lokinet/llarp/dns/server.cpp

#include "server.hpp"

#include "nm_platform.hpp"
#include "sd_platform.hpp"

#include <llarp/constants/apple.hpp>
#include <llarp/constants/platform.hpp>
#include <llarp/ev/udp_handle.hpp>

#include <oxen/log.hpp>
#include <unbound.h>
#include <uvw.hpp>

#include <memory>
#include <optional>
#include <stdexcept>
#include <utility>

namespace llarp::dns
{
  static auto logcat = log::Cat("dns");

  void
  QueryJob_Base::Cancel()
  {
    Message reply{m_Query};
    reply.AddServFail();
    SendReply(reply.ToBuffer());
  }

  /// sucks up udp packets from a bound socket and feeds it to a server
  class UDPReader : public PacketSource_Base, public std::enable_shared_from_this<UDPReader>
  {
    Server& m_DNS;
    std::shared_ptr<llarp::UDPHandle> m_udp;
    SockAddr m_LocalAddr;

   public:
    explicit UDPReader(Server& dns, const EventLoop_ptr& loop, llarp::SockAddr bindaddr)
        : m_DNS{dns}
    {
      m_udp = loop->make_udp([&](auto&, SockAddr src, llarp::OwnedBuffer buf) {
        if (src == m_LocalAddr)
          return;
        if (not m_DNS.MaybeHandlePacket(shared_from_this(), m_LocalAddr, src, std::move(buf)))
        {
          log::warning(logcat, "did not handle dns packet from {} to {}", src, m_LocalAddr);
        }
      });
      m_udp->listen(bindaddr);
      if (auto maybe_addr = BoundOn())
      {
        m_LocalAddr = *maybe_addr;
      }
      else
        throw std::runtime_error{"cannot find which address our dns socket is bound on"};
    }

    std::optional<SockAddr>
    BoundOn() const override
    {
      return m_udp->LocalAddr();
    }

    bool
    WouldLoop(const SockAddr& to, const SockAddr&) const override
    {
      return to != m_LocalAddr;
    }

    void
    SendTo(const SockAddr& to, const SockAddr&, llarp::OwnedBuffer buf) const override
    {
      m_udp->send(to, std::move(buf));
    }

    void
    Stop() override
    {
      m_udp->close();
    }
  };

  namespace libunbound
  {
    class Resolver;

    class Query : public QueryJob_Base, public std::enable_shared_from_this<Query>
    {
      std::shared_ptr<PacketSource_Base> src;
      SockAddr resolverAddr;
      SockAddr askerAddr;

     public:
      explicit Query(
          std::weak_ptr<Resolver> parent_,
          Message query,
          std::shared_ptr<PacketSource_Base> pktsrc,
          SockAddr toaddr,
          SockAddr fromaddr)
          : QueryJob_Base{std::move(query)}
          , src{std::move(pktsrc)}
          , resolverAddr{std::move(toaddr)}
          , askerAddr{std::move(fromaddr)}
          , parent{parent_}
      {}
      std::weak_ptr<Resolver> parent;
      int id{};

      void
      SendReply(llarp::OwnedBuffer replyBuf) override;
    };

    /// Resolver_Base that uses libunbound
    class Resolver final : public Resolver_Base, public std::enable_shared_from_this<Resolver>
    {
      ub_ctx* m_ctx = nullptr;
      std::weak_ptr<EventLoop> m_Loop;
#ifdef _WIN32
      // windows is dumb so we do ub mainloop in a thread
      std::thread runner;
      std::atomic<bool> running;
#else
      std::shared_ptr<uvw::PollHandle> m_Poller;
#endif

      std::optional<SockAddr> m_LocalAddr;
      std::unordered_set<std::shared_ptr<Query>> m_Pending;

      struct ub_result_deleter
      {
        void
        operator()(ub_result* ptr)
        {
          ::ub_resolve_free(ptr);
        }
      };

      const net::Platform*
      Net_ptr() const
      {
        return m_Loop.lock()->Net_ptr();
      }

      static void
      Callback(void* data, int err, ub_result* _result)
      {
        log::debug(logcat, "got dns response from libunbound");
        // take ownership of ub_result
        std::unique_ptr<ub_result, ub_result_deleter> result{_result};
        // borrow query
        auto* query = static_cast<Query*>(data);
        if (err)
        {
          // some kind of error from upstream
          log::warning(logcat, "Upstream DNS failure: {}", ub_strerror(err));
          query->Cancel();
          return;
        }

        log::trace(logcat, "queueing dns response from libunbound to userland");

        // rewrite response
        OwnedBuffer pkt{(const byte_t*)result->answer_packet, (size_t)result->answer_len};
        llarp_buffer_t buf{pkt};
        MessageHeader hdr;
        hdr.Decode(&buf);
        hdr.id = query->Underlying().hdr_id;
        buf.cur = buf.base;
        hdr.Encode(&buf);

        // send reply
        query->SendReply(std::move(pkt));
      }

      void
      AddUpstreamResolver(const SockAddr& dns)
      {
        std::string str = fmt::format("{}@{}", dns.hostString(false), dns.getPort());

        if (auto err = ub_ctx_set_fwd(m_ctx, str.c_str()))
        {
          throw std::runtime_error{
              fmt::format("cannot use {} as upstream dns: {}", str, ub_strerror(err))};
        }
      }

      bool
      ConfigureAppleTrampoline(const SockAddr& dns)
      {
        // On Apple, when we turn on exit mode, we tear down and then reestablish the unbound
        // resolver: in exit mode, we set use upstream to a localhost trampoline that redirects
        // packets through the tunnel.  In non-exit mode, we directly use the upstream, so we look
        // here for a reconfiguration to use the trampoline port to check which state we're in.
        //
        // We have to do all this crap because we can't directly connect to upstream from here:
        // within the network extension, macOS ignores the tunnel we are managing and so, if we
        // didn't do this, all our DNS queries would leak out around the tunnel.  Instead we have to
        // bounce things through the objective C trampoline code (which is what actually handles the
        // upstream querying) so that it can call into Apple's special snowflake API to set up a
        // socket that has the magic Apple snowflake sauce added on top so that it actually routes
        // through the tunnel instead of around it.
        //
        // But the trampoline *always* tries to send the packet through the tunnel, and that will
        // only work in exit mode.
        //
        // All of this macos behaviour is all carefully and explicitly documented by Apple with
        // plenty of examples and other exposition, of course, just like all of their wonderful new
        // APIs to reinvent standard unix interfaces with half-baked replacements.

        if constexpr (platform::is_apple)
        {
          if (dns.hostString() == "127.0.0.1" and dns.getPort() == apple::dns_trampoline_port)
          {
            // macOS is stupid: the default (0.0.0.0) fails with "send failed: Can't assign
            // requested address" when unbound tries to connect to the localhost address using a
            // source address of 0.0.0.0.  Yay apple.
            SetOpt("outgoing-interface:", "127.0.0.1");

            // The trampoline expects just a single source port (and sends everything back to it).
            SetOpt("outgoing-range:", "1");
            SetOpt("outgoing-port-avoid:", "0-65535");
            SetOpt("outgoing-port-permit:", "{}", apple::dns_trampoline_source_port);
            return true;
          }
        }
        return false;
      }

      void
      ConfigureUpstream(const llarp::DnsConfig& conf)
      {
        bool is_apple_tramp = false;

        // set up forward dns
        for (const auto& dns : conf.upstream_dns)
        {
          AddUpstreamResolver(dns);
          is_apple_tramp = is_apple_tramp or ConfigureAppleTrampoline(dns);
        }

        if (auto maybe_addr = conf.query_bind; maybe_addr and not is_apple_tramp)
        {
          SockAddr addr{*maybe_addr};
          std::string host{addr.hostString()};

          if (addr.getPort() == 0)
          {
            // unbound manages their own sockets because of COURSE it does. so we find an open port
            // on our system and use it so we KNOW what it is before giving it to unbound to
            // explicitly bind to JUST that port.

            auto fd = socket(AF_INET, SOCK_DGRAM, IPPROTO_UDP);
#ifdef _WIN32
            if (fd == INVALID_SOCKET)
#else
            if (fd == -1)
#endif
            {
              throw std::invalid_argument{
                  fmt::format("Failed to create UDP socket for unbound: {}", strerror(errno))};
            }

#ifdef _WIN32
#define CLOSE closesocket
#else
#define CLOSE close
#endif
            if (0 != bind(fd, static_cast<const sockaddr*>(addr), addr.sockaddr_len()))
            {
              CLOSE(fd);
              throw std::invalid_argument{
                  fmt::format("Failed to bind UDP socket for unbound: {}", strerror(errno))};
            }
            struct sockaddr_storage sas;
            auto* sa = reinterpret_cast<struct sockaddr*>(&sas);
            socklen_t sa_len = sizeof(sas);
            int rc = getsockname(fd, sa, &sa_len);
            CLOSE(fd);
#undef CLOSE
            if (rc != 0)
            {
              throw std::invalid_argument{
                  fmt::format("Failed to query UDP port for unbound: {}", strerror(errno))};
            }
            addr = SockAddr{*sa};
          }
          m_LocalAddr = addr;

          log::info(logcat, "sending dns queries from {}:{}", host, addr.getPort());
          // set up query bind port if needed
          SetOpt("outgoing-interface:", host);
          SetOpt("outgoing-range:", "1");
          SetOpt("outgoing-port-avoid:", "0-65535");
          SetOpt("outgoing-port-permit:", "{}", addr.getPort());
        }
      }

      void
      SetOpt(const std::string& key, const std::string& val)
      {
        ub_ctx_set_option(m_ctx, key.c_str(), val.c_str());
      }

      // Wrapper around the above that takes 3+ arguments: the 2nd arg gets formatted with the
      // remaining args, and the formatted string passed to the above as `val`.
      template <typename... FmtArgs, std::enable_if_t<sizeof...(FmtArgs), int> = 0>
      void
      SetOpt(const std::string& key, std::string_view format, FmtArgs&&... args)
      {
        SetOpt(key, fmt::format(format, std::forward<FmtArgs>(args)...));
      }

      // Copy of the DNS config (a copy because on some platforms, like Apple, we change the applied
      // upstream DNS settings when turning on/off exit mode).
      llarp::DnsConfig m_conf;

     public:
      explicit Resolver(const EventLoop_ptr& loop, llarp::DnsConfig conf)
          : m_Loop{loop}, m_conf{std::move(conf)}
      {
        Up(m_conf);
      }

      ~Resolver() override
      {
        Down();
      }

      std::string_view
      ResolverName() const override
      {
        return "unbound";
      }

      virtual std::optional<SockAddr>
      GetLocalAddr() const override
      {
        return m_LocalAddr;
      }

      void
      RemovePending(const std::shared_ptr<Query>& query)
      {
        m_Pending.erase(query);
      }

      void
      Up(const llarp::DnsConfig& conf)
      {
        if (m_ctx)
          throw std::logic_error{"Internal error: attempt to Up() dns server multiple times"};

        m_ctx = ::ub_ctx_create();
        // set libunbound settings

        SetOpt("do-tcp:", "no");

        for (const auto& [k, v] : conf.extra_opts)
          SetOpt(k, v);

        // add host files
        for (const auto& file : conf.hostfiles)
        {
          const auto str = file.u8string();
          if (auto ret = ub_ctx_hosts(m_ctx, str.c_str()))
          {
            throw std::runtime_error{
                fmt::format("Failed to add host file {}: {}", file, ub_strerror(ret))};
          }
        }

        ConfigureUpstream(conf);

        // set async
        ub_ctx_async(m_ctx, 1);
        // setup mainloop
#ifdef _WIN32
        running = true;
        runner = std::thread{[this]() {
          while (running)
          {
            // poll and process callbacks it this thread
            if (ub_poll(m_ctx))
            {
              ub_process(m_ctx);
            }
            else  // nothing to do, sleep.
              std::this_thread::sleep_for(10ms);
          }
        }};
#else
        if (auto loop = m_Loop.lock())
        {
          if (auto loop_ptr = loop->MaybeGetUVWLoop())
          {
            m_Poller = loop_ptr->resource<uvw::PollHandle>(ub_fd(m_ctx));
            m_Poller->on<uvw::PollEvent>([this](auto&, auto&) { ub_process(m_ctx); });
            m_Poller->start(uvw::PollHandle::Event::READABLE);
            return;
          }
        }
        throw std::runtime_error{"no uvw loop"};
#endif
      }

      void
      Down() override
      {
#ifdef _WIN32
        if (running.exchange(false))
        {
          log::debug(logcat, "shutting down win32 dns thread");
          runner.join();
        }
#else
        if (m_Poller)
          m_Poller->close();
#endif
        if (m_ctx)
        {
          ::ub_ctx_delete(m_ctx);
          m_ctx = nullptr;

          // destroy any outstanding queries that unbound hasn't fired yet
          if (not m_Pending.empty())
          {
            log::debug(logcat, "cancelling {} pending queries", m_Pending.size());
            // We must copy because Cancel does a loop call to remove itself, but since we are
            // already in the main loop it happens immediately, which would invalidate our iterator
            // if we were looping through m_Pending at the time.
            auto copy = m_Pending;
            for (const auto& query : copy)
              query->Cancel();
          }
        }
      }

      int
      Rank() const override
      {
        return 10;
      }

      void
      ResetResolver(std::optional<std::vector<SockAddr>> replace_upstream) override
      {
        Down();
        if (replace_upstream)
          m_conf.upstream_dns = std::move(*replace_upstream);
        Up(m_conf);
      }

      template <typename Callable>
      void
      call(Callable&& f)
      {
        if (auto loop = m_Loop.lock())
          loop->call(std::forward<Callable>(f));
        else
          log::critical(logcat, "no mainloop?");
      }

      bool
      MaybeHookDNS(
          std::shared_ptr<PacketSource_Base> source,
          const Message& query,
          const SockAddr& to,
          const SockAddr& from) override
      {
        auto tmp = std::make_shared<Query>(weak_from_this(), query, source, to, from);
        // no questions, send fail
        if (query.questions.empty())
        {
          log::info(
              logcat,
              "dns from {} to {} has empty query questions, sending failure reply",
              from,
              to);
          tmp->Cancel();
          return true;
        }

        for (const auto& q : query.questions)
        {
          // dont process .loki or .snode
          if (q.HasTLD(".loki") or q.HasTLD(".snode"))
          {
            log::warning(
                logcat,
                "dns from {} to {} is for .loki or .snode but got to the unbound resolver, sending "
                "failure reply",
                from,
                to);
            tmp->Cancel();
            return true;
          }
        }
        if (not m_ctx)
        {
          // we are down
          log::debug(
              logcat,
              "dns from {} to {} got to the unbound resolver, but the resolver isn't set up, "
              "sending failure reply",
              from,
              to);
          tmp->Cancel();
          return true;
        }

#ifdef _WIN32
        if (not running)
        {
          // we are stopping the win32 thread
          log::debug(
              logcat,
              "dns from {} to {} got to the unbound resolver, but the resolver isn't running, "
              "sending failure reply",
              from,
              to);
          tmp->Cancel();
          return true;
        }
#endif
        const auto& q = query.questions[0];
        if (auto err = ub_resolve_async(
                m_ctx,
                q.Name().c_str(),
                q.qtype,
                q.qclass,
                tmp.get(),
                &Resolver::Callback,
                nullptr))
        {
          log::warning(
              logcat, "failed to send upstream query with libunbound: {}", ub_strerror(err));
          tmp->Cancel();
        }
        else
        {
          log::trace(logcat, "dns from {} to {} processing via libunbound", from, to);
          m_Pending.insert(std::move(tmp));
        }

        return true;
      }
    };

    void
    Query::SendReply(llarp::OwnedBuffer replyBuf)
    {
      if (m_Done.test_and_set())
        return;
      auto parent_ptr = parent.lock();
      if (parent_ptr)
      {
        parent_ptr->call(
            [self = shared_from_this(), parent_ptr = std::move(parent_ptr), buf = replyBuf.copy()] {
              log::trace(
                  logcat,
                  "forwarding dns response from libunbound to userland (resolverAddr: {}, "
                  "askerAddr: {})",
                  self->resolverAddr,
                  self->askerAddr);
              self->src->SendTo(self->askerAddr, self->resolverAddr, OwnedBuffer::copy_from(buf));
              // remove query
              parent_ptr->RemovePending(self);
            });
      }
      else
        log::error(logcat, "no parent");
    }
  }  // namespace libunbound

  Server::Server(EventLoop_ptr loop, llarp::DnsConfig conf, unsigned int netif)
      : m_Loop{std::move(loop)}
      , m_Config{std::move(conf)}
      , m_Platform{CreatePlatform()}
      , m_NetIfIndex{std::move(netif)}
  {}

  std::vector<std::weak_ptr<Resolver_Base>>
  Server::GetAllResolvers() const
  {
    return {m_Resolvers.begin(), m_Resolvers.end()};
  }

  void
  Server::Start()
  {
    // set up udp sockets
    for (const auto& addr : m_Config.bind_addr)
    {
      if (auto ptr = MakePacketSourceOn(addr, m_Config))
        AddPacketSource(std::move(ptr));
    }

    // add default resolver as needed
    if (auto ptr = MakeDefaultResolver())
      AddResolver(ptr);
  }

  std::shared_ptr<I_Platform>
  Server::CreatePlatform() const
  {
    auto plat = std::make_shared<Multi_Platform>();
    if constexpr (llarp::platform::has_systemd)
    {
      plat->add_impl(std::make_unique<SD_Platform_t>());
      plat->add_impl(std::make_unique<NM_Platform_t>());
    }
    return plat;
  }

  std::shared_ptr<PacketSource_Base>
  Server::MakePacketSourceOn(const llarp::SockAddr& addr, const llarp::DnsConfig&)
  {
    return std::make_shared<UDPReader>(*this, m_Loop, addr);
  }

  std::shared_ptr<Resolver_Base>
  Server::MakeDefaultResolver()
  {
    if (m_Config.upstream_dns.empty())
    {
      log::info(
          logcat,
          "explicitly no upstream dns providers specified, we will not resolve anything but .loki "
          "and .snode");
      return nullptr;
    }

    return std::make_shared<libunbound::Resolver>(m_Loop, m_Config);
  }

  std::vector<SockAddr>
  Server::BoundPacketSourceAddrs() const
  {
    std::vector<SockAddr> addrs;
    for (const auto& src : m_PacketSources)
    {
      if (auto ptr = src.lock())
        if (auto maybe_addr = ptr->BoundOn())
          addrs.emplace_back(*maybe_addr);
    }
    return addrs;
  }

  std::optional<SockAddr>
  Server::FirstBoundPacketSourceAddr() const
  {
    for (const auto& src : m_PacketSources)
    {
      if (auto ptr = src.lock())
        if (auto bound = ptr->BoundOn())
          return bound;
    }
    return std::nullopt;
  }

  void
  Server::AddResolver(std::weak_ptr<Resolver_Base> resolver)
  {
    m_Resolvers.insert(resolver);
  }

  void
  Server::AddResolver(std::shared_ptr<Resolver_Base> resolver)
  {
    m_OwnedResolvers.insert(resolver);
    AddResolver(std::weak_ptr<Resolver_Base>{resolver});
  }

  void
  Server::AddPacketSource(std::weak_ptr<PacketSource_Base> pkt)
  {
    m_PacketSources.push_back(pkt);
  }

  void
  Server::AddPacketSource(std::shared_ptr<PacketSource_Base> pkt)
  {
    AddPacketSource(std::weak_ptr<PacketSource_Base>{pkt});
    m_OwnedPacketSources.push_back(std::move(pkt));
  }

  void
  Server::Stop()
  {
    for (const auto& resolver : m_Resolvers)
    {
      if (auto ptr = resolver.lock())
        ptr->Down();
    }
  }

  void
  Server::Reset()
  {
    for (const auto& resolver : m_Resolvers)
    {
      if (auto ptr = resolver.lock())
        ptr->ResetResolver();
    }
  }

  void
  Server::SetDNSMode(bool all_queries)
  {
    if (auto maybe_addr = FirstBoundPacketSourceAddr())
      m_Platform->set_resolver(m_NetIfIndex, *maybe_addr, all_queries);
  }

  bool
  Server::MaybeHandlePacket(
      std::shared_ptr<PacketSource_Base> ptr,
      const SockAddr& to,
      const SockAddr& from,
      llarp::OwnedBuffer buf)
  {
    // dont process to prevent feedback loop
    if (ptr->WouldLoop(to, from))
    {
      log::warning(logcat, "preventing dns packet replay to={} from={}", to, from);
      return false;
    }

    auto maybe = MaybeParseDNSMessage(buf);
    if (not maybe)
    {
      log::warning(logcat, "invalid dns message format from {} to dns listener on {}", from, to);
      return false;
    }

    auto& msg = *maybe;
    // we don't provide a DoH resolver because it requires verified TLS
    // TLS needs X509/ASN.1-DER and opting into the Root CA Cabal
    // thankfully mozilla added a backdoor that allows ISPs to turn it off
    // so we disable DoH for firefox using mozilla's ISP backdoor
    // see: https://github.com/oxen-io/lokinet/issues/832
    for (const auto& q : msg.questions)
    {
      // is this firefox looking for their backdoor record?
      if (q.IsName("use-application-dns.net"))
      {
        // yea it is, let's turn off DoH because god is dead.
        msg.AddNXReply();
        // press F to pay respects and send it back where it came from
        ptr->SendTo(from, to, msg.ToBuffer());
        return true;
      }
    }

    for (const auto& resolver : m_Resolvers)
    {
      if (auto res_ptr = resolver.lock())
      {
        log::trace(
            logcat, "check resolver {} for dns from {} to {}", res_ptr->ResolverName(), from, to);
        if (res_ptr->MaybeHookDNS(ptr, msg, to, from))
        {
          log::trace(
              logcat, "resolver {} handling dns from {} to {}", res_ptr->ResolverName(), from, to);
          return true;
        }
      }
    }
    return false;
  }

}  // namespace llarp::dns