path: root/src/core/lib/iomgr/ev_epollex_linux.cc

/*
 *
 * Copyright 2017 gRPC authors.
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 *
 */

#include <grpc/support/port_platform.h>

#include "src/core/lib/iomgr/port.h"

#include <grpc/support/log.h>

/* This polling engine is only relevant on linux kernels supporting epoll() */
#ifdef GRPC_LINUX_EPOLL_CREATE1

#include "src/core/lib/iomgr/ev_epollex_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 <sys/syscall.h>
#include <unistd.h>

#include <grpc/support/alloc.h>
#include <grpc/support/string_util.h>

#include "src/core/lib/debug/stats.h"
#include "src/core/lib/gpr/spinlock.h"
#include "src/core/lib/gpr/tls.h"
#include "src/core/lib/gpr/useful.h"
#include "src/core/lib/gprpp/manual_constructor.h"
#include "src/core/lib/gprpp/mutex_lock.h"
#include "src/core/lib/iomgr/block_annotate.h"
#include "src/core/lib/iomgr/iomgr_internal.h"
#include "src/core/lib/iomgr/is_epollexclusive_available.h"
#include "src/core/lib/iomgr/lockfree_event.h"
#include "src/core/lib/iomgr/sys_epoll_wrapper.h"
#include "src/core/lib/iomgr/timer.h"
#include "src/core/lib/iomgr/wakeup_fd_posix.h"
#include "src/core/lib/profiling/timers.h"

// debug aid: create workers on the heap (allows asan to spot
// use-after-destruction)
//#define GRPC_EPOLLEX_CREATE_WORKERS_ON_HEAP 1

#define MAX_EPOLL_EVENTS 100
// TODO(juanlishen): We use a greater-than-one value here as a workaround fix to
// a keepalive ping timeout issue. We may want to revert https://github
// .com/grpc/grpc/pull/14943 once we figure out the root cause.
#define MAX_EPOLL_EVENTS_HANDLED_EACH_POLL_CALL 16
#define MAX_FDS_IN_CACHE 32

grpc_core::DebugOnlyTraceFlag grpc_trace_pollable_refcount(false,
                                                           "pollable_refcount");

/*******************************************************************************
 * pollable Declarations
 */

typedef enum { PO_MULTI, PO_FD, PO_EMPTY } pollable_type;

typedef struct pollable pollable;

typedef struct cached_fd {
  // Set to the grpc_fd's salt value. See 'salt' variable' in grpc_fd for more
  // details
  intptr_t salt;

  // The underlying fd
  int fd;

  // A recency time counter that helps to determine the LRU fd in the cache
  uint64_t last_used;
} cached_fd;

/// A pollable is something that can be polled: it has an epoll set to poll on,
/// and a wakeup fd for kicks
/// There are three broad types:
///  - PO_EMPTY - the empty pollable, used before file descriptors are added to
///               a pollset
///  - PO_FD - a pollable containing only one FD - used to optimize single-fd
///            pollsets (which are common with synchronous api usage)
///  - PO_MULTI - a pollable containing many fds
struct pollable {
  pollable_type type;  // immutable
  gpr_refcount refs;

  int epfd;
  grpc_wakeup_fd wakeup;

  // The following are relevant only for type PO_FD
  grpc_fd* owner_fd;       // Set to the owner_fd if the type is PO_FD
  gpr_mu owner_orphan_mu;  // Synchronizes access to owner_orphaned field
  bool owner_orphaned;     // Is the owner fd orphaned

  grpc_pollset_set* pollset_set;
  pollable* next;
  pollable* prev;

  gpr_mu mu;
  grpc_pollset_worker* root_worker;

  int event_cursor;
  int event_count;
  struct epoll_event events[MAX_EPOLL_EVENTS];

  // We may be calling pollable_add_fd() on the same (pollable, fd) multiple
  // times. To prevent pollable_add_fd() from making multiple sys calls to
  // epoll_ctl() to add the fd, we maintain a cache of what fds are already
  // present in the underlying epoll-set.
  //
  // Since this is not a correctness issue, we do not need to maintain all the
  // fds in the cache. Hence we just use an LRU cache of size 'MAX_FDS_IN_CACHE'
  //
  // NOTE: An ideal implementation of this should do the following:
  //  1) Add fds to the cache in pollable_add_fd() function (i.e whenever the fd
  //     is added to the pollable's epoll set)
  //  2) Remove the fd from the cache whenever the fd is removed from the
  //     underlying epoll set (i.e whenever fd_orphan() is called).
  //
  // Implementing (2) above (i.e removing fds from cache on fd_orphan) adds a
  // lot of complexity since an fd can be present in multiple pollables. So our
  // implementation ONLY DOES (1) and NOT (2).
  //
  // The cache_fd.salt variable helps here to maintain correctness (it serves as
  // an epoch that differentiates one grpc_fd from the other even though both of
  // them may have the same fd number)
  //
  // The following implements LRU-eviction cache of fds in this pollable
  cached_fd fd_cache[MAX_FDS_IN_CACHE];
  int fd_cache_size;
  uint64_t fd_cache_counter;  // Recency timer tick counter
};

static const char* pollable_type_string(pollable_type t) {
  switch (t) {
    case PO_MULTI:
      return "pollset";
    case PO_FD:
      return "fd";
    case PO_EMPTY:
      return "empty";
  }
  return "<invalid>";
}

static char* pollable_desc(pollable* p) {
  char* out;
  gpr_asprintf(&out, "type=%s epfd=%d wakeup=%d", pollable_type_string(p->type),
               p->epfd, p->wakeup.read_fd);
  return out;
}

/// Shared empty pollable - used by pollset to poll on until the first fd is
/// added
static pollable* g_empty_pollable;

static grpc_error* pollable_create(pollable_type type, pollable** p);
#ifdef NDEBUG
static pollable* pollable_ref(pollable* p);
static void pollable_unref(pollable* p);
#define POLLABLE_REF(p, r) pollable_ref(p)
#define POLLABLE_UNREF(p, r) pollable_unref(p)
#else
static pollable* pollable_ref(pollable* p, int line, const char* reason);
static void pollable_unref(pollable* p, int line, const char* reason);
#define POLLABLE_REF(p, r) pollable_ref((p), __LINE__, (r))
#define POLLABLE_UNREF(p, r) pollable_unref((p), __LINE__, (r))
#endif

/*******************************************************************************
 * Fd Declarations
 */

// Monotonically increasing Epoch counter that is assinged to each grpc_fd. See
// the description of 'salt' variable in 'grpc_fd' for more details
// TODO: (sreek/kpayson) gpr_atm is intptr_t which may not be wide-enough on
// 32-bit systems. Change this to int_64 - atleast on 32-bit systems
static gpr_atm g_fd_salt;

struct grpc_fd {
  int fd;

  // Since fd numbers can be reused (after old fds are closed), this serves as
  // an epoch that uniquely identifies this fd (i.e the pair (salt, fd) is
  // unique (until the salt counter (i.e g_fd_salt) overflows)
  intptr_t salt;

  // 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 orphan_mu;

  gpr_mu pollable_mu;
  pollable* pollable_obj;

  grpc_core::ManualConstructor<grpc_core::LockfreeEvent> read_closure;
  grpc_core::ManualConstructor<grpc_core::LockfreeEvent> write_closure;
  grpc_core::ManualConstructor<grpc_core::LockfreeEvent> error_closure;

  struct grpc_fd* freelist_next;
  grpc_closure* on_done_closure;

  grpc_iomgr_object iomgr_object;

  // Do we need to track EPOLLERR events separately?
  bool track_err;
};

static void fd_global_init(void);
static void fd_global_shutdown(void);

/*******************************************************************************
 * Pollset Declarations
 */

typedef struct {
  grpc_pollset_worker* next;
  grpc_pollset_worker* prev;
} pwlink;

typedef enum { PWLINK_POLLABLE = 0, PWLINK_POLLSET, PWLINK_COUNT } pwlinks;

struct grpc_pollset_worker {
  bool kicked;
  bool initialized_cv;
#ifndef NDEBUG
  // debug aid: which thread started this worker
  pid_t originator;
#endif
  gpr_cv cv;
  grpc_pollset* pollset;
  pollable* pollable_obj;

  pwlink links[PWLINK_COUNT];
};

struct grpc_pollset {
  gpr_mu mu;
  gpr_atm worker_count;
  pollable* active_pollable;
  bool kicked_without_poller;
  grpc_closure* shutdown_closure;
  bool already_shutdown;
  grpc_pollset_worker* root_worker;
  int containing_pollset_set_count;
};

/*******************************************************************************
 * Pollset-set Declarations
 */

struct grpc_pollset_set {
  gpr_refcount refs;
  gpr_mu mu;
  grpc_pollset_set* parent;

  size_t pollset_count;
  size_t pollset_capacity;
  grpc_pollset** pollsets;

  size_t fd_count;
  size_t fd_capacity;
  grpc_fd** fds;
};

/*******************************************************************************
 * 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;
}

/*******************************************************************************
 * 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.
 */

static grpc_fd* fd_freelist = nullptr;
static gpr_mu fd_freelist_mu;

#ifndef NDEBUG
#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) {
  if (grpc_trace_fd_refcount.enabled()) {
    gpr_log(GPR_DEBUG,
            "FD %d %p   ref %d %" PRIdPTR " -> %" PRIdPTR " [%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);
}

#ifndef NDEBUG
#define INVALIDATE_FD(fd) invalidate_fd(fd)
/* Since an fd is never really destroyed (i.e gpr_free() is not called), it is
 * hard to cases where fd fields are accessed even after calling fd_destroy().
 * The following invalidates fd fields to make catching such errors easier */
static void invalidate_fd(grpc_fd* fd) {
  fd->fd = -1;
  fd->salt = -1;
  gpr_atm_no_barrier_store(&fd->refst, -1);
  memset(&fd->orphan_mu, -1, sizeof(fd->orphan_mu));
  memset(&fd->pollable_mu, -1, sizeof(fd->pollable_mu));
  fd->pollable_obj = nullptr;
  fd->on_done_closure = nullptr;
  memset(&fd->iomgr_object, -1, sizeof(fd->iomgr_object));
  fd->track_err = false;
}
#else
#define INVALIDATE_FD(fd)
#endif

/* Uninitialize and add to the freelist */
static void fd_destroy(void* arg, grpc_error* error) {
  grpc_fd* fd = static_cast<grpc_fd*>(arg);
  grpc_iomgr_unregister_object(&fd->iomgr_object);
  POLLABLE_UNREF(fd->pollable_obj, "fd_pollable");
  gpr_mu_destroy(&fd->pollable_mu);
  gpr_mu_destroy(&fd->orphan_mu);

  fd->read_closure->DestroyEvent();
  fd->write_closure->DestroyEvent();
  fd->error_closure->DestroyEvent();

  INVALIDATE_FD(fd);

  /* Add the fd to the freelist */
  gpr_mu_lock(&fd_freelist_mu);
  fd->freelist_next = fd_freelist;
  fd_freelist = fd;
  gpr_mu_unlock(&fd_freelist_mu);
}

#ifndef NDEBUG
static void unref_by(grpc_fd* fd, int n, const char* reason, const char* file,
                     int line) {
  if (grpc_trace_fd_refcount.enabled()) {
    gpr_log(GPR_DEBUG,
            "FD %d %p unref %d %" PRIdPTR " -> %" PRIdPTR " [%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) {
#endif
  gpr_atm old = gpr_atm_full_fetch_add(&fd->refst, -n);
  if (old == n) {
    GRPC_CLOSURE_SCHED(
        GRPC_CLOSURE_CREATE(fd_destroy, fd, grpc_schedule_on_exec_ctx),
        GRPC_ERROR_NONE);
  } else {
    GPR_ASSERT(old > n);
  }
}

static void fd_global_init(void) { gpr_mu_init(&fd_freelist_mu); }

static void fd_global_shutdown(void) {
  // TODO(guantaol): We don't have a reasonable explanation about this
  // lock()/unlock() pattern. It can be a valid barrier if there is at most one
  // pending lock() at this point. Otherwise, there is still a possibility of
  // use-after-free race. Need to reason about the code and/or clean it up.
  gpr_mu_lock(&fd_freelist_mu);
  gpr_mu_unlock(&fd_freelist_mu);
  while (fd_freelist != nullptr) {
    grpc_fd* fd = fd_freelist;
    fd_freelist = fd_freelist->freelist_next;
    gpr_free(fd);
  }
  gpr_mu_destroy(&fd_freelist_mu);
}

static grpc_fd* fd_create(int fd, const char* name, bool track_err) {
  grpc_fd* new_fd = nullptr;

  gpr_mu_lock(&fd_freelist_mu);
  if (fd_freelist != nullptr) {
    new_fd = fd_freelist;
    fd_freelist = fd_freelist->freelist_next;
  }
  gpr_mu_unlock(&fd_freelist_mu);

  if (new_fd == nullptr) {
    new_fd = static_cast<grpc_fd*>(gpr_malloc(sizeof(grpc_fd)));
    new_fd->read_closure.Init();
    new_fd->write_closure.Init();
    new_fd->error_closure.Init();
  }

  new_fd->fd = fd;
  new_fd->salt = gpr_atm_no_barrier_fetch_add(&g_fd_salt, 1);
  gpr_atm_rel_store(&new_fd->refst, (gpr_atm)1);
  gpr_mu_init(&new_fd->orphan_mu);
  gpr_mu_init(&new_fd->pollable_mu);
  new_fd->pollable_obj = nullptr;
  new_fd->read_closure->InitEvent();
  new_fd->write_closure->InitEvent();
  new_fd->error_closure->InitEvent();
  new_fd->freelist_next = nullptr;
  new_fd->on_done_closure = nullptr;

  char* fd_name;
  gpr_asprintf(&fd_name, "%s fd=%d", name, fd);
  grpc_iomgr_register_object(&new_fd->iomgr_object, fd_name);
#ifndef NDEBUG
  if (grpc_trace_fd_refcount.enabled()) {
    gpr_log(GPR_DEBUG, "FD %d %p create %s", fd, new_fd, fd_name);
  }
#endif
  gpr_free(fd_name);

  new_fd->track_err = track_err;
  return new_fd;
}

static int fd_wrapped_fd(grpc_fd* fd) {
  int ret_fd = fd->fd;
  return (gpr_atm_acq_load(&fd->refst) & 1) ? ret_fd : -1;
}

static void fd_orphan(grpc_fd* fd, grpc_closure* on_done, int* release_fd,
                      const char* reason) {
  bool is_fd_closed = false;

  gpr_mu_lock(&fd->orphan_mu);

  // Get the fd->pollable_obj and set the owner_orphaned on that pollable to
  // true so that the pollable will no longer access its owner_fd field.
  gpr_mu_lock(&fd->pollable_mu);
  pollable* pollable_obj = fd->pollable_obj;
  gpr_mu_unlock(&fd->pollable_mu);

  if (pollable_obj) {
    gpr_mu_lock(&pollable_obj->owner_orphan_mu);
    pollable_obj->owner_orphaned = true;
  }

  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 != nullptr) {
    *release_fd = fd->fd;
  } else {
    close(fd->fd);
    is_fd_closed = true;
  }

  // TODO(sreek): handle fd removal (where is_fd_closed=false)
  if (!is_fd_closed) {
    GRPC_FD_TRACE("epoll_fd %p (%d) was orphaned but not closed.", fd, fd->fd);
  }

  /* 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);

  GRPC_CLOSURE_SCHED(fd->on_done_closure, GRPC_ERROR_NONE);

  if (pollable_obj) {
    gpr_mu_unlock(&pollable_obj->owner_orphan_mu);
  }

  gpr_mu_unlock(&fd->orphan_mu);

  UNREF_BY(fd, 2, reason); /* Drop the reference */
}

static bool fd_is_shutdown(grpc_fd* fd) {
  return fd->read_closure->IsShutdown();
}

/* Might be called multiple times */
static void fd_shutdown(grpc_fd* fd, grpc_error* why) {
  if (fd->read_closure->SetShutdown(GRPC_ERROR_REF(why))) {
    if (shutdown(fd->fd, SHUT_RDWR)) {
      if (errno != ENOTCONN) {
        gpr_log(GPR_ERROR, "Error shutting down fd %d. errno: %d",
                grpc_fd_wrapped_fd(fd), errno);
      }
    }
    fd->write_closure->SetShutdown(GRPC_ERROR_REF(why));
    fd->error_closure->SetShutdown(GRPC_ERROR_REF(why));
  }
  GRPC_ERROR_UNREF(why);
}

static void fd_notify_on_read(grpc_fd* fd, grpc_closure* closure) {
  fd->read_closure->NotifyOn(closure);
}

static void fd_notify_on_write(grpc_fd* fd, grpc_closure* closure) {
  fd->write_closure->NotifyOn(closure);
}

static void fd_notify_on_error(grpc_fd* fd, grpc_closure* closure) {
  fd->error_closure->NotifyOn(closure);
}

/*******************************************************************************
 * Pollable Definitions
 */

static grpc_error* pollable_create(pollable_type type, pollable** p) {
  *p = nullptr;

  int epfd = epoll_create1(EPOLL_CLOEXEC);
  if (epfd == -1) {
    return GRPC_OS_ERROR(errno, "epoll_create1");
  }
  GRPC_FD_TRACE("Pollable_create: created epfd: %d (type: %d)", epfd, type);
  *p = static_cast<pollable*>(gpr_malloc(sizeof(**p)));
  grpc_error* err = grpc_wakeup_fd_init(&(*p)->wakeup);
  if (err != GRPC_ERROR_NONE) {
    GRPC_FD_TRACE(
        "Pollable_create: closed epfd: %d (type: %d). wakeupfd_init error",
        epfd, type);
    close(epfd);
    gpr_free(*p);
    *p = nullptr;
    return err;
  }
  struct epoll_event ev;
  ev.events = static_cast<uint32_t>(EPOLLIN | EPOLLET);
  ev.data.ptr = (void*)(1 | (intptr_t) & (*p)->wakeup);
  if (epoll_ctl(epfd, EPOLL_CTL_ADD, (*p)->wakeup.read_fd, &ev) != 0) {
    err = GRPC_OS_ERROR(errno, "epoll_ctl");
    GRPC_FD_TRACE(
        "Pollable_create: closed epfd: %d (type: %d). epoll_ctl error", epfd,
        type);
    close(epfd);
    grpc_wakeup_fd_destroy(&(*p)->wakeup);
    gpr_free(*p);
    *p = nullptr;
    return err;
  }

  (*p)->type = type;
  gpr_ref_init(&(*p)->refs, 1);
  gpr_mu_init(&(*p)->mu);
  (*p)->epfd = epfd;
  (*p)->owner_fd = nullptr;
  gpr_mu_init(&(*p)->owner_orphan_mu);
  (*p)->owner_orphaned = false;
  (*p)->pollset_set = nullptr;
  (*p)->next = (*p)->prev = *p;
  (*p)->root_worker = nullptr;
  (*p)->event_cursor = 0;
  (*p)->event_count = 0;
  (*p)->fd_cache_size = 0;
  (*p)->fd_cache_counter = 0;
  return GRPC_ERROR_NONE;
}

#ifdef NDEBUG
static pollable* pollable_ref(pollable* p) {
#else
static pollable* pollable_ref(pollable* p, int line, const char* reason) {
  if (grpc_trace_pollable_refcount.enabled()) {
    int r = static_cast<int> gpr_atm_no_barrier_load(&p->refs.count);
    gpr_log(__FILE__, line, GPR_LOG_SEVERITY_DEBUG,
            "POLLABLE:%p   ref %d->%d %s", p, r, r + 1, reason);
  }
#endif
  gpr_ref(&p->refs);
  return p;
}

#ifdef NDEBUG
static void pollable_unref(pollable* p) {
#else
static void pollable_unref(pollable* p, int line, const char* reason) {
  if (p == nullptr) return;
  if (grpc_trace_pollable_refcount.enabled()) {
    int r = static_cast<int> gpr_atm_no_barrier_load(&p->refs.count);
    gpr_log(__FILE__, line, GPR_LOG_SEVERITY_DEBUG,
            "POLLABLE:%p unref %d->%d %s", p, r, r - 1, reason);
  }
#endif
  if (p != nullptr && gpr_unref(&p->refs)) {
    GRPC_FD_TRACE("pollable_unref: Closing epfd: %d", p->epfd);
    close(p->epfd);
    grpc_wakeup_fd_destroy(&p->wakeup);
    gpr_mu_destroy(&p->owner_orphan_mu);
    gpr_free(p);
  }
}

static grpc_error* pollable_add_fd(pollable* p, grpc_fd* fd) {
  grpc_error* error = GRPC_ERROR_NONE;
  static const char* err_desc = "pollable_add_fd";
  const int epfd = p->epfd;
  gpr_mu_lock(&p->mu);
  p->fd_cache_counter++;

  // Handle the case of overflow for our cache counter by
  // reseting the recency-counter on all cache objects
  if (p->fd_cache_counter == 0) {
    for (int i = 0; i < p->fd_cache_size; i++) {
      p->fd_cache[i].last_used = 0;
    }
  }

  int lru_idx = 0;
  for (int i = 0; i < p->fd_cache_size; i++) {
    if (p->fd_cache[i].fd == fd->fd && p->fd_cache[i].salt == fd->salt) {
      GRPC_STATS_INC_POLLSET_FD_CACHE_HITS();
      p->fd_cache[i].last_used = p->fd_cache_counter;
      gpr_mu_unlock(&p->mu);
      return GRPC_ERROR_NONE;
    } else if (p->fd_cache[i].last_used < p->fd_cache[lru_idx].last_used) {
      lru_idx = i;
    }
  }

  // Add to cache
  if (p->fd_cache_size < MAX_FDS_IN_CACHE) {
    lru_idx = p->fd_cache_size;
    p->fd_cache_size++;
  }
  p->fd_cache[lru_idx].fd = fd->fd;
  p->fd_cache[lru_idx].salt = fd->salt;
  p->fd_cache[lru_idx].last_used = p->fd_cache_counter;
  gpr_mu_unlock(&p->mu);

  if (grpc_polling_trace.enabled()) {
    gpr_log(GPR_INFO, "add fd %p (%d) to pollable %p", fd, fd->fd, p);
  }

  struct epoll_event ev_fd;
  ev_fd.events =
      static_cast<uint32_t>(EPOLLET | EPOLLIN | EPOLLOUT | EPOLLEXCLUSIVE);
  /* Use the second least significant bit of ev_fd.data.ptr to store track_err
   * to avoid synchronization issues when accessing it after receiving an event.
   * Accessing fd would be a data race there because the fd might have been
   * returned to the free list at that point. */
  ev_fd.data.ptr = reinterpret_cast<void*>(reinterpret_cast<intptr_t>(fd) |
                                           (fd->track_err ? 2 : 0));
  GRPC_STATS_INC_SYSCALL_EPOLL_CTL();
  if (epoll_ctl(epfd, EPOLL_CTL_ADD, fd->fd, &ev_fd) != 0) {
    switch (errno) {
      case EEXIST:
        break;
      default:
        append_error(&error, GRPC_OS_ERROR(errno, "epoll_ctl"), err_desc);
    }
  }

  return error;
}

/*******************************************************************************
 * Pollset Definitions
 */

GPR_TLS_DECL(g_current_thread_pollset);
GPR_TLS_DECL(g_current_thread_worker);

/* 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 pollable_create(PO_EMPTY, &g_empty_pollable);
}

static void pollset_global_shutdown(void) {
  POLLABLE_UNREF(g_empty_pollable, "g_empty_pollable");
  gpr_tls_destroy(&g_current_thread_pollset);
  gpr_tls_destroy(&g_current_thread_worker);
}

/* pollset->mu must be held while calling this function */
static void pollset_maybe_finish_shutdown(grpc_pollset* pollset) {
  if (grpc_polling_trace.enabled()) {
    gpr_log(GPR_INFO,
            "PS:%p (pollable:%p) maybe_finish_shutdown sc=%p (target:!NULL) "
            "rw=%p (target:NULL) cpsc=%d (target:0)",
            pollset, pollset->active_pollable, pollset->shutdown_closure,
            pollset->root_worker, pollset->containing_pollset_set_count);
  }
  if (pollset->shutdown_closure != nullptr && pollset->root_worker == nullptr &&
      pollset->containing_pollset_set_count == 0) {
    GPR_TIMER_MARK("pollset_finish_shutdown", 0);
    GRPC_CLOSURE_SCHED(pollset->shutdown_closure, GRPC_ERROR_NONE);
    pollset->shutdown_closure = nullptr;
    pollset->already_shutdown = true;
  }
}

/* pollset->mu must be held before calling this function,
 * pollset->active_pollable->mu & specific_worker->pollable_obj->mu must not be
 * held */
static grpc_error* kick_one_worker(grpc_pollset_worker* specific_worker) {
  GPR_TIMER_SCOPE("kick_one_worker", 0);
  pollable* p = specific_worker->pollable_obj;
  grpc_core::MutexLock lock(&p->mu);
  GPR_ASSERT(specific_worker != nullptr);
  if (specific_worker->kicked) {
    if (grpc_polling_trace.enabled()) {
      gpr_log(GPR_INFO, "PS:%p kicked_specific_but_already_kicked", p);
    }
    GRPC_STATS_INC_POLLSET_KICKED_AGAIN();
    return GRPC_ERROR_NONE;
  }
  if (gpr_tls_get(&g_current_thread_worker) == (intptr_t)specific_worker) {
    if (grpc_polling_trace.enabled()) {
      gpr_log(GPR_INFO, "PS:%p kicked_specific_but_awake", p);
    }
    GRPC_STATS_INC_POLLSET_KICK_OWN_THREAD();
    specific_worker->kicked = true;
    return GRPC_ERROR_NONE;
  }
  if (specific_worker == p->root_worker) {
    GRPC_STATS_INC_POLLSET_KICK_WAKEUP_FD();
    if (grpc_polling_trace.enabled()) {
      gpr_log(GPR_INFO, "PS:%p kicked_specific_via_wakeup_fd", p);
    }
    specific_worker->kicked = true;
    grpc_error* error = grpc_wakeup_fd_wakeup(&p->wakeup);
    return error;
  }
  if (specific_worker->initialized_cv) {
    GRPC_STATS_INC_POLLSET_KICK_WAKEUP_CV();
    if (grpc_polling_trace.enabled()) {
      gpr_log(GPR_INFO, "PS:%p kicked_specific_via_cv", p);
    }
    specific_worker->kicked = true;
    gpr_cv_signal(&specific_worker->cv);
    return GRPC_ERROR_NONE;
  }
  // we can get here during end_worker after removing specific_worker from the
  // pollable list but before removing it from the pollset list
  return GRPC_ERROR_NONE;
}

static grpc_error* pollset_kick(grpc_pollset* pollset,
                                grpc_pollset_worker* specific_worker) {
  GPR_TIMER_SCOPE("pollset_kick", 0);
  GRPC_STATS_INC_POLLSET_KICK();
  if (grpc_polling_trace.enabled()) {
    gpr_log(GPR_INFO,
            "PS:%p kick %p tls_pollset=%p tls_worker=%p pollset.root_worker=%p",
            pollset, specific_worker,
            (void*)gpr_tls_get(&g_current_thread_pollset),
            (void*)gpr_tls_get(&g_current_thread_worker), pollset->root_worker);
  }
  if (specific_worker == nullptr) {
    if (gpr_tls_get(&g_current_thread_pollset) != (intptr_t)pollset) {
      if (pollset->root_worker == nullptr) {
        if (grpc_polling_trace.enabled()) {
          gpr_log(GPR_INFO, "PS:%p kicked_any_without_poller", pollset);
        }
        GRPC_STATS_INC_POLLSET_KICKED_WITHOUT_POLLER();
        pollset->kicked_without_poller = true;
        return GRPC_ERROR_NONE;
      } else {
        // We've been asked to kick a poller, but we haven't been told which one
        // ... any will do
        // We look at the pollset worker list because:
        // 1. the pollable list may include workers from other pollers, so we'd
        //    need to do an O(N) search
        // 2. we'd additionally need to take the pollable lock, which we've so
        //    far avoided
        // Now, we would prefer to wake a poller in cv_wait, and not in
        // epoll_wait (since the latter would imply the need to do an additional
        // wakeup)
        // We know that if a worker is at the root of a pollable, it's (likely)
        // also the root of a pollset, and we know that if a worker is NOT at
        // the root of a pollset, it's (likely) not at the root of a pollable,
        // so we take our chances and choose the SECOND worker enqueued against
        // the pollset as a worker that's likely to be in cv_wait
        return kick_one_worker(
            pollset->root_worker->links[PWLINK_POLLSET].next);
      }
    } else {
      if (grpc_polling_trace.enabled()) {
        gpr_log(GPR_INFO, "PS:%p kicked_any_but_awake", pollset);
      }
      GRPC_STATS_INC_POLLSET_KICK_OWN_THREAD();
      return GRPC_ERROR_NONE;
    }
  } else {
    return kick_one_worker(specific_worker);
  }
}

static grpc_error* pollset_kick_all(grpc_pollset* pollset) {
  GPR_TIMER_SCOPE("pollset_kick_all", 0);
  grpc_error* error = GRPC_ERROR_NONE;
  const char* err_desc = "pollset_kick_all";
  grpc_pollset_worker* w = pollset->root_worker;
  if (w != nullptr) {
    do {
      GRPC_STATS_INC_POLLSET_KICK();
      append_error(&error, kick_one_worker(w), err_desc);
      w = w->links[PWLINK_POLLSET].next;
    } while (w != pollset->root_worker);
  }
  return error;
}

static void pollset_init(grpc_pollset* pollset, gpr_mu** mu) {
  gpr_mu_init(&pollset->mu);
  gpr_atm_no_barrier_store(&pollset->worker_count, 0);
  pollset->active_pollable = POLLABLE_REF(g_empty_pollable, "pollset");
  pollset->kicked_without_poller = false;
  pollset->shutdown_closure = nullptr;
  pollset->already_shutdown = false;
  pollset->root_worker = nullptr;
  pollset->containing_pollset_set_count = 0;
  *mu = &pollset->mu;
}

static int poll_deadline_to_millis_timeout(grpc_millis millis) {
  if (millis == GRPC_MILLIS_INF_FUTURE) return -1;
  grpc_millis delta = millis - grpc_core::ExecCtx::Get()->Now();
  if (delta > INT_MAX)
    return INT_MAX;
  else if (delta < 0)
    return 0;
  else
    return static_cast<int>(delta);
}

static void fd_become_readable(grpc_fd* fd) { fd->read_closure->SetReady(); }

static void fd_become_writable(grpc_fd* fd) { fd->write_closure->SetReady(); }

static void fd_has_errors(grpc_fd* fd) { fd->error_closure->SetReady(); }

/* Get the pollable_obj attached to this fd. If none is attached, create a new
 * pollable object (of type PO_FD), attach it to the fd and return it
 *
 * Note that if a pollable object is already attached to the fd, it may be of
 * either PO_FD or PO_MULTI type */
static grpc_error* get_fd_pollable(grpc_fd* fd, pollable** p) {
  gpr_mu_lock(&fd->pollable_mu);
  grpc_error* error = GRPC_ERROR_NONE;
  static const char* err_desc = "get_fd_pollable";
  if (fd->pollable_obj == nullptr) {
    if (append_error(&error, pollable_create(PO_FD, &fd->pollable_obj),
                     err_desc)) {
      fd->pollable_obj->owner_fd = fd;
      if (!append_error(&error, pollable_add_fd(fd->pollable_obj, fd),
                        err_desc)) {
        POLLABLE_UNREF(fd->pollable_obj, "fd_pollable");
        fd->pollable_obj = nullptr;
      }
    }
  }
  if (error == GRPC_ERROR_NONE) {
    GPR_ASSERT(fd->pollable_obj != nullptr);
    *p = POLLABLE_REF(fd->pollable_obj, "pollset");
  } else {
    GPR_ASSERT(fd->pollable_obj == nullptr);
    *p = nullptr;
  }
  gpr_mu_unlock(&fd->pollable_mu);
  return error;
}

/* pollset->po.mu lock must be held by the caller before calling this */
static void pollset_shutdown(grpc_pollset* pollset, grpc_closure* closure) {
  GPR_TIMER_SCOPE("pollset_shutdown", 0);
  GPR_ASSERT(pollset->shutdown_closure == nullptr);
  pollset->shutdown_closure = closure;
  GRPC_LOG_IF_ERROR("pollset_shutdown", pollset_kick_all(pollset));
  pollset_maybe_finish_shutdown(pollset);
}

static grpc_error* pollable_process_events(grpc_pollset* pollset,
                                           pollable* pollable_obj, bool drain) {
  GPR_TIMER_SCOPE("pollable_process_events", 0);
  static const char* err_desc = "pollset_process_events";
  // Use a simple heuristic to determine how many fd events to process
  // per loop iteration.  (events/workers)
  int handle_count = 1;
  int worker_count = gpr_atm_no_barrier_load(&pollset->worker_count);
  GPR_ASSERT(worker_count > 0);
  handle_count =
      (pollable_obj->event_count - pollable_obj->event_cursor) / worker_count;
  if (handle_count == 0) {
    handle_count = 1;
  } else if (handle_count > MAX_EPOLL_EVENTS_HANDLED_EACH_POLL_CALL) {
    handle_count = MAX_EPOLL_EVENTS_HANDLED_EACH_POLL_CALL;
  }
  grpc_error* error = GRPC_ERROR_NONE;
  for (int i = 0; (drain || i < handle_count) &&
                  pollable_obj->event_cursor != pollable_obj->event_count;
       i++) {
    int n = pollable_obj->event_cursor++;
    struct epoll_event* ev = &pollable_obj->events[n];
    void* data_ptr = ev->data.ptr;
    if (1 & (intptr_t)data_ptr) {
      if (grpc_polling_trace.enabled()) {
        gpr_log(GPR_INFO, "PS:%p got pollset_wakeup %p", pollset, data_ptr);
      }
      append_error(&error,
                   grpc_wakeup_fd_consume_wakeup(
                       (grpc_wakeup_fd*)((~static_cast<intptr_t>(1)) &
                                         (intptr_t)data_ptr)),
                   err_desc);
    } else {
      grpc_fd* fd =
          reinterpret_cast<grpc_fd*>(reinterpret_cast<intptr_t>(data_ptr) & ~2);
      bool track_err = reinterpret_cast<intptr_t>(data_ptr) & 2;
      bool cancel = (ev->events & EPOLLHUP) != 0;
      bool error = (ev->events & EPOLLERR) != 0;
      bool read_ev = (ev->events & (EPOLLIN | EPOLLPRI)) != 0;
      bool write_ev = (ev->events & EPOLLOUT) != 0;
      bool err_fallback = error && !track_err;

      if (grpc_polling_trace.enabled()) {
        gpr_log(GPR_INFO,
                "PS:%p got fd %p: cancel=%d read=%d "
                "write=%d",
                pollset, fd, cancel, read_ev, write_ev);
      }
      if (error && !err_fallback) {
        fd_has_errors(fd);
      }
      if (read_ev || cancel || err_fallback) {
        fd_become_readable(fd);
      }
      if (write_ev || cancel || err_fallback) {
        fd_become_writable(fd);
      }
    }
  }

  return error;
}

/* pollset_shutdown is guaranteed to be called before pollset_destroy. */
static void pollset_destroy(grpc_pollset* pollset) {
  POLLABLE_UNREF(pollset->active_pollable, "pollset");
  pollset->active_pollable = nullptr;
  gpr_mu_destroy(&pollset->mu);
}

static grpc_error* pollable_epoll(pollable* p, grpc_millis deadline) {
  GPR_TIMER_SCOPE("pollable_epoll", 0);
  int timeout = poll_deadline_to_millis_timeout(deadline);

  if (grpc_polling_trace.enabled()) {
    char* desc = pollable_desc(p);
    gpr_log(GPR_INFO, "POLLABLE:%p[%s] poll for %dms", p, desc, timeout);
    gpr_free(desc);
  }

  if (timeout != 0) {
    GRPC_SCHEDULING_START_BLOCKING_REGION;
  }
  int r;
  do {
    GRPC_STATS_INC_SYSCALL_POLL();
    r = epoll_wait(p->epfd, p->events, MAX_EPOLL_EVENTS, timeout);
  } while (r < 0 && errno == EINTR);
  if (timeout != 0) {
    GRPC_SCHEDULING_END_BLOCKING_REGION;
  }

  if (r < 0) return GRPC_OS_ERROR(errno, "epoll_wait");

  if (grpc_polling_trace.enabled()) {
    gpr_log(GPR_INFO, "POLLABLE:%p got %d events", p, r);
  }

  p->event_cursor = 0;
  p->event_count = r;

  return GRPC_ERROR_NONE;
}

/* Return true if first in list */
static bool worker_insert(grpc_pollset_worker** root_worker,
                          grpc_pollset_worker* worker, pwlinks link) {
  if (*root_worker == nullptr) {
    *root_worker = worker;
    worker->links[link].next = worker->links[link].prev = worker;
    return true;
  } else {
    worker->links[link].next = *root_worker;
    worker->links[link].prev = worker->links[link].next->links[link].prev;
    worker->links[link].next->links[link].prev = worker;
    worker->links[link].prev->links[link].next = worker;
    return false;
  }
}

/* returns the new root IFF the root changed */
typedef enum { WRR_NEW_ROOT, WRR_EMPTIED, WRR_REMOVED } worker_remove_result;

static worker_remove_result worker_remove(grpc_pollset_worker** root_worker,
                                          grpc_pollset_worker* worker,
                                          pwlinks link) {
  if (worker == *root_worker) {
    if (worker == worker->links[link].next) {
      *root_worker = nullptr;
      return WRR_EMPTIED;
    } else {
      *root_worker = worker->links[link].next;
      worker->links[link].prev->links[link].next = worker->links[link].next;
      worker->links[link].next->links[link].prev = worker->links[link].prev;
      return WRR_NEW_ROOT;
    }
  } else {
    worker->links[link].prev->links[link].next = worker->links[link].next;
    worker->links[link].next->links[link].prev = worker->links[link].prev;
    return WRR_REMOVED;
  }
}

/* Return true if this thread should poll */
static bool begin_worker(grpc_pollset* pollset, grpc_pollset_worker* worker,
                         grpc_pollset_worker** worker_hdl,
                         grpc_millis deadline) {
  GPR_TIMER_SCOPE("begin_worker", 0);
  bool do_poll =
      (pollset->shutdown_closure == nullptr && !pollset->already_shutdown);
  gpr_atm_no_barrier_fetch_add(&pollset->worker_count, 1);
  if (worker_hdl != nullptr) *worker_hdl = worker;
  worker->initialized_cv = false;
  worker->kicked = false;
  worker->pollset = pollset;
  worker->pollable_obj =
      POLLABLE_REF(pollset->active_pollable, "pollset_worker");
  worker_insert(&pollset->root_worker, worker, PWLINK_POLLSET);
  gpr_mu_lock(&worker->pollable_obj->mu);
  if (!worker_insert(&worker->pollable_obj->root_worker, worker,
                     PWLINK_POLLABLE)) {
    worker->initialized_cv = true;
    gpr_cv_init(&worker->cv);
    gpr_mu_unlock(&pollset->mu);
    if (grpc_polling_trace.enabled() &&
        worker->pollable_obj->root_worker != worker) {
      gpr_log(GPR_INFO, "PS:%p wait %p w=%p for %dms", pollset,
              worker->pollable_obj, worker,
              poll_deadline_to_millis_timeout(deadline));
    }
    while (do_poll && worker->pollable_obj->root_worker != worker) {
      if (gpr_cv_wait(&worker->cv, &worker->pollable_obj->mu,
                      grpc_millis_to_timespec(deadline, GPR_CLOCK_REALTIME))) {
        if (grpc_polling_trace.enabled()) {
          gpr_log(GPR_INFO, "PS:%p timeout_wait %p w=%p", pollset,
                  worker->pollable_obj, worker);
        }
        do_poll = false;
      } else if (worker->kicked) {
        if (grpc_polling_trace.enabled()) {
          gpr_log(GPR_INFO, "PS:%p wakeup %p w=%p", pollset,
                  worker->pollable_obj, worker);
        }
        do_poll = false;
      } else if (grpc_polling_trace.enabled() &&
                 worker->pollable_obj->root_worker != worker) {
        gpr_log(GPR_INFO, "PS:%p spurious_wakeup %p w=%p", pollset,
                worker->pollable_obj, worker);
      }
    }
    grpc_core::ExecCtx::Get()->InvalidateNow();
  } else {
    gpr_mu_unlock(&pollset->mu);
  }
  gpr_mu_unlock(&worker->pollable_obj->mu);

  return do_poll;
}

static void end_worker(grpc_pollset* pollset, grpc_pollset_worker* worker,
                       grpc_pollset_worker** worker_hdl) {
  GPR_TIMER_SCOPE("end_worker", 0);
  gpr_mu_lock(&pollset->mu);
  gpr_mu_lock(&worker->pollable_obj->mu);
  switch (worker_remove(&worker->pollable_obj->root_worker, worker,
                        PWLINK_POLLABLE)) {
    case WRR_NEW_ROOT: {
      // wakeup new poller
      grpc_pollset_worker* new_root = worker->pollable_obj->root_worker;
      GPR_ASSERT(new_root->initialized_cv);
      gpr_cv_signal(&new_root->cv);
      break;
    }
    case WRR_EMPTIED:
      if (pollset->active_pollable != worker->pollable_obj) {
        // pollable no longer being polled: flush events
        pollable_process_events(pollset, worker->pollable_obj, true);
      }
      break;
    case WRR_REMOVED:
      break;
  }
  gpr_mu_unlock(&worker->pollable_obj->mu);
  POLLABLE_UNREF(worker->pollable_obj, "pollset_worker");
  if (worker_remove(&pollset->root_worker, worker, PWLINK_POLLSET) ==
      WRR_EMPTIED) {
    pollset_maybe_finish_shutdown(pollset);
  }
  if (worker->initialized_cv) {
    gpr_cv_destroy(&worker->cv);
  }
  gpr_atm_no_barrier_fetch_add(&pollset->worker_count, -1);
}

#ifndef NDEBUG
static long gettid(void) { return syscall(__NR_gettid); }
#endif

/* pollset->mu lock must be held by the caller before calling this.
   The function pollset_work() may temporarily release the lock (pollset->po.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_pollset* pollset,
                                grpc_pollset_worker** worker_hdl,
                                grpc_millis deadline) {
  GPR_TIMER_SCOPE("pollset_work", 0);
#ifdef GRPC_EPOLLEX_CREATE_WORKERS_ON_HEAP
  grpc_pollset_worker* worker =
      (grpc_pollset_worker*)gpr_malloc(sizeof(*worker));
#define WORKER_PTR (worker)
#else
  grpc_pollset_worker worker;
#define WORKER_PTR (&worker)
#endif
#ifndef NDEBUG
  WORKER_PTR->originator = gettid();
#endif
  if (grpc_polling_trace.enabled()) {
    gpr_log(GPR_INFO,
            "PS:%p work hdl=%p worker=%p now=%" PRId64 " deadline=%" PRId64
            " kwp=%d pollable=%p",
            pollset, worker_hdl, WORKER_PTR, grpc_core::ExecCtx::Get()->Now(),
            deadline, pollset->kicked_without_poller, pollset->active_pollable);
  }
  static const char* err_desc = "pollset_work";
  grpc_error* error = GRPC_ERROR_NONE;
  if (pollset->kicked_without_poller) {
    pollset->kicked_without_poller = false;
  } else {
    if (begin_worker(pollset, WORKER_PTR, worker_hdl, deadline)) {
      gpr_tls_set(&g_current_thread_pollset, (intptr_t)pollset);
      gpr_tls_set(&g_current_thread_worker, (intptr_t)WORKER_PTR);
      if (WORKER_PTR->pollable_obj->event_cursor ==
          WORKER_PTR->pollable_obj->event_count) {
        append_error(&error, pollable_epoll(WORKER_PTR->pollable_obj, deadline),
                     err_desc);
      }
      append_error(
          &error,
          pollable_process_events(pollset, WORKER_PTR->pollable_obj, false),
          err_desc);
      grpc_core::ExecCtx::Get()->Flush();
      gpr_tls_set(&g_current_thread_pollset, 0);
      gpr_tls_set(&g_current_thread_worker, 0);
    }
    end_worker(pollset, WORKER_PTR, worker_hdl);
  }
#ifdef GRPC_EPOLLEX_CREATE_WORKERS_ON_HEAP
  gpr_free(worker);
#endif
#undef WORKER_PTR
  return error;
}

static grpc_error* pollset_transition_pollable_from_empty_to_fd_locked(
    grpc_pollset* pollset, grpc_fd* fd) {
  static const char* err_desc = "pollset_transition_pollable_from_empty_to_fd";
  grpc_error* error = GRPC_ERROR_NONE;
  if (grpc_polling_trace.enabled()) {
    gpr_log(GPR_INFO,
            "PS:%p add fd %p (%d); transition pollable from empty to fd",
            pollset, fd, fd->fd);
  }
  append_error(&error, pollset_kick_all(pollset), err_desc);
  POLLABLE_UNREF(pollset->active_pollable, "pollset");
  append_error(&error, get_fd_pollable(fd, &pollset->active_pollable),
               err_desc);
  return error;
}

static grpc_error* pollset_transition_pollable_from_fd_to_multi_locked(
    grpc_pollset* pollset, grpc_fd* and_add_fd) {
  static const char* err_desc = "pollset_transition_pollable_from_fd_to_multi";
  grpc_error* error = GRPC_ERROR_NONE;
  if (grpc_polling_trace.enabled()) {
    gpr_log(
        GPR_INFO,
        "PS:%p add fd %p (%d); transition pollable from fd %p to multipoller",
        pollset, and_add_fd, and_add_fd ? and_add_fd->fd : -1,
        pollset->active_pollable->owner_fd);
  }
  append_error(&error, pollset_kick_all(pollset), err_desc);
  grpc_fd* initial_fd = pollset->active_pollable->owner_fd;
  POLLABLE_UNREF(pollset->active_pollable, "pollset");
  pollset->active_pollable = nullptr;
  if (append_error(&error, pollable_create(PO_MULTI, &pollset->active_pollable),
                   err_desc)) {
    append_error(&error, pollable_add_fd(pollset->active_pollable, initial_fd),
                 err_desc);
    if (and_add_fd != nullptr) {
      append_error(&error,
                   pollable_add_fd(pollset->active_pollable, and_add_fd),
                   err_desc);
    }
  }
  return error;
}

/* expects pollsets locked, flag whether fd is locked or not */
static grpc_error* pollset_add_fd_locked(grpc_pollset* pollset, grpc_fd* fd) {
  grpc_error* error = GRPC_ERROR_NONE;
  pollable* po_at_start =
      POLLABLE_REF(pollset->active_pollable, "pollset_add_fd");
  switch (pollset->active_pollable->type) {
    case PO_EMPTY:
      /* empty pollable --> single fd pollable */
      error = pollset_transition_pollable_from_empty_to_fd_locked(pollset, fd);
      break;
    case PO_FD:
      gpr_mu_lock(&po_at_start->owner_orphan_mu);
      if (po_at_start->owner_orphaned) {
        error =
            pollset_transition_pollable_from_empty_to_fd_locked(pollset, fd);
      } else {
        /* fd --> multipoller */
        error =
            pollset_transition_pollable_from_fd_to_multi_locked(pollset, fd);
      }
      gpr_mu_unlock(&po_at_start->owner_orphan_mu);
      break;
    case PO_MULTI:
      error = pollable_add_fd(pollset->active_pollable, fd);
      break;
  }
  if (error != GRPC_ERROR_NONE) {
    POLLABLE_UNREF(pollset->active_pollable, "pollset");
    pollset->active_pollable = po_at_start;
  } else {
    POLLABLE_UNREF(po_at_start, "pollset_add_fd");
  }
  return error;
}

static grpc_error* pollset_as_multipollable_locked(grpc_pollset* pollset,
                                                   pollable** pollable_obj) {
  grpc_error* error = GRPC_ERROR_NONE;
  pollable* po_at_start =
      POLLABLE_REF(pollset->active_pollable, "pollset_as_multipollable");
  switch (pollset->active_pollable->type) {
    case PO_EMPTY:
      POLLABLE_UNREF(pollset->active_pollable, "pollset");
      error = pollable_create(PO_MULTI, &pollset->active_pollable);
      /* Any workers currently polling on this pollset must now be woked up so
       * that they can pick up the new active_pollable */
      if (grpc_polling_trace.enabled()) {
        gpr_log(GPR_INFO,
                "PS:%p active pollable transition from empty to multi",
                pollset);
      }
      static const char* err_desc =
          "pollset_as_multipollable_locked: empty -> multi";
      append_error(&error, pollset_kick_all(pollset), err_desc);
      break;
    case PO_FD:
      gpr_mu_lock(&po_at_start->owner_orphan_mu);
      if (po_at_start->owner_orphaned) {
        // Unlock before Unref'ing the pollable
        gpr_mu_unlock(&po_at_start->owner_orphan_mu);
        POLLABLE_UNREF(pollset->active_pollable, "pollset");
        error = pollable_create(PO_MULTI, &pollset->active_pollable);
      } else {
        error = pollset_transition_pollable_from_fd_to_multi_locked(pollset,
                                                                    nullptr);
        gpr_mu_unlock(&po_at_start->owner_orphan_mu);
      }
      break;
    case PO_MULTI:
      break;
  }
  if (error != GRPC_ERROR_NONE) {
    POLLABLE_UNREF(pollset->active_pollable, "pollset");
    pollset->active_pollable = po_at_start;
    *pollable_obj = nullptr;
  } else {
    *pollable_obj = POLLABLE_REF(pollset->active_pollable, "pollset_set");
    POLLABLE_UNREF(po_at_start, "pollset_as_multipollable");
  }
  return error;
}

static void pollset_add_fd(grpc_pollset* pollset, grpc_fd* fd) {
  GPR_TIMER_SCOPE("pollset_add_fd", 0);
  gpr_mu_lock(&pollset->mu);
  grpc_error* error = pollset_add_fd_locked(pollset, fd);
  gpr_mu_unlock(&pollset->mu);
  GRPC_LOG_IF_ERROR("pollset_add_fd", error);
}

/*******************************************************************************
 * Pollset-set Definitions
 */

static grpc_pollset_set* pss_lock_adam(grpc_pollset_set* pss) {
  gpr_mu_lock(&pss->mu);
  while (pss->parent != nullptr) {
    gpr_mu_unlock(&pss->mu);
    pss = pss->parent;
    gpr_mu_lock(&pss->mu);
  }
  return pss;
}

static grpc_pollset_set* pollset_set_create(void) {
  grpc_pollset_set* pss =
      static_cast<grpc_pollset_set*>(gpr_zalloc(sizeof(*pss)));
  gpr_mu_init(&pss->mu);
  gpr_ref_init(&pss->refs, 1);
  return pss;
}

static void pollset_set_unref(grpc_pollset_set* pss) {
  if (pss == nullptr) return;
  if (!gpr_unref(&pss->refs)) return;
  pollset_set_unref(pss->parent);
  gpr_mu_destroy(&pss->mu);
  for (size_t i = 0; i < pss->pollset_count; i++) {
    gpr_mu_lock(&pss->pollsets[i]->mu);
    if (0 == --pss->pollsets[i]->containing_pollset_set_count) {
      pollset_maybe_finish_shutdown(pss->pollsets[i]);
    }
    gpr_mu_unlock(&pss->pollsets[i]->mu);
  }
  for (size_t i = 0; i < pss->fd_count; i++) {
    UNREF_BY(pss->fds[i], 2, "pollset_set");
  }
  gpr_free(pss->pollsets);
  gpr_free(pss->fds);
  gpr_free(pss);
}

static void pollset_set_add_fd(grpc_pollset_set* pss, grpc_fd* fd) {
  GPR_TIMER_SCOPE("pollset_set_add_fd", 0);
  if (grpc_polling_trace.enabled()) {
    gpr_log(GPR_INFO, "PSS:%p: add fd %p (%d)", pss, fd, fd->fd);
  }
  grpc_error* error = GRPC_ERROR_NONE;
  static const char* err_desc = "pollset_set_add_fd";
  pss = pss_lock_adam(pss);
  for (size_t i = 0; i < pss->pollset_count; i++) {
    append_error(&error, pollable_add_fd(pss->pollsets[i]->active_pollable, fd),
                 err_desc);
  }
  if (pss->fd_count == pss->fd_capacity) {
    pss->fd_capacity = GPR_MAX(pss->fd_capacity * 2, 8);
    pss->fds = static_cast<grpc_fd**>(
        gpr_realloc(pss->fds, pss->fd_capacity * sizeof(*pss->fds)));
  }
  REF_BY(fd, 2, "pollset_set");
  pss->fds[pss->fd_count++] = fd;
  gpr_mu_unlock(&pss->mu);

  GRPC_LOG_IF_ERROR(err_desc, error);
}

static void pollset_set_del_fd(grpc_pollset_set* pss, grpc_fd* fd) {
  GPR_TIMER_SCOPE("pollset_set_del_fd", 0);
  if (grpc_polling_trace.enabled()) {
    gpr_log(GPR_INFO, "PSS:%p: del fd %p", pss, fd);
  }
  pss = pss_lock_adam(pss);
  size_t i;
  for (i = 0; i < pss->fd_count; i++) {
    if (pss->fds[i] == fd) {
      UNREF_BY(fd, 2, "pollset_set");
      break;
    }
  }
  GPR_ASSERT(i != pss->fd_count);
  for (; i < pss->fd_count - 1; i++) {
    pss->fds[i] = pss->fds[i + 1];
  }
  pss->fd_count--;
  gpr_mu_unlock(&pss->mu);
}

static void pollset_set_del_pollset(grpc_pollset_set* pss, grpc_pollset* ps) {
  GPR_TIMER_SCOPE("pollset_set_del_pollset", 0);
  if (grpc_polling_trace.enabled()) {
    gpr_log(GPR_INFO, "PSS:%p: del pollset %p", pss, ps);
  }
  pss = pss_lock_adam(pss);
  size_t i;
  for (i = 0; i < pss->pollset_count; i++) {
    if (pss->pollsets[i] == ps) {
      break;
    }
  }
  GPR_ASSERT(i != pss->pollset_count);
  for (; i < pss->pollset_count - 1; i++) {
    pss->pollsets[i] = pss->pollsets[i + 1];
  }
  pss->pollset_count--;
  gpr_mu_unlock(&pss->mu);
  gpr_mu_lock(&ps->mu);
  if (0 == --ps->containing_pollset_set_count) {
    pollset_maybe_finish_shutdown(ps);
  }
  gpr_mu_unlock(&ps->mu);
}

// add all fds to pollables, and output a new array of unorphaned out_fds
// assumes pollsets are multipollable
static grpc_error* add_fds_to_pollsets(grpc_fd** fds, size_t fd_count,
                                       grpc_pollset** pollsets,
                                       size_t pollset_count,
                                       const char* err_desc, grpc_fd** out_fds,
                                       size_t* out_fd_count) {
  GPR_TIMER_SCOPE("add_fds_to_pollsets", 0);
  grpc_error* error = GRPC_ERROR_NONE;
  for (size_t i = 0; i < fd_count; i++) {
    gpr_mu_lock(&fds[i]->orphan_mu);
    if ((gpr_atm_no_barrier_load(&fds[i]->refst) & 1) == 0) {
      gpr_mu_unlock(&fds[i]->orphan_mu);
      UNREF_BY(fds[i], 2, "pollset_set");
    } else {
      for (size_t j = 0; j < pollset_count; j++) {
        append_error(&error,
                     pollable_add_fd(pollsets[j]->active_pollable, fds[i]),
                     err_desc);
      }
      gpr_mu_unlock(&fds[i]->orphan_mu);
      out_fds[(*out_fd_count)++] = fds[i];
    }
  }
  return error;
}

static void pollset_set_add_pollset(grpc_pollset_set* pss, grpc_pollset* ps) {
  GPR_TIMER_SCOPE("pollset_set_add_pollset", 0);
  if (grpc_polling_trace.enabled()) {
    gpr_log(GPR_INFO, "PSS:%p: add pollset %p", pss, ps);
  }
  grpc_error* error = GRPC_ERROR_NONE;
  static const char* err_desc = "pollset_set_add_pollset";
  pollable* pollable_obj = nullptr;
  gpr_mu_lock(&ps->mu);
  if (!GRPC_LOG_IF_ERROR(err_desc,
                         pollset_as_multipollable_locked(ps, &pollable_obj))) {
    GPR_ASSERT(pollable_obj == nullptr);
    gpr_mu_unlock(&ps->mu);
    return;
  }
  ps->containing_pollset_set_count++;
  gpr_mu_unlock(&ps->mu);
  pss = pss_lock_adam(pss);
  size_t initial_fd_count = pss->fd_count;
  pss->fd_count = 0;
  append_error(&error,
               add_fds_to_pollsets(pss->fds, initial_fd_count, &ps, 1, err_desc,
                                   pss->fds, &pss->fd_count),
               err_desc);
  if (pss->pollset_count == pss->pollset_capacity) {
    pss->pollset_capacity = GPR_MAX(pss->pollset_capacity * 2, 8);
    pss->pollsets = static_cast<grpc_pollset**>(gpr_realloc(
        pss->pollsets, pss->pollset_capacity * sizeof(*pss->pollsets)));
  }
  pss->pollsets[pss->pollset_count++] = ps;
  gpr_mu_unlock(&pss->mu);
  POLLABLE_UNREF(pollable_obj, "pollset_set");

  GRPC_LOG_IF_ERROR(err_desc, error);
}

static void pollset_set_add_pollset_set(grpc_pollset_set* a,
                                        grpc_pollset_set* b) {
  GPR_TIMER_SCOPE("pollset_set_add_pollset_set", 0);
  if (grpc_polling_trace.enabled()) {
    gpr_log(GPR_INFO, "PSS: merge (%p, %p)", a, b);
  }
  grpc_error* error = GRPC_ERROR_NONE;
  static const char* err_desc = "pollset_set_add_fd";
  for (;;) {
    if (a == b) {
      // pollset ancestors are the same: nothing to do
      return;
    }
    if (a > b) {
      GPR_SWAP(grpc_pollset_set*, a, b);
    }
    gpr_mu* a_mu = &a->mu;
    gpr_mu* b_mu = &b->mu;
    gpr_mu_lock(a_mu);
    gpr_mu_lock(b_mu);
    if (a->parent != nullptr) {
      a = a->parent;
    } else if (b->parent != nullptr) {
      b = b->parent;
    } else {
      break;  // exit loop, both pollsets locked
    }
    gpr_mu_unlock(a_mu);
    gpr_mu_unlock(b_mu);
  }
  // try to do the least copying possible
  // TODO(sreek): there's probably a better heuristic here
  const size_t a_size = a->fd_count + a->pollset_count;
  const size_t b_size = b->fd_count + b->pollset_count;
  if (b_size > a_size) {
    GPR_SWAP(grpc_pollset_set*, a, b);
  }
  if (grpc_polling_trace.enabled()) {
    gpr_log(GPR_INFO, "PSS: parent %p to %p", b, a);
  }
  gpr_ref(&a->refs);
  b->parent = a;
  if (a->fd_capacity < a->fd_count + b->fd_count) {
    a->fd_capacity = GPR_MAX(2 * a->fd_capacity, a->fd_count + b->fd_count);
    a->fds = static_cast<grpc_fd**>(
        gpr_realloc(a->fds, a->fd_capacity * sizeof(*a->fds)));
  }
  size_t initial_a_fd_count = a->fd_count;
  a->fd_count = 0;
  append_error(
      &error,
      add_fds_to_pollsets(a->fds, initial_a_fd_count, b->pollsets,
                          b->pollset_count, "merge_a2b", a->fds, &a->fd_count),
      err_desc);
  append_error(
      &error,
      add_fds_to_pollsets(b->fds, b->fd_count, a->pollsets, a->pollset_count,
                          "merge_b2a", a->fds, &a->fd_count),
      err_desc);
  if (a->pollset_capacity < a->pollset_count + b->pollset_count) {
    a->pollset_capacity =
        GPR_MAX(2 * a->pollset_capacity, a->pollset_count + b->pollset_count);
    a->pollsets = static_cast<grpc_pollset**>(
        gpr_realloc(a->pollsets, a->pollset_capacity * sizeof(*a->pollsets)));
  }
  if (b->pollset_count > 0) {
    memcpy(a->pollsets + a->pollset_count, b->pollsets,
           b->pollset_count * sizeof(*b->pollsets));
  }
  a->pollset_count += b->pollset_count;
  gpr_free(b->fds);
  gpr_free(b->pollsets);
  b->fds = nullptr;
  b->pollsets = nullptr;
  b->fd_count = b->fd_capacity = b->pollset_count = b->pollset_capacity = 0;
  gpr_mu_unlock(&a->mu);
  gpr_mu_unlock(&b->mu);
}

static void pollset_set_del_pollset_set(grpc_pollset_set* bag,
                                        grpc_pollset_set* item) {}

/*******************************************************************************
 * Event engine binding
 */

static bool is_any_background_poller_thread(void) { return false; }

static void shutdown_background_closure(void) {}

static void shutdown_engine(void) {
  fd_global_shutdown();
  pollset_global_shutdown();
}

static const grpc_event_engine_vtable vtable = {
    sizeof(grpc_pollset),
    true,
    false,

    fd_create,
    fd_wrapped_fd,
    fd_orphan,
    fd_shutdown,
    fd_notify_on_read,
    fd_notify_on_write,
    fd_notify_on_error,
    fd_become_readable,
    fd_become_writable,
    fd_has_errors,
    fd_is_shutdown,

    pollset_init,
    pollset_shutdown,
    pollset_destroy,
    pollset_work,
    pollset_kick,
    pollset_add_fd,

    pollset_set_create,
    pollset_set_unref,  // destroy ==> unref 1 public ref
    pollset_set_add_pollset,
    pollset_set_del_pollset,
    pollset_set_add_pollset_set,
    pollset_set_del_pollset_set,
    pollset_set_add_fd,
    pollset_set_del_fd,

    is_any_background_poller_thread,
    shutdown_background_closure,
    shutdown_engine,
};

const grpc_event_engine_vtable* grpc_init_epollex_linux(
    bool explicitly_requested) {
  if (!grpc_has_wakeup_fd()) {
    gpr_log(GPR_ERROR, "Skipping epollex because of no wakeup fd.");
    return nullptr;
  }

  if (!grpc_is_epollexclusive_available()) {
    gpr_log(GPR_INFO, "Skipping epollex because it is not supported.");
    return nullptr;
  }

  fd_global_init();

  if (!GRPC_LOG_IF_ERROR("pollset_global_init", pollset_global_init())) {
    pollset_global_shutdown();
    fd_global_shutdown();
    return nullptr;
  }

  return &vtable;
}

#else /* defined(GRPC_LINUX_EPOLL_CREATE1) */
#if defined(GRPC_POSIX_SOCKET_EV_EPOLLEX)
#include "src/core/lib/iomgr/ev_epollex_linux.h"
/* If GRPC_LINUX_EPOLL_CREATE1 is not defined, it means
   epoll_create1 is not available. Return NULL */
const grpc_event_engine_vtable* grpc_init_epollex_linux(
    bool explicitly_requested) {
  return nullptr;
}
#endif /* defined(GRPC_POSIX_SOCKET_EV_EPOLLEX) */

#endif /* !defined(GRPC_LINUX_EPOLL_CREATE1) */