16 files changed, 397 insertions, 976 deletions

diff --git a/src/core/lib/iomgr/ev_poll_and_epoll_posix.c b/src/core/lib/iomgr/ev_poll_posix.c
index 3c8127e1a8..e2a21230b9 100644
--- a/src/core/lib/iomgr/ev_poll_and_epoll_posix.c
+++ b/src/core/lib/iomgr/ev_poll_posix.c
@@ -1,6 +1,6 @@
 /*
  *
- * Copyright 2015, Google Inc.
+ * Copyright 2015-2016, Google Inc.
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
@@ -31,22 +31,11 @@
  *
  */
 
-/* This file will be removed shortly: it's here to keep refactoring
- * steps simple and auditable.
- * It's the combination of the old files:
- *  - fd_posix.{h,c}
- *  - pollset_posix.{h,c}
- *  - pullset_multipoller_with_{poll,epoll}.{h,c}
- * The new version will be split into:
- *  - ev_poll_posix.{h,c}
- *  - ev_epoll_posix.{h,c}
- */
-
 #include <grpc/support/port_platform.h>
 
 #ifdef GPR_POSIX_SOCKET
 
-#include "src/core/lib/iomgr/ev_poll_and_epoll_posix.h"
+#include "src/core/lib/iomgr/ev_poll_posix.h"
 
 #include <assert.h>
 #include <errno.h>
@@ -70,6 +59,8 @@
  * FD declarations
  */
 
+grpc_wakeup_fd grpc_global_wakeup_fd;
+
 typedef struct grpc_fd_watcher {
   struct grpc_fd_watcher *next;
   struct grpc_fd_watcher *prev;
@@ -121,11 +112,12 @@ struct grpc_fd {
   grpc_closure *read_closure;
   grpc_closure *write_closure;
 
-  struct grpc_fd *freelist_next;
-
   grpc_closure *on_done_closure;
 
   grpc_iomgr_object iomgr_object;
+
+  /* The pollset that last noticed and notified that the fd is readable */
+  grpc_pollset *read_notifier_pollset;
 };
 
 /* Begin polling on an fd.
@@ -147,7 +139,8 @@ static uint32_t fd_begin_poll(grpc_fd *fd, grpc_pollset *pollset,
    if got_read or got_write are 1, also does the become_{readable,writable} as
    appropriate. */
 static void fd_end_poll(grpc_exec_ctx *exec_ctx, grpc_fd_watcher *rec,
-                        int got_read, int got_write);
+                        int got_read, int got_write,
+                        grpc_pollset *read_notifier_pollset);
 
 /* Return 1 if this fd is orphaned, 0 otherwise */
 static bool fd_is_orphaned(grpc_fd *fd);
@@ -167,9 +160,6 @@ static void fd_unref(grpc_fd *fd);
 #define GRPC_FD_UNREF(fd, reason) fd_unref(fd)
 #endif
 
-static void fd_global_init(void);
-static void fd_global_shutdown(void);
-
 #define CLOSURE_NOT_READY ((grpc_closure *)0)
 #define CLOSURE_READY ((grpc_closure *)1)
 
@@ -177,8 +167,6 @@ static void fd_global_shutdown(void);
  * pollset declarations
  */
 
-typedef struct grpc_pollset_vtable grpc_pollset_vtable;
-
 typedef struct grpc_cached_wakeup_fd {
   grpc_wakeup_fd fd;
   struct grpc_cached_wakeup_fd *next;
@@ -193,37 +181,21 @@ struct grpc_pollset_worker {
 };
 
 struct grpc_pollset {
-  /* pollsets under posix can mutate representation as fds are added and
-     removed.
-     For example, we may choose a poll() based implementation on linux for
-     few fds, and an epoll() based implementation for many fds */
-  const grpc_pollset_vtable *vtable;
   gpr_mu mu;
   grpc_pollset_worker root_worker;
-  int in_flight_cbs;
   int shutting_down;
   int called_shutdown;
   int kicked_without_pollers;
   grpc_closure *shutdown_done;
   grpc_closure_list idle_jobs;
-  union {
-    int fd;
-    void *ptr;
-  } data;
+  /* all polled fds */
+  size_t fd_count;
+  size_t fd_capacity;
+  grpc_fd **fds;
   /* Local cache of eventfds for workers */
   grpc_cached_wakeup_fd *local_wakeup_cache;
 };
 
-struct grpc_pollset_vtable {
-  void (*add_fd)(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
-                 struct grpc_fd *fd, int and_unlock_pollset);
-  void (*maybe_work_and_unlock)(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
-                                grpc_pollset_worker *worker,
-                                gpr_timespec deadline, gpr_timespec now);
-  void (*finish_shutdown)(grpc_pollset *pollset);
-  void (*destroy)(grpc_pollset *pollset);
-};
-
 /* Add an fd to a pollset */
 static void pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
                            struct grpc_fd *fd);
@@ -251,19 +223,10 @@ static void pollset_kick_ext(grpc_pollset *p,
                              grpc_pollset_worker *specific_worker,
                              uint32_t flags);
 
-/* turn a pollset into a multipoller: platform specific */
-typedef void (*platform_become_multipoller_type)(grpc_exec_ctx *exec_ctx,
-                                                 grpc_pollset *pollset,
-                                                 struct grpc_fd **fds,
-                                                 size_t fd_count);
-static platform_become_multipoller_type platform_become_multipoller;
-
 /* Return 1 if the pollset has active threads in pollset_work (pollset must
  * be locked) */
 static int pollset_has_workers(grpc_pollset *pollset);
 
-static void remove_fd_from_all_epoll_sets(int fd);
-
 /*******************************************************************************
  * pollset_set definitions
  */
@@ -288,69 +251,6 @@ struct grpc_pollset_set {
  * fd_posix.c
  */
 
-/* We need to keep a freelist not because of any concerns of malloc performance
- * but instead so that implementations with multiple threads in (for example)
- * epoll_wait deal with the race between pollset removal and incoming poll
- * notifications.
- *
- * The problem is that the poller ultimately holds a reference to this
- * object, so it is very difficult to know when is safe to free it, at least
- * without some expensive synchronization.
- *
- * If we keep the object freelisted, in the worst case losing this race just
- * becomes a spurious read notification on a reused fd.
- */
-/* TODO(klempner): We could use some form of polling generation count to know
- * when these are safe to free. */
-/* TODO(klempner): Consider disabling freelisting if we don't have multiple
- * threads in poll on the same fd */
-/* TODO(klempner): Batch these allocations to reduce fragmentation */
-static grpc_fd *fd_freelist = NULL;
-static gpr_mu fd_freelist_mu;
-
-static void freelist_fd(grpc_fd *fd) {
-  gpr_mu_lock(&fd_freelist_mu);
-  fd->freelist_next = fd_freelist;
-  fd_freelist = fd;
-  grpc_iomgr_unregister_object(&fd->iomgr_object);
-  gpr_mu_unlock(&fd_freelist_mu);
-}
-
-static grpc_fd *alloc_fd(int fd) {
-  grpc_fd *r = NULL;
-  gpr_mu_lock(&fd_freelist_mu);
-  if (fd_freelist != NULL) {
-    r = fd_freelist;
-    fd_freelist = fd_freelist->freelist_next;
-  }
-  gpr_mu_unlock(&fd_freelist_mu);
-  if (r == NULL) {
-    r = gpr_malloc(sizeof(grpc_fd));
-    gpr_mu_init(&r->mu);
-  }
-
-  gpr_mu_lock(&r->mu);
-  gpr_atm_rel_store(&r->refst, 1);
-  r->shutdown = 0;
-  r->read_closure = CLOSURE_NOT_READY;
-  r->write_closure = CLOSURE_NOT_READY;
-  r->fd = fd;
-  r->inactive_watcher_root.next = r->inactive_watcher_root.prev =
-      &r->inactive_watcher_root;
-  r->freelist_next = NULL;
-  r->read_watcher = r->write_watcher = NULL;
-  r->on_done_closure = NULL;
-  r->closed = 0;
-  r->released = 0;
-  gpr_mu_unlock(&r->mu);
-  return r;
-}
-
-static void destroy(grpc_fd *fd) {
-  gpr_mu_destroy(&fd->mu);
-  gpr_free(fd);
-}
-
 #ifdef GRPC_FD_REF_COUNT_DEBUG
 #define REF_BY(fd, n, reason) ref_by(fd, n, reason, __FILE__, __LINE__)
 #define UNREF_BY(fd, n, reason) unref_by(fd, n, reason, __FILE__, __LINE__)
@@ -380,27 +280,30 @@ static void unref_by(grpc_fd *fd, int n) {
 #endif
   old = gpr_atm_full_fetch_add(&fd->refst, -n);
   if (old == n) {
-    freelist_fd(fd);
+    gpr_mu_destroy(&fd->mu);
+    grpc_iomgr_unregister_object(&fd->iomgr_object);
+    gpr_free(fd);
   } else {
     GPR_ASSERT(old > n);
   }
 }
 
-static void fd_global_init(void) { gpr_mu_init(&fd_freelist_mu); }
-
-static void fd_global_shutdown(void) {
-  gpr_mu_lock(&fd_freelist_mu);
-  gpr_mu_unlock(&fd_freelist_mu);
-  while (fd_freelist != NULL) {
-    grpc_fd *fd = fd_freelist;
-    fd_freelist = fd_freelist->freelist_next;
-    destroy(fd);
-  }
-  gpr_mu_destroy(&fd_freelist_mu);
-}
-
 static grpc_fd *fd_create(int fd, const char *name) {
-  grpc_fd *r = alloc_fd(fd);
+  grpc_fd *r = gpr_malloc(sizeof(*r));
+  gpr_mu_init(&r->mu);
+  gpr_atm_rel_store(&r->refst, 1);
+  r->shutdown = 0;
+  r->read_closure = CLOSURE_NOT_READY;
+  r->write_closure = CLOSURE_NOT_READY;
+  r->fd = fd;
+  r->inactive_watcher_root.next = r->inactive_watcher_root.prev =
+      &r->inactive_watcher_root;
+  r->read_watcher = r->write_watcher = NULL;
+  r->on_done_closure = NULL;
+  r->closed = 0;
+  r->released = 0;
+  r->read_notifier_pollset = NULL;
+
   char *name2;
   gpr_asprintf(&name2, "%s fd=%d", name, fd);
   grpc_iomgr_register_object(&r->iomgr_object, name2);
@@ -415,6 +318,18 @@ static bool fd_is_orphaned(grpc_fd *fd) {
   return (gpr_atm_acq_load(&fd->refst) & 1) == 0;
 }
 
+/* Return the read-notifier pollset */
+static grpc_pollset *fd_get_read_notifier_pollset(grpc_exec_ctx *exec_ctx,
+                                                  grpc_fd *fd) {
+  grpc_pollset *notifier = NULL;
+
+  gpr_mu_lock(&fd->mu);
+  notifier = fd->read_notifier_pollset;
+  gpr_mu_unlock(&fd->mu);
+
+  return notifier;
+}
+
 static void pollset_kick_locked(grpc_fd_watcher *watcher) {
   gpr_mu_lock(&watcher->pollset->mu);
   GPR_ASSERT(watcher->worker);
@@ -456,8 +371,6 @@ static void close_fd_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd) {
   fd->closed = 1;
   if (!fd->released) {
     close(fd->fd);
-  } else {
-    remove_fd_from_all_epoll_sets(fd->fd);
   }
   grpc_exec_ctx_enqueue(exec_ctx, fd->on_done_closure, true, NULL);
 }
@@ -545,6 +458,11 @@ static int set_ready_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
   }
 }
 
+static void set_read_notifier_pollset_locked(
+    grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_pollset *read_notifier_pollset) {
+  fd->read_notifier_pollset = read_notifier_pollset;
+}
+
 static void fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd) {
   gpr_mu_lock(&fd->mu);
   GPR_ASSERT(!fd->shutdown);
@@ -620,7 +538,8 @@ static uint32_t fd_begin_poll(grpc_fd *fd, grpc_pollset *pollset,
 }
 
 static void fd_end_poll(grpc_exec_ctx *exec_ctx, grpc_fd_watcher *watcher,
-                        int got_read, int got_write) {
+                        int got_read, int got_write,
+                        grpc_pollset *read_notifier_pollset) {
   int was_polling = 0;
   int kick = 0;
   grpc_fd *fd = watcher->fd;
@@ -656,6 +575,9 @@ static void fd_end_poll(grpc_exec_ctx *exec_ctx, grpc_fd_watcher *watcher,
     if (set_ready_locked(exec_ctx, fd, &fd->read_closure)) {
       kick = 1;
     }
+    if (read_notifier_pollset != NULL) {
+      set_read_notifier_pollset_locked(exec_ctx, fd, read_notifier_pollset);
+    }
   }
   if (got_write) {
     if (set_ready_locked(exec_ctx, fd, &fd->write_closure)) {
@@ -680,12 +602,6 @@ static void fd_end_poll(grpc_exec_ctx *exec_ctx, grpc_fd_watcher *watcher,
 GPR_TLS_DECL(g_current_thread_poller);
 GPR_TLS_DECL(g_current_thread_worker);
 
-/** The alarm system needs to be able to wakeup 'some poller' sometimes
- *  (specifically when a new alarm needs to be triggered earlier than the next
- *  alarm 'epoch').
- *  This wakeup_fd gives us something to alert on when such a case occurs. */
-grpc_wakeup_fd grpc_global_wakeup_fd;
-
 static void remove_worker(grpc_pollset *p, grpc_pollset_worker *worker) {
   worker->prev->next = worker->next;
   worker->next->prev = worker->prev;
@@ -790,7 +706,6 @@ static void pollset_kick(grpc_pollset *p,
 static void pollset_global_init(void) {
   gpr_tls_init(&g_current_thread_poller);
   gpr_tls_init(&g_current_thread_worker);
-  grpc_wakeup_fd_global_init();
   grpc_wakeup_fd_init(&grpc_global_wakeup_fd);
 }
 
@@ -798,72 +713,78 @@ static void pollset_global_shutdown(void) {
   grpc_wakeup_fd_destroy(&grpc_global_wakeup_fd);
   gpr_tls_destroy(&g_current_thread_poller);
   gpr_tls_destroy(&g_current_thread_worker);
-  grpc_wakeup_fd_global_destroy();
 }
 
 static void kick_poller(void) { grpc_wakeup_fd_wakeup(&grpc_global_wakeup_fd); }
 
 /* main interface */
 
-static void become_basic_pollset(grpc_pollset *pollset, grpc_fd *fd_or_null);
-
 static void pollset_init(grpc_pollset *pollset, gpr_mu **mu) {
   gpr_mu_init(&pollset->mu);
   *mu = &pollset->mu;
   pollset->root_worker.next = pollset->root_worker.prev = &pollset->root_worker;
-  pollset->in_flight_cbs = 0;
   pollset->shutting_down = 0;
   pollset->called_shutdown = 0;
   pollset->kicked_without_pollers = 0;
   pollset->idle_jobs.head = pollset->idle_jobs.tail = NULL;
   pollset->local_wakeup_cache = NULL;
   pollset->kicked_without_pollers = 0;
-  become_basic_pollset(pollset, NULL);
+  pollset->fd_count = 0;
+  pollset->fd_capacity = 0;
+  pollset->fds = NULL;
 }
 
 static void pollset_destroy(grpc_pollset *pollset) {
-  GPR_ASSERT(pollset->in_flight_cbs == 0);
   GPR_ASSERT(!pollset_has_workers(pollset));
   GPR_ASSERT(pollset->idle_jobs.head == pollset->idle_jobs.tail);
-  pollset->vtable->destroy(pollset);
   while (pollset->local_wakeup_cache) {
     grpc_cached_wakeup_fd *next = pollset->local_wakeup_cache->next;
     grpc_wakeup_fd_destroy(&pollset->local_wakeup_cache->fd);
     gpr_free(pollset->local_wakeup_cache);
     pollset->local_wakeup_cache = next;
   }
+  gpr_free(pollset->fds);
   gpr_mu_destroy(&pollset->mu);
 }
 
 static void pollset_reset(grpc_pollset *pollset) {
   GPR_ASSERT(pollset->shutting_down);
-  GPR_ASSERT(pollset->in_flight_cbs == 0);
   GPR_ASSERT(!pollset_has_workers(pollset));
   GPR_ASSERT(pollset->idle_jobs.head == pollset->idle_jobs.tail);
-  pollset->vtable->destroy(pollset);
+  GPR_ASSERT(pollset->fd_count == 0);
   pollset->shutting_down = 0;
   pollset->called_shutdown = 0;
   pollset->kicked_without_pollers = 0;
-  become_basic_pollset(pollset, NULL);
 }
 
 static void pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
                            grpc_fd *fd) {
   gpr_mu_lock(&pollset->mu);
-  pollset->vtable->add_fd(exec_ctx, pollset, fd, 1);
-/* the following (enabled only in debug) will reacquire and then release
-   our lock - meaning that if the unlocking flag passed to add_fd above is
-   not respected, the code will deadlock (in a way that we have a chance of
-   debugging) */
-#ifndef NDEBUG
-  gpr_mu_lock(&pollset->mu);
+  size_t i;
+  /* TODO(ctiller): this is O(num_fds^2); maybe switch to a hash set here */
+  for (i = 0; i < pollset->fd_count; i++) {
+    if (pollset->fds[i] == fd) goto exit;
+  }
+  if (pollset->fd_count == pollset->fd_capacity) {
+    pollset->fd_capacity =
+        GPR_MAX(pollset->fd_capacity + 8, pollset->fd_count * 3 / 2);
+    pollset->fds =
+        gpr_realloc(pollset->fds, sizeof(grpc_fd *) * pollset->fd_capacity);
+  }
+  pollset->fds[pollset->fd_count++] = fd;
+  GRPC_FD_REF(fd, "multipoller");
+  pollset_kick(pollset, NULL);
+exit:
   gpr_mu_unlock(&pollset->mu);
-#endif
 }
 
 static void finish_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset) {
   GPR_ASSERT(grpc_closure_list_empty(pollset->idle_jobs));
-  pollset->vtable->finish_shutdown(pollset);
+  size_t i;
+  for (i = 0; i < pollset->fd_count; i++) {
+    GRPC_FD_UNREF(pollset->fds[i], "multipoller");
+  }
+  pollset->fd_count = 0;
   grpc_exec_ctx_enqueue(exec_ctx, pollset->shutdown_done, true, NULL);
 }
 
@@ -903,17 +824,11 @@ static void pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
     GPR_TIMER_MARK("pollset_work.shutting_down", 0);
     goto done;
   }
-  /* Give do_promote priority so we don't starve it out */
-  if (pollset->in_flight_cbs) {
-    GPR_TIMER_MARK("pollset_work.in_flight_cbs", 0);
-    gpr_mu_unlock(&pollset->mu);
-    locked = 0;
-    goto done;
-  }
   /* Start polling, and keep doing so while we're being asked to
      re-evaluate our pollers (this allows poll() based pollers to
      ensure they don't miss wakeups) */
   keep_polling = 1;
+  gpr_tls_set(&g_current_thread_poller, (intptr_t)pollset);
   while (keep_polling) {
     keep_polling = 0;
     if (!pollset->kicked_without_pollers) {
@@ -922,13 +837,91 @@ static void pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
         added_worker = 1;
         gpr_tls_set(&g_current_thread_worker, (intptr_t)&worker);
       }
-      gpr_tls_set(&g_current_thread_poller, (intptr_t)pollset);
       GPR_TIMER_BEGIN("maybe_work_and_unlock", 0);
-      pollset->vtable->maybe_work_and_unlock(exec_ctx, pollset, &worker,
-                                             deadline, now);
+#define POLLOUT_CHECK (POLLOUT | POLLHUP | POLLERR)
+#define POLLIN_CHECK (POLLIN | POLLHUP | POLLERR)
+
+      int timeout;
+      int r;
+      size_t i, fd_count;
+      nfds_t pfd_count;
+      /* TODO(ctiller): inline some elements to avoid an allocation */
+      grpc_fd_watcher *watchers;
+      struct pollfd *pfds;
+
+      timeout = poll_deadline_to_millis_timeout(deadline, now);
+      /* TODO(ctiller): perform just one malloc here if we exceed the inline
+       * case */
+      pfds = gpr_malloc(sizeof(*pfds) * (pollset->fd_count + 2));
+      watchers = gpr_malloc(sizeof(*watchers) * (pollset->fd_count + 2));
+      fd_count = 0;
+      pfd_count = 2;
+      pfds[0].fd = GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd);
+      pfds[0].events = POLLIN;
+      pfds[0].revents = 0;
+      pfds[1].fd = GRPC_WAKEUP_FD_GET_READ_FD(&worker.wakeup_fd->fd);
+      pfds[1].events = POLLIN;
+      pfds[1].revents = 0;
+      for (i = 0; i < pollset->fd_count; i++) {
+        if (fd_is_orphaned(pollset->fds[i])) {
+          GRPC_FD_UNREF(pollset->fds[i], "multipoller");
+        } else {
+          pollset->fds[fd_count++] = pollset->fds[i];
+          watchers[pfd_count].fd = pollset->fds[i];
+          GRPC_FD_REF(watchers[pfd_count].fd, "multipoller_start");
+          pfds[pfd_count].fd = pollset->fds[i]->fd;
+          pfds[pfd_count].revents = 0;
+          pfd_count++;
+        }
+      }
+      pollset->fd_count = fd_count;
+      gpr_mu_unlock(&pollset->mu);
+
+      for (i = 2; i < pfd_count; i++) {
+        grpc_fd *fd = watchers[i].fd;
+        pfds[i].events = (short)fd_begin_poll(fd, pollset, &worker, POLLIN,
+                                              POLLOUT, &watchers[i]);
+        GRPC_FD_UNREF(fd, "multipoller_start");
+      }
+
+      /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid
+         even going into the blocking annotation if possible */
+      GRPC_SCHEDULING_START_BLOCKING_REGION;
+      r = grpc_poll_function(pfds, pfd_count, timeout);
+      GRPC_SCHEDULING_END_BLOCKING_REGION;
+
+      if (r < 0) {
+        if (errno != EINTR) {
+          gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno));
+        }
+        for (i = 2; i < pfd_count; i++) {
+          fd_end_poll(exec_ctx, &watchers[i], 0, 0, NULL);
+        }
+      } else if (r == 0) {
+        for (i = 2; i < pfd_count; i++) {
+          fd_end_poll(exec_ctx, &watchers[i], 0, 0, NULL);
+        }
+      } else {
+        if (pfds[0].revents & POLLIN_CHECK) {
+          grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd);
+        }
+        if (pfds[1].revents & POLLIN_CHECK) {
+          grpc_wakeup_fd_consume_wakeup(&worker.wakeup_fd->fd);
+        }
+        for (i = 2; i < pfd_count; i++) {
+          if (watchers[i].fd == NULL) {
+            fd_end_poll(exec_ctx, &watchers[i], 0, 0, NULL);
+          } else {
+            fd_end_poll(exec_ctx, &watchers[i], pfds[i].revents & POLLIN_CHECK,
+                        pfds[i].revents & POLLOUT_CHECK, pollset);
+          }
+        }
+      }
+
+      gpr_free(pfds);
+      gpr_free(watchers);
       GPR_TIMER_END("maybe_work_and_unlock", 0);
       locked = 0;
-      gpr_tls_set(&g_current_thread_poller, 0);
     } else {
       GPR_TIMER_MARK("pollset_work.kicked_without_pollers", 0);
       pollset->kicked_without_pollers = 0;
@@ -956,7 +949,11 @@ static void pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
       }
       keep_polling = 1;
     }
+    if (keep_polling) {
+      now = gpr_now(now.clock_type);
+    }
   }
+  gpr_tls_set(&g_current_thread_poller, 0);
   if (added_worker) {
     remove_worker(pollset, &worker);
     gpr_tls_set(&g_current_thread_worker, 0);
@@ -968,7 +965,7 @@ static void pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
   if (pollset->shutting_down) {
     if (pollset_has_workers(pollset)) {
       pollset_kick(pollset, NULL);
-    } else if (!pollset->called_shutdown && pollset->in_flight_cbs == 0) {
+    } else if (!pollset->called_shutdown) {
       pollset->called_shutdown = 1;
       gpr_mu_unlock(&pollset->mu);
       finish_shutdown(exec_ctx, pollset);
@@ -998,8 +995,7 @@ static void pollset_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
   if (!pollset_has_workers(pollset)) {
     grpc_exec_ctx_enqueue_list(exec_ctx, &pollset->idle_jobs, NULL);
   }
-  if (!pollset->called_shutdown && pollset->in_flight_cbs == 0 &&
-      !pollset_has_workers(pollset)) {
+  if (!pollset->called_shutdown && !pollset_has_workers(pollset)) {
     pollset->called_shutdown = 1;
     finish_shutdown(exec_ctx, pollset);
   }
@@ -1022,722 +1018,6 @@ static int poll_deadline_to_millis_timeout(gpr_timespec deadline,
       timeout, gpr_time_from_nanos(GPR_NS_PER_MS - 1, GPR_TIMESPAN)));
 }
 
-/*
- * basic_pollset - a vtable that provides polling for zero or one file
- *                 descriptor via poll()
- */
-
-typedef struct grpc_unary_promote_args {
-  const grpc_pollset_vtable *original_vtable;
-  grpc_pollset *pollset;
-  grpc_fd *fd;
-  grpc_closure promotion_closure;
-} grpc_unary_promote_args;
-
-static void basic_do_promote(grpc_exec_ctx *exec_ctx, void *args,
-                             bool success) {
-  grpc_unary_promote_args *up_args = args;
-  const grpc_pollset_vtable *original_vtable = up_args->original_vtable;
-  grpc_pollset *pollset = up_args->pollset;
-  grpc_fd *fd = up_args->fd;
-
-  /*
-   * This is quite tricky. There are a number of cases to keep in mind here:
-   * 1. fd may have been orphaned
-   * 2. The pollset may no longer be a unary poller (and we can't let case #1
-   * leak to other pollset types!)
-   * 3. pollset's fd (which may have changed) may have been orphaned
-   * 4. The pollset may be shutting down.
-   */
-
-  gpr_mu_lock(&pollset->mu);
-  /* First we need to ensure that nobody is polling concurrently */
-  GPR_ASSERT(!pollset_has_workers(pollset));
-
-  gpr_free(up_args);
-  /* At this point the pollset may no longer be a unary poller. In that case
-   * we should just call the right add function and be done. */
-  /* TODO(klempner): If we're not careful this could cause infinite recursion.
-   * That's not a problem for now because empty_pollset has a trivial poller
-   * and we don't have any mechanism to unbecome multipoller. */
-  pollset->in_flight_cbs--;
-  if (pollset->shutting_down) {
-    /* We don't care about this pollset anymore. */
-    if (pollset->in_flight_cbs == 0 && !pollset->called_shutdown) {
-      pollset->called_shutdown = 1;
-      finish_shutdown(exec_ctx, pollset);
-    }
-  } else if (fd_is_orphaned(fd)) {
-    /* Don't try to add it to anything, we'll drop our ref on it below */
-  } else if (pollset->vtable != original_vtable) {
-    pollset->vtable->add_fd(exec_ctx, pollset, fd, 0);
-  } else if (fd != pollset->data.ptr) {
-    grpc_fd *fds[2];
-    fds[0] = pollset->data.ptr;
-    fds[1] = fd;
-
-    if (fds[0] && !fd_is_orphaned(fds[0])) {
-      platform_become_multipoller(exec_ctx, pollset, fds, GPR_ARRAY_SIZE(fds));
-      GRPC_FD_UNREF(fds[0], "basicpoll");
-    } else {
-      /* old fd is orphaned and we haven't cleaned it up until now, so remain a
-       * unary poller */
-      /* Note that it is possible that fds[1] is also orphaned at this point.
-       * That's okay, we'll correct it at the next add or poll. */
-      if (fds[0]) GRPC_FD_UNREF(fds[0], "basicpoll");
-      pollset->data.ptr = fd;
-      GRPC_FD_REF(fd, "basicpoll");
-    }
-  }
-
-  gpr_mu_unlock(&pollset->mu);
-
-  /* Matching ref in basic_pollset_add_fd */
-  GRPC_FD_UNREF(fd, "basicpoll_add");
-}
-
-static void basic_pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
-                                 grpc_fd *fd, int and_unlock_pollset) {
-  grpc_unary_promote_args *up_args;
-  GPR_ASSERT(fd);
-  if (fd == pollset->data.ptr) goto exit;
-
-  if (!pollset_has_workers(pollset)) {
-    /* Fast path -- no in flight cbs */
-    /* TODO(klempner): Comment this out and fix any test failures or establish
-     * they are due to timing issues */
-    grpc_fd *fds[2];
-    fds[0] = pollset->data.ptr;
-    fds[1] = fd;
-
-    if (fds[0] == NULL) {
-      pollset->data.ptr = fd;
-      GRPC_FD_REF(fd, "basicpoll");
-    } else if (!fd_is_orphaned(fds[0])) {
-      platform_become_multipoller(exec_ctx, pollset, fds, GPR_ARRAY_SIZE(fds));
-      GRPC_FD_UNREF(fds[0], "basicpoll");
-    } else {
-      /* old fd is orphaned and we haven't cleaned it up until now, so remain a
-       * unary poller */
-      GRPC_FD_UNREF(fds[0], "basicpoll");
-      pollset->data.ptr = fd;
-      GRPC_FD_REF(fd, "basicpoll");
-    }
-    goto exit;
-  }
-
-  /* Now we need to promote. This needs to happen when we're not polling. Since
-   * this may be called from poll, the wait needs to happen asynchronously. */
-  GRPC_FD_REF(fd, "basicpoll_add");
-  pollset->in_flight_cbs++;
-  up_args = gpr_malloc(sizeof(*up_args));
-  up_args->fd = fd;
-  up_args->original_vtable = pollset->vtable;
-  up_args->pollset = pollset;
-  up_args->promotion_closure.cb = basic_do_promote;
-  up_args->promotion_closure.cb_arg = up_args;
-
-  grpc_closure_list_add(&pollset->idle_jobs, &up_args->promotion_closure, 1);
-  pollset_kick(pollset, GRPC_POLLSET_KICK_BROADCAST);
-
-exit:
-  if (and_unlock_pollset) {
-    gpr_mu_unlock(&pollset->mu);
-  }
-}
-
-static void basic_pollset_maybe_work_and_unlock(grpc_exec_ctx *exec_ctx,
-                                                grpc_pollset *pollset,
-                                                grpc_pollset_worker *worker,
-                                                gpr_timespec deadline,
-                                                gpr_timespec now) {
-#define POLLOUT_CHECK (POLLOUT | POLLHUP | POLLERR)
-#define POLLIN_CHECK (POLLIN | POLLHUP | POLLERR)
-
-  struct pollfd pfd[3];
-  grpc_fd *fd;
-  grpc_fd_watcher fd_watcher;
-  int timeout;
-  int r;
-  nfds_t nfds;
-
-  fd = pollset->data.ptr;
-  if (fd && fd_is_orphaned(fd)) {
-    GRPC_FD_UNREF(fd, "basicpoll");
-    fd = pollset->data.ptr = NULL;
-  }
-  timeout = poll_deadline_to_millis_timeout(deadline, now);
-  pfd[0].fd = GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd);
-  pfd[0].events = POLLIN;
-  pfd[0].revents = 0;
-  pfd[1].fd = GRPC_WAKEUP_FD_GET_READ_FD(&worker->wakeup_fd->fd);
-  pfd[1].events = POLLIN;
-  pfd[1].revents = 0;
-  nfds = 2;
-  if (fd) {
-    pfd[2].fd = fd->fd;
-    pfd[2].revents = 0;
-    GRPC_FD_REF(fd, "basicpoll_begin");
-    gpr_mu_unlock(&pollset->mu);
-    pfd[2].events =
-        (short)fd_begin_poll(fd, pollset, worker, POLLIN, POLLOUT, &fd_watcher);
-    if (pfd[2].events != 0) {
-      nfds++;
-    }
-  } else {
-    gpr_mu_unlock(&pollset->mu);
-  }
-
-  /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid
-     even going into the blocking annotation if possible */
-  /* poll fd count (argument 2) is shortened by one if we have no events
-     to poll on - such that it only includes the kicker */
-  GPR_TIMER_BEGIN("poll", 0);
-  GRPC_SCHEDULING_START_BLOCKING_REGION;
-  r = grpc_poll_function(pfd, nfds, timeout);
-  GRPC_SCHEDULING_END_BLOCKING_REGION;
-  GPR_TIMER_END("poll", 0);
-
-  if (r < 0) {
-    if (errno != EINTR) {
-      gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno));
-    }
-    if (fd) {
-      fd_end_poll(exec_ctx, &fd_watcher, 0, 0);
-    }
-  } else if (r == 0) {
-    if (fd) {
-      fd_end_poll(exec_ctx, &fd_watcher, 0, 0);
-    }
-  } else {
-    if (pfd[0].revents & POLLIN_CHECK) {
-      grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd);
-    }
-    if (pfd[1].revents & POLLIN_CHECK) {
-      grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd->fd);
-    }
-    if (nfds > 2) {
-      fd_end_poll(exec_ctx, &fd_watcher, pfd[2].revents & POLLIN_CHECK,
-                  pfd[2].revents & POLLOUT_CHECK);
-    } else if (fd) {
-      fd_end_poll(exec_ctx, &fd_watcher, 0, 0);
-    }
-  }
-
-  if (fd) {
-    GRPC_FD_UNREF(fd, "basicpoll_begin");
-  }
-}
-
-static void basic_pollset_destroy(grpc_pollset *pollset) {
-  if (pollset->data.ptr != NULL) {
-    GRPC_FD_UNREF(pollset->data.ptr, "basicpoll");
-    pollset->data.ptr = NULL;
-  }
-}
-
-static const grpc_pollset_vtable basic_pollset = {
-    basic_pollset_add_fd, basic_pollset_maybe_work_and_unlock,
-    basic_pollset_destroy, basic_pollset_destroy};
-
-static void become_basic_pollset(grpc_pollset *pollset, grpc_fd *fd_or_null) {
-  pollset->vtable = &basic_pollset;
-  pollset->data.ptr = fd_or_null;
-  if (fd_or_null != NULL) {
-    GRPC_FD_REF(fd_or_null, "basicpoll");
-  }
-}
-
-/*******************************************************************************
- * pollset_multipoller_with_poll_posix.c
- */
-
-#ifndef GPR_LINUX_MULTIPOLL_WITH_EPOLL
-
-typedef struct {
-  /* all polled fds */
-  size_t fd_count;
-  size_t fd_capacity;
-  grpc_fd **fds;
-  /* fds that have been removed from the pollset explicitly */
-  size_t del_count;
-  size_t del_capacity;
-  grpc_fd **dels;
-} poll_hdr;
-
-static void multipoll_with_poll_pollset_add_fd(grpc_exec_ctx *exec_ctx,
-                                               grpc_pollset *pollset,
-                                               grpc_fd *fd,
-                                               int and_unlock_pollset) {
-  size_t i;
-  poll_hdr *h = pollset->data.ptr;
-  /* TODO(ctiller): this is O(num_fds^2); maybe switch to a hash set here */
-  for (i = 0; i < h->fd_count; i++) {
-    if (h->fds[i] == fd) goto exit;
-  }
-  if (h->fd_count == h->fd_capacity) {
-    h->fd_capacity = GPR_MAX(h->fd_capacity + 8, h->fd_count * 3 / 2);
-    h->fds = gpr_realloc(h->fds, sizeof(grpc_fd *) * h->fd_capacity);
-  }
-  h->fds[h->fd_count++] = fd;
-  GRPC_FD_REF(fd, "multipoller");
-exit:
-  if (and_unlock_pollset) {
-    gpr_mu_unlock(&pollset->mu);
-  }
-}
-
-static void multipoll_with_poll_pollset_maybe_work_and_unlock(
-    grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, grpc_pollset_worker *worker,
-    gpr_timespec deadline, gpr_timespec now) {
-#define POLLOUT_CHECK (POLLOUT | POLLHUP | POLLERR)
-#define POLLIN_CHECK (POLLIN | POLLHUP | POLLERR)
-
-  int timeout;
-  int r;
-  size_t i, j, fd_count;
-  nfds_t pfd_count;
-  poll_hdr *h;
-  /* TODO(ctiller): inline some elements to avoid an allocation */
-  grpc_fd_watcher *watchers;
-  struct pollfd *pfds;
-
-  h = pollset->data.ptr;
-  timeout = poll_deadline_to_millis_timeout(deadline, now);
-  /* TODO(ctiller): perform just one malloc here if we exceed the inline case */
-  pfds = gpr_malloc(sizeof(*pfds) * (h->fd_count + 2));
-  watchers = gpr_malloc(sizeof(*watchers) * (h->fd_count + 2));
-  fd_count = 0;
-  pfd_count = 2;
-  pfds[0].fd = GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd);
-  pfds[0].events = POLLIN;
-  pfds[0].revents = 0;
-  pfds[1].fd = GRPC_WAKEUP_FD_GET_READ_FD(&worker->wakeup_fd->fd);
-  pfds[1].events = POLLIN;
-  pfds[1].revents = 0;
-  for (i = 0; i < h->fd_count; i++) {
-    int remove = fd_is_orphaned(h->fds[i]);
-    for (j = 0; !remove && j < h->del_count; j++) {
-      if (h->fds[i] == h->dels[j]) remove = 1;
-    }
-    if (remove) {
-      GRPC_FD_UNREF(h->fds[i], "multipoller");
-    } else {
-      h->fds[fd_count++] = h->fds[i];
-      watchers[pfd_count].fd = h->fds[i];
-      GRPC_FD_REF(watchers[pfd_count].fd, "multipoller_start");
-      pfds[pfd_count].fd = h->fds[i]->fd;
-      pfds[pfd_count].revents = 0;
-      pfd_count++;
-    }
-  }
-  for (j = 0; j < h->del_count; j++) {
-    GRPC_FD_UNREF(h->dels[j], "multipoller_del");
-  }
-  h->del_count = 0;
-  h->fd_count = fd_count;
-  gpr_mu_unlock(&pollset->mu);
-
-  for (i = 2; i < pfd_count; i++) {
-    grpc_fd *fd = watchers[i].fd;
-    pfds[i].events = (short)fd_begin_poll(fd, pollset, worker, POLLIN, POLLOUT,
-                                          &watchers[i]);
-    GRPC_FD_UNREF(fd, "multipoller_start");
-  }
-
-  /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid
-     even going into the blocking annotation if possible */
-  GRPC_SCHEDULING_START_BLOCKING_REGION;
-  r = grpc_poll_function(pfds, pfd_count, timeout);
-  GRPC_SCHEDULING_END_BLOCKING_REGION;
-
-  if (r < 0) {
-    if (errno != EINTR) {
-      gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno));
-    }
-    for (i = 2; i < pfd_count; i++) {
-      fd_end_poll(exec_ctx, &watchers[i], 0, 0);
-    }
-  } else if (r == 0) {
-    for (i = 2; i < pfd_count; i++) {
-      fd_end_poll(exec_ctx, &watchers[i], 0, 0);
-    }
-  } else {
-    if (pfds[0].revents & POLLIN_CHECK) {
-      grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd);
-    }
-    if (pfds[1].revents & POLLIN_CHECK) {
-      grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd->fd);
-    }
-    for (i = 2; i < pfd_count; i++) {
-      if (watchers[i].fd == NULL) {
-        fd_end_poll(exec_ctx, &watchers[i], 0, 0);
-        continue;
-      }
-      fd_end_poll(exec_ctx, &watchers[i], pfds[i].revents & POLLIN_CHECK,
-                  pfds[i].revents & POLLOUT_CHECK);
-    }
-  }
-
-  gpr_free(pfds);
-  gpr_free(watchers);
-}
-
-static void multipoll_with_poll_pollset_finish_shutdown(grpc_pollset *pollset) {
-  size_t i;
-  poll_hdr *h = pollset->data.ptr;
-  for (i = 0; i < h->fd_count; i++) {
-    GRPC_FD_UNREF(h->fds[i], "multipoller");
-  }
-  for (i = 0; i < h->del_count; i++) {
-    GRPC_FD_UNREF(h->dels[i], "multipoller_del");
-  }
-  h->fd_count = 0;
-  h->del_count = 0;
-}
-
-static void multipoll_with_poll_pollset_destroy(grpc_pollset *pollset) {
-  poll_hdr *h = pollset->data.ptr;
-  multipoll_with_poll_pollset_finish_shutdown(pollset);
-  gpr_free(h->fds);
-  gpr_free(h->dels);
-  gpr_free(h);
-}
-
-static const grpc_pollset_vtable multipoll_with_poll_pollset = {
-    multipoll_with_poll_pollset_add_fd,
-    multipoll_with_poll_pollset_maybe_work_and_unlock,
-    multipoll_with_poll_pollset_finish_shutdown,
-    multipoll_with_poll_pollset_destroy};
-
-static void poll_become_multipoller(grpc_exec_ctx *exec_ctx,
-                                    grpc_pollset *pollset, grpc_fd **fds,
-                                    size_t nfds) {
-  size_t i;
-  poll_hdr *h = gpr_malloc(sizeof(poll_hdr));
-  pollset->vtable = &multipoll_with_poll_pollset;
-  pollset->data.ptr = h;
-  h->fd_count = nfds;
-  h->fd_capacity = nfds;
-  h->fds = gpr_malloc(nfds * sizeof(grpc_fd *));
-  h->del_count = 0;
-  h->del_capacity = 0;
-  h->dels = NULL;
-  for (i = 0; i < nfds; i++) {
-    h->fds[i] = fds[i];
-    GRPC_FD_REF(fds[i], "multipoller");
-  }
-}
-
-#endif /* !GPR_LINUX_MULTIPOLL_WITH_EPOLL */
-
-/*******************************************************************************
- * pollset_multipoller_with_epoll_posix.c
- */
-
-#ifdef GPR_LINUX_MULTIPOLL_WITH_EPOLL
-
-#include <errno.h>
-#include <poll.h>
-#include <string.h>
-#include <sys/epoll.h>
-#include <unistd.h>
-
-#include <grpc/support/alloc.h>
-#include <grpc/support/log.h>
-#include <grpc/support/useful.h>
-
-#include "src/core/lib/iomgr/ev_posix.h"
-#include "src/core/lib/profiling/timers.h"
-#include "src/core/lib/support/block_annotate.h"
-
-static void set_ready(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure **st) {
-  /* only one set_ready can be active at once (but there may be a racing
-     notify_on) */
-  gpr_mu_lock(&fd->mu);
-  set_ready_locked(exec_ctx, fd, st);
-  gpr_mu_unlock(&fd->mu);
-}
-
-static void fd_become_readable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) {
-  set_ready(exec_ctx, fd, &fd->read_closure);
-}
-
-static void fd_become_writable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) {
-  set_ready(exec_ctx, fd, &fd->write_closure);
-}
-
-struct epoll_fd_list {
-  int *epoll_fds;
-  size_t count;
-  size_t capacity;
-};
-
-static struct epoll_fd_list epoll_fd_global_list;
-static gpr_once init_epoll_fd_list_mu = GPR_ONCE_INIT;
-static gpr_mu epoll_fd_list_mu;
-
-static void init_mu(void) { gpr_mu_init(&epoll_fd_list_mu); }
-
-static void add_epoll_fd_to_global_list(int epoll_fd) {
-  gpr_once_init(&init_epoll_fd_list_mu, init_mu);
-
-  gpr_mu_lock(&epoll_fd_list_mu);
-  if (epoll_fd_global_list.count == epoll_fd_global_list.capacity) {
-    epoll_fd_global_list.capacity =
-        GPR_MAX((size_t)8, epoll_fd_global_list.capacity * 2);
-    epoll_fd_global_list.epoll_fds =
-        gpr_realloc(epoll_fd_global_list.epoll_fds,
-                    epoll_fd_global_list.capacity * sizeof(int));
-  }
-  epoll_fd_global_list.epoll_fds[epoll_fd_global_list.count++] = epoll_fd;
-  gpr_mu_unlock(&epoll_fd_list_mu);
-}
-
-static void remove_epoll_fd_from_global_list(int epoll_fd) {
-  gpr_mu_lock(&epoll_fd_list_mu);
-  GPR_ASSERT(epoll_fd_global_list.count > 0);
-  for (size_t i = 0; i < epoll_fd_global_list.count; i++) {
-    if (epoll_fd == epoll_fd_global_list.epoll_fds[i]) {
-      epoll_fd_global_list.epoll_fds[i] =
-          epoll_fd_global_list.epoll_fds[--(epoll_fd_global_list.count)];
-      break;
-    }
-  }
-  gpr_mu_unlock(&epoll_fd_list_mu);
-}
-
-static void remove_fd_from_all_epoll_sets(int fd) {
-  int err;
-  gpr_once_init(&init_epoll_fd_list_mu, init_mu);
-  gpr_mu_lock(&epoll_fd_list_mu);
-  if (epoll_fd_global_list.count == 0) {
-    gpr_mu_unlock(&epoll_fd_list_mu);
-    return;
-  }
-  for (size_t i = 0; i < epoll_fd_global_list.count; i++) {
-    err = epoll_ctl(epoll_fd_global_list.epoll_fds[i], EPOLL_CTL_DEL, fd, NULL);
-    if (err < 0 && errno != ENOENT) {
-      gpr_log(GPR_ERROR, "epoll_ctl del for %d failed: %s", fd,
-              strerror(errno));
-    }
-  }
-  gpr_mu_unlock(&epoll_fd_list_mu);
-}
-
-typedef struct {
-  grpc_pollset *pollset;
-  grpc_fd *fd;
-  grpc_closure closure;
-} delayed_add;
-
-typedef struct { int epoll_fd; } epoll_hdr;
-
-static void finally_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
-                           grpc_fd *fd) {
-  epoll_hdr *h = pollset->data.ptr;
-  struct epoll_event ev;
-  int err;
-  grpc_fd_watcher watcher;
-
-  /* We pretend to be polling whilst adding an fd to keep the fd from being
-     closed during the add. This may result in a spurious wakeup being assigned
-     to this pollset whilst adding, but that should be benign. */
-  GPR_ASSERT(fd_begin_poll(fd, pollset, NULL, 0, 0, &watcher) == 0);
-  if (watcher.fd != NULL) {
-    ev.events = (uint32_t)(EPOLLIN | EPOLLOUT | EPOLLET);
-    ev.data.ptr = fd;
-    err = epoll_ctl(h->epoll_fd, EPOLL_CTL_ADD, fd->fd, &ev);
-    if (err < 0) {
-      /* FDs may be added to a pollset multiple times, so EEXIST is normal. */
-      if (errno != EEXIST) {
-        gpr_log(GPR_ERROR, "epoll_ctl add for %d failed: %s", fd->fd,
-                strerror(errno));
-      }
-    }
-  }
-  fd_end_poll(exec_ctx, &watcher, 0, 0);
-}
-
-static void perform_delayed_add(grpc_exec_ctx *exec_ctx, void *arg,
-                                bool iomgr_status) {
-  delayed_add *da = arg;
-
-  if (!fd_is_orphaned(da->fd)) {
-    finally_add_fd(exec_ctx, da->pollset, da->fd);
-  }
-
-  gpr_mu_lock(&da->pollset->mu);
-  da->pollset->in_flight_cbs--;
-  if (da->pollset->shutting_down) {
-    /* We don't care about this pollset anymore. */
-    if (da->pollset->in_flight_cbs == 0 && !da->pollset->called_shutdown) {
-      da->pollset->called_shutdown = 1;
-      grpc_exec_ctx_enqueue(exec_ctx, da->pollset->shutdown_done, true, NULL);
-    }
-  }
-  gpr_mu_unlock(&da->pollset->mu);
-
-  GRPC_FD_UNREF(da->fd, "delayed_add");
-
-  gpr_free(da);
-}
-
-static void multipoll_with_epoll_pollset_add_fd(grpc_exec_ctx *exec_ctx,
-                                                grpc_pollset *pollset,
-                                                grpc_fd *fd,
-                                                int and_unlock_pollset) {
-  if (and_unlock_pollset) {
-    gpr_mu_unlock(&pollset->mu);
-    finally_add_fd(exec_ctx, pollset, fd);
-  } else {
-    delayed_add *da = gpr_malloc(sizeof(*da));
-    da->pollset = pollset;
-    da->fd = fd;
-    GRPC_FD_REF(fd, "delayed_add");
-    grpc_closure_init(&da->closure, perform_delayed_add, da);
-    pollset->in_flight_cbs++;
-    grpc_exec_ctx_enqueue(exec_ctx, &da->closure, true, NULL);
-  }
-}
-
-/* TODO(klempner): We probably want to turn this down a bit */
-#define GRPC_EPOLL_MAX_EVENTS 1000
-
-static void multipoll_with_epoll_pollset_maybe_work_and_unlock(
-    grpc_exec_ctx *exec_ctx, grpc_pollset *pollset, grpc_pollset_worker *worker,
-    gpr_timespec deadline, gpr_timespec now) {
-  struct epoll_event ep_ev[GRPC_EPOLL_MAX_EVENTS];
-  int ep_rv;
-  int poll_rv;
-  epoll_hdr *h = pollset->data.ptr;
-  int timeout_ms;
-  struct pollfd pfds[2];
-
-  /* If you want to ignore epoll's ability to sanely handle parallel pollers,
-   * for a more apples-to-apples performance comparison with poll, add a
-   * if (pollset->counter != 0) { return 0; }
-   * here.
-   */
-
-  gpr_mu_unlock(&pollset->mu);
-
-  timeout_ms = poll_deadline_to_millis_timeout(deadline, now);
-
-  pfds[0].fd = GRPC_WAKEUP_FD_GET_READ_FD(&worker->wakeup_fd->fd);
-  pfds[0].events = POLLIN;
-  pfds[0].revents = 0;
-  pfds[1].fd = h->epoll_fd;
-  pfds[1].events = POLLIN;
-  pfds[1].revents = 0;
-
-  /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid
-     even going into the blocking annotation if possible */
-  GPR_TIMER_BEGIN("poll", 0);
-  GRPC_SCHEDULING_START_BLOCKING_REGION;
-  poll_rv = grpc_poll_function(pfds, 2, timeout_ms);
-  GRPC_SCHEDULING_END_BLOCKING_REGION;
-  GPR_TIMER_END("poll", 0);
-
-  if (poll_rv < 0) {
-    if (errno != EINTR) {
-      gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno));
-    }
-  } else if (poll_rv == 0) {
-    /* do nothing */
-  } else {
-    if (pfds[0].revents) {
-      grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd->fd);
-    }
-    if (pfds[1].revents) {
-      do {
-        /* The following epoll_wait never blocks; it has a timeout of 0 */
-        ep_rv = epoll_wait(h->epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, 0);
-        if (ep_rv < 0) {
-          if (errno != EINTR) {
-            gpr_log(GPR_ERROR, "epoll_wait() failed: %s", strerror(errno));
-          }
-        } else {
-          int i;
-          for (i = 0; i < ep_rv; ++i) {
-            grpc_fd *fd = ep_ev[i].data.ptr;
-            /* TODO(klempner): We might want to consider making err and pri
-             * separate events */
-            int cancel = ep_ev[i].events & (EPOLLERR | EPOLLHUP);
-            int read_ev = ep_ev[i].events & (EPOLLIN | EPOLLPRI);
-            int write_ev = ep_ev[i].events & EPOLLOUT;
-            if (fd == NULL) {
-              grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd);
-            } else {
-              if (read_ev || cancel) {
-                fd_become_readable(exec_ctx, fd);
-              }
-              if (write_ev || cancel) {
-                fd_become_writable(exec_ctx, fd);
-              }
-            }
-          }
-        }
-      } while (ep_rv == GRPC_EPOLL_MAX_EVENTS);
-    }
-  }
-}
-
-static void multipoll_with_epoll_pollset_finish_shutdown(
-    grpc_pollset *pollset) {}
-
-static void multipoll_with_epoll_pollset_destroy(grpc_pollset *pollset) {
-  epoll_hdr *h = pollset->data.ptr;
-  close(h->epoll_fd);
-  remove_epoll_fd_from_global_list(h->epoll_fd);
-  gpr_free(h);
-}
-
-static const grpc_pollset_vtable multipoll_with_epoll_pollset = {
-    multipoll_with_epoll_pollset_add_fd,
-    multipoll_with_epoll_pollset_maybe_work_and_unlock,
-    multipoll_with_epoll_pollset_finish_shutdown,
-    multipoll_with_epoll_pollset_destroy};
-
-static void epoll_become_multipoller(grpc_exec_ctx *exec_ctx,
-                                     grpc_pollset *pollset, grpc_fd **fds,
-                                     size_t nfds) {
-  size_t i;
-  epoll_hdr *h = gpr_malloc(sizeof(epoll_hdr));
-  struct epoll_event ev;
-  int err;
-
-  pollset->vtable = &multipoll_with_epoll_pollset;
-  pollset->data.ptr = h;
-  h->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
-  if (h->epoll_fd < 0) {
-    /* TODO(klempner): Fall back to poll here, especially on ENOSYS */
-    gpr_log(GPR_ERROR, "epoll_create1 failed: %s", strerror(errno));
-    abort();
-  }
-  add_epoll_fd_to_global_list(h->epoll_fd);
-
-  ev.events = (uint32_t)(EPOLLIN | EPOLLET);
-  ev.data.ptr = NULL;
-  err = epoll_ctl(h->epoll_fd, EPOLL_CTL_ADD,
-                  GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd), &ev);
-  if (err < 0) {
-    gpr_log(GPR_ERROR, "epoll_ctl add for %d failed: %s",
-            GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd),
-            strerror(errno));
-  }
-
-  for (i = 0; i < nfds; i++) {
-    multipoll_with_epoll_pollset_add_fd(exec_ctx, pollset, fds[i], 0);
-  }
-}
-
-#else /* GPR_LINUX_MULTIPOLL_WITH_EPOLL */
-
-static void remove_fd_from_all_epoll_sets(int fd) {}
-
-#endif /* GPR_LINUX_MULTIPOLL_WITH_EPOLL */
-
 /*******************************************************************************
  * pollset_set_posix.c
  */
@@ -1885,10 +1165,7 @@ static void pollset_set_del_fd(grpc_exec_ctx *exec_ctx,
  * event engine binding
  */
 
-static void shutdown_engine(void) {
-  fd_global_shutdown();
-  pollset_global_shutdown();
-}
+static void shutdown_engine(void) { pollset_global_shutdown(); }
 
 static const grpc_event_engine_vtable vtable = {
     .pollset_size = sizeof(grpc_pollset),
@@ -1899,6 +1176,7 @@ static const grpc_event_engine_vtable vtable = {
     .fd_shutdown = fd_shutdown,
     .fd_notify_on_read = fd_notify_on_read,
     .fd_notify_on_write = fd_notify_on_write,
+    .fd_get_read_notifier_pollset = fd_get_read_notifier_pollset,
 
     .pollset_init = pollset_init,
     .pollset_shutdown = pollset_shutdown,
@@ -1922,13 +1200,7 @@ static const grpc_event_engine_vtable vtable = {
     .shutdown_engine = shutdown_engine,
 };
 
-const grpc_event_engine_vtable *grpc_init_poll_and_epoll_posix(void) {
-#ifdef GPR_LINUX_MULTIPOLL_WITH_EPOLL
-  platform_become_multipoller = epoll_become_multipoller;
-#else
-  platform_become_multipoller = poll_become_multipoller;
-#endif
-  fd_global_init();
+const grpc_event_engine_vtable *grpc_init_poll_posix(void) {
   pollset_global_init();
   return &vtable;
 }
diff --git a/src/core/lib/iomgr/ev_poll_and_epoll_posix.h b/src/core/lib/iomgr/ev_poll_posix.h
index 06d6dbf29d..291736a2db 100644
--- a/src/core/lib/iomgr/ev_poll_and_epoll_posix.h
+++ b/src/core/lib/iomgr/ev_poll_posix.h
@@ -1,6 +1,6 @@
 /*
  *
- * Copyright 2015, Google Inc.
+ * Copyright 2015-2016, Google Inc.
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
@@ -31,11 +31,11 @@
  *
  */
 
-#ifndef GRPC_CORE_LIB_IOMGR_EV_POLL_AND_EPOLL_POSIX_H
-#define GRPC_CORE_LIB_IOMGR_EV_POLL_AND_EPOLL_POSIX_H
+#ifndef GRPC_CORE_LIB_IOMGR_EV_POLL_POSIX_H
+#define GRPC_CORE_LIB_IOMGR_EV_POLL_POSIX_H
 
 #include "src/core/lib/iomgr/ev_posix.h"
 
-const grpc_event_engine_vtable *grpc_init_poll_and_epoll_posix(void);
+const grpc_event_engine_vtable *grpc_init_poll_posix(void);
 
-#endif /* GRPC_CORE_LIB_IOMGR_EV_POLL_AND_EPOLL_POSIX_H */
+#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 0eb95a2e09..95520b01d3 100644
--- a/src/core/lib/iomgr/ev_posix.c
+++ b/src/core/lib/iomgr/ev_posix.c
@@ -37,24 +37,103 @@
 
 #include "src/core/lib/iomgr/ev_posix.h"
 
+#include <string.h>
+
+#include <grpc/support/alloc.h>
 #include <grpc/support/log.h>
+#include <grpc/support/string_util.h>
+#include <grpc/support/useful.h>
+
+#include "src/core/lib/iomgr/ev_poll_posix.h"
+#include "src/core/lib/support/env.h"
 
-#include "src/core/lib/iomgr/ev_poll_and_epoll_posix.h"
+/** Default poll() function - a pointer so that it can be overridden by some
+ *  tests */
+grpc_poll_function_type grpc_poll_function = poll;
 
 static const grpc_event_engine_vtable *g_event_engine;
 
-grpc_poll_function_type grpc_poll_function = poll;
-grpc_wakeup_fd grpc_global_wakeup_fd;
+typedef const grpc_event_engine_vtable *(*event_engine_factory_fn)(void);
+
+typedef struct {
+  const char *name;
+  event_engine_factory_fn factory;
+} event_engine_factory;
+
+static const event_engine_factory g_factories[] = {
+    {"poll", grpc_init_poll_posix},
+};
+
+static void add(const char *beg, const char *end, char ***ss, size_t *ns) {
+  size_t n = *ns;
+  size_t np = n + 1;
+  char *s;
+  size_t len;
+  GPR_ASSERT(end >= beg);
+  len = (size_t)(end - beg);
+  s = gpr_malloc(len + 1);
+  memcpy(s, beg, len);
+  s[len] = 0;
+  *ss = gpr_realloc(*ss, sizeof(char **) * np);
+  (*ss)[n] = s;
+  *ns = np;
+}
+
+static void split(const char *s, char ***ss, size_t *ns) {
+  const char *c = strchr(s, ',');
+  if (c == NULL) {
+    add(s, s + strlen(s), ss, ns);
+  } else {
+    add(s, c, ss, ns);
+    split(c + 1, ss, ns);
+  }
+}
+
+static bool is(const char *want, const char *have) {
+  return 0 == strcmp(want, "all") || 0 == strcmp(want, have);
+}
+
+static void try_engine(const char *engine) {
+  for (size_t i = 0; i < GPR_ARRAY_SIZE(g_factories); i++) {
+    if (is(engine, g_factories[i].name)) {
+      if ((g_event_engine = g_factories[i].factory())) {
+        gpr_log(GPR_DEBUG, "Using polling engine: %s", g_factories[i].name);
+        return;
+      }
+    }
+  }
+}
 
 void grpc_event_engine_init(void) {
-  if ((g_event_engine = grpc_init_poll_and_epoll_posix())) {
-    return;
+  char *s = gpr_getenv("GRPC_POLL_STRATEGY");
+  if (s == NULL) {
+    s = gpr_strdup("all");
+  }
+
+  char **strings = NULL;
+  size_t nstrings = 0;
+  split(s, &strings, &nstrings);
+
+  for (size_t i = 0; g_event_engine == NULL && i < nstrings; i++) {
+    try_engine(strings[i]);
+  }
+
+  for (size_t i = 0; i < nstrings; i++) {
+    gpr_free(strings[i]);
+  }
+  gpr_free(strings);
+  gpr_free(s);
+
+  if (g_event_engine == NULL) {
+    gpr_log(GPR_ERROR, "No event engine could be initialized");
+    abort();
   }
-  gpr_log(GPR_ERROR, "No event engine could be initialized");
-  abort();
 }
 
-void grpc_event_engine_shutdown(void) { g_event_engine->shutdown_engine(); }
+void grpc_event_engine_shutdown(void) {
+  g_event_engine->shutdown_engine();
+  g_event_engine = NULL;
+}
 
 grpc_fd *grpc_fd_create(int fd, const char *name) {
   return g_event_engine->fd_create(fd, name);
@@ -83,6 +162,11 @@ void grpc_fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
   g_event_engine->fd_notify_on_write(exec_ctx, fd, closure);
 }
 
+grpc_pollset *grpc_fd_get_read_notifier_pollset(grpc_exec_ctx *exec_ctx,
+                                                grpc_fd *fd) {
+  return g_event_engine->fd_get_read_notifier_pollset(exec_ctx, fd);
+}
+
 size_t grpc_pollset_size(void) { return g_event_engine->pollset_size; }
 
 void grpc_pollset_init(grpc_pollset *pollset, gpr_mu **mu) {
diff --git a/src/core/lib/iomgr/ev_posix.h b/src/core/lib/iomgr/ev_posix.h
index 1fa9f5ef2d..344bf63438 100644
--- a/src/core/lib/iomgr/ev_posix.h
+++ b/src/core/lib/iomgr/ev_posix.h
@@ -55,6 +55,8 @@ typedef struct grpc_event_engine_vtable {
                             grpc_closure *closure);
   void (*fd_notify_on_write)(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
                              grpc_closure *closure);
+  grpc_pollset *(*fd_get_read_notifier_pollset)(grpc_exec_ctx *exec_ctx,
+                                                grpc_fd *fd);
 
   void (*pollset_init)(grpc_pollset *pollset, gpr_mu **mu);
   void (*pollset_shutdown)(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
@@ -137,6 +139,10 @@ void grpc_fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
 void grpc_fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
                              grpc_closure *closure);
 
+/* Return the read notifier pollset from the fd */
+grpc_pollset *grpc_fd_get_read_notifier_pollset(grpc_exec_ctx *exec_ctx,
+                                                grpc_fd *fd);
+
 /* pollset_posix functions */
 
 /* Add an fd to a pollset */
diff --git a/src/core/lib/iomgr/exec_ctx.c b/src/core/lib/iomgr/exec_ctx.c
index 2146c7dd1f..e451479073 100644
--- a/src/core/lib/iomgr/exec_ctx.c
+++ b/src/core/lib/iomgr/exec_ctx.c
@@ -39,6 +39,22 @@
 
 #include "src/core/lib/profiling/timers.h"
 
+bool grpc_exec_ctx_ready_to_finish(grpc_exec_ctx *exec_ctx) {
+  if (!exec_ctx->cached_ready_to_finish) {
+    exec_ctx->cached_ready_to_finish = exec_ctx->check_ready_to_finish(
+        exec_ctx, exec_ctx->check_ready_to_finish_arg);
+  }
+  return exec_ctx->cached_ready_to_finish;
+}
+
+bool grpc_never_ready_to_finish(grpc_exec_ctx *exec_ctx, void *arg_ignored) {
+  return false;
+}
+
+bool grpc_always_ready_to_finish(grpc_exec_ctx *exec_ctx, void *arg_ignored) {
+  return true;
+}
+
 #ifndef GRPC_EXECUTION_CONTEXT_SANITIZER
 bool grpc_exec_ctx_flush(grpc_exec_ctx *exec_ctx) {
   bool did_something = 0;
@@ -61,6 +77,7 @@ bool grpc_exec_ctx_flush(grpc_exec_ctx *exec_ctx) {
 }
 
 void grpc_exec_ctx_finish(grpc_exec_ctx *exec_ctx) {
+  exec_ctx->cached_ready_to_finish = true;
   grpc_exec_ctx_flush(exec_ctx);
 }
 
diff --git a/src/core/lib/iomgr/exec_ctx.h b/src/core/lib/iomgr/exec_ctx.h
index 976cc40347..9d47a262f8 100644
--- a/src/core/lib/iomgr/exec_ctx.h
+++ b/src/core/lib/iomgr/exec_ctx.h
@@ -53,6 +53,9 @@ typedef struct grpc_workqueue grpc_workqueue;
  *  - track a list of work that needs to be delayed until the top of the
  *    call stack (this provides a convenient mechanism to run callbacks
  *    without worrying about locking issues)
+ *  - provide a decision maker (via grpc_exec_ctx_ready_to_finish) that provides
+ *    signal as to whether a borrowed thread should continue to do work or
+ *    should actively try to finish up and get this thread back to its owner
  *
  *  CONVENTIONS:
  *  Instance of this must ALWAYS be constructed on the stack, never
@@ -63,18 +66,26 @@ typedef struct grpc_workqueue grpc_workqueue;
  */
 struct grpc_exec_ctx {
   grpc_closure_list closure_list;
+  bool cached_ready_to_finish;
+  void *check_ready_to_finish_arg;
+  bool (*check_ready_to_finish)(grpc_exec_ctx *exec_ctx, void *arg);
 };
 
-#define GRPC_EXEC_CTX_INIT \
-  { GRPC_CLOSURE_LIST_INIT }
+#define GRPC_EXEC_CTX_INIT_WITH_FINISH_CHECK(finish_check, finish_check_arg) \
+  { GRPC_CLOSURE_LIST_INIT, false, finish_check_arg, finish_check }
 #else
 struct grpc_exec_ctx {
-  int unused;
+  bool cached_ready_to_finish;
+  void *check_ready_to_finish_arg;
+  bool (*check_ready_to_finish)(grpc_exec_ctx *exec_ctx, void *arg);
 };
-#define GRPC_EXEC_CTX_INIT \
-  { 0 }
+#define GRPC_EXEC_CTX_INIT_WITH_FINISH_CHECK(finish_check, finish_check_arg) \
+  { false, finish_check_arg, finish_check }
 #endif
 
+#define GRPC_EXEC_CTX_INIT \
+  GRPC_EXEC_CTX_INIT_WITH_FINISH_CHECK(grpc_never_ready_to_finish, NULL)
+
 /** Flush any work that has been enqueued onto this grpc_exec_ctx.
  *  Caller must guarantee that no interfering locks are held.
  *  Returns true if work was performed, false otherwise. */
@@ -86,6 +97,14 @@ void grpc_exec_ctx_finish(grpc_exec_ctx *exec_ctx);
 void grpc_exec_ctx_enqueue(grpc_exec_ctx *exec_ctx, grpc_closure *closure,
                            bool success,
                            grpc_workqueue *offload_target_or_null);
+/** Returns true if we'd like to leave this execution context as soon as
+    possible: useful for deciding whether to do something more or not depending
+    on outside context */
+bool grpc_exec_ctx_ready_to_finish(grpc_exec_ctx *exec_ctx);
+/** A finish check that is never ready to finish */
+bool grpc_never_ready_to_finish(grpc_exec_ctx *exec_ctx, void *arg_ignored);
+/** A finish check that is always ready to finish */
+bool grpc_always_ready_to_finish(grpc_exec_ctx *exec_ctx, void *arg_ignored);
 /** Add a list of closures to be executed at the next flush/finish point.
  *  Leaves \a list empty. */
 void grpc_exec_ctx_enqueue_list(grpc_exec_ctx *exec_ctx,
diff --git a/src/core/lib/iomgr/iomgr.c b/src/core/lib/iomgr/iomgr.c
index 146663984d..60cef8ba77 100644
--- a/src/core/lib/iomgr/iomgr.c
+++ b/src/core/lib/iomgr/iomgr.c
@@ -166,8 +166,10 @@ bool grpc_iomgr_abort_on_leaks(void) {
   if (env == NULL) return false;
   static const char *truthy[] = {"yes",  "Yes",  "YES", "true",
                                  "True", "TRUE", "1"};
+  bool should_we = false;
   for (size_t i = 0; i < GPR_ARRAY_SIZE(truthy); i++) {
-    if (0 == strcmp(env, truthy[i])) return true;
+    if (0 == strcmp(env, truthy[i])) should_we = true;
   }
-  return false;
+  gpr_free(env);
+  return should_we;
 }
diff --git a/src/core/lib/iomgr/iomgr_posix.c b/src/core/lib/iomgr/iomgr_posix.c
index 016c501f75..cede97f4c6 100644
--- a/src/core/lib/iomgr/iomgr_posix.c
+++ b/src/core/lib/iomgr/iomgr_posix.c
@@ -41,12 +41,16 @@
 #include "src/core/lib/iomgr/tcp_posix.h"
 
 void grpc_iomgr_platform_init(void) {
+  grpc_wakeup_fd_global_init();
   grpc_event_engine_init();
   grpc_register_tracer("tcp", &grpc_tcp_trace);
 }
 
 void grpc_iomgr_platform_flush(void) {}
 
-void grpc_iomgr_platform_shutdown(void) { grpc_event_engine_shutdown(); }
+void grpc_iomgr_platform_shutdown(void) {
+  grpc_event_engine_shutdown();
+  grpc_wakeup_fd_global_destroy();
+}
 
 #endif /* GRPC_POSIX_SOCKET */
diff --git a/src/core/lib/iomgr/tcp_client_windows.c b/src/core/lib/iomgr/tcp_client_windows.c
index e41f67a04c..efa0140f54 100644
--- a/src/core/lib/iomgr/tcp_client_windows.c
+++ b/src/core/lib/iomgr/tcp_client_windows.c
@@ -63,39 +63,45 @@ typedef struct {
   grpc_endpoint **endpoint;
 } async_connect;
 
-static void async_connect_unlock_and_cleanup(async_connect *ac) {
+static void async_connect_unlock_and_cleanup(async_connect *ac,
+                                             grpc_winsocket *socket) {
   int done = (--ac->refs == 0);
   gpr_mu_unlock(&ac->mu);
   if (done) {
-    if (ac->socket != NULL) grpc_winsocket_destroy(ac->socket);
     gpr_mu_destroy(&ac->mu);
     gpr_free(ac->addr_name);
     gpr_free(ac);
   }
+  if (socket != NULL) grpc_winsocket_destroy(socket);
 }
 
 static void on_alarm(grpc_exec_ctx *exec_ctx, void *acp, bool occured) {
   async_connect *ac = acp;
   gpr_mu_lock(&ac->mu);
-  /* If the alarm didn't occur, it got cancelled. */
-  if (ac->socket != NULL && occured) {
+  if (ac->socket != NULL) {
     grpc_winsocket_shutdown(ac->socket);
   }
-  async_connect_unlock_and_cleanup(ac);
+  async_connect_unlock_and_cleanup(ac, ac->socket);
 }
 
 static void on_connect(grpc_exec_ctx *exec_ctx, void *acp, bool from_iocp) {
   async_connect *ac = acp;
   SOCKET sock = ac->socket->socket;
   grpc_endpoint **ep = ac->endpoint;
+  GPR_ASSERT(*ep == NULL);
   grpc_winsocket_callback_info *info = &ac->socket->write_info;
   grpc_closure *on_done = ac->on_done;
 
+  gpr_mu_lock(&ac->mu);
+  grpc_winsocket *socket = ac->socket;
+  ac->socket = NULL;
+  gpr_mu_unlock(&ac->mu);
+
   grpc_timer_cancel(exec_ctx, &ac->alarm);
 
   gpr_mu_lock(&ac->mu);
 
-  if (from_iocp) {
+  if (from_iocp && socket != NULL) {
     DWORD transfered_bytes = 0;
     DWORD flags;
     BOOL wsa_success = WSAGetOverlappedResult(sock, &info->overlapped,
@@ -107,12 +113,12 @@ static void on_connect(grpc_exec_ctx *exec_ctx, void *acp, bool from_iocp) {
               ac->addr_name, utf8_message);
       gpr_free(utf8_message);
     } else {
-      *ep = grpc_tcp_create(ac->socket, ac->addr_name);
-      ac->socket = NULL;
+      *ep = grpc_tcp_create(socket, ac->addr_name);
+      socket = NULL;
     }
   }
 
-  async_connect_unlock_and_cleanup(ac);
+  async_connect_unlock_and_cleanup(ac, socket);
   /* If the connection was aborted, the callback was already called when
      the deadline was met. */
   on_done->cb(exec_ctx, on_done->cb_arg, *ep != NULL);
@@ -138,6 +144,7 @@ void grpc_tcp_client_connect(grpc_exec_ctx *exec_ctx, grpc_closure *on_done,
   const char *message = NULL;
   char *utf8_message;
   grpc_winsocket_callback_info *info;
+  int last_error;
 
   *endpoint = NULL;
 
@@ -208,8 +215,10 @@ void grpc_tcp_client_connect(grpc_exec_ctx *exec_ctx, grpc_closure *on_done,
   return;
 
 failure:
-  utf8_message = gpr_format_message(WSAGetLastError());
+  last_error = WSAGetLastError();
+  utf8_message = gpr_format_message(last_error);
   gpr_log(GPR_ERROR, message, utf8_message);
+  gpr_log(GPR_ERROR, "last error = %d", last_error);
   gpr_free(utf8_message);
   if (socket != NULL) {
     grpc_winsocket_destroy(socket);
diff --git a/src/core/lib/iomgr/tcp_posix.c b/src/core/lib/iomgr/tcp_posix.c
index 7210aef5d5..e2869224f1 100644
--- a/src/core/lib/iomgr/tcp_posix.c
+++ b/src/core/lib/iomgr/tcp_posix.c
@@ -164,7 +164,7 @@ static void call_read_cb(grpc_exec_ctx *exec_ctx, grpc_tcp *tcp, int success) {
     for (i = 0; i < tcp->incoming_buffer->count; i++) {
       char *dump = gpr_dump_slice(tcp->incoming_buffer->slices[i],
                                   GPR_DUMP_HEX | GPR_DUMP_ASCII);
-      gpr_log(GPR_DEBUG, "READ %p: %s", tcp, dump);
+      gpr_log(GPR_DEBUG, "READ %p (peer=%s): %s", tcp, tcp->peer_string, dump);
       gpr_free(dump);
     }
   }
@@ -398,7 +398,7 @@ static void tcp_write(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep,
     for (i = 0; i < buf->count; i++) {
       char *data =
           gpr_dump_slice(buf->slices[i], GPR_DUMP_HEX | GPR_DUMP_ASCII);
-      gpr_log(GPR_DEBUG, "WRITE %p: %s", tcp, data);
+      gpr_log(GPR_DEBUG, "WRITE %p (peer=%s): %s", tcp, tcp->peer_string, data);
       gpr_free(data);
     }
   }
diff --git a/src/core/lib/iomgr/tcp_server.h b/src/core/lib/iomgr/tcp_server.h
index 99b9f29729..fee14ae661 100644
--- a/src/core/lib/iomgr/tcp_server.h
+++ b/src/core/lib/iomgr/tcp_server.h
@@ -52,6 +52,7 @@ typedef struct grpc_tcp_server_acceptor {
 /* Called for newly connected TCP connections. */
 typedef void (*grpc_tcp_server_cb)(grpc_exec_ctx *exec_ctx, void *arg,
                                    grpc_endpoint *ep,
+                                   grpc_pollset *accepting_pollset,
                                    grpc_tcp_server_acceptor *acceptor);
 
 /* Create a server, initially not bound to any ports. The caller owns one ref.
diff --git a/src/core/lib/iomgr/tcp_server_posix.c b/src/core/lib/iomgr/tcp_server_posix.c
index aaeb384f6e..909e34abc7 100644
--- a/src/core/lib/iomgr/tcp_server_posix.c
+++ b/src/core/lib/iomgr/tcp_server_posix.c
@@ -128,6 +128,9 @@ struct grpc_tcp_server {
   grpc_pollset **pollsets;
   /* number of pollsets in the pollsets array */
   size_t pollset_count;
+
+  /* next pollset to assign a channel to */
+  size_t next_pollset_to_assign;
 };
 
 grpc_tcp_server *grpc_tcp_server_create(grpc_closure *shutdown_complete) {
@@ -145,6 +148,7 @@ grpc_tcp_server *grpc_tcp_server_create(grpc_closure *shutdown_complete) {
   s->head = NULL;
   s->tail = NULL;
   s->nports = 0;
+  s->next_pollset_to_assign = 0;
   return s;
 }
 
@@ -310,13 +314,17 @@ static void on_read(grpc_exec_ctx *exec_ctx, void *arg, bool success) {
   grpc_tcp_listener *sp = arg;
   grpc_tcp_server_acceptor acceptor = {sp->server, sp->port_index,
                                        sp->fd_index};
+  grpc_pollset *read_notifier_pollset = NULL;
   grpc_fd *fdobj;
-  size_t i;
 
   if (!success) {
     goto error;
   }
 
+  read_notifier_pollset =
+      sp->server->pollsets[(sp->server->next_pollset_to_assign++) %
+                           sp->server->pollset_count];
+
   /* loop until accept4 returns EAGAIN, and then re-arm notification */
   for (;;) {
     struct sockaddr_storage addr;
@@ -349,16 +357,18 @@ static void on_read(grpc_exec_ctx *exec_ctx, void *arg, bool success) {
     }
 
     fdobj = grpc_fd_create(fd, name);
-    /* TODO(ctiller): revise this when we have server-side sharding
-       of channels -- we certainly should not be automatically adding every
-       incoming channel to every pollset owned by the server */
-    for (i = 0; i < sp->server->pollset_count; i++) {
-      grpc_pollset_add_fd(exec_ctx, sp->server->pollsets[i], fdobj);
+
+    if (read_notifier_pollset == NULL) {
+      gpr_log(GPR_ERROR, "Read notifier pollset is not set on the fd");
+      goto error;
     }
+
+    grpc_pollset_add_fd(exec_ctx, read_notifier_pollset, fdobj);
+
     sp->server->on_accept_cb(
         exec_ctx, sp->server->on_accept_cb_arg,
         grpc_tcp_create(fdobj, GRPC_TCP_DEFAULT_READ_SLICE_SIZE, addr_str),
-        &acceptor);
+        read_notifier_pollset, &acceptor);
 
     gpr_free(name);
     gpr_free(addr_str);
diff --git a/src/core/lib/iomgr/tcp_server_windows.c b/src/core/lib/iomgr/tcp_server_windows.c
index f32d53d240..8a21d930c8 100644
--- a/src/core/lib/iomgr/tcp_server_windows.c
+++ b/src/core/lib/iomgr/tcp_server_windows.c
@@ -379,9 +379,10 @@ static void on_accept(grpc_exec_ctx *exec_ctx, void *arg, bool from_iocp) {
 
   /* The only time we should call our callback, is where we successfully
      managed to accept a connection, and created an endpoint. */
-  if (ep)
-    sp->server->on_accept_cb(exec_ctx, sp->server->on_accept_cb_arg, ep,
+  if (ep) {
+    sp->server->on_accept_cb(exec_ctx, sp->server->on_accept_cb_arg, ep, NULL,
                              &acceptor);
+  }
   /* As we were notified from the IOCP of one and exactly one accept,
      the former socked we created has now either been destroy or assigned
      to the new connection. We need to create a new one for the next
@@ -508,34 +509,6 @@ int grpc_tcp_server_add_port(grpc_tcp_server *s, const void *addr,
   }
 }
 
-unsigned grpc_tcp_server_port_fd_count(grpc_tcp_server *s,
-                                       unsigned port_index) {
-  grpc_tcp_listener *sp;
-  for (sp = s->head; sp && port_index != 0; sp = sp->next, --port_index)
-    ;
-  if (sp) {
-    return 1;
-  } else {
-    return 0;
-  }
-}
-
-int grpc_tcp_server_port_fd(grpc_tcp_server *s, unsigned port_index,
-                            unsigned fd_index) {
-  grpc_tcp_listener *sp;
-  if (fd_index != 0) {
-    /* Windows implementation has only one fd per port_index. */
-    return -1;
-  }
-  for (sp = s->head; sp && port_index != 0; sp = sp->next, --port_index)
-    ;
-  if (sp) {
-    return _open_osfhandle((intptr_t)sp->socket->socket, 0);
-  } else {
-    return -1;
-  }
-}
-
 void grpc_tcp_server_start(grpc_exec_ctx *exec_ctx, grpc_tcp_server *s,
                            grpc_pollset **pollset, size_t pollset_count,
                            grpc_tcp_server_cb on_accept_cb,
diff --git a/src/core/lib/iomgr/tcp_windows.c b/src/core/lib/iomgr/tcp_windows.c
index a6715df5c9..f44bf5fda4 100644
--- a/src/core/lib/iomgr/tcp_windows.c
+++ b/src/core/lib/iomgr/tcp_windows.c
@@ -35,6 +35,8 @@
 
 #ifdef GPR_WINSOCK_SOCKET
 
+#include <limits.h>
+
 #include "src/core/lib/iomgr/sockaddr_windows.h"
 
 #include <grpc/support/alloc.h>
@@ -51,12 +53,20 @@
 #include "src/core/lib/iomgr/tcp_client.h"
 #include "src/core/lib/iomgr/timer.h"
 
+#if defined(__MSYS__) && defined(GPR_ARCH_64)
+/* Nasty workaround for nasty bug when using the 64 bits msys compiler
+   in conjunction with Microsoft Windows headers. */
+#define GRPC_FIONBIO _IOW('f', 126, uint32_t)
+#else
+#define GRPC_FIONBIO FIONBIO
+#endif
+
 static int set_non_block(SOCKET sock) {
   int status;
-  unsigned long param = 1;
+  uint32_t param = 1;
   DWORD ret;
-  status =
-      WSAIoctl(sock, FIONBIO, &param, sizeof(param), NULL, 0, &ret, NULL, NULL);
+  status = WSAIoctl(sock, GRPC_FIONBIO, &param, sizeof(param), NULL, 0, &ret,
+                    NULL, NULL);
   return status == 0;
 }
 
diff --git a/src/core/lib/iomgr/udp_server.c b/src/core/lib/iomgr/udp_server.c
index df6cf956d9..98ffccd59b 100644
--- a/src/core/lib/iomgr/udp_server.c
+++ b/src/core/lib/iomgr/udp_server.c
@@ -81,6 +81,7 @@ typedef struct {
   grpc_closure read_closure;
   grpc_closure destroyed_closure;
   grpc_udp_server_read_cb read_cb;
+  grpc_udp_server_orphan_cb orphan_cb;
 } server_port;
 
 /* the overall server */
@@ -168,6 +169,10 @@ static void deactivated_all_ports(grpc_exec_ctx *exec_ctx, grpc_udp_server *s) {
       server_port *sp = &s->ports[i];
       sp->destroyed_closure.cb = destroyed_port;
       sp->destroyed_closure.cb_arg = s;
+
+      GPR_ASSERT(sp->orphan_cb);
+      sp->orphan_cb(sp->emfd);
+
       grpc_fd_orphan(exec_ctx, sp->emfd, &sp->destroyed_closure, NULL,
                      "udp_listener_shutdown");
     }
@@ -268,7 +273,8 @@ static void on_read(grpc_exec_ctx *exec_ctx, void *arg, bool success) {
 
 static int add_socket_to_server(grpc_udp_server *s, int fd,
                                 const struct sockaddr *addr, size_t addr_len,
-                                grpc_udp_server_read_cb read_cb) {
+                                grpc_udp_server_read_cb read_cb,
+                                grpc_udp_server_orphan_cb orphan_cb) {
   server_port *sp;
   int port;
   char *addr_str;
@@ -292,6 +298,7 @@ static int add_socket_to_server(grpc_udp_server *s, int fd,
     memcpy(sp->addr.untyped, addr, addr_len);
     sp->addr_len = addr_len;
     sp->read_cb = read_cb;
+    sp->orphan_cb = orphan_cb;
     GPR_ASSERT(sp->emfd);
     gpr_mu_unlock(&s->mu);
     gpr_free(name);
@@ -301,7 +308,8 @@ static int add_socket_to_server(grpc_udp_server *s, int fd,
 }
 
 int grpc_udp_server_add_port(grpc_udp_server *s, const void *addr,
-                             size_t addr_len, grpc_udp_server_read_cb read_cb) {
+                             size_t addr_len, grpc_udp_server_read_cb read_cb,
+                             grpc_udp_server_orphan_cb orphan_cb) {
   int allocated_port1 = -1;
   int allocated_port2 = -1;
   unsigned i;
@@ -348,7 +356,8 @@ int grpc_udp_server_add_port(grpc_udp_server *s, const void *addr,
     addr = (struct sockaddr *)&wild6;
     addr_len = sizeof(wild6);
     fd = grpc_create_dualstack_socket(addr, SOCK_DGRAM, IPPROTO_UDP, &dsmode);
-    allocated_port1 = add_socket_to_server(s, fd, addr, addr_len, read_cb);
+    allocated_port1 =
+        add_socket_to_server(s, fd, addr, addr_len, read_cb, orphan_cb);
     if (fd >= 0 && dsmode == GRPC_DSMODE_DUALSTACK) {
       goto done;
     }
@@ -370,7 +379,8 @@ int grpc_udp_server_add_port(grpc_udp_server *s, const void *addr,
     addr = (struct sockaddr *)&addr4_copy;
     addr_len = sizeof(addr4_copy);
   }
-  allocated_port2 = add_socket_to_server(s, fd, addr, addr_len, read_cb);
+  allocated_port2 =
+      add_socket_to_server(s, fd, addr, addr_len, read_cb, orphan_cb);
 
 done:
   gpr_free(allocated_addr);
diff --git a/src/core/lib/iomgr/udp_server.h b/src/core/lib/iomgr/udp_server.h
index d8cf957a22..33c5ce11cd 100644
--- a/src/core/lib/iomgr/udp_server.h
+++ b/src/core/lib/iomgr/udp_server.h
@@ -48,6 +48,9 @@ typedef struct grpc_udp_server grpc_udp_server;
 typedef void (*grpc_udp_server_read_cb)(grpc_exec_ctx *exec_ctx, grpc_fd *emfd,
                                         struct grpc_server *server);
 
+/* Called when the grpc_fd is about to be orphaned (and the FD closed). */
+typedef void (*grpc_udp_server_orphan_cb)(grpc_fd *emfd);
+
 /* Create a server, initially not bound to any ports */
 grpc_udp_server *grpc_udp_server_create(void);
 
@@ -69,7 +72,8 @@ int grpc_udp_server_get_fd(grpc_udp_server *s, unsigned index);
 /* TODO(ctiller): deprecate this, and make grpc_udp_server_add_ports to handle
                   all of the multiple socket port matching logic in one place */
 int grpc_udp_server_add_port(grpc_udp_server *s, const void *addr,
-                             size_t addr_len, grpc_udp_server_read_cb read_cb);
+                             size_t addr_len, grpc_udp_server_read_cb read_cb,
+                             grpc_udp_server_orphan_cb orphan_cb);
 
 void grpc_udp_server_destroy(grpc_exec_ctx *exec_ctx, grpc_udp_server *server,
                              grpc_closure *on_done);