base/physicalsocketserver.cc

/*
 * libjingle
 * Copyright 2004--2005, Google Inc.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are met:
 *
 *  1. Redistributions of source code must retain the above copyright notice,
 *     this list of conditions and the following disclaimer.
 *  2. Redistributions in binary form must reproduce the above copyright notice,
 *     this list of conditions and the following disclaimer in the documentation
 *     and/or other materials provided with the distribution.
 *  3. The name of the author may not be used to endorse or promote products
 *     derived from this software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO
 * EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
 * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
 * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */

#if defined(_MSC_VER) && _MSC_VER < 1300
#pragma warning(disable:4786)
#endif

#include <assert.h>

#ifdef POSIX
#include <string.h>
#include <errno.h>
#include <fcntl.h>
#include <sys/time.h>
#include <sys/select.h>
#include <unistd.h>
#include <signal.h>
#endif

#ifdef WIN32
#define WIN32_LEAN_AND_MEAN
#include <windows.h>
#include <winsock2.h>
#include <ws2tcpip.h>
#undef SetPort
#endif

#include <algorithm>
#include <map>

#include "talk/base/basictypes.h"
#include "talk/base/byteorder.h"
#include "talk/base/common.h"
#include "talk/base/logging.h"
#include "talk/base/nethelpers.h"
#include "talk/base/physicalsocketserver.h"
#include "talk/base/timeutils.h"
#include "talk/base/winping.h"
#include "talk/base/win32socketinit.h"

// stm: this will tell us if we are on OSX
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif

#ifdef POSIX
#include <netinet/tcp.h>  // for TCP_NODELAY
#define IP_MTU 14 // Until this is integrated from linux/in.h to netinet/in.h
typedef void* SockOptArg;
#endif  // POSIX

#ifdef WIN32
typedef char* SockOptArg;
#endif

namespace talk_base {

#if defined(WIN32)
// Standard MTUs, from RFC 1191
const uint16 PACKET_MAXIMUMS[] = {
  65535,    // Theoretical maximum, Hyperchannel
  32000,    // Nothing
  17914,    // 16Mb IBM Token Ring
  8166,     // IEEE 802.4
  //4464,   // IEEE 802.5 (4Mb max)
  4352,     // FDDI
  //2048,   // Wideband Network
  2002,     // IEEE 802.5 (4Mb recommended)
  //1536,   // Expermental Ethernet Networks
  //1500,   // Ethernet, Point-to-Point (default)
  1492,     // IEEE 802.3
  1006,     // SLIP, ARPANET
  //576,    // X.25 Networks
  //544,    // DEC IP Portal
  //512,    // NETBIOS
  508,      // IEEE 802/Source-Rt Bridge, ARCNET
  296,      // Point-to-Point (low delay)
  68,       // Official minimum
  0,        // End of list marker
};

static const int IP_HEADER_SIZE = 20u;
static const int IPV6_HEADER_SIZE = 40u;
static const int ICMP_HEADER_SIZE = 8u;
static const int ICMP_PING_TIMEOUT_MILLIS = 10000u;
#endif

class PhysicalSocket : public AsyncSocket, public sigslot::has_slots<> {
 public:
  PhysicalSocket(PhysicalSocketServer* ss, SOCKET s = INVALID_SOCKET)
    : ss_(ss), s_(s), enabled_events_(0), error_(0),
      state_((s == INVALID_SOCKET) ? CS_CLOSED : CS_CONNECTED),
      resolver_(NULL) {
#ifdef WIN32
    // EnsureWinsockInit() ensures that winsock is initialized. The default
    // version of this function doesn't do anything because winsock is
    // initialized by constructor of a static object. If neccessary libjingle
    // users can link it with a different version of this function by replacing
    // win32socketinit.cc. See win32socketinit.cc for more details.
    EnsureWinsockInit();
#endif
    if (s_ != INVALID_SOCKET) {
      enabled_events_ = DE_READ | DE_WRITE;

      int type = SOCK_STREAM;
      socklen_t len = sizeof(type);
      VERIFY(0 == getsockopt(s_, SOL_SOCKET, SO_TYPE, (SockOptArg)&type, &len));
      udp_ = (SOCK_DGRAM == type);
    }
  }

  virtual ~PhysicalSocket() {
    Close();
  }

  // Creates the underlying OS socket (same as the "socket" function).
  virtual bool Create(int family, int type) {
    Close();
    s_ = ::socket(family, type, 0);
    udp_ = (SOCK_DGRAM == type);
    UpdateLastError();
    if (udp_)
      enabled_events_ = DE_READ | DE_WRITE;
    return s_ != INVALID_SOCKET;
  }

  SocketAddress GetLocalAddress() const {
    sockaddr_storage addr_storage = {0};
    socklen_t addrlen = sizeof(addr_storage);
    sockaddr* addr = reinterpret_cast<sockaddr*>(&addr_storage);
    int result = ::getsockname(s_, addr, &addrlen);
    SocketAddress address;
    if (result >= 0) {
      SocketAddressFromSockAddrStorage(addr_storage, &address);
    } else {
      LOG(LS_WARNING) << "GetLocalAddress: unable to get local addr, socket="
                      << s_;
    }
    return address;
  }

  SocketAddress GetRemoteAddress() const {
    sockaddr_storage addr_storage = {0};
    socklen_t addrlen = sizeof(addr_storage);
    sockaddr* addr = reinterpret_cast<sockaddr*>(&addr_storage);
    int result = ::getpeername(s_, addr, &addrlen);
    SocketAddress address;
    if (result >= 0) {
      SocketAddressFromSockAddrStorage(addr_storage, &address);
    } else {
      LOG(LS_WARNING) << "GetRemoteAddress: unable to get remote addr, socket="
                      << s_;
    }
    return address;
  }

  int Bind(const SocketAddress& bind_addr) {
    sockaddr_storage addr_storage;
    size_t len = bind_addr.ToSockAddrStorage(&addr_storage);
    sockaddr* addr = reinterpret_cast<sockaddr*>(&addr_storage);
    int err = ::bind(s_, addr, static_cast<int>(len));
    UpdateLastError();
#ifdef _DEBUG
    if (0 == err) {
      dbg_addr_ = "Bound @ ";
      dbg_addr_.append(GetLocalAddress().ToString());
    }
#endif  // _DEBUG
    return err;
  }

  int Connect(const SocketAddress& addr) {
    // TODO: Implicit creation is required to reconnect...
    // ...but should we make it more explicit?
    if (state_ != CS_CLOSED) {
      SetError(EALREADY);
      return SOCKET_ERROR;
    }
    if (addr.IsUnresolved()) {
      LOG(LS_VERBOSE) << "Resolving addr in PhysicalSocket::Connect";
      resolver_ = new AsyncResolver();
      resolver_->SignalDone.connect(this, &PhysicalSocket::OnResolveResult);
      resolver_->Start(addr);
      state_ = CS_CONNECTING;
      return 0;
    }

    return DoConnect(addr);
  }

  int DoConnect(const SocketAddress& connect_addr) {
    if ((s_ == INVALID_SOCKET) &&
        !Create(connect_addr.family(), SOCK_STREAM)) {
      return SOCKET_ERROR;
    }
    sockaddr_storage addr_storage;
    size_t len = connect_addr.ToSockAddrStorage(&addr_storage);
    sockaddr* addr = reinterpret_cast<sockaddr*>(&addr_storage);
    int err = ::connect(s_, addr, static_cast<int>(len));
    UpdateLastError();
    if (err == 0) {
      state_ = CS_CONNECTED;
    } else if (IsBlockingError(GetError())) {
      state_ = CS_CONNECTING;
      enabled_events_ |= DE_CONNECT;
    } else {
      return SOCKET_ERROR;
    }

    enabled_events_ |= DE_READ | DE_WRITE;
    return 0;
  }

  int GetError() const {
    CritScope cs(&crit_);
    return error_;
  }

  void SetError(int error) {
    CritScope cs(&crit_);
    error_ = error;
  }

  ConnState GetState() const {
    return state_;
  }

  int GetOption(Option opt, int* value) {
    int slevel;
    int sopt;
    if (TranslateOption(opt, &slevel, &sopt) == -1)
      return -1;
    socklen_t optlen = sizeof(*value);
    int ret = ::getsockopt(s_, slevel, sopt, (SockOptArg)value, &optlen);
    if (ret != -1 && opt == OPT_DONTFRAGMENT) {
#ifdef LINUX
      *value = (*value != IP_PMTUDISC_DONT) ? 1 : 0;
#endif
    }
    return ret;
  }

  int SetOption(Option opt, int value) {
    int slevel;
    int sopt;
    if (TranslateOption(opt, &slevel, &sopt) == -1)
      return -1;
    if (opt == OPT_DONTFRAGMENT) {
#ifdef LINUX
      value = (value) ? IP_PMTUDISC_DO : IP_PMTUDISC_DONT;
#endif
    }
    return ::setsockopt(s_, slevel, sopt, (SockOptArg)&value, sizeof(value));
  }

  int Send(const void *pv, size_t cb) {
    int sent = ::send(s_, reinterpret_cast<const char *>(pv), (int)cb,
#ifdef LINUX
        // Suppress SIGPIPE. Without this, attempting to send on a socket whose
        // other end is closed will result in a SIGPIPE signal being raised to
        // our process, which by default will terminate the process, which we
        // don't want. By specifying this flag, we'll just get the error EPIPE
        // instead and can handle the error gracefully.
        MSG_NOSIGNAL
#else
        0
#endif
        );
    UpdateLastError();
    MaybeRemapSendError();
    // We have seen minidumps where this may be false.
    ASSERT(sent <= static_cast<int>(cb));
    if ((sent < 0) && IsBlockingError(GetError())) {
      enabled_events_ |= DE_WRITE;
    }
    return sent;
  }

  int SendTo(const void* buffer, size_t length, const SocketAddress& addr) {
    sockaddr_storage saddr;
    size_t len = addr.ToSockAddrStorage(&saddr);
    int sent = ::sendto(
        s_, static_cast<const char *>(buffer), static_cast<int>(length),
#ifdef LINUX
        // Suppress SIGPIPE. See above for explanation.
        MSG_NOSIGNAL,
#else
        0,
#endif
        reinterpret_cast<sockaddr*>(&saddr), static_cast<int>(len));
    UpdateLastError();
    MaybeRemapSendError();
    // We have seen minidumps where this may be false.
    ASSERT(sent <= static_cast<int>(length));
    if ((sent < 0) && IsBlockingError(GetError())) {
      enabled_events_ |= DE_WRITE;
    }
    return sent;
  }

  int Recv(void* buffer, size_t length) {
    int received = ::recv(s_, static_cast<char*>(buffer),
                          static_cast<int>(length), 0);
    if ((received == 0) && (length != 0)) {
      // Note: on graceful shutdown, recv can return 0.  In this case, we
      // pretend it is blocking, and then signal close, so that simplifying
      // assumptions can be made about Recv.
      LOG(LS_WARNING) << "EOF from socket; deferring close event";
      // Must turn this back on so that the select() loop will notice the close
      // event.
      enabled_events_ |= DE_READ;
      SetError(EWOULDBLOCK);
      return SOCKET_ERROR;
    }
    UpdateLastError();
    int error = GetError();
    bool success = (received >= 0) || IsBlockingError(error);
    if (udp_ || success) {
      enabled_events_ |= DE_READ;
    }
    if (!success) {
      LOG_F(LS_VERBOSE) << "Error = " << error;
    }
    return received;
  }

  int RecvFrom(void* buffer, size_t length, SocketAddress *out_addr) {
    sockaddr_storage addr_storage;
    socklen_t addr_len = sizeof(addr_storage);
    sockaddr* addr = reinterpret_cast<sockaddr*>(&addr_storage);
    int received = ::recvfrom(s_, static_cast<char*>(buffer),
                              static_cast<int>(length), 0, addr, &addr_len);
    UpdateLastError();
    if ((received >= 0) && (out_addr != NULL))
      SocketAddressFromSockAddrStorage(addr_storage, out_addr);
    int error = GetError();
    bool success = (received >= 0) || IsBlockingError(error);
    if (udp_ || success) {
      enabled_events_ |= DE_READ;
    }
    if (!success) {
      LOG_F(LS_VERBOSE) << "Error = " << error;
    }
    return received;
  }

  int Listen(int backlog) {
    int err = ::listen(s_, backlog);
    UpdateLastError();
    if (err == 0) {
      state_ = CS_CONNECTING;
      enabled_events_ |= DE_ACCEPT;
#ifdef _DEBUG
      dbg_addr_ = "Listening @ ";
      dbg_addr_.append(GetLocalAddress().ToString());
#endif  // _DEBUG
    }
    return err;
  }

  AsyncSocket* Accept(SocketAddress *out_addr) {
    sockaddr_storage addr_storage;
    socklen_t addr_len = sizeof(addr_storage);
    sockaddr* addr = reinterpret_cast<sockaddr*>(&addr_storage);
    SOCKET s = ::accept(s_, addr, &addr_len);
    UpdateLastError();
    if (s == INVALID_SOCKET)
      return NULL;
    enabled_events_ |= DE_ACCEPT;
    if (out_addr != NULL)
      SocketAddressFromSockAddrStorage(addr_storage, out_addr);
    return ss_->WrapSocket(s);
  }

  int Close() {
    if (s_ == INVALID_SOCKET)
      return 0;
    int err = ::closesocket(s_);
    UpdateLastError();
    s_ = INVALID_SOCKET;
    state_ = CS_CLOSED;
    enabled_events_ = 0;
    if (resolver_) {
      resolver_->Destroy(false);
      resolver_ = NULL;
    }
    return err;
  }

  int EstimateMTU(uint16* mtu) {
    SocketAddress addr = GetRemoteAddress();
    if (addr.IsAny()) {
      SetError(ENOTCONN);
      return -1;
    }

#if defined(WIN32)
    // Gets the interface MTU (TTL=1) for the interface used to reach |addr|.
    WinPing ping;
    if (!ping.IsValid()) {
      SetError(EINVAL);  // can't think of a better error ID
      return -1;
    }
    int header_size = ICMP_HEADER_SIZE;
    if (addr.family() == AF_INET6) {
      header_size += IPV6_HEADER_SIZE;
    } else if (addr.family() == AF_INET) {
      header_size += IP_HEADER_SIZE;
    }

    for (int level = 0; PACKET_MAXIMUMS[level + 1] > 0; ++level) {
      int32 size = PACKET_MAXIMUMS[level] - header_size;
      WinPing::PingResult result = ping.Ping(addr.ipaddr(), size,
                                             ICMP_PING_TIMEOUT_MILLIS,
                                             1, false);
      if (result == WinPing::PING_FAIL) {
        SetError(EINVAL);  // can't think of a better error ID
        return -1;
      } else if (result != WinPing::PING_TOO_LARGE) {
        *mtu = PACKET_MAXIMUMS[level];
        return 0;
      }
    }

    ASSERT(false);
    return -1;
#elif defined(IOS) || defined(OSX)
    // No simple way to do this on Mac OS X.
    // SIOCGIFMTU would work if we knew which interface would be used, but
    // figuring that out is pretty complicated. For now we'll return an error
    // and let the caller pick a default MTU.
    SetError(EINVAL);
    return -1;
#elif defined(LINUX) || defined(ANDROID)
    // Gets the path MTU.
    int value;
    socklen_t vlen = sizeof(value);
    int err = getsockopt(s_, IPPROTO_IP, IP_MTU, &value, &vlen);
    if (err < 0) {
      UpdateLastError();
      return err;
    }

    ASSERT((0 <= value) && (value <= 65536));
    *mtu = value;
    return 0;
#elif defined(__native_client__)
    // Most socket operations, including this, will fail in NaCl's sandbox.
    error_ = EACCES;
    return -1;
#endif
  }

  SocketServer* socketserver() { return ss_; }

 protected:
  void OnResolveResult(AsyncResolverInterface* resolver) {
    if (resolver != resolver_) {
      return;
    }

    int error = resolver_->GetError();
    if (error == 0) {
      error = DoConnect(resolver_->address());
    } else {
      Close();
    }

    if (error) {
      SetError(error);
      SignalCloseEvent(this, error);
    }
  }

  void UpdateLastError() {
    SetError(LAST_SYSTEM_ERROR);
  }

  void MaybeRemapSendError() {
#if defined(OSX) || defined(IOS)
    // https://developer.apple.com/library/mac/documentation/Darwin/
    // Reference/ManPages/man2/sendto.2.html
    // ENOBUFS - The output queue for a network interface is full.
    // This generally indicates that the interface has stopped sending,
    // but may be caused by transient congestion.
    if (GetError() == ENOBUFS) {
      SetError(EWOULDBLOCK);
    }
#endif
  }

  static int TranslateOption(Option opt, int* slevel, int* sopt) {
    switch (opt) {
      case OPT_DONTFRAGMENT:
#ifdef WIN32
        *slevel = IPPROTO_IP;
        *sopt = IP_DONTFRAGMENT;
        break;
#elif defined(IOS) || defined(OSX) || defined(BSD) || defined(__native_client__)
        LOG(LS_WARNING) << "Socket::OPT_DONTFRAGMENT not supported.";
        return -1;
#elif defined(POSIX)
        *slevel = IPPROTO_IP;
        *sopt = IP_MTU_DISCOVER;
        break;
#endif
      case OPT_RCVBUF:
        *slevel = SOL_SOCKET;
        *sopt = SO_RCVBUF;
        break;
      case OPT_SNDBUF:
        *slevel = SOL_SOCKET;
        *sopt = SO_SNDBUF;
        break;
      case OPT_NODELAY:
        *slevel = IPPROTO_TCP;
        *sopt = TCP_NODELAY;
        break;
      case OPT_DSCP:
        LOG(LS_WARNING) << "Socket::OPT_DSCP not supported.";
        return -1;
      case OPT_RTP_SENDTIME_EXTN_ID:
        return -1;  // No logging is necessary as this not a OS socket option.
      default:
        ASSERT(false);
        return -1;
    }
    return 0;
  }

  PhysicalSocketServer* ss_;
  SOCKET s_;
  uint8 enabled_events_;
  bool udp_;
  int error_;
  // Protects |error_| that is accessed from different threads.
  mutable CriticalSection crit_;
  ConnState state_;
  AsyncResolver* resolver_;

#ifdef _DEBUG
  std::string dbg_addr_;
#endif  // _DEBUG;
};

#ifdef POSIX
class EventDispatcher : public Dispatcher {
 public:
  EventDispatcher(PhysicalSocketServer* ss) : ss_(ss), fSignaled_(false) {
    if (pipe(afd_) < 0)
      LOG(LERROR) << "pipe failed";
    ss_->Add(this);
  }

  virtual ~EventDispatcher() {
    ss_->Remove(this);
    close(afd_[0]);
    close(afd_[1]);
  }

  virtual void Signal() {
    CritScope cs(&crit_);
    if (!fSignaled_) {
      const uint8 b[1] = { 0 };
      if (VERIFY(1 == write(afd_[1], b, sizeof(b)))) {
        fSignaled_ = true;
      }
    }
  }

  virtual uint32 GetRequestedEvents() {
    return DE_READ;
  }

  virtual void OnPreEvent(uint32 ff) {
    // It is not possible to perfectly emulate an auto-resetting event with
    // pipes.  This simulates it by resetting before the event is handled.

    CritScope cs(&crit_);
    if (fSignaled_) {
      uint8 b[4];  // Allow for reading more than 1 byte, but expect 1.
      VERIFY(1 == read(afd_[0], b, sizeof(b)));
      fSignaled_ = false;
    }
  }

  virtual void OnEvent(uint32 ff, int err) {
    ASSERT(false);
  }

  virtual int GetDescriptor() {
    return afd_[0];
  }

  virtual bool IsDescriptorClosed() {
    return false;
  }

 private:
  PhysicalSocketServer *ss_;
  int afd_[2];
  bool fSignaled_;
  CriticalSection crit_;
};

// These two classes use the self-pipe trick to deliver POSIX signals to our
// select loop. This is the only safe, reliable, cross-platform way to do
// non-trivial things with a POSIX signal in an event-driven program (until
// proper pselect() implementations become ubiquitous).

class PosixSignalHandler {
 public:
  // POSIX only specifies 32 signals, but in principle the system might have
  // more and the programmer might choose to use them, so we size our array
  // for 128.
  static const int kNumPosixSignals = 128;

  // There is just a single global instance. (Signal handlers do not get any
  // sort of user-defined void * parameter, so they can't access anything that
  // isn't global.)
  static PosixSignalHandler* Instance() {
    LIBJINGLE_DEFINE_STATIC_LOCAL(PosixSignalHandler, instance, ());
    return &instance;
  }

  // Returns true if the given signal number is set.
  bool IsSignalSet(int signum) const {
    ASSERT(signum < ARRAY_SIZE(received_signal_));
    if (signum < ARRAY_SIZE(received_signal_)) {
      return received_signal_[signum];
    } else {
      return false;
    }
  }

  // Clears the given signal number.
  void ClearSignal(int signum) {
    ASSERT(signum < ARRAY_SIZE(received_signal_));
    if (signum < ARRAY_SIZE(received_signal_)) {
      received_signal_[signum] = false;
    }
  }

  // Returns the file descriptor to monitor for signal events.
  int GetDescriptor() const {
    return afd_[0];
  }

  // This is called directly from our real signal handler, so it must be
  // signal-handler-safe. That means it cannot assume anything about the
  // user-level state of the process, since the handler could be executed at any
  // time on any thread.
  void OnPosixSignalReceived(int signum) {
    if (signum >= ARRAY_SIZE(received_signal_)) {
      // We don't have space in our array for this.
      return;
    }
    // Set a flag saying we've seen this signal.
    received_signal_[signum] = true;
    // Notify application code that we got a signal.
    const uint8 b[1] = { 0 };
    if (-1 == write(afd_[1], b, sizeof(b))) {
      // Nothing we can do here. If there's an error somehow then there's
      // nothing we can safely do from a signal handler.
      // No, we can't even safely log it.
      // But, we still have to check the return value here. Otherwise,
      // GCC 4.4.1 complains ignoring return value. Even (void) doesn't help.
      return;
    }
  }

 private:
  PosixSignalHandler() {
    if (pipe(afd_) < 0) {
      LOG_ERR(LS_ERROR) << "pipe failed";
      return;
    }
    if (fcntl(afd_[0], F_SETFL, O_NONBLOCK) < 0) {
      LOG_ERR(LS_WARNING) << "fcntl #1 failed";
    }
    if (fcntl(afd_[1], F_SETFL, O_NONBLOCK) < 0) {
      LOG_ERR(LS_WARNING) << "fcntl #2 failed";
    }
    memset(const_cast<void *>(static_cast<volatile void *>(received_signal_)),
           0,
           sizeof(received_signal_));
  }

  ~PosixSignalHandler() {
    int fd1 = afd_[0];
    int fd2 = afd_[1];
    // We clobber the stored file descriptor numbers here or else in principle
    // a signal that happens to be delivered during application termination
    // could erroneously write a zero byte to an unrelated file handle in
    // OnPosixSignalReceived() if some other file happens to be opened later
    // during shutdown and happens to be given the same file descriptor number
    // as our pipe had. Unfortunately even with this precaution there is still a
    // race where that could occur if said signal happens to be handled
    // concurrently with this code and happens to have already read the value of
    // afd_[1] from memory before we clobber it, but that's unlikely.
    afd_[0] = -1;
    afd_[1] = -1;
    close(fd1);
    close(fd2);
  }

  int afd_[2];
  // These are boolean flags that will be set in our signal handler and read
  // and cleared from Wait(). There is a race involved in this, but it is
  // benign. The signal handler sets the flag before signaling the pipe, so
  // we'll never end up blocking in select() while a flag is still true.
  // However, if two of the same signal arrive close to each other then it's
  // possible that the second time the handler may set the flag while it's still
  // true, meaning that signal will be missed. But the first occurrence of it
  // will still be handled, so this isn't a problem.
  // Volatile is not necessary here for correctness, but this data _is_ volatile
  // so I've marked it as such.
  volatile uint8 received_signal_[kNumPosixSignals];
};

class PosixSignalDispatcher : public Dispatcher {
 public:
  PosixSignalDispatcher(PhysicalSocketServer *owner) : owner_(owner) {
    owner_->Add(this);
  }

  virtual ~PosixSignalDispatcher() {
    owner_->Remove(this);
  }

  virtual uint32 GetRequestedEvents() {
    return DE_READ;
  }

  virtual void OnPreEvent(uint32 ff) {
    // Events might get grouped if signals come very fast, so we read out up to
    // 16 bytes to make sure we keep the pipe empty.
    uint8 b[16];
    ssize_t ret = read(GetDescriptor(), b, sizeof(b));
    if (ret < 0) {
      LOG_ERR(LS_WARNING) << "Error in read()";
    } else if (ret == 0) {
      LOG(LS_WARNING) << "Should have read at least one byte";
    }
  }

  virtual void OnEvent(uint32 ff, int err) {
    for (int signum = 0; signum < PosixSignalHandler::kNumPosixSignals;
         ++signum) {
      if (PosixSignalHandler::Instance()->IsSignalSet(signum)) {
        PosixSignalHandler::Instance()->ClearSignal(signum);
        HandlerMap::iterator i = handlers_.find(signum);
        if (i == handlers_.end()) {
          // This can happen if a signal is delivered to our process at around
          // the same time as we unset our handler for it. It is not an error
          // condition, but it's unusual enough to be worth logging.
          LOG(LS_INFO) << "Received signal with no handler: " << signum;
        } else {
          // Otherwise, execute our handler.
          (*i->second)(signum);
        }
      }
    }
  }

  virtual int GetDescriptor() {
    return PosixSignalHandler::Instance()->GetDescriptor();
  }

  virtual bool IsDescriptorClosed() {
    return false;
  }

  void SetHandler(int signum, void (*handler)(int)) {
    handlers_[signum] = handler;
  }

  void ClearHandler(int signum) {
    handlers_.erase(signum);
  }

  bool HasHandlers() {
    return !handlers_.empty();
  }

 private:
  typedef std::map<int, void (*)(int)> HandlerMap;

  HandlerMap handlers_;
  // Our owner.
  PhysicalSocketServer *owner_;
};

class SocketDispatcher : public Dispatcher, public PhysicalSocket {
 public:
  explicit SocketDispatcher(PhysicalSocketServer *ss) : PhysicalSocket(ss) {
  }
  SocketDispatcher(SOCKET s, PhysicalSocketServer *ss) : PhysicalSocket(ss, s) {
  }

  virtual ~SocketDispatcher() {
    Close();
  }

  bool Initialize() {
    ss_->Add(this);
    fcntl(s_, F_SETFL, fcntl(s_, F_GETFL, 0) | O_NONBLOCK);
    return true;
  }

  virtual bool Create(int type) {
    return Create(AF_INET, type);
  }

  virtual bool Create(int family, int type) {
    // Change the socket to be non-blocking.
    if (!PhysicalSocket::Create(family, type))
      return false;

    return Initialize();
  }

  virtual int GetDescriptor() {
    return s_;
  }

  virtual bool IsDescriptorClosed() {
    // We don't have a reliable way of distinguishing end-of-stream
    // from readability.  So test on each readable call.  Is this
    // inefficient?  Probably.
    char ch;
    ssize_t res = ::recv(s_, &ch, 1, MSG_PEEK);
    if (res > 0) {
      // Data available, so not closed.
      return false;
    } else if (res == 0) {
      // EOF, so closed.
      return true;
    } else {  // error
      switch (errno) {
        // Returned if we've already closed s_.
        case EBADF:
        // Returned during ungraceful peer shutdown.
        case ECONNRESET:
          return true;
        default:
          // Assume that all other errors are just blocking errors, meaning the
          // connection is still good but we just can't read from it right now.
          // This should only happen when connecting (and at most once), because
          // in all other cases this function is only called if the file
          // descriptor is already known to be in the readable state. However,
          // it's not necessary a problem if we spuriously interpret a
          // "connection lost"-type error as a blocking error, because typically
          // the next recv() will get EOF, so we'll still eventually notice that
          // the socket is closed.
          LOG_ERR(LS_WARNING) << "Assuming benign blocking error";
          return false;
      }
    }
  }

  virtual uint32 GetRequestedEvents() {
    return enabled_events_;
  }

  virtual void OnPreEvent(uint32 ff) {
    if ((ff & DE_CONNECT) != 0)
      state_ = CS_CONNECTED;
    if ((ff & DE_CLOSE) != 0)
      state_ = CS_CLOSED;
  }

  virtual void OnEvent(uint32 ff, int err) {
    // Make sure we deliver connect/accept first. Otherwise, consumers may see
    // something like a READ followed by a CONNECT, which would be odd.
    if ((ff & DE_CONNECT) != 0) {
      enabled_events_ &= ~DE_CONNECT;
      SignalConnectEvent(this);
    }
    if ((ff & DE_ACCEPT) != 0) {
      enabled_events_ &= ~DE_ACCEPT;
      SignalReadEvent(this);
    }
    if ((ff & DE_READ) != 0) {
      enabled_events_ &= ~DE_READ;
      SignalReadEvent(this);
    }
    if ((ff & DE_WRITE) != 0) {
      enabled_events_ &= ~DE_WRITE;
      SignalWriteEvent(this);
    }
    if ((ff & DE_CLOSE) != 0) {
      // The socket is now dead to us, so stop checking it.
      enabled_events_ = 0;
      SignalCloseEvent(this, err);
    }
  }

  virtual int Close() {
    if (s_ == INVALID_SOCKET)
      return 0;

    ss_->Remove(this);
    return PhysicalSocket::Close();
  }
};

class FileDispatcher: public Dispatcher, public AsyncFile {
 public:
  FileDispatcher(int fd, PhysicalSocketServer *ss) : ss_(ss), fd_(fd) {
    set_readable(true);

    ss_->Add(this);

    fcntl(fd_, F_SETFL, fcntl(fd_, F_GETFL, 0) | O_NONBLOCK);
  }

  virtual ~FileDispatcher() {
    ss_->Remove(this);
  }

  SocketServer* socketserver() { return ss_; }

  virtual int GetDescriptor() {
    return fd_;
  }

  virtual bool IsDescriptorClosed() {
    return false;
  }

  virtual uint32 GetRequestedEvents() {
    return flags_;
  }

  virtual void OnPreEvent(uint32 ff) {
  }

  virtual void OnEvent(uint32 ff, int err) {
    if ((ff & DE_READ) != 0)
      SignalReadEvent(this);
    if ((ff & DE_WRITE) != 0)
      SignalWriteEvent(this);
    if ((ff & DE_CLOSE) != 0)
      SignalCloseEvent(this, err);
  }

  virtual bool readable() {
    return (flags_ & DE_READ) != 0;
  }

  virtual void set_readable(bool value) {
    flags_ = value ? (flags_ | DE_READ) : (flags_ & ~DE_READ);
  }

  virtual bool writable() {
    return (flags_ & DE_WRITE) != 0;
  }

  virtual void set_writable(bool value) {
    flags_ = value ? (flags_ | DE_WRITE) : (flags_ & ~DE_WRITE);
  }

 private:
  PhysicalSocketServer* ss_;
  int fd_;
  int flags_;
};

AsyncFile* PhysicalSocketServer::CreateFile(int fd) {
  return new FileDispatcher(fd, this);
}

#endif // POSIX

#ifdef WIN32
static uint32 FlagsToEvents(uint32 events) {
  uint32 ffFD = FD_CLOSE;
  if (events & DE_READ)
    ffFD |= FD_READ;
  if (events & DE_WRITE)
    ffFD |= FD_WRITE;
  if (events & DE_CONNECT)
    ffFD |= FD_CONNECT;
  if (events & DE_ACCEPT)
    ffFD |= FD_ACCEPT;
  return ffFD;
}

class EventDispatcher : public Dispatcher {
 public:
  EventDispatcher(PhysicalSocketServer *ss) : ss_(ss) {
    hev_ = WSACreateEvent();
    if (hev_) {
      ss_->Add(this);
    }
  }

  ~EventDispatcher() {
    if (hev_ != NULL) {
      ss_->Remove(this);
      WSACloseEvent(hev_);
      hev_ = NULL;
    }
  }

  virtual void Signal() {
    if (hev_ != NULL)
      WSASetEvent(hev_);
  }

  virtual uint32 GetRequestedEvents() {
    return 0;
  }

  virtual void OnPreEvent(uint32 ff) {
    WSAResetEvent(hev_);
  }

  virtual void OnEvent(uint32 ff, int err) {
  }

  virtual WSAEVENT GetWSAEvent() {
    return hev_;
  }

  virtual SOCKET GetSocket() {
    return INVALID_SOCKET;
  }

  virtual bool CheckSignalClose() { return false; }

private:
  PhysicalSocketServer* ss_;
  WSAEVENT hev_;
};

class SocketDispatcher : public Dispatcher, public PhysicalSocket {
 public:
  static int next_id_;
  int id_;
  bool signal_close_;
  int signal_err_;

  SocketDispatcher(PhysicalSocketServer* ss)
      : PhysicalSocket(ss),
        id_(0),
        signal_close_(false) {
  }

  SocketDispatcher(SOCKET s, PhysicalSocketServer* ss)
      : PhysicalSocket(ss, s),
        id_(0),
        signal_close_(false) {
  }

  virtual ~SocketDispatcher() {
    Close();
  }

  bool Initialize() {
    ASSERT(s_ != INVALID_SOCKET);
    // Must be a non-blocking
    u_long argp = 1;
    ioctlsocket(s_, FIONBIO, &argp);
    ss_->Add(this);
    return true;
  }

  virtual bool Create(int type) {
    return Create(AF_INET, type);
  }

  virtual bool Create(int family, int type) {
    // Create socket
    if (!PhysicalSocket::Create(family, type))
      return false;

    if (!Initialize())
      return false;

    do { id_ = ++next_id_; } while (id_ == 0);
    return true;
  }

  virtual int Close() {
    if (s_ == INVALID_SOCKET)
      return 0;

    id_ = 0;
    signal_close_ = false;
    ss_->Remove(this);
    return PhysicalSocket::Close();
  }

  virtual uint32 GetRequestedEvents() {
    return enabled_events_;
  }

  virtual void OnPreEvent(uint32 ff) {
    if ((ff & DE_CONNECT) != 0)
      state_ = CS_CONNECTED;
    // We set CS_CLOSED from CheckSignalClose.
  }

  virtual void OnEvent(uint32 ff, int err) {
    int cache_id = id_;
    // Make sure we deliver connect/accept first. Otherwise, consumers may see
    // something like a READ followed by a CONNECT, which would be odd.
    if (((ff & DE_CONNECT) != 0) && (id_ == cache_id)) {
      if (ff != DE_CONNECT)
        LOG(LS_VERBOSE) << "Signalled with DE_CONNECT: " << ff;
      enabled_events_ &= ~DE_CONNECT;
#ifdef _DEBUG
      dbg_addr_ = "Connected @ ";
      dbg_addr_.append(GetRemoteAddress().ToString());
#endif  // _DEBUG
      SignalConnectEvent(this);
    }
    if (((ff & DE_ACCEPT) != 0) && (id_ == cache_id)) {
      enabled_events_ &= ~DE_ACCEPT;
      SignalReadEvent(this);
    }
    if ((ff & DE_READ) != 0) {
      enabled_events_ &= ~DE_READ;
      SignalReadEvent(this);
    }
    if (((ff & DE_WRITE) != 0) && (id_ == cache_id)) {
      enabled_events_ &= ~DE_WRITE;
      SignalWriteEvent(this);
    }
    if (((ff & DE_CLOSE) != 0) && (id_ == cache_id)) {
      signal_close_ = true;
      signal_err_ = err;
    }
  }

  virtual WSAEVENT GetWSAEvent() {
    return WSA_INVALID_EVENT;
  }

  virtual SOCKET GetSocket() {
    return s_;
  }

  virtual bool CheckSignalClose() {
    if (!signal_close_)
      return false;

    char ch;
    if (recv(s_, &ch, 1, MSG_PEEK) > 0)
      return false;

    state_ = CS_CLOSED;
    signal_close_ = false;
    SignalCloseEvent(this, signal_err_);
    return true;
  }
};

int SocketDispatcher::next_id_ = 0;

#endif  // WIN32

// Sets the value of a boolean value to false when signaled.
class Signaler : public EventDispatcher {
 public:
  Signaler(PhysicalSocketServer* ss, bool* pf)
      : EventDispatcher(ss), pf_(pf) {
  }
  virtual ~Signaler() { }

  void OnEvent(uint32 ff, int err) {
    if (pf_)
      *pf_ = false;
  }

 private:
  bool *pf_;
};

PhysicalSocketServer::PhysicalSocketServer()
    : fWait_(false) {
  signal_wakeup_ = new Signaler(this, &fWait_);
#ifdef WIN32
  socket_ev_ = WSACreateEvent();
#endif
}

PhysicalSocketServer::~PhysicalSocketServer() {
#ifdef WIN32
  WSACloseEvent(socket_ev_);
#endif
#ifdef POSIX
  signal_dispatcher_.reset();
#endif
  delete signal_wakeup_;
  ASSERT(dispatchers_.empty());
}

void PhysicalSocketServer::WakeUp() {
  signal_wakeup_->Signal();
}

Socket* PhysicalSocketServer::CreateSocket(int type) {
  return CreateSocket(AF_INET, type);
}

Socket* PhysicalSocketServer::CreateSocket(int family, int type) {
  PhysicalSocket* socket = new PhysicalSocket(this);
  if (socket->Create(family, type)) {
    return socket;
  } else {
    delete socket;
    return 0;
  }
}

AsyncSocket* PhysicalSocketServer::CreateAsyncSocket(int type) {
  return CreateAsyncSocket(AF_INET, type);
}

AsyncSocket* PhysicalSocketServer::CreateAsyncSocket(int family, int type) {
  SocketDispatcher* dispatcher = new SocketDispatcher(this);
  if (dispatcher->Create(family, type)) {
    return dispatcher;
  } else {
    delete dispatcher;
    return 0;
  }
}

AsyncSocket* PhysicalSocketServer::WrapSocket(SOCKET s) {
  SocketDispatcher* dispatcher = new SocketDispatcher(s, this);
  if (dispatcher->Initialize()) {
    return dispatcher;
  } else {
    delete dispatcher;
    return 0;
  }
}

void PhysicalSocketServer::Add(Dispatcher *pdispatcher) {
  CritScope cs(&crit_);
  // Prevent duplicates. This can cause dead dispatchers to stick around.
  DispatcherList::iterator pos = std::find(dispatchers_.begin(),
                                           dispatchers_.end(),
                                           pdispatcher);
  if (pos != dispatchers_.end())
    return;
  dispatchers_.push_back(pdispatcher);
}

void PhysicalSocketServer::Remove(Dispatcher *pdispatcher) {
  CritScope cs(&crit_);
  DispatcherList::iterator pos = std::find(dispatchers_.begin(),
                                           dispatchers_.end(),
                                           pdispatcher);
  // We silently ignore duplicate calls to Add, so we should silently ignore
  // the (expected) symmetric calls to Remove. Note that this may still hide
  // a real issue, so we at least log a warning about it.
  if (pos == dispatchers_.end()) {
    LOG(LS_WARNING) << "PhysicalSocketServer asked to remove a unknown "
                    << "dispatcher, potentially from a duplicate call to Add.";
    return;
  }
  size_t index = pos - dispatchers_.begin();
  dispatchers_.erase(pos);
  for (IteratorList::iterator it = iterators_.begin(); it != iterators_.end();
       ++it) {
    if (index < **it) {
      --**it;
    }
  }
}

#ifdef POSIX
bool PhysicalSocketServer::Wait(int cmsWait, bool process_io) {
  // Calculate timing information

  struct timeval *ptvWait = NULL;
  struct timeval tvWait;
  struct timeval tvStop;
  if (cmsWait != kForever) {
    // Calculate wait timeval
    tvWait.tv_sec = cmsWait / 1000;
    tvWait.tv_usec = (cmsWait % 1000) * 1000;
    ptvWait = &tvWait;

    // Calculate when to return in a timeval
    gettimeofday(&tvStop, NULL);
    tvStop.tv_sec += tvWait.tv_sec;
    tvStop.tv_usec += tvWait.tv_usec;
    if (tvStop.tv_usec >= 1000000) {
      tvStop.tv_usec -= 1000000;
      tvStop.tv_sec += 1;
    }
  }

  // Zero all fd_sets. Don't need to do this inside the loop since
  // select() zeros the descriptors not signaled

  fd_set fdsRead;
  FD_ZERO(&fdsRead);
  fd_set fdsWrite;
  FD_ZERO(&fdsWrite);

  fWait_ = true;

  while (fWait_) {
    int fdmax = -1;
    {
      CritScope cr(&crit_);
      for (size_t i = 0; i < dispatchers_.size(); ++i) {
        // Query dispatchers for read and write wait state
        Dispatcher *pdispatcher = dispatchers_[i];
        ASSERT(pdispatcher);
        if (!process_io && (pdispatcher != signal_wakeup_))
          continue;
        int fd = pdispatcher->GetDescriptor();
        if (fd > fdmax)
          fdmax = fd;

        uint32 ff = pdispatcher->GetRequestedEvents();
        if (ff & (DE_READ | DE_ACCEPT))
          FD_SET(fd, &fdsRead);
        if (ff & (DE_WRITE | DE_CONNECT))
          FD_SET(fd, &fdsWrite);
      }
    }

    // Wait then call handlers as appropriate
    // < 0 means error
    // 0 means timeout
    // > 0 means count of descriptors ready
    int n = select(fdmax + 1, &fdsRead, &fdsWrite, NULL, ptvWait);

    // If error, return error.
    if (n < 0) {
      if (errno != EINTR) {
        LOG_E(LS_ERROR, EN, errno) << "select";
        return false;
      }
      // Else ignore the error and keep going. If this EINTR was for one of the
      // signals managed by this PhysicalSocketServer, the
      // PosixSignalDeliveryDispatcher will be in the signaled state in the next
      // iteration.
    } else if (n == 0) {
      // If timeout, return success
      return true;
    } else {
      // We have signaled descriptors
      CritScope cr(&crit_);
      for (size_t i = 0; i < dispatchers_.size(); ++i) {
        Dispatcher *pdispatcher = dispatchers_[i];
        int fd = pdispatcher->GetDescriptor();
        uint32 ff = 0;
        int errcode = 0;

        // Reap any error code, which can be signaled through reads or writes.
        // TODO: Should we set errcode if getsockopt fails?
        if (FD_ISSET(fd, &fdsRead) || FD_ISSET(fd, &fdsWrite)) {
          socklen_t len = sizeof(errcode);
          ::getsockopt(fd, SOL_SOCKET, SO_ERROR, &errcode, &len);
        }

        // Check readable descriptors. If we're waiting on an accept, signal
        // that. Otherwise we're waiting for data, check to see if we're
        // readable or really closed.
        // TODO: Only peek at TCP descriptors.
        if (FD_ISSET(fd, &fdsRead)) {
          FD_CLR(fd, &fdsRead);
          if (pdispatcher->GetRequestedEvents() & DE_ACCEPT) {
            ff |= DE_ACCEPT;
          } else if (errcode || pdispatcher->IsDescriptorClosed()) {
            ff |= DE_CLOSE;
          } else {
            ff |= DE_READ;
          }
        }

        // Check writable descriptors. If we're waiting on a connect, detect
        // success versus failure by the reaped error code.
        if (FD_ISSET(fd, &fdsWrite)) {
          FD_CLR(fd, &fdsWrite);
          if (pdispatcher->GetRequestedEvents() & DE_CONNECT) {
            if (!errcode) {
              ff |= DE_CONNECT;
            } else {
              ff |= DE_CLOSE;
            }
          } else {
            ff |= DE_WRITE;
          }
        }

        // Tell the descriptor about the event.
        if (ff != 0) {
          pdispatcher->OnPreEvent(ff);
          pdispatcher->OnEvent(ff, errcode);
        }
      }
    }

    // Recalc the time remaining to wait. Doing it here means it doesn't get
    // calced twice the first time through the loop
    if (ptvWait) {
      ptvWait->tv_sec = 0;
      ptvWait->tv_usec = 0;
      struct timeval tvT;
      gettimeofday(&tvT, NULL);
      if ((tvStop.tv_sec > tvT.tv_sec)
          || ((tvStop.tv_sec == tvT.tv_sec)
              && (tvStop.tv_usec > tvT.tv_usec))) {
        ptvWait->tv_sec = tvStop.tv_sec - tvT.tv_sec;
        ptvWait->tv_usec = tvStop.tv_usec - tvT.tv_usec;
        if (ptvWait->tv_usec < 0) {
          ASSERT(ptvWait->tv_sec > 0);
          ptvWait->tv_usec += 1000000;
          ptvWait->tv_sec -= 1;
        }
      }
    }
  }

  return true;
}

static void GlobalSignalHandler(int signum) {
  PosixSignalHandler::Instance()->OnPosixSignalReceived(signum);
}

bool PhysicalSocketServer::SetPosixSignalHandler(int signum,
                                                 void (*handler)(int)) {
  // If handler is SIG_IGN or SIG_DFL then clear our user-level handler,
  // otherwise set one.
  if (handler == SIG_IGN || handler == SIG_DFL) {
    if (!InstallSignal(signum, handler)) {
      return false;
    }
    if (signal_dispatcher_) {
      signal_dispatcher_->ClearHandler(signum);
      if (!signal_dispatcher_->HasHandlers()) {
        signal_dispatcher_.reset();
      }
    }
  } else {
    if (!signal_dispatcher_) {
      signal_dispatcher_.reset(new PosixSignalDispatcher(this));
    }
    signal_dispatcher_->SetHandler(signum, handler);
    if (!InstallSignal(signum, &GlobalSignalHandler)) {
      return false;
    }
  }
  return true;
}

Dispatcher* PhysicalSocketServer::signal_dispatcher() {
  return signal_dispatcher_.get();
}

bool PhysicalSocketServer::InstallSignal(int signum, void (*handler)(int)) {
  struct sigaction act;
  // It doesn't really matter what we set this mask to.
  if (sigemptyset(&act.sa_mask) != 0) {
    LOG_ERR(LS_ERROR) << "Couldn't set mask";
    return false;
  }
  act.sa_handler = handler;
#if !defined(__native_client__)
  // Use SA_RESTART so that our syscalls don't get EINTR, since we don't need it
  // and it's a nuisance. Though some syscalls still return EINTR and there's no
  // real standard for which ones. :(
  act.sa_flags = SA_RESTART;
#else
  act.sa_flags = 0;
#endif
  if (sigaction(signum, &act, NULL) != 0) {
    LOG_ERR(LS_ERROR) << "Couldn't set sigaction";
    return false;
  }
  return true;
}
#endif  // POSIX

#ifdef WIN32
bool PhysicalSocketServer::Wait(int cmsWait, bool process_io) {
  int cmsTotal = cmsWait;
  int cmsElapsed = 0;
  uint32 msStart = Time();

  fWait_ = true;
  while (fWait_) {
    std::vector<WSAEVENT> events;
    std::vector<Dispatcher *> event_owners;

    events.push_back(socket_ev_);

    {
      CritScope cr(&crit_);
      size_t i = 0;
      iterators_.push_back(&i);
      // Don't track dispatchers_.size(), because we want to pick up any new
      // dispatchers that were added while processing the loop.
      while (i < dispatchers_.size()) {
        Dispatcher* disp = dispatchers_[i++];
        if (!process_io && (disp != signal_wakeup_))
          continue;
        SOCKET s = disp->GetSocket();
        if (disp->CheckSignalClose()) {
          // We just signalled close, don't poll this socket
        } else if (s != INVALID_SOCKET) {
          WSAEventSelect(s,
                         events[0],
                         FlagsToEvents(disp->GetRequestedEvents()));
        } else {
          events.push_back(disp->GetWSAEvent());
          event_owners.push_back(disp);
        }
      }
      ASSERT(iterators_.back() == &i);
      iterators_.pop_back();
    }

    // Which is shorter, the delay wait or the asked wait?

    int cmsNext;
    if (cmsWait == kForever) {
      cmsNext = cmsWait;
    } else {
      cmsNext = _max(0, cmsTotal - cmsElapsed);
    }

    // Wait for one of the events to signal
    DWORD dw = WSAWaitForMultipleEvents(static_cast<DWORD>(events.size()),
                                        &events[0],
                                        false,
                                        cmsNext,
                                        false);

    if (dw == WSA_WAIT_FAILED) {
      // Failed?
      // TODO: need a better strategy than this!
      WSAGetLastError();
      ASSERT(false);
      return false;
    } else if (dw == WSA_WAIT_TIMEOUT) {
      // Timeout?
      return true;
    } else {
      // Figure out which one it is and call it
      CritScope cr(&crit_);
      int index = dw - WSA_WAIT_EVENT_0;
      if (index > 0) {
        --index; // The first event is the socket event
        event_owners[index]->OnPreEvent(0);
        event_owners[index]->OnEvent(0, 0);
      } else if (process_io) {
        size_t i = 0, end = dispatchers_.size();
        iterators_.push_back(&i);
        iterators_.push_back(&end);  // Don't iterate over new dispatchers.
        while (i < end) {
          Dispatcher* disp = dispatchers_[i++];
          SOCKET s = disp->GetSocket();
          if (s == INVALID_SOCKET)
            continue;

          WSANETWORKEVENTS wsaEvents;
          int err = WSAEnumNetworkEvents(s, events[0], &wsaEvents);
          if (err == 0) {

#if LOGGING
            {
              if ((wsaEvents.lNetworkEvents & FD_READ) &&
                  wsaEvents.iErrorCode[FD_READ_BIT] != 0) {
                LOG(WARNING) << "PhysicalSocketServer got FD_READ_BIT error "
                             << wsaEvents.iErrorCode[FD_READ_BIT];
              }
              if ((wsaEvents.lNetworkEvents & FD_WRITE) &&
                  wsaEvents.iErrorCode[FD_WRITE_BIT] != 0) {
                LOG(WARNING) << "PhysicalSocketServer got FD_WRITE_BIT error "
                             << wsaEvents.iErrorCode[FD_WRITE_BIT];
              }
              if ((wsaEvents.lNetworkEvents & FD_CONNECT) &&
                  wsaEvents.iErrorCode[FD_CONNECT_BIT] != 0) {
                LOG(WARNING) << "PhysicalSocketServer got FD_CONNECT_BIT error "
                             << wsaEvents.iErrorCode[FD_CONNECT_BIT];
              }
              if ((wsaEvents.lNetworkEvents & FD_ACCEPT) &&
                  wsaEvents.iErrorCode[FD_ACCEPT_BIT] != 0) {
                LOG(WARNING) << "PhysicalSocketServer got FD_ACCEPT_BIT error "
                             << wsaEvents.iErrorCode[FD_ACCEPT_BIT];
              }
              if ((wsaEvents.lNetworkEvents & FD_CLOSE) &&
                  wsaEvents.iErrorCode[FD_CLOSE_BIT] != 0) {
                LOG(WARNING) << "PhysicalSocketServer got FD_CLOSE_BIT error "
                             << wsaEvents.iErrorCode[FD_CLOSE_BIT];
              }
            }
#endif
            uint32 ff = 0;
            int errcode = 0;
            if (wsaEvents.lNetworkEvents & FD_READ)
              ff |= DE_READ;
            if (wsaEvents.lNetworkEvents & FD_WRITE)
              ff |= DE_WRITE;
            if (wsaEvents.lNetworkEvents & FD_CONNECT) {
              if (wsaEvents.iErrorCode[FD_CONNECT_BIT] == 0) {
                ff |= DE_CONNECT;
              } else {
                ff |= DE_CLOSE;
                errcode = wsaEvents.iErrorCode[FD_CONNECT_BIT];
              }
            }
            if (wsaEvents.lNetworkEvents & FD_ACCEPT)
              ff |= DE_ACCEPT;
            if (wsaEvents.lNetworkEvents & FD_CLOSE) {
              ff |= DE_CLOSE;
              errcode = wsaEvents.iErrorCode[FD_CLOSE_BIT];
            }
            if (ff != 0) {
              disp->OnPreEvent(ff);
              disp->OnEvent(ff, errcode);
            }
          }
        }
        ASSERT(iterators_.back() == &end);
        iterators_.pop_back();
        ASSERT(iterators_.back() == &i);
        iterators_.pop_back();
      }

      // Reset the network event until new activity occurs
      WSAResetEvent(socket_ev_);
    }

    // Break?
    if (!fWait_)
      break;
    cmsElapsed = TimeSince(msStart);
    if ((cmsWait != kForever) && (cmsElapsed >= cmsWait)) {
       break;
    }
  }

  // Done
  return true;
}
#endif  // WIN32

}  // namespace talk_base