librats 2.1.4 → 2.2.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -1,6 +1,7 @@
1
1
  #include "librats/subsystems/hole_punch.h"
2
2
  #include "librats/node/dial_service.h"
3
3
  #include "librats/node/node_context.h"
4
+ #include "librats/peer/peer_info.h"
4
5
  #include "librats/util/logger.h"
5
6
 
6
7
  #include <algorithm>
@@ -135,6 +136,7 @@ void HolePunch::attach(NodeContext& ctx) {
135
136
  // does not know its own external one simply cannot punch — it can still relay.
136
137
  dialer_ = ctx.services.get<DialService>();
137
138
  external_ = ctx.services.get<ExternalAddressService>();
139
+ services_ = &ctx.services;
138
140
 
139
141
  // Offer punching to sibling modules. Registered even without a dialer: punch()
140
142
  // answers false on its own then, which is exactly what a caller expects from a
@@ -151,17 +153,30 @@ void HolePunch::attach(NodeContext& ctx) {
151
153
  // reconciles sessions against the peer set, so this is the fast path, not the
152
154
  // only one.
153
155
  network_->on_peer_connected([this](const Peer& peer) {
156
+ // A relayed link is not what a punch was for — it is the fallback the punch
157
+ // is trying to make unnecessary, and a punch may well have been started
158
+ // BECAUSE it came up (see subsystems/relay.h). Retiring the session here
159
+ // would call the upgrade off the moment it began.
160
+ const auto info = peer.info();
161
+ if (info && info->transport == TransportKind::Relay) return;
154
162
  std::lock_guard<std::mutex> lock(mutex_);
155
163
  sessions_.erase(peer.id());
156
164
  });
157
165
  }
158
166
 
159
167
  void HolePunch::start() {
168
+ // Resolved here rather than in attach(): Relay may be attached after us, and
169
+ // every attach() runs before any start(). Stored before running_ goes true, so
170
+ // an escalation on a reactor thread never sees a half-initialised state.
171
+ if (config_.relay_on_failure && services_)
172
+ relay_.store(services_->get<RelayService>());
173
+
160
174
  if (running_.exchange(true)) return;
161
175
  worker_ = std::thread([this] { loop(); });
162
176
  }
163
177
 
164
178
  void HolePunch::stop() {
179
+ relay_.store(nullptr);
165
180
  if (!running_.exchange(false)) return;
166
181
  cv_.notify_all();
167
182
  if (worker_.joinable()) worker_.join();
@@ -178,14 +193,20 @@ bool HolePunch::punch(const PeerId& target) {
178
193
  if (target == network_->local_id()) return false;
179
194
 
180
195
  // Already reachable the ordinary way. Checked before any state is created so a
181
- // caller can punch() freely on discovery without having to check first.
182
- for (const PeerId& id : network_->connected_peers())
196
+ // caller can punch() freely on discovery without having to check first. A
197
+ // RELAYED peer is deliberately not "already reachable": punching it is how the
198
+ // circuit gets replaced by a direct link (see subsystems/relay.h).
199
+ for (const PeerId& id : directly_connected())
183
200
  if (id == target) return false;
184
201
 
185
202
  if (config_.skip_when_endpoint_dependent && external_ &&
186
203
  external_->udp_mapping() == NatMapping::EndpointDependent) {
187
204
  LOG_DEBUG("punch", "Not punching to " << target.short_hex()
188
205
  << ": our own mapping is per-destination (symmetric NAT)");
206
+ // Nothing about this will improve with time or retries: no endpoint we can
207
+ // advertise is the one the target's packets would arrive on. This is
208
+ // exactly the case a relayed path exists for.
209
+ escalate_to_relay(target);
189
210
  return false;
190
211
  }
191
212
 
@@ -193,6 +214,7 @@ bool HolePunch::punch(const PeerId& target) {
193
214
  if (addresses.empty()) {
194
215
  LOG_DEBUG("punch", "Not punching to " << target.short_hex()
195
216
  << ": no external datagram endpoint to advertise yet");
217
+ escalate_to_relay(target);
196
218
  return false;
197
219
  }
198
220
 
@@ -210,8 +232,11 @@ bool HolePunch::punch(const PeerId& target) {
210
232
  }
211
233
 
212
234
  if (!send_connect(target, /*opening=*/true, nullptr)) {
213
- std::lock_guard<std::mutex> lock(mutex_);
214
- sessions_.erase(target); // nobody could carry it; nothing is in flight
235
+ {
236
+ std::lock_guard<std::mutex> lock(mutex_);
237
+ sessions_.erase(target); // nobody could carry it; nothing is in flight
238
+ }
239
+ escalate_to_relay(target);
215
240
  return false;
216
241
  }
217
242
  LOG_DEBUG("punch", "Punch rendezvous started with " << target.short_hex());
@@ -461,10 +486,12 @@ void HolePunch::service_sessions() {
461
486
 
462
487
  // Connected peers first, without the lock: a session whose target is now a peer
463
488
  // succeeded, however it got there (our burst, or theirs arriving as inbound).
464
- std::vector<PeerId> connected = network_->connected_peers();
489
+ // Direct links only — a relayed one is what a punch is trying to replace.
490
+ std::vector<PeerId> connected = directly_connected();
465
491
 
466
492
  struct Retry { PeerId target; PeerId via; bool have_via; };
467
493
  std::vector<Retry> retry;
494
+ std::vector<PeerId> exhausted; ///< targets to hand on to the relay, once unlocked
468
495
  {
469
496
  std::lock_guard<std::mutex> lock(mutex_);
470
497
 
@@ -496,7 +523,10 @@ void HolePunch::service_sessions() {
496
523
  // first). Calling the target off here would drop the retries it is
497
524
  // about to send, and block this node's own later punch to it, for
498
525
  // the whole cooldown. So a responder just forgets the round.
499
- if (initiator) begin_cooldown(target);
526
+ if (initiator) {
527
+ begin_cooldown(target);
528
+ exhausted.push_back(target);
529
+ }
500
530
  it = sessions_.erase(it);
501
531
  continue;
502
532
  }
@@ -510,9 +540,52 @@ void HolePunch::service_sessions() {
510
540
 
511
541
  for (const Retry& r : retry) {
512
542
  if (send_connect(r.target, /*opening=*/true, r.have_via ? &r.via : nullptr)) continue;
513
- std::lock_guard<std::mutex> lock(mutex_);
514
- sessions_.erase(r.target);
543
+ {
544
+ std::lock_guard<std::mutex> lock(mutex_);
545
+ sessions_.erase(r.target);
546
+ }
547
+ exhausted.push_back(r.target);
515
548
  }
549
+
550
+ // The rungs above this one are spent: no address we can advertise got through.
551
+ // Handing the target on is the difference between a peer that is unreachable
552
+ // and one that is merely expensive to reach.
553
+ for (const PeerId& target : exhausted) escalate_to_relay(target);
554
+ }
555
+
556
+ void HolePunch::escalate_to_relay(const PeerId& target) {
557
+ RelayService* relay = relay_.load();
558
+ if (!relay) return; // no Relay attached, or the fallback is off
559
+ if (!relay->connect_via_relay(target)) return;
560
+
561
+ LOG_DEBUG("punch", "Punching to " << target.short_hex()
562
+ << " is not going to work; looking for a relay instead");
563
+
564
+ // The target has just changed hands, and the cooldown a give-up started must not
565
+ // outlive that. It exists to stop us hammering a peer we cannot reach — a job
566
+ // the relay now owns, with an attempt timeout and a cooldown of its own.
567
+ //
568
+ // And if that attempt lands, the circuit is precisely the new information that
569
+ // makes another punch worth trying: the two ends are peers at last, so the
570
+ // rendezvous can travel over the very circuit carrying them, which is what
571
+ // Relay asks for the moment the circuit comes up (see subsystems/relay.h).
572
+ // Leaving the cooldown standing would refuse that upgrade before it started —
573
+ // and nothing would ever ask again, so the circuit would outlive its purpose and
574
+ // keep costing a third node bandwidth for the life of the peer.
575
+ //
576
+ // Bounded, not a loop: an upgrade punch that fails lands here again, and
577
+ // connect_via_relay then answers false — the target is already a peer — so the
578
+ // fresh cooldown stands. Taken after the call, never around it: Relay resolves
579
+ // the reverse direction through us, and this mutex must not be held into it.
580
+ std::lock_guard<std::mutex> lock(mutex_);
581
+ cooldown_.erase(target);
582
+ }
583
+
584
+ std::vector<PeerId> HolePunch::directly_connected() const {
585
+ std::vector<PeerId> ids;
586
+ for (const PeerInfo& info : network_->peers())
587
+ if (info.transport != TransportKind::Relay) ids.push_back(info.id);
588
+ return ids;
516
589
  }
517
590
 
518
591
  bool HolePunch::in_cooldown(const PeerId& target) const {
@@ -95,6 +95,7 @@
95
95
  #include "librats/peer/peer.h"
96
96
  #include "librats/peer/peer_id.h"
97
97
  #include "librats/subsystems/hole_punch_service.h"
98
+ #include "librats/subsystems/relay_service.h"
98
99
 
99
100
  #include <atomic>
100
101
  #include <chrono>
@@ -108,6 +109,7 @@
108
109
  namespace librats {
109
110
 
110
111
  class DialService;
112
+ class ServiceRegistry;
111
113
 
112
114
  /// Published as HolePunchService, so a module that discovers a peer it cannot dial
113
115
  /// (PeerExchange) can hand the id over without depending on this class.
@@ -163,6 +165,12 @@ public:
163
165
 
164
166
  /// Session bookkeeping cadence — retries, timeouts, cooldown expiry.
165
167
  std::chrono::milliseconds tick{250};
168
+
169
+ /// When a punch cannot be attempted at all, or has been given up on, hand
170
+ /// the target to RelayService — the next rung down the ladder (see
171
+ /// subsystems/relay.h). Costs nothing when no Relay is attached: the
172
+ /// service simply does not resolve.
173
+ bool relay_on_failure = true;
166
174
  };
167
175
 
168
176
  HolePunch();
@@ -235,6 +243,17 @@ private:
235
243
  void service_sessions();
236
244
 
237
245
  // — helpers —
246
+ /// Hand `target` to the relay module, if one is attached and the fallback is
247
+ /// on. Called when a punch is impossible or has run out of attempts — the
248
+ /// point at which a relayed path stops being the worse option and becomes the
249
+ /// only one. Retires the target's cooldown when the relay takes it on, so that
250
+ /// the upgrade punch a circuit asks for is not refused by the very give-up that
251
+ /// produced the circuit. Caller must NOT hold the mutex.
252
+ void escalate_to_relay(const PeerId& target);
253
+ /// Peers reachable over a DIRECT link. A relayed peer is deliberately not one:
254
+ /// a punch to it is an upgrade in progress, and counting the circuit as success
255
+ /// would retire the session before it had done anything.
256
+ std::vector<PeerId> directly_connected() const;
238
257
  std::vector<Address> own_punch_addresses() const;
239
258
  bool relay_budget_ok(const PeerId& from);
240
259
  bool in_cooldown(const PeerId& target) const; ///< caller holds mutex_
@@ -244,6 +263,11 @@ private:
244
263
  PeerNetwork* network_ = nullptr;
245
264
  DialService* dialer_ = nullptr;
246
265
  ExternalAddressService* external_ = nullptr;
266
+ ServiceRegistry* services_ = nullptr;
267
+ /// Resolved in start(), not attach(): Relay may be attached after us, and every
268
+ /// attach() runs before any start(). Atomic because it is read from reactor
269
+ /// threads and the worker alike.
270
+ std::atomic<RelayService*> relay_{nullptr};
247
271
 
248
272
  std::atomic<bool> running_{false};
249
273
  std::atomic<uint64_t> punches_started_{0};