Move the event handler on macos from poll to kqueue. Simpler and faster.

runtime/bin/eventhandler_macos.cc

 // Copyright (c) 2012, the Dart project authors.  Please see the AUTHORS file
 // for details. All rights reserved. Use of this source code is governed by a
 // BSD-style license that can be found in the LICENSE file.
 
 #include "bin/eventhandler.h"
 
 #include <errno.h>
-#include <poll.h>
 #include <pthread.h>
 #include <stdio.h>
 #include <string.h>
+#include <sys/event.h>
 #include <sys/time.h>
 #include <unistd.h>
 
 #include "bin/dartutils.h"
 #include "bin/fdutils.h"
 #include "bin/hashmap.h"
 #include "platform/utils.h"
 
 
 int64_t GetCurrentTimeMilliseconds() {
   struct timeval tv;
   if (gettimeofday(&tv, NULL) < 0) {
     UNREACHABLE();
     return 0;
   }
   return ((static_cast<int64_t>(tv.tv_sec) * 1000000) + tv.tv_usec) / 1000;
 }
 
 
 static const int kInterruptMessageSize = sizeof(InterruptMessage);
 static const int kInfinityTimeout = -1;
 static const int kTimerId = -1;
 
 
-intptr_t SocketData::GetPollEvents() {
-  // Do not ask for POLLERR and POLLHUP explicitly as they are
-  // triggered anyway.
-  intptr_t events = 0;
-  if (!IsClosedRead()) {
-    if ((mask_ & (1 << kInEvent)) != 0) {
-      events |= POLLIN;
-    }
-  }
-  if (!IsClosedWrite()) {
-    if ((mask_ & (1 << kOutEvent)) != 0) {
-      events |= POLLOUT;
-    }
-  }
-  return events;
+bool SocketData::HasReadEvent() {
+  return !IsClosedRead() && ((mask_ & (1 << kInEvent)) != 0);
 }
 
 
+bool SocketData::HasWriteEvent() {
+  return !IsClosedWrite() && ((mask_ & (1 << kOutEvent)) != 0);
+}
+
+
+// Unregister the file descriptor for a SocketData structure with kqueue.
+static void RemoveFromKqueue(intptr_t kqueue_fd_, SocketData* sd) {

[Søren Gjesse, 2012/02/08 10:40:36]

Maybe add constant kMaxChanges = 2.

[Mads Ager (google), 2012/02/08 12:15:15]

Done.

+  intptr_t changes = 0;
+  struct kevent events[2];
+  if (sd->read_tracked_by_kqueue()) {
+    EV_SET(events + changes, sd->fd(), EVFILT_READ, EV_DELETE, 0, 0, NULL);
+    ++changes;
+    sd->set_read_tracked_by_kqueue(false);
+  }
+  if (sd->write_tracked_by_kqueue()) {
+    EV_SET(events + changes, sd->fd(), EVFILT_WRITE, EV_DELETE, 0, 0, sd);
+    ++changes;
+    sd->set_write_tracked_by_kqueue(false);
+  }
+  if (changes > 0) {

[Søren Gjesse, 2012/02/08 10:40:36]

And ASSERT(changes <= kMaxChanges)

[Mads Ager (google), 2012/02/08 12:15:15]

Done.

+    int status =
+      TEMP_FAILURE_RETRY(kevent(kqueue_fd_, events, changes, NULL, 0, NULL));
+    if (status == -1) {
+      FATAL("Failed deleting events from kqueue");
+    }
+  }
+}
+
+
+// Update the kqueue registration for SocketData structure to reflect
+// the events currently of interest.
+static void UpdateKqueue(intptr_t kqueue_fd_, SocketData* sd) {
+  intptr_t changes = 0;

[Søren Gjesse, 2012/02/08 10:40:36]

Ditto constant.

[Mads Ager (google), 2012/02/08 12:15:15]

Done.

+  struct kevent events[2];
+  if (sd->port() != 0) {
+    // Register or unregister READ filter if needed.
+    if (sd->HasReadEvent()) {

[Søren Gjesse, 2012/02/08 10:40:36]

Is sd->read_tracked_by_kqueue() always false here? If so ASSERT otherwise skip
the EV_SET.

[Mads Ager (google), 2012/02/08 12:15:15]

It is not always false. What is there now is correct because an ADD operation
means 'modify if exists'. However, by checking explicitly we can sometimes avoid
a system call. I have added the explicit checking.

+      EV_SET(events + changes, sd->fd(), EVFILT_READ, EV_ADD, 0, 0, sd);
+      ++changes;
+      sd->set_read_tracked_by_kqueue(true);
+    } else if (sd->read_tracked_by_kqueue()) {
+      EV_SET(events + changes, sd->fd(), EVFILT_READ, EV_DELETE, 0, 0, NULL);
+      ++changes;
+      sd->set_read_tracked_by_kqueue(false);
+    }
+    // Register or unregister WRITE filter if needed.
+    if (sd->HasWriteEvent()) {

[Søren Gjesse, 2012/02/08 10:40:36]

Ditto, is sd->write_tracked_by_kqueue() is always false?

[Mads Ager (google), 2012/02/08 12:15:15]

Same as above. Added explicit checking.

+      EV_SET(events + changes, sd->fd(), EVFILT_WRITE, EV_ADD, 0, 0, sd);
+      ++changes;
+      sd->set_write_tracked_by_kqueue(true);
+    } else if (sd->write_tracked_by_kqueue()) {
+      EV_SET(events + changes, sd->fd(), EVFILT_WRITE, EV_DELETE, 0, 0, NULL);
+      ++changes;
+      sd->set_write_tracked_by_kqueue(false);
+    }
+  }
+  if (changes > 0) {
+    int status =
+      TEMP_FAILURE_RETRY(kevent(kqueue_fd_, events, changes, NULL, 0, NULL));
+    if (status == -1) {
+      FATAL("Failed updating kqueue");
+    }
+  }
+}
+
+
 EventHandlerImplementation::EventHandlerImplementation()
     : socket_map_(&HashMap::SamePointerValue, 16) {
   intptr_t result;
   result = TEMP_FAILURE_RETRY(pipe(interrupt_fds_));
   if (result != 0) {
     FATAL("Pipe creation failed");
   }
   FDUtils::SetNonBlocking(interrupt_fds_[0]);
   timeout_ = kInfinityTimeout;
   timeout_port_ = 0;
+
+  kqueue_fd_ = TEMP_FAILURE_RETRY(kqueue());
+  if (kqueue_fd_ == -1) {
+    FATAL("Failed creating kqueue");
+  }
+  // Register the interrupt_fd with the kqueue.
+  struct kevent event;
+  EV_SET(&event, interrupt_fds_[0], EVFILT_READ, EV_ADD, 0, 0, NULL);
+  int status = TEMP_FAILURE_RETRY(kevent(kqueue_fd_, &event, 1, NULL, 0, NULL));
+  if (status == -1) {
+    FATAL("Failed adding interrupt fd to kqueue");
+  }
 }
 
 
 EventHandlerImplementation::~EventHandlerImplementation() {
   TEMP_FAILURE_RETRY(close(interrupt_fds_[0]));
   TEMP_FAILURE_RETRY(close(interrupt_fds_[1]));
 }
 
 
 SocketData* EventHandlerImplementation::GetSocketData(intptr_t fd) {
   ASSERT(fd >= 0);
   HashMap::Entry* entry = socket_map_.Lookup(
       GetHashmapKeyFromFd(fd), GetHashmapHashFromFd(fd), true);
   ASSERT(entry != NULL);
   SocketData* sd = reinterpret_cast<SocketData*>(entry->value);
   if (sd == NULL) {
-    // If there is no data in the hash map for this file descriptor
-    // then this is inserting a new SocketData for the file descriptor.
+    // If there is no data in the hash map for this file descriptor a
+    // new SocketData for the file descriptor is inserted.
     sd = new SocketData(fd);
     entry->value = sd;
   }
   ASSERT(fd == sd->fd());
   return sd;
 }
 
 
 void EventHandlerImplementation::WakeupHandler(intptr_t id,
                                                Dart_Port dart_port,
                                                int64_t data) {
   InterruptMessage msg;
   msg.id = id;
   msg.dart_port = dart_port;
   msg.data = data;
   intptr_t result =
       FDUtils::WriteToBlocking(interrupt_fds_[1], &msg, kInterruptMessageSize);
   if (result != kInterruptMessageSize) {
     if (result == -1) {
       perror("Interrupt message failure:");
     }
     FATAL1("Interrupt message failure. Wrote %d bytes.", result);
   }
 }
 
 
-struct pollfd* EventHandlerImplementation::GetPollFds(intptr_t* pollfds_size) {
-  struct pollfd* pollfds;
-
-  // Calculate the number of file descriptors to poll on.
-  intptr_t numPollfds = 1;
-  for (HashMap::Entry* entry = socket_map_.Start();
-       entry != NULL;
-       entry = socket_map_.Next(entry)) {
-    SocketData* sd = reinterpret_cast<SocketData*>(entry->value);
-    if (sd->port() > 0 && sd->GetPollEvents() != 0) numPollfds++;
-  }
-
-  pollfds = reinterpret_cast<struct pollfd*>(calloc(sizeof(struct pollfd),
-                                                    numPollfds));
-  pollfds[0].fd = interrupt_fds_[0];
-  pollfds[0].events |= POLLIN;
-
-  int i = 1;
-  for (HashMap::Entry* entry = socket_map_.Start();
-       entry != NULL;
-       entry = socket_map_.Next(entry)) {
-    SocketData* sd = reinterpret_cast<SocketData*>(entry->value);
-    intptr_t events = sd->GetPollEvents();
-    if (sd->port() > 0 && events != 0) {
-      // Fd is added to the poll set.
-      pollfds[i].fd = sd->fd();
-      pollfds[i].events = events;
-      i++;
-    }
-  }
-  ASSERT(numPollfds == i);
-  *pollfds_size = i;
-
-  return pollfds;
-}
-
-
 bool EventHandlerImplementation::GetInterruptMessage(InterruptMessage* msg) {
   int total_read = 0;
   int bytes_read =
       TEMP_FAILURE_RETRY(read(interrupt_fds_[0], msg, kInterruptMessageSize));
   if (bytes_read < 0) {
     return false;
   }
   total_read = bytes_read;
   while (total_read < kInterruptMessageSize) {
     bytes_read = TEMP_FAILURE_RETRY(read(interrupt_fds_[0],
@@ skipping 11 matching lines @@
   while (GetInterruptMessage(&msg)) {
     if (msg.id == kTimerId) {
       timeout_ = msg.data;
       timeout_port_ = msg.dart_port;
     } else {
       SocketData* sd = GetSocketData(msg.id);
       if ((msg.data & (1 << kShutdownReadCommand)) != 0) {
         ASSERT(msg.data == (1 << kShutdownReadCommand));
         // Close the socket for reading.
         sd->ShutdownRead();
+        UpdateKqueue(kqueue_fd_, sd);
       } else if ((msg.data & (1 << kShutdownWriteCommand)) != 0) {
         ASSERT(msg.data == (1 << kShutdownWriteCommand));
         // Close the socket for writing.
         sd->ShutdownWrite();
+        UpdateKqueue(kqueue_fd_, sd);
       } else if ((msg.data & (1 << kCloseCommand)) != 0) {
         ASSERT(msg.data == (1 << kCloseCommand));
         // Close the socket and free system resources.
+        RemoveFromKqueue(kqueue_fd_, sd);
         intptr_t fd = sd->fd();
         sd->Close();
         socket_map_.Remove(GetHashmapKeyFromFd(fd), GetHashmapHashFromFd(fd));
         delete sd;
       } else {
         // Setup events to wait for.
         sd->SetPortAndMask(msg.dart_port, msg.data);
+        UpdateKqueue(kqueue_fd_, sd);
       }
     }
   }
 }
 
-#ifdef DEBUG_POLL
-static void PrintEventMask(struct pollfd* pollfd) {
-  printf("%d ", pollfd->fd);
-  if ((pollfd->revents & POLLIN) != 0) printf("POLLIN ");
-  if ((pollfd->revents & POLLPRI) != 0) printf("POLLPRI ");
-  if ((pollfd->revents & POLLOUT) != 0) printf("POLLOUT ");
-  if ((pollfd->revents & POLLERR) != 0) printf("POLLERR ");
-  if ((pollfd->revents & POLLHUP) != 0) printf("POLLHUP ");
-  if ((pollfd->revents & POLLNVAL) != 0) printf("POLLNVAL ");
-  int all_events = POLLIN | POLLPRI | POLLOUT | POLLERR | POLLHUP | POLLNVAL;
-  if ((pollfd->revents & ~all_events) != 0) {
-    printf("(and %08x) ", pollfd->revents & ~all_events);
-  }
-  printf("(available %d) ", FDUtils::AvailableBytes(pollfd->fd));
 
+#ifdef DEBUG_KQUEUE
+static void PrintEventMask(intptr_t fd, struct kevent* event) {
+  printf("%d ", fd);
+  if (event->filter == EVFILT_READ) printf("EVFILT_READ ");
+  if (event->filter == EVFILT_WRITE) printf("EVFILT_WRITE ");
+  printf("flags: %x: ", event->flags);
+  if ((event->flags & EV_EOF) != 0) printf("EV_EOF ");
+  if ((event->flags & EV_ERROR) != 0) printf("EV_ERROR ");
+  printf("(available %d) ", FDUtils::AvailableBytes(fd));
   printf("\n");
 }
 #endif
 
-intptr_t EventHandlerImplementation::GetPollEvents(struct pollfd* pollfd) {
-#ifdef DEBUG_POLL
-  if (pollfd->fd != interrupt_fds_[0]) PrintEventMask(pollfd);
+
+intptr_t EventHandlerImplementation::GetEvents(struct kevent* event,

[Søren Gjesse, 2012/02/08 10:40:36]

Nice that we don't need special handling of pipes and need recv calls any more.

[Mads Ager (google), 2012/02/08 12:15:15]

Yes, this is so much cleaner. :)

+                                               SocketData* sd) {
+#ifdef DEBUG_KQUEUE
+  PrintEventMask(sd->fd(), event);
 #endif
   intptr_t event_mask = 0;
-  SocketData* sd = GetSocketData(pollfd->fd);
   if (sd->IsListeningSocket()) {
-    // For listening sockets the POLLIN event indicate that there are
-    // connections ready for accept unless accompanied with one of the
-    // other flags.
-    if ((pollfd->revents & POLLIN) != 0) {
-      if ((pollfd->revents & POLLHUP) != 0) event_mask |= (1 << kCloseEvent);
-      if ((pollfd->revents & POLLERR) != 0) event_mask |= (1 << kErrorEvent);
+    // On a listening socket the READ event means that there are
+    // connections ready to be accepted.
+    if (event->filter == EVFILT_READ) {
+      if ((event->flags & EV_EOF) != 0) event_mask |= (1 << kCloseEvent);
+      if ((event->flags & EV_ERROR) != 0) event_mask |= (1 << kErrorEvent);
       if (event_mask == 0) event_mask |= (1 << kInEvent);
     }
   } else {
-    if ((pollfd->revents & POLLNVAL) != 0) {
-      return 0;
-    }
-
     // Prioritize data events over close and error events.
-    if ((pollfd->revents & POLLIN) != 0) {
-      if (FDUtils::AvailableBytes(pollfd->fd) != 0) {
-        event_mask = (1 << kInEvent);
-      } else if (((pollfd->revents & POLLHUP) != 0)) {
+    if (event->filter == EVFILT_READ) {
+      if (FDUtils::AvailableBytes(sd->fd()) != 0) {
+         event_mask = (1 << kInEvent);
+      } else if ((event->flags & EV_EOF) != 0) {
         event_mask = (1 << kCloseEvent);
         sd->MarkClosedRead();
-      } else if ((pollfd->revents & POLLERR) != 0) {
+      } else if ((event->flags & EV_ERROR) != 0) {
         event_mask = (1 << kErrorEvent);
-      } else {
-        if (sd->IsPipe()) {
-          // When reading from stdin (either from a terminal or piped
-          // input) treat POLLIN with 0 available bytes as
-          // end-of-file.
-          if (sd->fd() == STDIN_FILENO) {
-            event_mask = (1 << kCloseEvent);
-            sd->MarkClosedRead();
-          }
-        } else {
-          // If POLLIN is set with no available data and no POLLHUP use
-          // recv to peek for whether the other end of the socket
-          // actually closed.
-          char buffer;
-          ssize_t bytesPeeked =
-              TEMP_FAILURE_RETRY(recv(sd->fd(), &buffer, 1, MSG_PEEK));
-          ASSERT(EAGAIN == EWOULDBLOCK);
-          if (bytesPeeked == 0) {
-            event_mask = (1 << kCloseEvent);
-            sd->MarkClosedRead();
-          } else if (errno != EWOULDBLOCK) {
-            fprintf(stderr, "Error recv: %s\n", strerror(errno));
-          }
-        }
       }
     }
 
-    // On pipes POLLHUP is reported without POLLIN when there is no
-    // more data to read.
-    if (sd->IsPipe()) {
-      if (((pollfd->revents & POLLIN) == 0) &&
-          ((pollfd->revents & POLLHUP) != 0)) {
-        event_mask = (1 << kCloseEvent);
-        sd->MarkClosedRead();
-      }
-    }
-
-    if ((pollfd->revents & POLLOUT) != 0) {
-      if ((pollfd->revents & POLLERR) != 0) {
+    if (event->filter == EVFILT_WRITE) {
+      if ((event->flags & EV_ERROR) != 0) {
         event_mask = (1 << kErrorEvent);
         sd->MarkClosedWrite();
+      } else if ((event->flags & EV_EOF) != 0) {
+        // If the read end closed, close the write end as well,

[Søren Gjesse, 2012/02/08 10:40:36]

Maybe rephrase "If the read end closed, close the write end as well," to
something like "If the receiver closed for reading then close for writing".

[Mads Ager (google), 2012/02/08 12:15:15]

Done.

+        // update the registration with kqueue and do not report
+        // a write event.
+        sd->MarkClosedWrite();
+        UpdateKqueue(kqueue_fd_, sd);
       } else {
         event_mask |= (1 << kOutEvent);
       }
     }
   }
 
   return event_mask;
 }
 
 
-void EventHandlerImplementation::HandleEvents(struct pollfd* pollfds,
-                                              int pollfds_size,
-                                              int result_size) {
-  if ((pollfds[0].revents & POLLIN) != 0) {
-    result_size -= 1;
-  }
-  if (result_size > 0) {
-    for (int i = 1; i < pollfds_size; i++) {
-     /*
-      * The fd is unregistered. It gets re-registered when the request
-      * was handled by dart.
-      */
-      intptr_t event_mask = GetPollEvents(&pollfds[i]);
+void EventHandlerImplementation::HandleEvents(struct kevent* events,
+                                              int size) {
+  for (int i = 0; i < size; i++) {
+    if (events[i].udata != NULL) {
+      SocketData* sd = reinterpret_cast<SocketData*>(events[i].udata);
+      intptr_t event_mask = GetEvents(events + i, sd);
       if (event_mask != 0) {
-        intptr_t fd = pollfds[i].fd;
-        SocketData* sd = GetSocketData(fd);
+        // Unregister events for the file descriptor. Events will be
+        // registered again when the current event has been handled in
+        // Dart code.
+        RemoveFromKqueue(kqueue_fd_, sd);
         Dart_Port port = sd->port();
         ASSERT(port != 0);
-        sd->Unregister();
         DartUtils::PostInt32(port, event_mask);
       }
     }
   }
   HandleInterruptFd();
 }
 
 
 intptr_t EventHandlerImplementation::GetTimeout() {
   if (timeout_ == kInfinityTimeout) {
     return kInfinityTimeout;
   }
   intptr_t millis = timeout_ - GetCurrentTimeMilliseconds();
   return (millis < 0) ? 0 : millis;
 }
 
 
 void EventHandlerImplementation::HandleTimeout() {
   if (timeout_ != kInfinityTimeout) {
     intptr_t millis = timeout_ - GetCurrentTimeMilliseconds();
     if (millis <= 0) {
       DartUtils::PostNull(timeout_port_);
       timeout_ = kInfinityTimeout;
       timeout_port_ = 0;
     }
   }
 }
 
 
-void EventHandlerImplementation::Poll(uword args) {
-  intptr_t pollfds_size;
-  struct pollfd* pollfds;
+void EventHandlerImplementation::EventHandlerEntry(uword args) {
+  static const intptr_t kMaxEvents = 16;
+  struct kevent events[kMaxEvents];
   EventHandlerImplementation* handler =
       reinterpret_cast<EventHandlerImplementation*>(args);
   ASSERT(handler != NULL);
   while (1) {
-    pollfds = handler->GetPollFds(&pollfds_size);
     intptr_t millis = handler->GetTimeout();
-    intptr_t result = TEMP_FAILURE_RETRY(poll(pollfds, pollfds_size, millis));
-    ASSERT(EAGAIN == EWOULDBLOCK);
+    struct timespec ts;
+    int64_t secs = 0;
+    int64_t nanos = 0;
+    if (millis > 0) {
+      secs = millis / 1000;
+      nanos = (millis - (secs * 1000)) * 1000000;
+    }
+    ts.tv_sec = secs;
+    ts.tv_nsec = nanos;
+    intptr_t result = TEMP_FAILURE_RETRY(kevent(handler->kqueue_fd_,
+                                                NULL,
+                                                0,
+                                                events,
+                                                kMaxEvents,
+                                                &ts));
     if (result == -1) {
-      if (errno != EWOULDBLOCK) {
-        perror("Poll failed");
-      }
+      perror("kevent failed");
+      FATAL("kevent failed\n");
     } else {
       handler->HandleTimeout();
-      handler->HandleEvents(pollfds, pollfds_size, result);
+      handler->HandleEvents(events, result);
     }
-    free(pollfds);
   }
 }
 
 
 void EventHandlerImplementation::StartEventHandler() {
-  int result = dart::Thread::Start(&EventHandlerImplementation::Poll,
-                                   reinterpret_cast<uword>(this));
+  int result =
+      dart::Thread::Start(&EventHandlerImplementation::EventHandlerEntry,
+                          reinterpret_cast<uword>(this));
   if (result != 0) {
     FATAL1("Failed to start event handler thread %d", result);
   }
 }
 
 
 void EventHandlerImplementation::SendData(intptr_t id,
                                           Dart_Port dart_port,
                                           intptr_t data) {
   WakeupHandler(id, dart_port, data);
 }
 
 
 void* EventHandlerImplementation::GetHashmapKeyFromFd(intptr_t fd) {
   // The hashmap does not support keys with value 0.
   return reinterpret_cast<void*>(fd + 1);
 }
 
 
 uint32_t EventHandlerImplementation::GetHashmapHashFromFd(intptr_t fd) {
   // The hashmap does not support keys with value 0.
   return dart::Utils::WordHash(fd + 1);
 }