Merge github.com:grpc/grpc into delayed-write

34 files changed, 4634 insertions, 101 deletions

diff --git a/src/core/lib/iomgr/endpoint.h b/src/core/lib/iomgr/endpoint.h
index 3877ceb1e2..f9808bbda1 100644
--- a/src/core/lib/iomgr/endpoint.h
+++ b/src/core/lib/iomgr/endpoint.h
@@ -82,7 +82,7 @@ char *grpc_endpoint_get_peer(grpc_endpoint *ep);
 void grpc_endpoint_write(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep,
                          gpr_slice_buffer *slices, grpc_closure *cb);
 
-/* Causes any pending read/write callbacks to run immediately with
+/* Causes any pending and future read/write callbacks to run immediately with
    success==0 */
 void grpc_endpoint_shutdown(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep);
 void grpc_endpoint_destroy(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep);
diff --git a/src/core/lib/iomgr/error.c b/src/core/lib/iomgr/error.c
index 35154268f3..e20a0169c5 100644
--- a/src/core/lib/iomgr/error.c
+++ b/src/core/lib/iomgr/error.c
@@ -1,6 +1,6 @@
 /*
  *
- * Copyright 2015, Google Inc.
+ * Copyright 2016, Google Inc.
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
@@ -33,19 +33,23 @@
 
 #include "src/core/lib/iomgr/error.h"
 
+#include <inttypes.h>
 #include <stdbool.h>
 #include <string.h>
 
+#include <grpc/status.h>
 #include <grpc/support/alloc.h>
 #include <grpc/support/avl.h>
 #include <grpc/support/log.h>
 #include <grpc/support/string_util.h>
 #include <grpc/support/useful.h>
 
-#ifdef GPR_WIN32
-#include <grpc/support/log_win32.h>
+#ifdef GPR_WINDOWS
+#include <grpc/support/log_windows.h>
 #endif
 
+#include "src/core/lib/profiling/timers.h"
+
 static void destroy_integer(void *key) {}
 
 static void *copy_integer(void *key) { return key; }
@@ -88,22 +92,16 @@ static const gpr_avl_vtable avl_vtable_errs = {
 
 static const char *error_int_name(grpc_error_ints key) {
   switch (key) {
-    case GRPC_ERROR_INT_STATUS_CODE:
-      return "status_code";
     case GRPC_ERROR_INT_ERRNO:
       return "errno";
     case GRPC_ERROR_INT_FILE_LINE:
       return "file_line";
-    case GRPC_ERROR_INT_WARNING:
-      return "warning";
     case GRPC_ERROR_INT_STREAM_ID:
       return "stream_id";
     case GRPC_ERROR_INT_GRPC_STATUS:
       return "grpc_status";
     case GRPC_ERROR_INT_OFFSET:
       return "offset";
-    case GRPC_ERROR_INT_LIMIT:
-      return "limit";
     case GRPC_ERROR_INT_INDEX:
       return "index";
     case GRPC_ERROR_INT_SIZE:
@@ -118,6 +116,10 @@ static const char *error_int_name(grpc_error_ints key) {
       return "fd";
     case GRPC_ERROR_INT_WSA_ERROR:
       return "wsa_error";
+    case GRPC_ERROR_INT_HTTP_STATUS:
+      return "http_status";
+    case GRPC_ERROR_INT_LIMIT:
+      return "limit";
   }
   GPR_UNREACHABLE_CODE(return "unknown");
 }
@@ -216,6 +218,7 @@ void grpc_error_unref(grpc_error *err) {
 grpc_error *grpc_error_create(const char *file, int line, const char *desc,
                               grpc_error **referencing,
                               size_t num_referencing) {
+  GPR_TIMER_BEGIN("grpc_error_create", 0);
   grpc_error *err = gpr_malloc(sizeof(*err));
   if (err == NULL) {  // TODO(ctiller): make gpr_malloc return NULL
     return GRPC_ERROR_OOM;
@@ -241,57 +244,91 @@ grpc_error *grpc_error_create(const char *file, int line, const char *desc,
                            (void *)(uintptr_t)GRPC_ERROR_TIME_CREATED,
                            box_time(gpr_now(GPR_CLOCK_REALTIME)));
   gpr_ref_init(&err->refs, 1);
+  GPR_TIMER_END("grpc_error_create", 0);
   return err;
 }
 
 static grpc_error *copy_error_and_unref(grpc_error *in) {
+  GPR_TIMER_BEGIN("copy_error_and_unref", 0);
+  grpc_error *out;
   if (is_special(in)) {
-    if (in == GRPC_ERROR_NONE) return GRPC_ERROR_CREATE("no error");
-    if (in == GRPC_ERROR_OOM) return GRPC_ERROR_CREATE("oom");
-    if (in == GRPC_ERROR_CANCELLED) return GRPC_ERROR_CREATE("cancelled");
-    return GRPC_ERROR_CREATE("unknown");
-  }
-  grpc_error *out = gpr_malloc(sizeof(*out));
+    if (in == GRPC_ERROR_NONE)
+      out = GRPC_ERROR_CREATE("no error");
+    else if (in == GRPC_ERROR_OOM)
+      out = GRPC_ERROR_CREATE("oom");
+    else if (in == GRPC_ERROR_CANCELLED)
+      out =
+          grpc_error_set_int(GRPC_ERROR_CREATE("cancelled"),
+                             GRPC_ERROR_INT_GRPC_STATUS, GRPC_STATUS_CANCELLED);
+    else
+      out = GRPC_ERROR_CREATE("unknown");
+  } else {
+    out = gpr_malloc(sizeof(*out));
 #ifdef GRPC_ERROR_REFCOUNT_DEBUG
-  gpr_log(GPR_DEBUG, "%p create copying", out);
+    gpr_log(GPR_DEBUG, "%p create copying", out);
 #endif
-  out->ints = gpr_avl_ref(in->ints);
-  out->strs = gpr_avl_ref(in->strs);
-  out->errs = gpr_avl_ref(in->errs);
-  out->times = gpr_avl_ref(in->times);
-  out->next_err = in->next_err;
-  gpr_ref_init(&out->refs, 1);
-  GRPC_ERROR_UNREF(in);
+    out->ints = gpr_avl_ref(in->ints);
+    out->strs = gpr_avl_ref(in->strs);
+    out->errs = gpr_avl_ref(in->errs);
+    out->times = gpr_avl_ref(in->times);
+    out->next_err = in->next_err;
+    gpr_ref_init(&out->refs, 1);
+    GRPC_ERROR_UNREF(in);
+  }
+  GPR_TIMER_END("copy_error_and_unref", 0);
   return out;
 }
 
 grpc_error *grpc_error_set_int(grpc_error *src, grpc_error_ints which,
                                intptr_t value) {
+  GPR_TIMER_BEGIN("grpc_error_set_int", 0);
   grpc_error *new = copy_error_and_unref(src);
   new->ints = gpr_avl_add(new->ints, (void *)(uintptr_t)which, (void *)value);
+  GPR_TIMER_END("grpc_error_set_int", 0);
   return new;
 }
 
 bool grpc_error_get_int(grpc_error *err, grpc_error_ints which, intptr_t *p) {
+  GPR_TIMER_BEGIN("grpc_error_get_int", 0);
   void *pp;
+  if (is_special(err)) {
+    if (err == GRPC_ERROR_CANCELLED && which == GRPC_ERROR_INT_GRPC_STATUS) {
+      *p = GRPC_STATUS_CANCELLED;
+      GPR_TIMER_END("grpc_error_get_int", 0);
+      return true;
+    }
+    GPR_TIMER_END("grpc_error_get_int", 0);
+    return false;
+  }
   if (gpr_avl_maybe_get(err->ints, (void *)(uintptr_t)which, &pp)) {
     if (p != NULL) *p = (intptr_t)pp;
+    GPR_TIMER_END("grpc_error_get_int", 0);
     return true;
   }
+  GPR_TIMER_END("grpc_error_get_int", 0);
   return false;
 }
 
 grpc_error *grpc_error_set_str(grpc_error *src, grpc_error_strs which,
                                const char *value) {
+  GPR_TIMER_BEGIN("grpc_error_set_str", 0);
   grpc_error *new = copy_error_and_unref(src);
   new->strs =
       gpr_avl_add(new->strs, (void *)(uintptr_t)which, gpr_strdup(value));
+  GPR_TIMER_END("grpc_error_set_str", 0);
   return new;
 }
 
+const char *grpc_error_get_str(grpc_error *err, grpc_error_strs which) {
+  if (is_special(err)) return NULL;
+  return gpr_avl_get(err->strs, (void *)(uintptr_t)which);
+}
+
 grpc_error *grpc_error_add_child(grpc_error *src, grpc_error *child) {
+  GPR_TIMER_BEGIN("grpc_error_add_child", 0);
   grpc_error *new = copy_error_and_unref(src);
   new->errs = gpr_avl_add(new->errs, (void *)(new->next_err++), child);
+  GPR_TIMER_END("grpc_error_add_child", 0);
   return new;
 }
 
@@ -342,7 +379,7 @@ static char *key_time(void *p) {
 
 static char *fmt_int(void *p) {
   char *s;
-  gpr_asprintf(&s, "%lld", (intptr_t)p);
+  gpr_asprintf(&s, "%" PRIdPTR, (intptr_t)p);
   return s;
 }
 
@@ -424,7 +461,7 @@ static char *fmt_time(void *p) {
       pfx = "";
       break;
   }
-  gpr_asprintf(&out, "\"%s%d.%09d\"", pfx, tm.tv_sec, tm.tv_nsec);
+  gpr_asprintf(&out, "\"%s%" PRId64 ".%09d\"", pfx, tm.tv_sec, tm.tv_nsec);
   return out;
 }
 
@@ -483,6 +520,7 @@ void grpc_error_free_string(const char *str) {
 }
 
 const char *grpc_error_string(grpc_error *err) {
+  GPR_TIMER_BEGIN("grpc_error_string", 0);
   if (err == GRPC_ERROR_NONE) return no_error_string;
   if (err == GRPC_ERROR_OOM) return oom_error_string;
   if (err == GRPC_ERROR_CANCELLED) return cancelled_error_string;
@@ -499,7 +537,9 @@ const char *grpc_error_string(grpc_error *err) {
 
   qsort(kvs.kvs, kvs.num_kvs, sizeof(kv_pair), cmp_kvs);
 
-  return finish_kvs(&kvs);
+  const char *out = finish_kvs(&kvs);
+  GPR_TIMER_END("grpc_error_string", 0);
+  return out;
 }
 
 grpc_error *grpc_os_error(const char *file, int line, int err,
@@ -512,7 +552,7 @@ grpc_error *grpc_os_error(const char *file, int line, int err,
       GRPC_ERROR_STR_SYSCALL, call_name);
 }
 
-#ifdef GPR_WIN32
+#ifdef GPR_WINDOWS
 grpc_error *grpc_wsa_error(const char *file, int line, int err,
                            const char *call_name) {
   char *utf8_message = gpr_format_message(err);
diff --git a/src/core/lib/iomgr/error.h b/src/core/lib/iomgr/error.h
index ede159526e..13f898e31a 100644
--- a/src/core/lib/iomgr/error.h
+++ b/src/core/lib/iomgr/error.h
@@ -1,6 +1,6 @@
 /*
  *
- * Copyright 2015, Google Inc.
+ * Copyright 2016, Google Inc.
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
@@ -39,50 +39,86 @@
 
 #include <grpc/support/time.h>
 
-// Opaque representation of an error.
-// Errors are refcounted objects that represent the result of an operation.
-// Ownership laws:
-//  if a grpc_error is returned by a function, the caller owns a ref to that
-//    instance
-//  if a grpc_error is passed to a grpc_closure callback function (functions
-//    with the signature:
-//      void (*f)(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error))
-//    then those functions do not automatically own a ref to error
-//  if a grpc_error is passed to *ANY OTHER FUNCTION* then that function takes
-//    ownership of the error
+/// Opaque representation of an error.
+/// Errors are refcounted objects that represent the result of an operation.
+/// Ownership laws:
+///  if a grpc_error is returned by a function, the caller owns a ref to that
+///    instance
+///  if a grpc_error is passed to a grpc_closure callback function (functions
+///    with the signature:
+///      void (*f)(grpc_exec_ctx *exec_ctx, void *arg, grpc_error *error))
+///    then those functions do not automatically own a ref to error
+///  if a grpc_error is passed to *ANY OTHER FUNCTION* then that function takes
+///    ownership of the error
+/// Errors have:
+///  a set of ints, strings, and timestamps that describe the error
+///  always present are:
+///    GRPC_ERROR_STR_FILE, GRPC_ERROR_INT_FILE_LINE - source location the error
+///      was generated
+///    GRPC_ERROR_STR_DESCRIPTION - a human readable description of the error
+///    GRPC_ERROR_TIME_CREATED - a timestamp indicating when the error happened
+///  an error can also have children; these are other errors that are believed
+///    to have contributed to this one. By accumulating children, we can begin
+///    to root cause high level failures from low level failures, without having
+///    to derive execution paths from log lines
 typedef struct grpc_error grpc_error;
 
 typedef enum {
+  /// 'errno' from the operating system
   GRPC_ERROR_INT_ERRNO,
+  /// __LINE__ from the call site creating the error
   GRPC_ERROR_INT_FILE_LINE,
-  GRPC_ERROR_INT_STATUS_CODE,
-  GRPC_ERROR_INT_WARNING,
+  /// stream identifier: for errors that are associated with an individual
+  /// wire stream
   GRPC_ERROR_INT_STREAM_ID,
+  /// grpc status code representing this error
   GRPC_ERROR_INT_GRPC_STATUS,
+  /// offset into some binary blob (usually represented by
+  /// GRPC_ERROR_STR_RAW_BYTES) where the error occurred
   GRPC_ERROR_INT_OFFSET,
+  /// context sensitive index associated with the error
   GRPC_ERROR_INT_INDEX,
+  /// context sensitive size associated with the error
   GRPC_ERROR_INT_SIZE,
-  GRPC_ERROR_INT_LIMIT,
+  /// http2 error code associated with the error (see the HTTP2 RFC)
   GRPC_ERROR_INT_HTTP2_ERROR,
+  /// TSI status code associated with the error
   GRPC_ERROR_INT_TSI_CODE,
+  /// grpc_security_status associated with the error
   GRPC_ERROR_INT_SECURITY_STATUS,
+  /// WSAGetLastError() reported when this error occurred
   GRPC_ERROR_INT_WSA_ERROR,
+  /// File descriptor associated with this error
   GRPC_ERROR_INT_FD,
+  /// HTTP status (i.e. 404)
+  GRPC_ERROR_INT_HTTP_STATUS,
+  /// context sensitive limit associated with the error
+  GRPC_ERROR_INT_LIMIT,
 } grpc_error_ints;
 
 typedef enum {
+  /// top-level textual description of this error
   GRPC_ERROR_STR_DESCRIPTION,
+  /// source file in which this error occurred
   GRPC_ERROR_STR_FILE,
+  /// operating system description of this error
   GRPC_ERROR_STR_OS_ERROR,
+  /// syscall that generated this error
   GRPC_ERROR_STR_SYSCALL,
+  /// peer that we were trying to communicate when this error occurred
   GRPC_ERROR_STR_TARGET_ADDRESS,
+  /// grpc status message associated with this error
   GRPC_ERROR_STR_GRPC_MESSAGE,
+  /// hex dump (or similar) with the data that generated this error
   GRPC_ERROR_STR_RAW_BYTES,
+  /// tsi error string associated with this error
   GRPC_ERROR_STR_TSI_ERROR,
+  /// filename that we were trying to read/write when this error occurred
   GRPC_ERROR_STR_FILENAME,
 } grpc_error_strs;
 
 typedef enum {
+  /// timestamp of error creation
   GRPC_ERROR_TIME_CREATED,
 } grpc_error_times;
 
@@ -93,8 +129,17 @@ typedef enum {
 const char *grpc_error_string(grpc_error *error);
 void grpc_error_free_string(const char *str);
 
+/// Create an error - but use GRPC_ERROR_CREATE instead
 grpc_error *grpc_error_create(const char *file, int line, const char *desc,
                               grpc_error **referencing, size_t num_referencing);
+/// Create an error (this is the preferred way of generating an error that is
+///   not due to a system call - for system calls, use GRPC_OS_ERROR or
+///   GRPC_WSA_ERROR as appropriate)
+/// \a referencing is an array of num_referencing elements indicating one or
+/// more errors that are believed to have contributed to this one
+/// err = grpc_error_create(x, y, z, r, nr) is equivalent to:
+///   err = grpc_error_create(x, y, z, NULL, 0);
+///   for (i=0; i<nr; i++) err = grpc_error_add_child(err, r[i]);
 #define GRPC_ERROR_CREATE(desc) \
   grpc_error_create(__FILE__, __LINE__, desc, NULL, 0)
 
@@ -120,19 +165,26 @@ void grpc_error_unref(grpc_error *err);
 #endif
 
 grpc_error *grpc_error_set_int(grpc_error *src, grpc_error_ints which,
-                               intptr_t value);
+                               intptr_t value) GRPC_MUST_USE_RESULT;
 bool grpc_error_get_int(grpc_error *error, grpc_error_ints which, intptr_t *p);
 grpc_error *grpc_error_set_time(grpc_error *src, grpc_error_times which,
-                                gpr_timespec value);
+                                gpr_timespec value) GRPC_MUST_USE_RESULT;
 grpc_error *grpc_error_set_str(grpc_error *src, grpc_error_strs which,
-                               const char *value);
-grpc_error *grpc_error_add_child(grpc_error *src, grpc_error *child);
+                               const char *value) GRPC_MUST_USE_RESULT;
+const char *grpc_error_get_str(grpc_error *error, grpc_error_strs which);
+/// Add a child error: an error that is believed to have contributed to this
+/// error occurring. Allows root causing high level errors from lower level
+/// errors that contributed to them.
+grpc_error *grpc_error_add_child(grpc_error *src,
+                                 grpc_error *child) GRPC_MUST_USE_RESULT;
 grpc_error *grpc_os_error(const char *file, int line, int err,
-                          const char *call_name);
+                          const char *call_name) GRPC_MUST_USE_RESULT;
+/// create an error associated with errno!=0 (an 'operating system' error)
 #define GRPC_OS_ERROR(err, call_name) \
   grpc_os_error(__FILE__, __LINE__, err, call_name)
 grpc_error *grpc_wsa_error(const char *file, int line, int err,
-                           const char *call_name);
+                           const char *call_name) GRPC_MUST_USE_RESULT;
+/// windows only: create an error associated with WSAGetLastError()!=0
 #define GRPC_WSA_ERROR(err, call_name) \
   grpc_wsa_error(__FILE__, __LINE__, err, call_name)
 
diff --git a/src/core/lib/iomgr/ev_epoll_linux.c b/src/core/lib/iomgr/ev_epoll_linux.c
new file mode 100644
index 0000000000..cf0fe736a0
--- /dev/null
+++ b/src/core/lib/iomgr/ev_epoll_linux.c
@@ -0,0 +1,1872 @@
+/*
+ *
+ * Copyright 2016, Google Inc.
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above
+ * copyright notice, this list of conditions and the following disclaimer
+ * in the documentation and/or other materials provided with the
+ * distribution.
+ *     * Neither the name of Google Inc. nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ */
+
+#include <grpc/grpc_posix.h>
+#include <grpc/support/port_platform.h>
+
+/* This polling engine is only relevant on linux kernels supporting epoll() */
+#ifdef GPR_LINUX_EPOLL
+
+#include "src/core/lib/iomgr/ev_epoll_linux.h"
+
+#include <assert.h>
+#include <errno.h>
+#include <poll.h>
+#include <signal.h>
+#include <string.h>
+#include <sys/epoll.h>
+#include <sys/socket.h>
+#include <unistd.h>
+
+#include <grpc/support/alloc.h>
+#include <grpc/support/log.h>
+#include <grpc/support/string_util.h>
+#include <grpc/support/tls.h>
+#include <grpc/support/useful.h>
+
+#include "src/core/lib/iomgr/ev_posix.h"
+#include "src/core/lib/iomgr/iomgr_internal.h"
+#include "src/core/lib/iomgr/wakeup_fd_posix.h"
+#include "src/core/lib/profiling/timers.h"
+#include "src/core/lib/support/block_annotate.h"
+
+/* TODO: sreek - Move this to init.c and initialize this like other tracers. */
+static int grpc_polling_trace = 0; /* Disabled by default */
+#define GRPC_POLLING_TRACE(fmt, ...)       \
+  if (grpc_polling_trace) {                \
+    gpr_log(GPR_INFO, (fmt), __VA_ARGS__); \
+  }
+
+static int grpc_wakeup_signal = -1;
+static bool is_grpc_wakeup_signal_initialized = false;
+
+/* Implements the function defined in grpc_posix.h. This function might be
+ * called before even calling grpc_init() to set either a different signal to
+ * use. If signum == -1, then the use of signals is disabled */
+void grpc_use_signal(int signum) {
+  grpc_wakeup_signal = signum;
+  is_grpc_wakeup_signal_initialized = true;
+
+  if (grpc_wakeup_signal < 0) {
+    gpr_log(GPR_INFO,
+            "Use of signals is disabled. Epoll engine will not be used");
+  } else {
+    gpr_log(GPR_INFO, "epoll engine will be using signal: %d",
+            grpc_wakeup_signal);
+  }
+}
+
+struct polling_island;
+
+/*******************************************************************************
+ * Fd Declarations
+ */
+struct grpc_fd {
+  int fd;
+  /* refst format:
+       bit 0    : 1=Active / 0=Orphaned
+       bits 1-n : refcount
+     Ref/Unref by two to avoid altering the orphaned bit */
+  gpr_atm refst;
+
+  gpr_mu mu;
+
+  /* Indicates that the fd is shutdown and that any pending read/write closures
+     should fail */
+  bool shutdown;
+
+  /* The fd is either closed or we relinquished control of it. In either cases,
+     this indicates that the 'fd' on this structure is no longer valid */
+  bool orphaned;
+
+  /* TODO: sreek - Move this to a lockfree implementation */
+  grpc_closure *read_closure;
+  grpc_closure *write_closure;
+
+  /* The polling island to which this fd belongs to and the mutex protecting the
+     the field */
+  gpr_mu pi_mu;
+  struct polling_island *polling_island;
+
+  struct grpc_fd *freelist_next;
+  grpc_closure *on_done_closure;
+
+  /* The pollset that last noticed that the fd is readable */
+  grpc_pollset *read_notifier_pollset;
+
+  grpc_iomgr_object iomgr_object;
+};
+
+/* Reference counting for fds */
+// #define GRPC_FD_REF_COUNT_DEBUG
+#ifdef GRPC_FD_REF_COUNT_DEBUG
+static void fd_ref(grpc_fd *fd, const char *reason, const char *file, int line);
+static void fd_unref(grpc_fd *fd, const char *reason, const char *file,
+                     int line);
+#define GRPC_FD_REF(fd, reason) fd_ref(fd, reason, __FILE__, __LINE__)
+#define GRPC_FD_UNREF(fd, reason) fd_unref(fd, reason, __FILE__, __LINE__)
+#else
+static void fd_ref(grpc_fd *fd);
+static void fd_unref(grpc_fd *fd);
+#define GRPC_FD_REF(fd, reason) fd_ref(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)
+
+/*******************************************************************************
+ * Polling island Declarations
+ */
+
+// #define GRPC_PI_REF_COUNT_DEBUG
+#ifdef GRPC_PI_REF_COUNT_DEBUG
+
+#define PI_ADD_REF(p, r) pi_add_ref_dbg((p), (r), __FILE__, __LINE__)
+#define PI_UNREF(p, r) pi_unref_dbg((p), (r), __FILE__, __LINE__)
+
+#else /* defined(GRPC_PI_REF_COUNT_DEBUG) */
+
+#define PI_ADD_REF(p, r) pi_add_ref((p))
+#define PI_UNREF(p, r) pi_unref((p))
+
+#endif /* !defined(GPRC_PI_REF_COUNT_DEBUG) */
+
+typedef struct polling_island {
+  gpr_mu mu;
+  /* Ref count. Use PI_ADD_REF() and PI_UNREF() macros to increment/decrement
+     the refcount.
+     Once the ref count becomes zero, this structure is destroyed which means
+     we should ensure that there is never a scenario where a PI_ADD_REF() is
+     racing with a PI_UNREF() that just made the ref_count zero. */
+  gpr_refcount ref_count;
+
+  /* Pointer to the polling_island this merged into.
+   * merged_to value is only set once in polling_island's lifetime (and that too
+   * only if the island is merged with another island). Because of this, we can
+   * use gpr_atm type here so that we can do atomic access on this and reduce
+   * lock contention on 'mu' mutex.
+   *
+   * Note that if this field is not NULL (i.e not 0), all the remaining fields
+   * (except mu and ref_count) are invalid and must be ignored. */
+  gpr_atm merged_to;
+
+  /* The fd of the underlying epoll set */
+  int epoll_fd;
+
+  /* The file descriptors in the epoll set */
+  size_t fd_cnt;
+  size_t fd_capacity;
+  grpc_fd **fds;
+
+  /* Polling islands that are no longer needed are kept in a freelist so that
+     they can be reused. This field points to the next polling island in the
+     free list */
+  struct polling_island *next_free;
+} polling_island;
+
+/*******************************************************************************
+ * Pollset Declarations
+ */
+struct grpc_pollset_worker {
+  /* Thread id of this worker */
+  pthread_t pt_id;
+
+  /* Used to prevent a worker from getting kicked multiple times */
+  gpr_atm is_kicked;
+  struct grpc_pollset_worker *next;
+  struct grpc_pollset_worker *prev;
+};
+
+struct grpc_pollset {
+  gpr_mu mu;
+  grpc_pollset_worker root_worker;
+  bool kicked_without_pollers;
+
+  bool shutting_down;          /* Is the pollset shutting down ? */
+  bool finish_shutdown_called; /* Is the 'finish_shutdown_locked()' called ? */
+  grpc_closure *shutdown_done; /* Called after after shutdown is complete */
+
+  /* The polling island to which this pollset belongs to */
+  struct polling_island *polling_island;
+};
+
+/*******************************************************************************
+ * Pollset-set Declarations
+ */
+/* TODO: sreek - Change the pollset_set implementation such that a pollset_set
+ * directly points to a polling_island (and adding an fd/pollset/pollset_set to
+ * the current pollset_set would result in polling island merges. This would
+ * remove the need to maintain fd_count here. This will also significantly
+ * simplify the grpc_fd structure since we would no longer need to explicitly
+ * maintain the orphaned state */
+struct grpc_pollset_set {
+  gpr_mu mu;
+
+  size_t pollset_count;
+  size_t pollset_capacity;
+  grpc_pollset **pollsets;
+
+  size_t pollset_set_count;
+  size_t pollset_set_capacity;
+  struct grpc_pollset_set **pollset_sets;
+
+  size_t fd_count;
+  size_t fd_capacity;
+  grpc_fd **fds;
+};
+
+/*******************************************************************************
+ * Common helpers
+ */
+
+static void append_error(grpc_error **composite, grpc_error *error,
+                         const char *desc) {
+  if (error == GRPC_ERROR_NONE) return;
+  if (*composite == GRPC_ERROR_NONE) {
+    *composite = GRPC_ERROR_CREATE(desc);
+  }
+  *composite = grpc_error_add_child(*composite, error);
+}
+
+/*******************************************************************************
+ * Polling island Definitions
+ */
+
+/* The wakeup fd that is used to wake up all threads in a Polling island. This
+   is useful in the polling island merge operation where we need to wakeup all
+   the threads currently polling the smaller polling island (so that they can
+   start polling the new/merged polling island)
+
+   NOTE: This fd is initialized to be readable and MUST NOT be consumed i.e the
+   threads that woke up MUST NOT call grpc_wakeup_fd_consume_wakeup() */
+static grpc_wakeup_fd polling_island_wakeup_fd;
+
+/* Polling island freelist */
+static gpr_mu g_pi_freelist_mu;
+static polling_island *g_pi_freelist = NULL;
+
+static void polling_island_delete(); /* Forward declaration */
+
+#ifdef GRPC_TSAN
+/* Currently TSAN may incorrectly flag data races between epoll_ctl and
+   epoll_wait for any grpc_fd structs that are added to the epoll set via
+   epoll_ctl and are returned (within a very short window) via epoll_wait().
+
+   To work-around this race, we establish a happens-before relation between
+   the code just-before epoll_ctl() and the code after epoll_wait() by using
+   this atomic */
+gpr_atm g_epoll_sync;
+#endif /* defined(GRPC_TSAN) */
+
+#ifdef GRPC_PI_REF_COUNT_DEBUG
+void pi_add_ref(polling_island *pi);
+void pi_unref(polling_island *pi);
+
+void pi_add_ref_dbg(polling_island *pi, char *reason, char *file, int line) {
+  long old_cnt = gpr_atm_acq_load(&(pi->ref_count.count));
+  pi_add_ref(pi);
+  gpr_log(GPR_DEBUG, "Add ref pi: %p, old: %ld -> new:%ld (%s) - (%s, %d)",
+          (void *)pi, old_cnt, old_cnt + 1, reason, file, line);
+}
+
+void pi_unref_dbg(polling_island *pi, char *reason, char *file, int line) {
+  long old_cnt = gpr_atm_acq_load(&(pi->ref_count.count));
+  pi_unref(pi);
+  gpr_log(GPR_DEBUG, "Unref pi: %p, old:%ld -> new:%ld (%s) - (%s, %d)",
+          (void *)pi, old_cnt, (old_cnt - 1), reason, file, line);
+}
+#endif
+
+void pi_add_ref(polling_island *pi) { gpr_ref(&pi->ref_count); }
+
+void pi_unref(polling_island *pi) {
+  /* If ref count went to zero, delete the polling island.
+     Note that this deletion not be done under a lock. Once the ref count goes
+     to zero, we are guaranteed that no one else holds a reference to the
+     polling island (and that there is no racing pi_add_ref() call either).
+
+     Also, if we are deleting the polling island and the merged_to field is
+     non-empty, we should remove a ref to the merged_to polling island
+   */
+  if (gpr_unref(&pi->ref_count)) {
+    polling_island *next = (polling_island *)gpr_atm_acq_load(&pi->merged_to);
+    polling_island_delete(pi);
+    if (next != NULL) {
+      PI_UNREF(next, "pi_delete"); /* Recursive call */
+    }
+  }
+}
+
+/* The caller is expected to hold pi->mu lock before calling this function */
+static void polling_island_add_fds_locked(polling_island *pi, grpc_fd **fds,
+                                          size_t fd_count, bool add_fd_refs,
+                                          grpc_error **error) {
+  int err;
+  size_t i;
+  struct epoll_event ev;
+  char *err_msg;
+  const char *err_desc = "polling_island_add_fds";
+
+#ifdef GRPC_TSAN
+  /* See the definition of g_epoll_sync for more context */
+  gpr_atm_rel_store(&g_epoll_sync, (gpr_atm)0);
+#endif /* defined(GRPC_TSAN) */
+
+  for (i = 0; i < fd_count; i++) {
+    ev.events = (uint32_t)(EPOLLIN | EPOLLOUT | EPOLLET);
+    ev.data.ptr = fds[i];
+    err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_ADD, fds[i]->fd, &ev);
+
+    if (err < 0) {
+      if (errno != EEXIST) {
+        gpr_asprintf(
+            &err_msg,
+            "epoll_ctl (epoll_fd: %d) add fd: %d failed with error: %d (%s)",
+            pi->epoll_fd, fds[i]->fd, errno, strerror(errno));
+        append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc);
+        gpr_free(err_msg);
+      }
+
+      continue;
+    }
+
+    if (pi->fd_cnt == pi->fd_capacity) {
+      pi->fd_capacity = GPR_MAX(pi->fd_capacity + 8, pi->fd_cnt * 3 / 2);
+      pi->fds = gpr_realloc(pi->fds, sizeof(grpc_fd *) * pi->fd_capacity);
+    }
+
+    pi->fds[pi->fd_cnt++] = fds[i];
+    if (add_fd_refs) {
+      GRPC_FD_REF(fds[i], "polling_island");
+    }
+  }
+}
+
+/* The caller is expected to hold pi->mu before calling this */
+static void polling_island_add_wakeup_fd_locked(polling_island *pi,
+                                                grpc_wakeup_fd *wakeup_fd,
+                                                grpc_error **error) {
+  struct epoll_event ev;
+  int err;
+  char *err_msg;
+  const char *err_desc = "polling_island_add_wakeup_fd";
+
+  ev.events = (uint32_t)(EPOLLIN | EPOLLET);
+  ev.data.ptr = wakeup_fd;
+  err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_ADD,
+                  GRPC_WAKEUP_FD_GET_READ_FD(wakeup_fd), &ev);
+  if (err < 0 && errno != EEXIST) {
+    gpr_asprintf(&err_msg,
+                 "epoll_ctl (epoll_fd: %d) add wakeup fd: %d failed with "
+                 "error: %d (%s)",
+                 pi->epoll_fd,
+                 GRPC_WAKEUP_FD_GET_READ_FD(&grpc_global_wakeup_fd), errno,
+                 strerror(errno));
+    append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc);
+    gpr_free(err_msg);
+  }
+}
+
+/* The caller is expected to hold pi->mu lock before calling this function */
+static void polling_island_remove_all_fds_locked(polling_island *pi,
+                                                 bool remove_fd_refs,
+                                                 grpc_error **error) {
+  int err;
+  size_t i;
+  char *err_msg;
+  const char *err_desc = "polling_island_remove_fds";
+
+  for (i = 0; i < pi->fd_cnt; i++) {
+    err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_DEL, pi->fds[i]->fd, NULL);
+    if (err < 0 && errno != ENOENT) {
+      gpr_asprintf(&err_msg,
+                   "epoll_ctl (epoll_fd: %d) delete fds[%zu]: %d failed with "
+                   "error: %d (%s)",
+                   pi->epoll_fd, i, pi->fds[i]->fd, errno, strerror(errno));
+      append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc);
+      gpr_free(err_msg);
+    }
+
+    if (remove_fd_refs) {
+      GRPC_FD_UNREF(pi->fds[i], "polling_island");
+    }
+  }
+
+  pi->fd_cnt = 0;
+}
+
+/* The caller is expected to hold pi->mu lock before calling this function */
+static void polling_island_remove_fd_locked(polling_island *pi, grpc_fd *fd,
+                                            bool is_fd_closed,
+                                            grpc_error **error) {
+  int err;
+  size_t i;
+  char *err_msg;
+  const char *err_desc = "polling_island_remove_fd";
+
+  /* If fd is already closed, then it would have been automatically been removed
+     from the epoll set */
+  if (!is_fd_closed) {
+    err = epoll_ctl(pi->epoll_fd, EPOLL_CTL_DEL, fd->fd, NULL);
+    if (err < 0 && errno != ENOENT) {
+      gpr_asprintf(
+          &err_msg,
+          "epoll_ctl (epoll_fd: %d) del fd: %d failed with error: %d (%s)",
+          pi->epoll_fd, fd->fd, errno, strerror(errno));
+      append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc);
+      gpr_free(err_msg);
+    }
+  }
+
+  for (i = 0; i < pi->fd_cnt; i++) {
+    if (pi->fds[i] == fd) {
+      pi->fds[i] = pi->fds[--pi->fd_cnt];
+      GRPC_FD_UNREF(fd, "polling_island");
+      break;
+    }
+  }
+}
+
+/* Might return NULL in case of an error */
+static polling_island *polling_island_create(grpc_fd *initial_fd,
+                                             grpc_error **error) {
+  polling_island *pi = NULL;
+  char *err_msg;
+  const char *err_desc = "polling_island_create";
+
+  /* Try to get one from the polling island freelist */
+  gpr_mu_lock(&g_pi_freelist_mu);
+  if (g_pi_freelist != NULL) {
+    pi = g_pi_freelist;
+    g_pi_freelist = g_pi_freelist->next_free;
+    pi->next_free = NULL;
+  }
+  gpr_mu_unlock(&g_pi_freelist_mu);
+
+  /* Create new polling island if we could not get one from the free list */
+  if (pi == NULL) {
+    pi = gpr_malloc(sizeof(*pi));
+    gpr_mu_init(&pi->mu);
+    pi->fd_cnt = 0;
+    pi->fd_capacity = 0;
+    pi->fds = NULL;
+  }
+
+  gpr_ref_init(&pi->ref_count, 0);
+  gpr_atm_rel_store(&pi->merged_to, (gpr_atm)NULL);
+
+  pi->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
+
+  if (pi->epoll_fd < 0) {
+    gpr_asprintf(&err_msg, "epoll_create1 failed with error %d (%s)", errno,
+                 strerror(errno));
+    append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc);
+    gpr_free(err_msg);
+  } else {
+    polling_island_add_wakeup_fd_locked(pi, &grpc_global_wakeup_fd, error);
+    pi->next_free = NULL;
+
+    if (initial_fd != NULL) {
+      /* Lock the polling island here just in case we got this structure from
+         the freelist and the polling island lock was not released yet (by the
+         code that adds the polling island to the freelist) */
+      gpr_mu_lock(&pi->mu);
+      polling_island_add_fds_locked(pi, &initial_fd, 1, true, error);
+      gpr_mu_unlock(&pi->mu);
+    }
+  }
+
+  return pi;
+}
+
+static void polling_island_delete(polling_island *pi) {
+  GPR_ASSERT(pi->fd_cnt == 0);
+
+  gpr_atm_rel_store(&pi->merged_to, (gpr_atm)NULL);
+
+  close(pi->epoll_fd);
+  pi->epoll_fd = -1;
+
+  gpr_mu_lock(&g_pi_freelist_mu);
+  pi->next_free = g_pi_freelist;
+  g_pi_freelist = pi;
+  gpr_mu_unlock(&g_pi_freelist_mu);
+}
+
+/* Attempts to gets the last polling island in the linked list (liked by the
+ * 'merged_to' field). Since this does not lock the polling island, there are no
+ * guarantees that the island returned is the last island */
+static polling_island *polling_island_maybe_get_latest(polling_island *pi) {
+  polling_island *next = (polling_island *)gpr_atm_acq_load(&pi->merged_to);
+  while (next != NULL) {
+    pi = next;
+    next = (polling_island *)gpr_atm_acq_load(&pi->merged_to);
+  }
+
+  return pi;
+}
+
+/* Gets the lock on the *latest* polling island i.e the last polling island in
+   the linked list (linked by the 'merged_to' field). Call gpr_mu_unlock on the
+   returned polling island's mu.
+   Usage: To lock/unlock polling island "pi", do the following:
+      polling_island *pi_latest = polling_island_lock(pi);
+      ...
+      ... critical section ..
+      ...
+      gpr_mu_unlock(&pi_latest->mu); // NOTE: use pi_latest->mu. NOT pi->mu */
+static polling_island *polling_island_lock(polling_island *pi) {
+  polling_island *next = NULL;
+
+  while (true) {
+    next = (polling_island *)gpr_atm_acq_load(&pi->merged_to);
+    if (next == NULL) {
+      /* Looks like 'pi' is the last node in the linked list but unless we check
+         this by holding the pi->mu lock, we cannot be sure (i.e without the
+         pi->mu lock, we don't prevent island merges).
+         To be absolutely sure, check once more by holding the pi->mu lock */
+      gpr_mu_lock(&pi->mu);
+      next = (polling_island *)gpr_atm_acq_load(&pi->merged_to);
+      if (next == NULL) {
+        /* pi is infact the last node and we have the pi->mu lock. we're done */
+        break;
+      }
+
+      /* pi->merged_to is not NULL i.e pi isn't the last node anymore. pi->mu
+       * isn't the lock we are interested in. Continue traversing the list */
+      gpr_mu_unlock(&pi->mu);
+    }
+
+    pi = next;
+  }
+
+  return pi;
+}
+
+/* Gets the lock on the *latest* polling islands in the linked lists pointed by
+   *p and *q (and also updates *p and *q to point to the latest polling islands)
+
+   This function is needed because calling the following block of code to obtain
+   locks on polling islands (*p and *q) is prone to deadlocks.
+     {
+       polling_island_lock(*p, true);
+       polling_island_lock(*q, true);
+     }
+
+   Usage/example:
+     polling_island *p1;
+     polling_island *p2;
+     ..
+     polling_island_lock_pair(&p1, &p2);
+     ..
+     .. Critical section with both p1 and p2 locked
+     ..
+     // Release locks: Always call polling_island_unlock_pair() to release locks
+     polling_island_unlock_pair(p1, p2);
+*/
+static void polling_island_lock_pair(polling_island **p, polling_island **q) {
+  polling_island *pi_1 = *p;
+  polling_island *pi_2 = *q;
+  polling_island *next_1 = NULL;
+  polling_island *next_2 = NULL;
+
+  /* The algorithm is simple:
+      - Go to the last polling islands in the linked lists *pi_1 and *pi_2 (and
+        keep updating pi_1 and pi_2)
+      - Then obtain locks on the islands by following a lock order rule of
+        locking polling_island with lower address first
+           Special case: Before obtaining the locks, check if pi_1 and pi_2 are
+           pointing to the same island. If that is the case, we can just call
+           polling_island_lock()
+      - After obtaining both the locks, double check that the polling islands
+        are still the last polling islands in their respective linked lists
+        (this is because there might have been polling island merges before
+        we got the lock)
+      - If the polling islands are the last islands, we are done. If not,
+        release the locks and continue the process from the first step */
+  while (true) {
+    next_1 = (polling_island *)gpr_atm_acq_load(&pi_1->merged_to);
+    while (next_1 != NULL) {
+      pi_1 = next_1;
+      next_1 = (polling_island *)gpr_atm_acq_load(&pi_1->merged_to);
+    }
+
+    next_2 = (polling_island *)gpr_atm_acq_load(&pi_2->merged_to);
+    while (next_2 != NULL) {
+      pi_2 = next_2;
+      next_2 = (polling_island *)gpr_atm_acq_load(&pi_2->merged_to);
+    }
+
+    if (pi_1 == pi_2) {
+      pi_1 = pi_2 = polling_island_lock(pi_1);
+      break;
+    }
+
+    if (pi_1 < pi_2) {
+      gpr_mu_lock(&pi_1->mu);
+      gpr_mu_lock(&pi_2->mu);
+    } else {
+      gpr_mu_lock(&pi_2->mu);
+      gpr_mu_lock(&pi_1->mu);
+    }
+
+    next_1 = (polling_island *)gpr_atm_acq_load(&pi_1->merged_to);
+    next_2 = (polling_island *)gpr_atm_acq_load(&pi_2->merged_to);
+    if (next_1 == NULL && next_2 == NULL) {
+      break;
+    }
+
+    gpr_mu_unlock(&pi_1->mu);
+    gpr_mu_unlock(&pi_2->mu);
+  }
+
+  *p = pi_1;
+  *q = pi_2;
+}
+
+static void polling_island_unlock_pair(polling_island *p, polling_island *q) {
+  if (p == q) {
+    gpr_mu_unlock(&p->mu);
+  } else {
+    gpr_mu_unlock(&p->mu);
+    gpr_mu_unlock(&q->mu);
+  }
+}
+
+static polling_island *polling_island_merge(polling_island *p,
+                                            polling_island *q,
+                                            grpc_error **error) {
+  /* Get locks on both the polling islands */
+  polling_island_lock_pair(&p, &q);
+
+  if (p != q) {
+    /* Make sure that p points to the polling island with fewer fds than q */
+    if (p->fd_cnt > q->fd_cnt) {
+      GPR_SWAP(polling_island *, p, q);
+    }
+
+    /* Merge p with q i.e move all the fds from p (The one with fewer fds) to q
+       Note that the refcounts on the fds being moved will not change here.
+       This is why the last param in the following two functions is 'false') */
+    polling_island_add_fds_locked(q, p->fds, p->fd_cnt, false, error);
+    polling_island_remove_all_fds_locked(p, false, error);
+
+    /* Wakeup all the pollers (if any) on p so that they pickup this change */
+    polling_island_add_wakeup_fd_locked(p, &polling_island_wakeup_fd, error);
+
+    /* Add the 'merged_to' link from p --> q */
+    gpr_atm_rel_store(&p->merged_to, (gpr_atm)q);
+    PI_ADD_REF(q, "pi_merge"); /* To account for the new incoming ref from p */
+  }
+  /* else if p == q, nothing needs to be done */
+
+  polling_island_unlock_pair(p, q);
+
+  /* Return the merged polling island (Note that no merge would have happened
+     if p == q which is ok) */
+  return q;
+}
+
+static grpc_error *polling_island_global_init() {
+  grpc_error *error = GRPC_ERROR_NONE;
+
+  gpr_mu_init(&g_pi_freelist_mu);
+  g_pi_freelist = NULL;
+
+  error = grpc_wakeup_fd_init(&polling_island_wakeup_fd);
+  if (error == GRPC_ERROR_NONE) {
+    error = grpc_wakeup_fd_wakeup(&polling_island_wakeup_fd);
+  }
+
+  return error;
+}
+
+static void polling_island_global_shutdown() {
+  polling_island *next;
+  gpr_mu_lock(&g_pi_freelist_mu);
+  gpr_mu_unlock(&g_pi_freelist_mu);
+  while (g_pi_freelist != NULL) {
+    next = g_pi_freelist->next_free;
+    gpr_mu_destroy(&g_pi_freelist->mu);
+    gpr_free(g_pi_freelist->fds);
+    gpr_free(g_pi_freelist);
+    g_pi_freelist = next;
+  }
+  gpr_mu_destroy(&g_pi_freelist_mu);
+
+  grpc_wakeup_fd_destroy(&polling_island_wakeup_fd);
+}
+
+/*******************************************************************************
+ * Fd Definitions
+ */
+
+/* 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.
+ */
+
+/* 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. */
+
+/* TODO: sreek: Right now, this wakes up all pollers. In future we should make
+ * sure to wake up one polling thread (which can wake up other threads if
+ * needed) */
+grpc_wakeup_fd grpc_global_wakeup_fd;
+
+static grpc_fd *fd_freelist = NULL;
+static gpr_mu fd_freelist_mu;
+
+#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__)
+static void ref_by(grpc_fd *fd, int n, const char *reason, const char *file,
+                   int line) {
+  gpr_log(GPR_DEBUG, "FD %d %p   ref %d %ld -> %ld [%s; %s:%d]", fd->fd,
+          (void *)fd, n, gpr_atm_no_barrier_load(&fd->refst),
+          gpr_atm_no_barrier_load(&fd->refst) + n, reason, file, line);
+#else
+#define REF_BY(fd, n, reason) ref_by(fd, n)
+#define UNREF_BY(fd, n, reason) unref_by(fd, n)
+static void ref_by(grpc_fd *fd, int n) {
+#endif
+  GPR_ASSERT(gpr_atm_no_barrier_fetch_add(&fd->refst, n) > 0);
+}
+
+#ifdef GRPC_FD_REF_COUNT_DEBUG
+static void unref_by(grpc_fd *fd, int n, const char *reason, const char *file,
+                     int line) {
+  gpr_atm old;
+  gpr_log(GPR_DEBUG, "FD %d %p unref %d %ld -> %ld [%s; %s:%d]", fd->fd,
+          (void *)fd, n, gpr_atm_no_barrier_load(&fd->refst),
+          gpr_atm_no_barrier_load(&fd->refst) - n, reason, file, line);
+#else
+static void unref_by(grpc_fd *fd, int n) {
+  gpr_atm old;
+#endif
+  old = gpr_atm_full_fetch_add(&fd->refst, -n);
+  if (old == n) {
+    /* Add the fd to the freelist */
+    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);
+  } else {
+    GPR_ASSERT(old > n);
+  }
+}
+
+/* Increment refcount by two to avoid changing the orphan bit */
+#ifdef GRPC_FD_REF_COUNT_DEBUG
+static void fd_ref(grpc_fd *fd, const char *reason, const char *file,
+                   int line) {
+  ref_by(fd, 2, reason, file, line);
+}
+
+static void fd_unref(grpc_fd *fd, const char *reason, const char *file,
+                     int line) {
+  unref_by(fd, 2, reason, file, line);
+}
+#else
+static void fd_ref(grpc_fd *fd) { ref_by(fd, 2); }
+static void fd_unref(grpc_fd *fd) { unref_by(fd, 2); }
+#endif
+
+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;
+    gpr_mu_destroy(&fd->mu);
+    gpr_free(fd);
+  }
+  gpr_mu_destroy(&fd_freelist_mu);
+}
+
+static grpc_fd *fd_create(int fd, const char *name) {
+  grpc_fd *new_fd = NULL;
+
+  gpr_mu_lock(&fd_freelist_mu);
+  if (fd_freelist != NULL) {
+    new_fd = fd_freelist;
+    fd_freelist = fd_freelist->freelist_next;
+  }
+  gpr_mu_unlock(&fd_freelist_mu);
+
+  if (new_fd == NULL) {
+    new_fd = gpr_malloc(sizeof(grpc_fd));
+    gpr_mu_init(&new_fd->mu);
+    gpr_mu_init(&new_fd->pi_mu);
+  }
+
+  /* Note: It is not really needed to get the new_fd->mu lock here. If this is a
+     newly created fd (or an fd we got from the freelist), no one else would be
+     holding a lock to it anyway. */
+  gpr_mu_lock(&new_fd->mu);
+
+  gpr_atm_rel_store(&new_fd->refst, (gpr_atm)1);
+  new_fd->fd = fd;
+  new_fd->shutdown = false;
+  new_fd->orphaned = false;
+  new_fd->read_closure = CLOSURE_NOT_READY;
+  new_fd->write_closure = CLOSURE_NOT_READY;
+  new_fd->polling_island = NULL;
+  new_fd->freelist_next = NULL;
+  new_fd->on_done_closure = NULL;
+  new_fd->read_notifier_pollset = NULL;
+
+  gpr_mu_unlock(&new_fd->mu);
+
+  char *fd_name;
+  gpr_asprintf(&fd_name, "%s fd=%d", name, fd);
+  grpc_iomgr_register_object(&new_fd->iomgr_object, fd_name);
+#ifdef GRPC_FD_REF_COUNT_DEBUG
+  gpr_log(GPR_DEBUG, "FD %d %p create %s", fd, (void *)new_fd, fd_name);
+#endif
+  gpr_free(fd_name);
+  return new_fd;
+}
+
+static bool fd_is_orphaned(grpc_fd *fd) {
+  return (gpr_atm_acq_load(&fd->refst) & 1) == 0;
+}
+
+static int fd_wrapped_fd(grpc_fd *fd) {
+  int ret_fd = -1;
+  gpr_mu_lock(&fd->mu);
+  if (!fd->orphaned) {
+    ret_fd = fd->fd;
+  }
+  gpr_mu_unlock(&fd->mu);
+
+  return ret_fd;
+}
+
+static void fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
+                      grpc_closure *on_done, int *release_fd,
+                      const char *reason) {
+  bool is_fd_closed = false;
+  grpc_error *error = GRPC_ERROR_NONE;
+
+  gpr_mu_lock(&fd->mu);
+  fd->on_done_closure = on_done;
+
+  /* If release_fd is not NULL, we should be relinquishing control of the file
+     descriptor fd->fd (but we still own the grpc_fd structure). */
+  if (release_fd != NULL) {
+    *release_fd = fd->fd;
+  } else {
+    close(fd->fd);
+    is_fd_closed = true;
+  }
+
+  fd->orphaned = true;
+
+  /* Remove the active status but keep referenced. We want this grpc_fd struct
+     to be alive (and not added to freelist) until the end of this function */
+  REF_BY(fd, 1, reason);
+
+  /* Remove the fd from the polling island:
+     - Get a lock on the latest polling island (i.e the last island in the
+       linked list pointed by fd->polling_island). This is the island that
+       would actually contain the fd
+     - Remove the fd from the latest polling island
+     - Unlock the latest polling island
+     - Set fd->polling_island to NULL (but remove the ref on the polling island
+       before doing this.) */
+  gpr_mu_lock(&fd->pi_mu);
+  if (fd->polling_island != NULL) {
+    polling_island *pi_latest = polling_island_lock(fd->polling_island);
+    polling_island_remove_fd_locked(pi_latest, fd, is_fd_closed, &error);
+    gpr_mu_unlock(&pi_latest->mu);
+
+    PI_UNREF(fd->polling_island, "fd_orphan");
+    fd->polling_island = NULL;
+  }
+  gpr_mu_unlock(&fd->pi_mu);
+
+  grpc_exec_ctx_sched(exec_ctx, fd->on_done_closure, error, NULL);
+
+  gpr_mu_unlock(&fd->mu);
+  UNREF_BY(fd, 2, reason); /* Drop the reference */
+  GRPC_LOG_IF_ERROR("fd_orphan", GRPC_ERROR_REF(error));
+}
+
+static grpc_error *fd_shutdown_error(bool shutdown) {
+  if (!shutdown) {
+    return GRPC_ERROR_NONE;
+  } else {
+    return GRPC_ERROR_CREATE("FD shutdown");
+  }
+}
+
+static void notify_on_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
+                             grpc_closure **st, grpc_closure *closure) {
+  if (fd->shutdown) {
+    grpc_exec_ctx_sched(exec_ctx, closure, GRPC_ERROR_CREATE("FD shutdown"),
+                        NULL);
+  } else 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_sched(exec_ctx, closure, fd_shutdown_error(fd->shutdown),
+                        NULL);
+  } 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();
+  }
+}
+
+/* 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_sched(exec_ctx, *st, fd_shutdown_error(fd->shutdown), NULL);
+    *st = CLOSURE_NOT_READY;
+    return 1;
+  }
+}
+
+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 bool fd_is_shutdown(grpc_fd *fd) {
+  gpr_mu_lock(&fd->mu);
+  const bool r = fd->shutdown;
+  gpr_mu_unlock(&fd->mu);
+  return r;
+}
+
+/* Might be called multiple times */
+static void fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd) {
+  gpr_mu_lock(&fd->mu);
+  /* Do the actual shutdown only once */
+  if (!fd->shutdown) {
+    fd->shutdown = true;
+
+    shutdown(fd->fd, SHUT_RDWR);
+    /* Flush any pending read and write closures. Since fd->shutdown is 'true'
+       at this point, the closures would be called with 'success = false' */
+    set_ready_locked(exec_ctx, fd, &fd->read_closure);
+    set_ready_locked(exec_ctx, fd, &fd->write_closure);
+  }
+  gpr_mu_unlock(&fd->mu);
+}
+
+static void fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
+                              grpc_closure *closure) {
+  gpr_mu_lock(&fd->mu);
+  notify_on_locked(exec_ctx, fd, &fd->read_closure, closure);
+  gpr_mu_unlock(&fd->mu);
+}
+
+static void fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
+                               grpc_closure *closure) {
+  gpr_mu_lock(&fd->mu);
+  notify_on_locked(exec_ctx, fd, &fd->write_closure, closure);
+  gpr_mu_unlock(&fd->mu);
+}
+
+/*******************************************************************************
+ * Pollset Definitions
+ */
+GPR_TLS_DECL(g_current_thread_pollset);
+GPR_TLS_DECL(g_current_thread_worker);
+static __thread bool g_initialized_sigmask;
+static __thread sigset_t g_orig_sigmask;
+
+static void sig_handler(int sig_num) {
+#ifdef GRPC_EPOLL_DEBUG
+  gpr_log(GPR_INFO, "Received signal %d", sig_num);
+#endif
+}
+
+static void poller_kick_init() { signal(grpc_wakeup_signal, sig_handler); }
+
+/* Global state management */
+static grpc_error *pollset_global_init(void) {
+  gpr_tls_init(&g_current_thread_pollset);
+  gpr_tls_init(&g_current_thread_worker);
+  poller_kick_init();
+  return grpc_wakeup_fd_init(&grpc_global_wakeup_fd);
+}
+
+static void pollset_global_shutdown(void) {
+  grpc_wakeup_fd_destroy(&grpc_global_wakeup_fd);
+  gpr_tls_destroy(&g_current_thread_pollset);
+  gpr_tls_destroy(&g_current_thread_worker);
+}
+
+static grpc_error *pollset_worker_kick(grpc_pollset_worker *worker) {
+  grpc_error *err = GRPC_ERROR_NONE;
+
+  /* Kick the worker only if it was not already kicked */
+  if (gpr_atm_no_barrier_cas(&worker->is_kicked, (gpr_atm)0, (gpr_atm)1)) {
+    GRPC_POLLING_TRACE(
+        "pollset_worker_kick: Kicking worker: %p (thread id: %ld)",
+        (void *)worker, worker->pt_id);
+    int err_num = pthread_kill(worker->pt_id, grpc_wakeup_signal);
+    if (err_num != 0) {
+      err = GRPC_OS_ERROR(err_num, "pthread_kill");
+    }
+  }
+  return err;
+}
+
+/* Return 1 if the pollset has active threads in pollset_work (pollset must
+ * be locked) */
+static int pollset_has_workers(grpc_pollset *p) {
+  return p->root_worker.next != &p->root_worker;
+}
+
+static void remove_worker(grpc_pollset *p, grpc_pollset_worker *worker) {
+  worker->prev->next = worker->next;
+  worker->next->prev = worker->prev;
+}
+
+static grpc_pollset_worker *pop_front_worker(grpc_pollset *p) {
+  if (pollset_has_workers(p)) {
+    grpc_pollset_worker *w = p->root_worker.next;
+    remove_worker(p, w);
+    return w;
+  } else {
+    return NULL;
+  }
+}
+
+static void push_back_worker(grpc_pollset *p, grpc_pollset_worker *worker) {
+  worker->next = &p->root_worker;
+  worker->prev = worker->next->prev;
+  worker->prev->next = worker->next->prev = worker;
+}
+
+static void push_front_worker(grpc_pollset *p, grpc_pollset_worker *worker) {
+  worker->prev = &p->root_worker;
+  worker->next = worker->prev->next;
+  worker->prev->next = worker->next->prev = worker;
+}
+
+/* p->mu must be held before calling this function */
+static grpc_error *pollset_kick(grpc_pollset *p,
+                                grpc_pollset_worker *specific_worker) {
+  GPR_TIMER_BEGIN("pollset_kick", 0);
+  grpc_error *error = GRPC_ERROR_NONE;
+  const char *err_desc = "Kick Failure";
+  grpc_pollset_worker *worker = specific_worker;
+  if (worker != NULL) {
+    if (worker == GRPC_POLLSET_KICK_BROADCAST) {
+      if (pollset_has_workers(p)) {
+        GPR_TIMER_BEGIN("pollset_kick.broadcast", 0);
+        for (worker = p->root_worker.next; worker != &p->root_worker;
+             worker = worker->next) {
+          if (gpr_tls_get(&g_current_thread_worker) != (intptr_t)worker) {
+            append_error(&error, pollset_worker_kick(worker), err_desc);
+          }
+        }
+        GPR_TIMER_END("pollset_kick.broadcast", 0);
+      } else {
+        p->kicked_without_pollers = true;
+      }
+    } else {
+      GPR_TIMER_MARK("kicked_specifically", 0);
+      if (gpr_tls_get(&g_current_thread_worker) != (intptr_t)worker) {
+        append_error(&error, pollset_worker_kick(worker), err_desc);
+      }
+    }
+  } else if (gpr_tls_get(&g_current_thread_pollset) != (intptr_t)p) {
+    /* Since worker == NULL, it means that we can kick "any" worker on this
+       pollset 'p'. If 'p' happens to be the same pollset this thread is
+       currently polling (i.e in pollset_work() function), then there is no need
+       to kick any other worker since the current thread can just absorb the
+       kick. This is the reason why we enter this case only when
+       g_current_thread_pollset is != p */
+
+    GPR_TIMER_MARK("kick_anonymous", 0);
+    worker = pop_front_worker(p);
+    if (worker != NULL) {
+      GPR_TIMER_MARK("finally_kick", 0);
+      push_back_worker(p, worker);
+      append_error(&error, pollset_worker_kick(worker), err_desc);
+    } else {
+      GPR_TIMER_MARK("kicked_no_pollers", 0);
+      p->kicked_without_pollers = true;
+    }
+  }
+
+  GPR_TIMER_END("pollset_kick", 0);
+  GRPC_LOG_IF_ERROR("pollset_kick", GRPC_ERROR_REF(error));
+  return error;
+}
+
+static grpc_error *kick_poller(void) {
+  return grpc_wakeup_fd_wakeup(&grpc_global_wakeup_fd);
+}
+
+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->kicked_without_pollers = false;
+
+  pollset->shutting_down = false;
+  pollset->finish_shutdown_called = false;
+  pollset->shutdown_done = NULL;
+
+  pollset->polling_island = NULL;
+}
+
+/* Convert a timespec to milliseconds:
+   - Very small or negative poll times are clamped to zero to do a non-blocking
+     poll (which becomes spin polling)
+   - Other small values are rounded up to one millisecond
+   - Longer than a millisecond polls are rounded up to the next nearest
+     millisecond to avoid spinning
+   - Infinite timeouts are converted to -1 */
+static int poll_deadline_to_millis_timeout(gpr_timespec deadline,
+                                           gpr_timespec now) {
+  gpr_timespec timeout;
+  static const int64_t max_spin_polling_us = 10;
+  if (gpr_time_cmp(deadline, gpr_inf_future(deadline.clock_type)) == 0) {
+    return -1;
+  }
+
+  if (gpr_time_cmp(deadline, gpr_time_add(now, gpr_time_from_micros(
+                                                   max_spin_polling_us,
+                                                   GPR_TIMESPAN))) <= 0) {
+    return 0;
+  }
+  timeout = gpr_time_sub(deadline, now);
+  return gpr_time_to_millis(gpr_time_add(
+      timeout, gpr_time_from_nanos(GPR_NS_PER_MS - 1, GPR_TIMESPAN)));
+}
+
+static void fd_become_readable(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
+                               grpc_pollset *notifier) {
+  /* Need the fd->mu since we might be racing with fd_notify_on_read */
+  gpr_mu_lock(&fd->mu);
+  set_ready_locked(exec_ctx, fd, &fd->read_closure);
+  fd->read_notifier_pollset = notifier;
+  gpr_mu_unlock(&fd->mu);
+}
+
+static void fd_become_writable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) {
+  /* Need the fd->mu since we might be racing with fd_notify_on_write */
+  gpr_mu_lock(&fd->mu);
+  set_ready_locked(exec_ctx, fd, &fd->write_closure);
+  gpr_mu_unlock(&fd->mu);
+}
+
+static void pollset_release_polling_island(grpc_pollset *ps, char *reason) {
+  if (ps->polling_island != NULL) {
+    PI_UNREF(ps->polling_island, reason);
+  }
+  ps->polling_island = NULL;
+}
+
+static void finish_shutdown_locked(grpc_exec_ctx *exec_ctx,
+                                   grpc_pollset *pollset) {
+  /* The pollset cannot have any workers if we are at this stage */
+  GPR_ASSERT(!pollset_has_workers(pollset));
+
+  pollset->finish_shutdown_called = true;
+
+  /* Release the ref and set pollset->polling_island to NULL */
+  pollset_release_polling_island(pollset, "ps_shutdown");
+  grpc_exec_ctx_sched(exec_ctx, pollset->shutdown_done, GRPC_ERROR_NONE, NULL);
+}
+
+/* pollset->mu lock must be held by the caller before calling this */
+static void pollset_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
+                             grpc_closure *closure) {
+  GPR_TIMER_BEGIN("pollset_shutdown", 0);
+  GPR_ASSERT(!pollset->shutting_down);
+  pollset->shutting_down = true;
+  pollset->shutdown_done = closure;
+  pollset_kick(pollset, GRPC_POLLSET_KICK_BROADCAST);
+
+  /* If the pollset has any workers, we cannot call finish_shutdown_locked()
+     because it would release the underlying polling island. In such a case, we
+     let the last worker call finish_shutdown_locked() from pollset_work() */
+  if (!pollset_has_workers(pollset)) {
+    GPR_ASSERT(!pollset->finish_shutdown_called);
+    GPR_TIMER_MARK("pollset_shutdown.finish_shutdown_locked", 0);
+    finish_shutdown_locked(exec_ctx, pollset);
+  }
+  GPR_TIMER_END("pollset_shutdown", 0);
+}
+
+/* pollset_shutdown is guaranteed to be called before pollset_destroy. So other
+ * than destroying the mutexes, there is nothing special that needs to be done
+ * here */
+static void pollset_destroy(grpc_pollset *pollset) {
+  GPR_ASSERT(!pollset_has_workers(pollset));
+  gpr_mu_destroy(&pollset->mu);
+}
+
+static void pollset_reset(grpc_pollset *pollset) {
+  GPR_ASSERT(pollset->shutting_down);
+  GPR_ASSERT(!pollset_has_workers(pollset));
+  pollset->shutting_down = false;
+  pollset->finish_shutdown_called = false;
+  pollset->kicked_without_pollers = false;
+  pollset->shutdown_done = NULL;
+  pollset_release_polling_island(pollset, "ps_reset");
+}
+
+#define GRPC_EPOLL_MAX_EVENTS 1000
+/* Note: sig_mask contains the signal mask to use *during* epoll_wait() */
+static void pollset_work_and_unlock(grpc_exec_ctx *exec_ctx,
+                                    grpc_pollset *pollset,
+                                    grpc_pollset_worker *worker, int timeout_ms,
+                                    sigset_t *sig_mask, grpc_error **error) {
+  struct epoll_event ep_ev[GRPC_EPOLL_MAX_EVENTS];
+  int epoll_fd = -1;
+  int ep_rv;
+  polling_island *pi = NULL;
+  char *err_msg;
+  const char *err_desc = "pollset_work_and_unlock";
+  GPR_TIMER_BEGIN("pollset_work_and_unlock", 0);
+
+  /* We need to get the epoll_fd to wait on. The epoll_fd is in inside the
+     latest polling island pointed by pollset->polling_island.
+
+     Since epoll_fd is immutable, we can read it without obtaining the polling
+     island lock. There is however a possibility that the polling island (from
+     which we got the epoll_fd) got merged with another island while we are
+     in this function. This is still okay because in such a case, we will wakeup
+     right-away from epoll_wait() and pick up the latest polling_island the next
+     this function (i.e pollset_work_and_unlock()) is called */
+
+  if (pollset->polling_island == NULL) {
+    pollset->polling_island = polling_island_create(NULL, error);
+    if (pollset->polling_island == NULL) {
+      GPR_TIMER_END("pollset_work_and_unlock", 0);
+      return; /* Fatal error. We cannot continue */
+    }
+
+    PI_ADD_REF(pollset->polling_island, "ps");
+    GRPC_POLLING_TRACE("pollset_work: pollset: %p created new pi: %p",
+                       (void *)pollset, (void *)pollset->polling_island);
+  }
+
+  pi = polling_island_maybe_get_latest(pollset->polling_island);
+  epoll_fd = pi->epoll_fd;
+
+  /* Update the pollset->polling_island since the island being pointed by
+     pollset->polling_island maybe older than the one pointed by pi) */
+  if (pollset->polling_island != pi) {
+    /* Always do PI_ADD_REF before PI_UNREF because PI_UNREF may cause the
+       polling island to be deleted */
+    PI_ADD_REF(pi, "ps");
+    PI_UNREF(pollset->polling_island, "ps");
+    pollset->polling_island = pi;
+  }
+
+  /* Add an extra ref so that the island does not get destroyed (which means
+     the epoll_fd won't be closed) while we are are doing an epoll_wait() on the
+     epoll_fd */
+  PI_ADD_REF(pi, "ps_work");
+  gpr_mu_unlock(&pollset->mu);
+
+  do {
+    ep_rv = epoll_pwait(epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, timeout_ms,
+                        sig_mask);
+    if (ep_rv < 0) {
+      if (errno != EINTR) {
+        gpr_asprintf(&err_msg,
+                     "epoll_wait() epoll fd: %d failed with error: %d (%s)",
+                     epoll_fd, errno, strerror(errno));
+        append_error(error, GRPC_OS_ERROR(errno, err_msg), err_desc);
+      } else {
+        /* We were interrupted. Save an interation by doing a zero timeout
+           epoll_wait to see if there are any other events of interest */
+        GRPC_POLLING_TRACE(
+            "pollset_work: pollset: %p, worker: %p received kick",
+            (void *)pollset, (void *)worker);
+        ep_rv = epoll_wait(epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, 0);
+      }
+    }
+
+#ifdef GRPC_TSAN
+    /* See the definition of g_poll_sync for more details */
+    gpr_atm_acq_load(&g_epoll_sync);
+#endif /* defined(GRPC_TSAN) */
+
+    for (int i = 0; i < ep_rv; ++i) {
+      void *data_ptr = ep_ev[i].data.ptr;
+      if (data_ptr == &grpc_global_wakeup_fd) {
+        append_error(error,
+                     grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd),
+                     err_desc);
+      } else if (data_ptr == &polling_island_wakeup_fd) {
+        GRPC_POLLING_TRACE(
+            "pollset_work: pollset: %p, worker: %p polling island (epoll_fd: "
+            "%d) got merged",
+            (void *)pollset, (void *)worker, epoll_fd);
+        /* This means that our polling island is merged with a different
+           island. We do not have to do anything here since the subsequent call
+           to the function pollset_work_and_unlock() will pick up the correct
+           epoll_fd */
+      } else {
+        grpc_fd *fd = data_ptr;
+        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 (read_ev || cancel) {
+          fd_become_readable(exec_ctx, fd, pollset);
+        }
+        if (write_ev || cancel) {
+          fd_become_writable(exec_ctx, fd);
+        }
+      }
+    }
+  } while (ep_rv == GRPC_EPOLL_MAX_EVENTS);
+
+  GPR_ASSERT(pi != NULL);
+
+  /* Before leaving, release the extra ref we added to the polling island. It
+     is important to use "pi" here (i.e our old copy of pollset->polling_island
+     that we got before releasing the polling island lock). This is because
+     pollset->polling_island pointer might get udpated in other parts of the
+     code when there is an island merge while we are doing epoll_wait() above */
+  PI_UNREF(pi, "ps_work");
+
+  GPR_TIMER_END("pollset_work_and_unlock", 0);
+}
+
+/* pollset->mu lock must be held by the caller before calling this.
+   The function pollset_work() may temporarily release the lock (pollset->mu)
+   during the course of its execution but it will always re-acquire the lock and
+   ensure that it is held by the time the function returns */
+static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
+                                grpc_pollset_worker **worker_hdl,
+                                gpr_timespec now, gpr_timespec deadline) {
+  GPR_TIMER_BEGIN("pollset_work", 0);
+  grpc_error *error = GRPC_ERROR_NONE;
+  int timeout_ms = poll_deadline_to_millis_timeout(deadline, now);
+
+  sigset_t new_mask;
+
+  grpc_pollset_worker worker;
+  worker.next = worker.prev = NULL;
+  worker.pt_id = pthread_self();
+  gpr_atm_no_barrier_store(&worker.is_kicked, (gpr_atm)0);
+
+  *worker_hdl = &worker;
+
+  gpr_tls_set(&g_current_thread_pollset, (intptr_t)pollset);
+  gpr_tls_set(&g_current_thread_worker, (intptr_t)&worker);
+
+  if (pollset->kicked_without_pollers) {
+    /* If the pollset was kicked without pollers, pretend that the current
+       worker got the kick and skip polling. A kick indicates that there is some
+       work that needs attention like an event on the completion queue or an
+       alarm */
+    GPR_TIMER_MARK("pollset_work.kicked_without_pollers", 0);
+    pollset->kicked_without_pollers = 0;
+  } else if (!pollset->shutting_down) {
+    /* We use the posix-signal with number 'grpc_wakeup_signal' for waking up
+       (i.e 'kicking') a worker in the pollset. A 'kick' is a way to inform the
+       worker that there is some pending work that needs immediate attention
+       (like an event on the completion queue, or a polling island merge that
+       results in a new epoll-fd to wait on) and that the worker should not
+       spend time waiting in epoll_pwait().
+
+       A worker can be kicked anytime from the point it is added to the pollset
+       via push_front_worker() (or push_back_worker()) to the point it is
+       removed via remove_worker().
+       If the worker is kicked before/during it calls epoll_pwait(), it should
+       immediately exit from epoll_wait(). If the worker is kicked after it
+       returns from epoll_wait(), then nothing really needs to be done.
+
+       To accomplish this, we mask 'grpc_wakeup_signal' on this thread at all
+       times *except* when it is in epoll_pwait(). This way, the worker never
+       misses acting on a kick */
+
+    if (!g_initialized_sigmask) {
+      sigemptyset(&new_mask);
+      sigaddset(&new_mask, grpc_wakeup_signal);
+      pthread_sigmask(SIG_BLOCK, &new_mask, &g_orig_sigmask);
+      sigdelset(&g_orig_sigmask, grpc_wakeup_signal);
+      g_initialized_sigmask = true;
+      /* new_mask:       The new thread mask which blocks 'grpc_wakeup_signal'.
+                         This is the mask used at all times *except during
+                         epoll_wait()*"
+         g_orig_sigmask: The thread mask which allows 'grpc_wakeup_signal' and
+                         this is the mask to use *during epoll_wait()*
+
+         The new_mask is set on the worker before it is added to the pollset
+         (i.e before it can be kicked) */
+    }
+
+    push_front_worker(pollset, &worker); /* Add worker to pollset */
+
+    pollset_work_and_unlock(exec_ctx, pollset, &worker, timeout_ms,
+                            &g_orig_sigmask, &error);
+    grpc_exec_ctx_flush(exec_ctx);
+
+    gpr_mu_lock(&pollset->mu);
+
+    /* Note: There is no need to reset worker.is_kicked to 0 since we are no
+       longer going to use this worker */
+    remove_worker(pollset, &worker);
+  }
+
+  /* If we are the last worker on the pollset (i.e pollset_has_workers() is
+     false at this point) and the pollset is shutting down, we may have to
+     finish the shutdown process by calling finish_shutdown_locked().
+     See pollset_shutdown() for more details.
+
+     Note: Continuing to access pollset here is safe; it is the caller's
+     responsibility to not destroy a pollset when it has outstanding calls to
+     pollset_work() */
+  if (pollset->shutting_down && !pollset_has_workers(pollset) &&
+      !pollset->finish_shutdown_called) {
+    GPR_TIMER_MARK("pollset_work.finish_shutdown_locked", 0);
+    finish_shutdown_locked(exec_ctx, pollset);
+
+    gpr_mu_unlock(&pollset->mu);
+    grpc_exec_ctx_flush(exec_ctx);
+    gpr_mu_lock(&pollset->mu);
+  }
+
+  *worker_hdl = NULL;
+
+  gpr_tls_set(&g_current_thread_pollset, (intptr_t)0);
+  gpr_tls_set(&g_current_thread_worker, (intptr_t)0);
+
+  GPR_TIMER_END("pollset_work", 0);
+
+  GRPC_LOG_IF_ERROR("pollset_work", GRPC_ERROR_REF(error));
+  return error;
+}
+
+static void pollset_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
+                           grpc_fd *fd) {
+  grpc_error *error = GRPC_ERROR_NONE;
+
+  gpr_mu_lock(&pollset->mu);
+  gpr_mu_lock(&fd->pi_mu);
+
+  polling_island *pi_new = NULL;
+
+  /* 1) If fd->polling_island and pollset->polling_island are both non-NULL and
+   *    equal, do nothing.
+   * 2) If fd->polling_island and pollset->polling_island are both NULL, create
+   *    a new polling island (with a refcount of 2) and make the polling_island
+   *    fields in both fd and pollset to point to the new island
+   * 3) If one of fd->polling_island or pollset->polling_island is NULL, update
+   *    the NULL polling_island field to point to the non-NULL polling_island
+   *    field (ensure that the refcount on the polling island is incremented by
+   *    1 to account for the newly added reference)
+   * 4) Finally, if fd->polling_island and pollset->polling_island are non-NULL
+   *    and different, merge both the polling islands and update the
+   *    polling_island fields in both fd and pollset to point to the merged
+   *    polling island.
+   */
+  if (fd->polling_island == pollset->polling_island) {
+    pi_new = fd->polling_island;
+    if (pi_new == NULL) {
+      pi_new = polling_island_create(fd, &error);
+
+      GRPC_POLLING_TRACE(
+          "pollset_add_fd: Created new polling island. pi_new: %p (fd: %d, "
+          "pollset: %p)",
+          (void *)pi_new, fd->fd, (void *)pollset);
+    }
+  } else if (fd->polling_island == NULL) {
+    pi_new = polling_island_lock(pollset->polling_island);
+    polling_island_add_fds_locked(pi_new, &fd, 1, true, &error);
+    gpr_mu_unlock(&pi_new->mu);
+
+    GRPC_POLLING_TRACE(
+        "pollset_add_fd: fd->pi was NULL. pi_new: %p (fd: %d, pollset: %p, "
+        "pollset->pi: %p)",
+        (void *)pi_new, fd->fd, (void *)pollset,
+        (void *)pollset->polling_island);
+  } else if (pollset->polling_island == NULL) {
+    pi_new = polling_island_lock(fd->polling_island);
+    gpr_mu_unlock(&pi_new->mu);
+
+    GRPC_POLLING_TRACE(
+        "pollset_add_fd: pollset->pi was NULL. pi_new: %p (fd: %d, pollset: "
+        "%p, fd->pi: %p",
+        (void *)pi_new, fd->fd, (void *)pollset, (void *)fd->polling_island);
+  } else {
+    pi_new = polling_island_merge(fd->polling_island, pollset->polling_island,
+                                  &error);
+    GRPC_POLLING_TRACE(
+        "pollset_add_fd: polling islands merged. pi_new: %p (fd: %d, pollset: "
+        "%p, fd->pi: %p, pollset->pi: %p)",
+        (void *)pi_new, fd->fd, (void *)pollset, (void *)fd->polling_island,
+        (void *)pollset->polling_island);
+  }
+
+  /* At this point, pi_new is the polling island that both fd->polling_island
+     and pollset->polling_island must be pointing to */
+
+  if (fd->polling_island != pi_new) {
+    PI_ADD_REF(pi_new, "fd");
+    if (fd->polling_island != NULL) {
+      PI_UNREF(fd->polling_island, "fd");
+    }
+    fd->polling_island = pi_new;
+  }
+
+  if (pollset->polling_island != pi_new) {
+    PI_ADD_REF(pi_new, "ps");
+    if (pollset->polling_island != NULL) {
+      PI_UNREF(pollset->polling_island, "ps");
+    }
+    pollset->polling_island = pi_new;
+  }
+
+  gpr_mu_unlock(&fd->pi_mu);
+  gpr_mu_unlock(&pollset->mu);
+}
+
+/*******************************************************************************
+ * Pollset-set Definitions
+ */
+
+static grpc_pollset_set *pollset_set_create(void) {
+  grpc_pollset_set *pollset_set = gpr_malloc(sizeof(*pollset_set));
+  memset(pollset_set, 0, sizeof(*pollset_set));
+  gpr_mu_init(&pollset_set->mu);
+  return pollset_set;
+}
+
+static void pollset_set_destroy(grpc_pollset_set *pollset_set) {
+  size_t i;
+  gpr_mu_destroy(&pollset_set->mu);
+  for (i = 0; i < pollset_set->fd_count; i++) {
+    GRPC_FD_UNREF(pollset_set->fds[i], "pollset_set");
+  }
+  gpr_free(pollset_set->pollsets);
+  gpr_free(pollset_set->pollset_sets);
+  gpr_free(pollset_set->fds);
+  gpr_free(pollset_set);
+}
+
+static void pollset_set_add_fd(grpc_exec_ctx *exec_ctx,
+                               grpc_pollset_set *pollset_set, grpc_fd *fd) {
+  size_t i;
+  gpr_mu_lock(&pollset_set->mu);
+  if (pollset_set->fd_count == pollset_set->fd_capacity) {
+    pollset_set->fd_capacity = GPR_MAX(8, 2 * pollset_set->fd_capacity);
+    pollset_set->fds = gpr_realloc(
+        pollset_set->fds, pollset_set->fd_capacity * sizeof(*pollset_set->fds));
+  }
+  GRPC_FD_REF(fd, "pollset_set");
+  pollset_set->fds[pollset_set->fd_count++] = fd;
+  for (i = 0; i < pollset_set->pollset_count; i++) {
+    pollset_add_fd(exec_ctx, pollset_set->pollsets[i], fd);
+  }
+  for (i = 0; i < pollset_set->pollset_set_count; i++) {
+    pollset_set_add_fd(exec_ctx, pollset_set->pollset_sets[i], fd);
+  }
+  gpr_mu_unlock(&pollset_set->mu);
+}
+
+static void pollset_set_del_fd(grpc_exec_ctx *exec_ctx,
+                               grpc_pollset_set *pollset_set, grpc_fd *fd) {
+  size_t i;
+  gpr_mu_lock(&pollset_set->mu);
+  for (i = 0; i < pollset_set->fd_count; i++) {
+    if (pollset_set->fds[i] == fd) {
+      pollset_set->fd_count--;
+      GPR_SWAP(grpc_fd *, pollset_set->fds[i],
+               pollset_set->fds[pollset_set->fd_count]);
+      GRPC_FD_UNREF(fd, "pollset_set");
+      break;
+    }
+  }
+  for (i = 0; i < pollset_set->pollset_set_count; i++) {
+    pollset_set_del_fd(exec_ctx, pollset_set->pollset_sets[i], fd);
+  }
+  gpr_mu_unlock(&pollset_set->mu);
+}
+
+static void pollset_set_add_pollset(grpc_exec_ctx *exec_ctx,
+                                    grpc_pollset_set *pollset_set,
+                                    grpc_pollset *pollset) {
+  size_t i, j;
+  gpr_mu_lock(&pollset_set->mu);
+  if (pollset_set->pollset_count == pollset_set->pollset_capacity) {
+    pollset_set->pollset_capacity =
+        GPR_MAX(8, 2 * pollset_set->pollset_capacity);
+    pollset_set->pollsets =
+        gpr_realloc(pollset_set->pollsets, pollset_set->pollset_capacity *
+                                               sizeof(*pollset_set->pollsets));
+  }
+  pollset_set->pollsets[pollset_set->pollset_count++] = pollset;
+  for (i = 0, j = 0; i < pollset_set->fd_count; i++) {
+    if (fd_is_orphaned(pollset_set->fds[i])) {
+      GRPC_FD_UNREF(pollset_set->fds[i], "pollset_set");
+    } else {
+      pollset_add_fd(exec_ctx, pollset, pollset_set->fds[i]);
+      pollset_set->fds[j++] = pollset_set->fds[i];
+    }
+  }
+  pollset_set->fd_count = j;
+  gpr_mu_unlock(&pollset_set->mu);
+}
+
+static void pollset_set_del_pollset(grpc_exec_ctx *exec_ctx,
+                                    grpc_pollset_set *pollset_set,
+                                    grpc_pollset *pollset) {
+  size_t i;
+  gpr_mu_lock(&pollset_set->mu);
+  for (i = 0; i < pollset_set->pollset_count; i++) {
+    if (pollset_set->pollsets[i] == pollset) {
+      pollset_set->pollset_count--;
+      GPR_SWAP(grpc_pollset *, pollset_set->pollsets[i],
+               pollset_set->pollsets[pollset_set->pollset_count]);
+      break;
+    }
+  }
+  gpr_mu_unlock(&pollset_set->mu);
+}
+
+static void pollset_set_add_pollset_set(grpc_exec_ctx *exec_ctx,
+                                        grpc_pollset_set *bag,
+                                        grpc_pollset_set *item) {
+  size_t i, j;
+  gpr_mu_lock(&bag->mu);
+  if (bag->pollset_set_count == bag->pollset_set_capacity) {
+    bag->pollset_set_capacity = GPR_MAX(8, 2 * bag->pollset_set_capacity);
+    bag->pollset_sets =
+        gpr_realloc(bag->pollset_sets,
+                    bag->pollset_set_capacity * sizeof(*bag->pollset_sets));
+  }
+  bag->pollset_sets[bag->pollset_set_count++] = item;
+  for (i = 0, j = 0; i < bag->fd_count; i++) {
+    if (fd_is_orphaned(bag->fds[i])) {
+      GRPC_FD_UNREF(bag->fds[i], "pollset_set");
+    } else {
+      pollset_set_add_fd(exec_ctx, item, bag->fds[i]);
+      bag->fds[j++] = bag->fds[i];
+    }
+  }
+  bag->fd_count = j;
+  gpr_mu_unlock(&bag->mu);
+}
+
+static void pollset_set_del_pollset_set(grpc_exec_ctx *exec_ctx,
+                                        grpc_pollset_set *bag,
+                                        grpc_pollset_set *item) {
+  size_t i;
+  gpr_mu_lock(&bag->mu);
+  for (i = 0; i < bag->pollset_set_count; i++) {
+    if (bag->pollset_sets[i] == item) {
+      bag->pollset_set_count--;
+      GPR_SWAP(grpc_pollset_set *, bag->pollset_sets[i],
+               bag->pollset_sets[bag->pollset_set_count]);
+      break;
+    }
+  }
+  gpr_mu_unlock(&bag->mu);
+}
+
+/* Test helper functions
+ * */
+void *grpc_fd_get_polling_island(grpc_fd *fd) {
+  polling_island *pi;
+
+  gpr_mu_lock(&fd->pi_mu);
+  pi = fd->polling_island;
+  gpr_mu_unlock(&fd->pi_mu);
+
+  return pi;
+}
+
+void *grpc_pollset_get_polling_island(grpc_pollset *ps) {
+  polling_island *pi;
+
+  gpr_mu_lock(&ps->mu);
+  pi = ps->polling_island;
+  gpr_mu_unlock(&ps->mu);
+
+  return pi;
+}
+
+bool grpc_are_polling_islands_equal(void *p, void *q) {
+  polling_island *p1 = p;
+  polling_island *p2 = q;
+
+  /* Note: polling_island_lock_pair() may change p1 and p2 to point to the
+     latest polling islands in their respective linked lists */
+  polling_island_lock_pair(&p1, &p2);
+  polling_island_unlock_pair(p1, p2);
+
+  return p1 == p2;
+}
+
+/*******************************************************************************
+ * Event engine binding
+ */
+
+static void shutdown_engine(void) {
+  fd_global_shutdown();
+  pollset_global_shutdown();
+  polling_island_global_shutdown();
+}
+
+static const grpc_event_engine_vtable vtable = {
+    .pollset_size = sizeof(grpc_pollset),
+
+    .fd_create = fd_create,
+    .fd_wrapped_fd = fd_wrapped_fd,
+    .fd_orphan = fd_orphan,
+    .fd_shutdown = fd_shutdown,
+    .fd_is_shutdown = fd_is_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,
+    .pollset_reset = pollset_reset,
+    .pollset_destroy = pollset_destroy,
+    .pollset_work = pollset_work,
+    .pollset_kick = pollset_kick,
+    .pollset_add_fd = pollset_add_fd,
+
+    .pollset_set_create = pollset_set_create,
+    .pollset_set_destroy = pollset_set_destroy,
+    .pollset_set_add_pollset = pollset_set_add_pollset,
+    .pollset_set_del_pollset = pollset_set_del_pollset,
+    .pollset_set_add_pollset_set = pollset_set_add_pollset_set,
+    .pollset_set_del_pollset_set = pollset_set_del_pollset_set,
+    .pollset_set_add_fd = pollset_set_add_fd,
+    .pollset_set_del_fd = pollset_set_del_fd,
+
+    .kick_poller = kick_poller,
+
+    .shutdown_engine = shutdown_engine,
+};
+
+/* It is possible that GLIBC has epoll but the underlying kernel doesn't.
+ * Create a dummy epoll_fd to make sure epoll support is available */
+static bool is_epoll_available() {
+  int fd = epoll_create1(EPOLL_CLOEXEC);
+  if (fd < 0) {
+    gpr_log(
+        GPR_ERROR,
+        "epoll_create1 failed with error: %d. Not using epoll polling engine",
+        fd);
+    return false;
+  }
+  close(fd);
+  return true;
+}
+
+const grpc_event_engine_vtable *grpc_init_epoll_linux(void) {
+  /* If use of signals is disabled, we cannot use epoll engine*/
+  if (is_grpc_wakeup_signal_initialized && grpc_wakeup_signal < 0) {
+    return NULL;
+  }
+
+  if (!is_epoll_available()) {
+    return NULL;
+  }
+
+  if (!is_grpc_wakeup_signal_initialized) {
+    grpc_use_signal(SIGRTMIN + 2);
+  }
+
+  fd_global_init();
+
+  if (!GRPC_LOG_IF_ERROR("pollset_global_init", pollset_global_init())) {
+    return NULL;
+  }
+
+  if (!GRPC_LOG_IF_ERROR("polling_island_global_init",
+                         polling_island_global_init())) {
+    return NULL;
+  }
+
+  return &vtable;
+}
+
+#else /* defined(GPR_LINUX_EPOLL) */
+#if defined(GPR_POSIX_SOCKET)
+#include "src/core/lib/iomgr/ev_posix.h"
+/* If GPR_LINUX_EPOLL is not defined, it means epoll is not available. Return
+ * NULL */
+const grpc_event_engine_vtable *grpc_init_epoll_linux(void) { return NULL; }
+#endif /* defined(GPR_POSIX_SOCKET) */
+
+void grpc_use_signal(int signum) {}
+#endif /* !defined(GPR_LINUX_EPOLL) */
diff --git a/src/core/lib/iomgr/ev_epoll_linux.h b/src/core/lib/iomgr/ev_epoll_linux.h
new file mode 100644
index 0000000000..7a494aba19
--- /dev/null
+++ b/src/core/lib/iomgr/ev_epoll_linux.h
@@ -0,0 +1,47 @@
+/*
+ *
+ * Copyright 2015, Google Inc.
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above
+ * copyright notice, this list of conditions and the following disclaimer
+ * in the documentation and/or other materials provided with the
+ * distribution.
+ *     * Neither the name of Google Inc. nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ */
+
+#ifndef GRPC_CORE_LIB_IOMGR_EV_EPOLL_LINUX_H
+#define GRPC_CORE_LIB_IOMGR_EV_EPOLL_LINUX_H
+
+#include "src/core/lib/iomgr/ev_posix.h"
+
+const grpc_event_engine_vtable *grpc_init_epoll_linux(void);
+
+#ifdef GPR_LINUX_EPOLL
+void *grpc_fd_get_polling_island(grpc_fd *fd);
+void *grpc_pollset_get_polling_island(grpc_pollset *ps);
+bool grpc_are_polling_islands_equal(void *p, void *q);
+#endif /* defined(GPR_LINUX_EPOLL) */
+
+#endif /* GRPC_CORE_LIB_IOMGR_EV_EPOLL_LINUX_H */
diff --git a/src/core/lib/iomgr/ev_poll_and_epoll_posix.c b/src/core/lib/iomgr/ev_poll_and_epoll_posix.c
new file mode 100644
index 0000000000..9e306af5fa
--- /dev/null
+++ b/src/core/lib/iomgr/ev_poll_and_epoll_posix.c
@@ -0,0 +1,2043 @@
+/*
+ *
+ * Copyright 2015, Google Inc.
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above
+ * copyright notice, this list of conditions and the following disclaimer
+ * in the documentation and/or other materials provided with the
+ * distribution.
+ *     * Neither the name of Google Inc. nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ */
+
+/* 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 <assert.h>
+#include <errno.h>
+#include <poll.h>
+#include <string.h>
+#include <sys/socket.h>
+#include <unistd.h>
+
+#include <grpc/support/alloc.h>
+#include <grpc/support/log.h>
+#include <grpc/support/string_util.h>
+#include <grpc/support/tls.h>
+#include <grpc/support/useful.h>
+
+#include "src/core/lib/iomgr/iomgr_internal.h"
+#include "src/core/lib/iomgr/wakeup_fd_posix.h"
+#include "src/core/lib/profiling/timers.h"
+#include "src/core/lib/support/block_annotate.h"
+
+/*******************************************************************************
+ * FD declarations
+ */
+
+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;
+
+struct grpc_fd {
+  int fd;
+  /* refst format:
+     bit0:   1=active/0=orphaned
+     bit1-n: refcount
+     meaning that mostly we ref by two to avoid altering the orphaned bit,
+     and just unref by 1 when we're ready to flag the object as orphaned */
+  gpr_atm refst;
+
+  gpr_mu mu;
+  int shutdown;
+  int closed;
+  int released;
+
+  /* The watcher list.
+
+     The following watcher related fields are protected by watcher_mu.
+
+     An fd_watcher is an ephemeral object created when an fd wants to
+     begin polling, and destroyed after the poll.
+
+     It denotes the fd's interest in whether to read poll or write poll
+     or both or neither on this fd.
+
+     If a watcher is asked to poll for reads or writes, the read_watcher
+     or write_watcher fields are set respectively. A watcher may be asked
+     to poll for both, in which case both fields will be set.
+
+     read_watcher and write_watcher may be NULL if no watcher has been
+     asked to poll for reads or writes.
+
+     If an fd_watcher is not asked to poll for reads or writes, it's added
+     to a linked list of inactive watchers, rooted at inactive_watcher_root.
+     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. */
+  grpc_fd_watcher inactive_watcher_root;
+  grpc_fd_watcher *read_watcher;
+  grpc_fd_watcher *write_watcher;
+
+  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.
+   Registers that the given pollset is interested in this fd - so that if read
+   or writability interest changes, the pollset can be kicked to pick up that
+   new interest.
+   Return value is:
+     (fd_needs_read? read_mask : 0) | (fd_needs_write? write_mask : 0)
+   i.e. a combination of read_mask and write_mask determined by the fd's current
+   interest in said events.
+   Polling strategies that do not need to alter their behavior depending on the
+   fd's current interest (such as epoll) do not need to call this function.
+   MUST NOT be called with a pollset lock taken */
+static uint32_t fd_begin_poll(grpc_fd *fd, grpc_pollset *pollset,
+                              grpc_pollset_worker *worker, uint32_t read_mask,
+                              uint32_t write_mask, grpc_fd_watcher *rec);
+/* Complete polling previously started with fd_begin_poll
+   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. */
+static void fd_end_poll(grpc_exec_ctx *exec_ctx, grpc_fd_watcher *rec,
+                        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);
+
+/* Reference counting for fds */
+/*#define GRPC_FD_REF_COUNT_DEBUG*/
+#ifdef GRPC_FD_REF_COUNT_DEBUG
+static void fd_ref(grpc_fd *fd, const char *reason, const char *file, int line);
+static void fd_unref(grpc_fd *fd, const char *reason, const char *file,
+                     int line);
+#define GRPC_FD_REF(fd, reason) fd_ref(fd, reason, __FILE__, __LINE__)
+#define GRPC_FD_UNREF(fd, reason) fd_unref(fd, reason, __FILE__, __LINE__)
+#else
+static void fd_ref(grpc_fd *fd);
+static void fd_unref(grpc_fd *fd);
+#define GRPC_FD_REF(fd, reason) fd_ref(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)
+
+/*******************************************************************************
+ * 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;
+} grpc_cached_wakeup_fd;
+
+struct grpc_pollset_worker {
+  grpc_cached_wakeup_fd *wakeup_fd;
+  int reevaluate_polling_on_wakeup;
+  int kicked_specifically;
+  struct grpc_pollset_worker *next;
+  struct grpc_pollset_worker *prev;
+};
+
+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;
+  /* 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);
+  grpc_error *(*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);
+
+static void pollset_set_add_fd(grpc_exec_ctx *exec_ctx,
+                               grpc_pollset_set *pollset_set, grpc_fd *fd);
+
+/* Convert a timespec to milliseconds:
+   - very small or negative poll times are clamped to zero to do a
+     non-blocking poll (which becomes spin polling)
+   - other small values are rounded up to one millisecond
+   - longer than a millisecond polls are rounded up to the next nearest
+     millisecond to avoid spinning
+   - infinite timeouts are converted to -1 */
+static int 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 pollset_kick, with an extended set of flags (defined above)
+   -- mostly for fd_posix's use. */
+static grpc_error *pollset_kick_ext(grpc_pollset *p,
+                                    grpc_pollset_worker *specific_worker,
+                                    uint32_t flags) GRPC_MUST_USE_RESULT;
+
+/* 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
+ */
+
+struct grpc_pollset_set {
+  gpr_mu mu;
+
+  size_t pollset_count;
+  size_t pollset_capacity;
+  grpc_pollset **pollsets;
+
+  size_t pollset_set_count;
+  size_t pollset_set_capacity;
+  struct grpc_pollset_set **pollset_sets;
+
+  size_t fd_count;
+  size_t fd_capacity;
+  grpc_fd **fds;
+};
+
+/*******************************************************************************
+ * 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;
+  r->read_notifier_pollset = NULL;
+  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__)
+static void ref_by(grpc_fd *fd, int n, const char *reason, const char *file,
+                   int line) {
+  gpr_log(GPR_DEBUG, "FD %d %p   ref %d %d -> %d [%s; %s:%d]", fd->fd, fd, n,
+          gpr_atm_no_barrier_load(&fd->refst),
+          gpr_atm_no_barrier_load(&fd->refst) + n, reason, file, line);
+#else
+#define REF_BY(fd, n, reason) ref_by(fd, n)
+#define UNREF_BY(fd, n, reason) unref_by(fd, n)
+static void ref_by(grpc_fd *fd, int n) {
+#endif
+  GPR_ASSERT(gpr_atm_no_barrier_fetch_add(&fd->refst, n) > 0);
+}
+
+#ifdef GRPC_FD_REF_COUNT_DEBUG
+static void unref_by(grpc_fd *fd, int n, const char *reason, const char *file,
+                     int line) {
+  gpr_atm old;
+  gpr_log(GPR_DEBUG, "FD %d %p unref %d %d -> %d [%s; %s:%d]", fd->fd, fd, n,
+          gpr_atm_no_barrier_load(&fd->refst),
+          gpr_atm_no_barrier_load(&fd->refst) - n, reason, file, line);
+#else
+static void unref_by(grpc_fd *fd, int n) {
+  gpr_atm old;
+#endif
+  old = gpr_atm_full_fetch_add(&fd->refst, -n);
+  if (old == n) {
+    freelist_fd(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);
+  char *name2;
+  gpr_asprintf(&name2, "%s fd=%d", name, fd);
+  grpc_iomgr_register_object(&r->iomgr_object, name2);
+  gpr_free(name2);
+#ifdef GRPC_FD_REF_COUNT_DEBUG
+  gpr_log(GPR_DEBUG, "FD %d %p create %s", fd, r, name);
+#endif
+  return r;
+}
+
+static bool fd_is_orphaned(grpc_fd *fd) {
+  return (gpr_atm_acq_load(&fd->refst) & 1) == 0;
+}
+
+static grpc_error *pollset_kick_locked(grpc_fd_watcher *watcher) {
+  gpr_mu_lock(&watcher->pollset->mu);
+  GPR_ASSERT(watcher->worker);
+  grpc_error *err = pollset_kick_ext(watcher->pollset, watcher->worker,
+                                     GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP);
+  gpr_mu_unlock(&watcher->pollset->mu);
+  return err;
+}
+
+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);
+  } else if (fd->read_watcher) {
+    pollset_kick_locked(fd->read_watcher);
+  } else if (fd->write_watcher) {
+    pollset_kick_locked(fd->write_watcher);
+  }
+}
+
+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);
+  }
+  if (fd->read_watcher) {
+    pollset_kick_locked(fd->read_watcher);
+  }
+  if (fd->write_watcher && fd->write_watcher != fd->read_watcher) {
+    pollset_kick_locked(fd->write_watcher);
+  }
+}
+
+static int has_watchers(grpc_fd *fd) {
+  return fd->read_watcher != NULL || fd->write_watcher != NULL ||
+         fd->inactive_watcher_root.next != &fd->inactive_watcher_root;
+}
+
+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_sched(exec_ctx, fd->on_done_closure, GRPC_ERROR_NONE, NULL);
+}
+
+static int fd_wrapped_fd(grpc_fd *fd) {
+  if (fd->released || fd->closed) {
+    return -1;
+  } else {
+    return fd->fd;
+  }
+}
+
+static void fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
+                      grpc_closure *on_done, int *release_fd,
+                      const char *reason) {
+  fd->on_done_closure = on_done;
+  fd->released = release_fd != NULL;
+  if (!fd->released) {
+    shutdown(fd->fd, SHUT_RDWR);
+  } else {
+    *release_fd = fd->fd;
+  }
+  gpr_mu_lock(&fd->mu);
+  REF_BY(fd, 1, reason); /* remove active status, but keep referenced */
+  if (!has_watchers(fd)) {
+    close_fd_locked(exec_ctx, fd);
+  } else {
+    wake_all_watchers_locked(fd);
+  }
+  gpr_mu_unlock(&fd->mu);
+  UNREF_BY(fd, 2, reason); /* drop the reference */
+}
+
+/* increment refcount by two to avoid changing the orphan bit */
+#ifdef GRPC_FD_REF_COUNT_DEBUG
+static void fd_ref(grpc_fd *fd, const char *reason, const char *file,
+                   int line) {
+  ref_by(fd, 2, reason, file, line);
+}
+
+static void fd_unref(grpc_fd *fd, const char *reason, const char *file,
+                     int line) {
+  unref_by(fd, 2, reason, file, line);
+}
+#else
+static void fd_ref(grpc_fd *fd) { ref_by(fd, 2); }
+
+static void fd_unref(grpc_fd *fd) { unref_by(fd, 2); }
+#endif
+
+static grpc_error *fd_shutdown_error(bool shutdown) {
+  if (!shutdown) {
+    return GRPC_ERROR_NONE;
+  } else {
+    return GRPC_ERROR_CREATE("FD shutdown");
+  }
+}
+
+static void notify_on_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
+                             grpc_closure **st, grpc_closure *closure) {
+  if (fd->shutdown) {
+    grpc_exec_ctx_sched(exec_ctx, closure, GRPC_ERROR_CREATE("FD shutdown"),
+                        NULL);
+  } else 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_sched(exec_ctx, closure, fd_shutdown_error(fd->shutdown),
+                        NULL);
+    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();
+  }
+}
+
+/* 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_sched(exec_ctx, *st, fd_shutdown_error(fd->shutdown), NULL);
+    *st = CLOSURE_NOT_READY;
+    return 1;
+  }
+}
+
+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);
+  /* only shutdown once */
+  if (!fd->shutdown) {
+    fd->shutdown = 1;
+    /* signal read/write closed to OS so that future operations fail */
+    shutdown(fd->fd, SHUT_RDWR);
+    set_ready_locked(exec_ctx, fd, &fd->read_closure);
+    set_ready_locked(exec_ctx, fd, &fd->write_closure);
+  }
+  gpr_mu_unlock(&fd->mu);
+}
+
+static bool fd_is_shutdown(grpc_fd *fd) {
+  gpr_mu_lock(&fd->mu);
+  bool r = fd->shutdown;
+  gpr_mu_unlock(&fd->mu);
+  return r;
+}
+
+static void fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
+                              grpc_closure *closure) {
+  gpr_mu_lock(&fd->mu);
+  notify_on_locked(exec_ctx, fd, &fd->read_closure, closure);
+  gpr_mu_unlock(&fd->mu);
+}
+
+static void fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
+                               grpc_closure *closure) {
+  gpr_mu_lock(&fd->mu);
+  notify_on_locked(exec_ctx, fd, &fd->write_closure, closure);
+  gpr_mu_unlock(&fd->mu);
+}
+
+/* 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 uint32_t fd_begin_poll(grpc_fd *fd, grpc_pollset *pollset,
+                              grpc_pollset_worker *worker, uint32_t read_mask,
+                              uint32_t write_mask, grpc_fd_watcher *watcher) {
+  uint32_t 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->mu);
+
+  /* if we are shutdown, then don't add to the watcher set */
+  if (fd->shutdown) {
+    watcher->fd = NULL;
+    watcher->pollset = NULL;
+    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 */
+  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
+   */
+  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 && 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->mu);
+
+  return mask;
+}
+
+static void fd_end_poll(grpc_exec_ctx *exec_ctx, grpc_fd_watcher *watcher,
+                        int got_read, int got_write,
+                        grpc_pollset *read_notifier_pollset) {
+  int was_polling = 0;
+  int kick = 0;
+  grpc_fd *fd = watcher->fd;
+
+  if (fd == NULL) {
+    return;
+  }
+
+  gpr_mu_lock(&fd->mu);
+
+  if (watcher == fd->read_watcher) {
+    /* remove read watcher, kick if we still need a read */
+    was_polling = 1;
+    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;
+    if (!got_write) {
+      kick = 1;
+    }
+    fd->write_watcher = NULL;
+  }
+  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 (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)) {
+      kick = 1;
+    }
+  }
+  if (kick) {
+    maybe_wake_one_watcher_locked(fd);
+  }
+  if (fd_is_orphaned(fd) && !has_watchers(fd) && !fd->closed) {
+    close_fd_locked(exec_ctx, fd);
+  }
+  gpr_mu_unlock(&fd->mu);
+
+  GRPC_FD_UNREF(fd, "poll");
+}
+
+/*******************************************************************************
+ * pollset_posix.c
+ */
+
+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;
+}
+
+static int pollset_has_workers(grpc_pollset *p) {
+  return p->root_worker.next != &p->root_worker;
+}
+
+static grpc_pollset_worker *pop_front_worker(grpc_pollset *p) {
+  if (pollset_has_workers(p)) {
+    grpc_pollset_worker *w = p->root_worker.next;
+    remove_worker(p, w);
+    return w;
+  } else {
+    return NULL;
+  }
+}
+
+static void push_back_worker(grpc_pollset *p, grpc_pollset_worker *worker) {
+  worker->next = &p->root_worker;
+  worker->prev = worker->next->prev;
+  worker->prev->next = worker->next->prev = worker;
+}
+
+static void push_front_worker(grpc_pollset *p, grpc_pollset_worker *worker) {
+  worker->prev = &p->root_worker;
+  worker->next = worker->prev->next;
+  worker->prev->next = worker->next->prev = worker;
+}
+
+static void kick_append_error(grpc_error **composite, grpc_error *error) {
+  if (error == GRPC_ERROR_NONE) return;
+  if (*composite == GRPC_ERROR_NONE) {
+    *composite = GRPC_ERROR_CREATE("Kick Failure");
+  }
+  *composite = grpc_error_add_child(*composite, error);
+}
+
+static grpc_error *pollset_kick_ext(grpc_pollset *p,
+                                    grpc_pollset_worker *specific_worker,
+                                    uint32_t flags) {
+  GPR_TIMER_BEGIN("pollset_kick_ext", 0);
+  grpc_error *error = GRPC_ERROR_NONE;
+
+  /* pollset->mu already held */
+  if (specific_worker != NULL) {
+    if (specific_worker == GRPC_POLLSET_KICK_BROADCAST) {
+      GPR_TIMER_BEGIN("pollset_kick_ext.broadcast", 0);
+      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) {
+        kick_append_error(
+            &error, grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd));
+      }
+      p->kicked_without_pollers = true;
+      GPR_TIMER_END("pollset_kick_ext.broadcast", 0);
+    } else if (gpr_tls_get(&g_current_thread_worker) !=
+               (intptr_t)specific_worker) {
+      GPR_TIMER_MARK("different_thread_worker", 0);
+      if ((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) != 0) {
+        specific_worker->reevaluate_polling_on_wakeup = true;
+      }
+      specific_worker->kicked_specifically = true;
+      kick_append_error(&error,
+                        grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd));
+    } else if ((flags & GRPC_POLLSET_CAN_KICK_SELF) != 0) {
+      GPR_TIMER_MARK("kick_yoself", 0);
+      if ((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) != 0) {
+        specific_worker->reevaluate_polling_on_wakeup = true;
+      }
+      specific_worker->kicked_specifically = true;
+      kick_append_error(&error,
+                        grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd));
+    }
+  } else if (gpr_tls_get(&g_current_thread_poller) != (intptr_t)p) {
+    GPR_ASSERT((flags & GRPC_POLLSET_REEVALUATE_POLLING_ON_WAKEUP) == 0);
+    GPR_TIMER_MARK("kick_anonymous", 0);
+    specific_worker = pop_front_worker(p);
+    if (specific_worker != NULL) {
+      if (gpr_tls_get(&g_current_thread_worker) == (intptr_t)specific_worker) {
+        GPR_TIMER_MARK("kick_anonymous_not_self", 0);
+        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) ==
+                (intptr_t)specific_worker) {
+          push_back_worker(p, specific_worker);
+          specific_worker = NULL;
+        }
+      }
+      if (specific_worker != NULL) {
+        GPR_TIMER_MARK("finally_kick", 0);
+        push_back_worker(p, specific_worker);
+        kick_append_error(
+            &error, grpc_wakeup_fd_wakeup(&specific_worker->wakeup_fd->fd));
+      }
+    } else {
+      GPR_TIMER_MARK("kicked_no_pollers", 0);
+      p->kicked_without_pollers = true;
+    }
+  }
+
+  GPR_TIMER_END("pollset_kick_ext", 0);
+  return error;
+}
+
+static grpc_error *pollset_kick(grpc_pollset *p,
+                                grpc_pollset_worker *specific_worker) {
+  return pollset_kick_ext(p, specific_worker, 0);
+}
+
+/* global state management */
+
+static grpc_error *pollset_global_init(void) {
+  gpr_tls_init(&g_current_thread_poller);
+  gpr_tls_init(&g_current_thread_worker);
+  return grpc_wakeup_fd_init(&grpc_global_wakeup_fd);
+}
+
+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);
+}
+
+static grpc_error *kick_poller(void) {
+  return 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);
+}
+
+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_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);
+  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);
+  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);
+  grpc_exec_ctx_sched(exec_ctx, pollset->shutdown_done, GRPC_ERROR_NONE, NULL);
+}
+
+static void work_combine_error(grpc_error **composite, grpc_error *error) {
+  if (error == GRPC_ERROR_NONE) return;
+  if (*composite == GRPC_ERROR_NONE) {
+    *composite = GRPC_ERROR_CREATE("pollset_work");
+  }
+  *composite = grpc_error_add_child(*composite, error);
+}
+
+static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
+                                grpc_pollset_worker **worker_hdl,
+                                gpr_timespec now, gpr_timespec deadline) {
+  grpc_pollset_worker worker;
+  *worker_hdl = &worker;
+  grpc_error *error = GRPC_ERROR_NONE;
+
+  /* pollset->mu already held */
+  int added_worker = 0;
+  int locked = 1;
+  int queued_work = 0;
+  int keep_polling = 0;
+  GPR_TIMER_BEGIN("pollset_work", 0);
+  /* this must happen before we (potentially) drop pollset->mu */
+  worker.next = worker.prev = NULL;
+  worker.reevaluate_polling_on_wakeup = 0;
+  if (pollset->local_wakeup_cache != NULL) {
+    worker.wakeup_fd = pollset->local_wakeup_cache;
+    pollset->local_wakeup_cache = worker.wakeup_fd->next;
+  } else {
+    worker.wakeup_fd = gpr_malloc(sizeof(*worker.wakeup_fd));
+    error = grpc_wakeup_fd_init(&worker.wakeup_fd->fd);
+    if (error != GRPC_ERROR_NONE) {
+      return error;
+    }
+  }
+  worker.kicked_specifically = 0;
+  /* If there's work waiting for the pollset to be idle, and the
+     pollset is idle, then do that work */
+  if (!pollset_has_workers(pollset) &&
+      !grpc_closure_list_empty(pollset->idle_jobs)) {
+    GPR_TIMER_MARK("pollset_work.idle_jobs", 0);
+    grpc_exec_ctx_enqueue_list(exec_ctx, &pollset->idle_jobs, NULL);
+    goto done;
+  }
+  /* If we're shutting down then we don't execute any extended work */
+  if (pollset->shutting_down) {
+    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;
+  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_worker, (intptr_t)&worker);
+      }
+      gpr_tls_set(&g_current_thread_poller, (intptr_t)pollset);
+      GPR_TIMER_BEGIN("maybe_work_and_unlock", 0);
+      work_combine_error(&error,
+                         pollset->vtable->maybe_work_and_unlock(
+                             exec_ctx, pollset, &worker, deadline, now));
+      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;
+    }
+  /* 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 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 || worker.kicked_specifically) {
+        /* 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;
+    }
+  }
+  if (added_worker) {
+    remove_worker(pollset, &worker);
+    gpr_tls_set(&g_current_thread_worker, 0);
+  }
+  /* release wakeup fd to the local pool */
+  worker.wakeup_fd->next = pollset->local_wakeup_cache;
+  pollset->local_wakeup_cache = worker.wakeup_fd;
+  /* check shutdown conditions */
+  if (pollset->shutting_down) {
+    if (pollset_has_workers(pollset)) {
+      pollset_kick(pollset, NULL);
+    } else if (!pollset->called_shutdown && pollset->in_flight_cbs == 0) {
+      pollset->called_shutdown = 1;
+      gpr_mu_unlock(&pollset->mu);
+      finish_shutdown(exec_ctx, pollset);
+      grpc_exec_ctx_flush(exec_ctx);
+      /* Continuing to access pollset here is safe -- it is the caller's
+       * responsibility to not destroy when it has outstanding calls to
+       * pollset_work.
+       * TODO(dklempner): Can we refactor the shutdown logic to avoid this? */
+      gpr_mu_lock(&pollset->mu);
+    } else if (!grpc_closure_list_empty(pollset->idle_jobs)) {
+      grpc_exec_ctx_enqueue_list(exec_ctx, &pollset->idle_jobs, NULL);
+      gpr_mu_unlock(&pollset->mu);
+      grpc_exec_ctx_flush(exec_ctx);
+      gpr_mu_lock(&pollset->mu);
+    }
+  }
+  *worker_hdl = NULL;
+  GPR_TIMER_END("pollset_work", 0);
+  return error;
+}
+
+static void pollset_shutdown(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
+                             grpc_closure *closure) {
+  GPR_ASSERT(!pollset->shutting_down);
+  pollset->shutting_down = 1;
+  pollset->shutdown_done = closure;
+  pollset_kick(pollset, GRPC_POLLSET_KICK_BROADCAST);
+  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)) {
+    pollset->called_shutdown = 1;
+    finish_shutdown(exec_ctx, pollset);
+  }
+}
+
+static int poll_deadline_to_millis_timeout(gpr_timespec deadline,
+                                           gpr_timespec now) {
+  gpr_timespec timeout;
+  static const int64_t max_spin_polling_us = 10;
+  if (gpr_time_cmp(deadline, gpr_inf_future(deadline.clock_type)) == 0) {
+    return -1;
+  }
+  if (gpr_time_cmp(deadline, gpr_time_add(now, gpr_time_from_micros(
+                                                   max_spin_polling_us,
+                                                   GPR_TIMESPAN))) <= 0) {
+    return 0;
+  }
+  timeout = gpr_time_sub(deadline, now);
+  return gpr_time_to_millis(gpr_time_add(
+      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,
+                             grpc_error *error) {
+  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_append(&pollset->idle_jobs, &up_args->promotion_closure,
+                           GRPC_ERROR_NONE);
+  pollset_kick(pollset, GRPC_POLLSET_KICK_BROADCAST);
+
+exit:
+  if (and_unlock_pollset) {
+    gpr_mu_unlock(&pollset->mu);
+  }
+}
+
+static grpc_error *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;
+  grpc_error *error = GRPC_ERROR_NONE;
+
+  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) {
+      work_combine_error(&error, GRPC_OS_ERROR(errno, "poll"));
+    }
+    if (fd) {
+      fd_end_poll(exec_ctx, &fd_watcher, 0, 0, NULL);
+    }
+  } else if (r == 0) {
+    if (fd) {
+      fd_end_poll(exec_ctx, &fd_watcher, 0, 0, NULL);
+    }
+  } else {
+    if (pfd[0].revents & POLLIN_CHECK) {
+      work_combine_error(&error,
+                         grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd));
+    }
+    if (pfd[1].revents & POLLIN_CHECK) {
+      work_combine_error(&error,
+                         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, pollset);
+    } else if (fd) {
+      fd_end_poll(exec_ctx, &fd_watcher, 0, 0, NULL);
+    }
+  }
+
+  if (fd) {
+    GRPC_FD_UNREF(fd, "basicpoll_begin");
+  }
+
+  return error;
+}
+
+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 grpc_error *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;
+  grpc_error *error = GRPC_ERROR_NONE;
+
+  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) {
+      work_combine_error(&error, GRPC_OS_ERROR(errno, "poll"));
+    }
+    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) {
+      work_combine_error(&error,
+                         grpc_wakeup_fd_consume_wakeup(&grpc_global_wakeup_fd));
+    }
+    if (pfds[1].revents & POLLIN_CHECK) {
+      work_combine_error(&error,
+                         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);
+        continue;
+      }
+      fd_end_poll(exec_ctx, &watchers[i], pfds[i].revents & POLLIN_CHECK,
+                  pfds[i].revents & POLLOUT_CHECK, pollset);
+    }
+  }
+
+  gpr_free(pfds);
+  gpr_free(watchers);
+
+  return error;
+}
+
+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,
+                      grpc_pollset *read_notifier_pollset) {
+  /* 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);
+
+  /* A non-NULL read_notifier_pollset means that the fd is readable. */
+  if (read_notifier_pollset != NULL) {
+    /* Note: Since the fd might be a part of multiple pollsets, this might be
+     * called multiple times (for each time the fd becomes readable) and it is
+     * okay to set the fd's read-notifier pollset to anyone of these pollsets */
+    set_read_notifier_pollset_locked(exec_ctx, fd, read_notifier_pollset);
+  }
+
+  gpr_mu_unlock(&fd->mu);
+}
+
+static void fd_become_readable(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
+                               grpc_pollset *notifier_pollset) {
+  set_ready(exec_ctx, fd, &fd->read_closure, notifier_pollset);
+}
+
+static void fd_become_writable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) {
+  set_ready(exec_ctx, fd, &fd->write_closure, NULL);
+}
+
+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, NULL);
+}
+
+static void perform_delayed_add(grpc_exec_ctx *exec_ctx, void *arg,
+                                grpc_error *error) {
+  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_sched(exec_ctx, da->pollset->shutdown_done, GRPC_ERROR_NONE,
+                          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_sched(exec_ctx, &da->closure, GRPC_ERROR_NONE, NULL);
+  }
+}
+
+/* TODO(klempner): We probably want to turn this down a bit */
+#define GRPC_EPOLL_MAX_EVENTS 1000
+
+static grpc_error *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];
+  grpc_error *error = GRPC_ERROR_NONE;
+
+  /* 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) {
+      work_combine_error(&error, GRPC_OS_ERROR(errno, "poll"));
+    }
+  } else if (poll_rv == 0) {
+    /* do nothing */
+  } else {
+    if (pfds[0].revents) {
+      work_combine_error(&error,
+                         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) {
+            work_combine_error(&error, GRPC_OS_ERROR(errno, "epoll_wait"));
+          }
+        } 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) {
+              work_combine_error(&error, grpc_wakeup_fd_consume_wakeup(
+                                             &grpc_global_wakeup_fd));
+            } else {
+              if (read_ev || cancel) {
+                fd_become_readable(exec_ctx, fd, pollset);
+              }
+              if (write_ev || cancel) {
+                fd_become_writable(exec_ctx, fd);
+              }
+            }
+          }
+        }
+      } while (ep_rv == GRPC_EPOLL_MAX_EVENTS);
+    }
+  }
+  return error;
+}
+
+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
+ */
+
+static grpc_pollset_set *pollset_set_create(void) {
+  grpc_pollset_set *pollset_set = gpr_malloc(sizeof(*pollset_set));
+  memset(pollset_set, 0, sizeof(*pollset_set));
+  gpr_mu_init(&pollset_set->mu);
+  return pollset_set;
+}
+
+static void pollset_set_destroy(grpc_pollset_set *pollset_set) {
+  size_t i;
+  gpr_mu_destroy(&pollset_set->mu);
+  for (i = 0; i < pollset_set->fd_count; i++) {
+    GRPC_FD_UNREF(pollset_set->fds[i], "pollset_set");
+  }
+  gpr_free(pollset_set->pollsets);
+  gpr_free(pollset_set->pollset_sets);
+  gpr_free(pollset_set->fds);
+  gpr_free(pollset_set);
+}
+
+static void pollset_set_add_pollset(grpc_exec_ctx *exec_ctx,
+                                    grpc_pollset_set *pollset_set,
+                                    grpc_pollset *pollset) {
+  size_t i, j;
+  gpr_mu_lock(&pollset_set->mu);
+  if (pollset_set->pollset_count == pollset_set->pollset_capacity) {
+    pollset_set->pollset_capacity =
+        GPR_MAX(8, 2 * pollset_set->pollset_capacity);
+    pollset_set->pollsets =
+        gpr_realloc(pollset_set->pollsets, pollset_set->pollset_capacity *
+                                               sizeof(*pollset_set->pollsets));
+  }
+  pollset_set->pollsets[pollset_set->pollset_count++] = pollset;
+  for (i = 0, j = 0; i < pollset_set->fd_count; i++) {
+    if (fd_is_orphaned(pollset_set->fds[i])) {
+      GRPC_FD_UNREF(pollset_set->fds[i], "pollset_set");
+    } else {
+      pollset_add_fd(exec_ctx, pollset, pollset_set->fds[i]);
+      pollset_set->fds[j++] = pollset_set->fds[i];
+    }
+  }
+  pollset_set->fd_count = j;
+  gpr_mu_unlock(&pollset_set->mu);
+}
+
+static void pollset_set_del_pollset(grpc_exec_ctx *exec_ctx,
+                                    grpc_pollset_set *pollset_set,
+                                    grpc_pollset *pollset) {
+  size_t i;
+  gpr_mu_lock(&pollset_set->mu);
+  for (i = 0; i < pollset_set->pollset_count; i++) {
+    if (pollset_set->pollsets[i] == pollset) {
+      pollset_set->pollset_count--;
+      GPR_SWAP(grpc_pollset *, pollset_set->pollsets[i],
+               pollset_set->pollsets[pollset_set->pollset_count]);
+      break;
+    }
+  }
+  gpr_mu_unlock(&pollset_set->mu);
+}
+
+static void pollset_set_add_pollset_set(grpc_exec_ctx *exec_ctx,
+                                        grpc_pollset_set *bag,
+                                        grpc_pollset_set *item) {
+  size_t i, j;
+  gpr_mu_lock(&bag->mu);
+  if (bag->pollset_set_count == bag->pollset_set_capacity) {
+    bag->pollset_set_capacity = GPR_MAX(8, 2 * bag->pollset_set_capacity);
+    bag->pollset_sets =
+        gpr_realloc(bag->pollset_sets,
+                    bag->pollset_set_capacity * sizeof(*bag->pollset_sets));
+  }
+  bag->pollset_sets[bag->pollset_set_count++] = item;
+  for (i = 0, j = 0; i < bag->fd_count; i++) {
+    if (fd_is_orphaned(bag->fds[i])) {
+      GRPC_FD_UNREF(bag->fds[i], "pollset_set");
+    } else {
+      pollset_set_add_fd(exec_ctx, item, bag->fds[i]);
+      bag->fds[j++] = bag->fds[i];
+    }
+  }
+  bag->fd_count = j;
+  gpr_mu_unlock(&bag->mu);
+}
+
+static void pollset_set_del_pollset_set(grpc_exec_ctx *exec_ctx,
+                                        grpc_pollset_set *bag,
+                                        grpc_pollset_set *item) {
+  size_t i;
+  gpr_mu_lock(&bag->mu);
+  for (i = 0; i < bag->pollset_set_count; i++) {
+    if (bag->pollset_sets[i] == item) {
+      bag->pollset_set_count--;
+      GPR_SWAP(grpc_pollset_set *, bag->pollset_sets[i],
+               bag->pollset_sets[bag->pollset_set_count]);
+      break;
+    }
+  }
+  gpr_mu_unlock(&bag->mu);
+}
+
+static void pollset_set_add_fd(grpc_exec_ctx *exec_ctx,
+                               grpc_pollset_set *pollset_set, grpc_fd *fd) {
+  size_t i;
+  gpr_mu_lock(&pollset_set->mu);
+  if (pollset_set->fd_count == pollset_set->fd_capacity) {
+    pollset_set->fd_capacity = GPR_MAX(8, 2 * pollset_set->fd_capacity);
+    pollset_set->fds = gpr_realloc(
+        pollset_set->fds, pollset_set->fd_capacity * sizeof(*pollset_set->fds));
+  }
+  GRPC_FD_REF(fd, "pollset_set");
+  pollset_set->fds[pollset_set->fd_count++] = fd;
+  for (i = 0; i < pollset_set->pollset_count; i++) {
+    pollset_add_fd(exec_ctx, pollset_set->pollsets[i], fd);
+  }
+  for (i = 0; i < pollset_set->pollset_set_count; i++) {
+    pollset_set_add_fd(exec_ctx, pollset_set->pollset_sets[i], fd);
+  }
+  gpr_mu_unlock(&pollset_set->mu);
+}
+
+static void pollset_set_del_fd(grpc_exec_ctx *exec_ctx,
+                               grpc_pollset_set *pollset_set, grpc_fd *fd) {
+  size_t i;
+  gpr_mu_lock(&pollset_set->mu);
+  for (i = 0; i < pollset_set->fd_count; i++) {
+    if (pollset_set->fds[i] == fd) {
+      pollset_set->fd_count--;
+      GPR_SWAP(grpc_fd *, pollset_set->fds[i],
+               pollset_set->fds[pollset_set->fd_count]);
+      GRPC_FD_UNREF(fd, "pollset_set");
+      break;
+    }
+  }
+  for (i = 0; i < pollset_set->pollset_set_count; i++) {
+    pollset_set_del_fd(exec_ctx, pollset_set->pollset_sets[i], fd);
+  }
+  gpr_mu_unlock(&pollset_set->mu);
+}
+
+/*******************************************************************************
+ * event engine binding
+ */
+
+static void shutdown_engine(void) {
+  fd_global_shutdown();
+  pollset_global_shutdown();
+}
+
+static const grpc_event_engine_vtable vtable = {
+    .pollset_size = sizeof(grpc_pollset),
+
+    .fd_create = fd_create,
+    .fd_wrapped_fd = fd_wrapped_fd,
+    .fd_orphan = fd_orphan,
+    .fd_shutdown = fd_shutdown,
+    .fd_is_shutdown = fd_is_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,
+    .pollset_reset = pollset_reset,
+    .pollset_destroy = pollset_destroy,
+    .pollset_work = pollset_work,
+    .pollset_kick = pollset_kick,
+    .pollset_add_fd = pollset_add_fd,
+
+    .pollset_set_create = pollset_set_create,
+    .pollset_set_destroy = pollset_set_destroy,
+    .pollset_set_add_pollset = pollset_set_add_pollset,
+    .pollset_set_del_pollset = pollset_set_del_pollset,
+    .pollset_set_add_pollset_set = pollset_set_add_pollset_set,
+    .pollset_set_del_pollset_set = pollset_set_del_pollset_set,
+    .pollset_set_add_fd = pollset_set_add_fd,
+    .pollset_set_del_fd = pollset_set_del_fd,
+
+    .kick_poller = kick_poller,
+
+    .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();
+  pollset_global_init();
+  return &vtable;
+}
+
+#endif
diff --git a/src/core/lib/iomgr/ev_poll_and_epoll_posix.h b/src/core/lib/iomgr/ev_poll_and_epoll_posix.h
new file mode 100644
index 0000000000..06d6dbf29d
--- /dev/null
+++ b/src/core/lib/iomgr/ev_poll_and_epoll_posix.h
@@ -0,0 +1,41 @@
+/*
+ *
+ * Copyright 2015, Google Inc.
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above
+ * copyright notice, this list of conditions and the following disclaimer
+ * in the documentation and/or other materials provided with the
+ * distribution.
+ *     * Neither the name of Google Inc. nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ */
+
+#ifndef GRPC_CORE_LIB_IOMGR_EV_POLL_AND_EPOLL_POSIX_H
+#define GRPC_CORE_LIB_IOMGR_EV_POLL_AND_EPOLL_POSIX_H
+
+#include "src/core/lib/iomgr/ev_posix.h"
+
+const grpc_event_engine_vtable *grpc_init_poll_and_epoll_posix(void);
+
+#endif /* GRPC_CORE_LIB_IOMGR_EV_POLL_AND_EPOLL_POSIX_H */
diff --git a/src/core/lib/iomgr/ev_poll_posix.c b/src/core/lib/iomgr/ev_poll_posix.c
index e5eb16c3be..45c0a5e954 100644
--- a/src/core/lib/iomgr/ev_poll_posix.c
+++ b/src/core/lib/iomgr/ev_poll_posix.c
@@ -59,8 +59,6 @@
  * FD declarations
  */
 
-grpc_wakeup_fd grpc_global_wakeup_fd;
-
 typedef struct grpc_fd_watcher {
   struct grpc_fd_watcher *next;
   struct grpc_fd_watcher *prev;
@@ -432,7 +430,10 @@ static grpc_error *fd_shutdown_error(bool shutdown) {
 
 static void notify_on_locked(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
                              grpc_closure **st, grpc_closure *closure) {
-  if (*st == CLOSURE_NOT_READY) {
+  if (fd->shutdown) {
+    grpc_exec_ctx_sched(exec_ctx, closure, GRPC_ERROR_CREATE("FD shutdown"),
+                        NULL);
+  } else if (*st == CLOSURE_NOT_READY) {
     /* not ready ==> switch to a waiting state by setting the closure */
     *st = closure;
   } else if (*st == CLOSURE_READY) {
@@ -475,13 +476,24 @@ static void set_read_notifier_pollset_locked(
 
 static void fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd) {
   gpr_mu_lock(&fd->mu);
-  GPR_ASSERT(!fd->shutdown);
-  fd->shutdown = 1;
-  set_ready_locked(exec_ctx, fd, &fd->read_closure);
-  set_ready_locked(exec_ctx, fd, &fd->write_closure);
+  /* only shutdown once */
+  if (!fd->shutdown) {
+    fd->shutdown = 1;
+    /* signal read/write closed to OS so that future operations fail */
+    shutdown(fd->fd, SHUT_RDWR);
+    set_ready_locked(exec_ctx, fd, &fd->read_closure);
+    set_ready_locked(exec_ctx, fd, &fd->write_closure);
+  }
   gpr_mu_unlock(&fd->mu);
 }
 
+static bool fd_is_shutdown(grpc_fd *fd) {
+  gpr_mu_lock(&fd->mu);
+  bool r = fd->shutdown;
+  gpr_mu_unlock(&fd->mu);
+  return r;
+}
+
 static void fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
                               grpc_closure *closure) {
   gpr_mu_lock(&fd->mu);
@@ -717,6 +729,7 @@ static grpc_error *pollset_kick_ext(grpc_pollset *p,
   }
 
   GPR_TIMER_END("pollset_kick_ext", 0);
+  GRPC_LOG_IF_ERROR("pollset_kick_ext", GRPC_ERROR_REF(error));
   return error;
 }
 
@@ -845,6 +858,7 @@ static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
     worker.wakeup_fd = gpr_malloc(sizeof(*worker.wakeup_fd));
     error = grpc_wakeup_fd_init(&worker.wakeup_fd->fd);
     if (error != GRPC_ERROR_NONE) {
+      GRPC_LOG_IF_ERROR("pollset_work", GRPC_ERROR_REF(error));
       return error;
     }
   }
@@ -1024,6 +1038,7 @@ static grpc_error *pollset_work(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
   }
   *worker_hdl = NULL;
   GPR_TIMER_END("pollset_work", 0);
+  GRPC_LOG_IF_ERROR("pollset_work", GRPC_ERROR_REF(error));
   return error;
 }
 
@@ -1215,6 +1230,7 @@ static const grpc_event_engine_vtable vtable = {
     .fd_wrapped_fd = fd_wrapped_fd,
     .fd_orphan = fd_orphan,
     .fd_shutdown = fd_shutdown,
+    .fd_is_shutdown = fd_is_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,
@@ -1242,7 +1258,9 @@ static const grpc_event_engine_vtable vtable = {
 };
 
 const grpc_event_engine_vtable *grpc_init_poll_posix(void) {
-  pollset_global_init();
+  if (!GRPC_LOG_IF_ERROR("pollset_global_init", pollset_global_init())) {
+    return NULL;
+  }
   return &vtable;
 }
 
diff --git a/src/core/lib/iomgr/ev_posix.c b/src/core/lib/iomgr/ev_posix.c
index ad8eaab454..a3c1e9db9a 100644
--- a/src/core/lib/iomgr/ev_posix.c
+++ b/src/core/lib/iomgr/ev_posix.c
@@ -44,6 +44,8 @@
 #include <grpc/support/string_util.h>
 #include <grpc/support/useful.h>
 
+#include "src/core/lib/iomgr/ev_epoll_linux.h"
+#include "src/core/lib/iomgr/ev_poll_and_epoll_posix.h"
 #include "src/core/lib/iomgr/ev_poll_posix.h"
 #include "src/core/lib/support/env.h"
 
@@ -52,6 +54,7 @@
 grpc_poll_function_type grpc_poll_function = poll;
 
 static const grpc_event_engine_vtable *g_event_engine;
+static const char *g_poll_strategy_name = NULL;
 
 typedef const grpc_event_engine_vtable *(*event_engine_factory_fn)(void);
 
@@ -61,7 +64,9 @@ typedef struct {
 } event_engine_factory;
 
 static const event_engine_factory g_factories[] = {
+    {"epoll", grpc_init_epoll_linux},
     {"poll", grpc_init_poll_posix},
+    {"legacy", grpc_init_poll_and_epoll_posix},
 };
 
 static void add(const char *beg, const char *end, char ***ss, size_t *ns) {
@@ -97,6 +102,7 @@ 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())) {
+        g_poll_strategy_name = g_factories[i].name;
         gpr_log(GPR_DEBUG, "Using polling engine: %s", g_factories[i].name);
         return;
       }
@@ -104,6 +110,9 @@ static void try_engine(const char *engine) {
   }
 }
 
+/* Call this only after calling grpc_event_engine_init() */
+const char *grpc_get_poll_strategy_name() { return g_poll_strategy_name; }
+
 void grpc_event_engine_init(void) {
   char *s = gpr_getenv("GRPC_POLL_STRATEGY");
   if (s == NULL) {
@@ -152,6 +161,10 @@ void grpc_fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd) {
   g_event_engine->fd_shutdown(exec_ctx, fd);
 }
 
+bool grpc_fd_is_shutdown(grpc_fd *fd) {
+  return g_event_engine->fd_is_shutdown(fd);
+}
+
 void grpc_fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
                             grpc_closure *closure) {
   g_event_engine->fd_notify_on_read(exec_ctx, fd, closure);
@@ -162,11 +175,6 @@ 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 3058f22c91..579c84ef70 100644
--- a/src/core/lib/iomgr/ev_posix.h
+++ b/src/core/lib/iomgr/ev_posix.h
@@ -55,6 +55,7 @@ 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);
+  bool (*fd_is_shutdown)(grpc_fd *fd);
   grpc_pollset *(*fd_get_read_notifier_pollset)(grpc_exec_ctx *exec_ctx,
                                                 grpc_fd *fd);
 
@@ -98,6 +99,9 @@ typedef struct grpc_event_engine_vtable {
 void grpc_event_engine_init(void);
 void grpc_event_engine_shutdown(void);
 
+/* Return the name of the poll strategy */
+const char *grpc_get_poll_strategy_name();
+
 /* Create a wrapped file descriptor.
    Requires fd is a non-blocking file descriptor.
    This takes ownership of closing fd. */
@@ -116,7 +120,10 @@ int grpc_fd_wrapped_fd(grpc_fd *fd);
 void grpc_fd_orphan(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_closure *on_done,
                     int *release_fd, const char *reason);
 
-/* Cause any current callbacks to error out with GRPC_CALLBACK_CANCELLED. */
+/* Has grpc_fd_shutdown been called on an fd? */
+bool grpc_fd_is_shutdown(grpc_fd *fd);
+
+/* Cause any current and future callbacks to fail. */
 void grpc_fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd);
 
 /* Register read interest, causing read_cb to be called once when fd becomes
diff --git a/src/core/lib/iomgr/iocp_windows.c b/src/core/lib/iomgr/iocp_windows.c
index 1faa8eeedf..2532e52e48 100644
--- a/src/core/lib/iomgr/iocp_windows.c
+++ b/src/core/lib/iomgr/iocp_windows.c
@@ -39,7 +39,7 @@
 
 #include <grpc/support/alloc.h>
 #include <grpc/support/log.h>
-#include <grpc/support/log_win32.h>
+#include <grpc/support/log_windows.h>
 #include <grpc/support/thd.h>
 
 #include "src/core/lib/iomgr/iocp_windows.h"
diff --git a/src/core/lib/iomgr/iomgr.c b/src/core/lib/iomgr/iomgr.c
index 60cef8ba77..89292a153e 100644
--- a/src/core/lib/iomgr/iomgr.c
+++ b/src/core/lib/iomgr/iomgr.c
@@ -96,7 +96,8 @@ void grpc_iomgr_shutdown(void) {
             gpr_time_sub(gpr_now(GPR_CLOCK_REALTIME), last_warning_time),
             gpr_time_from_seconds(1, GPR_TIMESPAN)) >= 0) {
       if (g_root_object.next != &g_root_object) {
-        gpr_log(GPR_DEBUG, "Waiting for %d iomgr objects to be destroyed",
+        gpr_log(GPR_DEBUG,
+                "Waiting for %" PRIuPTR " iomgr objects to be destroyed",
                 count_objects());
       }
       last_warning_time = gpr_now(GPR_CLOCK_REALTIME);
@@ -114,9 +115,9 @@ void grpc_iomgr_shutdown(void) {
       if (gpr_cv_wait(&g_rcv, &g_mu, short_deadline)) {
         if (gpr_time_cmp(gpr_now(GPR_CLOCK_REALTIME), shutdown_deadline) > 0) {
           if (g_root_object.next != &g_root_object) {
-            gpr_log(GPR_DEBUG,
-                    "Failed to free %d iomgr objects before shutdown deadline: "
-                    "memory leaks are likely",
+            gpr_log(GPR_DEBUG, "Failed to free %" PRIuPTR
+                               " iomgr objects before shutdown deadline: "
+                               "memory leaks are likely",
                     count_objects());
             dump_objects("LEAKED");
             if (grpc_iomgr_abort_on_leaks()) {
diff --git a/src/core/lib/iomgr/iomgr_windows.c b/src/core/lib/iomgr/iomgr_windows.c
index 398517fc75..7653f6e635 100644
--- a/src/core/lib/iomgr/iomgr_windows.c
+++ b/src/core/lib/iomgr/iomgr_windows.c
@@ -35,7 +35,7 @@
 
 #ifdef GPR_WINSOCK_SOCKET
 
-#include "src/core/lib/iomgr/sockaddr_win32.h"
+#include "src/core/lib/iomgr/sockaddr_windows.h"
 
 #include <grpc/support/log.h>
 
diff --git a/src/core/lib/iomgr/network_status_tracker.c b/src/core/lib/iomgr/network_status_tracker.c
new file mode 100644
index 0000000000..38a1c9b7d4
--- /dev/null
+++ b/src/core/lib/iomgr/network_status_tracker.c
@@ -0,0 +1,121 @@
+/*
+ *
+ * Copyright 2015, Google Inc.
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above
+ * copyright notice, this list of conditions and the following disclaimer
+ * in the documentation and/or other materials provided with the
+ * distribution.
+ *     * Neither the name of Google Inc. nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ */
+
+#include <grpc/support/alloc.h>
+#include <grpc/support/log.h>
+#include "src/core/lib/iomgr/endpoint.h"
+
+typedef struct endpoint_ll_node {
+  grpc_endpoint *ep;
+  struct endpoint_ll_node *next;
+} endpoint_ll_node;
+
+static endpoint_ll_node *head = NULL;
+static gpr_mu g_endpoint_mutex;
+static bool g_init_done = false;
+
+void grpc_initialize_network_status_monitor() {
+  g_init_done = true;
+  gpr_mu_init(&g_endpoint_mutex);
+  // TODO(makarandd): Install callback with OS to monitor network status.
+}
+
+void grpc_destroy_network_status_monitor() {
+  for (endpoint_ll_node *curr = head; curr != NULL;) {
+    endpoint_ll_node *next = curr->next;
+    gpr_free(curr);
+    curr = next;
+  }
+  gpr_mu_destroy(&g_endpoint_mutex);
+}
+
+void grpc_network_status_register_endpoint(grpc_endpoint *ep) {
+  if (!g_init_done) {
+    grpc_initialize_network_status_monitor();
+  }
+  gpr_mu_lock(&g_endpoint_mutex);
+  if (head == NULL) {
+    head = (endpoint_ll_node *)gpr_malloc(sizeof(endpoint_ll_node));
+    head->ep = ep;
+    head->next = NULL;
+  } else {
+    endpoint_ll_node *prev_head = head;
+    head = (endpoint_ll_node *)gpr_malloc(sizeof(endpoint_ll_node));
+    head->ep = ep;
+    head->next = prev_head;
+  }
+  gpr_mu_unlock(&g_endpoint_mutex);
+}
+
+void grpc_network_status_unregister_endpoint(grpc_endpoint *ep) {
+  gpr_mu_lock(&g_endpoint_mutex);
+  GPR_ASSERT(head);
+  bool found = false;
+  endpoint_ll_node *prev = head;
+  // if we're unregistering the head, just move head to the next
+  if (ep == head->ep) {
+    head = head->next;
+    gpr_free(prev);
+    found = true;
+  } else {
+    for (endpoint_ll_node *curr = head->next; curr != NULL; curr = curr->next) {
+      if (ep == curr->ep) {
+        prev->next = curr->next;
+        gpr_free(curr);
+        found = true;
+        break;
+      }
+      prev = curr;
+    }
+  }
+  gpr_mu_unlock(&g_endpoint_mutex);
+  GPR_ASSERT(found);
+}
+
+// Walk the linked-list from head and execute shutdown. It is possible that
+// other threads might be in the process of shutdown as well, but that has
+// no side effect since endpoint shutdown is idempotent.
+void grpc_network_status_shutdown_all_endpoints() {
+  gpr_mu_lock(&g_endpoint_mutex);
+  if (head == NULL) {
+    gpr_mu_unlock(&g_endpoint_mutex);
+    return;
+  }
+  grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
+
+  for (endpoint_ll_node *curr = head; curr != NULL; curr = curr->next) {
+    curr->ep->vtable->shutdown(&exec_ctx, curr->ep);
+  }
+  gpr_mu_unlock(&g_endpoint_mutex);
+  grpc_exec_ctx_finish(&exec_ctx);
+}
diff --git a/src/core/lib/iomgr/network_status_tracker.h b/src/core/lib/iomgr/network_status_tracker.h
new file mode 100644
index 0000000000..74a1aa8135
--- /dev/null
+++ b/src/core/lib/iomgr/network_status_tracker.h
@@ -0,0 +1,41 @@
+/*
+ *
+ * Copyright 2016, Google Inc.
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above
+ * copyright notice, this list of conditions and the following disclaimer
+ * in the documentation and/or other materials provided with the
+ * distribution.
+ *     * Neither the name of Google Inc. nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ */
+
+#ifndef GRPC_CORE_LIB_IOMGR_NETWORK_STATUS_TRACKER_H
+#define GRPC_CORE_LIB_IOMGR_NETWORK_STATUS_TRACKER_H
+#include "src/core/lib/iomgr/endpoint.h"
+
+void grpc_network_status_register_endpoint(grpc_endpoint *ep);
+void grpc_network_status_unregister_endpoint(grpc_endpoint *ep);
+void grpc_network_status_shutdown_all_endpoints();
+#endif /* GRPC_CORE_LIB_IOMGR_NETWORK_STATUS_TRACKER_H */
diff --git a/src/core/lib/iomgr/polling_entity.c b/src/core/lib/iomgr/polling_entity.c
new file mode 100644
index 0000000000..d1686aa12f
--- /dev/null
+++ b/src/core/lib/iomgr/polling_entity.c
@@ -0,0 +1,104 @@
+/*
+ *
+ * Copyright 2016, Google Inc.
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above
+ * copyright notice, this list of conditions and the following disclaimer
+ * in the documentation and/or other materials provided with the
+ * distribution.
+ *     * Neither the name of Google Inc. nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ */
+
+#include <grpc/support/alloc.h>
+#include <grpc/support/log.h>
+
+#include "src/core/lib/iomgr/polling_entity.h"
+
+grpc_polling_entity grpc_polling_entity_create_from_pollset_set(
+    grpc_pollset_set *pollset_set) {
+  grpc_polling_entity pollent;
+  pollent.pollent.pollset_set = pollset_set;
+  pollent.tag = POPS_POLLSET_SET;
+  return pollent;
+}
+
+grpc_polling_entity grpc_polling_entity_create_from_pollset(
+    grpc_pollset *pollset) {
+  grpc_polling_entity pollent;
+  pollent.pollent.pollset = pollset;
+  pollent.tag = POPS_POLLSET;
+  return pollent;
+}
+
+grpc_pollset *grpc_polling_entity_pollset(grpc_polling_entity *pollent) {
+  if (pollent->tag == POPS_POLLSET) {
+    return pollent->pollent.pollset;
+  }
+  return NULL;
+}
+
+grpc_pollset_set *grpc_polling_entity_pollset_set(
+    grpc_polling_entity *pollent) {
+  if (pollent->tag == POPS_POLLSET_SET) {
+    return pollent->pollent.pollset_set;
+  }
+  return NULL;
+}
+
+bool grpc_polling_entity_is_empty(const grpc_polling_entity *pollent) {
+  return pollent->tag == POPS_NONE;
+}
+
+void grpc_polling_entity_add_to_pollset_set(grpc_exec_ctx *exec_ctx,
+                                            grpc_polling_entity *pollent,
+                                            grpc_pollset_set *pss_dst) {
+  if (pollent->tag == POPS_POLLSET) {
+    GPR_ASSERT(pollent->pollent.pollset != NULL);
+    grpc_pollset_set_add_pollset(exec_ctx, pss_dst, pollent->pollent.pollset);
+  } else if (pollent->tag == POPS_POLLSET_SET) {
+    GPR_ASSERT(pollent->pollent.pollset_set != NULL);
+    grpc_pollset_set_add_pollset_set(exec_ctx, pss_dst,
+                                     pollent->pollent.pollset_set);
+  } else {
+    gpr_log(GPR_ERROR, "Invalid grpc_polling_entity tag '%d'", pollent->tag);
+    abort();
+  }
+}
+
+void grpc_polling_entity_del_from_pollset_set(grpc_exec_ctx *exec_ctx,
+                                              grpc_polling_entity *pollent,
+                                              grpc_pollset_set *pss_dst) {
+  if (pollent->tag == POPS_POLLSET) {
+    GPR_ASSERT(pollent->pollent.pollset != NULL);
+    grpc_pollset_set_del_pollset(exec_ctx, pss_dst, pollent->pollent.pollset);
+  } else if (pollent->tag == POPS_POLLSET_SET) {
+    GPR_ASSERT(pollent->pollent.pollset_set != NULL);
+    grpc_pollset_set_del_pollset_set(exec_ctx, pss_dst,
+                                     pollent->pollent.pollset_set);
+  } else {
+    gpr_log(GPR_ERROR, "Invalid grpc_polling_entity tag '%d'", pollent->tag);
+    abort();
+  }
+}
diff --git a/src/core/lib/iomgr/polling_entity.h b/src/core/lib/iomgr/polling_entity.h
new file mode 100644
index 0000000000..e81531053c
--- /dev/null
+++ b/src/core/lib/iomgr/polling_entity.h
@@ -0,0 +1,81 @@
+/*
+ *
+ * Copyright 2015, Google Inc.
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above
+ * copyright notice, this list of conditions and the following disclaimer
+ * in the documentation and/or other materials provided with the
+ * distribution.
+ *     * Neither the name of Google Inc. nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ */
+
+#ifndef GRPC_CORE_LIB_IOMGR_POLLING_ENTITY_H
+#define GRPC_CORE_LIB_IOMGR_POLLING_ENTITY_H
+
+#include "src/core/lib/iomgr/pollset.h"
+#include "src/core/lib/iomgr/pollset_set.h"
+
+/* A grpc_polling_entity is a pollset-or-pollset_set container. It allows
+ * functions that
+ * accept a pollset XOR a pollset_set to do so through an abstract interface.
+ * No ownership is taken. */
+
+typedef struct grpc_polling_entity {
+  union {
+    grpc_pollset *pollset;
+    grpc_pollset_set *pollset_set;
+  } pollent;
+  enum pops_tag { POPS_NONE, POPS_POLLSET, POPS_POLLSET_SET } tag;
+} grpc_polling_entity;
+
+grpc_polling_entity grpc_polling_entity_create_from_pollset_set(
+    grpc_pollset_set *pollset_set);
+grpc_polling_entity grpc_polling_entity_create_from_pollset(
+    grpc_pollset *pollset);
+
+/** If \a pollent contains a pollset, return it. Otherwise, return NULL */
+grpc_pollset *grpc_polling_entity_pollset(grpc_polling_entity *pollent);
+
+/** If \a pollent contains a pollset_set, return it. Otherwise, return NULL */
+grpc_pollset_set *grpc_polling_entity_pollset_set(grpc_polling_entity *pollent);
+
+bool grpc_polling_entity_is_empty(const grpc_polling_entity *pollent);
+
+/** Add the pollset or pollset_set in \a pollent to the destination pollset_set
+ * \a
+ * pss_dst */
+void grpc_polling_entity_add_to_pollset_set(grpc_exec_ctx *exec_ctx,
+                                            grpc_polling_entity *pollent,
+                                            grpc_pollset_set *pss_dst);
+
+/** Delete the pollset or pollset_set in \a pollent from the destination
+ * pollset_set \a
+ * pss_dst */
+void grpc_polling_entity_del_from_pollset_set(grpc_exec_ctx *exec_ctx,
+                                              grpc_polling_entity *pollent,
+                                              grpc_pollset_set *pss_dst);
+/* pollset_set specific */
+
+#endif /* GRPC_CORE_LIB_IOMGR_POLLING_ENTITY_H */
diff --git a/src/core/lib/iomgr/pollset_set_windows.c b/src/core/lib/iomgr/pollset_set_windows.c
index 89f60b92fb..a35a9766fc 100644
--- a/src/core/lib/iomgr/pollset_set_windows.c
+++ b/src/core/lib/iomgr/pollset_set_windows.c
@@ -32,12 +32,15 @@
  */
 
 #include <grpc/support/port_platform.h>
+#include <stdint.h>
 
 #ifdef GPR_WINSOCK_SOCKET
 
 #include "src/core/lib/iomgr/pollset_set_windows.h"
 
-grpc_pollset_set* grpc_pollset_set_create(void) { return NULL; }
+grpc_pollset_set* grpc_pollset_set_create(void) {
+  return (grpc_pollset_set*)((intptr_t)0xdeafbeef);
+}
 
 void grpc_pollset_set_destroy(grpc_pollset_set* pollset_set) {}
 
diff --git a/src/core/lib/iomgr/resolve_address_windows.c b/src/core/lib/iomgr/resolve_address_windows.c
index fac4d13a27..2af8af82dc 100644
--- a/src/core/lib/iomgr/resolve_address_windows.c
+++ b/src/core/lib/iomgr/resolve_address_windows.c
@@ -43,7 +43,7 @@
 #include <grpc/support/alloc.h>
 #include <grpc/support/host_port.h>
 #include <grpc/support/log.h>
-#include <grpc/support/log_win32.h>
+#include <grpc/support/log_windows.h>
 #include <grpc/support/string_util.h>
 #include <grpc/support/thd.h>
 #include <grpc/support/time.h>
diff --git a/src/core/lib/iomgr/sockaddr.h b/src/core/lib/iomgr/sockaddr.h
index 891a2f094f..5563d0b8a6 100644
--- a/src/core/lib/iomgr/sockaddr.h
+++ b/src/core/lib/iomgr/sockaddr.h
@@ -36,8 +36,8 @@
 
 #include <grpc/support/port_platform.h>
 
-#ifdef GPR_WIN32
-#include "src/core/lib/iomgr/sockaddr_win32.h"
+#ifdef GPR_WINDOWS
+#include "src/core/lib/iomgr/sockaddr_windows.h"
 #endif
 
 #ifdef GPR_POSIX_SOCKETADDR
diff --git a/src/core/lib/iomgr/sockaddr_win32.h b/src/core/lib/iomgr/sockaddr_windows.h
index 02aeae7619..971db5b32b 100644
--- a/src/core/lib/iomgr/sockaddr_win32.h
+++ b/src/core/lib/iomgr/sockaddr_windows.h
@@ -31,8 +31,8 @@
  *
  */
 
-#ifndef GRPC_CORE_LIB_IOMGR_SOCKADDR_WIN32_H
-#define GRPC_CORE_LIB_IOMGR_SOCKADDR_WIN32_H
+#ifndef GRPC_CORE_LIB_IOMGR_SOCKADDR_WINDOWS_H
+#define GRPC_CORE_LIB_IOMGR_SOCKADDR_WINDOWS_H
 
 #include <winsock2.h>
 #include <ws2tcpip.h>
@@ -40,4 +40,4 @@
 // must be included after the above
 #include <mswsock.h>
 
-#endif /* GRPC_CORE_LIB_IOMGR_SOCKADDR_WIN32_H */
+#endif /* GRPC_CORE_LIB_IOMGR_SOCKADDR_WINDOWS_H */
diff --git a/src/core/lib/iomgr/socket_utils_common_posix.c b/src/core/lib/iomgr/socket_utils_common_posix.c
index bd8ef00c85..d2f6261e2a 100644
--- a/src/core/lib/iomgr/socket_utils_common_posix.c
+++ b/src/core/lib/iomgr/socket_utils_common_posix.c
@@ -49,6 +49,7 @@
 #include <sys/types.h>
 #include <unistd.h>
 
+#include <grpc/support/alloc.h>
 #include <grpc/support/host_port.h>
 #include <grpc/support/log.h>
 #include <grpc/support/port_platform.h>
@@ -116,6 +117,20 @@ grpc_error *grpc_set_socket_ipv6_recvpktinfo_if_possible(int fd) {
   return GRPC_ERROR_NONE;
 }
 
+grpc_error *grpc_set_socket_sndbuf(int fd, int buffer_size_bytes) {
+  return 0 == setsockopt(fd, SOL_SOCKET, SO_SNDBUF, &buffer_size_bytes,
+                         sizeof(buffer_size_bytes))
+             ? GRPC_ERROR_NONE
+             : GRPC_OS_ERROR(errno, "setsockopt(SO_SNDBUF)");
+}
+
+grpc_error *grpc_set_socket_rcvbuf(int fd, int buffer_size_bytes) {
+  return 0 == setsockopt(fd, SOL_SOCKET, SO_RCVBUF, &buffer_size_bytes,
+                         sizeof(buffer_size_bytes))
+             ? GRPC_ERROR_NONE
+             : GRPC_OS_ERROR(errno, "setsockopt(SO_RCVBUF)");
+}
+
 /* set a socket to close on exec */
 grpc_error *grpc_set_socket_cloexec(int fd, int close_on_exec) {
   int oldflags = fcntl(fd, F_GETFD, 0);
@@ -238,6 +253,16 @@ static int set_socket_dualstack(int fd) {
   }
 }
 
+static grpc_error *error_for_fd(int fd, const struct sockaddr *addr) {
+  if (fd >= 0) return GRPC_ERROR_NONE;
+  char *addr_str;
+  grpc_sockaddr_to_string(&addr_str, addr, 0);
+  grpc_error *err = grpc_error_set_str(GRPC_OS_ERROR(errno, "socket"),
+                                       GRPC_ERROR_STR_TARGET_ADDRESS, addr_str);
+  gpr_free(addr_str);
+  return err;
+}
+
 grpc_error *grpc_create_dualstack_socket(const struct sockaddr *addr, int type,
                                          int protocol,
                                          grpc_dualstack_mode *dsmode,
@@ -258,7 +283,7 @@ grpc_error *grpc_create_dualstack_socket(const struct sockaddr *addr, int type,
     /* If this isn't an IPv4 address, then return whatever we've got. */
     if (!grpc_sockaddr_is_v4mapped(addr, NULL)) {
       *dsmode = GRPC_DSMODE_IPV6;
-      return GRPC_ERROR_NONE;
+      return error_for_fd(*newfd, addr);
     }
     /* Fall back to AF_INET. */
     if (*newfd >= 0) {
@@ -268,10 +293,7 @@ grpc_error *grpc_create_dualstack_socket(const struct sockaddr *addr, int type,
   }
   *dsmode = family == AF_INET ? GRPC_DSMODE_IPV4 : GRPC_DSMODE_NONE;
   *newfd = socket(family, type, protocol);
-  if (*newfd == -1) {
-    return GRPC_OS_ERROR(errno, "socket");
-  }
-  return GRPC_ERROR_NONE;
+  return error_for_fd(*newfd, addr);
 }
 
 #endif
diff --git a/src/core/lib/iomgr/socket_utils_posix.h b/src/core/lib/iomgr/socket_utils_posix.h
index 607fbc7a79..7bcc2219ae 100644
--- a/src/core/lib/iomgr/socket_utils_posix.h
+++ b/src/core/lib/iomgr/socket_utils_posix.h
@@ -69,20 +69,23 @@ grpc_error *grpc_set_socket_reuse_port(int fd, int reuse);
 int grpc_ipv6_loopback_available(void);
 
 /* Tries to set SO_NOSIGPIPE if available on this platform.
-   Returns 1 on success, 0 on failure.
    If SO_NO_SIGPIPE is not available, returns 1. */
 grpc_error *grpc_set_socket_no_sigpipe_if_possible(int fd);
 
 /* Tries to set IP_PKTINFO if available on this platform.
-   Returns 1 on success, 0 on failure.
    If IP_PKTINFO is not available, returns 1. */
 grpc_error *grpc_set_socket_ip_pktinfo_if_possible(int fd);
 
 /* Tries to set IPV6_RECVPKTINFO if available on this platform.
-   Returns 1 on success, 0 on failure.
    If IPV6_RECVPKTINFO is not available, returns 1. */
 grpc_error *grpc_set_socket_ipv6_recvpktinfo_if_possible(int fd);
 
+/* Tries to set the socket's send buffer to given size. */
+grpc_error *grpc_set_socket_sndbuf(int fd, int buffer_size_bytes);
+
+/* Tries to set the socket's receive buffer to given size. */
+grpc_error *grpc_set_socket_rcvbuf(int fd, int buffer_size_bytes);
+
 /* An enum to keep track of IPv4/IPv6 socket modes.
 
    Currently, this information is only used when a socket is first created, but
diff --git a/src/core/lib/iomgr/socket_windows.c b/src/core/lib/iomgr/socket_windows.c
index fbc1bb5301..d7d5f6f157 100644
--- a/src/core/lib/iomgr/socket_windows.c
+++ b/src/core/lib/iomgr/socket_windows.c
@@ -42,7 +42,7 @@
 
 #include <grpc/support/alloc.h>
 #include <grpc/support/log.h>
-#include <grpc/support/log_win32.h>
+#include <grpc/support/log_windows.h>
 #include <grpc/support/string_util.h>
 
 #include "src/core/lib/iomgr/iocp_windows.h"
diff --git a/src/core/lib/iomgr/tcp_client_windows.c b/src/core/lib/iomgr/tcp_client_windows.c
index 6134d4476b..562cb9c6bf 100644
--- a/src/core/lib/iomgr/tcp_client_windows.c
+++ b/src/core/lib/iomgr/tcp_client_windows.c
@@ -35,11 +35,11 @@
 
 #ifdef GPR_WINSOCK_SOCKET
 
-#include "src/core/lib/iomgr/sockaddr_win32.h"
+#include "src/core/lib/iomgr/sockaddr_windows.h"
 
 #include <grpc/support/alloc.h>
 #include <grpc/support/log.h>
-#include <grpc/support/log_win32.h>
+#include <grpc/support/log_windows.h>
 #include <grpc/support/slice_buffer.h>
 #include <grpc/support/useful.h>
 
diff --git a/src/core/lib/iomgr/tcp_posix.c b/src/core/lib/iomgr/tcp_posix.c
index 2f7c50ce35..2ab45e33ce 100644
--- a/src/core/lib/iomgr/tcp_posix.c
+++ b/src/core/lib/iomgr/tcp_posix.c
@@ -35,6 +35,7 @@
 
 #ifdef GPR_POSIX_SOCKET
 
+#include "src/core/lib/iomgr/network_status_tracker.h"
 #include "src/core/lib/iomgr/tcp_posix.h"
 
 #include <errno.h>
@@ -152,6 +153,7 @@ static void tcp_ref(grpc_tcp *tcp) { gpr_ref(&tcp->refcount); }
 #endif
 
 static void tcp_destroy(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep) {
+  grpc_network_status_unregister_endpoint(ep);
   grpc_tcp *tcp = (grpc_tcp *)ep;
   TCP_UNREF(exec_ctx, tcp, "destroy");
 }
@@ -410,7 +412,10 @@ static void tcp_write(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep,
 
   if (buf->length == 0) {
     GPR_TIMER_END("tcp_write", 0);
-    grpc_exec_ctx_sched(exec_ctx, cb, GRPC_ERROR_NONE, NULL);
+    grpc_exec_ctx_sched(exec_ctx, cb, grpc_fd_is_shutdown(tcp->em_fd)
+                                          ? GRPC_ERROR_CREATE("EOF")
+                                          : GRPC_ERROR_NONE,
+                        NULL);
     return;
   }
   tcp->outgoing_buffer = buf;
@@ -471,6 +476,8 @@ grpc_endpoint *grpc_tcp_create(grpc_fd *em_fd, size_t slice_size,
   tcp->write_closure.cb = tcp_handle_write;
   tcp->write_closure.cb_arg = tcp;
   gpr_slice_buffer_init(&tcp->last_read_buffer);
+  /* Tell network status tracker about new endpoint */
+  grpc_network_status_register_endpoint(&tcp->base);
 
   return &tcp->base;
 }
diff --git a/src/core/lib/iomgr/tcp_server_posix.c b/src/core/lib/iomgr/tcp_server_posix.c
index 8c551785c1..eda2585c46 100644
--- a/src/core/lib/iomgr/tcp_server_posix.c
+++ b/src/core/lib/iomgr/tcp_server_posix.c
@@ -595,9 +595,9 @@ grpc_error *grpc_tcp_server_add_port(grpc_tcp_server *s, const void *addr,
       if (port == 0 && sp != NULL) {
         grpc_sockaddr_set_port((struct sockaddr *)&wild4, sp->port);
       }
-      addr = (struct sockaddr *)&wild4;
-      addr_len = sizeof(wild4);
     }
+    addr = (struct sockaddr *)&wild4;
+    addr_len = sizeof(wild4);
   }
 
   errs[1] = grpc_create_dualstack_socket(addr, SOCK_STREAM, 0, &dsmode, &fd);
diff --git a/src/core/lib/iomgr/tcp_server_windows.c b/src/core/lib/iomgr/tcp_server_windows.c
index b4f28a4c3b..7b0966704c 100644
--- a/src/core/lib/iomgr/tcp_server_windows.c
+++ b/src/core/lib/iomgr/tcp_server_windows.c
@@ -41,7 +41,7 @@
 
 #include <grpc/support/alloc.h>
 #include <grpc/support/log.h>
-#include <grpc/support/log_win32.h>
+#include <grpc/support/log_windows.h>
 #include <grpc/support/string_util.h>
 #include <grpc/support/sync.h>
 #include <grpc/support/time.h>
@@ -103,6 +103,7 @@ struct grpc_tcp_server {
 /* Public function. Allocates the proper data structures to hold a
    grpc_tcp_server. */
 grpc_error *grpc_tcp_server_create(grpc_closure *shutdown_complete,
+                                   const grpc_channel_args *args,
                                    grpc_tcp_server **server) {
   grpc_tcp_server *s = gpr_malloc(sizeof(grpc_tcp_server));
   gpr_ref_init(&s->refs, 1);
@@ -396,7 +397,7 @@ static grpc_error *add_socket_to_server(grpc_tcp_server *s, SOCKET sock,
                                         size_t addr_len, unsigned port_index,
                                         grpc_tcp_listener **listener) {
   grpc_tcp_listener *sp = NULL;
-  int port;
+  int port = -1;
   int status;
   GUID guid = WSAID_ACCEPTEX;
   DWORD ioctl_num_bytes;
diff --git a/src/core/lib/iomgr/tcp_windows.c b/src/core/lib/iomgr/tcp_windows.c
index b2003675be..37ab59021e 100644
--- a/src/core/lib/iomgr/tcp_windows.c
+++ b/src/core/lib/iomgr/tcp_windows.c
@@ -37,11 +37,12 @@
 
 #include <limits.h>
 
-#include "src/core/lib/iomgr/sockaddr_win32.h"
+#include "src/core/lib/iomgr/network_status_tracker.h"
+#include "src/core/lib/iomgr/sockaddr_windows.h"
 
 #include <grpc/support/alloc.h>
 #include <grpc/support/log.h>
-#include <grpc/support/log_win32.h>
+#include <grpc/support/log_windows.h>
 #include <grpc/support/slice_buffer.h>
 #include <grpc/support/string_util.h>
 #include <grpc/support/useful.h>
@@ -378,6 +379,7 @@ static void win_shutdown(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep) {
 }
 
 static void win_destroy(grpc_exec_ctx *exec_ctx, grpc_endpoint *ep) {
+  grpc_network_status_unregister_endpoint(ep);
   grpc_tcp *tcp = (grpc_tcp *)ep;
   TCP_UNREF(tcp, "destroy");
 }
@@ -401,6 +403,9 @@ grpc_endpoint *grpc_tcp_create(grpc_winsocket *socket, char *peer_string) {
   grpc_closure_init(&tcp->on_read, on_read, tcp);
   grpc_closure_init(&tcp->on_write, on_write, tcp);
   tcp->peer_string = gpr_strdup(peer_string);
+  /* Tell network status tracking code about the new endpoint */
+  grpc_network_status_register_endpoint(&tcp->base);
+
   return &tcp->base;
 }
 
diff --git a/src/core/lib/iomgr/timer.c b/src/core/lib/iomgr/timer.c
index 7b0ef550a7..9975fa1671 100644
--- a/src/core/lib/iomgr/timer.c
+++ b/src/core/lib/iomgr/timer.c
@@ -287,8 +287,8 @@ static int refill_queue(shard_type *shard, gpr_timespec now) {
   return !grpc_timer_heap_is_empty(&shard->heap);
 }
 
-/* This pops the next non-cancelled timer with deadline <= now from the queue,
-   or returns NULL if there isn't one.
+/* This pops the next non-cancelled timer with deadline <= now from the
+   queue, or returns NULL if there isn't one.
    REQUIRES: shard->mu locked */
 static grpc_timer *pop_one(shard_type *shard, gpr_timespec now) {
   grpc_timer *timer;
diff --git a/src/core/lib/iomgr/udp_server.c b/src/core/lib/iomgr/udp_server.c
index 98ffccd59b..1ebccf2ee2 100644
--- a/src/core/lib/iomgr/udp_server.c
+++ b/src/core/lib/iomgr/udp_server.c
@@ -210,6 +210,8 @@ static int prepare_socket(int fd, const struct sockaddr *addr,
                           size_t addr_len) {
   struct sockaddr_storage sockname_temp;
   socklen_t sockname_len;
+  /* Set send/receive socket buffers to 1 MB */
+  int buffer_size_bytes = 1024 * 1024;
 
   if (fd < 0) {
     goto error;
@@ -239,6 +241,18 @@ static int prepare_socket(int fd, const struct sockaddr *addr,
     goto error;
   }
 
+  if (!grpc_set_socket_sndbuf(fd, buffer_size_bytes)) {
+    gpr_log(GPR_ERROR, "Failed to set send buffer size to %d bytes",
+            buffer_size_bytes);
+    goto error;
+  }
+
+  if (!grpc_set_socket_rcvbuf(fd, buffer_size_bytes)) {
+    gpr_log(GPR_ERROR, "Failed to set receive buffer size to %d bytes",
+            buffer_size_bytes);
+    goto error;
+  }
+
   return grpc_sockaddr_get_port((struct sockaddr *)&sockname_temp);
 
 error:
diff --git a/src/core/lib/iomgr/wakeup_fd_eventfd.c b/src/core/lib/iomgr/wakeup_fd_eventfd.c
index 667b4a5f90..95f6102330 100644
--- a/src/core/lib/iomgr/wakeup_fd_eventfd.c
+++ b/src/core/lib/iomgr/wakeup_fd_eventfd.c
@@ -84,8 +84,10 @@ static void eventfd_destroy(grpc_wakeup_fd* fd_info) {
 }
 
 static int eventfd_check_availability(void) {
-  /* TODO(klempner): Actually check if eventfd is available */
-  return 1;
+  const int efd = eventfd(0, 0);
+  const int is_available = efd >= 0;
+  if (is_available) close(efd);
+  return is_available;
 }
 
 const grpc_wakeup_fd_vtable grpc_specialized_wakeup_fd_vtable = {
diff --git a/src/core/lib/iomgr/workqueue.h b/src/core/lib/iomgr/workqueue.h
index f4a2cfa4fd..ac765375eb 100644
--- a/src/core/lib/iomgr/workqueue.h
+++ b/src/core/lib/iomgr/workqueue.h
@@ -44,7 +44,7 @@
 #include "src/core/lib/iomgr/workqueue_posix.h"
 #endif
 
-#ifdef GPR_WIN32
+#ifdef GPR_WINDOWS
 #include "src/core/lib/iomgr/workqueue_windows.h"
 #endif
 
diff --git a/src/core/lib/iomgr/workqueue_windows.c b/src/core/lib/iomgr/workqueue_windows.c
index c3c0446a57..275f040b1c 100644
--- a/src/core/lib/iomgr/workqueue_windows.c
+++ b/src/core/lib/iomgr/workqueue_windows.c
@@ -33,8 +33,8 @@
 
 #include <grpc/support/port_platform.h>
 
-#ifdef GPR_WIN32
+#ifdef GPR_WINDOWS
 
 #include "src/core/lib/iomgr/workqueue.h"
 
-#endif /* GPR_WIN32 */
+#endif /* GPR_WINDOWS */