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/init.h>
  31 #include <linux/jiffies.h>
  32 #include <linux/kernel.h>
  33 #include <linux/kref.h>
  34 #include <linux/kthread.h>
  35 #include <linux/list.h>
  36 #include <linux/netdevice.h>
  37 #include <linux/param.h>
  38 #include <linux/printk.h>
  39 #include <linux/random.h>
  40 #include <linux/rculist.h>
  41 #include <linux/rcupdate.h>
  42 #include <linux/sched.h>
  43 #include <linux/skbuff.h>
  44 #include <linux/slab.h>
  45 #include <linux/spinlock.h>
  46 #include <linux/stddef.h>
  47 #include <linux/string.h>
  48 #include <linux/timer.h>
  49 #include <linux/wait.h>
  50 #include <linux/workqueue.h>
  51 #include <uapi/linux/batman_adv.h>
  52
  53 #include "hard-interface.h"
  54 #include "log.h"
  55 #include "netlink.h"
  56 #include "originator.h"
  57 #include "packet.h"
  58 #include "send.h"
  59
  60 /**
  61  * BATADV_TP_DEF_TEST_LENGTH - Default test length if not specified by the user
  62  *  in milliseconds
  63  */
  64 #define BATADV_TP_DEF_TEST_LENGTH 10000
  65
  66 /**
  67  * BATADV_TP_AWND - Advertised window by the receiver (in bytes)
  68  */
  69 #define BATADV_TP_AWND 0x20000000
  70
  71 /**
  72  * BATADV_TP_RECV_TIMEOUT - Receiver activity timeout. If the receiver does not
  73  *  get anything for such amount of milliseconds, the connection is killed
  74  */
  75 #define BATADV_TP_RECV_TIMEOUT 1000
  76
  77 /**
  78  * BATADV_TP_MAX_RTO - Maximum sender timeout. If the sender RTO gets beyond
  79  * such amound of milliseconds, the receiver is considered unreachable and the
  80  * connection is killed
  81  */
  82 #define BATADV_TP_MAX_RTO 30000
  83
  84 /**
  85  * BATADV_TP_FIRST_SEQ - First seqno of each session. The number is rather high
  86  *  in order to immediately trigger a wrap around (test purposes)
  87  */
  88 #define BATADV_TP_FIRST_SEQ ((u32)-1 - 2000)
  89
  90 /**
  91  * BATADV_TP_PLEN - length of the payload (data after the batadv_unicast header)
  92  *  to simulate
  93  */
  94 #define BATADV_TP_PLEN (BATADV_TP_PACKET_LEN - ETH_HLEN - \
  95                         sizeof(struct batadv_unicast_packet))
  96
  97 static u8 batadv_tp_prerandom[4096] __read_mostly;
  98
  99 /**
 100  * batadv_tp_session_cookie - generate session cookie based on session ids
 101  * @session: TP session identifier
 102  * @icmp_uid: icmp pseudo uid of the tp session
 103  *
 104  * Return: 32 bit tp_meter session cookie
 105  */
 106 static u32 batadv_tp_session_cookie(const u8 session[2], u8 icmp_uid)
 107 {
 108         u32 cookie;
 109
 110         cookie = icmp_uid << 16;
 111         cookie |= session[0] << 8;
 112         cookie |= session[1];
 113
 114         return cookie;
 115 }
 116
 117 /**
 118  * batadv_tp_cwnd - compute the new cwnd size
 119  * @base: base cwnd size value
 120  * @increment: the value to add to base to get the new size
 121  * @min: minumim cwnd value (usually MSS)
 122  *
 123  * Return the new cwnd size and ensures it does not exceed the Advertised
 124  * Receiver Window size. It is wrap around safe.
 125  * For details refer to Section 3.1 of RFC5681
 126  *
 127  * Return: new congestion window size in bytes
 128  */
 129 static u32 batadv_tp_cwnd(u32 base, u32 increment, u32 min)
 130 {
 131         u32 new_size = base + increment;
 132
 133         /* check for wrap-around */
 134         if (new_size < base)
 135                 new_size = (u32)ULONG_MAX;
 136
 137         new_size = min_t(u32, new_size, BATADV_TP_AWND);
 138
 139         return max_t(u32, new_size, min);
 140 }
 141
 142 /**
 143  * batadv_tp_updated_cwnd - update the Congestion Windows
 144  * @tp_vars: the private data of the current TP meter session
 145  * @mss: maximum segment size of transmission
 146  *
 147  * 1) if the session is in Slow Start, the CWND has to be increased by 1
 148  * MSS every unique received ACK
 149  * 2) if the session is in Congestion Avoidance, the CWND has to be
 150  * increased by MSS * MSS / CWND for every unique received ACK
 151  */
 152 static void batadv_tp_update_cwnd(struct batadv_tp_vars *tp_vars, u32 mss)
 153 {
 154         spin_lock_bh(&tp_vars->cwnd_lock);
 155
 156         /* slow start... */
 157         if (tp_vars->cwnd <= tp_vars->ss_threshold) {
 158                 tp_vars->dec_cwnd = 0;
 159                 tp_vars->cwnd = batadv_tp_cwnd(tp_vars->cwnd, mss, mss);
 160                 spin_unlock_bh(&tp_vars->cwnd_lock);
 161                 return;
 162         }
 163
 164         /* increment CWND at least of 1 (section 3.1 of RFC5681) */
 165         tp_vars->dec_cwnd += max_t(u32, 1U << 3,
 166                                    ((mss * mss) << 6) / (tp_vars->cwnd << 3));
 167         if (tp_vars->dec_cwnd < (mss << 3)) {
 168                 spin_unlock_bh(&tp_vars->cwnd_lock);
 169                 return;
 170         }
 171
 172         tp_vars->cwnd = batadv_tp_cwnd(tp_vars->cwnd, mss, mss);
 173         tp_vars->dec_cwnd = 0;
 174
 175         spin_unlock_bh(&tp_vars->cwnd_lock);
 176 }
 177
 178 /**
 179  * batadv_tp_update_rto - calculate new retransmission timeout
 180  * @tp_vars: the private data of the current TP meter session
 181  * @new_rtt: new roundtrip time in msec
 182  */
 183 static void batadv_tp_update_rto(struct batadv_tp_vars *tp_vars,
 184                                  u32 new_rtt)
 185 {
 186         long m = new_rtt;
 187
 188         /* RTT update
 189          * Details in Section 2.2 and 2.3 of RFC6298
 190          *
 191          * It's tricky to understand. Don't lose hair please.
 192          * Inspired by tcp_rtt_estimator() tcp_input.c
 193          */
 194         if (tp_vars->srtt != 0) {
 195                 m -= (tp_vars->srtt >> 3); /* m is now error in rtt est */
 196                 tp_vars->srtt += m; /* rtt = 7/8 srtt + 1/8 new */
 197                 if (m < 0)
 198                         m = -m;
 199
 200                 m -= (tp_vars->rttvar >> 2);
 201                 tp_vars->rttvar += m; /* mdev ~= 3/4 rttvar + 1/4 new */
 202         } else {
 203                 /* first measure getting in */
 204                 tp_vars->srtt = m << 3; /* take the measured time to be srtt */
 205                 tp_vars->rttvar = m << 1; /* new_rtt / 2 */
 206         }
 207
 208         /* rto = srtt + 4 * rttvar.
 209          * rttvar is scaled by 4, therefore doesn't need to be multiplied
 210          */
 211         tp_vars->rto = (tp_vars->srtt >> 3) + tp_vars->rttvar;
 212 }
 213
 214 /**
 215  * batadv_tp_batctl_notify - send client status result to client
 216  * @reason: reason for tp meter session stop
 217  * @dst: destination of tp_meter session
 218  * @bat_priv: the bat priv with all the soft interface information
 219  * @start_time: start of transmission in jiffies
 220  * @total_sent: bytes acked to the receiver
 221  * @cookie: cookie of tp_meter session
 222  */
 223 static void batadv_tp_batctl_notify(enum batadv_tp_meter_reason reason,
 224                                     const u8 *dst, struct batadv_priv *bat_priv,
 225                                     unsigned long start_time, u64 total_sent,
 226                                     u32 cookie)
 227 {
 228         u32 test_time;
 229         u8 result;
 230         u32 total_bytes;
 231
 232         if (!batadv_tp_is_error(reason)) {
 233                 result = BATADV_TP_REASON_COMPLETE;
 234                 test_time = jiffies_to_msecs(jiffies - start_time);
 235                 total_bytes = total_sent;
 236         } else {
 237                 result = reason;
 238                 test_time = 0;
 239                 total_bytes = 0;
 240         }
 241
 242         batadv_netlink_tpmeter_notify(bat_priv, dst, result, test_time,
 243                                       total_bytes, cookie);
 244 }
 245
 246 /**
 247  * batadv_tp_batctl_error_notify - send client error result to client
 248  * @reason: reason for tp meter session stop
 249  * @dst: destination of tp_meter session
 250  * @bat_priv: the bat priv with all the soft interface information
 251  * @cookie: cookie of tp_meter session
 252  */
 253 static void batadv_tp_batctl_error_notify(enum batadv_tp_meter_reason reason,
 254                                           const u8 *dst,
 255                                           struct batadv_priv *bat_priv,
 256                                           u32 cookie)
 257 {
 258         batadv_tp_batctl_notify(reason, dst, bat_priv, 0, 0, cookie);
 259 }
 260
 261 /**
 262  * batadv_tp_list_find - find a tp_vars object in the global list
 263  * @bat_priv: the bat priv with all the soft interface information
 264  * @dst: the other endpoint MAC address to look for
 265  *
 266  * Look for a tp_vars object matching dst as end_point and return it after
 267  * having incremented the refcounter. Return NULL is not found
 268  *
 269  * Return: matching tp_vars or NULL when no tp_vars with @dst was found
 270  */
 271 static struct batadv_tp_vars *batadv_tp_list_find(struct batadv_priv *bat_priv,
 272                                                   const u8 *dst)
 273 {
 274         struct batadv_tp_vars *pos, *tp_vars = NULL;
 275
 276         rcu_read_lock();
 277         hlist_for_each_entry_rcu(pos, &bat_priv->tp_list, list) {
 278                 if (!batadv_compare_eth(pos->other_end, dst))
 279                         continue;
 280
 281                 /* most of the time this function is invoked during the normal
 282                  * process..it makes sens to pay more when the session is
 283                  * finished and to speed the process up during the measurement
 284                  */
 285                 if (unlikely(!kref_get_unless_zero(&pos->refcount)))
 286                         continue;
 287
 288                 tp_vars = pos;
 289                 break;
 290         }
 291         rcu_read_unlock();
 292
 293         return tp_vars;
 294 }
 295
 296 /**
 297  * batadv_tp_list_find_session - find tp_vars session object in the global list
 298  * @bat_priv: the bat priv with all the soft interface information
 299  * @dst: the other endpoint MAC address to look for
 300  * @session: session identifier
 301  *
 302  * Look for a tp_vars object matching dst as end_point, session as tp meter
 303  * session and return it after having incremented the refcounter. Return NULL
 304  * is not found
 305  *
 306  * Return: matching tp_vars or NULL when no tp_vars was found
 307  */
 308 static struct batadv_tp_vars *
 309 batadv_tp_list_find_session(struct batadv_priv *bat_priv, const u8 *dst,
 310                             const u8 *session)
 311 {
 312         struct batadv_tp_vars *pos, *tp_vars = NULL;
 313
 314         rcu_read_lock();
 315         hlist_for_each_entry_rcu(pos, &bat_priv->tp_list, list) {
 316                 if (!batadv_compare_eth(pos->other_end, dst))
 317                         continue;
 318
 319                 if (memcmp(pos->session, session, sizeof(pos->session)) != 0)
 320                         continue;
 321
 322                 /* most of the time this function is invoked during the normal
 323                  * process..it makes sense to pay more when the session is
 324                  * finished and to speed the process up during the measurement
 325                  */
 326                 if (unlikely(!kref_get_unless_zero(&pos->refcount)))
 327                         continue;
 328
 329                 tp_vars = pos;
 330                 break;
 331         }
 332         rcu_read_unlock();
 333
 334         return tp_vars;
 335 }
 336
 337 /**
 338  * batadv_tp_vars_release - release batadv_tp_vars from lists and queue for
 339  *  free after rcu grace period
 340  * @ref: kref pointer of the batadv_tp_vars
 341  */
 342 static void batadv_tp_vars_release(struct kref *ref)
 343 {
 344         struct batadv_tp_vars *tp_vars;
 345         struct batadv_tp_unacked *un, *safe;
 346
 347         tp_vars = container_of(ref, struct batadv_tp_vars, refcount);
 348
 349         /* lock should not be needed because this object is now out of any
 350          * context!
 351          */
 352         spin_lock_bh(&tp_vars->unacked_lock);
 353         list_for_each_entry_safe(un, safe, &tp_vars->unacked_list, list) {
 354                 list_del(&un->list);
 355                 kfree(un);
 356         }
 357         spin_unlock_bh(&tp_vars->unacked_lock);
 358
 359         kfree_rcu(tp_vars, rcu);
 360 }
 361
 362 /**
 363  * batadv_tp_vars_put - decrement the batadv_tp_vars refcounter and possibly
 364  *  release it
 365  * @tp_vars: the private data of the current TP meter session to be free'd
 366  */
 367 static void batadv_tp_vars_put(struct batadv_tp_vars *tp_vars)
 368 {
 369         kref_put(&tp_vars->refcount, batadv_tp_vars_release);
 370 }
 371
 372 /**
 373  * batadv_tp_sender_cleanup - cleanup sender data and drop and timer
 374  * @bat_priv: the bat priv with all the soft interface information
 375  * @tp_vars: the private data of the current TP meter session to cleanup
 376  */
 377 static void batadv_tp_sender_cleanup(struct batadv_priv *bat_priv,
 378                                      struct batadv_tp_vars *tp_vars)
 379 {
 380         cancel_delayed_work(&tp_vars->finish_work);
 381
 382         spin_lock_bh(&tp_vars->bat_priv->tp_list_lock);
 383         hlist_del_rcu(&tp_vars->list);
 384         spin_unlock_bh(&tp_vars->bat_priv->tp_list_lock);
 385
 386         /* drop list reference */
 387         batadv_tp_vars_put(tp_vars);
 388
 389         atomic_dec(&tp_vars->bat_priv->tp_num);
 390
 391         /* kill the timer and remove its reference */
 392         del_timer_sync(&tp_vars->timer);
 393         /* the worker might have rearmed itself therefore we kill it again. Note
 394          * that if the worker should run again before invoking the following
 395          * del_timer(), it would not re-arm itself once again because the status
 396          * is OFF now
 397          */
 398         del_timer(&tp_vars->timer);
 399         batadv_tp_vars_put(tp_vars);
 400 }
 401
 402 /**
 403  * batadv_tp_sender_end - print info about ended session and inform client
 404  * @bat_priv: the bat priv with all the soft interface information
 405  * @tp_vars: the private data of the current TP meter session
 406  */
 407 static void batadv_tp_sender_end(struct batadv_priv *bat_priv,
 408                                  struct batadv_tp_vars *tp_vars)
 409 {
 410         u32 session_cookie;
 411
 412         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
 413                    "Test towards %pM finished..shutting down (reason=%d)\n",
 414                    tp_vars->other_end, tp_vars->reason);
 415
 416         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
 417                    "Last timing stats: SRTT=%ums RTTVAR=%ums RTO=%ums\n",
 418                    tp_vars->srtt >> 3, tp_vars->rttvar >> 2, tp_vars->rto);
 419
 420         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
 421                    "Final values: cwnd=%u ss_threshold=%u\n",
 422                    tp_vars->cwnd, tp_vars->ss_threshold);
 423
 424         session_cookie = batadv_tp_session_cookie(tp_vars->session,
 425                                                   tp_vars->icmp_uid);
 426
 427         batadv_tp_batctl_notify(tp_vars->reason,
 428                                 tp_vars->other_end,
 429                                 bat_priv,
 430                                 tp_vars->start_time,
 431                                 atomic64_read(&tp_vars->tot_sent),
 432                                 session_cookie);
 433 }
 434
 435 /**
 436  * batadv_tp_sender_shutdown - let sender thread/timer stop gracefully
 437  * @tp_vars: the private data of the current TP meter session
 438  * @reason: reason for tp meter session stop
 439  */
 440 static void batadv_tp_sender_shutdown(struct batadv_tp_vars *tp_vars,
 441                                       enum batadv_tp_meter_reason reason)
 442 {
 443         if (!atomic_dec_and_test(&tp_vars->sending))
 444                 return;
 445
 446         tp_vars->reason = reason;
 447 }
 448
 449 /**
 450  * batadv_tp_sender_finish - stop sender session after test_length was reached
 451  * @work: delayed work reference of the related tp_vars
 452  */
 453 static void batadv_tp_sender_finish(struct work_struct *work)
 454 {
 455         struct delayed_work *delayed_work;
 456         struct batadv_tp_vars *tp_vars;
 457
 458         delayed_work = to_delayed_work(work);
 459         tp_vars = container_of(delayed_work, struct batadv_tp_vars,
 460                                finish_work);
 461
 462         batadv_tp_sender_shutdown(tp_vars, BATADV_TP_REASON_COMPLETE);
 463 }
 464
 465 /**
 466  * batadv_tp_reset_sender_timer - reschedule the sender timer
 467  * @tp_vars: the private TP meter data for this session
 468  *
 469  * Reschedule the timer using tp_vars->rto as delay
 470  */
 471 static void batadv_tp_reset_sender_timer(struct batadv_tp_vars *tp_vars)
 472 {
 473         /* most of the time this function is invoked while normal packet
 474          * reception...
 475          */
 476         if (unlikely(atomic_read(&tp_vars->sending) == 0))
 477                 /* timer ref will be dropped in batadv_tp_sender_cleanup */
 478                 return;
 479
 480         mod_timer(&tp_vars->timer, jiffies + msecs_to_jiffies(tp_vars->rto));
 481 }
 482
 483 /**
 484  * batadv_tp_sender_timeout - timer that fires in case of packet loss
 485  * @arg: address of the related tp_vars
 486  *
 487  * If fired it means that there was packet loss.
 488  * Switch to Slow Start, set the ss_threshold to half of the current cwnd and
 489  * reset the cwnd to 3*MSS
 490  */
 491 static void batadv_tp_sender_timeout(unsigned long arg)
 492 {
 493         struct batadv_tp_vars *tp_vars = (struct batadv_tp_vars *)arg;
 494         struct batadv_priv *bat_priv = tp_vars->bat_priv;
 495
 496         if (atomic_read(&tp_vars->sending) == 0)
 497                 return;
 498
 499         /* if the user waited long enough...shutdown the test */
 500         if (unlikely(tp_vars->rto >= BATADV_TP_MAX_RTO)) {
 501                 batadv_tp_sender_shutdown(tp_vars,
 502                                           BATADV_TP_REASON_DST_UNREACHABLE);
 503                 return;
 504         }
 505
 506         /* RTO exponential backoff
 507          * Details in Section 5.5 of RFC6298
 508          */
 509         tp_vars->rto <<= 1;
 510
 511         spin_lock_bh(&tp_vars->cwnd_lock);
 512
 513         tp_vars->ss_threshold = tp_vars->cwnd >> 1;
 514         if (tp_vars->ss_threshold < BATADV_TP_PLEN * 2)
 515                 tp_vars->ss_threshold = BATADV_TP_PLEN * 2;
 516
 517         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
 518                    "Meter: RTO fired during test towards %pM! cwnd=%u new ss_thr=%u, resetting last_sent to %u\n",
 519                    tp_vars->other_end, tp_vars->cwnd, tp_vars->ss_threshold,
 520                    atomic_read(&tp_vars->last_acked));
 521
 522         tp_vars->cwnd = BATADV_TP_PLEN * 3;
 523
 524         spin_unlock_bh(&tp_vars->cwnd_lock);
 525
 526         /* resend the non-ACKed packets.. */
 527         tp_vars->last_sent = atomic_read(&tp_vars->last_acked);
 528         wake_up(&tp_vars->more_bytes);
 529
 530         batadv_tp_reset_sender_timer(tp_vars);
 531 }
 532
 533 /**
 534  * batadv_tp_fill_prerandom - Fill buffer with prefetched random bytes
 535  * @tp_vars: the private TP meter data for this session
 536  * @buf: Buffer to fill with bytes
 537  * @nbytes: amount of pseudorandom bytes
 538  */
 539 static void batadv_tp_fill_prerandom(struct batadv_tp_vars *tp_vars,
 540                                      u8 *buf, size_t nbytes)
 541 {
 542         u32 local_offset;
 543         size_t bytes_inbuf;
 544         size_t to_copy;
 545         size_t pos = 0;
 546
 547         spin_lock_bh(&tp_vars->prerandom_lock);
 548         local_offset = tp_vars->prerandom_offset;
 549         tp_vars->prerandom_offset += nbytes;
 550         tp_vars->prerandom_offset %= sizeof(batadv_tp_prerandom);
 551         spin_unlock_bh(&tp_vars->prerandom_lock);
 552
 553         while (nbytes) {
 554                 local_offset %= sizeof(batadv_tp_prerandom);
 555                 bytes_inbuf = sizeof(batadv_tp_prerandom) - local_offset;
 556                 to_copy = min(nbytes, bytes_inbuf);
 557
 558                 memcpy(&buf[pos], &batadv_tp_prerandom[local_offset], to_copy);
 559                 pos += to_copy;
 560                 nbytes -= to_copy;
 561                 local_offset = 0;
 562         }
 563 }
 564
 565 /**
 566  * batadv_tp_send_msg - send a single message
 567  * @tp_vars: the private TP meter data for this session
 568  * @src: source mac address
 569  * @orig_node: the originator of the destination
 570  * @seqno: sequence number of this packet
 571  * @len: length of the entire packet
 572  * @session: session identifier
 573  * @uid: local ICMP "socket" index
 574  * @timestamp: timestamp in jiffies which is replied in ack
 575  *
 576  * Create and send a single TP Meter message.
 577  *
 578  * Return: 0 on success, BATADV_TP_REASON_DST_UNREACHABLE if the destination is
 579  * not reachable, BATADV_TP_REASON_MEMORY_ERROR if the packet couldn't be
 580  * allocated
 581  */
 582 static int batadv_tp_send_msg(struct batadv_tp_vars *tp_vars, const u8 *src,
 583                               struct batadv_orig_node *orig_node,
 584                               u32 seqno, size_t len, const u8 *session,
 585                               int uid, u32 timestamp)
 586 {
 587         struct batadv_icmp_tp_packet *icmp;
 588         struct sk_buff *skb;
 589         int r;
 590         u8 *data;
 591         size_t data_len;
 592
 593         skb = netdev_alloc_skb_ip_align(NULL, len + ETH_HLEN);
 594         if (unlikely(!skb))
 595                 return BATADV_TP_REASON_MEMORY_ERROR;
 596
 597         skb_reserve(skb, ETH_HLEN);
 598         icmp = skb_put(skb, sizeof(*icmp));
 599
 600         /* fill the icmp header */
 601         ether_addr_copy(icmp->dst, orig_node->orig);
 602         ether_addr_copy(icmp->orig, src);
 603         icmp->version = BATADV_COMPAT_VERSION;
 604         icmp->packet_type = BATADV_ICMP;
 605         icmp->ttl = BATADV_TTL;
 606         icmp->msg_type = BATADV_TP;
 607         icmp->uid = uid;
 608
 609         icmp->subtype = BATADV_TP_MSG;
 610         memcpy(icmp->session, session, sizeof(icmp->session));
 611         icmp->seqno = htonl(seqno);
 612         icmp->timestamp = htonl(timestamp);
 613
 614         data_len = len - sizeof(*icmp);
 615         data = skb_put(skb, data_len);
 616         batadv_tp_fill_prerandom(tp_vars, data, data_len);
 617
 618         r = batadv_send_skb_to_orig(skb, orig_node, NULL);
 619         if (r == NET_XMIT_SUCCESS)
 620                 return 0;
 621
 622         return BATADV_TP_REASON_CANT_SEND;
 623 }
 624
 625 /**
 626  * batadv_tp_recv_ack - ACK receiving function
 627  * @bat_priv: the bat priv with all the soft interface information
 628  * @skb: the buffer containing the received packet
 629  *
 630  * Process a received TP ACK packet
 631  */
 632 static void batadv_tp_recv_ack(struct batadv_priv *bat_priv,
 633                                const struct sk_buff *skb)
 634 {
 635         struct batadv_hard_iface *primary_if = NULL;
 636         struct batadv_orig_node *orig_node = NULL;
 637         const struct batadv_icmp_tp_packet *icmp;
 638         struct batadv_tp_vars *tp_vars;
 639         size_t packet_len, mss;
 640         u32 rtt, recv_ack, cwnd;
 641         unsigned char *dev_addr;
 642
 643         packet_len = BATADV_TP_PLEN;
 644         mss = BATADV_TP_PLEN;
 645         packet_len += sizeof(struct batadv_unicast_packet);
 646
 647         icmp = (struct batadv_icmp_tp_packet *)skb->data;
 648
 649         /* find the tp_vars */
 650         tp_vars = batadv_tp_list_find_session(bat_priv, icmp->orig,
 651                                               icmp->session);
 652         if (unlikely(!tp_vars))
 653                 return;
 654
 655         if (unlikely(atomic_read(&tp_vars->sending) == 0))
 656                 goto out;
 657
 658         /* old ACK? silently drop it.. */
 659         if (batadv_seq_before(ntohl(icmp->seqno),
 660                               (u32)atomic_read(&tp_vars->last_acked)))
 661                 goto out;
 662
 663         primary_if = batadv_primary_if_get_selected(bat_priv);
 664         if (unlikely(!primary_if))
 665                 goto out;
 666
 667         orig_node = batadv_orig_hash_find(bat_priv, icmp->orig);
 668         if (unlikely(!orig_node))
 669                 goto out;
 670
 671         /* update RTO with the new sampled RTT, if any */
 672         rtt = jiffies_to_msecs(jiffies) - ntohl(icmp->timestamp);
 673         if (icmp->timestamp && rtt)
 674                 batadv_tp_update_rto(tp_vars, rtt);
 675
 676         /* ACK for new data... reset the timer */
 677         batadv_tp_reset_sender_timer(tp_vars);
 678
 679         recv_ack = ntohl(icmp->seqno);
 680
 681         /* check if this ACK is a duplicate */
 682         if (atomic_read(&tp_vars->last_acked) == recv_ack) {
 683                 atomic_inc(&tp_vars->dup_acks);
 684                 if (atomic_read(&tp_vars->dup_acks) != 3)
 685                         goto out;
 686
 687                 if (recv_ack >= tp_vars->recover)
 688                         goto out;
 689
 690                 /* if this is the third duplicate ACK do Fast Retransmit */
 691                 batadv_tp_send_msg(tp_vars, primary_if->net_dev->dev_addr,
 692                                    orig_node, recv_ack, packet_len,
 693                                    icmp->session, icmp->uid,
 694                                    jiffies_to_msecs(jiffies));
 695
 696                 spin_lock_bh(&tp_vars->cwnd_lock);
 697
 698                 /* Fast Recovery */
 699                 tp_vars->fast_recovery = true;
 700                 /* Set recover to the last outstanding seqno when Fast Recovery
 701                  * is entered. RFC6582, Section 3.2, step 1
 702                  */
 703                 tp_vars->recover = tp_vars->last_sent;
 704                 tp_vars->ss_threshold = tp_vars->cwnd >> 1;
 705                 batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
 706                            "Meter: Fast Recovery, (cur cwnd=%u) ss_thr=%u last_sent=%u recv_ack=%u\n",
 707                            tp_vars->cwnd, tp_vars->ss_threshold,
 708                            tp_vars->last_sent, recv_ack);
 709                 tp_vars->cwnd = batadv_tp_cwnd(tp_vars->ss_threshold, 3 * mss,
 710                                                mss);
 711                 tp_vars->dec_cwnd = 0;
 712                 tp_vars->last_sent = recv_ack;
 713
 714                 spin_unlock_bh(&tp_vars->cwnd_lock);
 715         } else {
 716                 /* count the acked data */
 717                 atomic64_add(recv_ack - atomic_read(&tp_vars->last_acked),
 718                              &tp_vars->tot_sent);
 719                 /* reset the duplicate ACKs counter */
 720                 atomic_set(&tp_vars->dup_acks, 0);
 721
 722                 if (tp_vars->fast_recovery) {
 723                         /* partial ACK */
 724                         if (batadv_seq_before(recv_ack, tp_vars->recover)) {
 725                                 /* this is another hole in the window. React
 726                                  * immediately as specified by NewReno (see
 727                                  * Section 3.2 of RFC6582 for details)
 728                                  */
 729                                 dev_addr = primary_if->net_dev->dev_addr;
 730                                 batadv_tp_send_msg(tp_vars, dev_addr,
 731                                                    orig_node, recv_ack,
 732                                                    packet_len, icmp->session,
 733                                                    icmp->uid,
 734                                                    jiffies_to_msecs(jiffies));
 735                                 tp_vars->cwnd = batadv_tp_cwnd(tp_vars->cwnd,
 736                                                                mss, mss);
 737                         } else {
 738                                 tp_vars->fast_recovery = false;
 739                                 /* set cwnd to the value of ss_threshold at the
 740                                  * moment that Fast Recovery was entered.
 741                                  * RFC6582, Section 3.2, step 3
 742                                  */
 743                                 cwnd = batadv_tp_cwnd(tp_vars->ss_threshold, 0,
 744                                                       mss);
 745                                 tp_vars->cwnd = cwnd;
 746                         }
 747                         goto move_twnd;
 748                 }
 749
 750                 if (recv_ack - atomic_read(&tp_vars->last_acked) >= mss)
 751                         batadv_tp_update_cwnd(tp_vars, mss);
 752 move_twnd:
 753                 /* move the Transmit Window */
 754                 atomic_set(&tp_vars->last_acked, recv_ack);
 755         }
 756
 757         wake_up(&tp_vars->more_bytes);
 758 out:
 759         if (likely(primary_if))
 760                 batadv_hardif_put(primary_if);
 761         if (likely(orig_node))
 762                 batadv_orig_node_put(orig_node);
 763         if (likely(tp_vars))
 764                 batadv_tp_vars_put(tp_vars);
 765 }
 766
 767 /**
 768  * batadv_tp_avail - check if congestion window is not full
 769  * @tp_vars: the private data of the current TP meter session
 770  * @payload_len: size of the payload of a single message
 771  *
 772  * Return: true when congestion window is not full, false otherwise
 773  */
 774 static bool batadv_tp_avail(struct batadv_tp_vars *tp_vars,
 775                             size_t payload_len)
 776 {
 777         u32 win_left, win_limit;
 778
 779         win_limit = atomic_read(&tp_vars->last_acked) + tp_vars->cwnd;
 780         win_left = win_limit - tp_vars->last_sent;
 781
 782         return win_left >= payload_len;
 783 }
 784
 785 /**
 786  * batadv_tp_wait_available - wait until congestion window becomes free or
 787  *  timeout is reached
 788  * @tp_vars: the private data of the current TP meter session
 789  * @plen: size of the payload of a single message
 790  *
 791  * Return: 0 if the condition evaluated to false after the timeout elapsed,
 792  *  1 if the condition evaluated to true after the timeout elapsed, the
 793  *  remaining jiffies (at least 1) if the condition evaluated to true before
 794  *  the timeout elapsed, or -ERESTARTSYS if it was interrupted by a signal.
 795  */
 796 static int batadv_tp_wait_available(struct batadv_tp_vars *tp_vars, size_t plen)
 797 {
 798         int ret;
 799
 800         ret = wait_event_interruptible_timeout(tp_vars->more_bytes,
 801                                                batadv_tp_avail(tp_vars, plen),
 802                                                HZ / 10);
 803
 804         return ret;
 805 }
 806
 807 /**
 808  * batadv_tp_send - main sending thread of a tp meter session
 809  * @arg: address of the related tp_vars
 810  *
 811  * Return: nothing, this function never returns
 812  */
 813 static int batadv_tp_send(void *arg)
 814 {
 815         struct batadv_tp_vars *tp_vars = arg;
 816         struct batadv_priv *bat_priv = tp_vars->bat_priv;
 817         struct batadv_hard_iface *primary_if = NULL;
 818         struct batadv_orig_node *orig_node = NULL;
 819         size_t payload_len, packet_len;
 820         int err = 0;
 821
 822         if (unlikely(tp_vars->role != BATADV_TP_SENDER)) {
 823                 err = BATADV_TP_REASON_DST_UNREACHABLE;
 824                 tp_vars->reason = err;
 825                 goto out;
 826         }
 827
 828         orig_node = batadv_orig_hash_find(bat_priv, tp_vars->other_end);
 829         if (unlikely(!orig_node)) {
 830                 err = BATADV_TP_REASON_DST_UNREACHABLE;
 831                 tp_vars->reason = err;
 832                 goto out;
 833         }
 834
 835         primary_if = batadv_primary_if_get_selected(bat_priv);
 836         if (unlikely(!primary_if)) {
 837                 err = BATADV_TP_REASON_DST_UNREACHABLE;
 838                 tp_vars->reason = err;
 839                 goto out;
 840         }
 841
 842         /* assume that all the hard_interfaces have a correctly
 843          * configured MTU, so use the soft_iface MTU as MSS.
 844          * This might not be true and in that case the fragmentation
 845          * should be used.
 846          * Now, try to send the packet as it is
 847          */
 848         payload_len = BATADV_TP_PLEN;
 849         BUILD_BUG_ON(sizeof(struct batadv_icmp_tp_packet) > BATADV_TP_PLEN);
 850
 851         batadv_tp_reset_sender_timer(tp_vars);
 852
 853         /* queue the worker in charge of terminating the test */
 854         queue_delayed_work(batadv_event_workqueue, &tp_vars->finish_work,
 855                            msecs_to_jiffies(tp_vars->test_length));
 856
 857         while (atomic_read(&tp_vars->sending) != 0) {
 858                 if (unlikely(!batadv_tp_avail(tp_vars, payload_len))) {
 859                         batadv_tp_wait_available(tp_vars, payload_len);
 860                         continue;
 861                 }
 862
 863                 /* to emulate normal unicast traffic, add to the payload len
 864                  * the size of the unicast header
 865                  */
 866                 packet_len = payload_len + sizeof(struct batadv_unicast_packet);
 867
 868                 err = batadv_tp_send_msg(tp_vars, primary_if->net_dev->dev_addr,
 869                                          orig_node, tp_vars->last_sent,
 870                                          packet_len,
 871                                          tp_vars->session, tp_vars->icmp_uid,
 872                                          jiffies_to_msecs(jiffies));
 873
 874                 /* something went wrong during the preparation/transmission */
 875                 if (unlikely(err && err != BATADV_TP_REASON_CANT_SEND)) {
 876                         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
 877                                    "Meter: %s() cannot send packets (%d)\n",
 878                                    __func__, err);
 879                         /* ensure nobody else tries to stop the thread now */
 880                         if (atomic_dec_and_test(&tp_vars->sending))
 881                                 tp_vars->reason = err;
 882                         break;
 883                 }
 884
 885                 /* right-shift the TWND */
 886                 if (!err)
 887                         tp_vars->last_sent += payload_len;
 888
 889                 cond_resched();
 890         }
 891
 892 out:
 893         if (likely(primary_if))
 894                 batadv_hardif_put(primary_if);
 895         if (likely(orig_node))
 896                 batadv_orig_node_put(orig_node);
 897
 898         batadv_tp_sender_end(bat_priv, tp_vars);
 899         batadv_tp_sender_cleanup(bat_priv, tp_vars);
 900
 901         batadv_tp_vars_put(tp_vars);
 902
 903         do_exit(0);
 904 }
 905
 906 /**
 907  * batadv_tp_start_kthread - start new thread which manages the tp meter sender
 908  * @tp_vars: the private data of the current TP meter session
 909  */
 910 static void batadv_tp_start_kthread(struct batadv_tp_vars *tp_vars)
 911 {
 912         struct task_struct *kthread;
 913         struct batadv_priv *bat_priv = tp_vars->bat_priv;
 914         u32 session_cookie;
 915
 916         kref_get(&tp_vars->refcount);
 917         kthread = kthread_create(batadv_tp_send, tp_vars, "kbatadv_tp_meter");
 918         if (IS_ERR(kthread)) {
 919                 session_cookie = batadv_tp_session_cookie(tp_vars->session,
 920                                                           tp_vars->icmp_uid);
 921                 pr_err("batadv: cannot create tp meter kthread\n");
 922                 batadv_tp_batctl_error_notify(BATADV_TP_REASON_MEMORY_ERROR,
 923                                               tp_vars->other_end,
 924                                               bat_priv, session_cookie);
 925
 926                 /* drop reserved reference for kthread */
 927                 batadv_tp_vars_put(tp_vars);
 928
 929                 /* cleanup of failed tp meter variables */
 930                 batadv_tp_sender_cleanup(bat_priv, tp_vars);
 931                 return;
 932         }
 933
 934         wake_up_process(kthread);
 935 }
 936
 937 /**
 938  * batadv_tp_start - start a new tp meter session
 939  * @bat_priv: the bat priv with all the soft interface information
 940  * @dst: the receiver MAC address
 941  * @test_length: test length in milliseconds
 942  * @cookie: session cookie
 943  */
 944 void batadv_tp_start(struct batadv_priv *bat_priv, const u8 *dst,
 945                      u32 test_length, u32 *cookie)
 946 {
 947         struct batadv_tp_vars *tp_vars;
 948         u8 session_id[2];
 949         u8 icmp_uid;
 950         u32 session_cookie;
 951
 952         get_random_bytes(session_id, sizeof(session_id));
 953         get_random_bytes(&icmp_uid, 1);
 954         session_cookie = batadv_tp_session_cookie(session_id, icmp_uid);
 955         *cookie = session_cookie;
 956
 957         /* look for an already existing test towards this node */
 958         spin_lock_bh(&bat_priv->tp_list_lock);
 959         tp_vars = batadv_tp_list_find(bat_priv, dst);
 960         if (tp_vars) {
 961                 spin_unlock_bh(&bat_priv->tp_list_lock);
 962                 batadv_tp_vars_put(tp_vars);
 963                 batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
 964                            "Meter: test to or from the same node already ongoing, aborting\n");
 965                 batadv_tp_batctl_error_notify(BATADV_TP_REASON_ALREADY_ONGOING,
 966                                               dst, bat_priv, session_cookie);
 967                 return;
 968         }
 969
 970         if (!atomic_add_unless(&bat_priv->tp_num, 1, BATADV_TP_MAX_NUM)) {
 971                 spin_unlock_bh(&bat_priv->tp_list_lock);
 972                 batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
 973                            "Meter: too many ongoing sessions, aborting (SEND)\n");
 974                 batadv_tp_batctl_error_notify(BATADV_TP_REASON_TOO_MANY, dst,
 975                                               bat_priv, session_cookie);
 976                 return;
 977         }
 978
 979         tp_vars = kmalloc(sizeof(*tp_vars), GFP_ATOMIC);
 980         if (!tp_vars) {
 981                 spin_unlock_bh(&bat_priv->tp_list_lock);
 982                 batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
 983                            "Meter: %s cannot allocate list elements\n",
 984                            __func__);
 985                 batadv_tp_batctl_error_notify(BATADV_TP_REASON_MEMORY_ERROR,
 986                                               dst, bat_priv, session_cookie);
 987                 return;
 988         }
 989
 990         /* initialize tp_vars */
 991         ether_addr_copy(tp_vars->other_end, dst);
 992         kref_init(&tp_vars->refcount);
 993         tp_vars->role = BATADV_TP_SENDER;
 994         atomic_set(&tp_vars->sending, 1);
 995         memcpy(tp_vars->session, session_id, sizeof(session_id));
 996         tp_vars->icmp_uid = icmp_uid;
 997
 998         tp_vars->last_sent = BATADV_TP_FIRST_SEQ;
 999         atomic_set(&tp_vars->last_acked, BATADV_TP_FIRST_SEQ);
1000         tp_vars->fast_recovery = false;
1001         tp_vars->recover = BATADV_TP_FIRST_SEQ;
1002
1003         /* initialise the CWND to 3*MSS (Section 3.1 in RFC5681).
1004          * For batman-adv the MSS is the size of the payload received by the
1005          * soft_interface, hence its MTU
1006          */
1007         tp_vars->cwnd = BATADV_TP_PLEN * 3;
1008         /* at the beginning initialise the SS threshold to the biggest possible
1009          * window size, hence the AWND size
1010          */
1011         tp_vars->ss_threshold = BATADV_TP_AWND;
1012
1013         /* RTO initial value is 3 seconds.
1014          * Details in Section 2.1 of RFC6298
1015          */
1016         tp_vars->rto = 1000;
1017         tp_vars->srtt = 0;
1018         tp_vars->rttvar = 0;
1019
1020         atomic64_set(&tp_vars->tot_sent, 0);
1021
1022         kref_get(&tp_vars->refcount);
1023         setup_timer(&tp_vars->timer, batadv_tp_sender_timeout,
1024                     (unsigned long)tp_vars);
1025
1026         tp_vars->bat_priv = bat_priv;
1027         tp_vars->start_time = jiffies;
1028
1029         init_waitqueue_head(&tp_vars->more_bytes);
1030
1031         spin_lock_init(&tp_vars->unacked_lock);
1032         INIT_LIST_HEAD(&tp_vars->unacked_list);
1033
1034         spin_lock_init(&tp_vars->cwnd_lock);
1035
1036         tp_vars->prerandom_offset = 0;
1037         spin_lock_init(&tp_vars->prerandom_lock);
1038
1039         kref_get(&tp_vars->refcount);
1040         hlist_add_head_rcu(&tp_vars->list, &bat_priv->tp_list);
1041         spin_unlock_bh(&bat_priv->tp_list_lock);
1042
1043         tp_vars->test_length = test_length;
1044         if (!tp_vars->test_length)
1045                 tp_vars->test_length = BATADV_TP_DEF_TEST_LENGTH;
1046
1047         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
1048                    "Meter: starting throughput meter towards %pM (length=%ums)\n",
1049                    dst, test_length);
1050
1051         /* init work item for finished tp tests */
1052         INIT_DELAYED_WORK(&tp_vars->finish_work, batadv_tp_sender_finish);
1053
1054         /* start tp kthread. This way the write() call issued from userspace can
1055          * happily return and avoid to block
1056          */
1057         batadv_tp_start_kthread(tp_vars);
1058
1059         /* don't return reference to new tp_vars */
1060         batadv_tp_vars_put(tp_vars);
1061 }
1062
1063 /**
1064  * batadv_tp_stop - stop currently running tp meter session
1065  * @bat_priv: the bat priv with all the soft interface information
1066  * @dst: the receiver MAC address
1067  * @return_value: reason for tp meter session stop
1068  */
1069 void batadv_tp_stop(struct batadv_priv *bat_priv, const u8 *dst,
1070                     u8 return_value)
1071 {
1072         struct batadv_orig_node *orig_node;
1073         struct batadv_tp_vars *tp_vars;
1074
1075         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
1076                    "Meter: stopping test towards %pM\n", dst);
1077
1078         orig_node = batadv_orig_hash_find(bat_priv, dst);
1079         if (!orig_node)
1080                 return;
1081
1082         tp_vars = batadv_tp_list_find(bat_priv, orig_node->orig);
1083         if (!tp_vars) {
1084                 batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
1085                            "Meter: trying to interrupt an already over connection\n");
1086                 goto out;
1087         }
1088
1089         batadv_tp_sender_shutdown(tp_vars, return_value);
1090         batadv_tp_vars_put(tp_vars);
1091 out:
1092         batadv_orig_node_put(orig_node);
1093 }
1094
1095 /**
1096  * batadv_tp_reset_receiver_timer - reset the receiver shutdown timer
1097  * @tp_vars: the private data of the current TP meter session
1098  *
1099  * start the receiver shutdown timer or reset it if already started
1100  */
1101 static void batadv_tp_reset_receiver_timer(struct batadv_tp_vars *tp_vars)
1102 {
1103         mod_timer(&tp_vars->timer,
1104                   jiffies + msecs_to_jiffies(BATADV_TP_RECV_TIMEOUT));
1105 }
1106
1107 /**
1108  * batadv_tp_receiver_shutdown - stop a tp meter receiver when timeout is
1109  *  reached without received ack
1110  * @arg: address of the related tp_vars
1111  */
1112 static void batadv_tp_receiver_shutdown(unsigned long arg)
1113 {
1114         struct batadv_tp_vars *tp_vars = (struct batadv_tp_vars *)arg;
1115         struct batadv_tp_unacked *un, *safe;
1116         struct batadv_priv *bat_priv;
1117
1118         bat_priv = tp_vars->bat_priv;
1119
1120         /* if there is recent activity rearm the timer */
1121         if (!batadv_has_timed_out(tp_vars->last_recv_time,
1122                                   BATADV_TP_RECV_TIMEOUT)) {
1123                 /* reset the receiver shutdown timer */
1124                 batadv_tp_reset_receiver_timer(tp_vars);
1125                 return;
1126         }
1127
1128         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
1129                    "Shutting down for inactivity (more than %dms) from %pM\n",
1130                    BATADV_TP_RECV_TIMEOUT, tp_vars->other_end);
1131
1132         spin_lock_bh(&tp_vars->bat_priv->tp_list_lock);
1133         hlist_del_rcu(&tp_vars->list);
1134         spin_unlock_bh(&tp_vars->bat_priv->tp_list_lock);
1135
1136         /* drop list reference */
1137         batadv_tp_vars_put(tp_vars);
1138
1139         atomic_dec(&bat_priv->tp_num);
1140
1141         spin_lock_bh(&tp_vars->unacked_lock);
1142         list_for_each_entry_safe(un, safe, &tp_vars->unacked_list, list) {
1143                 list_del(&un->list);
1144                 kfree(un);
1145         }
1146         spin_unlock_bh(&tp_vars->unacked_lock);
1147
1148         /* drop reference of timer */
1149         batadv_tp_vars_put(tp_vars);
1150 }
1151
1152 /**
1153  * batadv_tp_send_ack - send an ACK packet
1154  * @bat_priv: the bat priv with all the soft interface information
1155  * @dst: the mac address of the destination originator
1156  * @seq: the sequence number to ACK
1157  * @timestamp: the timestamp to echo back in the ACK
1158  * @session: session identifier
1159  * @socket_index: local ICMP socket identifier
1160  *
1161  * Return: 0 on success, a positive integer representing the reason of the
1162  * failure otherwise
1163  */
1164 static int batadv_tp_send_ack(struct batadv_priv *bat_priv, const u8 *dst,
1165                               u32 seq, __be32 timestamp, const u8 *session,
1166                               int socket_index)
1167 {
1168         struct batadv_hard_iface *primary_if = NULL;
1169         struct batadv_orig_node *orig_node;
1170         struct batadv_icmp_tp_packet *icmp;
1171         struct sk_buff *skb;
1172         int r, ret;
1173
1174         orig_node = batadv_orig_hash_find(bat_priv, dst);
1175         if (unlikely(!orig_node)) {
1176                 ret = BATADV_TP_REASON_DST_UNREACHABLE;
1177                 goto out;
1178         }
1179
1180         primary_if = batadv_primary_if_get_selected(bat_priv);
1181         if (unlikely(!primary_if)) {
1182                 ret = BATADV_TP_REASON_DST_UNREACHABLE;
1183                 goto out;
1184         }
1185
1186         skb = netdev_alloc_skb_ip_align(NULL, sizeof(*icmp) + ETH_HLEN);
1187         if (unlikely(!skb)) {
1188                 ret = BATADV_TP_REASON_MEMORY_ERROR;
1189                 goto out;
1190         }
1191
1192         skb_reserve(skb, ETH_HLEN);
1193         icmp = skb_put(skb, sizeof(*icmp));
1194         icmp->packet_type = BATADV_ICMP;
1195         icmp->version = BATADV_COMPAT_VERSION;
1196         icmp->ttl = BATADV_TTL;
1197         icmp->msg_type = BATADV_TP;
1198         ether_addr_copy(icmp->dst, orig_node->orig);
1199         ether_addr_copy(icmp->orig, primary_if->net_dev->dev_addr);
1200         icmp->uid = socket_index;
1201
1202         icmp->subtype = BATADV_TP_ACK;
1203         memcpy(icmp->session, session, sizeof(icmp->session));
1204         icmp->seqno = htonl(seq);
1205         icmp->timestamp = timestamp;
1206
1207         /* send the ack */
1208         r = batadv_send_skb_to_orig(skb, orig_node, NULL);
1209         if (unlikely(r < 0) || r == NET_XMIT_DROP) {
1210                 ret = BATADV_TP_REASON_DST_UNREACHABLE;
1211                 goto out;
1212         }
1213         ret = 0;
1214
1215 out:
1216         if (likely(orig_node))
1217                 batadv_orig_node_put(orig_node);
1218         if (likely(primary_if))
1219                 batadv_hardif_put(primary_if);
1220
1221         return ret;
1222 }
1223
1224 /**
1225  * batadv_tp_handle_out_of_order - store an out of order packet
1226  * @tp_vars: the private data of the current TP meter session
1227  * @skb: the buffer containing the received packet
1228  *
1229  * Store the out of order packet in the unacked list for late processing. This
1230  * packets are kept in this list so that they can be ACKed at once as soon as
1231  * all the previous packets have been received
1232  *
1233  * Return: true if the packed has been successfully processed, false otherwise
1234  */
1235 static bool batadv_tp_handle_out_of_order(struct batadv_tp_vars *tp_vars,
1236                                           const struct sk_buff *skb)
1237 {
1238         const struct batadv_icmp_tp_packet *icmp;
1239         struct batadv_tp_unacked *un, *new;
1240         u32 payload_len;
1241         bool added = false;
1242
1243         new = kmalloc(sizeof(*new), GFP_ATOMIC);
1244         if (unlikely(!new))
1245                 return false;
1246
1247         icmp = (struct batadv_icmp_tp_packet *)skb->data;
1248
1249         new->seqno = ntohl(icmp->seqno);
1250         payload_len = skb->len - sizeof(struct batadv_unicast_packet);
1251         new->len = payload_len;
1252
1253         spin_lock_bh(&tp_vars->unacked_lock);
1254         /* if the list is empty immediately attach this new object */
1255         if (list_empty(&tp_vars->unacked_list)) {
1256                 list_add(&new->list, &tp_vars->unacked_list);
1257                 goto out;
1258         }
1259
1260         /* otherwise loop over the list and either drop the packet because this
1261          * is a duplicate or store it at the right position.
1262          *
1263          * The iteration is done in the reverse way because it is likely that
1264          * the last received packet (the one being processed now) has a bigger
1265          * seqno than all the others already stored.
1266          */
1267         list_for_each_entry_reverse(un, &tp_vars->unacked_list, list) {
1268                 /* check for duplicates */
1269                 if (new->seqno == un->seqno) {
1270                         if (new->len > un->len)
1271                                 un->len = new->len;
1272                         kfree(new);
1273                         added = true;
1274                         break;
1275                 }
1276
1277                 /* look for the right position */
1278                 if (batadv_seq_before(new->seqno, un->seqno))
1279                         continue;
1280
1281                 /* as soon as an entry having a bigger seqno is found, the new
1282                  * one is attached _after_ it. In this way the list is kept in
1283                  * ascending order
1284                  */
1285                 list_add_tail(&new->list, &un->list);
1286                 added = true;
1287                 break;
1288         }
1289
1290         /* received packet with smallest seqno out of order; add it to front */
1291         if (!added)
1292                 list_add(&new->list, &tp_vars->unacked_list);
1293
1294 out:
1295         spin_unlock_bh(&tp_vars->unacked_lock);
1296
1297         return true;
1298 }
1299
1300 /**
1301  * batadv_tp_ack_unordered - update number received bytes in current stream
1302  *  without gaps
1303  * @tp_vars: the private data of the current TP meter session
1304  */
1305 static void batadv_tp_ack_unordered(struct batadv_tp_vars *tp_vars)
1306 {
1307         struct batadv_tp_unacked *un, *safe;
1308         u32 to_ack;
1309
1310         /* go through the unacked packet list and possibly ACK them as
1311          * well
1312          */
1313         spin_lock_bh(&tp_vars->unacked_lock);
1314         list_for_each_entry_safe(un, safe, &tp_vars->unacked_list, list) {
1315                 /* the list is ordered, therefore it is possible to stop as soon
1316                  * there is a gap between the last acked seqno and the seqno of
1317                  * the packet under inspection
1318                  */
1319                 if (batadv_seq_before(tp_vars->last_recv, un->seqno))
1320                         break;
1321
1322                 to_ack = un->seqno + un->len - tp_vars->last_recv;
1323
1324                 if (batadv_seq_before(tp_vars->last_recv, un->seqno + un->len))
1325                         tp_vars->last_recv += to_ack;
1326
1327                 list_del(&un->list);
1328                 kfree(un);
1329         }
1330         spin_unlock_bh(&tp_vars->unacked_lock);
1331 }
1332
1333 /**
1334  * batadv_tp_init_recv - return matching or create new receiver tp_vars
1335  * @bat_priv: the bat priv with all the soft interface information
1336  * @icmp: received icmp tp msg
1337  *
1338  * Return: corresponding tp_vars or NULL on errors
1339  */
1340 static struct batadv_tp_vars *
1341 batadv_tp_init_recv(struct batadv_priv *bat_priv,
1342                     const struct batadv_icmp_tp_packet *icmp)
1343 {
1344         struct batadv_tp_vars *tp_vars;
1345
1346         spin_lock_bh(&bat_priv->tp_list_lock);
1347         tp_vars = batadv_tp_list_find_session(bat_priv, icmp->orig,
1348                                               icmp->session);
1349         if (tp_vars)
1350                 goto out_unlock;
1351
1352         if (!atomic_add_unless(&bat_priv->tp_num, 1, BATADV_TP_MAX_NUM)) {
1353                 batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
1354                            "Meter: too many ongoing sessions, aborting (RECV)\n");
1355                 goto out_unlock;
1356         }
1357
1358         tp_vars = kmalloc(sizeof(*tp_vars), GFP_ATOMIC);
1359         if (!tp_vars)
1360                 goto out_unlock;
1361
1362         ether_addr_copy(tp_vars->other_end, icmp->orig);
1363         tp_vars->role = BATADV_TP_RECEIVER;
1364         memcpy(tp_vars->session, icmp->session, sizeof(tp_vars->session));
1365         tp_vars->last_recv = BATADV_TP_FIRST_SEQ;
1366         tp_vars->bat_priv = bat_priv;
1367         kref_init(&tp_vars->refcount);
1368
1369         spin_lock_init(&tp_vars->unacked_lock);
1370         INIT_LIST_HEAD(&tp_vars->unacked_list);
1371
1372         kref_get(&tp_vars->refcount);
1373         hlist_add_head_rcu(&tp_vars->list, &bat_priv->tp_list);
1374
1375         kref_get(&tp_vars->refcount);
1376         setup_timer(&tp_vars->timer, batadv_tp_receiver_shutdown,
1377                     (unsigned long)tp_vars);
1378
1379         batadv_tp_reset_receiver_timer(tp_vars);
1380
1381 out_unlock:
1382         spin_unlock_bh(&bat_priv->tp_list_lock);
1383
1384         return tp_vars;
1385 }
1386
1387 /**
1388  * batadv_tp_recv_msg - process a single data message
1389  * @bat_priv: the bat priv with all the soft interface information
1390  * @skb: the buffer containing the received packet
1391  *
1392  * Process a received TP MSG packet
1393  */
1394 static void batadv_tp_recv_msg(struct batadv_priv *bat_priv,
1395                                const struct sk_buff *skb)
1396 {
1397         const struct batadv_icmp_tp_packet *icmp;
1398         struct batadv_tp_vars *tp_vars;
1399         size_t packet_size;
1400         u32 seqno;
1401
1402         icmp = (struct batadv_icmp_tp_packet *)skb->data;
1403
1404         seqno = ntohl(icmp->seqno);
1405         /* check if this is the first seqno. This means that if the
1406          * first packet is lost, the tp meter does not work anymore!
1407          */
1408         if (seqno == BATADV_TP_FIRST_SEQ) {
1409                 tp_vars = batadv_tp_init_recv(bat_priv, icmp);
1410                 if (!tp_vars) {
1411                         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
1412                                    "Meter: seqno != BATADV_TP_FIRST_SEQ cannot initiate connection\n");
1413                         goto out;
1414                 }
1415         } else {
1416                 tp_vars = batadv_tp_list_find_session(bat_priv, icmp->orig,
1417                                                       icmp->session);
1418                 if (!tp_vars) {
1419                         batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
1420                                    "Unexpected packet from %pM!\n",
1421                                    icmp->orig);
1422                         goto out;
1423                 }
1424         }
1425
1426         if (unlikely(tp_vars->role != BATADV_TP_RECEIVER)) {
1427                 batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
1428                            "Meter: dropping packet: not expected (role=%u)\n",
1429                            tp_vars->role);
1430                 goto out;
1431         }
1432
1433         tp_vars->last_recv_time = jiffies;
1434
1435         /* if the packet is a duplicate, it may be the case that an ACK has been
1436          * lost. Resend the ACK
1437          */
1438         if (batadv_seq_before(seqno, tp_vars->last_recv))
1439                 goto send_ack;
1440
1441         /* if the packet is out of order enqueue it */
1442         if (ntohl(icmp->seqno) != tp_vars->last_recv) {
1443                 /* exit immediately (and do not send any ACK) if the packet has
1444                  * not been enqueued correctly
1445                  */
1446                 if (!batadv_tp_handle_out_of_order(tp_vars, skb))
1447                         goto out;
1448
1449                 /* send a duplicate ACK */
1450                 goto send_ack;
1451         }
1452
1453         /* if everything was fine count the ACKed bytes */
1454         packet_size = skb->len - sizeof(struct batadv_unicast_packet);
1455         tp_vars->last_recv += packet_size;
1456
1457         /* check if this ordered message filled a gap.... */
1458         batadv_tp_ack_unordered(tp_vars);
1459
1460 send_ack:
1461         /* send the ACK. If the received packet was out of order, the ACK that
1462          * is going to be sent is a duplicate (the sender will count them and
1463          * possibly enter Fast Retransmit as soon as it has reached 3)
1464          */
1465         batadv_tp_send_ack(bat_priv, icmp->orig, tp_vars->last_recv,
1466                            icmp->timestamp, icmp->session, icmp->uid);
1467 out:
1468         if (likely(tp_vars))
1469                 batadv_tp_vars_put(tp_vars);
1470 }
1471
1472 /**
1473  * batadv_tp_meter_recv - main TP Meter receiving function
1474  * @bat_priv: the bat priv with all the soft interface information
1475  * @skb: the buffer containing the received packet
1476  */
1477 void batadv_tp_meter_recv(struct batadv_priv *bat_priv, struct sk_buff *skb)
1478 {
1479         struct batadv_icmp_tp_packet *icmp;
1480
1481         icmp = (struct batadv_icmp_tp_packet *)skb->data;
1482
1483         switch (icmp->subtype) {
1484         case BATADV_TP_MSG:
1485                 batadv_tp_recv_msg(bat_priv, skb);
1486                 break;
1487         case BATADV_TP_ACK:
1488                 batadv_tp_recv_ack(bat_priv, skb);
1489                 break;
1490         default:
1491                 batadv_dbg(BATADV_DBG_TP_METER, bat_priv,
1492                            "Received unknown TP Metric packet type %u\n",
1493                            icmp->subtype);
1494         }
1495         consume_skb(skb);
1496 }
1497
1498 /**
1499  * batadv_tp_meter_init - initialize global tp_meter structures
1500  */
1501 void __init batadv_tp_meter_init(void)
1502 {
1503         get_random_bytes(batadv_tp_prerandom, sizeof(batadv_tp_prerandom));
1504 }