diff --git a/src/unix/udp.c b/src/unix/udp.c
index d99fe104..d8b07e18 100644
--- a/src/unix/udp.c
+++ b/src/unix/udp.c
@@ -38,10 +38,9 @@
 
 
 static void uv__udp_run_completed(uv_udp_t* handle);
-static void uv__udp_run_pending(uv_udp_t* handle);
 static void uv__udp_io(uv_loop_t* loop, uv__io_t* w, unsigned int revents);
-static void uv__udp_recvmsg(uv_loop_t* loop, uv__io_t* w, unsigned int revents);
-static void uv__udp_sendmsg(uv_loop_t* loop, uv__io_t* w, unsigned int revents);
+static void uv__udp_recvmsg(uv_udp_t* handle);
+static void uv__udp_sendmsg(uv_udp_t* handle);
 static int uv__udp_maybe_deferred_bind(uv_udp_t* handle,
                                        int domain,
                                        unsigned int flags);
@@ -65,25 +64,19 @@ void uv__udp_finish_close(uv_udp_t* handle) {
   assert(!uv__io_active(&handle->io_watcher, UV__POLLIN | UV__POLLOUT));
   assert(handle->io_watcher.fd == -1);
 
-  uv__udp_run_completed(handle);
-
   while (!QUEUE_EMPTY(&handle->write_queue)) {
     q = QUEUE_HEAD(&handle->write_queue);
     QUEUE_REMOVE(q);
 
     req = QUEUE_DATA(q, uv_udp_send_t, queue);
-    uv__req_unregister(handle->loop, req);
-
-    if (req->bufs != req->bufsml)
-      free(req->bufs);
-    req->bufs = NULL;
-
-    if (req->send_cb != NULL)
-      req->send_cb(req, -ECANCELED);
+    req->status = -ECANCELED;
+    QUEUE_INSERT_TAIL(&handle->write_completed_queue, &req->queue);
   }
 
-  handle->send_queue_size = 0;
-  handle->send_queue_count = 0;
+  uv__udp_run_completed(handle);
+
+  assert(handle->send_queue_size == 0);
+  assert(handle->send_queue_count == 0);
 
   /* Now tear down the handle. */
   handle->recv_cb = NULL;
@@ -92,52 +85,6 @@ void uv__udp_finish_close(uv_udp_t* handle) {
 }
 
 
-static void uv__udp_run_pending(uv_udp_t* handle) {
-  uv_udp_send_t* req;
-  QUEUE* q;
-  struct msghdr h;
-  ssize_t size;
-
-  while (!QUEUE_EMPTY(&handle->write_queue)) {
-    q = QUEUE_HEAD(&handle->write_queue);
-    assert(q != NULL);
-
-    req = QUEUE_DATA(q, uv_udp_send_t, queue);
-    assert(req != NULL);
-
-    memset(&h, 0, sizeof h);
-    h.msg_name = &req->addr;
-    h.msg_namelen = (req->addr.ss_family == AF_INET6 ?
-      sizeof(struct sockaddr_in6) : sizeof(struct sockaddr_in));
-    h.msg_iov = (struct iovec*) req->bufs;
-    h.msg_iovlen = req->nbufs;
-
-    do {
-      size = sendmsg(handle->io_watcher.fd, &h, 0);
-    }
-    while (size == -1 && errno == EINTR);
-
-    /* TODO try to write once or twice more in the
-     * hope that the socket becomes readable again?
-     */
-    if (size == -1 && (errno == EAGAIN || errno == EWOULDBLOCK))
-      break;
-
-    req->status = (size == -1 ? -errno : size);
-
-    /* Sending a datagram is an atomic operation: either all data
-     * is written or nothing is (and EMSGSIZE is raised). That is
-     * why we don't handle partial writes. Just pop the request
-     * off the write queue and onto the completed queue, done.
-     */
-    handle->send_queue_size -= uv__count_bufs(req->bufs, req->nbufs);
-    handle->send_queue_count--;
-    QUEUE_REMOVE(&req->queue);
-    QUEUE_INSERT_TAIL(&handle->write_completed_queue, &req->queue);
-  }
-}
-
-
 static void uv__udp_run_completed(uv_udp_t* handle) {
   uv_udp_send_t* req;
   QUEUE* q;
@@ -149,6 +96,9 @@ static void uv__udp_run_completed(uv_udp_t* handle) {
     req = QUEUE_DATA(q, uv_udp_send_t, queue);
     uv__req_unregister(handle->loop, req);
 
+    handle->send_queue_size -= uv__count_bufs(req->bufs, req->nbufs);
+    handle->send_queue_count--;
+
     if (req->bufs != req->bufsml)
       free(req->bufs);
     req->bufs = NULL;
@@ -164,33 +114,40 @@ static void uv__udp_run_completed(uv_udp_t* handle) {
     else
       req->send_cb(req, req->status);
   }
+
+  if (QUEUE_EMPTY(&handle->write_queue)) {
+    /* Pending queue and completion queue empty, stop watcher. */
+    uv__io_stop(handle->loop, &handle->io_watcher, UV__POLLOUT);
+    if (!uv__io_active(&handle->io_watcher, UV__POLLIN))
+      uv__handle_stop(handle);
+  }
 }
 
 
 static void uv__udp_io(uv_loop_t* loop, uv__io_t* w, unsigned int revents) {
-  if (revents & UV__POLLIN)
-    uv__udp_recvmsg(loop, w, revents);
+  uv_udp_t* handle;
 
-  if (revents & UV__POLLOUT)
-    uv__udp_sendmsg(loop, w, revents);
+  handle = container_of(w, uv_udp_t, io_watcher);
+  assert(handle->type == UV_UDP);
+
+  if (revents & UV__POLLIN)
+    uv__udp_recvmsg(handle);
+
+  if (revents & UV__POLLOUT) {
+    uv__udp_sendmsg(handle);
+    uv__udp_run_completed(handle);
+  }
 }
 
 
-static void uv__udp_recvmsg(uv_loop_t* loop,
-                            uv__io_t* w,
-                            unsigned int revents) {
+static void uv__udp_recvmsg(uv_udp_t* handle) {
   struct sockaddr_storage peer;
   struct msghdr h;
-  uv_udp_t* handle;
   ssize_t nread;
   uv_buf_t buf;
   int flags;
   int count;
 
-  handle = container_of(w, uv_udp_t, io_watcher);
-  assert(handle->type == UV_UDP);
-  assert(revents & UV__POLLIN);
-
   assert(handle->recv_cb != NULL);
   assert(handle->alloc_cb != NULL);
 
@@ -247,35 +204,47 @@ static void uv__udp_recvmsg(uv_loop_t* loop,
 }
 
 
-static void uv__udp_sendmsg(uv_loop_t* loop,
-                            uv__io_t* w,
-                            unsigned int revents) {
-  uv_udp_t* handle;
-
-  handle = container_of(w, uv_udp_t, io_watcher);
-  assert(handle->type == UV_UDP);
-  assert(revents & UV__POLLOUT);
+static void uv__udp_sendmsg(uv_udp_t* handle) {
+  uv_udp_send_t* req;
+  QUEUE* q;
+  struct msghdr h;
+  ssize_t size;
 
   assert(!QUEUE_EMPTY(&handle->write_queue)
       || !QUEUE_EMPTY(&handle->write_completed_queue));
 
-  /* Write out pending data first. */
-  uv__udp_run_pending(handle);
+  while (!QUEUE_EMPTY(&handle->write_queue)) {
+    q = QUEUE_HEAD(&handle->write_queue);
+    assert(q != NULL);
 
-  /* Drain 'request completed' queue. */
-  uv__udp_run_completed(handle);
+    req = QUEUE_DATA(q, uv_udp_send_t, queue);
+    assert(req != NULL);
 
-  if (!QUEUE_EMPTY(&handle->write_completed_queue)) {
-    /* Schedule completion callbacks. */
+    memset(&h, 0, sizeof h);
+    h.msg_name = &req->addr;
+    h.msg_namelen = (req->addr.ss_family == AF_INET6 ?
+      sizeof(struct sockaddr_in6) : sizeof(struct sockaddr_in));
+    h.msg_iov = (struct iovec*) req->bufs;
+    h.msg_iovlen = req->nbufs;
+
+    do {
+      size = sendmsg(handle->io_watcher.fd, &h, 0);
+    } while (size == -1 && errno == EINTR);
+
+    if (size == -1 && (errno == EAGAIN || errno == EWOULDBLOCK))
+      break;
+
+    req->status = (size == -1 ? -errno : size);
+
+    /* Sending a datagram is an atomic operation: either all data
+     * is written or nothing is (and EMSGSIZE is raised). That is
+     * why we don't handle partial writes. Just pop the request
+     * off the write queue and onto the completed queue, done.
+     */
+    QUEUE_REMOVE(&req->queue);
+    QUEUE_INSERT_TAIL(&handle->write_completed_queue, &req->queue);
     uv__io_feed(handle->loop, &handle->io_watcher);
   }
-  else if (QUEUE_EMPTY(&handle->write_queue)) {
-    /* Pending queue and completion queue empty, stop watcher. */
-    uv__io_stop(loop, &handle->io_watcher, UV__POLLOUT);
-
-    if (!uv__io_active(&handle->io_watcher, UV__POLLIN))
-      uv__handle_stop(handle);
-  }
 }
 
 
@@ -415,6 +384,7 @@ int uv__udp_send(uv_udp_send_t* req,
                  unsigned int addrlen,
                  uv_udp_send_cb send_cb) {
   int err;
+  int empty_queue;
 
   assert(nbufs > 0);
 
@@ -422,8 +392,13 @@ int uv__udp_send(uv_udp_send_t* req,
   if (err)
     return err;
 
-  uv__req_init(handle->loop, req, UV_UDP_SEND);
+  /* It's legal for send_queue_count > 0 even when the write_queue is empty;
+   * it means there are error-state requests in the write_completed_queue that
+   * will touch up send_queue_size/count later.
+   */
+  empty_queue = (handle->send_queue_count == 0);
 
+  uv__req_init(handle->loop, req, UV_UDP_SEND);
   assert(addrlen <= sizeof(req->addr));
   memcpy(&req->addr, addr, addrlen);
   req->send_cb = send_cb;
@@ -441,9 +416,13 @@ int uv__udp_send(uv_udp_send_t* req,
   handle->send_queue_size += uv__count_bufs(req->bufs, req->nbufs);
   handle->send_queue_count++;
   QUEUE_INSERT_TAIL(&handle->write_queue, &req->queue);
-  uv__io_start(handle->loop, &handle->io_watcher, UV__POLLOUT);
   uv__handle_start(handle);
 
+  if (empty_queue)
+    uv__udp_sendmsg(handle);
+  else
+    uv__io_start(handle->loop, &handle->io_watcher, UV__POLLOUT);
+
   return 0;
 }