17 files changed, 639 insertions, 88 deletions

diff --git a/src/core/lib/iomgr/error.c b/src/core/lib/iomgr/error.c
index e366961936..31c80260f8 100644
--- a/src/core/lib/iomgr/error.c
+++ b/src/core/lib/iomgr/error.c
@@ -324,6 +324,64 @@ const char *grpc_error_get_str(grpc_error *err, grpc_error_strs which) {
   return gpr_avl_get(err->strs, (void *)(uintptr_t)which);
 }
 
+typedef struct {
+  grpc_error *error;
+  grpc_status_code code;
+  const char *msg;
+} special_error_status_map;
+static special_error_status_map error_status_map[] = {
+    {GRPC_ERROR_NONE, GRPC_STATUS_OK, ""},
+    {GRPC_ERROR_CANCELLED, GRPC_STATUS_CANCELLED, "RPC cancelled"},
+    {GRPC_ERROR_OOM, GRPC_STATUS_RESOURCE_EXHAUSTED, "Out of memory"},
+};
+
+static grpc_error *recursively_find_error_with_status(grpc_error *error,
+                                                      intptr_t *status) {
+  // If the error itself has a status code, return it.
+  if (grpc_error_get_int(error, GRPC_ERROR_INT_GRPC_STATUS, status)) {
+    return error;
+  }
+  // Otherwise, search through its children.
+  intptr_t key = 0;
+  while (true) {
+    grpc_error *child_error = gpr_avl_get(error->errs, (void *)key++);
+    if (child_error == NULL) break;
+    grpc_error *result =
+        recursively_find_error_with_status(child_error, status);
+    if (result != NULL) return result;
+  }
+  return NULL;
+}
+
+void grpc_error_get_status(grpc_error *error, grpc_status_code *code,
+                           const char **msg) {
+  // Handle special errors via the static map.
+  for (size_t i = 0; i < GPR_ARRAY_SIZE(error_status_map); ++i) {
+    if (error == error_status_map[i].error) {
+      *code = error_status_map[i].code;
+      *msg = error_status_map[i].msg;
+      return;
+    }
+  }
+  // Populate code.
+  // Start with the parent error and recurse through the tree of children
+  // until we find the first one that has a status code.
+  intptr_t status = GRPC_STATUS_UNKNOWN;  // Default in case we don't find one.
+  grpc_error *found_error = recursively_find_error_with_status(error, &status);
+  *code = (grpc_status_code)status;
+  // Now populate msg.
+  // If we found an error with a status code above, use that; otherwise,
+  // fall back to using the parent error.
+  if (found_error == NULL) found_error = error;
+  // If the error has a status message, use it.  Otherwise, fall back to
+  // the error description.
+  *msg = grpc_error_get_str(found_error, GRPC_ERROR_STR_GRPC_MESSAGE);
+  if (*msg == NULL) {
+    *msg = grpc_error_get_str(found_error, GRPC_ERROR_STR_DESCRIPTION);
+    if (*msg == NULL) *msg = "uknown error";  // Just in case.
+  }
+}
+
 grpc_error *grpc_error_add_child(grpc_error *src, grpc_error *child) {
   GPR_TIMER_BEGIN("grpc_error_add_child", 0);
   grpc_error *new = copy_error_and_unref(src);
diff --git a/src/core/lib/iomgr/error.h b/src/core/lib/iomgr/error.h
index 6c769accdb..00ace8a7a9 100644
--- a/src/core/lib/iomgr/error.h
+++ b/src/core/lib/iomgr/error.h
@@ -37,8 +37,13 @@
 #include <stdbool.h>
 #include <stdint.h>
 
+#include <grpc/status.h>
 #include <grpc/support/time.h>
 
+#ifdef __cplusplus
+extern "C" {
+#endif
+
 /// Opaque representation of an error.
 /// Errors are refcounted objects that represent the result of an operation.
 /// Ownership laws:
@@ -175,6 +180,13 @@ grpc_error *grpc_error_set_str(grpc_error *src, grpc_error_strs which,
 /// Returns NULL if the specified string is not set.
 /// Caller does NOT own return value.
 const char *grpc_error_get_str(grpc_error *error, grpc_error_strs which);
+
+/// A utility function to get the status code and message to be returned
+/// to the application.  If not set in the top-level message, looks
+/// through child errors until it finds the first one with these attributes.
+void grpc_error_get_status(grpc_error *error, grpc_status_code *code,
+                           const char **msg);
+
 /// Add a child error: an error that is believed to have contributed to this
 /// error occurring. Allows root causing high level errors from lower level
 /// errors that contributed to them.
@@ -196,4 +208,8 @@ bool grpc_log_if_error(const char *what, grpc_error *error, const char *file,
 #define GRPC_LOG_IF_ERROR(what, error) \
   grpc_log_if_error((what), (error), __FILE__, __LINE__)
 
+#ifdef __cplusplus
+}
+#endif
+
 #endif /* GRPC_CORE_LIB_IOMGR_ERROR_H */
diff --git a/src/core/lib/iomgr/ev_epoll_linux.c b/src/core/lib/iomgr/ev_epoll_linux.c
index 740920d760..ad1cfc2031 100644
--- a/src/core/lib/iomgr/ev_epoll_linux.c
+++ b/src/core/lib/iomgr/ev_epoll_linux.c
@@ -1892,12 +1892,16 @@ const grpc_event_engine_vtable *grpc_init_epoll_linux(void) {
     return NULL;
   }
 
+  if (!grpc_has_wakeup_fd()) {
+    return NULL;
+  }
+
   if (!is_epoll_available()) {
     return NULL;
   }
 
   if (!is_grpc_wakeup_signal_initialized) {
-    grpc_use_signal(SIGRTMIN + 2);
+    grpc_use_signal(SIGRTMIN + 6);
   }
 
   fd_global_init();
diff --git a/src/core/lib/iomgr/ev_poll_posix.c b/src/core/lib/iomgr/ev_poll_posix.c
index 16a5e3083e..f137e4dacc 100644
--- a/src/core/lib/iomgr/ev_poll_posix.c
+++ b/src/core/lib/iomgr/ev_poll_posix.c
@@ -47,10 +47,12 @@
 #include <grpc/support/alloc.h>
 #include <grpc/support/log.h>
 #include <grpc/support/string_util.h>
+#include <grpc/support/thd.h>
 #include <grpc/support/tls.h>
 #include <grpc/support/useful.h>
 
 #include "src/core/lib/iomgr/iomgr_internal.h"
+#include "src/core/lib/iomgr/wakeup_fd_cv.h"
 #include "src/core/lib/iomgr/wakeup_fd_posix.h"
 #include "src/core/lib/profiling/timers.h"
 #include "src/core/lib/support/block_annotate.h"
@@ -246,6 +248,28 @@ struct grpc_pollset_set {
 };
 
 /*******************************************************************************
+ * condition variable polling definitions
+ */
+
+#define CV_POLL_PERIOD_MS 1000
+#define CV_DEFAULT_TABLE_SIZE 16
+
+typedef enum poll_status_t { INPROGRESS, COMPLETED, CANCELLED } poll_status_t;
+
+typedef struct poll_args {
+  gpr_refcount refcount;
+  gpr_cv *cv;
+  struct pollfd *fds;
+  nfds_t nfds;
+  int timeout;
+  int retval;
+  int err;
+  gpr_atm status;
+} poll_args;
+
+cv_fd_table g_cvfds;
+
+/*******************************************************************************
  * fd_posix.c
  */
 
@@ -1236,10 +1260,211 @@ static void pollset_set_del_fd(grpc_exec_ctx *exec_ctx,
 }
 
 /*******************************************************************************
+ * Condition Variable polling extensions
+ */
+
+static void decref_poll_args(poll_args *args) {
+  if (gpr_unref(&args->refcount)) {
+    gpr_free(args->fds);
+    gpr_cv_destroy(args->cv);
+    gpr_free(args->cv);
+    gpr_free(args);
+  }
+}
+
+// Poll in a background thread
+static void run_poll(void *arg) {
+  int timeout, retval;
+  poll_args *pargs = (poll_args *)arg;
+  while (gpr_atm_no_barrier_load(&pargs->status) == INPROGRESS) {
+    if (pargs->timeout < 0) {
+      timeout = CV_POLL_PERIOD_MS;
+    } else {
+      timeout = GPR_MIN(CV_POLL_PERIOD_MS, pargs->timeout);
+      pargs->timeout -= timeout;
+    }
+    retval = g_cvfds.poll(pargs->fds, pargs->nfds, timeout);
+    if (retval != 0 || pargs->timeout == 0) {
+      pargs->retval = retval;
+      pargs->err = errno;
+      break;
+    }
+  }
+  gpr_mu_lock(&g_cvfds.mu);
+  if (gpr_atm_no_barrier_load(&pargs->status) == INPROGRESS) {
+    // Signal main thread that the poll completed
+    gpr_atm_no_barrier_store(&pargs->status, COMPLETED);
+    gpr_cv_signal(pargs->cv);
+  }
+  decref_poll_args(pargs);
+  g_cvfds.pollcount--;
+  if (g_cvfds.shutdown && g_cvfds.pollcount == 0) {
+    gpr_cv_signal(&g_cvfds.shutdown_complete);
+  }
+  gpr_mu_unlock(&g_cvfds.mu);
+}
+
+// This function overrides poll() to handle condition variable wakeup fds
+static int cvfd_poll(struct pollfd *fds, nfds_t nfds, int timeout) {
+  unsigned int i;
+  int res, idx;
+  gpr_cv *pollcv;
+  cv_node *cvn, *prev;
+  nfds_t nsockfds = 0;
+  gpr_thd_id t_id;
+  gpr_thd_options opt;
+  poll_args *pargs = NULL;
+  gpr_mu_lock(&g_cvfds.mu);
+  pollcv = gpr_malloc(sizeof(gpr_cv));
+  gpr_cv_init(pollcv);
+  for (i = 0; i < nfds; i++) {
+    fds[i].revents = 0;
+    if (fds[i].fd < 0 && (fds[i].events & POLLIN)) {
+      idx = FD_TO_IDX(fds[i].fd);
+      cvn = gpr_malloc(sizeof(cv_node));
+      cvn->cv = pollcv;
+      cvn->next = g_cvfds.cvfds[idx].cvs;
+      g_cvfds.cvfds[idx].cvs = cvn;
+      // We should return immediately if there are pending events,
+      // but we still need to call poll() to check for socket events
+      if (g_cvfds.cvfds[idx].is_set) {
+        timeout = 0;
+      }
+    } else if (fds[i].fd >= 0) {
+      nsockfds++;
+    }
+  }
+
+  if (nsockfds > 0) {
+    pargs = gpr_malloc(sizeof(struct poll_args));
+    // Both the main thread and calling thread get a reference
+    gpr_ref_init(&pargs->refcount, 2);
+    pargs->cv = pollcv;
+    pargs->fds = gpr_malloc(sizeof(struct pollfd) * nsockfds);
+    pargs->nfds = nsockfds;
+    pargs->timeout = timeout;
+    pargs->retval = 0;
+    pargs->err = 0;
+    gpr_atm_no_barrier_store(&pargs->status, INPROGRESS);
+    idx = 0;
+    for (i = 0; i < nfds; i++) {
+      if (fds[i].fd >= 0) {
+        pargs->fds[idx].fd = fds[i].fd;
+        pargs->fds[idx].events = fds[i].events;
+        pargs->fds[idx].revents = 0;
+        idx++;
+      }
+    }
+    g_cvfds.pollcount++;
+    opt = gpr_thd_options_default();
+    gpr_thd_options_set_detached(&opt);
+    gpr_thd_new(&t_id, &run_poll, pargs, &opt);
+    // We want the poll() thread to trigger the deadline, so wait forever here
+    gpr_cv_wait(pollcv, &g_cvfds.mu, gpr_inf_future(GPR_CLOCK_MONOTONIC));
+    if (gpr_atm_no_barrier_load(&pargs->status) == COMPLETED) {
+      res = pargs->retval;
+      errno = pargs->err;
+    } else {
+      res = 0;
+      errno = 0;
+      gpr_atm_no_barrier_store(&pargs->status, CANCELLED);
+    }
+  } else {
+    gpr_timespec deadline = gpr_now(GPR_CLOCK_REALTIME);
+    deadline =
+        gpr_time_add(deadline, gpr_time_from_millis(timeout, GPR_TIMESPAN));
+    gpr_cv_wait(pollcv, &g_cvfds.mu, deadline);
+    res = 0;
+  }
+
+  idx = 0;
+  for (i = 0; i < nfds; i++) {
+    if (fds[i].fd < 0 && (fds[i].events & POLLIN)) {
+      cvn = g_cvfds.cvfds[FD_TO_IDX(fds[i].fd)].cvs;
+      prev = NULL;
+      while (cvn->cv != pollcv) {
+        prev = cvn;
+        cvn = cvn->next;
+        GPR_ASSERT(cvn);
+      }
+      if (!prev) {
+        g_cvfds.cvfds[FD_TO_IDX(fds[i].fd)].cvs = cvn->next;
+      } else {
+        prev->next = cvn->next;
+      }
+      gpr_free(cvn);
+
+      if (g_cvfds.cvfds[FD_TO_IDX(fds[i].fd)].is_set) {
+        fds[i].revents = POLLIN;
+        if (res >= 0) res++;
+      }
+    } else if (fds[i].fd >= 0 &&
+               gpr_atm_no_barrier_load(&pargs->status) == COMPLETED) {
+      fds[i].revents = pargs->fds[idx].revents;
+      idx++;
+    }
+  }
+
+  if (pargs) {
+    decref_poll_args(pargs);
+  } else {
+    gpr_cv_destroy(pollcv);
+    gpr_free(pollcv);
+  }
+  gpr_mu_unlock(&g_cvfds.mu);
+
+  return res;
+}
+
+static void global_cv_fd_table_init() {
+  gpr_mu_init(&g_cvfds.mu);
+  gpr_mu_lock(&g_cvfds.mu);
+  gpr_cv_init(&g_cvfds.shutdown_complete);
+  g_cvfds.shutdown = 0;
+  g_cvfds.pollcount = 0;
+  g_cvfds.size = CV_DEFAULT_TABLE_SIZE;
+  g_cvfds.cvfds = gpr_malloc(sizeof(fd_node) * CV_DEFAULT_TABLE_SIZE);
+  g_cvfds.free_fds = NULL;
+  for (int i = 0; i < CV_DEFAULT_TABLE_SIZE; i++) {
+    g_cvfds.cvfds[i].is_set = 0;
+    g_cvfds.cvfds[i].cvs = NULL;
+    g_cvfds.cvfds[i].next_free = g_cvfds.free_fds;
+    g_cvfds.free_fds = &g_cvfds.cvfds[i];
+  }
+  // Override the poll function with one that supports cvfds
+  g_cvfds.poll = grpc_poll_function;
+  grpc_poll_function = &cvfd_poll;
+  gpr_mu_unlock(&g_cvfds.mu);
+}
+
+static void global_cv_fd_table_shutdown() {
+  gpr_mu_lock(&g_cvfds.mu);
+  g_cvfds.shutdown = 1;
+  // Attempt to wait for all abandoned poll() threads to terminate
+  // Not doing so will result in reported memory leaks
+  if (g_cvfds.pollcount > 0) {
+    int res = gpr_cv_wait(&g_cvfds.shutdown_complete, &g_cvfds.mu,
+                          gpr_time_add(gpr_now(GPR_CLOCK_REALTIME),
+                                       gpr_time_from_seconds(3, GPR_TIMESPAN)));
+    GPR_ASSERT(res == 0);
+  }
+  gpr_cv_destroy(&g_cvfds.shutdown_complete);
+  grpc_poll_function = g_cvfds.poll;
+  gpr_free(g_cvfds.cvfds);
+  gpr_mu_unlock(&g_cvfds.mu);
+  gpr_mu_destroy(&g_cvfds.mu);
+}
+
+/*******************************************************************************
  * event engine binding
  */
 
-static void shutdown_engine(void) { pollset_global_shutdown(); }
+static void shutdown_engine(void) {
+  pollset_global_shutdown();
+  if (grpc_cv_wakeup_fds_enabled()) {
+    global_cv_fd_table_shutdown();
+  }
+}
 
 static const grpc_event_engine_vtable vtable = {
     .pollset_size = sizeof(grpc_pollset),
@@ -1277,7 +1502,21 @@ static const grpc_event_engine_vtable vtable = {
 };
 
 const grpc_event_engine_vtable *grpc_init_poll_posix(void) {
+  if (!grpc_has_wakeup_fd()) {
+    return NULL;
+  }
+  if (!GRPC_LOG_IF_ERROR("pollset_global_init", pollset_global_init())) {
+    return NULL;
+  }
+  return &vtable;
+}
+
+const grpc_event_engine_vtable *grpc_init_poll_cv_posix(void) {
+  global_cv_fd_table_init();
+  grpc_enable_cv_wakeup_fds(1);
   if (!GRPC_LOG_IF_ERROR("pollset_global_init", pollset_global_init())) {
+    global_cv_fd_table_shutdown();
+    grpc_enable_cv_wakeup_fds(0);
     return NULL;
   }
   return &vtable;
diff --git a/src/core/lib/iomgr/ev_poll_posix.h b/src/core/lib/iomgr/ev_poll_posix.h
index 291736a2db..202ffca14c 100644
--- a/src/core/lib/iomgr/ev_poll_posix.h
+++ b/src/core/lib/iomgr/ev_poll_posix.h
@@ -37,5 +37,6 @@
 #include "src/core/lib/iomgr/ev_posix.h"
 
 const grpc_event_engine_vtable *grpc_init_poll_posix(void);
+const grpc_event_engine_vtable *grpc_init_poll_cv_posix(void);
 
 #endif /* GRPC_CORE_LIB_IOMGR_EV_POLL_POSIX_H */
diff --git a/src/core/lib/iomgr/ev_posix.c b/src/core/lib/iomgr/ev_posix.c
index 6536672685..0637f80421 100644
--- a/src/core/lib/iomgr/ev_posix.c
+++ b/src/core/lib/iomgr/ev_posix.c
@@ -66,6 +66,7 @@ typedef struct {
 static const event_engine_factory g_factories[] = {
     {"epoll", grpc_init_epoll_linux},
     {"poll", grpc_init_poll_posix},
+    {"poll-cv", grpc_init_poll_cv_posix},
     {"legacy", grpc_init_poll_and_epoll_posix},
 };
 
diff --git a/src/core/lib/iomgr/tcp_client_posix.c b/src/core/lib/iomgr/tcp_client_posix.c
index 80c7a3f128..3496b6094f 100644
--- a/src/core/lib/iomgr/tcp_client_posix.c
+++ b/src/core/lib/iomgr/tcp_client_posix.c
@@ -146,61 +146,57 @@ static void on_writable(grpc_exec_ctx *exec_ctx, void *acp, grpc_error *error) {
   grpc_timer_cancel(exec_ctx, &ac->alarm);
 
   gpr_mu_lock(&ac->mu);
-  if (error == GRPC_ERROR_NONE) {
-    do {
-      so_error_size = sizeof(so_error);
-      err = getsockopt(grpc_fd_wrapped_fd(fd), SOL_SOCKET, SO_ERROR, &so_error,
-                       &so_error_size);
-    } while (err < 0 && errno == EINTR);
-    if (err < 0) {
-      error = GRPC_OS_ERROR(errno, "getsockopt");
-      goto finish;
-    } else if (so_error != 0) {
-      if (so_error == ENOBUFS) {
-        /* We will get one of these errors if we have run out of
-           memory in the kernel for the data structures allocated
-           when you connect a socket.  If this happens it is very
-           likely that if we wait a little bit then try again the
-           connection will work (since other programs or this
-           program will close their network connections and free up
-           memory).  This does _not_ indicate that there is anything
-           wrong with the server we are connecting to, this is a
-           local problem.
-
-           If you are looking at this code, then chances are that
-           your program or another program on the same computer
-           opened too many network connections.  The "easy" fix:
-           don't do that! */
-        gpr_log(GPR_ERROR, "kernel out of buffers");
-        gpr_mu_unlock(&ac->mu);
-        grpc_fd_notify_on_write(exec_ctx, fd, &ac->write_closure);
-        return;
-      } else {
-        switch (so_error) {
-          case ECONNREFUSED:
-            error = grpc_error_set_int(error, GRPC_ERROR_INT_ERRNO, errno);
-            error = grpc_error_set_str(error, GRPC_ERROR_STR_OS_ERROR,
-                                       "Connection refused");
-            break;
-          default:
-            error = GRPC_OS_ERROR(errno, "getsockopt(SO_ERROR)");
-            break;
-        }
-        goto finish;
-      }
-    } else {
-      grpc_pollset_set_del_fd(exec_ctx, ac->interested_parties, fd);
-      *ep = grpc_tcp_create(fd, GRPC_TCP_DEFAULT_READ_SLICE_SIZE, ac->addr_str);
-      fd = NULL;
-      goto finish;
-    }
-  } else {
+  if (error != GRPC_ERROR_NONE) {
     error =
         grpc_error_set_str(error, GRPC_ERROR_STR_OS_ERROR, "Timeout occurred");
     goto finish;
   }
 
-  GPR_UNREACHABLE_CODE(return );
+  do {
+    so_error_size = sizeof(so_error);
+    err = getsockopt(grpc_fd_wrapped_fd(fd), SOL_SOCKET, SO_ERROR, &so_error,
+                     &so_error_size);
+  } while (err < 0 && errno == EINTR);
+  if (err < 0) {
+    error = GRPC_OS_ERROR(errno, "getsockopt");
+    goto finish;
+  }
+
+  switch (so_error) {
+    case 0:
+      grpc_pollset_set_del_fd(exec_ctx, ac->interested_parties, fd);
+      *ep = grpc_tcp_create(fd, GRPC_TCP_DEFAULT_READ_SLICE_SIZE, ac->addr_str);
+      fd = NULL;
+      break;
+    case ENOBUFS:
+      /* We will get one of these errors if we have run out of
+         memory in the kernel for the data structures allocated
+         when you connect a socket.  If this happens it is very
+         likely that if we wait a little bit then try again the
+         connection will work (since other programs or this
+         program will close their network connections and free up
+         memory).  This does _not_ indicate that there is anything
+         wrong with the server we are connecting to, this is a
+         local problem.
+
+         If you are looking at this code, then chances are that
+         your program or another program on the same computer
+         opened too many network connections.  The "easy" fix:
+         don't do that! */
+      gpr_log(GPR_ERROR, "kernel out of buffers");
+      gpr_mu_unlock(&ac->mu);
+      grpc_fd_notify_on_write(exec_ctx, fd, &ac->write_closure);
+      return;
+    case ECONNREFUSED:
+      /* This error shouldn't happen for anything other than connect(). */
+      error = GRPC_OS_ERROR(so_error, "connect");
+      break;
+    default:
+      /* We don't really know which syscall triggered the problem here,
+         so punt by reporting getsockopt(). */
+      error = GRPC_OS_ERROR(so_error, "getsockopt(SO_ERROR)");
+      break;
+  }
 
 finish:
   if (fd != NULL) {
diff --git a/src/core/lib/iomgr/tcp_server.h b/src/core/lib/iomgr/tcp_server.h
index 5a25d39a0c..9a390699b4 100644
--- a/src/core/lib/iomgr/tcp_server.h
+++ b/src/core/lib/iomgr/tcp_server.h
@@ -101,8 +101,8 @@ grpc_tcp_server *grpc_tcp_server_ref(grpc_tcp_server *s);
 void grpc_tcp_server_shutdown_starting_add(grpc_tcp_server *s,
                                            grpc_closure *shutdown_starting);
 
-/* If the refcount drops to zero, delete s, and call (exec_ctx==NULL) or enqueue
-   a call (exec_ctx!=NULL) to shutdown_complete. */
+/* If the refcount drops to zero, enqueue calls on exec_ctx to
+   shutdown_listeners and delete s. */
 void grpc_tcp_server_unref(grpc_exec_ctx *exec_ctx, grpc_tcp_server *s);
 
 /* Shutdown the fds of listeners. */
diff --git a/src/core/lib/iomgr/tcp_server_posix.c b/src/core/lib/iomgr/tcp_server_posix.c
index 2d3f6cf9a7..73df5477e6 100644
--- a/src/core/lib/iomgr/tcp_server_posix.c
+++ b/src/core/lib/iomgr/tcp_server_posix.c
@@ -191,6 +191,9 @@ grpc_error *grpc_tcp_server_create(grpc_closure *shutdown_complete,
 }
 
 static void finish_shutdown(grpc_exec_ctx *exec_ctx, grpc_tcp_server *s) {
+  gpr_mu_lock(&s->mu);
+  GPR_ASSERT(s->shutdown);
+  gpr_mu_unlock(&s->mu);
   if (s->shutdown_complete != NULL) {
     grpc_exec_ctx_sched(exec_ctx, s->shutdown_complete, GRPC_ERROR_NONE, NULL);
   }
@@ -652,6 +655,7 @@ unsigned grpc_tcp_server_port_fd_count(grpc_tcp_server *s,
                                        unsigned port_index) {
   unsigned num_fds = 0;
   grpc_tcp_listener *sp;
+  gpr_mu_lock(&s->mu);
   for (sp = s->head; sp && port_index != 0; sp = sp->next) {
     if (!sp->is_sibling) {
       --port_index;
@@ -659,12 +663,15 @@ unsigned grpc_tcp_server_port_fd_count(grpc_tcp_server *s,
   }
   for (; sp; sp = sp->sibling, ++num_fds)
     ;
+  gpr_mu_unlock(&s->mu);
   return num_fds;
 }
 
 int grpc_tcp_server_port_fd(grpc_tcp_server *s, unsigned port_index,
                             unsigned fd_index) {
   grpc_tcp_listener *sp;
+  int fd;
+  gpr_mu_lock(&s->mu);
   for (sp = s->head; sp && port_index != 0; sp = sp->next) {
     if (!sp->is_sibling) {
       --port_index;
@@ -673,10 +680,12 @@ int grpc_tcp_server_port_fd(grpc_tcp_server *s, unsigned port_index,
   for (; sp && fd_index != 0; sp = sp->sibling, --fd_index)
     ;
   if (sp) {
-    return sp->fd;
+    fd = sp->fd;
   } else {
-    return -1;
+    fd = -1;
   }
+  gpr_mu_unlock(&s->mu);
+  return fd;
 }
 
 void grpc_tcp_server_start(grpc_exec_ctx *exec_ctx, grpc_tcp_server *s,
@@ -722,7 +731,7 @@ void grpc_tcp_server_start(grpc_exec_ctx *exec_ctx, grpc_tcp_server *s,
 }
 
 grpc_tcp_server *grpc_tcp_server_ref(grpc_tcp_server *s) {
-  gpr_ref(&s->refs);
+  gpr_ref_non_zero(&s->refs);
   return s;
 }
 
@@ -736,19 +745,11 @@ void grpc_tcp_server_shutdown_starting_add(grpc_tcp_server *s,
 
 void grpc_tcp_server_unref(grpc_exec_ctx *exec_ctx, grpc_tcp_server *s) {
   if (gpr_unref(&s->refs)) {
-    /* Complete shutdown_starting work before destroying. */
-    grpc_exec_ctx local_exec_ctx = GRPC_EXEC_CTX_INIT;
+    grpc_tcp_server_shutdown_listeners(exec_ctx, s);
     gpr_mu_lock(&s->mu);
-    grpc_exec_ctx_enqueue_list(&local_exec_ctx, &s->shutdown_starting, NULL);
+    grpc_exec_ctx_enqueue_list(exec_ctx, &s->shutdown_starting, NULL);
     gpr_mu_unlock(&s->mu);
-    if (exec_ctx == NULL) {
-      grpc_exec_ctx_flush(&local_exec_ctx);
-      tcp_server_destroy(&local_exec_ctx, s);
-      grpc_exec_ctx_finish(&local_exec_ctx);
-    } else {
-      grpc_exec_ctx_finish(&local_exec_ctx);
-      tcp_server_destroy(exec_ctx, s);
-    }
+    tcp_server_destroy(exec_ctx, s);
   }
 }
 
diff --git a/src/core/lib/iomgr/tcp_server_windows.c b/src/core/lib/iomgr/tcp_server_windows.c
index 1b125e7005..4ff05601fa 100644
--- a/src/core/lib/iomgr/tcp_server_windows.c
+++ b/src/core/lib/iomgr/tcp_server_windows.c
@@ -139,7 +139,7 @@ static void finish_shutdown(grpc_exec_ctx *exec_ctx, grpc_tcp_server *s) {
 }
 
 grpc_tcp_server *grpc_tcp_server_ref(grpc_tcp_server *s) {
-  gpr_ref(&s->refs);
+  gpr_ref_non_zero(&s->refs);
   return s;
 }
 
@@ -174,19 +174,11 @@ static void tcp_server_destroy(grpc_exec_ctx *exec_ctx, grpc_tcp_server *s) {
 
 void grpc_tcp_server_unref(grpc_exec_ctx *exec_ctx, grpc_tcp_server *s) {
   if (gpr_unref(&s->refs)) {
-    /* Complete shutdown_starting work before destroying. */
-    grpc_exec_ctx local_exec_ctx = GRPC_EXEC_CTX_INIT;
+    grpc_tcp_server_shutdown_listeners(exec_ctx, s);
     gpr_mu_lock(&s->mu);
-    grpc_exec_ctx_enqueue_list(&local_exec_ctx, &s->shutdown_starting, NULL);
+    grpc_exec_ctx_enqueue_list(exec_ctx, &s->shutdown_starting, NULL);
     gpr_mu_unlock(&s->mu);
-    if (exec_ctx == NULL) {
-      grpc_exec_ctx_flush(&local_exec_ctx);
-      tcp_server_destroy(&local_exec_ctx, s);
-      grpc_exec_ctx_finish(&local_exec_ctx);
-    } else {
-      grpc_exec_ctx_finish(&local_exec_ctx);
-      tcp_server_destroy(exec_ctx, s);
-    }
+    tcp_server_destroy(exec_ctx, s);
   }
 }
 
diff --git a/src/core/lib/iomgr/tcp_windows.c b/src/core/lib/iomgr/tcp_windows.c
index 35054c42b5..448a72671c 100644
--- a/src/core/lib/iomgr/tcp_windows.c
+++ b/src/core/lib/iomgr/tcp_windows.c
@@ -319,6 +319,7 @@ static void win_write(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep,
                             ? GRPC_ERROR_NONE
                             : GRPC_WSA_ERROR(info->wsa_error, "WSASend");
     grpc_exec_ctx_sched(exec_ctx, cb, error, NULL);
+    if (allocated) gpr_free(allocated);
     return;
   }
 
diff --git a/src/core/lib/iomgr/udp_server.c b/src/core/lib/iomgr/udp_server.c
index 48032412a2..edf7b133e9 100644
--- a/src/core/lib/iomgr/udp_server.c
+++ b/src/core/lib/iomgr/udp_server.c
@@ -38,7 +38,6 @@
 
 #include <grpc/support/port_platform.h>
 
-#ifdef GRPC_NEED_UDP
 #ifdef GPR_POSIX_SOCKET
 
 #include "src/core/lib/iomgr/udp_server.h"
@@ -171,6 +170,8 @@ static void deactivated_all_ports(grpc_exec_ctx *exec_ctx, grpc_udp_server *s) {
       sp->destroyed_closure.cb = destroyed_port;
       sp->destroyed_closure.cb_arg = s;
 
+      /* Call the orphan_cb to signal that the FD is about to be closed and
+       * should no longer be used. */
       GPR_ASSERT(sp->orphan_cb);
       sp->orphan_cb(sp->emfd);
 
@@ -197,6 +198,12 @@ void grpc_udp_server_destroy(grpc_exec_ctx *exec_ctx, grpc_udp_server *s,
   /* shutdown all fd's */
   if (s->active_ports) {
     for (i = 0; i < s->nports; i++) {
+      server_port *sp = &s->ports[i];
+      /* Call the orphan_cb to signal that the FD is about to be closed and
+       * should no longer be used. */
+      GPR_ASSERT(sp->orphan_cb);
+      sp->orphan_cb(sp->emfd);
+
       grpc_fd_shutdown(exec_ctx, s->ports[i].emfd);
     }
     gpr_mu_unlock(&s->mu);
@@ -439,4 +446,3 @@ void grpc_udp_server_start(grpc_exec_ctx *exec_ctx, grpc_udp_server *s,
 }
 
 #endif
-#endif
diff --git a/src/core/lib/iomgr/wakeup_fd_cv.c b/src/core/lib/iomgr/wakeup_fd_cv.c
new file mode 100644
index 0000000000..b4165208ed
--- /dev/null
+++ b/src/core/lib/iomgr/wakeup_fd_cv.c
@@ -0,0 +1,118 @@
+/*
+ *
+ * Copyright 2016, Google Inc.
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ *     * 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.
+ *     * Neither the name of Google Inc. nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "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 COPYRIGHT
+ * OWNER OR CONTRIBUTORS 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.
+ *
+ */
+
+#include <grpc/support/port_platform.h>
+
+#ifdef GPR_POSIX_WAKEUP_FD
+
+#include "src/core/lib/iomgr/wakeup_fd_cv.h"
+
+#include <errno.h>
+#include <string.h>
+
+#include <grpc/support/alloc.h>
+#include <grpc/support/log.h>
+#include <grpc/support/sync.h>
+#include <grpc/support/thd.h>
+#include <grpc/support/time.h>
+#include <grpc/support/useful.h>
+
+#define MAX_TABLE_RESIZE 256
+
+extern cv_fd_table g_cvfds;
+
+static grpc_error* cv_fd_init(grpc_wakeup_fd* fd_info) {
+  unsigned int i, newsize;
+  int idx;
+  gpr_mu_lock(&g_cvfds.mu);
+  if (!g_cvfds.free_fds) {
+    newsize = GPR_MIN(g_cvfds.size * 2, g_cvfds.size + MAX_TABLE_RESIZE);
+    g_cvfds.cvfds = gpr_realloc(g_cvfds.cvfds, sizeof(fd_node) * newsize);
+    for (i = g_cvfds.size; i < newsize; i++) {
+      g_cvfds.cvfds[i].is_set = 0;
+      g_cvfds.cvfds[i].cvs = NULL;
+      g_cvfds.cvfds[i].next_free = g_cvfds.free_fds;
+      g_cvfds.free_fds = &g_cvfds.cvfds[i];
+    }
+    g_cvfds.size = newsize;
+  }
+
+  idx = (int)(g_cvfds.free_fds - g_cvfds.cvfds);
+  g_cvfds.free_fds = g_cvfds.free_fds->next_free;
+  g_cvfds.cvfds[idx].cvs = NULL;
+  g_cvfds.cvfds[idx].is_set = 0;
+  fd_info->read_fd = IDX_TO_FD(idx);
+  fd_info->write_fd = -1;
+  gpr_mu_unlock(&g_cvfds.mu);
+  return GRPC_ERROR_NONE;
+}
+
+static grpc_error* cv_fd_wakeup(grpc_wakeup_fd* fd_info) {
+  cv_node* cvn;
+  gpr_mu_lock(&g_cvfds.mu);
+  g_cvfds.cvfds[FD_TO_IDX(fd_info->read_fd)].is_set = 1;
+  cvn = g_cvfds.cvfds[FD_TO_IDX(fd_info->read_fd)].cvs;
+  while (cvn) {
+    gpr_cv_signal(cvn->cv);
+    cvn = cvn->next;
+  }
+  gpr_mu_unlock(&g_cvfds.mu);
+  return GRPC_ERROR_NONE;
+}
+
+static grpc_error* cv_fd_consume(grpc_wakeup_fd* fd_info) {
+  gpr_mu_lock(&g_cvfds.mu);
+  g_cvfds.cvfds[FD_TO_IDX(fd_info->read_fd)].is_set = 0;
+  gpr_mu_unlock(&g_cvfds.mu);
+  return GRPC_ERROR_NONE;
+}
+
+static void cv_fd_destroy(grpc_wakeup_fd* fd_info) {
+  if (fd_info->read_fd == 0) {
+    return;
+  }
+  gpr_mu_lock(&g_cvfds.mu);
+  // Assert that there are no active pollers
+  GPR_ASSERT(!g_cvfds.cvfds[FD_TO_IDX(fd_info->read_fd)].cvs);
+  g_cvfds.cvfds[FD_TO_IDX(fd_info->read_fd)].next_free = g_cvfds.free_fds;
+  g_cvfds.free_fds = &g_cvfds.cvfds[FD_TO_IDX(fd_info->read_fd)];
+  gpr_mu_unlock(&g_cvfds.mu);
+}
+
+static int cv_check_availability(void) { return 1; }
+
+const grpc_wakeup_fd_vtable grpc_cv_wakeup_fd_vtable = {
+    cv_fd_init, cv_fd_consume, cv_fd_wakeup, cv_fd_destroy,
+    cv_check_availability};
+
+#endif /* GPR_POSIX_WAKUP_FD */
diff --git a/src/core/lib/iomgr/wakeup_fd_cv.h b/src/core/lib/iomgr/wakeup_fd_cv.h
new file mode 100644
index 0000000000..ac16be1750
--- /dev/null
+++ b/src/core/lib/iomgr/wakeup_fd_cv.h
@@ -0,0 +1,80 @@
+/*
+ *
+ * Copyright 2016, Google Inc.
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ *     * 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.
+ *     * Neither the name of Google Inc. nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "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 COPYRIGHT
+ * OWNER OR CONTRIBUTORS 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.
+ *
+ */
+
+/*
+ * wakeup_fd_cv uses condition variables to implement wakeup fds.
+ *
+ * It is intended for use only in cases when eventfd() and pipe() are not
+ * available.  It can only be used with the "poll" engine.
+ *
+ * Implementation:
+ * A global table of cv wakeup fds is mantained.  A cv wakeup fd is a negative
+ * file descriptor.  poll() is then run in a background thread with only the
+ * real socket fds while we wait on a condition variable trigged by either the
+ * poll() completion or a wakeup_fd() call.
+ *
+ */
+
+#ifndef GRPC_CORE_LIB_IOMGR_WAKEUP_FD_CV_H
+#define GRPC_CORE_LIB_IOMGR_WAKEUP_FD_CV_H
+
+#include <grpc/support/sync.h>
+
+#include "src/core/lib/iomgr/ev_posix.h"
+
+#define FD_TO_IDX(fd) (-(fd)-1)
+#define IDX_TO_FD(idx) (-(idx)-1)
+
+typedef struct cv_node {
+  gpr_cv* cv;
+  struct cv_node* next;
+} cv_node;
+
+typedef struct fd_node {
+  int is_set;
+  cv_node* cvs;
+  struct fd_node* next_free;
+} fd_node;
+
+typedef struct cv_fd_table {
+  gpr_mu mu;
+  int pollcount;
+  int shutdown;
+  gpr_cv shutdown_complete;
+  fd_node* cvfds;
+  fd_node* free_fds;
+  unsigned int size;
+  grpc_poll_function_type poll;
+} cv_fd_table;
+
+#endif /* GRPC_CORE_LIB_IOMGR_WAKEUP_FD_CV_H */
diff --git a/src/core/lib/iomgr/wakeup_fd_pipe.c b/src/core/lib/iomgr/wakeup_fd_pipe.c
index 4e5dbdcb73..d0ea216aa0 100644
--- a/src/core/lib/iomgr/wakeup_fd_pipe.c
+++ b/src/core/lib/iomgr/wakeup_fd_pipe.c
@@ -47,11 +47,10 @@
 
 static grpc_error* pipe_init(grpc_wakeup_fd* fd_info) {
   int pipefd[2];
-  /* TODO(klempner): Make this nonfatal */
   int r = pipe(pipefd);
   if (0 != r) {
     gpr_log(GPR_ERROR, "pipe creation failed (%d): %s", errno, strerror(errno));
-    abort();
+    return GRPC_OS_ERROR(errno, "pipe");
   }
   grpc_error* err;
   err = grpc_set_socket_nonblocking(pipefd[0], 1);
@@ -95,8 +94,13 @@ static void pipe_destroy(grpc_wakeup_fd* fd_info) {
 }
 
 static int pipe_check_availability(void) {
-  /* Assume that pipes are always available. */
-  return 1;
+  grpc_wakeup_fd fd;
+  if (pipe_init(&fd) == GRPC_ERROR_NONE) {
+    pipe_destroy(&fd);
+    return 1;
+  } else {
+    return 0;
+  }
 }
 
 const grpc_wakeup_fd_vtable grpc_pipe_wakeup_fd_vtable = {
diff --git a/src/core/lib/iomgr/wakeup_fd_posix.c b/src/core/lib/iomgr/wakeup_fd_posix.c
index 046208abc8..5c894bef37 100644
--- a/src/core/lib/iomgr/wakeup_fd_posix.c
+++ b/src/core/lib/iomgr/wakeup_fd_posix.c
@@ -36,37 +36,66 @@
 #ifdef GPR_POSIX_WAKEUP_FD
 
 #include <stddef.h>
+#include "src/core/lib/iomgr/wakeup_fd_cv.h"
 #include "src/core/lib/iomgr/wakeup_fd_pipe.h"
 #include "src/core/lib/iomgr/wakeup_fd_posix.h"
 
+extern grpc_wakeup_fd_vtable grpc_cv_wakeup_fd_vtable;
 static const grpc_wakeup_fd_vtable *wakeup_fd_vtable = NULL;
+
 int grpc_allow_specialized_wakeup_fd = 1;
+int grpc_allow_pipe_wakeup_fd = 1;
+
+int has_real_wakeup_fd = 1;
+int cv_wakeup_fds_enabled = 0;
 
 void grpc_wakeup_fd_global_init(void) {
   if (grpc_allow_specialized_wakeup_fd &&
       grpc_specialized_wakeup_fd_vtable.check_availability()) {
     wakeup_fd_vtable = &grpc_specialized_wakeup_fd_vtable;
-  } else {
+  } else if (grpc_allow_pipe_wakeup_fd &&
+             grpc_pipe_wakeup_fd_vtable.check_availability()) {
     wakeup_fd_vtable = &grpc_pipe_wakeup_fd_vtable;
+  } else {
+    has_real_wakeup_fd = 0;
   }
 }
 
 void grpc_wakeup_fd_global_destroy(void) { wakeup_fd_vtable = NULL; }
 
+int grpc_has_wakeup_fd(void) { return has_real_wakeup_fd; }
+
+int grpc_cv_wakeup_fds_enabled(void) { return cv_wakeup_fds_enabled; }
+
+void grpc_enable_cv_wakeup_fds(int enable) { cv_wakeup_fds_enabled = enable; }
+
 grpc_error *grpc_wakeup_fd_init(grpc_wakeup_fd *fd_info) {
+  if (cv_wakeup_fds_enabled) {
+    return grpc_cv_wakeup_fd_vtable.init(fd_info);
+  }
   return wakeup_fd_vtable->init(fd_info);
 }
 
 grpc_error *grpc_wakeup_fd_consume_wakeup(grpc_wakeup_fd *fd_info) {
+  if (cv_wakeup_fds_enabled) {
+    return grpc_cv_wakeup_fd_vtable.consume(fd_info);
+  }
   return wakeup_fd_vtable->consume(fd_info);
 }
 
 grpc_error *grpc_wakeup_fd_wakeup(grpc_wakeup_fd *fd_info) {
+  if (cv_wakeup_fds_enabled) {
+    return grpc_cv_wakeup_fd_vtable.wakeup(fd_info);
+  }
   return wakeup_fd_vtable->wakeup(fd_info);
 }
 
 void grpc_wakeup_fd_destroy(grpc_wakeup_fd *fd_info) {
-  wakeup_fd_vtable->destroy(fd_info);
+  if (cv_wakeup_fds_enabled) {
+    grpc_cv_wakeup_fd_vtable.destroy(fd_info);
+  } else {
+    wakeup_fd_vtable->destroy(fd_info);
+  }
 }
 
 #endif /* GPR_POSIX_WAKEUP_FD */
diff --git a/src/core/lib/iomgr/wakeup_fd_posix.h b/src/core/lib/iomgr/wakeup_fd_posix.h
index e269f242d8..71d32d97ba 100644
--- a/src/core/lib/iomgr/wakeup_fd_posix.h
+++ b/src/core/lib/iomgr/wakeup_fd_posix.h
@@ -71,6 +71,10 @@ void grpc_wakeup_fd_global_destroy(void);
  * purposes only.*/
 void grpc_wakeup_fd_global_init_force_fallback(void);
 
+int grpc_has_wakeup_fd(void);
+int grpc_cv_wakeup_fds_enabled(void);
+void grpc_enable_cv_wakeup_fds(int enable);
+
 typedef struct grpc_wakeup_fd grpc_wakeup_fd;
 
 typedef struct grpc_wakeup_fd_vtable {
@@ -88,6 +92,7 @@ struct grpc_wakeup_fd {
 };
 
 extern int grpc_allow_specialized_wakeup_fd;
+extern int grpc_allow_pipe_wakeup_fd;
 
 #define GRPC_WAKEUP_FD_GET_READ_FD(fd_info) ((fd_info)->read_fd)