3 * most of the swift's state machine
5 * Created by Victor Grishchenko on 3/6/09.
6 * Copyright 2009-2012 TECHNISCHE UNIVERSITEIT DELFT. All rights reserved.
11 #include <algorithm> // kill it
17 using namespace swift;
20 struct event_base *Channel::evbase;
21 struct event Channel::evrecv;
23 #define DEBUGTRAFFIC 0
25 /** Arno: Victor's design allows a sender to choose some data to push to
26 * a receiver, if that receiver is not HINTing at data. Should be disabled
27 * when the receiver has a download rate limit.
29 #define ENABLE_SENDERSIZE_PUSH 0
32 /** Arno, 2011-11-24: When rate limit is on and the download is in progress
33 * we send HINTs for 2 chunks at the moment. This constant can be used to
34 * get greater granularity. Set to 0 for original behaviour.
36 #define HINT_GRANULARITY 16 // chunks
41 - randomized testing of advanced ops (new testcase)
44 void Channel::AddPeakHashes (struct evbuffer *evb) {
45 for(int i=0; i<file().peak_count(); i++) {
46 bin_t peak = file().peak(i);
47 evbuffer_add_8(evb, SWIFT_HASH);
48 evbuffer_add_32be(evb, bin_toUInt32(peak));
49 evbuffer_add_hash(evb, file().peak_hash(i));
50 char bin_name_buf[32];
51 dprintf("%s #%u +phash %s\n",tintstr(),id_,peak.str(bin_name_buf));
56 void Channel::AddUncleHashes (struct evbuffer *evb, bin_t pos) {
58 char bin_name_buf2[32];
59 dprintf("%s #%u +uncle hash for %s\n",tintstr(),id_,pos.str(bin_name_buf2));
61 bin_t peak = file().peak_for(pos);
62 while (pos!=peak && ((NOW&3)==3 || !pos.parent().contains(data_out_cap_)) &&
63 ack_in_.is_empty(pos.parent()) ) {
64 bin_t uncle = pos.sibling();
65 evbuffer_add_8(evb, SWIFT_HASH);
66 evbuffer_add_32be(evb, bin_toUInt32(uncle));
67 evbuffer_add_hash(evb, file().hash(uncle) );
68 char bin_name_buf[32];
69 dprintf("%s #%u +hash %s\n",tintstr(),id_,uncle.str(bin_name_buf));
75 bin_t Channel::ImposeHint () {
76 uint64_t twist = peer_channel_id_; // got no hints, send something randomly
78 twist &= file().peak(0).toUInt(); // FIXME may make it semi-seq here
80 bin_t my_pick = binmap_t::find_complement(ack_in_, file().ack_out(), twist);
82 my_pick.to_twisted(twist);
83 while (my_pick.base_length()>max(1,(int)cwnd_))
84 my_pick = my_pick.left();
86 return my_pick.twisted(twist);
90 bin_t Channel::DequeueHint (bool *retransmitptr) {
91 bin_t send = bin_t::NONE;
93 // Arno, 2012-01-23: Extra protection against channel loss, don't send DATA
94 if (last_recv_time_ < NOW-(3*TINT_SEC))
97 // Arno, 2012-01-18: Reenable Victor's retransmit
98 if (!data_out_tmo_.empty())
100 tintbin tb = data_out_tmo_.front();
102 data_out_tmo_.pop_front();
103 *retransmitptr = true;
106 *retransmitptr = false;
108 if (ENABLE_SENDERSIZE_PUSH && send.is_none() && hint_in_.empty() && last_recv_time_>NOW-rtt_avg_-TINT_SEC) {
109 bin_t my_pick = ImposeHint(); // FIXME move to the loop
110 if (!my_pick.is_none()) {
111 hint_in_.push_back(my_pick);
112 char bin_name_buf[32];
113 dprintf("%s #%u *hint %s\n",tintstr(),id_,my_pick.str(bin_name_buf));
117 while (!hint_in_.empty() && send.is_none()) {
118 bin_t hint = hint_in_.front().bin;
119 tint time = hint_in_.front().time;
120 hint_in_.pop_front();
121 while (!hint.is_base()) { // FIXME optimize; possible attack
122 hint_in_.push_front(tintbin(time,hint.right()));
125 //if (time < NOW-TINT_SEC*3/2 )
126 // continue; bad idea
127 if (!ack_in_.is_filled(hint))
131 // Arno, 2012-03-09: Is mucho expensive on busy server.
132 //for(int i=0; i<hint_in_.size(); i++)
133 // mass += hint_in_[i].bin.base_length();
134 char bin_name_buf[32];
135 dprintf("%s #%u dequeued %s [%lli]\n",tintstr(),id_,send.str(bin_name_buf),mass);
140 void Channel::AddHandshake (struct evbuffer *evb) {
141 if (!peer_channel_id_) { // initiating
142 evbuffer_add_8(evb, SWIFT_HASH);
143 evbuffer_add_32be(evb, bin_toUInt32(bin_t::ALL));
144 evbuffer_add_hash(evb, file().root_hash());
145 dprintf("%s #%u +hash ALL %s\n",
146 tintstr(),id_,file().root_hash().hex().c_str());
148 evbuffer_add_8(evb, SWIFT_HANDSHAKE);
150 if (send_control_==CLOSE_CONTROL) {
154 encoded = EncodeID(id_);
155 evbuffer_add_32be(evb, encoded);
156 dprintf("%s #%u +hs %x\n",tintstr(),id_,encoded);
161 void Channel::Send () {
163 dprintf("%s #%u Send called \n",tintstr(),id_);
165 struct evbuffer *evb = evbuffer_new();
166 evbuffer_add_32be(evb, peer_channel_id_);
167 bin_t data = bin_t::NONE;
168 int evbnonadplen = 0;
169 if ( is_established() ) {
170 if (send_control_!=CLOSE_CONTROL) {
171 // FIXME: seeder check
174 if (!file().is_complete()) {
176 /* Gertjan fix: 7aeea65f3efbb9013f601b22a57ee4a423f1a94d
177 "Only call Reschedule for 'reverse PEX' if the channel is in keep-alive mode"
185 // Arno: send explicit close
190 AddHave(evb); // Arno, 2011-10-28: from AddHandShake. Why double?
195 lastsendwaskeepalive_ = (evbuffer_get_length(evb) == 4);
197 if (evbuffer_get_length(evb)==4) {// only the channel id; bare keep-alive
200 dprintf("%s #%u sent %ib %s:%x\n",
201 tintstr(),id_,(int)evbuffer_get_length(evb),peer().str(),
203 int r = SendTo(socket_,peer(),evb);
205 print_error("can't send datagram");
208 last_send_time_ = NOW;
215 void Channel::AddHint (struct evbuffer *evb) {
218 // Policy is to not send hints when we are above speed limit
219 if (transfer().GetCurrentSpeed(DDIR_DOWNLOAD) > transfer().GetMaxSpeed(DDIR_DOWNLOAD)) {
221 fprintf(stderr,"hint: forbidden#");
226 // 1. Calc max of what we are allowed to request, uncongested bandwidth wise
227 tint plan_for = max(TINT_SEC,rtt_avg_*4);
229 tint timed_out = NOW - plan_for*2;
230 while ( !hint_out_.empty() && hint_out_.front().time < timed_out ) {
231 hint_out_size_ -= hint_out_.front().bin.base_length();
232 hint_out_.pop_front();
235 int first_plan_pck = max ( (tint)1, plan_for / dip_avg_ );
237 // Riccardo, 2012-04-04: Actually allowed is max minus what we already asked for
238 int queue_allowed_hints = max(0,first_plan_pck-(int)hint_out_size_);
242 // 2. Calc max of what is allowed by the rate limiter
243 int rate_allowed_hints = LONG_MAX;
244 if (transfer().GetMaxSpeed(DDIR_DOWNLOAD) < DBL_MAX)
246 uint64_t rough_global_hint_out_size = 0; // rough estimate, as hint_out_ clean up is not done for all channels
247 std::set<Channel *>::iterator iter;
248 for (iter=transfer().mychannels_.begin(); iter!=transfer().mychannels_.end(); iter++)
252 rough_global_hint_out_size += c->hint_out_size_;
255 // Policy: this channel is allowed to hint at the limit - global_hinted_at
256 // Handle MaxSpeed = unlimited
257 double rate_hints_limit_float = transfer().GetMaxSpeed(DDIR_DOWNLOAD)/((double)file().chunk_size());
259 int rate_hints_limit = (int)min((double)LONG_MAX,rate_hints_limit_float);
261 // Actually allowed is max minus what we already asked for, globally (=all channels)
262 rate_allowed_hints = max(0,rate_hints_limit-(int)rough_global_hint_out_size);
265 // 3. Take the smallest allowance from rate and queue limit
266 uint64_t plan_pck = (uint64_t)min(rate_allowed_hints,queue_allowed_hints);
268 // 4. Ask allowance in blocks of chunks to get pipelining going from serving peer.
269 if (hint_out_size_ == 0 || plan_pck > HINT_GRANULARITY)
271 bin_t hint = transfer().picker().Pick(ack_in_,plan_pck,NOW+plan_for*2);
272 if (!hint.is_none()) {
276 fprintf(stderr,"hint c%d: ask %s\n", id(), hint.str(binstr) );
278 evbuffer_add_8(evb, SWIFT_HINT);
279 evbuffer_add_32be(evb, bin_toUInt32(hint));
280 char bin_name_buf[32];
281 dprintf("%s #%u +hint %s [%lli]\n",tintstr(),id_,hint.str(bin_name_buf),hint_out_size_);
282 hint_out_.push_back(hint);
283 hint_out_size_ += hint.base_length();
284 //fprintf(stderr,"send c%d: HINTLEN %i\n", id(), hint.base_length());
285 //fprintf(stderr,"HL %i ", hint.base_length());
288 dprintf("%s #%u Xhint\n",tintstr(),id_);
294 bin_t Channel::AddData (struct evbuffer *evb) {
296 if (transfer().GetCurrentSpeed(DDIR_UPLOAD) > transfer().GetMaxSpeed(DDIR_UPLOAD)) {
297 transfer().OnSendNoData();
301 if (!file().size()) // know nothing
304 bin_t tosend = bin_t::NONE;
305 bool isretransmit = false;
306 tint luft = send_interval_>>4; // may wake up a bit earlier
307 if (data_out_.size()<cwnd_ &&
308 last_data_out_time_+send_interval_<=NOW+luft) {
309 tosend = DequeueHint(&isretransmit);
310 if (tosend.is_none()) {
311 dprintf("%s #%u sendctrl no idea what data to send\n",tintstr(),id_);
312 if (send_control_!=KEEP_ALIVE_CONTROL && send_control_!=CLOSE_CONTROL)
313 SwitchSendControl(KEEP_ALIVE_CONTROL);
316 dprintf("%s #%u sendctrl wait cwnd %f data_out %i next %s\n",
317 tintstr(),id_,cwnd_,(int)data_out_.size(),tintstr(last_data_out_time_+NOW-send_interval_));
319 if (tosend.is_none())// && (last_data_out_time_>NOW-TINT_SEC || data_out_.empty()))
320 return bin_t::NONE; // once in a while, empty data is sent just to check rtt FIXED
322 if (ack_in_.is_empty() && file().size())
326 if (file().get_check_netwvshash())
327 AddUncleHashes(evb,tosend);
329 if (!ack_in_.is_empty()) // TODO: cwnd_>1
330 data_out_cap_ = tosend;
332 // Arno, 2011-11-03: May happen when first data packet is sent to empty
333 // leech, then peak + uncle hashes may be so big that they don't fit in eth
334 // frame with DATA. Send 2 datagrams then, one with peaks so they have
335 // a better chance of arriving. Optimistic violation of atomic datagram
337 if (file().chunk_size() == SWIFT_DEFAULT_CHUNK_SIZE && evbuffer_get_length(evb) > SWIFT_MAX_NONDATA_DGRAM_SIZE) {
338 dprintf("%s #%u fsent %ib %s:%x\n",
339 tintstr(),id_,(int)evbuffer_get_length(evb),peer().str(),
341 int ret = Channel::SendTo(socket_,peer(),evb); // kind of fragmentation
343 raw_bytes_up_ += ret;
344 evbuffer_add_32be(evb, peer_channel_id_);
347 if (file().chunk_size() != SWIFT_DEFAULT_CHUNK_SIZE && isretransmit) {
349 * Arno, 2012-01-17: We observe strange behaviour when using
350 * fragmented UDP packets. When ULANC sends a specific datagram ("995"),
351 * the 2nd IP packet carrying it gets lost structurally. When
352 * downloading from the same asset hosted on a Linux 32-bit machine
353 * using a Win7 32-bit client (behind a NAT), one specific full
354 * datagram never gets delivered (6970 one before do). A workaround
355 * is to add some random data to the datagram. Hence we introduce
356 * the SWIFT_RANDOMIZE message, that is added to the datagram carrying
357 * the DATA on a retransmit.
360 fprintf(stderr,"AddData: retransmit of randomized chunk %s\n",tosend.str(binstr) );
361 evbuffer_add_8(evb, SWIFT_RANDOMIZE);
362 evbuffer_add_32be(evb, (int)rand() );
365 evbuffer_add_8(evb, SWIFT_DATA);
366 evbuffer_add_32be(evb, bin_toUInt32(tosend));
368 struct evbuffer_iovec vec;
369 if (evbuffer_reserve_space(evb, file().chunk_size(), &vec, 1) < 0) {
370 print_error("error on evbuffer_reserve_space");
373 size_t r = pread(file().file_descriptor(),(char *)vec.iov_base,
374 file().chunk_size(),tosend.base_offset()*file().chunk_size());
375 // TODO: corrupted data, retries, caching
377 print_error("error on reading");
379 evbuffer_commit_space(evb, &vec, 1);
382 // assert(dgram.space()>=r+4+1);
384 if (evbuffer_commit_space(evb, &vec, 1) < 0) {
385 print_error("error on evbuffer_commit_space");
389 last_data_out_time_ = NOW;
390 data_out_.push_back(tosend);
392 global_bytes_up += r;
394 char bin_name_buf[32];
395 dprintf("%s #%u +data %s\n",tintstr(),id_,tosend.str(bin_name_buf));
398 // ARNOSMPTODO: count overhead bytes too? Move to Send() then.
399 transfer_->OnSendData(file().chunk_size());
405 void Channel::AddAck (struct evbuffer *evb) {
406 if (data_in_==tintbin())
407 //if (data_in_.bin==bin64_t::NONE)
409 // sometimes, we send a HAVE (e.g. in case the peer did repetitive send)
410 evbuffer_add_8(evb, data_in_.time==TINT_NEVER?SWIFT_HAVE:SWIFT_ACK);
411 evbuffer_add_32be(evb, bin_toUInt32(data_in_.bin));
412 if (data_in_.time!=TINT_NEVER)
413 evbuffer_add_64be(evb, data_in_.time);
417 fprintf(stderr,"send c%d: ACK %i\n", id(), bin_toUInt32(data_in_.bin));
419 have_out_.set(data_in_.bin);
420 char bin_name_buf[32];
421 dprintf("%s #%u +ack %s %s\n",
422 tintstr(),id_,data_in_.bin.str(bin_name_buf),tintstr(data_in_.time));
423 if (data_in_.bin.layer()>2)
424 data_in_dbl_ = data_in_.bin;
426 //fprintf(stderr,"data_in_ c%d\n", id() );
427 data_in_ = tintbin();
428 //data_in_ = tintbin(NOW,bin64_t::NONE);
432 void Channel::AddHave (struct evbuffer *evb) {
433 if (!data_in_dbl_.is_none()) { // TODO: do redundancy better
434 evbuffer_add_8(evb, SWIFT_HAVE);
435 evbuffer_add_32be(evb, bin_toUInt32(data_in_dbl_));
436 data_in_dbl_=bin_t::NONE;
439 fprintf(stderr,"send c%d: HAVE ",id() );
440 for(int count=0; count<4; count++) {
441 bin_t ack = binmap_t::find_complement(have_out_, file().ack_out(), 0); // FIXME: do rotating queue
444 ack = file().ack_out().cover(ack);
446 evbuffer_add_8(evb, SWIFT_HAVE);
447 evbuffer_add_32be(evb, bin_toUInt32(ack));
450 fprintf(stderr," %i", bin_toUInt32(ack));
452 char bin_name_buf[32];
453 dprintf("%s #%u +have %s\n",tintstr(),id_,ack.str(bin_name_buf));
456 fprintf(stderr,"\n");
461 void Channel::Recv (struct evbuffer *evb) {
462 dprintf("%s #%u recvd %ib\n",tintstr(),id_,(int)evbuffer_get_length(evb)+4);
465 lastrecvwaskeepalive_ = (evbuffer_get_length(evb) == 0);
466 if (lastrecvwaskeepalive_)
467 // Update speed measurements such that they decrease when DL stops
468 transfer().OnRecvData(0);
470 if (last_send_time_ && rtt_avg_==TINT_SEC && dev_avg_==0) {
471 rtt_avg_ = NOW - last_send_time_;
474 dprintf("%s #%u sendctrl rtt init %lli\n",tintstr(),id_,rtt_avg_);
477 bin_t data = evbuffer_get_length(evb) ? bin_t::NONE : bin_t::ALL;
480 fprintf(stderr,"recv c%d: size %d ", id(), evbuffer_get_length(evb));
482 while (evbuffer_get_length(evb)) {
483 uint8_t type = evbuffer_remove_8(evb);
486 fprintf(stderr," %d", type);
489 case SWIFT_HANDSHAKE: OnHandshake(evb); break;
490 case SWIFT_DATA: data=OnData(evb); break;
491 case SWIFT_HAVE: OnHave(evb); break;
492 case SWIFT_ACK: OnAck(evb); break;
493 case SWIFT_HASH: OnHash(evb); break;
494 case SWIFT_HINT: OnHint(evb); break;
495 case SWIFT_PEX_ADD: OnPex(evb); break;
496 case SWIFT_PEX_REQ: OnPexReq(); break;
497 case SWIFT_RANDOMIZE: OnRandomize(evb); break; //FRAGRAND
499 dprintf("%s #%u ?msg id unknown %i\n",tintstr(),id_,(int)type);
505 fprintf(stderr,"\n");
509 last_recv_time_ = NOW;
510 sent_since_recv_ = 0;
515 * Arno: FAXME: HASH+DATA should be handled as a transaction: only when the
516 * hashes check out should they be stored in the hashtree, otherwise revert.
518 void Channel::OnHash (struct evbuffer *evb) {
519 bin_t pos = bin_fromUInt32(evbuffer_remove_32be(evb));
520 Sha1Hash hash = evbuffer_remove_hash(evb);
521 file().OfferHash(pos,hash);
522 char bin_name_buf[32];
523 dprintf("%s #%u -hash %s\n",tintstr(),id_,pos.str(bin_name_buf));
525 //fprintf(stderr,"HASH %lli hex %s\n",pos.toUInt(), hash.hex().c_str() );
529 void Channel::CleanHintOut (bin_t pos) {
531 while (hi<hint_out_.size() && !hint_out_[hi].bin.contains(pos))
533 if (hi==hint_out_.size())
534 return; // something not hinted or hinted in far past
535 while (hi--) { // removing likely snubbed hints
536 hint_out_size_ -= hint_out_.front().bin.base_length();
537 hint_out_.pop_front();
539 while (hint_out_.front().bin!=pos) {
540 tintbin f = hint_out_.front();
542 assert (f.bin.contains(pos));
550 hint_out_.front().bin = f.bin.sibling();
551 hint_out_.push_front(f);
553 hint_out_.pop_front();
558 bin_t Channel::OnData (struct evbuffer *evb) { // TODO: HAVE NONE for corrupted data
560 char bin_name_buf[32];
561 bin_t pos = bin_fromUInt32(evbuffer_remove_32be(evb));
563 // Arno: Assuming DATA last message in datagram
564 if (evbuffer_get_length(evb) > file().chunk_size()) {
565 dprintf("%s #%u !data chunk size mismatch %s: exp %lu got " PRISIZET "\n",tintstr(),id_,pos.str(bin_name_buf), file().chunk_size(), evbuffer_get_length(evb));
566 fprintf(stderr,"WARNING: chunk size mismatch: exp %lu got " PRISIZET "\n",file().chunk_size(), evbuffer_get_length(evb));
569 int length = (evbuffer_get_length(evb) < file().chunk_size()) ? evbuffer_get_length(evb) : file().chunk_size();
570 if (!file().ack_out().is_empty(pos)) {
571 // Arno, 2012-01-24: print message for duplicate
572 dprintf("%s #%u Ddata %s\n",tintstr(),id_,pos.str(bin_name_buf));
573 evbuffer_drain(evb, length);
574 data_in_ = tintbin(TINT_NEVER,transfer().ack_out().cover(pos));
576 // Arno, 2012-01-24: Make sure data interarrival periods don't get
577 // screwed up because of these (ignored) duplicates.
581 uint8_t *data = evbuffer_pullup(evb, length);
582 data_in_ = tintbin(NOW,bin_t::NONE);
583 if (!file().OfferData(pos, (char*)data, length)) {
584 evbuffer_drain(evb, length);
585 char bin_name_buf[32];
586 dprintf("%s #%u !data %s\n",tintstr(),id_,pos.str(bin_name_buf));
589 evbuffer_drain(evb, length);
590 dprintf("%s #%u -data %s\n",tintstr(),id_,pos.str(bin_name_buf));
593 fprintf(stderr,"$ ");
595 bin_t cover = transfer().ack_out().cover(pos);
596 for(int i=0; i<transfer().cb_installed; i++)
597 if (cover.layer()>=transfer().cb_agg[i])
598 transfer().callbacks[i](transfer().fd(),cover); // FIXME
599 if (cover.layer() >= 5) // Arno: tested with 32K, presently = 2 ** 5 * chunk_size CHUNKSIZE
600 transfer().OnRecvData( pow((double)2,(double)5)*((double)file().chunk_size()) );
605 bytes_down_ += length;
606 global_bytes_down += length;
611 void Channel::UpdateDIP(bin_t pos)
613 if (!pos.is_none()) {
614 if (last_data_in_time_) {
615 tint dip = NOW - last_data_in_time_;
616 dip_avg_ = ( dip_avg_*3 + dip ) >> 2;
618 last_data_in_time_ = NOW;
623 void Channel::OnAck (struct evbuffer *evb) {
624 bin_t ackd_pos = bin_fromUInt32(evbuffer_remove_32be(evb));
625 tint peer_time = evbuffer_remove_64be(evb); // FIXME 32
626 // FIXME FIXME: wrap around here
627 if (ackd_pos.is_none())
628 return; // likely, broken chunk/ insufficient hashes
629 if (file().size() && ackd_pos.base_offset()>=file().size_in_chunks()) {
630 char bin_name_buf[32];
631 eprintf("invalid ack: %s\n",ackd_pos.str(bin_name_buf));
634 ack_in_.set(ackd_pos);
636 //fprintf(stderr,"OnAck: got bin %s is_complete %d\n", ackd_pos.str(), (int)ack_in_.is_complete_arno( file().ack_out().get_height() ));
639 // find an entry for the send (data out) event
640 while ( di<data_out_.size() && ( data_out_[di]==tintbin() ||
641 !ackd_pos.contains(data_out_[di].bin) ) )
643 // FUTURE: delayed acks
644 // rule out retransmits
645 while ( ri<data_out_tmo_.size() && !ackd_pos.contains(data_out_tmo_[ri].bin) )
647 char bin_name_buf[32];
648 dprintf("%s #%u %cack %s %lli\n",tintstr(),id_,
649 di==data_out_.size()?'?':'-',ackd_pos.str(bin_name_buf),peer_time);
650 if (di!=data_out_.size() && ri==data_out_tmo_.size()) { // not a retransmit
651 // round trip time calculations
652 tint rtt = NOW-data_out_[di].time;
653 rtt_avg_ = (rtt_avg_*7 + rtt) >> 3;
654 dev_avg_ = ( dev_avg_*3 + tintabs(rtt-rtt_avg_) ) >> 2;
655 assert(data_out_[di].time!=TINT_NEVER);
656 // one-way delay calculations
657 tint owd = peer_time - data_out_[di].time;
658 owd_cur_bin_ = 0;//(owd_cur_bin_+1) & 3;
659 owd_current_[owd_cur_bin_] = owd;
660 if ( owd_min_bin_start_+TINT_SEC*30 < NOW ) {
661 owd_min_bin_start_ = NOW;
662 owd_min_bin_ = (owd_min_bin_+1) & 3;
663 owd_min_bins_[owd_min_bin_] = TINT_NEVER;
665 if (owd_min_bins_[owd_min_bin_]>owd)
666 owd_min_bins_[owd_min_bin_] = owd;
667 dprintf("%s #%u sendctrl rtt %lli dev %lli based on %s\n",
668 tintstr(),id_,rtt_avg_,dev_avg_,data_out_[di].bin.str(bin_name_buf));
670 // early loss detection by packet reordering
671 for (int re=0; re<di-MAX_REORDERING; re++) {
672 if (data_out_[re]==tintbin())
674 ack_not_rcvd_recent_++;
675 data_out_tmo_.push_back(data_out_[re].bin);
676 dprintf("%s #%u Rdata %s\n",tintstr(),id_,data_out_.front().bin.str(bin_name_buf));
677 data_out_cap_ = bin_t::ALL;
678 data_out_[re] = tintbin();
681 if (di!=data_out_.size())
682 data_out_[di]=tintbin();
683 // clear zeroed items
684 while (!data_out_.empty() && ( data_out_.front()==tintbin() ||
685 ack_in_.is_filled(data_out_.front().bin) ) )
686 data_out_.pop_front();
687 assert(data_out_.empty() || data_out_.front().time!=TINT_NEVER);
691 void Channel::TimeoutDataOut ( ) {
692 // losses: timeouted packets
693 tint timeout = NOW - ack_timeout();
694 while (!data_out_.empty() &&
695 ( data_out_.front().time<timeout || data_out_.front()==tintbin() ) ) {
696 if (data_out_.front()!=tintbin() && ack_in_.is_empty(data_out_.front().bin)) {
697 ack_not_rcvd_recent_++;
698 data_out_cap_ = bin_t::ALL;
699 data_out_tmo_.push_back(data_out_.front().bin);
700 char bin_name_buf[32];
701 dprintf("%s #%u Tdata %s\n",tintstr(),id_,data_out_.front().bin.str(bin_name_buf));
703 data_out_.pop_front();
705 // clear retransmit queue of older items
706 while (!data_out_tmo_.empty() && data_out_tmo_.front().time<NOW-MAX_POSSIBLE_RTT)
707 data_out_tmo_.pop_front();
711 void Channel::OnHave (struct evbuffer *evb) {
712 bin_t ackd_pos = bin_fromUInt32(evbuffer_remove_32be(evb));
713 if (ackd_pos.is_none())
714 return; // wow, peer has hashes
717 if (ENABLE_VOD_PIECEPICKER) {
718 // Ric: check if we should set the size in the file transfer
719 if (transfer().availability().size() <= 0 && file().size() > 0)
721 transfer().availability().setSize(file().size_in_chunks());
723 // Ric: update the availability if needed
724 transfer().availability().set(id_, ack_in_, ackd_pos);
727 ack_in_.set(ackd_pos);
728 char bin_name_buf[32];
729 dprintf("%s #%u -have %s\n",tintstr(),id_,ackd_pos.str(bin_name_buf));
731 //fprintf(stderr,"OnHave: got bin %s is_complete %d\n", ackd_pos.str(), IsComplete() );
736 void Channel::OnHint (struct evbuffer *evb) {
737 bin_t hint = bin_fromUInt32(evbuffer_remove_32be(evb));
738 // FIXME: wake up here
739 hint_in_.push_back(hint);
740 char bin_name_buf[32];
741 dprintf("%s #%u -hint %s\n",tintstr(),id_,hint.str(bin_name_buf));
745 void Channel::OnHandshake (struct evbuffer *evb) {
747 uint32_t pcid = evbuffer_remove_32be(evb);
748 dprintf("%s #%u -hs %x\n",tintstr(),id_,pcid);
750 if (is_established() && pcid == 0) {
751 // Arno: received explicit close
752 peer_channel_id_ = 0; // == established -> false
757 peer_channel_id_ = pcid;
758 // self-connection check
760 uint32_t try_id = DecodeID(peer_channel_id_);
761 if (channel(try_id) && !channel(try_id)->peer_channel_id_) {
762 peer_channel_id_ = 0;
764 return; // this is a self-connection
768 // FUTURE: channel forking
769 if (is_established())
770 dprintf("%s #%u established %s\n", tintstr(), id_, peer().str());
774 void Channel::OnPex (struct evbuffer *evb) {
775 uint32_t ipv4 = evbuffer_remove_32be(evb);
776 uint16_t port = evbuffer_remove_16be(evb);
777 Address addr(ipv4,port);
778 dprintf("%s #%u -pex %s\n",tintstr(),id_,addr.str());
779 if (transfer().OnPexIn(addr))
780 useless_pex_count_ = 0;
783 dprintf("%s #%u already channel to %s\n", tintstr(),id_,addr.str());
784 useless_pex_count_++;
786 pex_request_outstanding_ = false;
791 void Channel::OnRandomize (struct evbuffer *evb) {
792 dprintf("%s #%u -rand\n",tintstr(),id_ );
793 // Payload is 4 random bytes
794 uint32_t r = evbuffer_remove_32be(evb);
798 void Channel::AddPex (struct evbuffer *evb) {
799 // Gertjan fix: Reverse PEX
800 // PEX messages sent to facilitate NAT/FW puncturing get priority
801 if (!reverse_pex_out_.empty()) {
803 tintbin pex_peer = reverse_pex_out_.front();
804 reverse_pex_out_.pop_front();
805 if (channels[(int) pex_peer.bin.toUInt()] == NULL)
807 Address a = channels[(int) pex_peer.bin.toUInt()]->peer();
808 // Arno, 2012-02-28: Don't send private addresses to non-private peers.
809 if (!a.is_private() || (a.is_private() && peer().is_private()))
811 evbuffer_add_8(evb, SWIFT_PEX_ADD);
812 evbuffer_add_32be(evb, a.ipv4());
813 evbuffer_add_16be(evb, a.port());
814 dprintf("%s #%u +pex (reverse) %s\n",tintstr(),id_,a.str());
816 } while (!reverse_pex_out_.empty() && (SWIFT_MAX_NONDATA_DGRAM_SIZE-evbuffer_get_length(evb)) >= 7);
818 // Arno: 2012-02-23: Don't think this is right. Bit of DoS thing,
819 // that you only get back the addr of people that got your addr.
827 // Arno, 2012-02-28: Don't send private addresses to non-private peers.
828 int chid = 0, tries=0;
832 // Arno, 2011-10-03: Choosing Gertjan's RandomChannel over RevealChannel here.
833 chid = transfer().RandomChannel(id_);
834 if (chid==-1 || chid==id_ || tries > 5) {
835 pex_requested_ = false;
838 a = channels[chid]->peer();
839 if (!a.is_private() || (a.is_private() && peer().is_private()))
844 evbuffer_add_8(evb, SWIFT_PEX_ADD);
845 evbuffer_add_32be(evb, a.ipv4());
846 evbuffer_add_16be(evb, a.port());
847 dprintf("%s #%u +pex %s\n",tintstr(),id_,a.str());
849 pex_requested_ = false;
850 /* Ensure that we don't add the same id to the reverse_pex_out_ queue
852 for (tbqueue::iterator i = channels[chid]->reverse_pex_out_.begin();
853 i != channels[chid]->reverse_pex_out_.end(); i++)
854 if ((int) (i->bin.toUInt()) == id_)
857 dprintf("%s #%u adding pex for channel %u at time %s\n", tintstr(), chid,
858 id_, tintstr(NOW + 2 * TINT_SEC));
859 // Arno, 2011-10-03: should really be a queue of (tint,channel id(= uint32_t)) pairs.
860 channels[chid]->reverse_pex_out_.push_back(tintbin(NOW + 2 * TINT_SEC, bin_t(id_)));
861 if (channels[chid]->send_control_ == KEEP_ALIVE_CONTROL &&
862 channels[chid]->next_send_time_ > NOW + 2 * TINT_SEC)
863 channels[chid]->Reschedule();
866 void Channel::OnPexReq(void) {
867 dprintf("%s #%u -pex req\n", tintstr(), id_);
868 if (NOW > MIN_PEX_REQUEST_INTERVAL + last_pex_request_time_)
869 pex_requested_ = true;
872 void Channel::AddPexReq(struct evbuffer *evb) {
873 // Rate limit the number of PEX requests
874 if (NOW < next_pex_request_time_)
877 // If no answer has been received from a previous request, count it as useless
878 if (pex_request_outstanding_)
879 useless_pex_count_++;
881 pex_request_outstanding_ = false;
883 // Initiate at most SWIFT_MAX_CONNECTIONS connections
884 if (transfer().hs_in_.size() >= SWIFT_MAX_CONNECTIONS ||
885 // Check whether this channel has been providing useful peer information
886 useless_pex_count_ > 2)
888 // Arno, 2012-02-23: Fix: Code doesn't recover from useless_pex_count_ > 2,
889 // let's just try again in 30s
890 useless_pex_count_ = 0;
891 next_pex_request_time_ = NOW + 30 * TINT_SEC;
896 dprintf("%s #%u +pex req\n", tintstr(), id_);
897 evbuffer_add_8(evb, SWIFT_PEX_REQ);
898 /* Add a little more than the minimum interval, such that the other party is
899 less likely to drop it due to too high rate */
900 next_pex_request_time_ = NOW + MIN_PEX_REQUEST_INTERVAL * 1.1;
901 pex_request_outstanding_ = true;
907 * Channel class methods
910 void Channel::LibeventReceiveCallback(evutil_socket_t fd, short event, void *arg) {
911 // Called by libevent when a datagram is received on the socket
914 event_add(&evrecv, NULL);
917 #define NUM_DATAGRAMS 1
919 void Channel::RecvDatagram (evutil_socket_t socket) {
920 struct evbuffer *pevb[NUM_DATAGRAMS];
921 for (int i=0; i<NUM_DATAGRAMS; ++i)
922 pevb[i] = evbuffer_new();
924 //FIXME: make this more readable
926 addr.addr = (struct sockaddr_mptp *) calloc(1, sizeof(struct sockaddr_mptp) + NUM_DATAGRAMS * sizeof(struct mptp_dest));
927 addr.addr->count = NUM_DATAGRAMS;
928 RecvFrom(socket, addr, pevb);
930 printf("Got %d addresses\n", addr.addr->count);
931 for (; i<addr.addr->count; ++i) {
932 struct evbuffer *evb = pevb[i];
934 fromi.set_ipv4(addr.addr->dests[i].addr);
935 fromi.set_port(addr.addr->dests[i].port);
936 size_t evboriglen = evbuffer_get_length(evb);
937 #define return_log(...) { fprintf(stderr,__VA_ARGS__); evbuffer_free(evb); return; }
938 if (evbuffer_get_length(evb)<4)
939 return_log("socket layer weird: datagram shorter than 4 bytes from %s (prob ICMP unreach)\n",fromi.str());
940 uint32_t mych = evbuffer_remove_32be(evb);
942 Channel* channel = NULL;
943 if (mych==0) { // peer initiates handshake
944 if (evbuffer_get_length(evb)<1+4+1+4+Sha1Hash::SIZE)
945 return_log ("%s #0 incorrect size %i initial handshake packet %s\n",
946 tintstr(),(int)evbuffer_get_length(evb),fromi.str());
947 uint8_t hashid = evbuffer_remove_8(evb);
948 if (hashid!=SWIFT_HASH)
949 return_log ("%s #0 no hash in the initial handshake %s\n",
950 tintstr(),fromi.str());
951 bin_t pos = bin_fromUInt32(evbuffer_remove_32be(evb));
953 return_log ("%s #0 that is not the root hash %s\n",tintstr(),fromi.str());
954 hash = evbuffer_remove_hash(evb);
955 FileTransfer* ft = FileTransfer::Find(hash);
957 return_log ("%s #0 hash %s unknown, requested by %s\n",tintstr(),hash.hex().c_str(),fromi.str());
958 dprintf("%s #0 -hash ALL %s\n",tintstr(),hash.hex().c_str());
960 // Arno, 2012-02-27: Check for duplicate channel
961 Channel* existchannel = ft->FindChannel(fromi,NULL);
964 // Arno: 2011-10-13: Ignore if established, otherwise consider
965 // it a concurrent connection attempt.
966 if (existchannel->is_established()) {
967 // ARNOTODO: Read complete handshake here so we know whether
968 // attempt is to new channel or to existing. Currently read
971 return_log("%s #0 have a channel already to %s\n",tintstr(),fromi.str());
973 channel = existchannel;
974 //fprintf(stderr,"Channel::RecvDatagram: HANDSHAKE: reuse channel %s\n", channel->peer_.str() );
977 if (channel == NULL) {
978 //fprintf(stderr,"Channel::RecvDatagram: HANDSHAKE: create new channel %s\n", addr.str() );
979 channel = new Channel(ft, socket, fromi);
981 //fprintf(stderr,"CHANNEL INCOMING DEF hass %s is id %d\n",hash.hex().c_str(),channel->id());
983 } else { // peer responds to my handshake (and other messages)
984 mych = DecodeID(mych);
985 if (mych>=channels.size())
986 return_log("%s invalid channel #%u, %s\n",tintstr(),mych,fromi.str());
987 channel = channels[mych];
989 return_log ("%s #%u is already closed\n",tintstr(),mych);
990 if (channel->IsDiffSenderOrDuplicate(fromi,mych)) {
991 channel->Schedule4Close();
994 channel->own_id_mentioned_ = true;
996 channel->raw_bytes_down_ += evboriglen;
997 //dprintf("recvd %i bytes for %i\n",data.size(),channel->id);
998 bool wasestablished = channel->is_established();
1000 dprintf("%s #%u peer %s recv_peer %s addr %s\n", tintstr(),mych, channel->peer().str(), channel->recv_peer().str(), fromi.str() );
1006 if (wasestablished && !channel->is_established()) {
1007 // Arno, 2012-01-26: Received an explict close, clean up channel, safely.
1008 channel->Schedule4Close();
1011 for (; i<NUM_DATAGRAMS; ++i)
1012 evbuffer_free(pevb[i]);
1018 * Channel instance methods
1021 void Channel::CloseChannelByAddress(const Address &addr)
1023 // fprintf(stderr,"CloseChannelByAddress: address is %s\n", addr.str() );
1024 std::vector<Channel *>::iterator iter;
1025 for (iter = channels.begin(); iter != channels.end(); iter++)
1028 if (c != NULL && c->peer_ == addr)
1030 // ARNOSMPTODO: will do another send attempt before not being
1032 c->peer_channel_id_ = 0; // established->false, do no more sending
1033 c->Schedule4Close();
1040 void Channel::Close () {
1042 this->SwitchSendControl(CLOSE_CONTROL);
1044 if (is_established())
1045 this->Send(); // Arno: send explicit close
1047 if (ENABLE_VOD_PIECEPICKER) {
1048 // Ric: remove it's binmap from the availability
1049 transfer().availability().remove(id_, ack_in_);
1053 // Arno: ensure LibeventSendCallback is no longer called with ptr to this Channel
1058 void Channel::Reschedule () {
1060 // Arno: CAREFUL: direct send depends on diff between next_send_time_ and
1061 // NOW to be 0, so any calls to Time in between may put things off. Sigh.
1063 next_send_time_ = NextSendTime();
1064 if (next_send_time_!=TINT_NEVER) {
1066 assert(next_send_time_<NOW+TINT_MIN);
1067 tint duein = next_send_time_-NOW;
1069 // Arno, 2011-10-18: libevent's timer implementation appears to be
1070 // really slow, i.e., timers set for 100 usec from now get called
1071 // at least two times later :-( Hence, for sends after receives
1072 // perform them directly.
1073 dprintf("%s #%u requeue direct send\n",tintstr(),id_);
1074 LibeventSendCallback(-1,EV_TIMEOUT,this);
1077 if (evsend_ptr_ != NULL) {
1078 struct timeval duetv = *tint2tv(duein);
1079 evtimer_add(evsend_ptr_,&duetv);
1080 dprintf("%s #%u requeue for %s in %lli\n",tintstr(),id_,tintstr(next_send_time_), duein);
1083 dprintf("%s #%u cannot requeue for %s, closed\n",tintstr(),id_,tintstr(next_send_time_));
1087 dprintf("%s #%u resched, will close\n",tintstr(),id_);
1088 this->Schedule4Close();
1094 * Channel class methods
1096 void Channel::LibeventSendCallback(int fd, short event, void *arg) {
1098 // Called by libevent when it is the requested send time.
1100 Channel * sender = (Channel*) arg;
1101 if (NOW<sender->next_send_time_-TINT_MSEC)
1102 dprintf("%s #%u suspicious send %s<%s\n",tintstr(),
1103 sender->id(),tintstr(NOW),tintstr(sender->next_send_time_));
1104 if (sender->next_send_time_ != TINT_NEVER)