[project @ 2003-06-26 12:22:59 by stolz]
[ghc-hetmet.git] / ghc / rts / Select.c
index 87f3267..677fdd2 100644 (file)
@@ -1,23 +1,26 @@
 /* -----------------------------------------------------------------------------
- * $Id: Select.c,v 1.7 2000/01/13 12:40:16 simonmar Exp $
+ * $Id: Select.c,v 1.29 2003/06/26 12:22:59 stolz Exp $
  *
- * (c) The GHC Team 1995-1999
+ * (c) The GHC Team 1995-2002
  *
  * Support for concurrent non-blocking I/O and thread waiting.
  *
  * ---------------------------------------------------------------------------*/
 
+
 /* we're outside the realms of POSIX here... */
-#define NON_POSIX_SOURCE
+/* #include "PosixSource.h" */
 
 #include "Rts.h"
 #include "Schedule.h"
 #include "RtsUtils.h"
 #include "RtsFlags.h"
+#include "Timer.h"
 #include "Itimer.h"
 #include "Signals.h"
+#include "Capability.h"
 
-# if defined(HAVE_SYS_TYPES_H)
+# ifdef HAVE_SYS_TYPES_H
 #  include <sys/types.h>
 # endif
 
 #  include <sys/time.h>
 # endif
 
-nat ticks_since_select = 0;
+#include <errno.h>
+#include <string.h>
+
+#ifdef HAVE_UNISTD_H
+#include <unistd.h>
+#endif
+
+/* last timestamp */
+nat timestamp = 0;
+
+#ifdef RTS_SUPPORTS_THREADS
+static rtsBool isWorkerBlockedInAwaitEvent = rtsFalse;
+static rtsBool workerWakeupPending = rtsFalse;
+static int workerWakeupPipe[2];
+static rtsBool workerWakeupInited = rtsFalse;
+#endif
+
+/* There's a clever trick here to avoid problems when the time wraps
+ * around.  Since our maximum delay is smaller than 31 bits of ticks
+ * (it's actually 31 bits of microseconds), we can safely check
+ * whether a timer has expired even if our timer will wrap around
+ * before the target is reached, using the following formula:
+ *
+ *        (int)((uint)current_time - (uint)target_time) < 0
+ *
+ * if this is true, then our time has expired.
+ * (idea due to Andy Gill).
+ */
+rtsBool
+wakeUpSleepingThreads(nat ticks)
+{
+    StgTSO *tso;
+    rtsBool flag = rtsFalse;
+
+    while (sleeping_queue != END_TSO_QUEUE &&
+          (int)(ticks - sleeping_queue->block_info.target) > 0) {
+       tso = sleeping_queue;
+       sleeping_queue = tso->link;
+       tso->why_blocked = NotBlocked;
+       tso->link = END_TSO_QUEUE;
+       IF_DEBUG(scheduler,belch("Waking up sleeping thread %d\n", tso->id));
+       PUSH_ON_RUN_QUEUE(tso);
+       flag = rtsTrue;
+    }
+    return flag;
+}
 
 /* Argument 'wait' says whether to wait for I/O to become available,
  * or whether to just check and return immediately.  If there are
@@ -33,206 +81,274 @@ nat ticks_since_select = 0;
  * otherwise we wait (see Schedule.c).
  *
  * SMP note: must be called with sched_mutex locked.
+ *
+ * Windows: select only works on sockets, so this doesn't really work,
+ * though it makes things better than before. MsgWaitForMultipleObjects
+ * should really be used, though it only seems to work for read handles,
+ * not write handles.
+ *
  */
 void
 awaitEvent(rtsBool wait)
 {
-#ifdef mingw32_TARGET_OS
-/*
- * Win32 doesn't support select(). ToDo: use MsgWaitForMultipleObjects()
- * to achieve (similar) effect.
- *
- */
-    return;
-#else
-
     StgTSO *tso, *prev, *next;
     rtsBool ready;
     fd_set rfd,wfd;
     int numFound;
-    nat min, delta;
     int maxfd = -1;
-   
+    rtsBool select_succeeded = rtsTrue;
+    rtsBool unblock_all = rtsFalse;
     struct timeval tv;
-#ifndef linux_TARGET_OS
-    struct timeval tv_before,tv_after;
-#endif
+    lnat min, ticks;
 
-    IF_DEBUG(scheduler,belch("Checking for threads blocked on I/O...\n"));
+    tv.tv_sec  = 0;
+    tv.tv_usec = 0;
+    
+    IF_DEBUG(scheduler,
+            belch("scheduler: checking for threads blocked on I/O");
+            if (wait) {
+                belch(" (waiting)");
+            }
+            belch("\n");
+            );
 
-    /* see how long it's been since we last checked the blocked queue.
-     * ToDo: make this check atomic, so we don't lose any ticks.
+    /* loop until we've woken up some threads.  This loop is needed
+     * because the select timing isn't accurate, we sometimes sleep
+     * for a while but not long enough to wake up a thread in
+     * a threadDelay.
      */
-    delta = ticks_since_select;
-    ticks_since_select = 0;
-    delta = delta * TICK_MILLISECS * 1000;
-
-    min = wait == rtsTrue ? 0x7fffffff : 0;
-
-    /* 
-     * Collect all of the fd's that we're interested in, and capture
-     * the minimum waiting time (in microseconds) for the delayed threads.
-     */
-    FD_ZERO(&rfd);
-    FD_ZERO(&wfd);
-
-    for(tso = blocked_queue_hd; tso != END_TSO_QUEUE; tso = next) {
-      next = tso->link;
-
-      switch (tso->why_blocked) {
-      case BlockedOnRead:
-       { 
-         int fd = tso->block_info.fd;
-         maxfd = (fd > maxfd) ? fd : maxfd;
-         FD_SET(fd, &rfd);
-         continue;
-       }
+    do {
 
-      case BlockedOnWrite:
-       { 
-         int fd = tso->block_info.fd;
-         maxfd = (fd > maxfd) ? fd : maxfd;
-         FD_SET(fd, &wfd);
-         continue;
-       }
-
-      case BlockedOnDelay:
-       {
-         if (tso->block_info.delay < min)
-           min = tso->block_info.delay;
-         continue;
-       }
+      ticks = timestamp = getourtimeofday();
+      if (wakeUpSleepingThreads(ticks)) { 
+         return;
+      }
 
-      default:
-       barf("AwaitEvent");
+      if (!wait) {
+         min = 0;
+      } else if (sleeping_queue != END_TSO_QUEUE) {
+         min = (sleeping_queue->block_info.target - ticks) 
+             * TICK_MILLISECS * 1000;
+      } else {
+         min = 0x7ffffff;
       }
-    }
 
-    /* Release the scheduler lock while we do the poll.
-     * this means that someone might muck with the blocked_queue
-     * while we do this, but it shouldn't matter:
-     *
-     *   - another task might poll for I/O and remove one
-     *     or more threads from the blocked_queue.
-     *   - more I/O threads may be added to blocked_queue.
-     *   - more delayed threads may be added to blocked_queue. We'll
-     *     just subtract delta from their delays after the poll.
-     *
-     * I believe none of these cases lead to trouble --SDM.
-     */
-    RELEASE_LOCK(&sched_mutex);
+      /* 
+       * Collect all of the fd's that we're interested in
+       */
+      FD_ZERO(&rfd);
+      FD_ZERO(&wfd);
 
-    /* Check for any interesting events */
+      for(tso = blocked_queue_hd; tso != END_TSO_QUEUE; tso = next) {
+       next = tso->link;
 
-    tv.tv_sec = min / 1000000;
-    tv.tv_usec = min % 1000000;
+       switch (tso->why_blocked) {
+       case BlockedOnRead:
+         { 
+           int fd = tso->block_info.fd;
+           maxfd = (fd > maxfd) ? fd : maxfd;
+           FD_SET(fd, &rfd);
+           continue;
+         }
 
-#ifndef linux_TARGET_OS
-    gettimeofday(&tv_before, (struct timezone *) NULL);
-#endif
+       case BlockedOnWrite:
+         { 
+           int fd = tso->block_info.fd;
+           maxfd = (fd > maxfd) ? fd : maxfd;
+           FD_SET(fd, &wfd);
+           continue;
+         }
 
-    while (!interrupted &&
-          (numFound = select(maxfd+1, &rfd, &wfd, NULL, &tv)) < 0) {
-      if (errno != EINTR) {
-       /* fflush(stdout); */
-       perror("select");
-       barf("select failed");
+       default:
+         barf("AwaitEvent");
+       }
       }
-      ACQUIRE_LOCK(&sched_mutex);
 
-      /* We got a signal; could be one of ours.  If so, we need
-       * to start up the signal handler straight away, otherwise
-       * we could block for a long time before the signal is
-       * serviced.
-       */
-      if (signals_pending()) {
-       start_signal_handlers();
-       RELEASE_LOCK(&sched_mutex);
-       break;
+#ifdef RTS_SUPPORTS_THREADS
+      if(!workerWakeupInited) {
+          pipe(workerWakeupPipe);
+          workerWakeupInited = rtsTrue;
       }
-
-      /* If new runnable threads have arrived, stop waiting for
-       * I/O and run them.
+      FD_SET(workerWakeupPipe[0], &rfd);
+      maxfd = workerWakeupPipe[0] > maxfd ? workerWakeupPipe[0] : maxfd;
+#endif
+      
+      /* Release the scheduler lock while we do the poll.
+       * this means that someone might muck with the blocked_queue
+       * while we do this, but it shouldn't matter:
+       *
+       *   - another task might poll for I/O and remove one
+       *     or more threads from the blocked_queue.
+       *   - more I/O threads may be added to blocked_queue.
+       *   - more delayed threads may be added to blocked_queue. We'll
+       *     just subtract delta from their delays after the poll.
+       *
+       * I believe none of these cases lead to trouble --SDM.
        */
-      if (run_queue_hd != END_TSO_QUEUE) {
-       RELEASE_LOCK(&sched_mutex);
-       break;
-      }
-
+      
+#ifdef RTS_SUPPORTS_THREADS
+      isWorkerBlockedInAwaitEvent = rtsTrue;
+      workerWakeupPending = rtsFalse;
+#endif
       RELEASE_LOCK(&sched_mutex);
-    }  
 
-#ifdef linux_TARGET_OS
-    /* on Linux, tv is set to indicate the amount of time not
-     * slept, so we don't need to gettimeofday() to find out.
-     */
-    delta += min - (tv.tv_sec * 1000000 + tv.tv_usec);
-#else
-    gettimeofday(&tv_after, (struct timezone *) NULL);
-    delta += (tv_after.tv_sec - tv_before.tv_sec) * 1000000 +
-      tv_after.tv_usec - tv_before.tv_usec;
+      /* Check for any interesting events */
+      
+      tv.tv_sec  = min / 1000000;
+      tv.tv_usec = min % 1000000;
+
+      while ((numFound = select(maxfd+1, &rfd, &wfd, NULL, &tv)) < 0) {
+         if (errno != EINTR) {
+           /* Handle bad file descriptors by unblocking all the
+              waiting threads. Why? Because a thread might have been
+              a bit naughty and closed a file descriptor while another
+              was blocked waiting. This is less-than-good programming
+              practice, but having the RTS as a result fall over isn't
+              acceptable, so we simply unblock all the waiting threads
+              should we see a bad file descriptor & give the threads
+              a chance to clean up their act. 
+              
+              Note: assume here that threads becoming unblocked
+              will try to read/write the file descriptor before trying
+              to issue a threadWaitRead/threadWaitWrite again (==> an
+              IOError will result for the thread that's got the bad
+              file descriptor.) Hence, there's no danger of a bad
+              file descriptor being repeatedly select()'ed on, so
+              the RTS won't loop.
+           */
+           if ( errno == EBADF ) {
+             unblock_all = rtsTrue;
+             break;
+           } else {
+             fprintf(stderr,"%d\n", errno);
+             fflush(stderr);
+             perror("select");
+             barf("select failed");
+           }
+         }
+         ACQUIRE_LOCK(&sched_mutex);
+#ifdef RTS_SUPPORTS_THREADS
+          isWorkerBlockedInAwaitEvent = rtsFalse;
 #endif
 
-#if 0
-    if (delta != 0) { fprintf(stderr,"waited: %d %d %d\n", min, delta,
-                             interrupted); }
+         /* We got a signal; could be one of ours.  If so, we need
+          * to start up the signal handler straight away, otherwise
+          * we could block for a long time before the signal is
+          * serviced.
+          */
+#if defined(RTS_USER_SIGNALS)
+         if (signals_pending()) {
+             RELEASE_LOCK(&sched_mutex); /* ToDo: kill */
+             startSignalHandlers();
+             ACQUIRE_LOCK(&sched_mutex);
+             return; /* still hold the lock */
+         }
 #endif
 
-    ACQUIRE_LOCK(&sched_mutex);
+         /* we were interrupted, return to the scheduler immediately.
+          */
+         if (interrupted) {
+             return; /* still hold the lock */
+         }
+         
+         /* check for threads that need waking up 
+          */
+         wakeUpSleepingThreads(getourtimeofday());
+         
+         /* If new runnable threads have arrived, stop waiting for
+          * I/O and run them.
+          */
+         if (run_queue_hd != END_TSO_QUEUE) {
+             return; /* still hold the lock */
+         }
+         
+#ifdef RTS_SUPPORTS_THREADS
+         /* If another worker thread wants to take over,
+          * return to the scheduler
+          */
+         if (needToYieldToReturningWorker()) {
+             return; /* still hold the lock */
+         }
+#endif
+         
+#ifdef RTS_SUPPORTS_THREADS
+          isWorkerBlockedInAwaitEvent = rtsTrue;
+#endif
+         RELEASE_LOCK(&sched_mutex);
+      }
 
-    /*
-      Step through the waiting queue, unblocking every thread that now has
-      a file descriptor in a ready state.
+      ACQUIRE_LOCK(&sched_mutex);
 
-      For the delayed threads, decrement the number of microsecs
-      we've been blocked for. Unblock the threads that have thusly expired.
-     */
+      /* Step through the waiting queue, unblocking every thread that now has
+       * a file descriptor in a ready state.
+       */
 
-    prev = NULL;
-    for(tso = blocked_queue_hd; tso != END_TSO_QUEUE; tso = next) {
-      next = tso->link;
-      switch (tso->why_blocked) {
-      case BlockedOnRead:
-       ready = FD_ISSET(tso->block_info.fd, &rfd);
-       break;
-       
-      case BlockedOnWrite:
-       ready = FD_ISSET(tso->block_info.fd, &wfd);
-       break;
-       
-      case BlockedOnDelay:
-       if (tso->block_info.delay > delta) {
-         tso->block_info.delay -= delta;
-         ready = 0;
-       } else {
-         tso->block_info.delay = 0;
-         ready = 1;
-       }
-       break;
-       
-      default:
-       barf("awaitEvent");
+      prev = NULL;
+      if (select_succeeded || unblock_all) {
+         for(tso = blocked_queue_hd; tso != END_TSO_QUEUE; tso = next) {
+             next = tso->link;
+             switch (tso->why_blocked) {
+             case BlockedOnRead:
+                 ready = unblock_all || FD_ISSET(tso->block_info.fd, &rfd);
+                 break;
+             case BlockedOnWrite:
+                 ready = unblock_all || FD_ISSET(tso->block_info.fd, &wfd);
+                 break;
+             default:
+                 barf("awaitEvent");
+             }
+      
+             if (ready) {
+                 IF_DEBUG(scheduler,belch("Waking up blocked thread %d\n", tso->id));
+                 tso->why_blocked = NotBlocked;
+                 tso->link = END_TSO_QUEUE;
+                 PUSH_ON_RUN_QUEUE(tso);
+             } else {
+                 if (prev == NULL)
+                     blocked_queue_hd = tso;
+                 else
+                     prev->link = tso;
+                 prev = tso;
+             }
+         }
+
+         if (prev == NULL)
+             blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
+         else {
+             prev->link = END_TSO_QUEUE;
+             blocked_queue_tl = prev;
+         }
       }
       
-      if (ready) {
-       IF_DEBUG(scheduler,belch("Waking up thread %d\n", tso->id));
-       tso->why_blocked = NotBlocked;
-       tso->link = END_TSO_QUEUE;
-       PUSH_ON_RUN_QUEUE(tso);
-      } else {
-       if (prev == NULL)
-         blocked_queue_hd = tso;
-       else
-         prev->link = tso;
-       prev = tso;
+#if defined(RTS_SUPPORTS_THREADS)
+       // if we were woken up by wakeBlockedWorkerThread,
+       // read the dummy byte from the pipe
+      if(select_succeeded && FD_ISSET(workerWakeupPipe[0], &rfd)) {
+          unsigned char dummy;
+          wait = rtsFalse;
+          read(workerWakeupPipe[0],&dummy,1);
       }
-    }
+#endif
+    } while (wait && !interrupted && run_queue_hd == END_TSO_QUEUE);
+}
+
 
-    if (prev == NULL)
-      blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE;
-    else {
-      prev->link = END_TSO_QUEUE;
-      blocked_queue_tl = prev;
+#ifdef RTS_SUPPORTS_THREADS
+/* wakeBlockedWorkerThread
+ *
+ * If a worker thread is currently blocked within awaitEvent,
+ * wake it.
+ * Must be called with sched_mutex held.
+ */
+void
+wakeBlockedWorkerThread()
+{
+    if(isWorkerBlockedInAwaitEvent && !workerWakeupPending) {
+       unsigned char dummy = 42;       // Any value will do here
+       
+                       // write something so that select() wakes up
+       write(workerWakeupPipe[1],&dummy,1);
+       workerWakeupPending = rtsTrue;
     }
-#endif
 }
+#endif