwnd value (usually MSS)
 121  *
 122  * Return the new cwnd size and ensures it does not exceed the Advertised
 123  * Receiver Window size. It is wrap around safe.
 124  * For details refer to Section 3.1 of RFC5681
 125  *
 126  * Return: new congestion window size in bytes
 127  */
 128 static u32 batadv_tp_cwnd(u32 base, u32 increment, u32 min)
 129 {
 130         u32 new_size = base + increment;
 131
 132         /* check for wrap-around */
 133         if (new_size < base)
 134                 new_size = (u32)ULONG_MAX;
 135
 136         new_size = min_t(u32, new_size, BATADV_TP_AWND);
 137
 138         return max_t(u32, new_size, min);
 139 }
 140
 141 /**
 142  * batadv_tp_updated_cwnd - update the Congestion Windows
 143  * @tp_vars: the private data of the current TP meter session
 144  * @mss: maximum segment size of transmission
 145  *
 146  * 1) if the session is in Slow Start, the CWND has to be increased by 1
 147  * MSS every unique received ACK
 148  * 2) if the session is in Congestion Avoidance, the CWND has to be
 149  * increased by MSS * MSS / CWND for every unique received ACK
 150  */
 151 static void batadv_tp_update_cwnd(struct batadv_tp_vars *tp_vars, u32 mss)
 152 {
 153         spin_lock_bh(&tp_vars->cwnd_lock);
 154
 155         /* slow start... */
 156         if (tp_vars->cwnd <= tp_vars->ss_threshold) {
 157                 tp_vars->dec_cwnd = 0;
 158                 tp_vars->cwnd = batadv_tp_cwnd(tp_vars->cwnd, mss, mss);
 159                 spin_unlock_bh(&tp_vars->cwnd_lock);
 160                 return;
 161         }
 162
 163         /* increment CWND at least of 1 (section 3.1 of RFC5681) */
 164         tp_vars->dec_cwnd += max_t(u32, 1U << 3,
 165                                    ((mss * mss) << 6) / (tp_vars->cwnd << 3));
 166         if (tp_vars->dec_cwnd < (mss << 3)) {
 167                 spin_unlock_bh(&tp_vars->cwnd_lock);
 168                 return;
 169         }
 170
 171         tp_vars->cwnd = batadv_tp_cwnd(tp_vars->cwnd, mss, mss);
 172         tp_vars->dec_cwnd = 0;
 173
 174         spin_unlock_bh(&tp_vars->cwnd_lock);
 175 }
 176
 177 /**
 178  * batadv_tp_update_rto - calculate new retransmission timeout
 179  * @tp_vars: the private data of the current TP meter session
 180  * @new_rtt: new roundtrip time in msec
 181  */
 182 static void batadv_tp_update_rto(struct batadv_tp_vars *tp_vars,
 183                                  u32 new_rtt)
 184 {
 185         long m = new_rtt;
 186
 187         /* RTT update
 188          * Details in Section 2.2 and 2.3 of RFC6298
 189          *
 190          * It's tricky to understand. Don't lose hair please.
 191          * Inspired by tcp_rtt_estimator() tcp_input.c
 192          */
 193         if (tp_vars->srtt != 0) {
 194                 m -= (tp_vars->srtt >> 3); /* m is now error in rtt est */
 195                 tp_vars->srtt += m; /* rtt = 7/8 srtt + 1/8 new */
 196                 if (m < 0)
 197                         m = -m;
 198
 199                 m -= (tp_vars->rttvar >> 2);
 200                 tp_vars->rttvar += m; /* mdev ~= 3/4 rttvar + 1/4 new */
 201         } else {
 202                 /* first measure getting in */
 203                 tp_vars->srtt = m << 3; /* take the measured time to be srtt */
 204                 tp_vars->rttvar = m << 1; /* new_rtt / 2 */
 205         }
 206
 207         /* rto = srtt + 4 * rttvar.
 208          * rttvar is scaled by 4, therefore doesn't need to be multiplied
 209          */
 210         tp_vars->rto = (tp_vars->srtt >> 3) + tp_vars->rttvar;
 211 }
 212
 213 /**
 214  * batadv_tp_batctl_notify - send client status result to client
 215  * @reason: reason for tp meter session stop
 216  * @dst: destination of tp_meter session
 217  * @bat_priv: the bat priv with all the soft interface information
 218  * @start_time: start of transmission in jiffies
 219  * @total_sent: bytes acked to the receiver
 220  * @cookie: cookie of tp_meter session
 221  */
 222 static void batadv_tp_batctl_notify(enum batadv_tp_meter_reason reason,
 223                                     const u8 *dst, struct batadv_priv *bat_priv,
 224                                     unsigned long start_time, u64 total_sent,
 225                                     u32 cookie)
 226 {
 227         u32 test_time;
 228         u8 result;
 229         u32 total_bytes;
 230
 231         if (!batadv_tp_is_error(reason)) {
 232                 result = BATADV_TP_REASON_COMPLETE;
 233                 test_time = jiffies_to_msecs(jiffies - start_time);
 234                 total_bytes = total_sent;
 235         } else {
 236                 result = reason;
 237                 test_time = 0;
 238                 total_bytes = 0;
 239         }
 240
 241         batadv_netlink_tpmeter_notify(bat_priv, dst, result, test_time,
 242                                       total_bytes, cookie);
 243 }
 244
 245 /**
 246  * batadv_tp_batctl_error_notify - send client error result to client
 247  * @reason: reason for tp meter session stop
 248  * @dst: destination of tp_meter session
 249  * @bat_priv: the bat priv with all the soft interface information
 250  * @cookie: cookie of tp_meter session
 251  */
 252 static void batadv_tp_batctl_error_notify(enum batadv_tp_meter_reason reason,
 253                                           const u8 *dst,
 254                                           struct batadv_priv *bat_priv,
 255                                           u32 cookie)
 256 {
 257         batadv_tp_batctl_notify(reason, dst, bat_priv, 0, 0, cookie);
 258 }
 259
 260 /**
 261  * batadv_tp_list_find - find a tp_vars object in the global list
 262  * @bat_priv: the bat priv with all the soft interface information
 263  * @dst: the other endpoint MAC address to look for
 264  *
 265  * Look for a tp_vars object matching dst as end_point and return it after
 266  * having incremented the refcounter. Return NULL is not found
 267  *
 268  * Return: matching tp_vars or NULL when no tp_vars with @dst was found
 269  */
 270 static struct batadv_tp_vars *batadv_tp_list_find(struct batadv_priv *bat_priv,
 271                                                   const u8 *dst)
 272 {
 273         struct batadv_tp_vars *pos, *tp_vars = NULL;
 274
 275         rcu_read_lock();
 276         hlist_for_each_entry_rcu(pos, &bat_priv->tp_list, list) {
 277                 if (!batadv_compare_eth(pos->other_end, dst))
 278                         continue;
 279
 280                 /* most of the time this function is invoked during the normal
 281                  * process..it makes sens to pay more when the session is
 282                  * finished and to speed the process up during the measurement
 283                  */
 284                 if (unlikely(!kref_get_unless_zero(&pos->refcount)))
 285                         continue;
 286
 287                 tp_vars = pos;
 288                 break;
 289         }
 290         rcu_read_unlock();
 291
 292         return tp_vars;
 293 }
 294
 295 /**
 296  * batadv_tp_list_find_session - find tp_vars session object in the global list
 297  * @bat_priv: the bat priv with all the soft interface information
 298  * @dst: the other endpoint MAC address to look for
 299  * @session: session identifier
 300  *
 301  * Look for a tp_vars object matching dst as end_point, session as tp meter
 302  * session and return it after having incremented the refcounter. Return NULL
 303  * is not found
 304  *
 305  * Return: matching tp_vars or NULL when no tp_vars was found
 306  */
 307 static struct batadv_tp_vars *
 308 batadv_tp_list_find_session(struct batadv_priv *bat_priv, const u8 *dst,
 309                             const u8 *session)
 310 {
 311         struct batadv_tp_vars *pos, *tp_vars = NULL;
 312
 313         rcu_read_lock();
 314         hlist_for_each_entry_rcu(pos, &bat_priv->tp_list, list) {
 315                 if (!batadv_compare_eth(pos->other_end, dst))
 316                         continue;
 317
 318                 if (memcmp(pos->session, session, sizeof(pos->session)) != 0)
 319                         continue;
 320
 321                 /* most of the time this function is invoked during the normal
 322                  * process..it makes sense to pay more when the session is
 323                  * finished and to speed the process up during the measurement
 324                  */
 325                 if (unlikely(!kref_get_unless_zero(&pos->refcount)))
 326                         continue;
 327
 328                 tp_vars = pos;
 329                 break;
 330         }
 331         rcu_read_unlock();
 332
 333         return tp_vars;
 334 }
 335
 336 /**
 337  * batadv_tp_vars_release - release batadv_tp_vars from lists and queue for
 338  *  free after rcu grace period
 339  * @ref: kref pointer of the batadv_tp_vars
 340  */
 341 static void batadv_tp_vars_release(struct kref *ref)
 342 {
 343         struct batadv_tp_vars *tp_vars;
 344         struct batadv_tp_unacked *un, *safe;
 345
 346         tp_vars = container_of(ref, struct batadv_tp_vars, refcount);
 347
 348         /* lock should not be needed because this object is now out of any
 349          * context!
 350          */
 351         spin_lock_bh(&tp_vars->unacked_lock);
 352         list_for_each_entry_safe(un, safe, &tp_vars->unacked_list, list) {
 353                 list_del(&un->list);
 354                 kfree(un);
 355         }
 356         spin_unlock_bh(&tp_vars->unacked_lock);
 357
 358         kfree_rcu(tp_vars, rcu);
 359 }
 360
 361 /**
 362  * batadv_tp_vars_put - decrement the batadv_tp_vars refcounter and possibly
 363  *  release it
 364  * @tp_vars: the private data of the current TP meter session to be free'd
 365  */
 366 static void batadv_tp_vars_put(struct batadv_tp_vars *tp_vars)
 367 {
 368         kref_put(&tp_vars->refcount, batadv_tp_vars_release);
 369 }
 370
 371 /**
 372  * batadv_tp_sender_cleanup - cleanup sender data and drop and timer
 373  * @bat_priv: the bat priv with all the soft interface information
 374  * @tp_vars: the private data of the current TP meter session to cleanup
 375  */
 376 static void batadv_tp_sender_cleanup(struct batadv_priv *bat_priv,
 377                                      struct batadv_tp_vars *tp_vars)
 378 {
 379         cancel_delayed_work(&tp_vars->finish_work);
 380
 381         spin_lock_bh(&tp_vars->bat_priv->tp_list_lock);
 382         hlist_del_rcu(&tp_vars->list);
 383         spin_unlock_bh(&tp_vars->bat_priv->tp_list_lock);
 384
 385         /* drop list reference */
 386         batadv_tp_vars_put(tp_vars);
 387
 388         atomic_dec(&tp_vars->bat_priv->tp_num);
 389
 390         /* kill the timer and remove its reference */
 391         del_timer_sync(&tp_vars->timer);
 392         /* the worker might have rearmed itself therefore we kill it again. Note
 393          * that if the worker should run again before invoking the following
 394          * del_timer(), it would not re-arm itself once again because the status
 395          * is OFF now
 396          */
 397         del_timer(&tp_vars->timer);
 398         batadv_tp_vars_put(tp_vars);
 399 }
 400
 401 /**
 402  * batadv_tp_sender_end - print info about ended session and inform client
 403  * @bat_priv: the bat priv with all the soft interface information
 404  * @tp_vars: the private data of the current TP meter session
 405  */
 406 static void batadv_tp_sender_end(struct batadv_priv *bat_priv,
 407                                  struct batadv_tp_vars *tp_vars)
 408 {
 409         u32 session_cookie;
 410
 411         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
 412                    "Test towards %pM finished..shutting down (reason=%d)\n",
 413                    tp_vars->other_end, tp_vars->reason);
 414
 415         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
 416                    "Last timing stats: SRTT=%ums RTTVAR=%ums RTO=%ums\n",
 417                    tp_vars->srtt >> 3, tp_vars->rttvar >> 2, tp_vars->rto);
 418
 419         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
 420                    "Final values: cwnd=%u ss_threshold=%u\n",
 421                    tp_vars->cwnd, tp_vars->ss_threshold);
 422
 423         session_cookie = batadv_tp_session_cookie(tp_vars->session,
 424                                                   tp_vars->icmp_uid);
 425
 426         batadv_tp_batctl_notify(tp_vars->reason,
 427                                 tp_vars->other_end,
 428                                 bat_priv,
 429                                 tp_vars->start_time,
 430                                 atomic64_read(&tp_vars->tot_sent),
 431                                 session_cookie);
 432 }
 433
 434 /**
 435  * batadv_tp_sender_shutdown - let sender thread/timer stop gracefully
 436  * @tp_vars: the private data of the current TP meter session
 437  * @reason: reason for tp meter session stop
 438  */
 439 static void batadv_tp_sender_shutdown(struct batadv_tp_vars *tp_vars,
 440                                       enum batadv_tp_meter_reason reason)
 441 {
 442         if (!atomic_dec_and_test(&tp_vars->sending))
 443                 return;
 444
 445         tp_vars->reason = reason;
 446 }
 447
 448 /**
 449  * batadv_tp_sender_finish - stop sender session after test_length was reached
 450  * @work: delayed work reference of the related tp_vars
 451  */
 452 static void batadv_tp_sender_finish(struct work_struct *work)
 453 {
 454         struct delayed_work *delayed_work;
 455         struct batadv_tp_vars *tp_vars;
 456
 457         delayed_work = to_delayed_work(work);
 458         tp_vars = container_of(delayed_work, struct batadv_tp_vars,
 459                                finish_work);
 460
 461         batadv_tp_sender_shutdown(tp_vars, BATADV_TP_REASON_COMPLETE);
 462 }
 463
 464 /**
 465  * batadv_tp_reset_sender_timer - reschedule the sender timer
 466  * @tp_vars: the private TP meter data for this session
 467  *
 468  * Reschedule the timer using tp_vars->rto as delay
 469  */
 470 static void batadv_tp_reset_sender_timer(struct batadv_tp_vars *tp_vars)
 471 {
 472         /* most of the time this function is invoked while normal packet
 473          * reception...
 474          */
 475         if (unlikely(atomic_read(&tp_vars->sending) == 0))
 476                 /* timer ref will be dropped in batadv_tp_sender_cleanup */
 477                 return;
 478
 479         mod_timer(&tp_vars->timer, jiffies + msecs_to_jiffies(tp_vars->rto));
 480 }
 481
 482 /**
 483  * batadv_tp_sender_timeout - timer that fires in case of packet loss
 484  * @arg: address of the related tp_vars
 485  *
 486  * If fired it means that there was packet loss.
 487  * Switch to Slow Start, set the ss_threshold to half of the current cwnd and
 488  * reset the cwnd to 3*MSS
 489  */
 490 static void batadv_tp_sender_timeout(unsigned long arg)
 491 {
 492         struct batadv_tp_vars *tp_vars = (struct batadv_tp_vars *)arg;
 493         struct batadv_priv *bat_priv = tp_vars->bat_priv;
 494
 495         if (atomic_read(&tp_vars->sending) == 0)
 496                 return;
 497
 498         /* if the user waited long enough...shutdown the test */
 499         if (unlikely(tp_vars->rto >= BATADV_TP_MAX_RTO)) {
 500                 batadv_tp_sender_shutdown(tp_vars,
 501                                           BATADV_TP_REASON_DST_UNREACHABLE);
 502                 return;
 503         }
 504
 505         /* RTO exponential backoff
 506          * Details in Section 5.5 of RFC6298
 507          */
 508         tp_vars->rto <<= 1;
 509
 510         spin_lock_bh(&tp_vars->cwnd_lock);
 511
 512         tp_vars->ss_threshold = tp_vars->cwnd >> 1;
 513         if (tp_vars->ss_threshold < BATADV_TP_PLEN * 2)
 514                 tp_vars->ss_threshold = BATADV_TP_PLEN * 2;
 515
 516         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
 517                    "Meter: RTO fired during test towards %pM! cwnd=%u new ss_thr=%u, resetting last_sent to %u\n",
 518                    tp_vars->other_end, tp_vars->cwnd, tp_vars->ss_threshold,
 519                    atomic_read(&tp_vars->last_acked));
 520
 521         tp_vars->cwnd = BATADV_TP_PLEN * 3;
 522
 523         spin_unlock_bh(&tp_vars->cwnd_lock);
 524
 525         /* resend the non-ACKed packets.. */
 526         tp_vars->last_sent = atomic_read(&tp_vars->last_acked);
 527         wake_up(&tp_vars->more_bytes);
 528
 529         batadv_tp_reset_sender_timer(tp_vars);
 530 }
 531
 532 /**
 533  * batadv_tp_fill_prerandom - Fill buffer with prefetched random bytes
 534  * @tp_vars: the private TP meter data for this session
 535  * @buf: Buffer to fill with bytes
 536  * @nbytes: amount of pseudorandom bytes
 537  */
 538 static void batadv_tp_fill_prerandom(struct batadv_tp_vars *tp_vars,
 539                                      u8 *buf, size_t nbytes)
 540 {
 541         u32 local_offset;
 542         size_t bytes_inbuf;
 543         size_t to_copy;
 544         size_t pos = 0;
 545
 546         spin_lock_bh(&tp_vars->prerandom_lock);
 547         local_offset = tp_vars->prerandom_offset;
 548         tp_vars->prerandom_offset += nbytes;
 549         tp_vars->prerandom_offset %= sizeof(batadv_tp_prerandom);
 550         spin_unlock_bh(&tp_vars->prerandom_lock);
 551
 552         while (nbytes) {
 553                 local_offset %= sizeof(batadv_tp_prerandom);
 554                 bytes_inbuf = sizeof(batadv_tp_prerandom) - local_offset;
 555                 to_copy = min(nbytes, bytes_inbuf);
 556
 557                 memcpy(&buf[pos], &batadv_tp_prerandom[local_offset], to_copy);
 558                 pos += to_copy;
 559                 nbytes -= to_copy;
 560                 local_offset = 0;
 561         }
 562 }
 563
 564 /**
 565  * batadv_tp_send_msg - send a single message
 566  * @tp_vars: the private TP meter data for this session
 567  * @src: source mac address
 568  * @orig_node: the originator of the destination
 569  * @seqno: sequence number of this packet
 570  * @len: length of the entire packet
 571  * @session: session identifier
 572  * @uid: local ICMP "socket" index
 573  * @timestamp: timestamp in jiffies which is replied in ack
 574  *
 575  * Create and send a single TP Meter message.
 576  *
 577  * Return: 0 on success, BATADV_TP_REASON_DST_UNREACHABLE if the destination is
 578  * not reachable, BATADV_TP_REASON_MEMORY_ERROR if the packet couldn't be
 579  * allocated
 580  */
 581 static int batadv_tp_send_msg(struct batadv_tp_vars *tp_vars, const u8 *src,
 582                               struct batadv_orig_node *orig_node,
 583                               u32 seqno, size_t len, const u8 *session,
 584                               int uid, u32 timestamp)
 585 {
 586         struct batadv_icmp_tp_packet *icmp;
 587         struct sk_buff *skb;
 588         int r;
 589         u8 *data;
 590         size_t data_len;
 591
 592         skb = netdev_alloc_skb_ip_align(NULL, len + ETH_HLEN);
 593         if (unlikely(!skb))
 594                 return BATADV_TP_REASON_MEMORY_ERROR;
 595
 596         skb_reserve(skb, ETH_HLEN);
 597         icmp = (struct batadv_icmp_tp_packet *)skb_put(skb, sizeof(*icmp));
 598
 599         /* fill the icmp header */
 600         ether_addr_copy(icmp->dst, orig_node->orig);
 601         ether_addr_copy(icmp->orig, src);
 602         icmp->version = BATADV_COMPAT_VERSION;
 603         icmp->packet_type = BATADV_ICMP;
 604         icmp->ttl = BATADV_TTL;
 605         icmp->msg_type = BATADV_TP;
 606         icmp->uid = uid;
 607
 608         icmp->subtype = BATADV_TP_MSG;
 609         memcpy(icmp->session, session, sizeof(icmp->session));
 610         icmp->seqno = htonl(seqno);
 611         icmp->timestamp = htonl(timestamp);
 612
 613         data_len = len - sizeof(*icmp);
 614         data = (u8 *)skb_put(skb, data_len);
 615         batadv_tp_fill_prerandom(tp_vars, data, data_len);
 616
 617         r = batadv_send_skb_to_orig(skb, orig_node, NULL);
 618         if (r == NET_XMIT_SUCCESS)
 619                 return 0;
 620
 621         return BATADV_TP_REASON_CANT_SEND;
 622 }
 623
 624 /**
 625  * batadv_tp_recv_ack - ACK receiving function
 626  * @bat_priv: the bat priv with all the soft interface information
 627  * @skb: the buffer containing the received packet
 628  *
 629  * Process a received TP ACK packet
 630  */
 631 static void batadv_tp_recv_ack(struct batadv_priv *bat_priv,
 632                                const struct sk_buff *skb)
 633 {
 634         struct batadv_hard_iface *primary_if = NULL;
 635         struct batadv_orig_node *orig_node = NULL;
 636         const struct batadv_icmp_tp_packet *icmp;
 637         struct batadv_tp_vars *tp_vars;
 638         size_t packet_len, mss;
 639         u32 rtt, recv_ack, cwnd;
 640         unsigned char *dev_addr;
 641
 642         packet_len = BATADV_TP_PLEN;
 643         mss = BATADV_TP_PLEN;
 644         packet_len += sizeof(struct batadv_unicast_packet);
 645
 646         icmp = (struct batadv_icmp_tp_packet *)skb->data;
 647
 648         /* find the tp_vars */
 649         tp_vars = batadv_tp_list_find_session(bat_priv, icmp->orig,
 650                                               icmp->session);
 651         if (unlikely(!tp_vars))
 652                 return;
 653
 654         if (unlikely(atomic_read(&tp_vars->sending) == 0))
 655                 goto out;
 656
 657         /* old ACK? silently drop it.. */
 658         if (batadv_seq_before(ntohl(icmp->seqno),
 659                               (u32)atomic_read(&tp_vars->last_acked)))
 660                 goto out;
 661
 662         primary_if = batadv_primary_if_get_selected(bat_priv);
 663         if (unlikely(!primary_if))
 664                 goto out;
 665
 666         orig_node = batadv_orig_hash_find(bat_priv, icmp->orig);
 667         if (unlikely(!orig_node))
 668                 goto out;
 669
 670         /* update RTO with the new sampled RTT, if any */
 671         rtt = jiffies_to_msecs(jiffies) - ntohl(icmp->timestamp);
 672         if (icmp->timestamp && rtt)
 673                 batadv_tp_update_rto(tp_vars, rtt);
 674
 675         /* ACK for new data... reset the timer */
 676         batadv_tp_reset_sender_timer(tp_vars);
 677
 678         recv_ack = ntohl(icmp->seqno);
 679
 680         /* check if this ACK is a duplicate */
 681         if (atomic_read(&tp_vars->last_acked) == recv_ack) {
 682                 atomic_inc(&tp_vars->dup_acks);
 683                 if (atomic_read(&tp_vars->dup_acks) != 3)
 684                         goto out;
 685
 686                 if (recv_ack >= tp_vars->recover)
 687                         goto out;
 688
 689                 /* if this is the third duplicate ACK do Fast Retransmit */
 690                 batadv_tp_send_msg(tp_vars, primary_if->net_dev->dev_addr,
 691                                    orig_node, recv_ack, packet_len,
 692                                    icmp->session, icmp->uid,
 693                                    jiffies_to_msecs(jiffies));
 694
 695                 spin_lock_bh(&tp_vars->cwnd_lock);
 696
 697                 /* Fast Recovery */
 698                 tp_vars->fast_recovery = true;
 699                 /* Set recover to the last outstanding seqno when Fast Recovery
 700                  * is entered. RFC6582, Section 3.2, step 1
 701                  */
 702                 tp_vars->recover = tp_vars->last_sent;
 703                 tp_vars->ss_threshold = tp_vars->cwnd >> 1;
 704                 batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
 705                            "Meter: Fast Recovery, (cur cwnd=%u) ss_thr=%u last_sent=%u recv_ack=%u\n",
 706                            tp_vars->cwnd, tp_vars->ss_threshold,
 707                            tp_vars->last_sent, recv_ack);
 708                 tp_vars->cwnd = batadv_tp_cwnd(tp_vars->ss_threshold, 3 * mss,
 709                                                mss);
 710                 tp_vars->dec_cwnd = 0;
 711                 tp_vars->last_sent = recv_ack;
 712
 713                 spin_unlock_bh(&tp_vars->cwnd_lock);
 714         } else {
 715                 /* count the acked data */
 716                 atomic64_add(recv_ack - atomic_read(&tp_vars->last_acked),
 717                              &tp_vars->tot_sent);
 718                 /* reset the duplicate ACKs counter */
 719                 atomic_set(&tp_vars->dup_acks, 0);
 720
 721                 if (tp_vars->fast_recovery) {
 722                         /* partial ACK */
 723                         if (batadv_seq_before(recv_ack, tp_vars->recover)) {
 724                                 /* this is another hole in the window. React
 725                                  * immediately as specified by NewReno (see
 726                                  * Section 3.2 of RFC6582 for details)
 727                                  */
 728                                 dev_addr = primary_if->net_dev->dev_addr;
 729                                 batadv_tp_send_msg(tp_vars, dev_addr,
 730                                                    orig_node, recv_ack,
 731                                                    packet_len, icmp->session,
 732                                                    icmp->uid,
 733                                                    jiffies_to_msecs(jiffies));
 734                                 tp_vars->cwnd = batadv_tp_cwnd(tp_vars->cwnd,
 735                                                                mss, mss);
 736                         } else {
 737                                 tp_vars->fast_recovery = false;
 738                                 /* set cwnd to the value of ss_threshold at the
 739                                  * moment that Fast Recovery was entered.
 740                                  * RFC6582, Section 3.2, step 3
 741                                  */
 742                                 cwnd = batadv_tp_cwnd(tp_vars->ss_threshold, 0,
 743                                                       mss);
 744                                 tp_vars->cwnd = cwnd;
 745                         }
 746                         goto move_twnd;
 747                 }
 748
 749                 if (recv_ack - atomic_read(&tp_vars->last_acked) >= mss)
 750                         batadv_tp_update_cwnd(tp_vars, mss);
 751 move_twnd:
 752                 /* move the Transmit Window */
 753                 atomic_set(&tp_vars->last_acked, recv_ack);
 754         }
 755
 756         wake_up(&tp_vars->more_bytes);
 757 out:
 758         if (likely(primary_if))
 759                 batadv_hardif_put(primary_if);
 760         if (likely(orig_node))
 761                 batadv_orig_node_put(orig_node);
 762         if (likely(tp_vars))
 763                 batadv_tp_vars_put(tp_vars);
 764 }
 765
 766 /**
 767  * batadv_tp_avail - check if congestion window is not full
 768  * @tp_vars: the private data of the current TP meter session
 769  * @payload_len: size of the payload of a single message
 770  *
 771  * Return: true when congestion window is not full, false otherwise
 772  */
 773 static bool batadv_tp_avail(struct batadv_tp_vars *tp_vars,
 774                             size_t payload_len)
 775 {
 776         u32 win_left, win_limit;
 777
 778         win_limit = atomic_read(&tp_vars->last_acked) + tp_vars->cwnd;
 779         win_left = win_limit - tp_vars->last_sent;
 780
 781         return win_left >= payload_len;
 782 }
 783
 784 /**
 785  * batadv_tp_wait_available - wait until congestion window becomes free or
 786  *  timeout is reached
 787  * @tp_vars: the private data of the current TP meter session
 788  * @plen: size of the payload of a single message
 789  *
 790  * Return: 0 if the condition evaluated to false after the timeout elapsed,
 791  *  1 if the condition evaluated to true after the timeout elapsed, the
 792  *  remaining jiffies (at least 1) if the condition evaluated to true before
 793  *  the timeout elapsed, or -ERESTARTSYS if it was interrupted by a signal.
 794  */
 795 static int batadv_tp_wait_available(struct batadv_tp_vars *tp_vars, size_t plen)
 796 {
 797         int ret;
 798
 799         ret = wait_event_interruptible_timeout(tp_vars->more_bytes,
 800                                                batadv_tp_avail(tp_vars, plen),
 801                                                HZ / 10);
 802
 803         return ret;
 804 }
 805
 806 /**
 807  * batadv_tp_send - main sending thread of a tp meter session
 808  * @arg: address of the related tp_vars
 809  *
 810  * Return: nothing, this function never returns
 811  */
 812 static int batadv_tp_send(void *arg)
 813 {
 814         struct batadv_tp_vars *tp_vars = arg;
 815         struct batadv_priv *bat_priv = tp_vars->bat_priv;
 816         struct batadv_hard_iface *primary_if = NULL;
 817         struct batadv_orig_node *orig_node = NULL;
 818         size_t payload_len, packet_len;
 819         int err = 0;
 820
 821         if (unlikely(tp_vars->role != BATADV_TP_SENDER)) {
 822                 err = BATADV_TP_REASON_DST_UNREACHABLE;
 823                 tp_vars->reason = err;
 824                 goto out;
 825         }
 826
 827         orig_node = batadv_orig_hash_find(bat_priv, tp_vars->other_end);
 828         if (unlikely(!orig_node)) {
 829                 err = BATADV_TP_REASON_DST_UNREACHABLE;
 830                 tp_vars->reason = err;
 831                 goto out;
 832         }
 833
 834         primary_if = batadv_primary_if_get_selected(bat_priv);
 835         if (unlikely(!primary_if)) {
 836                 err = BATADV_TP_REASON_DST_UNREACHABLE;
 837                 tp_vars->reason = err;
 838                 goto out;
 839         }
 840
 841         /* assume that all the hard_interfaces have a correctly
 842          * configured MTU, so use the soft_iface MTU as MSS.
 843          * This might not be true and in that case the fragmentation
 844          * should be used.
 845          * Now, try to send the packet as it is
 846          */
 847         payload_len = BATADV_TP_PLEN;
 848         BUILD_BUG_ON(sizeof(struct batadv_icmp_tp_packet) > BATADV_TP_PLEN);
 849
 850         batadv_tp_reset_sender_timer(tp_vars);
 851
 852         /* queue the worker in charge of terminating the test */
 853         queue_delayed_work(batadv_event_workqueue, &tp_vars->finish_work,
 854                            msecs_to_jiffies(tp_vars->test_length));
 855
 856         while (atomic_read(&tp_vars->sending) != 0) {
 857                 if (unlikely(!batadv_tp_avail(tp_vars, payload_len))) {
 858                         batadv_tp_wait_available(tp_vars, payload_len);
 859                         continue;
 860                 }
 861
 862                 /* to emulate normal unicast traffic, add to the payload len
 863                  * the size of the unicast header
 864                  */
 865                 packet_len = payload_len + sizeof(struct batadv_unicast_packet);
 866
 867                 err = batadv_tp_send_msg(tp_vars, primary_if->net_dev->dev_addr,
 868                                          orig_node, tp_vars->last_sent,
 869                                          packet_len,
 870                                          tp_vars->session, tp_vars->icmp_uid,
 871                                          jiffies_to_msecs(jiffies));
 872
 873                 /* something went wrong during the preparation/transmission */
 874                 if (unlikely(err && err != BATADV_TP_REASON_CANT_SEND)) {
 875                         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
 876                                    "Meter: %s() cannot send packets (%d)\n",
 877                                    __func__, err);
 878                         /* ensure nobody else tries to stop the thread now */
 879                         if (atomic_dec_and_test(&tp_vars->sending))
 880                                 tp_vars->reason = err;
 881                         break;
 882                 }
 883
 884                 /* right-shift the TWND */
 885                 if (!err)
 886                         tp_vars->last_sent += payload_len;
 887
 888                 cond_resched();
 889         }
 890
 891 out:
 892         if (likely(primary_if))
 893                 batadv_hardif_put(primary_if);
 894         if (likely(orig_node))
 895                 batadv_orig_node_put(orig_node);
 896
 897         batadv_tp_sender_end(bat_priv, tp_vars);
 898         batadv_tp_sender_cleanup(bat_priv, tp_vars);
 899
 900         batadv_tp_vars_put(tp_vars);
 901
 902         do_exit(0);
 903 }
 904
 905 /**
 906  * batadv_tp_start_kthread - start new thread which manages the tp meter sender
 907  * @tp_vars: the private data of the current TP meter session
 908  */
 909 static void batadv_tp_start_kthread(struct batadv_tp_vars *tp_vars)
 910 {
 911         struct task_struct *kthread;
 912         struct batadv_priv *bat_priv = tp_vars->bat_priv;
 913         u32 session_cookie;
 914
 915         kref_get(&tp_vars->refcount);
 916         kthread = kthread_create(batadv_tp_send, tp_vars, "kbatadv_tp_meter");
 917         if (IS_ERR(kthread)) {
 918                 session_cookie = batadv_tp_session_cookie(tp_vars->session,
 919                                                           tp_vars->icmp_uid);
 920                 pr_err("batadv: cannot create tp meter kthread\n");
 921                 batadv_tp_batctl_error_notify(BATADV_TP_REASON_MEMORY_ERROR,
 922                                               tp_vars->other_end,
 923                                               bat_priv, session_cookie);
 924
 925                 /* drop reserved reference for kthread */
 926                 batadv_tp_vars_put(tp_vars);
 927
 928                 /* cleanup of failed tp meter variables */
 929                 batadv_tp_sender_cleanup(bat_priv, tp_vars);
 930                 return;
 931         }
 932
 933         wake_up_process(kthread);
 934 }
 935
 936 /**
 937  * batadv_tp_start - start a new tp meter session
 938  * @bat_priv: the bat priv with all the soft interface information
 939  * @dst: the receiver MAC address
 940  * @test_length: test length in milliseconds
 941  * @cookie: session cookie
 942  */
 943 void batadv_tp_start(struct batadv_priv *bat_priv, const u8 *dst,
 944                      u32 test_length, u32 *cookie)
 945 {
 946         struct batadv_tp_vars *tp_vars;
 947         u8 session_id[2];
 948         u8 icmp_uid;
 949         u32 session_cookie;
 950
 951         get_random_bytes(session_id, sizeof(session_id));
 952         get_random_bytes(&icmp_uid, 1);
 953         session_cookie = batadv_tp_session_cookie(session_id, icmp_uid);
 954         *cookie = session_cookie;
 955
 956         /* look for an already existing test towards this node */
 957         spin_lock_bh(&bat_priv->tp_list_lock);
 958         tp_vars = batadv_tp_list_find(bat_priv, dst);
 959         if (tp_vars) {
 960                 spin_unlock_bh(&bat_priv->tp_list_lock);
 961                 batadv_tp_vars_put(tp_vars);
 962                 batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
 963                            "Meter: test to or from the same node already ongoing, aborting\n");
 964                 batadv_tp_batctl_error_notify(BATADV_TP_REASON_ALREADY_ONGOING,
 965                                               dst, bat_priv, session_cookie);
 966                 return;
 967         }
 968
 969         if (!atomic_add_unless(&bat_priv->tp_num, 1, BATADV_TP_MAX_NUM)) {
 970                 spin_unlock_bh(&bat_priv->tp_list_lock);
 971                 batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
 972                            "Meter: too many ongoing sessions, aborting (SEND)\n");
 973                 batadv_tp_batctl_error_notify(BATADV_TP_REASON_TOO_MANY, dst,
 974                                               bat_priv, session_cookie);
 975                 return;
 976         }
 977
 978         tp_vars = kmalloc(sizeof(*tp_vars), GFP_ATOMIC);
 979         if (!tp_vars) {
 980                 spin_unlock_bh(&bat_priv->tp_list_lock);
 981                 batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
 982                            "Meter: %s cannot allocate list elements\n",
 983                            __func__);
 984                 batadv_tp_batctl_error_notify(BATADV_TP_REASON_MEMORY_ERROR,
 985                                               dst, bat_priv, session_cookie);
 986                 return;
 987         }
 988
 989         /* initialize tp_vars */
 990         ether_addr_copy(tp_vars->other_end, dst);
 991         kref_init(&tp_vars->refcount);
 992         tp_vars->role = BATADV_TP_SENDER;
 993         atomic_set(&tp_vars->sending, 1);
 994         memcpy(tp_vars->session, session_id, sizeof(session_id));
 995         tp_vars->icmp_uid = icmp_uid;
 996
 997         tp_vars->last_sent = BATADV_TP_FIRST_SEQ;
 998         atomic_set(&tp_vars->last_acked, BATADV_TP_FIRST_SEQ);
 999         tp_vars->fast_recovery = false;
1000         tp_vars->recover = BATADV_TP_FIRST_SEQ;
1001
1002         /* initialise the CWND to 3*MSS (Section 3.1 in RFC5681).
1003          * For batman-adv the MSS is the size of the payload received by the
1004          * soft_interface, hence its MTU
1005          */
1006         tp_vars->cwnd = BATADV_TP_PLEN * 3;
1007         /* at the beginning initialise the SS threshold to the biggest possible
1008          * window size, hence the AWND size
1009          */
1010         tp_vars->ss_threshold = BATADV_TP_AWND;
1011
1012         /* RTO initial value is 3 seconds.
1013          * Details in Section 2.1 of RFC6298
1014          */
1015         tp_vars->rto = 1000;
1016         tp_vars->srtt = 0;
1017         tp_vars->rttvar = 0;
1018
1019         atomic64_set(&tp_vars->tot_sent, 0);
1020
1021         kref_get(&tp_vars->refcount);
1022         setup_timer(&tp_vars->timer, batadv_tp_sender_timeout,
1023                     (unsigned long)tp_vars);
1024
1025         tp_vars->bat_priv = bat_priv;
1026         tp_vars->start_time = jiffies;
1027
1028         init_waitqueue_head(&tp_vars->more_bytes);
1029
1030         spin_lock_init(&tp_vars->unacked_lock);
1031         INIT_LIST_HEAD(&tp_vars->unacked_list);
1032
1033         spin_lock_init(&tp_vars->cwnd_lock);
1034
1035         tp_vars->prerandom_offset = 0;
1036         spin_lock_init(&tp_vars->prerandom_lock);
1037
1038         kref_get(&tp_vars->refcount);
1039         hlist_add_head_rcu(&tp_vars->list, &bat_priv->tp_list);
1040         spin_unlock_bh(&bat_priv->tp_list_lock);
1041
1042         tp_vars->test_length = test_length;
1043         if (!tp_vars->test_length)
1044                 tp_vars->test_length = BATADV_TP_DEF_TEST_LENGTH;
1045
1046         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
1047                    "Meter: starting throughput meter towards %pM (length=%ums)\n",
1048                    dst, test_length);
1049
1050         /* init work item for finished tp tests */
1051         INIT_DELAYED_WORK(&tp_vars->finish_work, batadv_tp_sender_finish);
1052
1053         /* start tp kthread. This way the write() call issued from userspace can
1054          * happily return and avoid to block
1055          */
1056         batadv_tp_start_kthread(tp_vars);
1057
1058         /* don't return reference to new tp_vars */
1059         batadv_tp_vars_put(tp_vars);
1060 }
1061
1062 /**
1063  * batadv_tp_stop - stop currently running tp meter session
1064  * @bat_priv: the bat priv with all the soft interface information
1065  * @dst: the receiver MAC address
1066  * @return_value: reason for tp meter session stop
1067  */
1068 void batadv_tp_stop(struct batadv_priv *bat_priv, const u8 *dst,
1069                     u8 return_value)
1070 {
1071         struct batadv_orig_node *orig_node;
1072         struct batadv_tp_vars *tp_vars;
1073
1074         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
1075                    "Meter: stopping test towards %pM\n", dst);
1076
1077         orig_node = batadv_orig_hash_find(bat_priv, dst);
1078         if (!orig_node)
1079                 return;
1080
1081         tp_vars = batadv_tp_list_find(bat_priv, orig_node->orig);
1082         if (!tp_vars) {
1083                 batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
1084                            "Meter: trying to interrupt an already over connection\n");
1085                 goto out;
1086         }
1087
1088         batadv_tp_sender_shutdown(tp_vars, return_value);
1089         batadv_tp_vars_put(tp_vars);
1090 out:
1091         batadv_orig_node_put(orig_node);
1092 }
1093
1094 /**
1095  * batadv_tp_reset_receiver_timer - reset the receiver shutdown timer
1096  * @tp_vars: the private data of the current TP meter session
1097  *
1098  * start the receiver shutdown timer or reset it if already started
1099  */
1100 static void batadv_tp_reset_receiver_timer(struct batadv_tp_vars *tp_vars)
1101 {
1102         mod_timer(&tp_vars->timer,
1103                   jiffies + msecs_to_jiffies(BATADV_TP_RECV_TIMEOUT));
1104 }
1105
1106 /**
1107  * batadv_tp_receiver_shutdown - stop a tp meter receiver when timeout is
1108  *  reached without received ack
1109  * @arg: address of the related tp_vars
1110  */
1111 static void batadv_tp_receiver_shutdown(unsigned long arg)
1112 {
1113         struct batadv_tp_vars *tp_vars = (struct batadv_tp_vars *)arg;
1114         struct batadv_tp_unacked *un, *safe;
1115         struct batadv_priv *bat_priv;
1116
1117         bat_priv = tp_vars->bat_priv;
1118
1119         /* if there is recent activity rearm the timer */
1120         if (!batadv_has_timed_out(tp_vars->last_recv_time,
1121                                   BATADV_TP_RECV_TIMEOUT)) {
1122                 /* reset the receiver shutdown timer */
1123                 batadv_tp_reset_receiver_timer(tp_vars);
1124                 return;
1125         }
1126
1127         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
1128                    "Shutting down for inactivity (more than %dms) from %pM\n",
1129                    BATADV_TP_RECV_TIMEOUT, tp_vars->other_end);
1130
1131         spin_lock_bh(&tp_vars->bat_priv->tp_list_lock);
1132         hlist_del_rcu(&tp_vars->list);
1133         spin_unlock_bh(&tp_vars->bat_priv->tp_list_lock);
1134
1135         /* drop list reference */
1136         batadv_tp_vars_put(tp_vars);
1137
1138         atomic_dec(&bat_priv->tp_num);
1139
1140         spin_lock_bh(&tp_vars->unacked_lock);
1141         list_for_each_entry_safe(un, safe, &tp_vars->unacked_list, list) {
1142                 list_del(&un->list);
1143                 kfree(un);
1144         }
1145         spin_unlock_bh(&tp_vars->unacked_lock);
1146
1147         /* drop reference of timer */
1148         batadv_tp_vars_put(tp_vars);
1149 }
1150
1151 /**
1152  * batadv_tp_send_ack - send an ACK packet
1153  * @bat_priv: the bat priv with all the soft interface information
1154  * @dst: the mac address of the destination originator
1155  * @seq: the sequence number to ACK
1156  * @timestamp: the timestamp to echo back in the ACK
1157  * @session: session identifier
1158  * @socket_index: local ICMP socket identifier
1159  *
1160  * Return: 0 on success, a positive integer representing the reason of the
1161  * failure otherwise
1162  */
1163 static int batadv_tp_send_ack(struct batadv_priv *bat_priv, const u8 *dst,
1164                               u32 seq, __be32 timestamp, const u8 *session,
1165                               int socket_index)
1166 {
1167         struct batadv_hard_iface *primary_if = NULL;
1168         struct batadv_orig_node *orig_node;
1169         struct batadv_icmp_tp_packet *icmp;
1170         struct sk_buff *skb;
1171         int r, ret;
1172
1173         orig_node = batadv_orig_hash_find(bat_priv, dst);
1174         if (unlikely(!orig_node)) {
1175                 ret = BATADV_TP_REASON_DST_UNREACHABLE;
1176                 goto out;
1177         }
1178
1179         primary_if = batadv_primary_if_get_selected(bat_priv);
1180         if (unlikely(!primary_if)) {
1181                 ret = BATADV_TP_REASON_DST_UNREACHABLE;
1182                 goto out;
1183         }
1184
1185         skb = netdev_alloc_skb_ip_align(NULL, sizeof(*icmp) + ETH_HLEN);
1186         if (unlikely(!skb)) {
1187                 ret = BATADV_TP_REASON_MEMORY_ERROR;
1188                 goto out;
1189         }
1190
1191         skb_reserve(skb, ETH_HLEN);
1192         icmp = (struct batadv_icmp_tp_packet *)skb_put(skb, sizeof(*icmp));
1193         icmp->packet_type = BATADV_ICMP;
1194         icmp->version = BATADV_COMPAT_VERSION;
1195         icmp->ttl = BATADV_TTL;
1196         icmp->msg_type = BATADV_TP;
1197         ether_addr_copy(icmp->dst, orig_node->orig);
1198         ether_addr_copy(icmp->orig, primary_if->net_dev->dev_addr);
1199         icmp->uid = socket_index;
1200
1201         icmp->subtype = BATADV_TP_ACK;
1202         memcpy(icmp->session, session, sizeof(icmp->session));
1203         icmp->seqno = htonl(seq);
1204         icmp->timestamp = timestamp;
1205
1206         /* send the ack */
1207         r = batadv_send_skb_to_orig(skb, orig_node, NULL);
1208         if (unlikely(r < 0) || (r == NET_XMIT_DROP)) {
1209                 ret = BATADV_TP_REASON_DST_UNREACHABLE;
1210                 goto out;
1211         }
1212         ret = 0;
1213
1214 out:
1215         if (likely(orig_node))
1216                 batadv_orig_node_put(orig_node);
1217         if (likely(primary_if))
1218                 batadv_hardif_put(primary_if);
1219
1220         return ret;
1221 }
1222
1223 /**
1224  * batadv_tp_handle_out_of_order - store an out of order packet
1225  * @tp_vars: the private data of the current TP meter session
1226  * @skb: the buffer containing the received packet
1227  *
1228  * Store the out of order packet in the unacked list for late processing. This
1229  * packets are kept in this list so that they can be ACKed at once as soon as
1230  * all the previous packets have been received
1231  *
1232  * Return: true if the packed has been successfully processed, false otherwise
1233  */
1234 static bool batadv_tp_handle_out_of_order(struct batadv_tp_vars *tp_vars,
1235                                           const struct sk_buff *skb)
1236 {
1237         const struct batadv_icmp_tp_packet *icmp;
1238         struct batadv_tp_unacked *un, *new;
1239         u32 payload_len;
1240         bool added = false;
1241
1242         new = kmalloc(sizeof(*new), GFP_ATOMIC);
1243         if (unlikely(!new))
1244                 return false;
1245
1246         icmp = (struct batadv_icmp_tp_packet *)skb->data;
1247
1248         new->seqno = ntohl(icmp->seqno);
1249         payload_len = skb->len - sizeof(struct batadv_unicast_packet);
1250         new->len = payload_len;
1251
1252         spin_lock_bh(&tp_vars->unacked_lock);
1253         /* if the list is empty immediately attach this new object */
1254         if (list_empty(&tp_vars->unacked_list)) {
1255                 list_add(&new->list, &tp_vars->unacked_list);
1256                 goto out;
1257         }
1258
1259         /* otherwise loop over the list and either drop the packet because this
1260          * is a duplicate or store it at the right position.
1261          *
1262          * The iteration is done in the reverse way because it is likely that
1263          * the last received packet (the one being processed now) has a bigger
1264          * seqno than all the others already stored.
1265          */
1266         list_for_each_entry_reverse(un, &tp_vars->unacked_list, list) {
1267                 /* check for duplicates */
1268                 if (new->seqno == un->seqno) {
1269                         if (new->len > un->len)
1270                                 un->len = new->len;
1271                         kfree(new);
1272                         added = true;
1273                         break;
1274                 }
1275
1276                 /* look for the right position */
1277                 if (batadv_seq_before(new->seqno, un->seqno))
1278                         continue;
1279
1280                 /* as soon as an entry having a bigger seqno is found, the new
1281                  * one is attached _after_ it. In this way the list is kept in
1282                  * ascending order
1283                  */
1284                 list_add_tail(&new->list, &un->list);
1285                 added = true;
1286                 break;
1287         }
1288
1289         /* received packet with smallest seqno out of order; add it to front */
1290         if (!added)
1291                 list_add(&new->list, &tp_vars->unacked_list);
1292
1293 out:
1294         spin_unlock_bh(&tp_vars->unacked_lock);
1295
1296         return true;
1297 }
1298
1299 /**
1300  * batadv_tp_ack_unordered - update number received bytes in current stream
1301  *  without gaps
1302  * @tp_vars: the private data of the current TP meter session
1303  */
1304 static void batadv_tp_ack_unordered(struct batadv_tp_vars *tp_vars)
1305 {
1306         struct batadv_tp_unacked *un, *safe;
1307         u32 to_ack;
1308
1309         /* go through the unacked packet list and possibly ACK them as
1310          * well
1311          */
1312         spin_lock_bh(&tp_vars->unacked_lock);
1313         list_for_each_entry_safe(un, safe, &tp_vars->unacked_list, list) {
1314                 /* the list is ordered, therefore it is possible to stop as soon
1315                  * there is a gap between the last acked seqno and the seqno of
1316                  * the packet under inspection
1317                  */
1318                 if (batadv_seq_before(tp_vars->last_recv, un->seqno))
1319                         break;
1320
1321                 to_ack = un->seqno + un->len - tp_vars->last_recv;
1322
1323                 if (batadv_seq_before(tp_vars->last_recv, un->seqno + un->len))
1324                         tp_vars->last_recv += to_ack;
1325
1326                 list_del(&un->list);
1327                 kfree(un);
1328         }
1329         spin_unlock_bh(&tp_vars->unacked_lock);
1330 }
1331
1332 /**
1333  * batadv_tp_init_recv - return matching or create new receiver tp_vars
1334  * @bat_priv: the bat priv with all the soft interface information
1335  * @icmp: received icmp tp msg
1336  *
1337  * Return: corresponding tp_vars or NULL on errors
1338  */
1339 static struct batadv_tp_vars *
1340 batadv_tp_init_recv(struct batadv_priv *bat_priv,
1341                     const struct batadv_icmp_tp_packet *icmp)
1342 {
1343         struct batadv_tp_vars *tp_vars;
1344
1345         spin_lock_bh(&bat_priv->tp_list_lock);
1346         tp_vars = batadv_tp_list_find_session(bat_priv, icmp->orig,
1347                                               icmp->session);
1348         if (tp_vars)
1349                 goto out_unlock;
1350
1351         if (!atomic_add_unless(&bat_priv->tp_num, 1, BATADV_TP_MAX_NUM)) {
1352                 batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
1353                            "Meter: too many ongoing sessions, aborting (RECV)\n");
1354                 goto out_unlock;
1355         }
1356
1357         tp_vars = kmalloc(sizeof(*tp_vars), GFP_ATOMIC);
1358         if (!tp_vars)
1359                 goto out_unlock;
1360
1361         ether_addr_copy(tp_vars->other_end, icmp->orig);
1362         tp_vars->role = BATADV_TP_RECEIVER;
1363         memcpy(tp_vars->session, icmp->session, sizeof(tp_vars->session));
1364         tp_vars->last_recv = BATADV_TP_FIRST_SEQ;
1365         tp_vars->bat_priv = bat_priv;
1366         kref_init(&tp_vars->refcount);
1367
1368         spin_lock_init(&tp_vars->unacked_lock);
1369         INIT_LIST_HEAD(&tp_vars->unacked_list);
1370
1371         kref_get(&tp_vars->refcount);
1372         hlist_add_head_rcu(&tp_vars->list, &bat_priv->tp_list);
1373
1374         kref_get(&tp_vars->refcount);
1375         setup_timer(&tp_vars->timer, batadv_tp_receiver_shutdown,
1376                     (unsigned long)tp_vars);
1377
1378         batadv_tp_reset_receiver_timer(tp_vars);
1379
1380 out_unlock:
1381         spin_unlock_bh(&bat_priv->tp_list_lock);
1382
1383         return tp_vars;
1384 }
1385
1386 /**
1387  * batadv_tp_recv_msg - process a single data message
1388  * @bat_priv: the bat priv with all the soft interface information
1389  * @skb: the buffer containing the received packet
1390  *
1391  * Process a received TP MSG packet
1392  */
1393 static void batadv_tp_recv_msg(struct batadv_priv *bat_priv,
1394                                const struct sk_buff *skb)
1395 {
1396         const struct batadv_icmp_tp_packet *icmp;
1397         struct batadv_tp_vars *tp_vars;
1398         size_t packet_size;
1399         u32 seqno;
1400
1401         icmp = (struct batadv_icmp_tp_packet *)skb->data;
1402
1403         seqno = ntohl(icmp->seqno);
1404         /* check if this is the first seqno. This means that if the
1405          * first packet is lost, the tp meter does not work anymore!
1406          */
1407         if (seqno == BATADV_TP_FIRST_SEQ) {
1408                 tp_vars = batadv_tp_init_recv(bat_priv, icmp);
1409                 if (!tp_vars) {
1410                         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
1411                                    "Meter: seqno != BATADV_TP_FIRST_SEQ cannot initiate connection\n");
1412                         goto out;
1413                 }
1414         } else {
1415                 tp_vars = batadv_tp_list_find_session(bat_priv, icmp->orig,
1416                                                       icmp->session);
1417                 if (!tp_vars) {
1418                         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
1419                                    "Unexpected packet from %pM!\n",
1420                                    icmp->orig);
1421                         goto out;
1422                 }
1423         }
1424
1425         if (unlikely(tp_vars->role != BATADV_TP_RECEIVER)) {
1426                 batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
1427                            "Meter: dropping packet: not expected (role=%u)\n",
1428                            tp_vars->role);
1429                 goto out;
1430         }
1431
1432         tp_vars->last_recv_time = jiffies;
1433
1434         /* if the packet is a duplicate, it may be the case that an ACK has been
1435          * lost. Resend the ACK
1436          */
1437         if (batadv_seq_before(seqno, tp_vars->last_recv))
1438                 goto send_ack;
1439
1440         /* if the packet is out of order enqueue it */
1441         if (ntohl(icmp->seqno) != tp_vars->last_recv) {
1442                 /* exit immediately (and do not send any ACK) if the packet has
1443                  * not been enqueued correctly
1444                  */
1445                 if (!batadv_tp_handle_out_of_order(tp_vars, skb))
1446                         goto out;
1447
1448                 /* send a duplicate ACK */
1449                 goto send_ack;
1450         }
1451
1452         /* if everything was fine count the ACKed bytes */
1453         packet_size = skb->len - sizeof(struct batadv_unicast_packet);
1454         tp_vars->last_recv += packet_size;
1455
1456         /* check if this ordered message filled a gap.... */
1457         batadv_tp_ack_unordered(tp_vars);
1458
1459 send_ack:
1460         /* send the ACK. If the received packet was out of order, the ACK that
1461          * is going to be sent is a duplicate (the sender will count them and
1462          * possibly enter Fast Retransmit as soon as it has reached 3)
1463          */
1464         batadv_tp_send_ack(bat_priv, icmp->orig, tp_vars->last_recv,
1465                            icmp->timestamp, icmp->session, icmp->uid);
1466 out:
1467         if (likely(tp_vars))
1468                 batadv_tp_vars_put(tp_vars);
1469 }
1470
1471 /**
1472  * batadv_tp_meter_recv - main TP Meter receiving function
1473  * @bat_priv: the bat priv with all the soft interface information
1474  * @skb: the buffer containing the received packet
1475  */
1476 void batadv_tp_meter_recv(struct batadv_priv *bat_priv, struct sk_buff *skb)
1477 {
1478         struct batadv_icmp_tp_packet *icmp;
1479
1480         icmp = (struct batadv_icmp_tp_packet *)skb->data;
1481
1482         switch (icmp->subtype) {
1483         case BATADV_TP_MSG:
1484                 batadv_tp_recv_msg(bat_priv, skb);
1485                 break;
1486         case BATADV_TP_ACK:
1487                 batadv_tp_recv_ack(bat_priv, skb);
1488                 break;
1489         default:
1490                 batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
1491                            "Received unknown TP Metric packet type %u\n",
1492                            icmp->subtype);
1493         }
1494         consume_skb(skb);
1495 }
1496
1497 /**
1498  * batadv_tp_meter_init - initialize global tp_meter structures
1499  */
1500 void batadv_tp_meter_init(void)
1501 {
1502         get_random_bytes(batadv_tp_prerandom, sizeof(batadv_tp_prerandom));
1503 }