Merge pull request #3610 from ctiller/stressy

Fix thread stuck problem on Mac

6 files changed, 250 insertions, 181 deletions

diff --git a/src/core/iomgr/fd_posix.c b/src/core/iomgr/fd_posix.c
index b48b7f050a..231bc988a8 100644
--- a/src/core/iomgr/fd_posix.c
+++ b/src/core/iomgr/fd_posix.c
@@ -45,10 +45,8 @@
 #include <grpc/support/log.h>
 #include <grpc/support/useful.h>
 
-enum descriptor_state {
-  NOT_READY = 0,
-  READY = 1
-}; /* or a pointer to a closure to call */
+#define CLOSURE_NOT_READY ((grpc_closure *)0)
+#define CLOSURE_READY ((grpc_closure *)1)
 
 /* 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)
@@ -88,14 +86,13 @@ static grpc_fd *alloc_fd(int fd) {
   gpr_mu_unlock(&fd_freelist_mu);
   if (r == NULL) {
     r = gpr_malloc(sizeof(grpc_fd));
-    gpr_mu_init(&r->set_state_mu);
-    gpr_mu_init(&r->watcher_mu);
+    gpr_mu_init(&r->mu);
   }
 
   gpr_atm_rel_store(&r->refst, 1);
-  gpr_atm_rel_store(&r->readst, NOT_READY);
-  gpr_atm_rel_store(&r->writest, NOT_READY);
-  gpr_atm_rel_store(&r->shutdown, 0);
+  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;
@@ -107,8 +104,7 @@ static grpc_fd *alloc_fd(int fd) {
 }
 
 static void destroy(grpc_fd *fd) {
-  gpr_mu_destroy(&fd->set_state_mu);
-  gpr_mu_destroy(&fd->watcher_mu);
+  gpr_mu_destroy(&fd->mu);
   gpr_free(fd);
 }
 
@@ -173,39 +169,35 @@ int grpc_fd_is_orphaned(grpc_fd *fd) {
   return (gpr_atm_acq_load(&fd->refst) & 1) == 0;
 }
 
-static void pollset_kick_locked(grpc_pollset *pollset) {
-  gpr_mu_lock(GRPC_POLLSET_MU(pollset));
-  grpc_pollset_kick(pollset, NULL);
-  gpr_mu_unlock(GRPC_POLLSET_MU(pollset));
+static void pollset_kick_locked(grpc_fd_watcher *watcher) {
+  gpr_mu_lock(GRPC_POLLSET_MU(watcher->pollset));
+  GPR_ASSERT(watcher->worker);
+  grpc_pollset_kick_ext(watcher->pollset, watcher->worker,
+                        GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP);
+  gpr_mu_unlock(GRPC_POLLSET_MU(watcher->pollset));
 }
 
 static void maybe_wake_one_watcher_locked(grpc_fd *fd) {
   if (fd->inactive_watcher_root.next != &fd->inactive_watcher_root) {
-    pollset_kick_locked(fd->inactive_watcher_root.next->pollset);
+    pollset_kick_locked(fd->inactive_watcher_root.next);
   } else if (fd->read_watcher) {
-    pollset_kick_locked(fd->read_watcher->pollset);
+    pollset_kick_locked(fd->read_watcher);
   } else if (fd->write_watcher) {
-    pollset_kick_locked(fd->write_watcher->pollset);
+    pollset_kick_locked(fd->write_watcher);
   }
 }
 
-static void maybe_wake_one_watcher(grpc_fd *fd) {
-  gpr_mu_lock(&fd->watcher_mu);
-  maybe_wake_one_watcher_locked(fd);
-  gpr_mu_unlock(&fd->watcher_mu);
-}
-
 static void wake_all_watchers_locked(grpc_fd *fd) {
   grpc_fd_watcher *watcher;
   for (watcher = fd->inactive_watcher_root.next;
        watcher != &fd->inactive_watcher_root; watcher = watcher->next) {
-    pollset_kick_locked(watcher->pollset);
+    pollset_kick_locked(watcher);
   }
   if (fd->read_watcher) {
-    pollset_kick_locked(fd->read_watcher->pollset);
+    pollset_kick_locked(fd->read_watcher);
   }
   if (fd->write_watcher && fd->write_watcher != fd->read_watcher) {
-    pollset_kick_locked(fd->write_watcher->pollset);
+    pollset_kick_locked(fd->write_watcher);
   }
 }
 
@@ -218,7 +210,7 @@ void grpc_fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure *on_done,
                     const char *reason) {
   fd->on_done_closure = on_done;
   shutdown(fd->fd, SHUT_RDWR);
-  gpr_mu_lock(&fd->watcher_mu);
+  gpr_mu_lock(&fd->mu);
   REF_BY(fd, 1, reason); /* remove active status, but keep referenced */
   if (!has_watchers(fd)) {
     fd->closed = 1;
@@ -227,7 +219,7 @@ void grpc_fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure *on_done,
   } else {
     wake_all_watchers_locked(fd);
   }
-  gpr_mu_unlock(&fd->watcher_mu);
+  gpr_mu_unlock(&fd->mu);
   UNREF_BY(fd, 2, reason); /* drop the reference */
 }
 
@@ -247,136 +239,121 @@ void grpc_fd_ref(grpc_fd *fd) { ref_by(fd, 2); }
 void grpc_fd_unref(grpc_fd *fd) { unref_by(fd, 2); }
 #endif
 
-static void notify_on(grpc_exec_ctx *exec_ctx, grpc_fd *fd, gpr_atm *st,
-                      grpc_closure *closure) {
-  switch (gpr_atm_acq_load(st)) {
-    case NOT_READY:
-      /* There is no race if the descriptor is already ready, so we skip
-         the interlocked op in that case.  As long as the app doesn't
-         try to set the same upcall twice (which it shouldn't) then
-         oldval should never be anything other than READY or NOT_READY.  We
-         don't
-         check for user error on the fast path. */
-      if (gpr_atm_rel_cas(st, NOT_READY, (gpr_intptr)closure)) {
-        /* swap was successful -- the closure will run after the next
-           set_ready call.  NOTE: we don't have an ABA problem here,
-           since we should never have concurrent calls to the same
-           notify_on function. */
-        maybe_wake_one_watcher(fd);
-        return;
-      }
-    /* swap was unsuccessful due to an intervening set_ready call.
-       Fall through to the READY code below */
-    case READY:
-      GPR_ASSERT(gpr_atm_no_barrier_load(st) == READY);
-      gpr_atm_rel_store(st, NOT_READY);
-      grpc_exec_ctx_enqueue(exec_ctx, closure,
-                            !gpr_atm_acq_load(&fd->shutdown));
-      return;
-    default: /* WAITING */
-      /* upcallptr was set to a different closure.  This is an error! */
-      gpr_log(GPR_ERROR,
-              "User called a notify_on function with a previous callback still "
-              "pending");
-      abort();
+static void notify_on_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
+                             grpc_closure **st, grpc_closure *closure) {
+  if (*st == CLOSURE_NOT_READY) {
+    /* not ready ==> switch to a waiting state by setting the closure */
+    *st = closure;
+  } else if (*st == CLOSURE_READY) {
+    /* already ready ==> queue the closure to run immediately */
+    *st = CLOSURE_NOT_READY;
+    grpc_exec_ctx_enqueue(exec_ctx, closure, !fd->shutdown);
+    maybe_wake_one_watcher_locked(fd);
+  } else {
+    /* upcallptr was set to a different closure.  This is an error! */
+    gpr_log(GPR_ERROR,
+            "User called a notify_on function with a previous callback still "
+            "pending");
+    abort();
   }
-  gpr_log(GPR_ERROR, "Corrupt memory in &st->state");
-  abort();
 }
 
-static void set_ready_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
-                             gpr_atm *st) {
-  gpr_intptr state = gpr_atm_acq_load(st);
-
-  switch (state) {
-    case READY:
-      /* duplicate ready, ignore */
-      return;
-    case NOT_READY:
-      if (gpr_atm_rel_cas(st, NOT_READY, READY)) {
-        /* swap was successful -- the closure will run after the next
-           notify_on call. */
-        return;
-      }
-      /* swap was unsuccessful due to an intervening set_ready call.
-         Fall through to the WAITING code below */
-      state = gpr_atm_acq_load(st);
-    default: /* waiting */
-      GPR_ASSERT(gpr_atm_no_barrier_load(st) != READY &&
-                 gpr_atm_no_barrier_load(st) != NOT_READY);
-      grpc_exec_ctx_enqueue(exec_ctx, (grpc_closure *)state,
-                            !gpr_atm_acq_load(&fd->shutdown));
-      gpr_atm_rel_store(st, NOT_READY);
-      return;
+/* returns 1 if state becomes not ready */
+static int set_ready_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
+                            grpc_closure **st) {
+  if (*st == CLOSURE_READY) {
+    /* duplicate ready ==> ignore */
+    return 0;
+  } else if (*st == CLOSURE_NOT_READY) {
+    /* not ready, and not waiting ==> flag ready */
+    *st = CLOSURE_READY;
+    return 0;
+  } else {
+    /* waiting ==> queue closure */
+    grpc_exec_ctx_enqueue(exec_ctx, *st, !fd->shutdown);
+    *st = CLOSURE_NOT_READY;
+    return 1;
   }
 }
 
-static void set_ready(grpc_exec_ctx *exec_ctx, grpc_fd *fd, gpr_atm *st) {
+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->set_state_mu);
+  gpr_mu_lock(&fd->mu);
   set_ready_locked(exec_ctx, fd, st);
-  gpr_mu_unlock(&fd->set_state_mu);
+  gpr_mu_unlock(&fd->mu);
 }
 
 void grpc_fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd) {
-  gpr_mu_lock(&fd->set_state_mu);
+  gpr_mu_lock(&fd->mu);
   GPR_ASSERT(!gpr_atm_no_barrier_load(&fd->shutdown));
-  gpr_atm_rel_store(&fd->shutdown, 1);
-  set_ready_locked(exec_ctx, fd, &fd->readst);
-  set_ready_locked(exec_ctx, fd, &fd->writest);
-  gpr_mu_unlock(&fd->set_state_mu);
+  fd->shutdown = 1;
+  set_ready_locked(exec_ctx, fd, &fd->read_closure);
+  set_ready_locked(exec_ctx, fd, &fd->write_closure);
+  gpr_mu_unlock(&fd->mu);
 }
 
 void grpc_fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
                             grpc_closure *closure) {
-  notify_on(exec_ctx, fd, &fd->readst, closure);
+  gpr_mu_lock(&fd->mu);
+  notify_on_locked(exec_ctx, fd, &fd->read_closure, closure);
+  gpr_mu_unlock(&fd->mu);
 }
 
 void grpc_fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
                              grpc_closure *closure) {
-  notify_on(exec_ctx, fd, &fd->writest, closure);
+  gpr_mu_lock(&fd->mu);
+  notify_on_locked(exec_ctx, fd, &fd->write_closure, closure);
+  gpr_mu_unlock(&fd->mu);
 }
 
 gpr_uint32 grpc_fd_begin_poll(grpc_fd *fd, grpc_pollset *pollset,
-                              gpr_uint32 read_mask, gpr_uint32 write_mask,
-                              grpc_fd_watcher *watcher) {
+                              grpc_pollset_worker *worker, gpr_uint32 read_mask,
+                              gpr_uint32 write_mask, grpc_fd_watcher *watcher) {
   gpr_uint32 mask = 0;
+  grpc_closure *cur;
+  int requested;
   /* keep track of pollers that have requested our events, in case they change
    */
   GRPC_FD_REF(fd, "poll");
 
-  gpr_mu_lock(&fd->watcher_mu);
+  gpr_mu_lock(&fd->mu);
+
   /* if we are shutdown, then don't add to the watcher set */
   if (gpr_atm_no_barrier_load(&fd->shutdown)) {
     watcher->fd = NULL;
     watcher->pollset = NULL;
-    gpr_mu_unlock(&fd->watcher_mu);
+    watcher->worker = NULL;
+    gpr_mu_unlock(&fd->mu);
     GRPC_FD_UNREF(fd, "poll");
     return 0;
   }
+
   /* if there is nobody polling for read, but we need to, then start doing so */
-  if (read_mask && !fd->read_watcher &&
-      (gpr_uintptr)gpr_atm_acq_load(&fd->readst) > READY) {
+  cur = fd->read_closure;
+  requested = cur != CLOSURE_READY;
+  if (read_mask && fd->read_watcher == NULL && requested) {
     fd->read_watcher = watcher;
     mask |= read_mask;
   }
   /* if there is nobody polling for write, but we need to, then start doing so
    */
-  if (write_mask && !fd->write_watcher &&
-      (gpr_uintptr)gpr_atm_acq_load(&fd->writest) > READY) {
+  cur = fd->write_closure;
+  requested = cur != CLOSURE_READY;
+  if (write_mask && fd->write_watcher == NULL && requested) {
     fd->write_watcher = watcher;
     mask |= write_mask;
   }
   /* if not polling, remember this watcher in case we need someone to later */
-  if (mask == 0) {
+  if (mask == 0 && worker != NULL) {
     watcher->next = &fd->inactive_watcher_root;
     watcher->prev = watcher->next->prev;
     watcher->next->prev = watcher->prev->next = watcher;
   }
   watcher->pollset = pollset;
+  watcher->worker = worker;
   watcher->fd = fd;
-  gpr_mu_unlock(&fd->watcher_mu);
+  gpr_mu_unlock(&fd->mu);
 
   return mask;
 }
@@ -391,24 +368,39 @@ void grpc_fd_end_poll(grpc_exec_ctx *exec_ctx, grpc_fd_watcher *watcher,
     return;
   }
 
-  gpr_mu_lock(&fd->watcher_mu);
+  gpr_mu_lock(&fd->mu);
+
   if (watcher == fd->read_watcher) {
     /* remove read watcher, kick if we still need a read */
     was_polling = 1;
-    kick = kick || !got_read;
+    if (!got_read) {
+      kick = 1;
+    }
     fd->read_watcher = NULL;
   }
   if (watcher == fd->write_watcher) {
     /* remove write watcher, kick if we still need a write */
     was_polling = 1;
-    kick = kick || !got_write;
+    if (!got_write) {
+      kick = 1;
+    }
     fd->write_watcher = NULL;
   }
-  if (!was_polling) {
+  if (!was_polling && watcher->worker != NULL) {
     /* remove from inactive list */
     watcher->next->prev = watcher->prev;
     watcher->prev->next = watcher->next;
   }
+  if (got_read) {
+    if (set_ready_locked(exec_ctx, fd, &fd->read_closure)) {
+      kick = 1;
+    }
+  }
+  if (got_write) {
+    if (set_ready_locked(exec_ctx, fd, &fd->write_closure)) {
+      kick = 1;
+    }
+  }
   if (kick) {
     maybe_wake_one_watcher_locked(fd);
   }
@@ -417,17 +409,17 @@ void grpc_fd_end_poll(grpc_exec_ctx *exec_ctx, grpc_fd_watcher *watcher,
     close(fd->fd);
     grpc_exec_ctx_enqueue(exec_ctx, fd->on_done_closure, 1);
   }
-  gpr_mu_unlock(&fd->watcher_mu);
+  gpr_mu_unlock(&fd->mu);
 
   GRPC_FD_UNREF(fd, "poll");
 }
 
 void grpc_fd_become_readable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) {
-  set_ready(exec_ctx, fd, &fd->readst);
+  set_ready(exec_ctx, fd, &fd->read_closure);
 }
 
 void grpc_fd_become_writable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) {
-  set_ready(exec_ctx, fd, &fd->writest);
+  set_ready(exec_ctx, fd, &fd->write_closure);
 }
 
 #endif
diff --git a/src/core/iomgr/fd_posix.h b/src/core/iomgr/fd_posix.h
index 089aa4d717..dc917ebbc0 100644
--- a/src/core/iomgr/fd_posix.h
+++ b/src/core/iomgr/fd_posix.h
@@ -46,6 +46,7 @@ typedef struct grpc_fd_watcher {
   struct grpc_fd_watcher *next;
   struct grpc_fd_watcher *prev;
   grpc_pollset *pollset;
+  grpc_pollset_worker *worker;
   grpc_fd *fd;
 } grpc_fd_watcher;
 
@@ -58,8 +59,8 @@ struct grpc_fd {
      and just unref by 1 when we're ready to flag the object as orphaned */
   gpr_atm refst;
 
-  gpr_mu set_state_mu;
-  gpr_atm shutdown;
+  gpr_mu mu;
+  int shutdown;
   int closed;
 
   /* The watcher list.
@@ -84,18 +85,16 @@ struct grpc_fd {
      If at a later time there becomes need of a poller to poll, one of
      the inactive pollers may be kicked out of their poll loops to take
      that responsibility. */
-  gpr_mu watcher_mu;
   grpc_fd_watcher inactive_watcher_root;
   grpc_fd_watcher *read_watcher;
   grpc_fd_watcher *write_watcher;
 
-  gpr_atm readst;
-  gpr_atm writest;
+  grpc_closure *read_closure;
+  grpc_closure *write_closure;
 
   struct grpc_fd *freelist_next;
 
   grpc_closure *on_done_closure;
-  grpc_closure *shutdown_closures[2];
 
   grpc_iomgr_object iomgr_object;
 };
@@ -126,10 +125,12 @@ void grpc_fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure *on_done,
    fd's current interest (such as epoll) do not need to call this function.
    MUST NOT be called with a pollset lock taken */
 gpr_uint32 grpc_fd_begin_poll(grpc_fd *fd, grpc_pollset *pollset,
-                              gpr_uint32 read_mask, gpr_uint32 write_mask,
-                              grpc_fd_watcher *rec);
+                              grpc_pollset_worker *worker, gpr_uint32 read_mask,
+                              gpr_uint32 write_mask, grpc_fd_watcher *rec);
 /* Complete polling previously started with grpc_fd_begin_poll
-   MUST NOT be called with a pollset lock taken */
+   MUST NOT be called with a pollset lock taken
+   if got_read or got_write are 1, also does the become_{readable,writable} as
+   appropriate. */
 void grpc_fd_end_poll(grpc_exec_ctx *exec_ctx, grpc_fd_watcher *rec,
                       int got_read, int got_write);
 
diff --git a/src/core/iomgr/pollset_multipoller_with_epoll.c b/src/core/iomgr/pollset_multipoller_with_epoll.c
index faf0a6362b..ba9ba73f9d 100644
--- a/src/core/iomgr/pollset_multipoller_with_epoll.c
+++ b/src/core/iomgr/pollset_multipoller_with_epoll.c
@@ -72,7 +72,7 @@ static void finally_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
   /* 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(grpc_fd_begin_poll(fd, pollset, 0, 0, &watcher) == 0);
+  GPR_ASSERT(grpc_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;
diff --git a/src/core/iomgr/pollset_multipoller_with_poll_posix.c b/src/core/iomgr/pollset_multipoller_with_poll_posix.c
index 1356ebe7a0..faa6c14491 100644
--- a/src/core/iomgr/pollset_multipoller_with_poll_posix.c
+++ b/src/core/iomgr/pollset_multipoller_with_poll_posix.c
@@ -102,6 +102,9 @@ static void multipoll_with_poll_pollset_del_fd(grpc_exec_ctx *exec_ctx,
 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;
@@ -147,8 +150,8 @@ static void multipoll_with_poll_pollset_maybe_work_and_unlock(
   gpr_mu_unlock(&pollset->mu);
 
   for (i = 2; i < pfd_count; i++) {
-    pfds[i].events = (short)grpc_fd_begin_poll(watchers[i].fd, pollset, POLLIN,
-                                               POLLOUT, &watchers[i]);
+    pfds[i].events = (short)grpc_fd_begin_poll(watchers[i].fd, pollset, worker,
+                                               POLLIN, POLLOUT, &watchers[i]);
   }
 
   /* TODO(vpai): Consider first doing a 0 timeout poll here to avoid
@@ -157,34 +160,29 @@ static void multipoll_with_poll_pollset_maybe_work_and_unlock(
   r = grpc_poll_function(pfds, pfd_count, timeout);
   GRPC_SCHEDULING_END_BLOCKING_REGION;
 
-  for (i = 2; i < pfd_count; i++) {
-    grpc_fd_end_poll(exec_ctx, &watchers[i], pfds[i].revents & POLLIN,
-                     pfds[i].revents & POLLOUT);
-  }
-
   if (r < 0) {
-    if (errno != EINTR) {
-      gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno));
+    gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno));
+    for (i = 2; i < pfd_count; i++) {
+      grpc_fd_end_poll(exec_ctx, &watchers[i], 0, 0);
     }
   } else if (r == 0) {
-    /* do nothing */
+    for (i = 2; i < pfd_count; i++) {
+      grpc_fd_end_poll(exec_ctx, &watchers[i], 0, 0);
+    }
   } else {
-    if (pfds[0].revents & POLLIN) {
+    if (pfds[0].revents & POLLIN_CHECK) {
       grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd);
     }
-    if (pfds[1].revents & POLLIN) {
+    if (pfds[1].revents & POLLIN_CHECK) {
       grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd);
     }
     for (i = 2; i < pfd_count; i++) {
       if (watchers[i].fd == NULL) {
+        grpc_fd_end_poll(exec_ctx, &watchers[i], 0, 0);
         continue;
       }
-      if (pfds[i].revents & (POLLIN | POLLHUP | POLLERR)) {
-        grpc_fd_become_readable(exec_ctx, watchers[i].fd);
-      }
-      if (pfds[i].revents & (POLLOUT | POLLHUP | POLLERR)) {
-        grpc_fd_become_writable(exec_ctx, watchers[i].fd);
-      }
+      grpc_fd_end_poll(exec_ctx, &watchers[i], pfds[i].revents & POLLIN_CHECK,
+                       pfds[i].revents & POLLOUT_CHECK);
     }
   }
 
diff --git a/src/core/iomgr/pollset_posix.c b/src/core/iomgr/pollset_posix.c
index b663780a02..d056866d13 100644
--- a/src/core/iomgr/pollset_posix.c
+++ b/src/core/iomgr/pollset_posix.c
@@ -98,31 +98,63 @@ static void push_front_worker(grpc_pollset *p, grpc_pollset_worker *worker) {
   worker->prev->next = worker->next->prev = worker;
 }
 
-void grpc_pollset_kick(grpc_pollset *p, grpc_pollset_worker *specific_worker) {
+void grpc_pollset_kick_ext(grpc_pollset *p,
+                           grpc_pollset_worker *specific_worker,
+                           gpr_uint32 flags) {
   /* pollset->mu already held */
   if (specific_worker != NULL) {
     if (specific_worker == GRPC_POLLSET_KICK_BROADCAST) {
+      GPR_ASSERT((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) == 0);
       for (specific_worker = p->root_worker.next;
            specific_worker != &p->root_worker;
            specific_worker = specific_worker->next) {
         grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd);
       }
       p->kicked_without_pollers = 1;
+      return;
     } else if (gpr_tls_get(&g_current_thread_worker) !=
                (gpr_intptr)specific_worker) {
+      if ((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) != 0) {
+        specific_worker->reevaluate_polling_on_wakeup = 1;
+      }
+      grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd);
+      return;
+    } else if ((flags & GRPC_POLLSET_CAN_KICK_SELF) != 0) {
+      if ((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) != 0) {
+        specific_worker->reevaluate_polling_on_wakeup = 1;
+      }
       grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd);
+      return;
     }
   } else if (gpr_tls_get(&g_current_thread_poller) != (gpr_intptr)p) {
+    GPR_ASSERT((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) == 0);
     specific_worker = pop_front_worker(p);
     if (specific_worker != NULL) {
+      if (gpr_tls_get(&g_current_thread_worker) ==
+          (gpr_intptr)specific_worker) {
+        push_back_worker(p, specific_worker);
+        specific_worker = pop_front_worker(p);
+        if ((flags & GRPC_POLLSET_CAN_KICK_SELF) == 0 &&
+            gpr_tls_get(&g_current_thread_worker) ==
+                (gpr_intptr)specific_worker) {
+          push_back_worker(p, specific_worker);
+          return;
+        }
+      }
       push_back_worker(p, specific_worker);
       grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd);
+      return;
     } else {
       p->kicked_without_pollers = 1;
+      return;
     }
   }
 }
 
+void grpc_pollset_kick(grpc_pollset *p, grpc_pollset_worker *specific_worker) {
+  grpc_pollset_kick_ext(p, specific_worker, 0);
+}
+
 /* global state management */
 
 void grpc_pollset_global_init(void) {
@@ -195,52 +227,88 @@ void grpc_pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
   /* pollset->mu already held */
   int added_worker = 0;
   int locked = 1;
+  int queued_work = 0;
+  int keep_polling = 0;
   /* this must happen before we (potentially) drop pollset->mu */
   worker->next = worker->prev = NULL;
+  worker->reevaluate_polling_on_wakeup = 0;
   /* TODO(ctiller): pool these */
   grpc_wakeup_fd_init(&worker->wakeup_fd);
+  /* If there's work waiting for the pollset to be idle, and the
+     pollset is idle, then do that work */
   if (!grpc_pollset_has_workers(pollset) &&
       !grpc_closure_list_empty(pollset->idle_jobs)) {
     grpc_exec_ctx_enqueue_list(exec_ctx, &pollset->idle_jobs);
     goto done;
   }
+  /* Check alarms - these are a global resource so we just ping
+     each time through on every pollset.
+     May update deadline to ensure timely wakeups.
+     TODO(ctiller): can this work be localized? */
   if (grpc_alarm_check(exec_ctx, now, &deadline)) {
     gpr_mu_unlock(&pollset->mu);
     locked = 0;
     goto done;
   }
+  /* If we're shutting down then we don't execute any extended work */
   if (pollset->shutting_down) {
     goto done;
   }
+  /* Give do_promote priority so we don't starve it out */
   if (pollset->in_flight_cbs) {
-    /* Give do_promote priority so we don't starve it out */
     gpr_mu_unlock(&pollset->mu);
     locked = 0;
     goto done;
   }
-  if (!pollset->kicked_without_pollers) {
-    push_front_worker(pollset, worker);
-    added_worker = 1;
-    gpr_tls_set(&g_current_thread_poller, (gpr_intptr)pollset);
-    gpr_tls_set(&g_current_thread_worker, (gpr_intptr)worker);
-    pollset->vtable->maybe_work_and_unlock(exec_ctx, pollset, worker, deadline,
-                                           now);
-    locked = 0;
-    gpr_tls_set(&g_current_thread_poller, 0);
-    gpr_tls_set(&g_current_thread_worker, 0);
-  } else {
-    pollset->kicked_without_pollers = 0;
-  }
-done:
-  if (!locked) {
-    grpc_exec_ctx_flush(exec_ctx);
-    gpr_mu_lock(&pollset->mu);
-    locked = 1;
+  /* 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;
+  while (keep_polling) {
+    keep_polling = 0;
+    if (!pollset->kicked_without_pollers) {
+      if (!added_worker) {
+        push_front_worker(pollset, worker);
+        added_worker = 1;
+      }
+      gpr_tls_set(&g_current_thread_poller, (gpr_intptr)pollset);
+      gpr_tls_set(&g_current_thread_worker, (gpr_intptr)worker);
+      pollset->vtable->maybe_work_and_unlock(exec_ctx, pollset, worker,
+                                             deadline, now);
+      locked = 0;
+      gpr_tls_set(&g_current_thread_poller, 0);
+      gpr_tls_set(&g_current_thread_worker, 0);
+    } else {
+      pollset->kicked_without_pollers = 0;
+    }
+  /* Finished execution - start cleaning up.
+     Note that we may arrive here from outside the enclosing while() loop.
+     In that case we won't loop though as we haven't added worker to the
+     worker list, which means nobody could ask us to re-evaluate polling). */
+  done:
+    if (!locked) {
+      queued_work |= grpc_exec_ctx_flush(exec_ctx);
+      gpr_mu_lock(&pollset->mu);
+      locked = 1;
+    }
+    /* If we're forced to re-evaluate polling (via grpc_pollset_kick with
+       GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) then we land here and force
+       a loop */
+    if (worker->reevaluate_polling_on_wakeup) {
+      worker->reevaluate_polling_on_wakeup = 0;
+      pollset->kicked_without_pollers = 0;
+      if (queued_work) {
+        /* If there's queued work on the list, then set the deadline to be
+           immediate so we get back out of the polling loop quickly */
+        deadline = gpr_inf_past(GPR_CLOCK_MONOTONIC);
+      }
+      keep_polling = 1;
+    }
   }
-  grpc_wakeup_fd_destroy(&worker->wakeup_fd);
   if (added_worker) {
     remove_worker(pollset, worker);
   }
+  grpc_wakeup_fd_destroy(&worker->wakeup_fd);
   if (pollset->shutting_down) {
     if (grpc_pollset_has_workers(pollset)) {
       grpc_pollset_kick(pollset, NULL);
@@ -454,6 +522,9 @@ static void basic_pollset_maybe_work_and_unlock(grpc_exec_ctx *exec_ctx,
                                                 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;
@@ -479,8 +550,8 @@ static void basic_pollset_maybe_work_and_unlock(grpc_exec_ctx *exec_ctx,
     pfd[2].revents = 0;
     GRPC_FD_REF(fd, "basicpoll_begin");
     gpr_mu_unlock(&pollset->mu);
-    pfd[2].events =
-        (short)grpc_fd_begin_poll(fd, pollset, POLLIN, POLLOUT, &fd_watcher);
+    pfd[2].events = (short)grpc_fd_begin_poll(fd, pollset, worker, POLLIN,
+                                              POLLOUT, &fd_watcher);
     if (pfd[2].events != 0) {
       nfds++;
     }
@@ -497,31 +568,27 @@ static void basic_pollset_maybe_work_and_unlock(grpc_exec_ctx *exec_ctx,
   GRPC_SCHEDULING_END_BLOCKING_REGION;
   GRPC_TIMER_MARK(GRPC_PTAG_POLL_FINISHED, r);
 
-  if (fd) {
-    grpc_fd_end_poll(exec_ctx, &fd_watcher, pfd[2].revents & POLLIN,
-                     pfd[2].revents & POLLOUT);
-  }
-
   if (r < 0) {
-    if (errno != EINTR) {
-      gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno));
+    gpr_log(GPR_ERROR, "poll() failed: %s", strerror(errno));
+    if (fd) {
+      grpc_fd_end_poll(exec_ctx, &fd_watcher, 0, 0);
     }
   } else if (r == 0) {
-    /* do nothing */
+    if (fd) {
+      grpc_fd_end_poll(exec_ctx, &fd_watcher, 0, 0);
+    }
   } else {
-    if (pfd[0].revents & POLLIN) {
+    if (pfd[0].revents & POLLIN_CHECK) {
       grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd);
     }
-    if (pfd[1].revents & POLLIN) {
+    if (pfd[1].revents & POLLIN_CHECK) {
       grpc_wakeup_fd_consume_wakeup(&worker->wakeup_fd);
     }
     if (nfds > 2) {
-      if (pfd[2].revents & (POLLIN | POLLHUP | POLLERR)) {
-        grpc_fd_become_readable(exec_ctx, fd);
-      }
-      if (pfd[2].revents & (POLLOUT | POLLHUP | POLLERR)) {
-        grpc_fd_become_writable(exec_ctx, fd);
-      }
+      grpc_fd_end_poll(exec_ctx, &fd_watcher, pfd[2].revents & POLLIN_CHECK,
+                       pfd[2].revents & POLLOUT_CHECK);
+    } else if (fd) {
+      grpc_fd_end_poll(exec_ctx, &fd_watcher, 0, 0);
     }
   }
 
diff --git a/src/core/iomgr/pollset_posix.h b/src/core/iomgr/pollset_posix.h
index 83c5258539..34f76db2af 100644
--- a/src/core/iomgr/pollset_posix.h
+++ b/src/core/iomgr/pollset_posix.h
@@ -50,6 +50,7 @@ struct grpc_fd;
 
 typedef struct grpc_pollset_worker {
   grpc_wakeup_fd wakeup_fd;
+  int reevaluate_polling_on_wakeup;
   struct grpc_pollset_worker *next;
   struct grpc_pollset_worker *prev;
 } grpc_pollset_worker;
@@ -111,6 +112,16 @@ void grpc_kick_drain(grpc_pollset *p);
 int grpc_poll_deadline_to_millis_timeout(gpr_timespec deadline,
                                          gpr_timespec now);
 
+/* Allow kick to wakeup the currently polling worker */
+#define GRPC_POLLSET_CAN_KICK_SELF 1
+/* Force the wakee to repoll when awoken */
+#define GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP 2
+/* As per grpc_pollset_kick, with an extended set of flags (defined above)
+   -- mostly for fd_posix's use. */
+void grpc_pollset_kick_ext(grpc_pollset *p,
+                           grpc_pollset_worker *specific_worker,
+                           gpr_uint32 flags);
+
 /* turn a pollset into a multipoller: platform specific */
 typedef void (*grpc_platform_become_multipoller_type)(grpc_exec_ctx *exec_ctx,
                                                       grpc_pollset *pollset,