desktopEvents: Leave libICE rug firmly under libSM.

While the libICE spec's section on error handling suggests applications
close libICE connections in response to I/O errors, libSM (which
sits atop libICE) continues to refer to such deceased libICE
connections, and doing so during shutdown leads to an app crash.
(libSM should've registered an I/O error handler of its own which would
run before the application's, but it doesn't.  Oh well.)

To work around this, we'll detach the ICE connection from our
application event loop but leave its handle alone.

Signed-off-by: Dmitry Torokhov <dtor@vmware.com>

diff --git a/open-vm-tools/services/plugins/desktopEvents/sessionMgr.c b/open-vm-tools/services/plugins/desktopEvents/sessionMgr.c
index 1077c02d4df6cf1ee07d9809cfa5cdf9e9ec4ec7..3792591ab026aa8bca637af62ed2cec46e23a5b9 100644 (file)
--- a/open-vm-tools/services/plugins/desktopEvents/sessionMgr.c
+++ b/open-vm-tools/services/plugins/desktopEvents/sessionMgr.c
@@ -40,6 +40,33 @@
  *       but errors handled by libICE's default may exit().
  */
 
+/*
+ * PR 957938 - Handling libICE I/O Errors
+ *
+ * “Before the application I/O error handler is invoked, protocol libraries
+ * that were interested in being notified of I/O errors will have their Ice-
+ * IOErrorProc handlers invoked.
+ *
+ * “[...]
+ *
+ * “There are two ways of handling IO errors in ICElib: [...] The next time
+ * IceProcessMessages is called it will return a status of IceProcessMessages-
+ * IOError. At that time, the application should call IceCloseConnection.”¹
+ *
+ * Unfortunately libSM, while creating ICE connections of its own, does NOT
+ * register an I/O error handler.  So, when such an error occurs, libSM does
+ * NOT shut itself down as it should, and so we must take care NOT to close
+ * connections ourselves, even while advised by the libICE spec quoted above.
+ *
+ * Instead, when fed an I/O error, we'll simply log its occurrence and
+ * inform GLib to no longer monitor the ICE connection.  This will still
+ * allow our application to exit cleanly when receiving SIGTERM, a fatal
+ * X server I/O error, etc.
+ *
+ * 1. Inter-Client Exchange Protocol standard, §13. Error Handling
+ *    http://www.x.org/docs/ICE/ICElib.pdf
+ */
+
 
 /* Include first.  Sets G_LOG_DOMAIN. */
 #include "desktopEventsInt.h"
@@ -358,8 +385,8 @@ ICEIOErrorHandler(IceConn iceCnx)
  * @param[in]  cond     condition satisfied (ignored)
  * @param[in]  cbData   (ICEWatchCtx*) Channel context.
  *
- * @retval TRUE  Underlying iceCnx is still valid, so do not kill this source.
- * @retval FALSE Underlying iceCnx was destroyed, so remove this source.
+ * @retval TRUE  Event loop should continue to monitoring event source.
+ * @retval FALSE Event loop should cease monitoring event source.
  *
  ******************************************************************************
  */
@@ -384,8 +411,15 @@ ICEDispatch(GIOChannel *chn,
    case IceProcessMessagesSuccess:
       return TRUE;
    case IceProcessMessagesIOError:
-      IceCloseConnection(watchCtx->iceCnx);    // Let ICEWatch kill the source.
-      return TRUE;
+      /*
+       * See “Handling libICE I/O Errors” above.  watchCtx will float around
+       * until libSM calls IceCloseConnection, upon which ICEWatch will free
+       * those resources.
+       */
+      g_message("%s: encountered IceProcessMessagesIOError\n", __func__);
+      g_message("%s: detaching fd %d from application event loop\n",
+                __func__, IceConnectionNumber(watchCtx->iceCnx));
+      return FALSE;
    case IceProcessMessagesConnectionClosed:
       /*
        * iceCnx was invalidated, so we won't see another call to ICEWatch,
@@ -414,10 +448,10 @@ ICEDispatch(GIOChannel *chn,
  *    any data it may need to save until the connection is closed and the
  *    watch procedure is invoked again with opening set to False."
  *
- * @param[in]  iceCnx    Opaque ICE connection descriptor.
- * @parma[in]  cbData    (ToolsPluginData*) plugin data
- * @param[in]  opening   True if creating a connection, False if closing.
- * @param[in]  watchData See above.  New source will be stored here.
+ * @param[in]      iceCnx    Opaque ICE connection descriptor.
+ * @parma[in]      cbData    (ToolsPluginData*) plugin data
+ * @param[in]      opening   True if creating a connection, False if closing.
+ * @param[in,out]  watchData See above.  New source will be stored here.
  *
  ******************************************************************************
  */
@@ -434,6 +468,9 @@ ICEWatch(IceConn iceCnx,
    ctx = g_hash_table_lookup(pdata->_private, DE_PRIVATE_CTX);
    ASSERT(ctx);
 
+   g_debug("%s: fd %d opening %d\n", __func__, IceConnectionNumber(iceCnx),
+           opening);
+
    if (opening) {
       GIOChannel *iceChannel;
       GSource *iceSource;

diff --git a/open-vm-tools/services/plugins/desktopEvents/xioError.c b/open-vm-tools/services/plugins/desktopEvents/xioError.c
index 8bdd8ace4c4b6df8f813a707de0135d854ae4fe6..fe2acab6fd03add8c19aa24af37d7c473f03cd04 100644 (file)
--- a/open-vm-tools/services/plugins/desktopEvents/xioError.c
+++ b/open-vm-tools/services/plugins/desktopEvents/xioError.c
@@ -72,6 +72,7 @@ DEXIOErrorHandler(Display *dpy)
        * Inform clients capable of/interested in quick'n'dirty cleanup upon an
        * X I/O error.
        */
+      g_message("Emitting %s due to X I/O error.\n", TOOLS_CORE_SIG_XIOERROR);
       g_signal_emit_by_name(gCtx->serviceObj, TOOLS_CORE_SIG_XIOERROR, gCtx);
 
       /*