librats 1.0.2 → 2.1.4

package/native-src/src/librats/dht/routing_table.cpp

@@ -0,0 +1,559 @@
+#include "librats/dht/routing_table.h"
+#include "librats/dht/bep42.h"
+#include "librats/dht/log.h"
+
+#include <algorithm>
+#include <array>
+#include <iomanip>
+#include <sstream>
+#include <string>
+
+namespace librats {
+namespace dht {
+
+namespace {
+
+// Locate a contact by id within one set (live or replacements).
+std::vector<NodeEntry>::iterator find_by_id(std::vector<NodeEntry>& v, const NodeId& id) {
+    return std::find_if(v.begin(), v.end(), [&](const NodeEntry& n) { return n.id == id; });
+}
+
+// The worst contact in `v` (the one every other is "no worse than"), or end() if empty.
+std::vector<NodeEntry>::iterator worst_of(std::vector<NodeEntry>& v) {
+    auto worst = v.begin();
+    for (auto it = v.begin(); it != v.end(); ++it)
+        if (it->is_worse_than(*worst)) worst = it;
+    return worst;
+}
+
+// log2 of a power-of-two bucket size: 8 -> 3 ... 128 -> 7.
+constexpr int prefix_bits_for(int bucket_size) noexcept {
+    int bits = 0;
+    while ((1 << bits) < bucket_size) ++bits;
+    return bits;
+}
+
+// The largest live-set limit any bucket can have (the widest top bucket); see
+// RoutingTable::bucket_limit. The spread classifier yields prefix_bits_for(limit)
+// bits, so this caps the number of distinct sub-prefix slots.
+constexpr int kMaxBucketLimit = static_cast<int>(kBucketSize) * 16;
+
+// Classify a node by the few sub-prefix bits right after the bucket's shared prefix.
+// For a non-last bucket we skip the one bit that *defines* the bucket (it's the same
+// for every entry); the last bucket hasn't split yet and holds both sides, so we keep
+// it. The result indexes a "spread slot" in [0, bucket_size).
+uint8_t classify_prefix(int bucket_idx, bool last_bucket, int bucket_size, const NodeId& id) noexcept {
+    const int shift = bucket_idx + (last_bucket ? 0 : 1);
+    return static_cast<uint8_t>(bits_at(id, shift, prefix_bits_for(bucket_size)));
+}
+
+// True if every live node and `id` carry the same bit at position `idx`: a split at
+// that bit wouldn't separate anyone, so it would be pointless.
+bool all_on_same_side(const std::vector<NodeEntry>& live, const NodeId& id, int idx) noexcept {
+    const int byte = idx >> 3;
+    const int bit  = idx & 7;
+    const uint8_t mask = static_cast<uint8_t>(0x80u >> bit);
+    int count[2] = {0, 0};
+    ++count[(id[byte] & mask) ? 1 : 0];
+    for (const auto& n : live) ++count[(n.id[byte] & mask) ? 1 : 0];
+    return count[0] == 0 || count[1] == 0;
+}
+
+} // namespace
+
+RoutingTable::RoutingTable(const NodeId& self, bool extended)
+    : self_(self), extended_(extended) {
+    buckets_.emplace_back();  // always keep at least the catch-all bucket
+}
+
+int RoutingTable::bucket_index(const NodeId& id) const noexcept {
+    const int prefix = shared_prefix_bits(self_, id);
+    const int last = static_cast<int>(buckets_.size()) - 1;
+    return prefix < last ? prefix : last;
+}
+
+int RoutingTable::bucket_limit(int index) const noexcept {
+    if (!extended_) return static_cast<int>(kBucketSize);
+    // The top buckets hold more contacts (16k, 8k, 4k, 2k) — that's where the
+    // keyspace and the first lookup hop are densest.
+    static constexpr int kMult[4] = {16, 8, 4, 2};
+    return static_cast<int>(kBucketSize) * (index < 4 ? kMult[index] : 1);
+}
+
+void RoutingTable::heard_about(const NodeId& id, const Address& endpoint, bool verified) {
+    if (id == self_) return;
+    NodeEntry e(id, endpoint);  // unpinged until it replies
+    e.verified = verified;
+    add_node(e);
+}
+
+bool RoutingTable::node_seen(const NodeId& id, const Address& endpoint,
+                             uint16_t rtt, bool verified) {
+    if (id == self_) return false;
+    NodeEntry entry(id, endpoint);
+    entry.verified = verified;
+    entry.record_success(rtt);  // confirmed (fail_count = 0), folds in the RTT
+    return add_node(entry);
+}
+
+bool RoutingTable::add_node(NodeEntry e) {
+    if (e.id == self_) return false;
+    const int idx = bucket_index(e.id);
+    Bucket& b = buckets_[idx];
+
+    // Already live → refresh in place. A confirmed sighting also records success;
+    // mere hearsay must not reset a live contact's quality. A confirmed contact that
+    // moved to a new endpoint re-points the IP index too.
+    if (auto it = find_by_id(b.live, e.id); it != b.live.end()) {
+        it->verified = it->verified || e.verified;
+        if (e.pinged()) {
+            if (it->endpoint.ip != e.endpoint.ip) { ip_untrack(it->endpoint.ip); ip_track(e.endpoint.ip); }
+            it->endpoint = e.endpoint;
+            it->record_success(e.rtt);
+        }
+        return true;
+    }
+
+    // Sitting in the replacement cache: a confirmed sighting pulls it out to promote,
+    // carrying its accumulated quality; mere hearsay leaves it parked. The erase
+    // untracks its IP; the shared insert below re-tracks it (at its current endpoint),
+    // so a promotion is IP-neutral even if the endpoint changed.
+    bool promoted = false;
+    if (auto it = find_by_id(b.replacements, e.id); it != b.replacements.end()) {
+        if (!e.pinged()) return false;
+        NodeEntry carried = *it;
+        carried.endpoint = e.endpoint;
+        carried.verified = carried.verified || e.verified;
+        carried.record_success(e.rtt);
+        ip_untrack(it->endpoint.ip);
+        b.replacements.erase(it);
+        e = carried;
+        promoted = true;
+    }
+
+    // A genuinely new IP must pass the IP-diversity gate (rule 1 + rule 2, public IPs
+    // only). A promoted contact is already one we know, so it skips the gate.
+    if (!promoted && !passes_ip_diversity(b, idx, e)) return false;
+
+    const int limit = bucket_limit(idx);
+    if (static_cast<int>(b.live.size()) < limit) {
+        b.live.push_back(e);
+        ip_track(e.endpoint.ip);
+        return true;
+    }
+
+    // The bucket is full. A confirmed contact may split it or displace a weaker one;
+    // hearsay just waits in the replacement cache.
+    if (e.confirmed()) {
+        if (can_split(idx, e)) {
+            split_bucket();              // buckets idx and idx + 1 now exist
+            const bool added = add_node(e);  // re-home e against the now-deeper table
+            // The split freed live slots; pull the best standbys up into them, but
+            // only after e has claimed its rightful slot.
+            fill_from_replacements(idx);
+            fill_from_replacements(idx + 1);
+            return added;
+        }
+        if (replace_for_spread(b, idx, limit, e)) return true;
+    }
+    stash_replacement(b, e);
+    return false;
+}
+
+bool RoutingTable::can_split(int index, const NodeEntry& e) const {
+    return index + 1 == static_cast<int>(buckets_.size())          // only the last bucket splits
+        && static_cast<int>(buckets_.size()) < kBucketCount - 1     // bounded depth
+        && e.confirmed()
+        && (index == 0 || buckets_[index - 1].live.size() > 1)      // don't deepen into near-empty
+        && !all_on_same_side(buckets_[index].live, e.id, index);    // a split must actually separate
+}
+
+void RoutingTable::split_bucket() {
+    const int last = static_cast<int>(buckets_.size()) - 1;
+    buckets_.emplace_back();  // may reallocate buckets_ — hold no Bucket& across this
+
+    Bucket moved = std::move(buckets_[last]);
+    buckets_[last] = Bucket{};
+
+    // Contacts sharing exactly `last` prefix bits with us stay; closer ones (longer
+    // shared prefix) move to the new, deeper bucket.
+    const auto place = [&](std::vector<NodeEntry>& src, bool live) {
+        for (auto& n : src) {
+            const int dest = shared_prefix_bits(self_, n.id) <= last ? last : last + 1;
+            auto& bucket = buckets_[dest];
+            (live ? bucket.live : bucket.replacements).push_back(std::move(n));
+        }
+    };
+    place(moved.live, true);
+    place(moved.replacements, false);
+
+    // The deeper bucket has a smaller limit, so it may now overflow — spill the rest.
+    trim_to_limit(last);
+    trim_to_limit(last + 1);
+
+    // Structural milestone, not per-packet churn: a bucket deepens only a bounded number
+    // of times as the table grows, so this stays quiet (unlike add/seen/fail, which must
+    // never log). The live composition over time is the heartbeat's describe() job.
+    LOG_DEBUG("dht.route", "bucket split → " << buckets_.size() << " bucket(s): #" << last
+                           << " keeps " << buckets_[last].live.size() << ", #" << (last + 1)
+                           << " gets " << buckets_[last + 1].live.size() << " live");
+}
+
+bool RoutingTable::replace_for_spread(Bucket& b, int index, int limit, const NodeEntry& e) {
+    auto& live = b.live;
+    const bool last = (index + 1 == static_cast<int>(buckets_.size()));
+
+    // 1) A node that has actually failed a query is the first to go.
+    NodeEntry* stale = nullptr;
+    for (auto& n : live)
+        if (n.pinged() && n.fail_count > 0 && (!stale || n.fail_count > stale->fail_count))
+            stale = &n;
+    if (stale) { assign(*stale, e); return true; }
+
+    // 2) Otherwise keep a spread of sub-prefixes. Group the live nodes by their slot.
+    // classify_prefix yields prefix_bits_for(limit) bits, so the slot index is in
+    // [0, 2^prefix_bits). Size the table to the largest possible bucket limit so a
+    // wider top bucket can never index out of bounds (tied to bucket_limit's cap).
+    static constexpr int kMaxSlots = 1 << prefix_bits_for(kMaxBucketLimit);
+    static_assert(kMaxSlots <= 256, "slot index must fit in the uint8_t classify_prefix returns");
+    const uint8_t want = classify_prefix(index, last, limit, e.id);
+    std::array<std::vector<NodeEntry*>, kMaxSlots> slot{};
+    for (auto& n : live) slot[classify_prefix(index, last, limit, n.id)].push_back(&n);
+
+    const auto worst_in = [](const std::vector<NodeEntry*>& v) {
+        NodeEntry* worst = v.front();
+        for (auto* n : v) if (n->is_worse_than(*worst)) worst = n;
+        return worst;
+    };
+
+    if (!slot[want].empty()) {
+        // Our slot is taken: displace its worst node only if we're strictly better.
+        NodeEntry* worst = worst_in(slot[want]);
+        if (worst->is_worse_than(e)) { assign(*worst, e); return true; }
+        return false;
+    }
+
+    // Our slot is empty: make room by dropping the worst node from an over-full slot,
+    // so a healthy contact alone in its slot is never sacrificed.
+    NodeEntry* victim = nullptr;
+    for (const auto& s : slot)
+        if (s.size() > 1)
+            for (auto* n : s)
+                if (!victim || n->is_worse_than(*victim)) victim = n;
+    if (victim) { assign(*victim, e); return true; }
+    return false;
+}
+
+void RoutingTable::stash_replacement(Bucket& b, const NodeEntry& entry) {
+    auto& r = b.replacements;
+    if (auto it = find_by_id(r, entry.id); it != r.end()) {
+        assign(*it, entry);
+        return;
+    }
+    if (r.size() < kBucketSize) {
+        r.push_back(entry);
+        ip_track(entry.endpoint.ip);
+        return;
+    }
+    auto worst = worst_of(r);
+    if (worst != r.end() && worst->is_worse_than(entry)) assign(*worst, entry);
+}
+
+void RoutingTable::trim_to_limit(int index) {
+    Bucket& b = buckets_[index];
+    const int limit = bucket_limit(index);
+    // Spilling a live contact into replacements keeps it in the table, so the IP index
+    // is untouched. Only the final overflow — a contact dropped from the table — untracks.
+    while (static_cast<int>(b.live.size()) > limit) {
+        auto worst = worst_of(b.live);
+        b.replacements.push_back(std::move(*worst));
+        b.live.erase(worst);
+    }
+    while (b.replacements.size() > kBucketSize) {
+        auto worst = worst_of(b.replacements);
+        ip_untrack(worst->endpoint.ip);
+        b.replacements.erase(worst);
+    }
+}
+
+void RoutingTable::fill_from_replacements(int index) {
+    Bucket& b = buckets_[index];
+    const int limit = bucket_limit(index);
+    while (static_cast<int>(b.live.size()) < limit && !b.replacements.empty()) {
+        // Promote the best standby — the one no other standby is better than, so
+        // confirmed contacts go before unpinged hearsay.
+        auto best = b.replacements.begin();
+        for (auto it = b.replacements.begin(); it != b.replacements.end(); ++it)
+            if (best->is_worse_than(*it)) best = it;
+        b.live.push_back(std::move(*best));
+        b.replacements.erase(best);
+    }
+}
+
+void RoutingTable::node_failed(const NodeId& id, const Address& endpoint) {
+    const int idx = bucket_index(id);
+    Bucket& b = buckets_[idx];
+
+    auto it = find_by_id(b.live, id);
+    if (it == b.live.end()) {
+        // A standby we were probing can fail too — drop it if it's spent. Only when the
+        // endpoint matches, though: a different node claiming the same id must not knock
+        // out the standby we already have for it.
+        if (auto r = find_by_id(b.replacements, id);
+            r != b.replacements.end() && r->endpoint == endpoint) {
+            r->record_failure();
+            if (!r->pinged() || r->fail_count >= kMaxFailCount) {
+                ip_untrack(r->endpoint.ip);
+                b.replacements.erase(r);
+            }
+        }
+        return;
+    }
+
+    // The id is live but now points at a different endpoint — a more recent sighting from
+    // another node overwrote it. This timeout belongs to the old node, so ignore it
+    // rather than penalise (or evict) the current, healthy contact.
+    if (it->endpoint != endpoint) return;
+
+    it->record_failure();
+
+    // A liveness check already failed here, so if a standby is ready, evict the
+    // failed contact and let the best standbys take the freed slot(s).
+    if (!b.replacements.empty()) {
+        ip_untrack(it->endpoint.ip);
+        b.live.erase(it);
+        fill_from_replacements(idx);
+        prune_empty_back();
+        return;
+    }
+
+    // No standby: keep giving it chances until it's clearly dead.
+    if (!it->pinged() || it->fail_count >= kMaxFailCount) {
+        ip_untrack(it->endpoint.ip);
+        b.live.erase(it);
+    }
+    prune_empty_back();
+}
+
+std::vector<NodeEntry> RoutingTable::find_closest(const NodeId& target, std::size_t count,
+                                                  bool include_unconfirmed) const {
+    std::vector<NodeEntry> result;
+    if (count == 0) return result;
+
+    const auto take = [&](const Bucket& b) {
+        for (const auto& n : b.live)
+            if (include_unconfirmed || n.confirmed()) result.push_back(n);
+    };
+
+    // By the XOR metric's prefix structure, the buckets form distance tiers around the
+    // target's own bucket `start`:
+    //   - bucket `start` is the *closest* tier: every contact in it is strictly nearer
+    //     than any contact in a deeper bucket, so if it alone yields `count` we're done.
+    //   - the deeper buckets (> start) are all tied at the same leading distance bit and
+    //     are unordered among themselves, so they must be taken as one whole tier or not
+    //     at all — we can stop before them, never partway through.
+    //   - the shallower buckets (< start) are each a strictly farther tier in order, so
+    //     they can be walked closest-first with an early stop the moment we have enough.
+    const int start = bucket_index(target);
+    result.reserve(count + static_cast<std::size_t>(buckets_[start].live.size()));
+
+    take(buckets_[start]);
+    if (result.size() < count)  // closest tier short -> the deeper tier is needed in full
+        for (int i = start + 1; i < static_cast<int>(buckets_.size()); ++i) take(buckets_[i]);
+    for (int i = start - 1; i >= 0 && result.size() < count; --i) take(buckets_[i]);
+
+    const std::size_t k = (std::min)(count, result.size());
+    std::partial_sort(result.begin(), result.begin() + k, result.end(),
+        [&](const NodeEntry& a, const NodeEntry& c) { return closer_to(a.id, c.id, target); });
+    if (result.size() > count) result.resize(count);
+    return result;
+}
+
+std::optional<NodeEntry> RoutingTable::next_to_refresh(std::chrono::steady_clock::time_point now) {
+    // Order by last touch: a never-contacted contact (last_seen == min) sorts first,
+    // then the least-recently-touched.
+    NodeEntry* best = nullptr;
+    for (auto& b : buckets_)
+        for (auto& n : b.live)
+            if (!best || n.last_seen < best->last_seen) best = &n;
+
+    if (!best) return std::nullopt;
+
+    // Stamp it as touched now: we're about to probe it, so the next refresh moves on to
+    // a different contact instead of re-probing this one while its probe is in flight.
+    // A reply refreshes last_seen again; a timeout drops the contact (node_failed).
+    best->last_seen = now;
+    return *best;
+}
+
+std::size_t RoutingTable::size() const {
+    std::size_t n = 0;
+    for (const auto& b : buckets_) n += b.live.size();
+    return n;
+}
+
+void RoutingTable::set_self(const NodeId& id) {
+    if (id == self_) return;
+
+    // Drain every contact, re-root the table to a single bucket, then re-home each
+    // one — bucket indices are all relative to self_ and are now stale.
+    std::vector<NodeEntry> all;
+    for (auto& b : buckets_) {
+        for (auto& n : b.live) all.push_back(std::move(n));
+        for (auto& n : b.replacements) all.push_back(std::move(n));
+    }
+
+    self_ = id;
+    buckets_.clear();
+    buckets_.emplace_back();
+    ip_count_.clear();  // rebuilt as add_node re-tracks each contact
+    for (const auto& n : all) add_node(n);
+}
+
+std::vector<NodeEntry> RoutingTable::good_contacts() const {
+    std::vector<NodeEntry> out;
+    for (const auto& b : buckets_)
+        for (const auto& n : b.live)
+            if (n.confirmed()) out.push_back(n);
+    return out;
+}
+
+void RoutingTable::load_contacts(const std::vector<NodeEntry>& contacts) {
+    for (const auto& n : contacts) add_node(n);
+}
+
+std::string RoutingTable::describe() const {
+    // One compact line per contact: "<id8> <ip:port> <rtt> <flags>", where flags are
+    //   ?  = only heard about (never replied)   !N = N consecutive failures   v = BEP 42-verified
+    const auto fmt_contact = [](const NodeEntry& n) {
+        std::ostringstream s;
+        s << short_hex(n.id) << ' ' << n.endpoint.to_string() << ' ';
+        if (n.rtt == NodeEntry::kRttUnknown) s << "-ms";
+        else                                 s << n.rtt << "ms";
+        if (!n.pinged())           s << " ?";
+        else if (n.fail_count > 0) s << " !" << static_cast<int>(n.fail_count);
+        if (n.verified)            s << " v";
+        return s.str();
+    };
+    // A 10-cell bar showing how full the live set is relative to this bucket's limit.
+    const auto fill_bar = [](std::size_t live, int limit) {
+        constexpr int width = 10;
+        int filled = limit > 0 ? static_cast<int>((live * width + limit / 2) / limit) : 0;
+        if (filled > width) filled = width;
+        return std::string(static_cast<std::size_t>(filled), '#') +
+               std::string(static_cast<std::size_t>(width - filled), '.');
+    };
+
+    std::ostringstream os;
+    os << "routing table: " << size() << " node(s) / " << buckets_.size()
+       << " bucket(s)  self " << short_hex(self_);
+    for (int i = 0; i < static_cast<int>(buckets_.size()); ++i) {
+        const Bucket& b = buckets_[i];
+        const int limit = bucket_limit(i);
+
+        std::ostringstream cnt;
+        cnt << b.live.size() << '/' << limit;
+
+        os << "\n  #" << std::right << std::setw(2) << i << "  "
+           << std::left << std::setw(7) << cnt.str()
+           << "repl " << std::right << std::setw(2) << b.replacements.size() << "  "
+           << '[' << fill_bar(b.live.size(), limit) << "]  ";
+
+        if (b.live.empty()) { os << "(empty)"; continue; }
+
+        // best = the contact no other is better than; worst = the first to be evicted.
+        const NodeEntry* best  = &b.live.front();
+        const NodeEntry* worst = &b.live.front();
+        for (const auto& n : b.live) {
+            if (best->is_worse_than(n)) best = &n;
+            if (n.is_worse_than(*worst)) worst = &n;
+        }
+        os << "best " << fmt_contact(*best);
+        if (b.live.size() > 1) os << "  worst " << fmt_contact(*worst);
+    }
+    return os.str();
+}
+
+void RoutingTable::prune_empty_back() {
+    while (buckets_.size() > 1
+           && buckets_.back().live.empty()
+           && buckets_.back().replacements.empty())
+        buckets_.pop_back();
+}
+
+// -- IP-diversity admission ---------------------------------------------------
+
+void RoutingTable::ip_track(const IpAddress& ip) {
+    if (is_public_address(ip)) ++ip_count_[ip];
+}
+
+void RoutingTable::ip_untrack(const IpAddress& ip) {
+    if (!is_public_address(ip)) return;
+    auto it = ip_count_.find(ip);
+    if (it == ip_count_.end()) return;         // defensive: never underflow past zero
+    if (--it->second <= 0) ip_count_.erase(it);
+}
+
+void RoutingTable::assign(NodeEntry& slot, const NodeEntry& e) {
+    ip_untrack(slot.endpoint.ip);              // the displaced contact leaves the table
+    slot = e;
+    ip_track(e.endpoint.ip);                   // ... and the new one takes its place
+}
+
+std::optional<RoutingTable::Located> RoutingTable::find_by_endpoint(const Address& ep) {
+    for (int i = 0; i < static_cast<int>(buckets_.size()); ++i) {
+        Bucket& b = buckets_[i];
+        for (std::size_t j = 0; j < b.live.size(); ++j)
+            if (b.live[j].endpoint == ep) return Located{i, true, j};
+        for (std::size_t j = 0; j < b.replacements.size(); ++j)
+            if (b.replacements[j].endpoint == ep) return Located{i, false, j};
+    }
+    return std::nullopt;
+}
+
+bool RoutingTable::passes_ip_diversity(Bucket& b, int idx, const NodeEntry& e) {
+    const IpAddress& ip = e.endpoint.ip;
+    if (!is_public_address(ip)) return true;   // private/loopback/CGNAT: unconstrained
+
+    // Rule 1 — one contact per IP across the whole table. The same-id sighting was
+    // already handled in add_node, so a hit here means this IP is held under a
+    // *different* id: a routing-table poisoning signal.
+    if (ip_count_.find(ip) != ip_count_.end()) {
+        auto found = find_by_endpoint(e.endpoint);
+        if (!found)       return false;        // same IP, different port -> ignore newcomer
+        if (!e.pinged())  return false;        // unconfirmed id swap  -> ignore (poison)
+
+        // A *confirmed* contact is answering from an endpoint we already hold under
+        // another id. The old entry is now suspect, so drop it and don't trust the new
+        // one either — libtorrent's conservative "evict on changed id". The freed live
+        // slot refills from that bucket's standbys.
+        Bucket& eb = buckets_[found->bucket];
+        auto& set  = found->live ? eb.live : eb.replacements;
+        ip_untrack(set[found->index].endpoint.ip);
+        set.erase(set.begin() + static_cast<std::ptrdiff_t>(found->index));
+        if (found->live) fill_from_replacements(found->bucket);
+        prune_empty_back();
+        return false;
+    }
+
+    // Rule 2 — at most one contact per /24 (v4) or /64 (v6) within this one bucket.
+    for (const auto& n : b.live)
+        if (ip_too_close(n.endpoint.ip, ip)) return false;
+    for (const auto& n : b.replacements)
+        if (ip_too_close(n.endpoint.ip, ip)) return false;
+    return true;
+}
+
+bool RoutingTable::ip_index_consistent() const {
+    std::unordered_map<IpAddress, int> expected;
+    for (const auto& b : buckets_) {
+        for (const auto& n : b.live)
+            if (is_public_address(n.endpoint.ip)) ++expected[n.endpoint.ip];
+        for (const auto& n : b.replacements)
+            if (is_public_address(n.endpoint.ip)) ++expected[n.endpoint.ip];
+    }
+    return expected == ip_count_;
+}
+
+} // namespace dht
+} // namespace librats