lokinet/llarp/ev.hpp

#ifndef LLARP_EV_HPP
#define LLARP_EV_HPP

#include <ev.h>
#include <codel.hpp>
#include <threading.hpp>

// writev
#ifndef _WIN32
#include <sys/uio.h>
#endif

#include <algorithm>
#include <buffer.h>
#include <deque>
#include <list>
#include <unistd.h>

#ifdef _WIN32
#include <win32_up.h>
#include <win32_upoll.h>
// io packet for TUN read/write
struct asio_evt_pkt
{
  OVERLAPPED pkt = {
      0, 0, 0, 0, nullptr};  // must be first, since this is part of the IO call
  bool write = false;        // true, or false if read pkt
  size_t sz;  // if this doesn't match what is in the packet, note the error
};
#else

#if defined(__FreeBSD__) || defined(__OpenBSD__) || defined(__NetBSD__) \
    || (__APPLE__ && __MACH__)
#include <sys/event.h>
#endif

#include <sys/un.h>
#endif

#ifndef MAX_WRITE_QUEUE_SIZE
#define MAX_WRITE_QUEUE_SIZE (1024UL)
#endif

#ifndef EV_READ_BUF_SZ
#define EV_READ_BUF_SZ (4 * 1024UL)
#endif
#ifndef EV_WRITE_BUF_SZ
#define EV_WRITE_BUF_SZ (2 * 1024UL)
#endif

namespace llarp
{
#ifdef _WIN32
  struct win32_ev_io
  {
    struct WriteBuffer
    {
      llarp_time_t timestamp = 0;
      size_t bufsz;
      byte_t buf[EV_WRITE_BUF_SZ] = {0};

      WriteBuffer() = default;

      WriteBuffer(const byte_t* ptr, size_t sz)
      {
        if(sz <= sizeof(buf))
        {
          bufsz = sz;
          memcpy(buf, ptr, bufsz);
        }
        else
          bufsz = 0;
      }

      struct GetTime
      {
        llarp_time_t
        operator()(const WriteBuffer& buf) const
        {
          return buf.timestamp;
        }
      };

      struct GetNow
      {
        llarp_ev_loop* loop;
        GetNow(llarp_ev_loop* l) : loop(l)
        {
        }

        llarp_time_t
        operator()() const
        {
          return llarp_ev_loop_time_now_ms(loop);
        }
      };

      struct PutTime
      {
        llarp_ev_loop* loop;
        PutTime(llarp_ev_loop* l) : loop(l)
        {
        }
        void
        operator()(WriteBuffer& buf)
        {
          buf.timestamp = llarp_ev_loop_time_now_ms(loop);
        }
      };

      struct Compare
      {
        bool
        operator()(const WriteBuffer& left, const WriteBuffer& right) const
        {
          return left.timestamp < right.timestamp;
        }
      };
    };

    using LossyWriteQueue_t =
        llarp::util::CoDelQueue< WriteBuffer, WriteBuffer::GetTime,
                                 WriteBuffer::PutTime, WriteBuffer::Compare,
                                 WriteBuffer::GetNow, llarp::util::NullMutex,
                                 llarp::util::NullLock, 5, 100, 128 >;

    using LosslessWriteQueue_t = std::deque< WriteBuffer >;

    union {
      intptr_t socket;
      HANDLE tun;
    } fd;

    int flags   = 0;
    bool is_tun = false;

    win32_ev_io(intptr_t f)
    {
      fd.socket = f;
    }

    /// for tun
    win32_ev_io(HANDLE f, LossyWriteQueue_t* q) : m_LossyWriteQueue(q)
    {
      fd.tun = f;
    }

    /// for tcp
    win32_ev_io(intptr_t f, LosslessWriteQueue_t* q) : m_BlockingWriteQueue(q)
    {
      fd.socket = f;
    }

    virtual void
    error()
    {
      llarp::LogError(strerror(errno));
    }

    virtual int
    read(byte_t* buf, size_t sz) = 0;

    virtual int
    sendto(const sockaddr* dst, const void* data, size_t sz)
    {
      (void)(dst);
      (void)(data);
      (void)(sz);
      return -1;
    };

    /// return false if we want to deregister and remove ourselves
    virtual bool
    tick()
    {
      return true;
    };

    /// used for tun interface and tcp conn
    virtual ssize_t
    do_write(void* data, size_t sz)
    {
      if(this->is_tun)
      {
        DWORD x;
        bool r;
        asio_evt_pkt* pkt = new asio_evt_pkt;
        pkt->sz           = sz;
        pkt->write        = true;
        int e             = 0;
        r                 = WriteFile(fd.tun, data, sz, &x, &pkt->pkt);
        if(r)  // we returned immediately
          return x;
        e = GetLastError();
        if(e == ERROR_IO_PENDING)
          return sz;
        else
          return -1;
      }
      return uwrite(fd.socket, (char*)data, sz);
    }

    bool
    queue_write(const byte_t* buf, size_t sz)
    {
      if(m_LossyWriteQueue)
      {
        m_LossyWriteQueue->Emplace(buf, sz);
        return true;
      }
      else if(m_BlockingWriteQueue)
      {
        m_BlockingWriteQueue->emplace_back(buf, sz);
        return true;
      }
      else
        return false;
    }

    virtual void
    flush_write()
    {
      flush_write_buffers(0);
    }

    /// called in event loop when fd is ready for writing
    /// requeues anything not written
    /// this assumes fd is set to non blocking
    virtual void
    flush_write_buffers(size_t amount)
    {
      if(m_LossyWriteQueue)
        m_LossyWriteQueue->Process([&](WriteBuffer& buffer) {
          do_write(buffer.buf, buffer.bufsz);
          // if we would block we save the entries for later
          // discard entry
        });
      else if(m_BlockingWriteQueue)
      {
        if(amount)
        {
          while(amount && m_BlockingWriteQueue->size())
          {
            auto& itr      = m_BlockingWriteQueue->front();
            ssize_t result = do_write(itr.buf, std::min(amount, itr.bufsz));
            if(result == -1)
              return;
            ssize_t dlt = itr.bufsz - result;
            if(dlt > 0)
            {
              // queue remaining to front of queue
              WriteBuffer buff(itr.buf + dlt, itr.bufsz - dlt);
              m_BlockingWriteQueue->pop_front();
              m_BlockingWriteQueue->push_front(buff);
              // TODO: errno?
              return;
            }
            m_BlockingWriteQueue->pop_front();
            amount -= result;
          }
        }
        else
        {
          // write buffers
          while(m_BlockingWriteQueue->size())
          {
            auto& itr      = m_BlockingWriteQueue->front();
            ssize_t result = do_write(itr.buf, itr.bufsz);
            if(result == -1)
              return;
            ssize_t dlt = itr.bufsz - result;
            if(dlt > 0)
            {
              // queue remaining to front of queue
              WriteBuffer buff(itr.buf + dlt, itr.bufsz - dlt);
              m_BlockingWriteQueue->pop_front();
              m_BlockingWriteQueue->push_front(buff);
              // TODO: errno?
              return;
            }
            m_BlockingWriteQueue->pop_front();
            int wsaerr = WSAGetLastError();
            int syserr = GetLastError();
            if(wsaerr == WSA_IO_PENDING || wsaerr == WSAEWOULDBLOCK
               || syserr == 997 || syserr == 21)
            {
              SetLastError(0);
              WSASetLastError(0);
              return;
            }
          }
        }
      }
      /// reset errno
      SetLastError(0);
      WSASetLastError(0);
    }

    std::unique_ptr< LossyWriteQueue_t > m_LossyWriteQueue;
    std::unique_ptr< LosslessWriteQueue_t > m_BlockingWriteQueue;

    virtual ~win32_ev_io()
    {
      uclose(fd.socket);
    };
  };
#endif

  struct posix_ev_io
  {
    struct WriteBuffer
    {
      llarp_time_t timestamp = 0;
      size_t bufsz;
      byte_t buf[EV_WRITE_BUF_SZ];

      WriteBuffer() = default;

      WriteBuffer(const byte_t* ptr, size_t sz)
      {
        if(sz <= sizeof(buf))
        {
          bufsz = sz;
          memcpy(buf, ptr, bufsz);
        }
        else
          bufsz = 0;
      }

      struct GetTime
      {
        llarp_time_t
        operator()(const WriteBuffer& buf) const
        {
          return buf.timestamp;
        }
      };

      struct GetNow
      {
        llarp_ev_loop* loop;
        GetNow(llarp_ev_loop* l) : loop(l)
        {
        }

        llarp_time_t
        operator()() const
        {
          return llarp_ev_loop_time_now_ms(loop);
        }
      };

      struct PutTime
      {
        llarp_ev_loop* loop;
        PutTime(llarp_ev_loop* l) : loop(l)
        {
        }
        void
        operator()(WriteBuffer& buf)
        {
          buf.timestamp = llarp_ev_loop_time_now_ms(loop);
        }
      };

      struct Compare
      {
        bool
        operator()(const WriteBuffer& left, const WriteBuffer& right) const
        {
          return left.timestamp < right.timestamp;
        }
      };
    };

    using LossyWriteQueue_t =
        llarp::util::CoDelQueue< WriteBuffer, WriteBuffer::GetTime,
                                 WriteBuffer::PutTime, WriteBuffer::Compare,
                                 WriteBuffer::GetNow, llarp::util::NullMutex,
                                 llarp::util::NullLock, 5, 100, 1024 >;

    using LosslessWriteQueue_t = std::deque< WriteBuffer >;

    int fd;
    int flags = 0;
#if defined(__FreeBSD__) || defined(__OpenBSD__) || defined(__NetBSD__) \
    || (__APPLE__ && __MACH__)
    struct kevent change;
#endif

    posix_ev_io(int f) : fd(f)
    {
    }

    /// for tun
    posix_ev_io(int f, LossyWriteQueue_t* q) : fd(f), m_LossyWriteQueue(q)
    {
    }

    /// for tcp
    posix_ev_io(int f, LosslessWriteQueue_t* q) : fd(f), m_BlockingWriteQueue(q)
    {
    }

    virtual void
    error()
    {
      llarp::LogError(strerror(errno));
    }

    virtual int
    read(byte_t* buf, size_t sz) = 0;

    virtual int
    sendto(__attribute__((unused)) const sockaddr* dst,
           __attribute__((unused)) const void* data,
           __attribute__((unused)) size_t sz)
    {
      return -1;
    };

    /// return false if we want to deregister and remove ourselves
    virtual bool
    tick()
    {
      return true;
    };

    /// used for tun interface and tcp conn
    virtual ssize_t
    do_write(void* data, size_t sz)
    {
      return write(fd, data, sz);
    }

    bool
    queue_write(const byte_t* buf, size_t sz)
    {
      if(m_LossyWriteQueue)
      {
        m_LossyWriteQueue->Emplace(buf, sz);
        return true;
      }
      else if(m_BlockingWriteQueue)
      {
        m_BlockingWriteQueue->emplace_back(buf, sz);
        return true;
      }
      else
        return false;
    }

    virtual void
    flush_write()
    {
      flush_write_buffers(0);
    }

    virtual void
    before_flush_write(){};

    /// called in event loop when fd is ready for writing
    /// requeues anything not written
    /// this assumes fd is set to non blocking
    virtual void
    flush_write_buffers(size_t amount)
    {
      before_flush_write();
      if(m_LossyWriteQueue)
      {
        m_LossyWriteQueue->Process([&](WriteBuffer& buffer) {
          do_write(buffer.buf, buffer.bufsz);
          // if we would block we save the entries for later
          // discard entry
        });
      }
      else if(m_BlockingWriteQueue)
      {
        if(amount)
        {
          while(amount && m_BlockingWriteQueue->size())
          {
            auto& itr      = m_BlockingWriteQueue->front();
            ssize_t result = do_write(itr.buf, std::min(amount, itr.bufsz));
            if(result == -1)
              return;
            ssize_t dlt = itr.bufsz - result;
            if(dlt > 0)
            {
              // queue remaining to front of queue
              WriteBuffer buff(itr.buf + dlt, itr.bufsz - dlt);
              m_BlockingWriteQueue->pop_front();
              m_BlockingWriteQueue->push_front(buff);
              // TODO: errno?
              return;
            }
            m_BlockingWriteQueue->pop_front();
            amount -= result;
          }
        }
        else
        {
          // write buffers
          while(m_BlockingWriteQueue->size())
          {
            auto& itr      = m_BlockingWriteQueue->front();
            ssize_t result = do_write(itr.buf, itr.bufsz);
            if(result == -1)
              return;
            ssize_t dlt = itr.bufsz - result;
            if(dlt > 0)
            {
              // queue remaining to front of queue
              WriteBuffer buff(itr.buf + dlt, itr.bufsz - dlt);
              m_BlockingWriteQueue->pop_front();
              m_BlockingWriteQueue->push_front(buff);
              // TODO: errno?
              return;
            }
            m_BlockingWriteQueue->pop_front();
            if(errno == EAGAIN || errno == EWOULDBLOCK)
            {
              errno = 0;
              return;
            }
          }
        }
      }
      /// reset errno
      errno = 0;
    }

    std::unique_ptr< LossyWriteQueue_t > m_LossyWriteQueue;
    std::unique_ptr< LosslessWriteQueue_t > m_BlockingWriteQueue;

    virtual ~posix_ev_io()
    {
      close(fd);
    };
  };

// finally create aliases by platform
#ifdef _WIN32
  using ev_io = win32_ev_io;

#define UNIX_PATH_MAX 108

  typedef struct sockaddr_un
  {
    ADDRESS_FAMILY sun_family;    /* AF_UNIX */
    char sun_path[UNIX_PATH_MAX]; /* pathname */
  } SOCKADDR_UN, *PSOCKADDR_UN;

#else
  using ev_io = posix_ev_io;
#endif

  // wew, managed to get away with using
  // 'int fd' across all platforms
  // since we're operating entirely
  // on sockets
  struct tcp_conn : public ev_io
  {
    sockaddr_storage _addr;
    bool _shouldClose     = false;
    bool _calledConnected = false;
    llarp_tcp_conn tcp;
    // null if inbound otherwise outbound
    llarp_tcp_connecter* _conn;

    /// inbound
    tcp_conn(llarp_ev_loop* loop, int fd)
        : ev_io(fd, new LosslessWriteQueue_t{}), _conn(nullptr)
    {
      tcp.impl   = this;
      tcp.loop   = loop;
      tcp.closed = nullptr;
      tcp.user   = nullptr;
      tcp.read   = nullptr;
      tcp.tick   = nullptr;
    }

    /// outbound
    tcp_conn(llarp_ev_loop* loop, int fd, const sockaddr* addr,
             llarp_tcp_connecter* conn)
        : ev_io(fd, new LosslessWriteQueue_t{}), _conn(conn)
    {
      socklen_t slen = sizeof(sockaddr_in);
      if(addr->sa_family == AF_INET6)
        slen = sizeof(sockaddr_in6);
      else if(addr->sa_family == AF_UNIX)
        slen = sizeof(sockaddr_un);
      memcpy(&_addr, addr, slen);
      tcp.impl   = this;
      tcp.loop   = loop;
      tcp.closed = nullptr;
      tcp.user   = nullptr;
      tcp.read   = nullptr;
      tcp.tick   = nullptr;
    }

    virtual ~tcp_conn()
    {
    }

    /// start connecting
    void
    connect();

    /// calls connected hooks
    void
    connected()
    {
      // we are connected yeh boi
      if(_conn)
      {
        if(_conn->connected && !_calledConnected)
          _conn->connected(_conn, &tcp);
      }
      _calledConnected = true;
    }

    void
    flush_write();

    void
    flush_write_buffers(size_t a)
    {
      connected();
      ev_io::flush_write_buffers(a);
    }

    void
    error()
    {
      if(_conn)
      {
#ifndef _WIN32
        llarp::LogError("tcp_conn error: ", strerror(errno));
#else
        char ebuf[1024];
        int err = WSAGetLastError();
        FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM, nullptr, err, LANG_NEUTRAL,
                      ebuf, 1024, nullptr);
        llarp::LogError("tcp_conn error: ", ebuf);
#endif
        if(_conn->error)
          _conn->error(_conn);
      }
    }

    virtual ssize_t
    do_write(void* buf, size_t sz);

    virtual int
    read(byte_t* buf, size_t sz);

    bool
    tick();
  };

  struct tcp_serv : public ev_io
  {
    llarp_ev_loop* loop;
    llarp_tcp_acceptor* tcp;
    tcp_serv(llarp_ev_loop* l, int fd, llarp_tcp_acceptor* t)
        : ev_io(fd), loop(l), tcp(t)
    {
      tcp->impl = this;
    }

    bool
    tick()
    {
      if(tcp->tick)
        tcp->tick(tcp);
      return true;
    }

    /// actually does accept() :^)
    virtual int
    read(byte_t*, size_t);
  };

};  // namespace llarp

#ifdef _WIN32
struct llarp_fd_promise
{
  void
  Set(int)
  {
  }

  int
  Get()
  {
    return -1;
  }
};
#else
struct llarp_fd_promise
{
  llarp_fd_promise(std::promise< int >* p) : _impl(p)
  {
  }
  std::promise< int >* _impl;

  void
  Set(int fd)
  {
    _impl->set_value(fd);
  }

  int
  Get()
  {
    auto future = _impl->get_future();
    future.wait();
    return future.get();
  }
};
#endif

// this (nearly!) abstract base class
// is overriden for each platform
struct llarp_ev_loop
{
  byte_t readbuf[EV_READ_BUF_SZ] = {0};
  llarp_time_t _now              = 0;
  virtual bool
  init() = 0;
  virtual int
  run() = 0;

  virtual int
  tick(int ms) = 0;

  virtual void
  stop() = 0;

  virtual bool
  udp_listen(llarp_udp_io* l, const sockaddr* src) = 0;

  virtual llarp::ev_io*
  create_udp(llarp_udp_io* l, const sockaddr* src) = 0;

  virtual bool
  udp_close(llarp_udp_io* l) = 0;
  /// deregister event listener
  virtual bool
  close_ev(llarp::ev_io* ev) = 0;

  virtual llarp::ev_io*
  create_tun(llarp_tun_io* tun) = 0;

  virtual llarp::ev_io*
  bind_tcp(llarp_tcp_acceptor* tcp, const sockaddr* addr) = 0;

  /// return false on socket error (non blocking)
  virtual bool
  tcp_connect(llarp_tcp_connecter* tcp, const sockaddr* addr) = 0;

  /// register event listener
  virtual bool
  add_ev(llarp::ev_io* ev, bool write) = 0;

  virtual bool
  running() const = 0;

  virtual ~llarp_ev_loop(){};

  std::list< std::unique_ptr< llarp::ev_io > > handlers;

  void
  tick_listeners()
  {
    auto itr = handlers.begin();
    while(itr != handlers.end())
    {
      if((*itr)->tick())
        ++itr;
      else
      {
        close_ev(itr->get());
        itr = handlers.erase(itr);
      }
    }
  }
};

#endif