New epoll-based polling engine with dedicated poller threads

4 files changed, 1544 insertions, 0 deletions

diff --git a/src/core/lib/iomgr/ev_epoll_thread_pool_linux.c b/src/core/lib/iomgr/ev_epoll_thread_pool_linux.c
new file mode 100644
index 0000000000..7a9e20b1ea
--- /dev/null
+++ b/src/core/lib/iomgr/ev_epoll_thread_pool_linux.c
@@ -0,0 +1,1499 @@
+/*
+ *
+ * Copyright 2017, 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 "src/core/lib/iomgr/port.h"
+
+/* This polling engine is only relevant on linux kernels supporting epoll() */
+#ifdef GRPC_LINUX_EPOLL
+
+#include "src/core/lib/iomgr/ev_epoll_thread_pool_linux.h"
+
+#include <assert.h>
+#include <errno.h>
+#include <limits.h>
+#include <poll.h>
+#include <pthread.h>
+#include <string.h>
+#include <sys/epoll.h>
+#include <sys/socket.h>
+#include <unistd.h>
+
+#include <grpc/support/alloc.h>
+#include <grpc/support/cpu.h>
+#include <grpc/support/log.h>
+#include <grpc/support/string_util.h>
+#include <grpc/support/thd.h>
+#include <grpc/support/tls.h>
+#include <grpc/support/useful.h>
+
+#include "src/core/lib/iomgr/ev_posix.h"
+#include "src/core/lib/iomgr/iomgr_internal.h"
+#include "src/core/lib/iomgr/lockfree_event.h"
+#include "src/core/lib/iomgr/timer.h"
+#include "src/core/lib/iomgr/wakeup_fd_posix.h"
+#include "src/core/lib/iomgr/workqueue.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__); \
+  }
+
+/* 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) */
+static grpc_wakeup_fd global_wakeup_fd;
+
+struct polling_island;
+
+/*******************************************************************************
+ * Fd Declarations
+ */
+struct grpc_fd {
+  gpr_mu mu;
+  struct polling_island *pi;
+
+  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;
+
+  /* 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;
+
+  gpr_atm read_closure;
+  gpr_atm write_closure;
+
+  struct grpc_fd *freelist_next;
+  grpc_closure *on_done_closure;
+
+  /* The pollset that last noticed that the fd is readable. The actual type
+   * stored in this is (grpc_pollset *) */
+  gpr_atm 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);
+
+/*******************************************************************************
+ * Polling island Declarations
+ */
+
+#ifdef GRPC_WORKQUEUE_REFCOUNT_DEBUG
+
+#define PI_ADD_REF(p, r) pi_add_ref_dbg((p), (r), __FILE__, __LINE__)
+#define PI_UNREF(exec_ctx, p, r) \
+  pi_unref_dbg((exec_ctx), (p), (r), __FILE__, __LINE__)
+
+#else /* defined(GRPC_WORKQUEUE_REFCOUNT_DEBUG) */
+
+#define PI_ADD_REF(p, r) pi_add_ref((p))
+#define PI_UNREF(exec_ctx, p, r) pi_unref((exec_ctx), (p))
+
+#endif /* !defined(GRPC_PI_REF_COUNT_DEBUG) */
+
+/* This is also used as grpc_workqueue (by directly casting it) */
+typedef struct polling_island {
+  grpc_closure_scheduler workqueue_scheduler;
+
+  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_atm ref_count;
+
+  /* Number of threads currently polling on this island */
+  gpr_atm poller_count;
+  /* Mutex guarding the read end of the workqueue (must be held to pop from
+   * workqueue_items) */
+  gpr_mu workqueue_read_mu;
+  /* Queue of closures to be executed */
+  gpr_mpscq workqueue_items;
+  /* Count of items in workqueue_items */
+  gpr_atm workqueue_item_count;
+  /* Wakeup fd used to wake pollers to check the contents of workqueue_items */
+  grpc_wakeup_fd workqueue_wakeup_fd;
+
+  /* The fd of the underlying epoll set */
+  int epoll_fd;
+
+  /* The file descriptors in the epoll set */
+  /* TODO: sreek - We no longer need this (and since no other structure in this
+   * polling engine keeps a reference to grpc_fd, we actually no longer need a
+   * ref count field in FD. Just a flag to say wheter it is orphaned or not */
+  size_t fd_cnt;
+  size_t fd_capacity;
+  grpc_fd **fds;
+} polling_island;
+
+/*******************************************************************************
+ * Pollset Declarations
+ */
+struct grpc_pollset_worker {
+  gpr_cv kick_cv;
+
+  struct grpc_pollset_worker *next;
+  struct grpc_pollset_worker *prev;
+};
+
+struct grpc_pollset {
+  gpr_mu mu;
+  struct polling_island *pi;
+
+  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 */
+  gpr_atm is_shutdown;
+};
+
+/*******************************************************************************
+ * Pollset-set Declarations
+ */
+struct grpc_pollset_set {
+  void *no_op;
+};
+
+/*****************************************************************************
+ * Dedicated polling threads and pollsets - Declarations
+ */
+
+size_t g_num_pollsets = 0;
+struct grpc_pollset *g_pollsets = NULL;
+gpr_thd_id *g_poller_threads = NULL;
+
+static void add_fd_to_global_pollset(grpc_fd *fd);
+static void init_dedicated_pollsets();
+static void poller_thread_loop(void *arg);
+static void start_dedicated_poller_threads();
+static void shutdown_dedicated_poller_threads();
+
+/*******************************************************************************
+ * Common helpers
+ */
+
+static bool append_error(grpc_error **composite, grpc_error *error,
+                         const char *desc) {
+  if (error == GRPC_ERROR_NONE) return true;
+  if (*composite == GRPC_ERROR_NONE) {
+    *composite = GRPC_ERROR_CREATE_FROM_COPIED_STRING(desc);
+  }
+  *composite = grpc_error_add_child(*composite, error);
+  return false;
+}
+
+/*******************************************************************************
+ * 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;
+
+/* The polling island being polled right now.
+   See comments in workqueue_maybe_wakeup for why this is tracked. */
+static __thread polling_island *g_current_thread_polling_island;
+
+/* Forward declaration */
+static void polling_island_delete(grpc_exec_ctx *exec_ctx, polling_island *pi);
+static void workqueue_enqueue(grpc_exec_ctx *exec_ctx, grpc_closure *closure,
+                              grpc_error *error);
+
+#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) */
+
+static const grpc_closure_scheduler_vtable workqueue_scheduler_vtable = {
+    workqueue_enqueue, workqueue_enqueue, "workqueue"};
+
+static void pi_add_ref(polling_island *pi);
+static void pi_unref(grpc_exec_ctx *exec_ctx, polling_island *pi);
+
+#ifdef GRPC_WORKQUEUE_REFCOUNT_DEBUG
+static void pi_add_ref_dbg(polling_island *pi, const char *reason,
+                           const char *file, int line) {
+  long old_cnt = gpr_atm_acq_load(&pi->ref_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);
+}
+
+static void pi_unref_dbg(grpc_exec_ctx *exec_ctx, polling_island *pi,
+                         const char *reason, const char *file, int line) {
+  long old_cnt = gpr_atm_acq_load(&pi->ref_count);
+  pi_unref(exec_ctx, 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);
+}
+
+static grpc_workqueue *workqueue_ref(grpc_workqueue *workqueue,
+                                     const char *file, int line,
+                                     const char *reason) {
+  if (workqueue != NULL) {
+    pi_add_ref_dbg((polling_island *)workqueue, reason, file, line);
+  }
+  return workqueue;
+}
+
+static void workqueue_unref(grpc_exec_ctx *exec_ctx, grpc_workqueue *workqueue,
+                            const char *file, int line, const char *reason) {
+  if (workqueue != NULL) {
+    pi_unref_dbg(exec_ctx, (polling_island *)workqueue, reason, file, line);
+  }
+}
+#else
+static grpc_workqueue *workqueue_ref(grpc_workqueue *workqueue) {
+  if (workqueue != NULL) {
+    pi_add_ref((polling_island *)workqueue);
+  }
+  return workqueue;
+}
+
+static void workqueue_unref(grpc_exec_ctx *exec_ctx,
+                            grpc_workqueue *workqueue) {
+  if (workqueue != NULL) {
+    pi_unref(exec_ctx, (polling_island *)workqueue);
+  }
+}
+#endif
+
+static void pi_add_ref(polling_island *pi) {
+  gpr_atm_no_barrier_fetch_add(&pi->ref_count, 1);
+}
+
+static void pi_unref(grpc_exec_ctx *exec_ctx, polling_island *pi) {
+  /* If ref count went to zero, delete the polling island. This deletion is
+     not done under a lock since 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).*/
+  if (1 == gpr_atm_full_fetch_add(&pi->ref_count, -1)) {
+    polling_island_delete(exec_ctx, pi);
+  }
+}
+
+/* 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(&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_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_exec_ctx *exec_ctx,
+                                             grpc_fd *initial_fd,
+                                             grpc_error **error) {
+  polling_island *pi = NULL;
+  const char *err_desc = "polling_island_create";
+
+  *error = GRPC_ERROR_NONE;
+
+  pi = gpr_malloc(sizeof(*pi));
+  pi->workqueue_scheduler.vtable = &workqueue_scheduler_vtable;
+  gpr_mu_init(&pi->mu);
+  pi->fd_cnt = 0;
+  pi->fd_capacity = 0;
+  pi->fds = NULL;
+  pi->epoll_fd = -1;
+
+  gpr_mu_init(&pi->workqueue_read_mu);
+  gpr_mpscq_init(&pi->workqueue_items);
+  gpr_atm_rel_store(&pi->workqueue_item_count, 0);
+
+  gpr_atm_rel_store(&pi->ref_count, 0);
+  gpr_atm_rel_store(&pi->poller_count, 0);
+
+  if (!append_error(error, grpc_wakeup_fd_init(&pi->workqueue_wakeup_fd),
+                    err_desc)) {
+    goto done;
+  }
+
+  pi->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
+
+  if (pi->epoll_fd < 0) {
+    append_error(error, GRPC_OS_ERROR(errno, "epoll_create1"), err_desc);
+    goto done;
+  }
+
+  polling_island_add_wakeup_fd_locked(pi, &global_wakeup_fd, error);
+  polling_island_add_wakeup_fd_locked(pi, &pi->workqueue_wakeup_fd, error);
+
+  if (initial_fd != NULL) {
+    polling_island_add_fds_locked(pi, &initial_fd, 1, true, error);
+  }
+
+done:
+  if (*error != GRPC_ERROR_NONE) {
+    polling_island_delete(exec_ctx, pi);
+    pi = NULL;
+  }
+  return pi;
+}
+
+static void polling_island_delete(grpc_exec_ctx *exec_ctx, polling_island *pi) {
+  GPR_ASSERT(pi->fd_cnt == 0);
+
+  if (pi->epoll_fd >= 0) {
+    close(pi->epoll_fd);
+  }
+  GPR_ASSERT(gpr_atm_no_barrier_load(&pi->workqueue_item_count) == 0);
+  gpr_mu_destroy(&pi->workqueue_read_mu);
+  gpr_mpscq_destroy(&pi->workqueue_items);
+  gpr_mu_destroy(&pi->mu);
+  grpc_wakeup_fd_destroy(&pi->workqueue_wakeup_fd);
+
+  gpr_free(pi->fds);
+  gpr_free(pi);
+}
+
+static void workqueue_maybe_wakeup(polling_island *pi) {
+  /* If this thread is the current poller, then it may be that it's about to
+     decrement the current poller count, so we need to look past this thread */
+  bool is_current_poller = (g_current_thread_polling_island == pi);
+  gpr_atm min_current_pollers_for_wakeup = is_current_poller ? 1 : 0;
+  gpr_atm current_pollers = gpr_atm_no_barrier_load(&pi->poller_count);
+  /* Only issue a wakeup if it's likely that some poller could come in and take
+     it right now. Note that since we do an anticipatory mpscq_pop every poll
+     loop, it's ok if we miss the wakeup here, as we'll get the work item when
+     the next poller enters anyway. */
+  if (current_pollers > min_current_pollers_for_wakeup) {
+    GRPC_LOG_IF_ERROR("workqueue_wakeup_fd",
+                      grpc_wakeup_fd_wakeup(&pi->workqueue_wakeup_fd));
+  }
+}
+
+static void workqueue_enqueue(grpc_exec_ctx *exec_ctx, grpc_closure *closure,
+                              grpc_error *error) {
+  GPR_TIMER_BEGIN("workqueue.enqueue", 0);
+  grpc_workqueue *workqueue = (grpc_workqueue *)closure->scheduler;
+  /* take a ref to the workqueue: otherwise it can happen that whatever events
+   * this kicks off ends up destroying the workqueue before this function
+   * completes */
+  GRPC_WORKQUEUE_REF(workqueue, "enqueue");
+  polling_island *pi = (polling_island *)workqueue;
+  gpr_atm last = gpr_atm_no_barrier_fetch_add(&pi->workqueue_item_count, 1);
+  closure->error_data.error = error;
+  gpr_mpscq_push(&pi->workqueue_items, &closure->next_data.atm_next);
+  if (last == 0) {
+    workqueue_maybe_wakeup(pi);
+  }
+
+  GRPC_WORKQUEUE_UNREF(exec_ctx, workqueue, "enqueue");
+  GPR_TIMER_END("workqueue.enqueue", 0);
+}
+
+static grpc_closure_scheduler *workqueue_scheduler(grpc_workqueue *workqueue) {
+  polling_island *pi = (polling_island *)workqueue;
+  return workqueue == NULL ? grpc_schedule_on_exec_ctx
+                           : &pi->workqueue_scheduler;
+}
+
+static grpc_error *polling_island_global_init() {
+  grpc_error *error = GRPC_ERROR_NONE;
+
+  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() {
+  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. */
+
+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);
+
+    grpc_lfev_destroy(&fd->read_closure);
+    grpc_lfev_destroy(&fd->write_closure);
+
+    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);
+  }
+
+  /* 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);
+  new_fd->pi = NULL;
+
+  gpr_atm_rel_store(&new_fd->refst, (gpr_atm)1);
+  new_fd->fd = fd;
+  new_fd->orphaned = false;
+  grpc_lfev_init(&new_fd->read_closure);
+  grpc_lfev_init(&new_fd->write_closure);
+  gpr_atm_no_barrier_store(&new_fd->read_notifier_pollset, (gpr_atm)NULL);
+
+  new_fd->freelist_next = NULL;
+  new_fd->on_done_closure = 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);
+  add_fd_to_global_pollset(new_fd);
+  return new_fd;
+}
+
+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;
+  polling_island *unref_pi = NULL;
+
+  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->pi). 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->pi to NULL (but remove the ref on the polling island
+       before doing this.) */
+  if (fd->pi != NULL) {
+    polling_island *pi = fd->pi;
+    gpr_mu_lock(&pi->mu);
+    polling_island_remove_fd_locked(pi, fd, is_fd_closed, &error);
+    gpr_mu_unlock(&pi->mu);
+
+    unref_pi = fd->pi;
+    fd->pi = NULL;
+  }
+
+  grpc_closure_sched(exec_ctx, fd->on_done_closure, GRPC_ERROR_REF(error));
+
+  gpr_mu_unlock(&fd->mu);
+  UNREF_BY(fd, 2, reason); /* Drop the reference */
+  if (unref_pi != NULL) {
+    /* Unref stale polling island here, outside the fd lock above.
+       The polling island owns a workqueue which owns an fd, and unreffing
+       inside the lock can cause an eventual lock loop that makes TSAN very
+       unhappy. */
+    PI_UNREF(exec_ctx, unref_pi, "fd_orphan");
+  }
+  GRPC_LOG_IF_ERROR("fd_orphan", GRPC_ERROR_REF(error));
+  GRPC_ERROR_UNREF(error);
+}
+
+static grpc_pollset *fd_get_read_notifier_pollset(grpc_exec_ctx *exec_ctx,
+                                                  grpc_fd *fd) {
+  gpr_atm notifier = gpr_atm_acq_load(&fd->read_notifier_pollset);
+  return (grpc_pollset *)notifier;
+}
+
+static bool fd_is_shutdown(grpc_fd *fd) {
+  return grpc_lfev_is_shutdown(&fd->read_closure);
+}
+
+/* Might be called multiple times */
+static void fd_shutdown(grpc_exec_ctx *exec_ctx, grpc_fd *fd, grpc_error *why) {
+  if (grpc_lfev_set_shutdown(exec_ctx, &fd->read_closure,
+                             GRPC_ERROR_REF(why))) {
+    shutdown(fd->fd, SHUT_RDWR);
+    grpc_lfev_set_shutdown(exec_ctx, &fd->write_closure, GRPC_ERROR_REF(why));
+  }
+  GRPC_ERROR_UNREF(why);
+}
+
+static void fd_notify_on_read(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
+                              grpc_closure *closure) {
+  grpc_lfev_notify_on(exec_ctx, &fd->read_closure, closure);
+}
+
+static void fd_notify_on_write(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
+                               grpc_closure *closure) {
+  grpc_lfev_notify_on(exec_ctx, &fd->write_closure, closure);
+}
+
+static grpc_workqueue *fd_get_workqueue(grpc_fd *fd) {
+  return NULL;
+}
+
+/*******************************************************************************
+ * Pollset Definitions
+ */
+GPR_TLS_DECL(g_current_thread_pollset);
+GPR_TLS_DECL(g_current_thread_worker);
+
+static void pollset_worker_init(grpc_pollset_worker *worker) {
+  worker->next = worker->prev = NULL;
+  gpr_cv_init(&worker->kick_cv);
+}
+
+/* Global state management */
+static grpc_error *pollset_global_init(void) {
+  gpr_tls_init(&g_current_thread_pollset);
+  gpr_tls_init(&g_current_thread_worker);
+  return grpc_wakeup_fd_init(&global_wakeup_fd);
+}
+
+static void pollset_global_shutdown(void) {
+  grpc_wakeup_fd_destroy(&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) {
+  gpr_cv_signal(&worker->kick_cv);
+  return GRPC_ERROR_NONE;
+}
+
+/* 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(&global_wakeup_fd);
+}
+
+static void pollset_init(grpc_pollset *pollset, gpr_mu **mu) {
+  gpr_mu_init(&pollset->mu);
+  *mu = &pollset->mu;
+  pollset->pi = NULL;
+
+  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;
+  gpr_atm_no_barrier_store(&pollset->is_shutdown, 0);
+}
+
+static void fd_become_readable(grpc_exec_ctx *exec_ctx, grpc_fd *fd,
+                               grpc_pollset *notifier) {
+  grpc_lfev_set_ready(exec_ctx, &fd->read_closure);
+
+  /* Note, it is possible that fd_become_readable might be called twice with
+     different 'notifier's when an fd becomes readable and it is in two epoll
+     sets (This can happen briefly during polling island merges). In such cases
+     it does not really matter which notifer is set as the read_notifier_pollset
+     (They would both point to the same polling island anyway) */
+  /* Use release store to match with acquire load in fd_get_read_notifier */
+  gpr_atm_rel_store(&fd->read_notifier_pollset, (gpr_atm)notifier);
+}
+
+static void fd_become_writable(grpc_exec_ctx *exec_ctx, grpc_fd *fd) {
+  grpc_lfev_set_ready(exec_ctx, &fd->write_closure);
+}
+
+static void pollset_release_polling_island(grpc_exec_ctx *exec_ctx,
+                                           grpc_pollset *ps, char *reason) {
+  if (ps->pi != NULL) {
+    PI_UNREF(exec_ctx, ps->pi, reason);
+  }
+  ps->pi = 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->pi to NULL */
+  pollset_release_polling_island(exec_ctx, pollset, "ps_shutdown");
+  grpc_closure_sched(exec_ctx, pollset->shutdown_done, GRPC_ERROR_NONE);
+}
+
+/* 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 bool maybe_do_workqueue_work(grpc_exec_ctx *exec_ctx,
+                                    polling_island *pi) {
+  if (gpr_mu_trylock(&pi->workqueue_read_mu)) {
+    gpr_mpscq_node *n = gpr_mpscq_pop(&pi->workqueue_items);
+    gpr_mu_unlock(&pi->workqueue_read_mu);
+    if (n != NULL) {
+      if (gpr_atm_full_fetch_add(&pi->workqueue_item_count, -1) > 1) {
+        workqueue_maybe_wakeup(pi);
+      }
+      grpc_closure *c = (grpc_closure *)n;
+      grpc_error *error = c->error_data.error;
+#ifndef NDEBUG
+      c->scheduled = false;
+#endif
+      c->cb(exec_ctx, c->cb_arg, error);
+      GRPC_ERROR_UNREF(error);
+      return true;
+    } else if (gpr_atm_no_barrier_load(&pi->workqueue_item_count) > 0) {
+      /* n == NULL might mean there's work but it's not available to be popped
+       * yet - try to ensure another workqueue wakes up to check shortly if so
+       */
+      workqueue_maybe_wakeup(pi);
+    }
+  }
+  return false;
+}
+
+#define GRPC_EPOLL_MAX_EVENTS 100
+static void pollset_do_epoll_pwait(grpc_exec_ctx *exec_ctx, int epoll_fd,
+                                   grpc_pollset *pollset, polling_island *pi,
+                                   grpc_error **error) {
+  struct epoll_event ep_ev[GRPC_EPOLL_MAX_EVENTS];
+  int ep_rv;
+  char *err_msg;
+  const char *err_desc = "pollset_work_and_unlock";
+
+  int timeout_ms = -1;
+
+  GRPC_SCHEDULING_START_BLOCKING_REGION;
+  // gpr_log(GPR_ERROR, "epoll_wait(%d)..", epoll_fd);
+  ep_rv = epoll_wait(epoll_fd, ep_ev, GRPC_EPOLL_MAX_EVENTS, timeout_ms);
+  /* gpr_log(GPR_ERROR, "epoll_wait(%d) returned: %d (errno: %d - %s)",
+     epoll_fd, ep_rv, errno, strerror(errno)); */
+
+  GRPC_SCHEDULING_END_BLOCKING_REGION;
+
+  if (ep_rv < 0) {
+    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);
+  }
+
+#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 == &global_wakeup_fd) {
+      grpc_timer_consume_kick();
+      append_error(error, grpc_wakeup_fd_consume_wakeup(&global_wakeup_fd),
+                   err_desc);
+    } else if (data_ptr == &pi->workqueue_wakeup_fd) {
+      append_error(error,
+                   grpc_wakeup_fd_consume_wakeup(&pi->workqueue_wakeup_fd),
+                   err_desc);
+      maybe_do_workqueue_work(exec_ctx, pi);
+    } else if (data_ptr == &polling_island_wakeup_fd) {
+      gpr_atm_rel_store(&pollset->is_shutdown, 1);
+      gpr_log(GPR_INFO, "pollset poller: shutdown set");
+    } 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);
+      }
+    }
+  }
+}
+
+static void pollset_work_and_unlock(grpc_exec_ctx *exec_ctx,
+                                    grpc_pollset *pollset, grpc_error **error) {
+  int epoll_fd = -1;
+  polling_island *pi = NULL;
+  GPR_TIMER_BEGIN("pollset_work_and_unlock", 0);
+
+  /* Since epoll_fd is immutable, it is safe to read it without a lock on the
+     polling island. */
+
+  if (pollset->pi == NULL) {
+    pollset->pi = polling_island_create(exec_ctx, NULL, error);
+    if (pollset->pi == NULL) {
+      GPR_TIMER_END("pollset_work_and_unlock", 0);
+      return; /* Fatal error. Cannot continue */
+    }
+
+    PI_ADD_REF(pollset->pi, "ps");
+    GRPC_POLLING_TRACE("pollset_work: pollset: %p created new pi: %p",
+                       (void *)pollset, (void *)pollset->pi);
+  }
+
+  pi = pollset->pi;
+  epoll_fd = pi->epoll_fd;
+
+  /* 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);
+
+  /* If we get some workqueue work to do, it might end up completing an item on
+     the completion queue, so there's no need to poll... so we skip that and
+     redo the complete loop to verify */
+  if (!maybe_do_workqueue_work(exec_ctx, pi)) {
+    gpr_atm_no_barrier_fetch_add(&pi->poller_count, 1);
+    g_current_thread_polling_island = pi;
+    pollset_do_epoll_pwait(exec_ctx, epoll_fd, pollset, pi, error);
+    g_current_thread_polling_island = NULL;
+    gpr_atm_no_barrier_fetch_add(&pi->poller_count, -1);
+  }
+
+  /* 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->pi
+     that we got before releasing the polling island lock). This is because
+     pollset->pi 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(exec_ctx, 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;
+
+  grpc_pollset_worker worker;
+  pollset_worker_init(&worker);
+
+  if (worker_hdl) *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) {
+    push_front_worker(pollset, &worker);
+
+    gpr_cv_wait(&worker.kick_cv, &pollset->mu,
+                gpr_convert_clock_type(deadline, GPR_CLOCK_REALTIME));
+    /* pollset->mu locked here */
+
+    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);
+  }
+
+  if (worker_hdl) *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 *ps,
+                           grpc_fd *fd) {
+  GPR_TIMER_BEGIN("pollset_add_fd", 0);
+
+  grpc_error *error = GRPC_ERROR_NONE;
+  polling_island *pi_new = NULL;
+
+  gpr_mu_lock(&ps->mu);
+  gpr_mu_lock(&fd->mu);
+
+  /* fd MUST have a NULL polling island */
+  GPR_ASSERT(fd->pi == NULL);
+
+  /* Early out if we are trying to add an 'fd' to a 'pollset' but the fd is
+   * already orphaned */
+  if (fd->orphaned) {
+    gpr_mu_unlock(&ps->mu);
+    gpr_mu_unlock(&fd->mu);
+    return;
+  }
+
+  pi_new = ps->pi;
+  if (pi_new == NULL) {
+    /* Unlock before creating a new polling island: the polling island will
+       create a workqueue which creates a file descriptor, and holding an fd
+       lock here can eventually cause a loop to appear to TSAN (making it
+       unhappy). We don't think it's a real loop (there's an epoch point
+       where that loop possibility disappears), but the advantages of
+       keeping TSAN happy outweigh any performance advantage we might have
+       by keeping the lock held. */
+    gpr_mu_unlock(&fd->mu);
+    pi_new = polling_island_create(exec_ctx, fd, &error);
+    gpr_mu_lock(&fd->mu);
+
+    GRPC_POLLING_TRACE(
+        "pollset_add_fd: Created new polling island: %p (ps: %p, fd: %d",
+        (void *)pi_new, (void *)ps, fd->fd);
+  } else {
+    gpr_mu_lock(&pi_new->mu);
+    polling_island_add_fds_locked(pi_new, &fd, 1, true, &error);
+    gpr_mu_unlock(&pi_new->mu);
+
+    GRPC_POLLING_TRACE("pollset_add_fd: ps->pi = %p. Add fd: %d",
+                       (void *)pi_new, fd->fd);
+  }
+
+  PI_ADD_REF(pi_new, "fd");
+  fd->pi = pi_new;
+
+  GPR_ASSERT((ps->pi == NULL) || (ps->pi == pi_new));
+  if (ps->pi == NULL) {
+    PI_ADD_REF(pi_new, "pollset");
+    ps->pi = pi_new;
+  }
+
+  gpr_mu_unlock(&ps->mu);
+  gpr_mu_unlock(&fd->mu);
+
+  GRPC_LOG_IF_ERROR("pollset_add_fd", error);
+  GPR_TIMER_END("pollset_add_fd", 0);
+}
+
+static void pollset_add_fd_no_op(grpc_exec_ctx *exec_ctx, grpc_pollset *pollset,
+                                 grpc_fd *fd) {
+  /* Nothing to do */
+}
+
+/*******************************************************************************
+ * Pollset-set Definitions
+ */
+grpc_pollset_set g_dummy_pollset_set;
+static grpc_pollset_set *pollset_set_create(void) {
+  return &g_dummy_pollset_set;
+}
+
+static void pollset_set_destroy(grpc_exec_ctx *exec_ctx,
+                                grpc_pollset_set *pss) {
+  /* Nothing to do */
+}
+
+static void pollset_set_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset_set *pss,
+                               grpc_fd *fd) {
+  /* Nothing to do */
+}
+
+static void pollset_set_del_fd(grpc_exec_ctx *exec_ctx, grpc_pollset_set *pss,
+                               grpc_fd *fd) {
+  /* Nothing to do */
+}
+
+static void pollset_set_add_pollset(grpc_exec_ctx *exec_ctx,
+                                    grpc_pollset_set *pss, grpc_pollset *ps) {
+  /* Nothing to do */
+}
+
+static void pollset_set_del_pollset(grpc_exec_ctx *exec_ctx,
+                                    grpc_pollset_set *pss, grpc_pollset *ps) {
+  /* Nothing to do */
+}
+
+static void pollset_set_add_pollset_set(grpc_exec_ctx *exec_ctx,
+                                        grpc_pollset_set *bag,
+                                        grpc_pollset_set *item) {
+  /* Nothing to do */
+}
+
+static void pollset_set_del_pollset_set(grpc_exec_ctx *exec_ctx,
+                                        grpc_pollset_set *bag,
+                                        grpc_pollset_set *item) {
+  /* Nothing to do */
+}
+
+/*******************************************************************************
+ * Event engine binding
+ */
+
+static void shutdown_engine(void) {
+  shutdown_dedicated_poller_threads();
+  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,
+    .fd_get_workqueue = fd_get_workqueue,
+
+    .pollset_init = pollset_init,
+    .pollset_shutdown = pollset_shutdown,
+    .pollset_destroy = pollset_destroy,
+    .pollset_work = pollset_work,
+    .pollset_kick = pollset_kick,
+    .pollset_add_fd = pollset_add_fd_no_op,
+
+    .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,
+
+    .workqueue_ref = workqueue_ref,
+    .workqueue_unref = workqueue_unref,
+    .workqueue_scheduler = workqueue_scheduler,
+
+    .shutdown_engine = shutdown_engine,
+};
+
+/*****************************************************************************
+ * Dedicated polling threads and pollsets - Definitions
+ */
+static void add_fd_to_global_pollset(grpc_fd *fd) {
+  size_t idx = ((size_t)rand()) % g_num_pollsets;
+  grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
+  pollset_add_fd(&exec_ctx, &g_pollsets[idx], fd);
+  grpc_exec_ctx_finish(&exec_ctx);
+}
+
+static void init_dedicated_pollsets() {
+  gpr_mu *temp_mu;
+
+  g_num_pollsets = (size_t)gpr_cpu_num_cores();
+  g_pollsets = (grpc_pollset *)malloc(g_num_pollsets * sizeof(grpc_pollset));
+  for (size_t i = 0; i < g_num_pollsets; i++) {
+    pollset_init(&g_pollsets[i], &temp_mu);
+  }
+
+  gpr_log(GPR_INFO, "Created %ld pollsets", g_num_pollsets);
+}
+
+static void poller_thread_loop(void *arg) {
+  grpc_exec_ctx exec_ctx = GRPC_EXEC_CTX_INIT;
+  grpc_error *error = GRPC_ERROR_NONE;
+  grpc_pollset *ps = (grpc_pollset *)arg;
+
+  while (!gpr_atm_acq_load(&ps->is_shutdown)) {
+    gpr_mu_lock(&ps->mu);
+    pollset_work_and_unlock(&exec_ctx, ps, &error);
+    grpc_exec_ctx_flush(&exec_ctx);
+  }
+
+  grpc_exec_ctx_finish(&exec_ctx);
+}
+
+/* g_pollsets MUST be initialized before calling this */
+static void start_dedicated_poller_threads() {
+  GPR_ASSERT(g_pollsets);
+  gpr_log(GPR_ERROR, "Starting poller threads");
+
+  /* One thread per pollset */
+  g_poller_threads = (gpr_thd_id *)malloc(g_num_pollsets * sizeof(gpr_thd_id));
+  gpr_thd_options options = gpr_thd_options_default();
+  gpr_thd_options_set_joinable(&options);
+
+  for (size_t i = 0; i < g_num_pollsets; i++) {
+    gpr_thd_new(&g_poller_threads[i], poller_thread_loop,
+                (void *)&g_pollsets[i], &options);
+  }
+}
+
+static void shutdown_dedicated_poller_threads() {
+  GPR_ASSERT(g_poller_threads);
+  grpc_error *error = GRPC_ERROR_NONE;
+
+  gpr_log(GPR_INFO, "Shutting down pollers");
+
+  for (size_t i = 0; i < g_num_pollsets; i++) {
+    gpr_mu_lock(&g_pollsets[i].mu);
+    polling_island *pi = g_pollsets[i].pi;
+    GPR_ASSERT(pi);
+    gpr_mu_lock(&pi->mu);
+    polling_island_add_wakeup_fd_locked(pi, &polling_island_wakeup_fd, &error);
+    gpr_mu_unlock(&pi->mu);
+  }
+
+  for (size_t i = 0; i < g_num_pollsets; i++) {
+    gpr_thd_join(g_poller_threads[i]);
+  }
+}
+
+/****************************************************************************/
+
+/* 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_thread_pool_linux(void) {
+  if (!grpc_has_wakeup_fd()) {
+    return NULL;
+  }
+
+  if (!is_epoll_available()) {
+    return NULL;
+  }
+
+  init_dedicated_pollsets();
+
+  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;
+  }
+
+  /* TODO (sreek): Maynot be a good idea to start threads here (especially if
+   * this engine doesn't get picked. Consider introducing an engine_init
+   * function in the vtable */
+  start_dedicated_poller_threads();
+  return &vtable;
+}
+
+#else /* defined(GRPC_LINUX_EPOLL) */
+#if defined(GRPC_POSIX_SOCKET)
+#include "src/core/lib/iomgr/ev_posix.h"
+/* If GRPC_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(GRPC_POSIX_SOCKET) */
+#endif /* !defined(GRPC_LINUX_EPOLL) */
diff --git a/src/core/lib/iomgr/ev_epoll_thread_pool_linux.h b/src/core/lib/iomgr/ev_epoll_thread_pool_linux.h
new file mode 100644
index 0000000000..6743dcfe3f
--- /dev/null
+++ b/src/core/lib/iomgr/ev_epoll_thread_pool_linux.h
@@ -0,0 +1,42 @@
+/*
+ *
+ * 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_THREAD_POOL_LINUX_H
+#define GRPC_CORE_LIB_IOMGR_EV_EPOLL_THREAD_POOL_LINUX_H
+
+#include "src/core/lib/iomgr/ev_posix.h"
+#include "src/core/lib/iomgr/port.h"
+
+const grpc_event_engine_vtable *grpc_init_epoll_thread_pool_linux(void);
+
+#endif /* GRPC_CORE_LIB_IOMGR_EV_EPOLL_THREAD_POOL_LINUX_H */
diff --git a/src/core/lib/iomgr/ev_posix.c b/src/core/lib/iomgr/ev_posix.c
index 13409a4de8..4742b2e571 100644
--- a/src/core/lib/iomgr/ev_posix.c
+++ b/src/core/lib/iomgr/ev_posix.c
@@ -45,6 +45,7 @@
 #include <grpc/support/useful.h>
 
 #include "src/core/lib/iomgr/ev_epoll_linux.h"
+#include "src/core/lib/iomgr/ev_epoll_thread_pool_linux.h"
 #include "src/core/lib/iomgr/ev_poll_posix.h"
 #include "src/core/lib/support/env.h"
 
@@ -66,6 +67,7 @@ typedef struct {
 
 static const event_engine_factory g_factories[] = {
     {"epoll", grpc_init_epoll_linux},
+    {"epoll-threadpool", grpc_init_epoll_thread_pool_linux},
     {"poll", grpc_init_poll_posix},
     {"poll-cv", grpc_init_poll_cv_posix},
 };
diff --git a/src/python/grpcio/grpc_core_dependencies.py b/src/python/grpcio/grpc_core_dependencies.py
index d2a570cc87..b249d2ee11 100644
--- a/src/python/grpcio/grpc_core_dependencies.py
+++ b/src/python/grpcio/grpc_core_dependencies.py
@@ -98,6 +98,7 @@ CORE_SOURCE_FILES = [
   'src/core/lib/iomgr/endpoint_pair_windows.c',
   'src/core/lib/iomgr/error.c',
   'src/core/lib/iomgr/ev_epoll_linux.c',
+  'src/core/lib/iomgr/ev_epoll_thread_pool_linux.c',
   'src/core/lib/iomgr/ev_poll_posix.c',
   'src/core/lib/iomgr/ev_posix.c',
   'src/core/lib/iomgr/exec_ctx.c',