Merge branch 'poll'

Closes: #316, #401
Ref: #265, #349

config-unix.mk

@@ -37,6 +37,7 @@ OBJS += src/unix/fs.o
 OBJS += src/unix/idle.o
 OBJS += src/unix/loop.o
 OBJS += src/unix/pipe.o
+OBJS += src/unix/poll.o
 OBJS += src/unix/prepare.o
 OBJS += src/unix/process.o
 OBJS += src/unix/stream.o

include/uv-private/uv-unix.h

@@ -45,6 +45,8 @@ typedef struct {
 
 typedef int uv_file;
 
+typedef int uv_os_sock_t;
+
 #define UV_ONCE_INIT PTHREAD_ONCE_INIT
 
 typedef pthread_once_t uv_once_t;
@@ -162,6 +164,11 @@ typedef void* uv_lib_t;
   const char* pipe_fname; /* strdup'ed */
 
 
+/* UV_POLL */
+#define UV_POLL_PRIVATE_FIELDS        \
+  ev_io io_watcher;
+
+
 /* UV_PREPARE */ \
 #define UV_PREPARE_PRIVATE_FIELDS \
   ev_prepare prepare_watcher; \

include/uv-private/uv-win.h

@@ -146,6 +146,8 @@ typedef struct _AFD_POLL_INFO {
   AFD_POLL_HANDLE_INFO Handles[1];
 } AFD_POLL_INFO, *PAFD_POLL_INFO;
 
+#define UV_MSAFD_PROVIDER_COUNT 3
+
 
 /**
  * It should be possible to cast uv_buf_t[] to WSABUF[]
@@ -158,6 +160,8 @@ typedef struct uv_buf_t {
 
 typedef int uv_file;
 
+typedef SOCKET uv_os_sock_t;
+
 typedef HANDLE uv_thread_t;
 
 typedef CRITICAL_SECTION uv_mutex_t;
@@ -219,6 +223,9 @@ RB_HEAD(uv_timer_tree_s, uv_timer_s);
   uv_prepare_t* next_prepare_handle;                                          \
   uv_check_t* next_check_handle;                                              \
   uv_idle_t* next_idle_handle;                                                \
+  /* This handle holds the peer sockets for the fast variant of uv_poll_t */  \
+  SOCKET poll_peer_sockets[UV_MSAFD_PROVIDER_COUNT];                          \
+  /* State used by uv_ares. */                                                \
   ares_channel ares_chan;                                                     \
   int ares_active_sockets;                                                    \
   uv_timer_t ares_polling_timer;                                              \
@@ -237,6 +244,7 @@ RB_HEAD(uv_timer_tree_s, uv_timer_s);
   UV_ARES_CLEANUP_REQ,                    \
   UV_FS_EVENT_REQ,                        \
   UV_GETADDRINFO_REQ,                     \
+  UV_POLL_REQ,                            \
   UV_PROCESS_EXIT,                        \
   UV_PROCESS_CLOSE,                       \
   UV_READ,                                \
@@ -386,6 +394,21 @@ RB_HEAD(uv_timer_tree_s, uv_timer_s);
   COORD saved_position;                   \
   WORD saved_attributes;
 
+#define UV_POLL_PRIVATE_FIELDS            \
+  SOCKET socket;                          \
+  /* Used in fast mode */                 \
+  SOCKET peer_socket;                     \
+  AFD_POLL_INFO afd_poll_info_1;          \
+  AFD_POLL_INFO afd_poll_info_2;          \
+  /* Used in fast and slow mode. */       \
+  uv_req_t poll_req_1;                    \
+  uv_req_t poll_req_2;                    \
+  unsigned char submitted_events_1;       \
+  unsigned char submitted_events_2;       \
+  unsigned char mask_events_1;            \
+  unsigned char mask_events_2;            \
+  unsigned char events;
+
 #define UV_TIMER_PRIVATE_FIELDS           \
   RB_ENTRY(uv_timer_s) tree_entry;        \
   int64_t due;                            \

include/uv.h

@@ -164,6 +164,7 @@ typedef enum {
 #define XX(uc, lc) UV_##uc,
   UV_HANDLE_TYPE_MAP(XX)
 #undef XX
+  UV_POLL,
   UV_FILE,
   UV_HANDLE_TYPE_PRIVATE
   UV_HANDLE_TYPE_MAX
@@ -189,6 +190,7 @@ typedef struct uv_tcp_s uv_tcp_t;
 typedef struct uv_udp_s uv_udp_t;
 typedef struct uv_pipe_s uv_pipe_t;
 typedef struct uv_tty_s uv_tty_t;
+typedef struct uv_poll_s uv_poll_t;
 typedef struct uv_timer_s uv_timer_t;
 typedef struct uv_prepare_s uv_prepare_t;
 typedef struct uv_check_s uv_check_t;
@@ -288,6 +290,7 @@ typedef void (*uv_connect_cb)(uv_connect_t* req, int status);
 typedef void (*uv_shutdown_cb)(uv_shutdown_t* req, int status);
 typedef void (*uv_connection_cb)(uv_stream_t* server, int status);
 typedef void (*uv_close_cb)(uv_handle_t* handle);
+typedef void (*uv_poll_cb)(uv_poll_t* handle, int status, int events);
 typedef void (*uv_timer_cb)(uv_timer_t* handle, int status);
 /* TODO: do these really need a status argument? */
 typedef void (*uv_async_cb)(uv_async_t* handle, int status);
@@ -926,6 +929,74 @@ UV_EXTERN void uv_pipe_connect(uv_connect_t* req, uv_pipe_t* handle,
 UV_EXTERN void uv_pipe_pending_instances(uv_pipe_t* handle, int count);
 
 
+/*
+ * uv_poll_t is a subclass of uv_handle_t.
+ *
+ * The uv_poll watcher is used to watch file descriptors for readability and
+ * writability, similar to the purpose of poll(2).
+ *
+ * The purpose of uv_poll is to enable integrating external libraries that
+ * rely on the event loop to signal it about the socket status changes, like
+ * c-ares or libssh2. Using uv_poll_t for any other other purpose is not
+ * recommended; uv_tcp_t, uv_udp_t, etc. provide an implementation that is
+ * much faster and more scalable than what can be achieved with uv_poll_t,
+ * especially on Windows.
+ *
+ * It is possible that uv_poll occasionally signals that a file descriptor is
+ * readable or writable even when it isn't. The user should therefore always
+ * be prepared to handle EAGAIN or equivalent when it attempts to read from or
+ * write to the fd.
+ *
+ * It is not okay to have multiple active uv_poll watchers for the same socket.
+ * This can cause libuv to busyloop or otherwise malfunction.
+ *
+ * The user should not close a file descriptor while it is being polled by an
+ * active uv_poll watcher. This can cause the poll watcher to report an error,
+ * but it might also start polling another socket. However the fd can be safely
+ * closed immediately after a call to uv_poll_stop() or uv_close().
+ *
+ * On windows only sockets can be polled with uv_poll. On unix any file
+ * descriptor that would be accepted by poll(2) can be used with uv_poll.
+ */
+struct uv_poll_s {
+  UV_HANDLE_FIELDS
+  uv_poll_cb poll_cb;
+  UV_POLL_PRIVATE_FIELDS
+};
+
+enum uv_poll_event {
+  UV_READABLE = 1,
+  UV_WRITABLE = 2
+};
+
+/* Initialize the poll watcher using a file descriptor. */
+UV_EXTERN int uv_poll_init(uv_loop_t* loop, uv_poll_t* handle, int fd);
+
+/* Initialize the poll watcher using a socket descriptor. On unix this is */
+/* identical to uv_poll_init. On windows it takes a SOCKET handle. */
+UV_EXTERN int uv_poll_init_socket(uv_loop_t* loop, uv_poll_t* handle,
+    uv_os_sock_t socket);
+
+/*
+ * Starts polling the file descriptor. `events` is a bitmask consisting made up
+ * of UV_READABLE and UV_WRITABLE. As soon as an event is detected the callback
+ * will be called with `status` set to 0, and the detected events set en the
+ * `events` field.
+ *
+ * If an error happens while polling status may be set to -1 and the error
+ * code can be retrieved with uv_last_error. The user should not close the
+ * socket while uv_poll is active. If the user does that anyway, the callback
+ * *may* be called reporting an error status, but this is not guaranteed.
+ *
+ * Calling uv_poll_start on an uv_poll watcher that is already active is fine.
+ * Doing so will update the events mask that is being watched for.
+ */
+UV_EXTERN int uv_poll_start(uv_poll_t* handle, int events, uv_poll_cb cb);
+
+/* Stops polling the file descriptor. */
+UV_EXTERN int uv_poll_stop(uv_poll_t* handle);
+
+
 /*
  * uv_prepare_t is a subclass of uv_handle_t.
  *
@@ -1543,6 +1614,7 @@ struct uv_counters_s {
   uint64_t udp_init;
   uint64_t pipe_init;
   uint64_t tty_init;
+  uint64_t poll_init;
   uint64_t prepare_init;
   uint64_t check_init;
   uint64_t idle_init;

src/unix/core.c

@@ -107,6 +107,10 @@ void uv_close(uv_handle_t* handle, uv_close_cb close_cb) {
     uv__fs_event_close((uv_fs_event_t*)handle);
     break;
 
+  case UV_POLL:
+    uv__poll_close((uv_poll_t*)handle);
+    break;
+
   default:
     assert(0);
   }
@@ -249,6 +253,9 @@ void uv__finish_close(uv_handle_t* handle) {
     case UV_FS_EVENT:
       break;
 
+    case UV_POLL:
+      break;
+
     default:
       assert(0);
       break;
@@ -285,6 +292,8 @@ int64_t uv_now(uv_loop_t* loop) {
 
 int uv_is_active(const uv_handle_t* handle) {
   switch (handle->type) {
+  case UV_POLL:
+    return uv__poll_active((const uv_poll_t*)handle);
   case UV_CHECK:
     return uv__check_active((const uv_check_t*)handle);
   case UV_IDLE:

src/unix/internal.h

@@ -141,6 +141,10 @@ int uv__tcp_keepalive(uv_tcp_t* handle, int enable, unsigned int delay);
 int uv_pipe_listen(uv_pipe_t* handle, int backlog, uv_connection_cb cb);
 void uv__pipe_accept(EV_P_ ev_io* watcher, int revents);
 
+/* poll */
+void uv__poll_close(uv_poll_t* handle);
+int uv__poll_active(const uv_poll_t* handle);
+
 /* various */
 int uv__check_active(const uv_check_t* handle);
 int uv__idle_active(const uv_idle_t* handle);

src/unix/poll.c

@@ -0,0 +1,130 @@
+/* Copyright Joyent, Inc. and other Node contributors. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "uv.h"
+#include "internal.h"
+
+#include <unistd.h>
+#include <assert.h>
+#include <errno.h>
+
+
+static void uv__poll_io(EV_P_ ev_io* watcher, int ev_events) {
+  uv_poll_t* handle = watcher->data;
+  int events;
+
+  if (ev_events & EV_ERROR) {
+    /* An error happened. Libev has implicitly stopped the watcher, but we */
+    /* need to fix the refcount. */
+    uv_ref(handle->loop);
+    uv__set_sys_error(handle->loop, EBADF);
+    handle->poll_cb(handle, -1, 0);
+    return;
+  }
+
+  assert(ev_events & (EV_READ | EV_WRITE));
+  assert((ev_events & ~(EV_READ | EV_WRITE)) == 0);
+
+  events = 0;
+  if (ev_events & EV_READ)
+    events |= UV_READABLE;
+  if (ev_events & EV_WRITE)
+    events |= UV_WRITABLE;
+
+  handle->poll_cb(handle, 0, events);
+}
+
+
+int uv_poll_init(uv_loop_t* loop, uv_poll_t* handle, int fd) {
+  uv__handle_init(loop, (uv_handle_t*) handle, UV_POLL);
+  loop->counters.poll_init++;
+
+  handle->fd = fd;
+  handle->poll_cb = NULL;
+
+  ev_init(&handle->io_watcher, uv__poll_io);
+  handle->io_watcher.data = handle;
+
+  return 0;
+}
+
+
+int uv_poll_init_socket(uv_loop_t* loop, uv_poll_t* handle,
+    uv_os_sock_t socket) {
+  return uv_poll_init(loop, handle, socket);
+}
+
+
+static void uv__poll_stop(uv_poll_t* handle) {
+  if (ev_is_active(&handle->io_watcher)) {
+    ev_io_stop(handle->loop->ev, &handle->io_watcher);
+    uv_ref(handle->loop);
+  }
+}
+
+
+int uv_poll_stop(uv_poll_t* handle) {
+  assert(!(handle->flags & (UV_CLOSING | UV_CLOSED)));
+  uv__poll_stop(handle);
+  return 0;
+}
+
+
+int uv_poll_start(uv_poll_t* handle, int events, uv_poll_cb poll_cb) {
+  int ev_events;
+  int was_active;
+
+  assert((events & ~(UV_READABLE | UV_WRITABLE)) == 0);
+  assert(!(handle->flags & (UV_CLOSING | UV_CLOSED)));
+
+  if (events == 0) {
+    uv__poll_stop(handle);
+    return 0;
+  }
+
+  ev_events = 0;
+  if (events & UV_READABLE)
+    ev_events |= EV_READ;
+  if (events & UV_WRITABLE)
+    ev_events |= EV_WRITE;
+
+  was_active = ev_is_active(&handle->io_watcher);
+
+  ev_io_set(&handle->io_watcher, handle->fd, ev_events);
+  ev_io_start(handle->loop->ev, &handle->io_watcher);
+
+  if (!was_active)
+    uv_unref(handle->loop);
+
+  handle->poll_cb = poll_cb;
+
+  return 0;
+}
+
+
+void uv__poll_close(uv_poll_t* handle) {
+  uv__poll_stop(handle);
+}
+
+
+int uv__poll_active(const uv_poll_t* handle) {
+  return ev_is_active(&handle->io_watcher);
+}

src/win/core.c

@@ -84,6 +84,8 @@ static void uv_loop_init(uv_loop_t* loop) {
   loop->next_check_handle = NULL;
   loop->next_idle_handle = NULL;
 
+  memset(&loop->poll_peer_sockets, 0, sizeof loop->poll_peer_sockets);
+
   loop->ares_active_sockets = 0;
   loop->ares_chan = NULL;
 
@@ -130,6 +132,14 @@ uv_loop_t* uv_loop_new(void) {
 
 void uv_loop_delete(uv_loop_t* loop) {
   if (loop != &uv_default_loop_) {
+    int i;
+    for (i = 0; i < ARRAY_SIZE(loop->poll_peer_sockets); i++) {
+      SOCKET sock = loop->poll_peer_sockets[i];
+      if (sock != 0 && sock != INVALID_SOCKET) {
+        closesocket(sock);
+      }
+    }
+
     free(loop);
   }
 }

src/win/handle.c

@@ -64,6 +64,9 @@ int uv_is_active(const uv_handle_t* handle) {
     case UV_CHECK:
       return (handle->flags & UV_HANDLE_ACTIVE) ? 1 : 0;
 
+    case UV_POLL:
+      return ((uv_poll_t*) handle)->events != 0;
+
     default:
       return 1;
   }
@@ -112,6 +115,10 @@ void uv_close(uv_handle_t* handle, uv_close_cb cb) {
       }
       return;
 
+    case UV_POLL:
+      uv_poll_close(handle->loop, (uv_poll_t*) handle);
+      return;
+
     case UV_TIMER:
       uv_timer_stop((uv_timer_t*)handle);
       uv_want_endgame(loop, handle);
@@ -195,6 +202,10 @@ void uv_process_endgames(uv_loop_t* loop) {
         uv_udp_endgame(loop, (uv_udp_t*) handle);
         break;
 
+      case UV_POLL:
+        uv_poll_endgame(loop, (uv_poll_t*) handle);
+        break;
+
       case UV_TIMER:
         uv_timer_endgame(loop, (uv_timer_t*) handle);
         break;

src/win/internal.h

@@ -84,6 +84,9 @@ void uv_process_timers(uv_loop_t* loop);
 #define UV_HANDLE_TTY_SAVED_POSITION            0x02000000
 #define UV_HANDLE_TTY_SAVED_ATTRIBUTES          0x04000000
 
+/* Only used by uv_poll_t handles. */
+#define UV_HANDLE_POLL_SLOW                     0x02000000
+
 
 void uv_want_endgame(uv_loop_t* loop, uv_handle_t* handle);
 void uv_process_endgames(uv_loop_t* loop);
@@ -230,6 +233,16 @@ void uv_process_tty_connect_req(uv_loop_t* loop, uv_tty_t* handle,
 void uv_tty_endgame(uv_loop_t* loop, uv_tty_t* handle);
 
 
+/*
+ * Poll watchers
+ */
+void uv_process_poll_req(uv_loop_t* loop, uv_poll_t* handle,
+    uv_req_t* req);
+
+void uv_poll_close(uv_loop_t* loop, uv_poll_t* handle);
+void uv_poll_endgame(uv_loop_t* loop, uv_poll_t* handle);
+
+
 /*
  * Loop watchers
  */

src/win/poll.c

@@ -0,0 +1,609 @@
+/* Copyright Joyent, Inc. and other Node contributors. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+
+#include <assert.h>
+#include <io.h>
+
+#include "uv.h"
+#include "../uv-common.h"
+#include "internal.h"
+
+
+static const GUID uv_msafd_provider_ids[UV_MSAFD_PROVIDER_COUNT] = {
+  {0xe70f1aa0, 0xab8b, 0x11cf,
+      {0x8c, 0xa3, 0x00, 0x80, 0x5f, 0x48, 0xa1, 0x92}},
+  {0xf9eab0c0, 0x26d4, 0x11d0,
+      {0xbb, 0xbf, 0x00, 0xaa, 0x00, 0x6c, 0x34, 0xe4}},
+  {0x9fc48064, 0x7298, 0x43e4,
+      {0xb7, 0xbd, 0x18, 0x1f, 0x20, 0x89, 0x79, 0x2a}}
+};
+
+typedef struct uv_single_fd_set_s {
+  unsigned int fd_count;
+  SOCKET fd_array[1];
+} uv_single_fd_set_t;
+
+
+static void uv__fast_poll_submit_poll_req(uv_loop_t* loop, uv_poll_t* handle) {
+  uv_req_t* req;
+  AFD_POLL_INFO* afd_poll_info;
+  DWORD result;
+
+  /* Find a yet unsubmitted req to submit. */
+  if (handle->submitted_events_1 == 0) {
+    req = &handle->poll_req_1;
+    afd_poll_info = &handle->afd_poll_info_1;
+    handle->submitted_events_1 = handle->events;
+    handle->mask_events_1 = 0;
+    handle->mask_events_2 = handle->events;
+  } else if (handle->submitted_events_2 == 0) {
+    req = &handle->poll_req_2;
+    afd_poll_info = &handle->afd_poll_info_2;
+    handle->submitted_events_2 = handle->events;
+    handle->mask_events_1 = handle->events;
+    handle->mask_events_2 = 0;
+  } else {
+    assert(0);
+  }
+
+  /* Setting Exclusive to TRUE makes the other poll request return if there */
+  /* is any. */
+  afd_poll_info->Exclusive = TRUE;
+  afd_poll_info->NumberOfHandles = 1;
+  afd_poll_info->Timeout.QuadPart = INT64_MAX;
+  afd_poll_info->Handles[0].Handle = (HANDLE) handle->socket;
+  afd_poll_info->Handles[0].Status = 0;
+  afd_poll_info->Handles[0].Events = 0;
+
+  if (handle->events & UV_READABLE) {
+    afd_poll_info->Handles[0].Events |= AFD_POLL_RECEIVE |
+        AFD_POLL_DISCONNECT | AFD_POLL_ACCEPT | AFD_POLL_ABORT;
+  }
+  if (handle->events & UV_WRITABLE) {
+    afd_poll_info->Handles[0].Events |= AFD_POLL_SEND | AFD_POLL_CONNECT_FAIL;
+  }
+
+  memset(&req->overlapped, 0, sizeof req->overlapped);
+
+  result = uv_msafd_poll((SOCKET) handle->peer_socket,
+                         afd_poll_info,
+                         &req->overlapped);
+  if (result != 0 && WSAGetLastError() != WSA_IO_PENDING) {
+    /* Queue this req, reporting an error. */
+    SET_REQ_ERROR(req, WSAGetLastError());
+    uv_insert_pending_req(loop, req);
+  }
+}
+
+
+static int uv__fast_poll_cancel_poll_reqs(uv_loop_t* loop, uv_poll_t* handle) {
+  AFD_POLL_INFO afd_poll_info;
+  DWORD result;
+  HANDLE event;
+  OVERLAPPED overlapped;
+
+  event = CreateEvent(NULL, TRUE, FALSE, NULL);
+  if (event == NULL) {
+    uv__set_sys_error(loop, GetLastError());
+    return -1;
+  }
+
+  afd_poll_info.Exclusive = TRUE;
+  afd_poll_info.NumberOfHandles = 1;
+  afd_poll_info.Timeout.QuadPart = INT64_MAX;
+  afd_poll_info.Handles[0].Handle = (HANDLE) handle->socket;
+  afd_poll_info.Handles[0].Status = 0;
+  afd_poll_info.Handles[0].Events = AFD_POLL_ALL;
+
+  memset(&overlapped, 0, sizeof overlapped);
+  overlapped.hEvent = (HANDLE) ((uintptr_t) event & 1);
+
+  result = uv_msafd_poll(handle->socket,
+                         &afd_poll_info,
+                         &overlapped);
+
+  if (result == SOCKET_ERROR) {
+    DWORD error = WSAGetLastError();
+    if (error != WSA_IO_PENDING) {
+      uv__set_sys_error(loop, WSAGetLastError());
+      CloseHandle(event);
+      return -1;
+    }
+  }
+
+  CloseHandle(event);
+  return 0;
+}
+
+
+static void uv__fast_poll_process_poll_req(uv_loop_t* loop, uv_poll_t* handle,
+    uv_req_t* req) {
+  unsigned char mask_events;
+  AFD_POLL_INFO* afd_poll_info;
+
+  if (req == &handle->poll_req_1) {
+    afd_poll_info = &handle->afd_poll_info_1;
+    handle->submitted_events_1 = 0;
+    mask_events = handle->mask_events_1;
+  } else if (req == &handle->poll_req_2) {
+    afd_poll_info = &handle->afd_poll_info_2;
+    handle->submitted_events_2 = 0;
+    mask_events = handle->mask_events_2;
+  } else {
+    assert(0);
+  }
+
+  /* Report an error unless the select was just interrupted. */
+  if (!REQ_SUCCESS(req)) {
+    DWORD error = GET_REQ_SOCK_ERROR(req);
+    if (error != WSAEINTR && handle->events != 0) {
+      handle->events = 0; /* Stop the watcher */
+      uv__set_sys_error(loop, error);
+      handle->poll_cb(handle, -1, 0);
+    }
+
+  } else if (afd_poll_info->NumberOfHandles >= 1) {
+    unsigned char events = 0;
+
+    if ((afd_poll_info->Handles[0].Events & (AFD_POLL_RECEIVE |
+        AFD_POLL_DISCONNECT | AFD_POLL_ACCEPT | AFD_POLL_ABORT)) != 0) {
+      events |= UV_READABLE;
+    }
+    if ((afd_poll_info->Handles[0].Events & (AFD_POLL_SEND |
+        AFD_POLL_CONNECT_FAIL)) != 0) {
+      events |= UV_WRITABLE;
+    }
+
+    events &= handle->events & ~mask_events;
+
+    if (afd_poll_info->Handles[0].Events & AFD_POLL_LOCAL_CLOSE) {
+      /* Stop polling. */
+      handle->events = 0;
+    }
+
+    if (events != 0) {
+      handle->poll_cb(handle, 0, events);
+    }
+  }
+
+  if ((handle->events & ~(handle->submitted_events_1 |
+      handle->submitted_events_2)) != 0) {
+    uv__fast_poll_submit_poll_req(loop, handle);
+  } else if ((handle->flags & UV_HANDLE_CLOSING) &&
+             handle->submitted_events_1 == 0 &&
+             handle->submitted_events_2 == 0) {
+    uv_want_endgame(loop, (uv_handle_t*) handle);
+  }
+}
+
+
+static int uv__fast_poll_set(uv_loop_t* loop, uv_poll_t* handle, int events) {
+  assert(handle->type == UV_POLL);
+  assert(!(handle->flags & UV_HANDLE_CLOSING));
+  assert((events & ~(UV_READABLE | UV_WRITABLE)) == 0);
+
+  handle->events = events;
+  if ((handle->events & ~(handle->submitted_events_1 |
+      handle->submitted_events_2)) != 0) {
+    uv__fast_poll_submit_poll_req(handle->loop, handle);
+  }
+
+  return 0;
+}
+
+
+static void uv__fast_poll_close(uv_loop_t* loop, uv_poll_t* handle) {
+  handle->events = 0;
+
+  if (handle->submitted_events_1 == 0 &&
+      handle->submitted_events_2 == 0) {
+    uv_want_endgame(loop, (uv_handle_t*) handle);
+  } else {
+    /* Try to cancel outstanding poll requests. */
+    if (pCancelIoEx) {
+      /* Use CancelIoEx to cancel poll requests if available. */
+      if (handle->submitted_events_1)
+        pCancelIoEx((HANDLE) handle->socket, &handle->poll_req_1.overlapped);
+      if (handle->submitted_events_2)
+        pCancelIoEx((HANDLE) handle->socket, &handle->poll_req_2.overlapped);
+    } else if (handle->submitted_events_1 | handle->submitted_events_2) {
+      /* Execute another unique poll to force the others to return. */
+      uv__fast_poll_cancel_poll_reqs(loop, handle);
+    }
+  }
+}
+
+
+static SOCKET uv__fast_poll_create_peer_socket(HANDLE iocp,
+    WSAPROTOCOL_INFOW* protocol_info) {
+  SOCKET sock = 0;
+
+  sock = WSASocketW(protocol_info->iAddressFamily,
+                    protocol_info->iSocketType,
+                    protocol_info->iProtocol,
+                    protocol_info,
+                    0,
+                    WSA_FLAG_OVERLAPPED);
+  if (sock == INVALID_SOCKET) {
+    return INVALID_SOCKET;
+  }
+
+  if (!SetHandleInformation((HANDLE) sock, HANDLE_FLAG_INHERIT, 0)) {
+    goto error;
+  };
+
+  if (CreateIoCompletionPort((HANDLE) sock,
+                             iocp,
+                             (ULONG_PTR) sock,
+                             0) == NULL) {
+    goto error;
+  }
+
+  return sock;
+
+ error:
+  closesocket(sock);
+  return INVALID_SOCKET;
+}
+
+
+static SOCKET uv__fast_poll_get_peer_socket(uv_loop_t* loop,
+    WSAPROTOCOL_INFOW* protocol_info) {
+  int index, i;
+  SOCKET peer_socket;
+
+  index = -1;
+  for (i = 0; i < ARRAY_SIZE(uv_msafd_provider_ids); i++) {
+    if (memcmp((void*) &protocol_info->ProviderId,
+               (void*) &uv_msafd_provider_ids[i],
+               sizeof protocol_info->ProviderId) == 0) {
+      index = i;
+    }
+  }
+
+  /* Check if the protocol uses an msafd socket. */
+  if (index < 0) {
+    return INVALID_SOCKET;
+  }
+
+  /* If we didn't (try) to create a peer socket yet, try to make one. Don't */
+  /* try again if the peer socket creation failed earlier for the same */
+  /* protocol. */
+  peer_socket = loop->poll_peer_sockets[index];
+  if (peer_socket == 0) {
+    peer_socket = uv__fast_poll_create_peer_socket(loop->iocp, protocol_info);
+    loop->poll_peer_sockets[index] = peer_socket;
+  }
+
+  return peer_socket;
+}
+
+
+static DWORD WINAPI uv__slow_poll_thread_proc(void* arg) {
+  uv_req_t* req = (uv_req_t*) arg;
+  uv_poll_t* handle = (uv_poll_t*) req->data;
+  unsigned char events, reported_events;
+  int r;
+  uv_single_fd_set_t rfds, wfds, efds;
+  struct timeval timeout;
+
+  assert(handle->type == UV_POLL);
+  assert(req->type == UV_POLL_REQ);
+
+  if (req == &handle->poll_req_1) {
+    events = handle->submitted_events_1;
+  } else if (req == &handle->poll_req_2) {
+    events = handle->submitted_events_2;
+  } else {
+    assert(0);
+  }
+
+  if (handle->events & UV_READABLE) {
+    rfds.fd_count = 1;
+    rfds.fd_array[0] = handle->socket;
+  } else {
+    rfds.fd_count = 0;
+  }
+
+  if (handle->events & UV_WRITABLE) {
+    wfds.fd_count = 1;
+    wfds.fd_array[0] = handle->socket;
+    efds.fd_count = 1;
+    efds.fd_array[0] = handle->socket;
+  } else {
+    wfds.fd_count = 0;
+    efds.fd_count = 0;
+  }
+
+  /* Make the select() time out after 3 minutes. If select() hangs because */
+  /* the user closed the socket, we will at least not hang indefinitely. */
+  timeout.tv_sec = 3 * 60;
+  timeout.tv_usec = 0;
+
+  r = select(1, (fd_set*) &rfds, (fd_set*) &wfds, (fd_set*) &efds, &timeout);
+  if (r == SOCKET_ERROR) {
+    /* Queue this req, reporting an error. */
+    SET_REQ_ERROR(&handle->poll_req_1, WSAGetLastError());
+    POST_COMPLETION_FOR_REQ(handle->loop, req);
+    return 0;
+  }
+
+  reported_events = 0;
+
+  if (r > 0) {
+    if (rfds.fd_count > 0) {
+      assert(rfds.fd_count == 1);
+      assert(rfds.fd_array[0] == handle->socket);
+      reported_events |= UV_READABLE;
+    }
+
+    if (wfds.fd_count > 0) {
+      assert(wfds.fd_count == 1);
+      assert(wfds.fd_array[0] == handle->socket);
+      reported_events |= UV_WRITABLE;
+    } else if (efds.fd_count > 0) {
+      assert(efds.fd_count == 1);
+      assert(efds.fd_array[0] == handle->socket);
+      reported_events |= UV_WRITABLE;
+    }
+  }
+
+  SET_REQ_SUCCESS(req);
+  req->overlapped.InternalHigh = (DWORD) reported_events;
+  POST_COMPLETION_FOR_REQ(handle->loop, req);
+
+  return 0;
+}
+
+
+static void uv__slow_poll_submit_poll_req(uv_loop_t* loop, uv_poll_t* handle) {
+  uv_req_t* req;
+
+  /* Find a yet unsubmitted req to submit. */
+  if (handle->submitted_events_1 == 0) {
+    req = &handle->poll_req_1;
+    handle->submitted_events_1 = handle->events;
+    handle->mask_events_1 = 0;
+    handle->mask_events_2 = handle->events;
+  } else if (handle->submitted_events_2 == 0) {
+    req = &handle->poll_req_2;
+    handle->submitted_events_2 = handle->events;
+    handle->mask_events_1 = handle->events;
+    handle->mask_events_2 = 0;
+  } else {
+    assert(0);
+  }
+
+  if (!QueueUserWorkItem(uv__slow_poll_thread_proc,
+                         (void*) req,
+                         WT_EXECUTELONGFUNCTION)) {
+    /* Make this req pending, reporting an error. */
+    SET_REQ_ERROR(req, GetLastError());
+    uv_insert_pending_req(loop, req);
+  }
+}
+
+
+
+static void uv__slow_poll_process_poll_req(uv_loop_t* loop, uv_poll_t* handle,
+    uv_req_t* req) {
+  unsigned char mask_events;
+  if (req == &handle->poll_req_1) {
+    handle->submitted_events_1 = 0;
+    mask_events = handle->mask_events_1;
+  } else if (req == &handle->poll_req_2) {
+    handle->submitted_events_2 = 0;
+    mask_events = handle->mask_events_2;
+  } else {
+    assert(0);
+  }
+
+  if (!REQ_SUCCESS(req)) {
+    /* Error. */
+    if (handle->events != 0) {
+      handle->events = 0; /* Stop the watcher */
+      uv__set_sys_error(loop, GET_REQ_ERROR(req));
+      handle->poll_cb(handle, -1, 0);
+    }
+  } else {
+    /* Got some events. */
+    int events = req->overlapped.InternalHigh & handle->events & ~mask_events;
+    if (events != 0) {
+      handle->poll_cb(handle, 0, events);
+    }
+  }
+
+  if ((handle->events & ~(handle->submitted_events_1 |
+      handle->submitted_events_2)) != 0) {
+    uv__slow_poll_submit_poll_req(loop, handle);
+  } else if ((handle->flags & UV_HANDLE_CLOSING) &&
+             handle->submitted_events_1 == 0 &&
+             handle->submitted_events_2 == 0) {
+    uv_want_endgame(loop, (uv_handle_t*) handle);
+  }
+}
+
+
+static int uv__slow_poll_set(uv_loop_t* loop, uv_poll_t* handle, int events) {
+  assert(handle->type == UV_POLL);
+  assert(!(handle->flags & UV_HANDLE_CLOSING));
+  assert((events & ~(UV_READABLE | UV_WRITABLE)) == 0);
+
+  handle->events = events;
+  if ((handle->events &
+      ~(handle->submitted_events_1 | handle->submitted_events_2)) != 0) {
+    uv__slow_poll_submit_poll_req(handle->loop, handle);
+  }
+
+  return 0;
+}
+
+
+static void uv__slow_poll_close(uv_loop_t* loop, uv_poll_t* handle) {
+  handle->events = 0;
+
+  if (handle->submitted_events_1 == 0 &&
+      handle->submitted_events_2 == 0) {
+    uv_want_endgame(loop, (uv_handle_t*) handle);
+  }
+}
+
+
+int uv_poll_init(uv_loop_t* loop, uv_poll_t* handle, int fd) {
+  return uv_poll_init_socket(loop, handle, (SOCKET) _get_osfhandle(fd));
+}
+
+
+int uv_poll_init_socket(uv_loop_t* loop, uv_poll_t* handle,
+    uv_os_sock_t socket) {
+  WSAPROTOCOL_INFOW protocol_info;
+  int len;
+  SOCKET peer_socket, base_socket;
+  DWORD bytes;
+
+  /* Try to obtain a base handle for the socket. This increases this chances */
+  /* that we find an AFD handle and are able to use the fast poll mechanism. */
+  /* This will always fail on windows XP/2k3, since they don't support the */
+  /* SIO_BASE_HANDLE ioctl. */
+#ifndef NDEBUG
+  base_socket = INVALID_SOCKET;
+#endif
+
+  if (WSAIoctl(socket,
+               SIO_BASE_HANDLE,
+               NULL,
+               0,
+               &base_socket,
+               sizeof base_socket,
+               &bytes,
+               NULL,
+               NULL) == 0) {
+    assert(base_socket != 0 && base_socket != INVALID_SOCKET);
+    socket = base_socket;
+  }
+
+  handle->type = UV_POLL;
+  handle->socket = socket;
+  handle->loop = loop;
+  handle->flags = 0;
+  handle->events = 0;
+
+  /* Obtain protocol information about the socket. */
+  len = sizeof protocol_info;
+  if (getsockopt(socket,
+                 SOL_SOCKET,
+                 SO_PROTOCOL_INFOW,
+                 (char*) &protocol_info,
+                 &len) != 0) {
+    uv__set_sys_error(loop, WSAGetLastError());
+    return -1;
+  }
+
+  /* Get the peer socket that is needed to enable fast poll. If the returned */
+  /* value is NULL, the protocol is not implemented by MSAFD and we'll have */
+  /* to use slow mode. */
+  peer_socket = uv__fast_poll_get_peer_socket(loop, &protocol_info);
+
+  if (peer_socket != INVALID_SOCKET) {
+    /* Initialize fast poll specific fields. */
+    handle->peer_socket = peer_socket;
+  } else {
+    /* Initialize slow poll specific fields. */
+    handle->flags |= UV_HANDLE_POLL_SLOW;
+  }
+
+  /* Intialize 2 poll reqs. */
+  handle->submitted_events_1 = 0;
+  uv_req_init(loop, (uv_req_t*) &(handle->poll_req_1));
+  handle->poll_req_1.type = UV_POLL_REQ;
+  handle->poll_req_1.data = handle;
+
+  handle->submitted_events_2 = 0;
+  uv_req_init(loop, (uv_req_t*) &(handle->poll_req_2));
+  handle->poll_req_2.type = UV_POLL_REQ;
+  handle->poll_req_2.data = handle;
+
+  uv_ref(loop);
+
+  loop->counters.handle_init++;
+  loop->counters.poll_init++;
+
+  return 0;
+}
+
+
+int uv_poll_start(uv_poll_t* handle, int events, uv_poll_cb cb) {
+  if (!(handle->flags & UV_HANDLE_POLL_SLOW)) {
+    if (uv__fast_poll_set(handle->loop, handle, events) < 0)
+      return -1;
+  } else {
+    if (uv__slow_poll_set(handle->loop, handle, events) < 0)
+      return -1;
+  }
+
+  handle->poll_cb = cb;
+  return 0;
+}
+
+
+int uv_poll_stop(uv_poll_t* handle) {
+  if (!(handle->flags & UV_HANDLE_POLL_SLOW)) {
+    return uv__fast_poll_set(handle->loop, handle, 0);
+  } else {
+    return uv__slow_poll_set(handle->loop, handle, 0);
+  }
+}
+
+
+void uv_process_poll_req(uv_loop_t* loop, uv_poll_t* handle, uv_req_t* req) {
+  if (!(handle->flags & UV_HANDLE_POLL_SLOW)) {
+    uv__fast_poll_process_poll_req(loop, handle, req);
+  } else {
+    uv__slow_poll_process_poll_req(loop, handle, req);
+  }
+}
+
+
+void uv_poll_close(uv_loop_t* loop, uv_poll_t* handle) {
+  if (!(handle->flags & UV_HANDLE_POLL_SLOW)) {
+    uv__fast_poll_close(loop, handle);
+  } else {
+    uv__slow_poll_close(loop, handle);
+  }
+}
+
+
+void uv_poll_endgame(uv_loop_t* loop, uv_poll_t* handle) {
+  assert(handle->flags & UV_HANDLE_CLOSING);
+  assert(!(handle->flags & UV_HANDLE_CLOSED));
+
+  assert(handle->submitted_events_1 == 0);
+  assert(handle->submitted_events_2 == 0);
+
+  handle->flags |= UV_HANDLE_CLOSED;
+
+  if (handle->close_cb) {
+    handle->close_cb((uv_handle_t*)handle);
+  }
+
+  uv_unref(loop);
+}

src/win/req.c

@@ -135,6 +135,10 @@ void uv_process_reqs(uv_loop_t* loop) {
         uv_process_async_wakeup_req(loop, (uv_async_t*) req->data, req);
         break;
 
+      case UV_POLL_REQ:
+        uv_process_poll_req(loop, (uv_poll_t*) req->data, req);
+        break;
+
       case UV_ARES_EVENT_REQ:
         uv_process_ares_event_req(loop, (uv_ares_action_t*) req->data, req);
         break;

test/benchmark-sizes.c

@@ -36,5 +36,6 @@ BENCHMARK_IMPL(sizes) {
   LOGF("uv_async_t: %u bytes\n", (unsigned int) sizeof(uv_async_t));
   LOGF("uv_timer_t: %u bytes\n", (unsigned int) sizeof(uv_timer_t));
   LOGF("uv_process_t: %u bytes\n", (unsigned int) sizeof(uv_process_t));
+  LOGF("uv_poll_t: %u bytes\n", (unsigned int) sizeof(uv_poll_t));
   return 0;
 }

test/test-list.h

@@ -159,6 +159,8 @@ TEST_DECLARE   (strlcpy)
 TEST_DECLARE   (strlcat)
 TEST_DECLARE   (counters_init)
 TEST_DECLARE   (dlerror)
+TEST_DECLARE   (poll_duplex)
+TEST_DECLARE   (poll_unidirectional)
 #ifdef _WIN32
 TEST_DECLARE   (spawn_detect_pipe_name_collisions_on_windows)
 TEST_DECLARE   (argument_escaping)
@@ -314,6 +316,9 @@ TASK_LIST_START
   TEST_ENTRY  (getsockname_tcp)
   TEST_ENTRY  (getsockname_udp)
 
+  TEST_ENTRY  (poll_duplex)
+  TEST_ENTRY  (poll_unidirectional)
+
   TEST_ENTRY  (spawn_exit_code)
   TEST_ENTRY  (spawn_stdout)
   TEST_ENTRY  (spawn_stdin)

test/test-poll.c

@@ -0,0 +1,573 @@
+/* Copyright Joyent, Inc. and other Node contributors. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <errno.h>
+
+#ifndef _WIN32
+# include <fcntl.h>
+# include <sys/socket.h>
+# include <unistd.h>
+#endif
+
+#include "uv.h"
+#include "task.h"
+
+
+#define NUM_CLIENTS 5
+#define TRANSFER_BYTES (1 << 16)
+
+#undef MIN
+#define MIN(a, b) (((a) < (b)) ? (a) : (b));
+
+
+typedef enum {
+  UNIDIRECTIONAL,
+  DUPLEX
+} test_mode_t;
+
+typedef struct connection_context_s {
+  uv_poll_t poll_handle;
+  uv_timer_t timer_handle;
+  uv_os_sock_t sock;
+  size_t read, sent;
+  int is_server_connection;
+  int open_handles;
+  int got_fin, sent_fin;
+  unsigned int events, delayed_events;
+} connection_context_t;
+
+typedef struct server_context_s {
+  uv_poll_t poll_handle;
+  uv_os_sock_t sock;
+  int connections;
+} server_context_t;
+
+
+static void delay_timer_cb(uv_timer_t* timer, int status);
+
+
+static test_mode_t test_mode = DUPLEX;
+
+static int closed_connections = 0;
+
+static int valid_writable_wakeups = 0;
+static int spurious_writable_wakeups = 0;
+
+
+static int got_eagain() {
+#ifdef _WIN32
+  return WSAGetLastError() == WSAEWOULDBLOCK;
+#else
+  return errno == EAGAIN
+      || errno == EINPROGRESS
+#ifdef EWOULDBLOCK
+      || errno == EWOULDBLOCK;
+#endif
+      ;
+#endif
+}
+
+
+static void set_nonblocking(uv_os_sock_t sock) {
+  int r;
+#ifdef _WIN32
+  unsigned long on = 1;
+  r = ioctlsocket(sock, FIONBIO, &on);
+  ASSERT(r == 0);
+#else
+  int flags = fcntl(sock, F_GETFL, 0);
+  ASSERT(flags >= 0);
+  r = fcntl(sock, F_SETFL, flags | O_NONBLOCK);
+  ASSERT(r >= 0);
+#endif
+}
+
+
+static uv_os_sock_t create_nonblocking_bound_socket(
+    struct sockaddr_in bind_addr) {
+  uv_os_sock_t sock;
+  int r;
+
+  sock = socket(AF_INET, SOCK_STREAM, IPPROTO_IP);
+#ifdef _WIN32
+  ASSERT(sock != INVALID_SOCKET);
+#else
+  ASSERT(sock >= 0);
+#endif
+
+  set_nonblocking(sock);
+
+#ifndef _WIN32
+  {
+    /* Allow reuse of the port. */
+    int yes = 1;
+    r = setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, &yes, sizeof yes);
+    ASSERT(r == 0);
+  }
+#endif
+
+  r = bind(sock, (const struct sockaddr*) &bind_addr, sizeof bind_addr);
+  ASSERT(r == 0);
+
+  return sock;
+}
+
+
+static void close_socket(uv_os_sock_t sock) {
+  int r;
+#ifdef _WIN32
+  r = closesocket(sock);
+#else
+  r = close(sock);
+#endif
+  ASSERT(r == 0);
+}
+
+
+static connection_context_t* create_connection_context(
+    uv_os_sock_t sock, int is_server_connection) {
+  int r;
+  connection_context_t* context;
+
+  context = (connection_context_t*) malloc(sizeof *context);
+  ASSERT(context != NULL);
+
+  context->sock = sock;
+  context->is_server_connection = is_server_connection;
+  context->read = 0;
+  context->sent = 0;
+  context->open_handles = 0;
+  context->events = 0;
+  context->delayed_events = 0;
+  context->got_fin = 0;
+  context->sent_fin = 0;
+
+  r = uv_poll_init_socket(uv_default_loop(), &context->poll_handle, sock);
+  context->open_handles++;
+  context->poll_handle.data = context;
+  ASSERT(r == 0);
+
+  r = uv_timer_init(uv_default_loop(), &context->timer_handle);
+  context->open_handles++;
+  context->timer_handle.data = context;
+  ASSERT(r == 0);
+
+  return context;
+}
+
+
+static void connection_close_cb(uv_handle_t* handle) {
+  connection_context_t* context = (connection_context_t*) handle->data;
+
+  if (--context->open_handles == 0) {
+    if (test_mode == DUPLEX || context->is_server_connection) {
+      ASSERT(context->read == TRANSFER_BYTES);
+    } else {
+      ASSERT(context->read == 0);
+    }
+
+    if (test_mode == DUPLEX || !context->is_server_connection) {
+      ASSERT(context->sent == TRANSFER_BYTES);
+    } else {
+      ASSERT(context->sent == 0);
+    }
+
+    closed_connections++;
+
+    free(context);
+  }
+}
+
+
+static void destroy_connection_context(connection_context_t* context) {
+  uv_close((uv_handle_t*) &context->poll_handle, connection_close_cb);
+  uv_close((uv_handle_t*) &context->timer_handle, connection_close_cb);
+}
+
+
+static void connection_poll_cb(uv_poll_t* handle, int status, int events) {
+  connection_context_t* context = (connection_context_t*) handle->data;
+  int new_events;
+  int r;
+
+  ASSERT(status == 0);
+  ASSERT(events & context->events);
+  ASSERT(!(events & ~context->events));
+
+  new_events = context->events;
+
+  if (events & UV_READABLE) {
+    int action = rand() % 7;
+
+    switch (action) {
+      case 0:
+      case 1: {
+        /* Read a couple of bytes. */
+        static char buffer[74];
+        r = recv(context->sock, buffer, sizeof buffer, 0);
+        ASSERT(r >= 0);
+
+        if (r > 0) {
+          context->read += r;
+        } else {
+          /* Got FIN. */
+          context->got_fin = 1;
+          new_events &= ~UV_READABLE;
+        }
+
+        break;
+      }
+
+      case 2:
+      case 3: {
+        /* Read until EAGAIN. */
+        static char buffer[931];
+        r = recv(context->sock, buffer, sizeof buffer, 0);
+        ASSERT(r >= 0);
+
+        while (r > 0) {
+          context->read += r;
+          r = recv(context->sock, buffer, sizeof buffer, 0);
+        }
+
+        if (r == 0) {
+          /* Got FIN. */
+          context->got_fin = 1;
+          new_events &= ~UV_READABLE;
+        } else {
+          ASSERT(got_eagain());
+        }
+
+        break;
+      }
+
+      case 4:
+        /* Ignore. */
+        break;
+
+      case 5:
+        /* Stop reading for a while. Restart in timer callback. */
+        new_events &= ~UV_READABLE;
+        if (!uv_is_active((uv_handle_t*) &context->timer_handle)) {
+          context->delayed_events = UV_READABLE;
+          uv_timer_start(&context->timer_handle, delay_timer_cb, 10, 0);
+        } else {
+          context->delayed_events |= UV_READABLE;
+        }
+        break;
+
+      case 6:
+        /* Fudge with the event mask. */
+        uv_poll_start(&context->poll_handle, UV_WRITABLE, connection_poll_cb);
+        uv_poll_start(&context->poll_handle, UV_READABLE, connection_poll_cb);
+        context->events = UV_READABLE;
+        break;
+
+      default:
+        ASSERT(0);
+    }
+  }
+
+  if (events & UV_WRITABLE) {
+    if (context->sent < TRANSFER_BYTES &&
+        !(test_mode == UNIDIRECTIONAL && context->is_server_connection)) {
+      /* We have to send more bytes. */
+      int action = rand() % 7;
+
+      switch (action) {
+        case 0:
+        case 1: {
+          /* Send a couple of bytes. */
+          static char buffer[103];
+
+          int send_bytes = MIN(TRANSFER_BYTES - context->sent, sizeof buffer);
+          ASSERT(send_bytes > 0);
+
+          r = send(context->sock, buffer, send_bytes, 0);
+
+          if (r < 0) {
+            ASSERT(got_eagain());
+            spurious_writable_wakeups++;
+            break;
+          }
+
+          ASSERT(r > 0);
+          context->sent += r;
+          valid_writable_wakeups++;
+          break;
+        }
+
+        case 2:
+        case 3: {
+          /* Send until EAGAIN. */
+          static char buffer[1234];
+
+          int send_bytes = MIN(TRANSFER_BYTES - context->sent, sizeof buffer);
+          ASSERT(send_bytes > 0);
+
+          r = send(context->sock, buffer, send_bytes, 0);
+
+          if (r < 0) {
+            ASSERT(got_eagain());
+            spurious_writable_wakeups++;
+            break;
+          }
+
+          ASSERT(r > 0);
+          valid_writable_wakeups++;
+          context->sent += r;
+
+          while (context->sent < TRANSFER_BYTES) {
+            send_bytes = MIN(TRANSFER_BYTES - context->sent, sizeof buffer);
+            ASSERT(send_bytes > 0);
+
+            r = send(context->sock, buffer, send_bytes, 0);
+
+            if (r <= 0) break;
+            context->sent += r;
+          }
+          ASSERT(r > 0 || got_eagain());
+          break;
+        }
+
+        case 4:
+          /* Ignore. */
+         break;
+
+        case 5:
+          /* Stop sending for a while. Restart in timer callback. */
+          new_events &= ~UV_WRITABLE;
+          if (!uv_is_active((uv_handle_t*) &context->timer_handle)) {
+            context->delayed_events = UV_WRITABLE;
+            uv_timer_start(&context->timer_handle, delay_timer_cb, 100, 0);
+          } else {
+            context->delayed_events |= UV_WRITABLE;
+          }
+          break;
+
+        case 6:
+          /* Fudge with the event mask. */
+          uv_poll_start(&context->poll_handle,
+                        UV_READABLE,
+                        connection_poll_cb);
+          uv_poll_start(&context->poll_handle,
+                        UV_WRITABLE,
+                        connection_poll_cb);
+          context->events = UV_WRITABLE;
+          break;
+
+        default:
+          ASSERT(0);
+      }
+
+    } else {
+      /* Nothing more to write. Send FIN. */
+      int r;
+#ifdef _WIN32
+      r = shutdown(context->sock, SD_SEND);
+#else
+      r = shutdown(context->sock, SHUT_WR);
+#endif
+      ASSERT(r == 0);
+      context->sent_fin = 1;
+      new_events &= ~UV_WRITABLE;
+    }
+  }
+
+  if (context->got_fin && context->sent_fin) {
+    /* Sent and received FIN. Close and destroy context. */
+    close_socket(context->sock);
+    destroy_connection_context(context);
+    context->events = 0;
+
+  } else if (new_events != context->events) {
+    /* Poll mask changed. Call uv_poll_start again. */
+    context->events = new_events;
+    uv_poll_start(handle, new_events, connection_poll_cb);
+  }
+
+  /* Assert that uv_is_active works correctly for poll handles. */
+  if (context->events != 0) {
+    ASSERT(uv_is_active((uv_handle_t*) handle));
+  } else {
+    ASSERT(!uv_is_active((uv_handle_t*) handle));
+  }
+}
+
+
+static void delay_timer_cb(uv_timer_t* timer, int status) {
+  connection_context_t* context = (connection_context_t*) timer->data;
+  int r;
+
+  /* Timer should auto stop. */
+  ASSERT(!uv_is_active((uv_handle_t*) timer));
+
+  /* Add the requested events to the poll mask. */
+  ASSERT(context->delayed_events != 0);
+  context->events |= context->delayed_events;
+  context->delayed_events = 0;
+
+  r = uv_poll_start(&context->poll_handle,
+                    context->events,
+                    connection_poll_cb);
+  ASSERT(r == 0);
+}
+
+
+static server_context_t* create_server_context(
+    uv_os_sock_t sock) {
+  int r;
+  server_context_t* context;
+
+  context = (server_context_t*) malloc(sizeof *context);
+  ASSERT(context != NULL);
+
+  context->sock = sock;
+  context->connections = 0;
+
+  r = uv_poll_init_socket(uv_default_loop(), &context->poll_handle, sock);
+  context->poll_handle.data = context;
+  ASSERT(r == 0);
+
+  return context;
+}
+
+
+static void server_close_cb(uv_handle_t* handle) {
+  server_context_t* context = (server_context_t*) handle->data;
+  free(context);
+}
+
+
+static void destroy_server_context(server_context_t* context) {
+  uv_close((uv_handle_t*) &context->poll_handle, server_close_cb);
+}
+
+
+static void server_poll_cb(uv_poll_t* handle, int status, int events) {
+  server_context_t* server_context = (server_context_t*)
+                                          handle->data;
+  connection_context_t* connection_context;
+  struct sockaddr_in addr;
+  socklen_t addr_len;
+  uv_os_sock_t sock;
+  int r;
+
+  addr_len = sizeof addr;
+  sock = accept(server_context->sock, (struct sockaddr*) &addr, &addr_len);
+#ifdef _WIN32
+  ASSERT(sock != INVALID_SOCKET);
+#else
+  ASSERT(sock >= 0);
+#endif
+
+  set_nonblocking(sock);
+
+  connection_context = create_connection_context(sock, 1);
+  connection_context->events = UV_READABLE | UV_WRITABLE;
+  r = uv_poll_start(&connection_context->poll_handle,
+                    UV_READABLE | UV_WRITABLE,
+                    connection_poll_cb);
+  ASSERT(r == 0);
+
+  if (++server_context->connections == NUM_CLIENTS) {
+    close_socket(server_context->sock);
+    destroy_server_context(server_context);
+  }
+}
+
+
+static void start_server() {
+  uv_os_sock_t sock;
+  server_context_t* context;
+  int r;
+
+  sock = create_nonblocking_bound_socket(uv_ip4_addr("127.0.0.1", TEST_PORT));
+  context = create_server_context(sock);
+
+  r = listen(sock, 100);
+  ASSERT(r == 0);
+
+  r = uv_poll_start(&context->poll_handle, UV_READABLE, server_poll_cb);
+  ASSERT(r == 0);
+}
+
+
+static void start_client() {
+  uv_os_sock_t sock;
+  connection_context_t* context;
+  struct sockaddr_in server_addr = uv_ip4_addr("127.0.0.1", TEST_PORT);
+  int r;
+
+  sock = create_nonblocking_bound_socket(uv_ip4_addr("0.0.0.0", 0));
+  context = create_connection_context(sock, 0);
+
+  context->events = UV_READABLE | UV_WRITABLE;
+  r = uv_poll_start(&context->poll_handle,
+                    UV_READABLE | UV_WRITABLE,
+                    connection_poll_cb);
+  ASSERT(r == 0);
+
+  r = connect(sock, (struct sockaddr*) &server_addr, sizeof server_addr);
+  ASSERT(r == 0 || got_eagain());
+}
+
+
+static void start_poll_test() {
+  int i, r;
+
+#ifdef _WIN32
+  {
+    struct WSAData wsa_data;
+    r = WSAStartup(MAKEWORD(2, 2), &wsa_data);
+    ASSERT(r == 0);
+  }
+#endif
+
+  start_server();
+
+  for (i = 0; i < NUM_CLIENTS; i++)
+    start_client();
+
+  r = uv_run(uv_default_loop());
+  ASSERT(r == 0);
+
+  /* Assert that at most one percent of the writable wakeups was spurious. */
+  ASSERT(spurious_writable_wakeups == 0 ||
+         (valid_writable_wakeups + spurious_writable_wakeups) /
+         spurious_writable_wakeups > 100);
+
+  ASSERT(closed_connections == NUM_CLIENTS * 2);
+}
+
+
+TEST_IMPL(poll_duplex) {
+  test_mode = DUPLEX;
+  start_poll_test();
+  return 0;
+}
+
+
+TEST_IMPL(poll_unidirectional) {
+  test_mode = UNIDIRECTIONAL;
+  start_poll_test();
+  return 0;
+}

uv.gyp

@@ -145,6 +145,7 @@
             'src/win/loop-watcher.c',
             'src/win/pipe.c',
             'src/win/thread.c',
+            'src/win/poll.c',
             'src/win/process.c',
             'src/win/req.c',
             'src/win/stream.c',
@@ -198,6 +199,7 @@
             'src/unix/internal.h',
             'src/unix/loop.c',
             'src/unix/pipe.c',
+            'src/unix/poll.c',
             'src/unix/prepare.c',
             'src/unix/process.c',
             'src/unix/stream.c',
@@ -326,6 +328,7 @@
         'test/test-pipe-bind-error.c',
         'test/test-pipe-connect-error.c',
         'test/test-platform-output.c',
+        'test/test-poll.c',
         'test/test-process-title.c',
         'test/test-ref.c',
         'test/test-shutdown-close.c',