[project @ 2002-11-21 17:54:17 by simonpj]

[ghc-hetmet.git] / ghc / rts / Select.c

diff --git a/ghc/rts/Select.c b/ghc/rts/Select.c
index 0408772..dc19cbf 100644 (file)
--- a/ghc/rts/Select.c
+++ b/ghc/rts/Select.c
@@ -1,7 +1,7 @@
 /* -----------------------------------------------------------------------------
- * $Id: Select.c,v 1.19 2001/11/13 13:38:02 simonmar Exp $
+ * $Id: Select.c,v 1.22 2002/07/24 03:38:58 sof Exp $
  *
- * (c) The GHC Team 1995-1999
+ * (c) The GHC Team 1995-2002
  *
  * Support for concurrent non-blocking I/O and thread waiting.
  *
@@ -29,6 +29,9 @@
 #  include <windows.h>
 # endif
 
+#include <errno.h>
+#include <string.h>
+
 /* last timestamp */
 nat timestamp = 0;
 
@@ -86,12 +89,13 @@ awaitEvent(rtsBool wait)
     int maxfd = -1;
 #endif
     rtsBool select_succeeded = rtsTrue;
+    rtsBool unblock_all = rtsFalse;
     struct timeval tv;
     lnat min, ticks;
 
     tv.tv_sec  = 0;
     tv.tv_usec = 0;
-
+    
     IF_DEBUG(scheduler,
             belch("scheduler: checking for threads blocked on I/O");
             if (wait) {
@@ -174,11 +178,32 @@ awaitEvent(rtsBool wait)
 
       while ((numFound = select(maxfd+1, &rfd, &wfd, NULL, &tv)) < 0) {
          if (errno != EINTR) {
-
-             printf("%d\n", errno);
-             fflush(stdout);
-             perror("select");
+           /* 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");
+           }
          }
 #else /* on mingwin */
       while (1) {
@@ -227,15 +252,15 @@ awaitEvent(rtsBool wait)
        */
 
       prev = NULL;
-      if (select_succeeded) {
+      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 = FD_ISSET(tso->block_info.fd, &rfd);
+                 ready = unblock_all || FD_ISSET(tso->block_info.fd, &rfd);
                  break;
              case BlockedOnWrite:
-                 ready = FD_ISSET(tso->block_info.fd, &wfd);
+                 ready = unblock_all || FD_ISSET(tso->block_info.fd, &wfd);
                  break;
              default:
                  barf("awaitEvent");