src/core/lib/iomgr/ev_epoll1_linux.c

/*
 *
 * Copyright 2016, Google Inc.
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are
 * met:
 *
 *     * Redistributions of source code must retain the above copyright
 * notice, this list of conditions and the following disclaimer.
 *     * Redistributions in binary form must reproduce the above
 * copyright notice, this list of conditions and the following disclaimer
 * in the documentation and/or other materials provided with the
 * distribution.
 *     * Neither the name of Google Inc. nor the names of its
 * contributors may be used to endorse or promote products derived from
 * this software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *
 */

#include "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_linux.h"

#include <assert.h>
#include <errno.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/log.h>
#include <grpc/support/string_util.h>
#include <grpc/support/tls.h>
#include <grpc/support/useful.h>

#include "src/core/lib/iomgr/ev_posix.h"
#include "src/core/lib/iomgr/iomgr_internal.h"
#include "src/core/lib/iomgr/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: 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;
static int g_epfd;

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

struct grpc_fd {
  int fd;

  /* 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;
};

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

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

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

typedef enum {
  PWL_POLLSET,
  PWL_POLLABLE,
  POLLSET_WORKER_LINK_COUNT
} pollset_worker_links;

struct grpc_pollset_worker {
  bool kicked;
  bool initialized_cv;
  pollset_worker_link links[POLLSET_WORKER_LINK_COUNT];
  gpr_cv cv;
};

struct grpc_pollset {
  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 */
};

/*******************************************************************************
 * Pollset-set Declarations
 */
struct grpc_pollset_set {};

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

/* 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->po.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->po.mu);
  }

  /* Note: It is not really needed to get the new_fd->po.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->po.mu);
  new_fd->po.pi = NULL;
#ifdef PO_DEBUG
  new_fd->po.obj_type = POLL_OBJ_FD;
#endif

  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->po.mu);

  char *fd_name;
  gpr_asprintf(&fd_name, "%s fd=%d", name, fd);
  grpc_iomgr_register_object(&new_fd->iomgr_object, fd_name);
#ifdef GRPC_FD_REF_COUNT_DEBUG
  gpr_log(GPR_DEBUG, "FD %d %p create %s", fd, (void *)new_fd, fd_name);
#endif
  gpr_free(fd_name);
  return new_fd;
}

static int fd_wrapped_fd(grpc_fd *fd) {
  int ret_fd = -1;
  gpr_mu_lock(&fd->po.mu);
  if (!fd->orphaned) {
    ret_fd = fd->fd;
  }
  gpr_mu_unlock(&fd->po.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->po.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->po.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->po.pi to NULL (but remove the ref on the polling island
       before doing this.) */
  if (fd->po.pi != NULL) {
    polling_island *pi_latest = polling_island_lock(fd->po.pi);
    polling_island_remove_fd_locked(pi_latest, fd, is_fd_closed, &error);
    gpr_mu_unlock(&pi_latest->mu);

    unref_pi = fd->po.pi;
    fd->po.pi = NULL;
  }

  grpc_closure_sched(exec_ctx, fd->on_done_closure, GRPC_ERROR_REF(error));

  gpr_mu_unlock(&fd->po.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) {
  gpr_mu_lock(&fd->po.mu);
  grpc_workqueue *workqueue =
      GRPC_WORKQUEUE_REF((grpc_workqueue *)fd->po.pi, "fd_get_workqueue");
  gpr_mu_unlock(&fd->po.mu);
  return workqueue;
}

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

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

/* Return true if last in list */
typedef enum { EMPTIED, NEW_ROOT, REMOVED } worker_remove_result;

static worker_remove_result worker_remove(grpc_pollset_worker **root,
                                          pollset_worker_links link,
                                          grpc_pollset_worker *worker) {
  if (worker == *root) {
    if (worker == worker->links[link].next) {
      *root = NULL;
      return EMPTIED;
    } else {
      *root = 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 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 REMOVED;
  }
}

GPR_TLS_DECL(g_current_thread_pollset);
GPR_TLS_DECL(g_current_thread_worker);

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

static grpc_pollset_set *pollset_set_create(void) {
  return (grpc_pollset_set *)((intptr_t)0xdeafbeef);
}

static void pollset_set_destroy(grpc_exec_ctx *exec_ctx,
                                grpc_pollset_set *pss) {}

static void pollset_set_add_fd(grpc_exec_ctx *exec_ctx, grpc_pollset_set *pss,
                               grpc_fd *fd) {}

static void pollset_set_del_fd(grpc_exec_ctx *exec_ctx, grpc_pollset_set *pss,
                               grpc_fd *fd) {}

static void pollset_set_add_pollset(grpc_exec_ctx *exec_ctx,
                                    grpc_pollset_set *pss, grpc_pollset *ps) {}

static void pollset_set_del_pollset(grpc_exec_ctx *exec_ctx,
                                    grpc_pollset_set *pss, grpc_pollset *ps) {}

static void pollset_set_add_pollset_set(grpc_exec_ctx *exec_ctx,
                                        grpc_pollset_set *bag,
                                        grpc_pollset_set *item) {}

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

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

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

static const grpc_event_engine_vtable vtable = {
    .pollset_size = sizeof(grpc_pollset),

    .fd_create = fd_create,
    .fd_wrapped_fd = fd_wrapped_fd,
    .fd_orphan = fd_orphan,
    .fd_shutdown = fd_shutdown,
    .fd_is_shutdown = fd_is_shutdown,
    .fd_notify_on_read = fd_notify_on_read,
    .fd_notify_on_write = fd_notify_on_write,
    .fd_get_read_notifier_pollset = fd_get_read_notifier_pollset,
    .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,

    .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,
};

/* 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_epoll1_linux(void) {
  /* If use of signals is disabled, we cannot use epoll engine*/
  if (is_grpc_wakeup_signal_initialized && grpc_wakeup_signal < 0) {
    return NULL;
  }

  if (!grpc_has_wakeup_fd()) {
    return NULL;
  }

  if (!is_epoll_available()) {
    return NULL;
  }

  if (!is_grpc_wakeup_signal_initialized) {
    grpc_use_signal(SIGRTMIN + 6);
  }

  fd_global_init();

  if (!GRPC_LOG_IF_ERROR("pollset_global_init", pollset_global_init())) {
    return NULL;
  }

  if (!GRPC_LOG_IF_ERROR("polling_island_global_init",
                         polling_island_global_init())) {
    return NULL;
  }

  return &vtable;
}

#else /* defined(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_epoll1_linux(void) { return NULL; }
#endif /* defined(GRPC_POSIX_SOCKET) */
#endif /* !defined(GRPC_LINUX_EPOLL) */