[Linux-ia64] [PATCH] Fix for the X->text->X console crash

[Linux-ia64] [PATCH] Fix for the X->text->X console crash

[Chris Ahna]

A patch which should fix the famous X->text->X console switch crashes is
appended.  It seems that the kernel was sending a 'console reacquired'
signal to X _before_ doing a screen blank operation.  This allowed X to
be configuring the video card from one processor while the kernel
cleared parts of the VGA range from another.  Interleaving these types
of accesses apparently caused the system crashes (at least that's what
the hardware guys with PCI bus analyzers saw and the problem goes away
when running UP :-)

I posted this patch to linux-kernel last night because it doesn't seem
to be an ia64 specific problem.  Andrea Arcangeli said he thought the
patch looked alright, so that gives me some more confidence in its
correctness.  I hope Linus and Alan will be applying this soon, but in
the meantime it should fix the problem when applied to your local
kernels.  The patch is against 2.4.13-ia64-011024.

Chris


--- linux-2.4.13-ia64-011024-pristine/drivers/char/vt.c	Mon Sep 17 22:52:35 2001
+++ linux-2.4.13-ia64-011024-dev/drivers/char/vt.c	Thu Oct 25 20:46:08 2001
@@ -1184,6 +1184,24 @@
 	switch_screen(new_console);
 
 	/*
+	 * This can't appear below a successful kill_proc().  If it did,
+	 * then the *blank_screen operation could occur while X, having
+	 * received acqsig, is waking up on another processor.  This
+	 * condition can lead to overlapping accesses to the VGA range
+	 * and the framebuffer (causing system lockups).
+	 *
+	 * To account for this we duplicate this code below only if the
+	 * controlling process is gone and we've called reset_vc.
+	 */
+	if (old_vc_mode != vt_cons[new_console]->vc_mode)
+	{
+		if (vt_cons[new_console]->vc_mode = KD_TEXT)
+			unblank_screen();
+		else
+			do_blank_screen(1);
+	}
+
+	/*
 	 * If this new console is under process control, send it a signal
 	 * telling it that it has acquired. Also check if it has died and
 	 * clean up (similar to logic employed in change_console())
@@ -1209,19 +1227,15 @@
 		 * to account for and tracking tty count may be undesirable.
 		 */
 		        reset_vc(new_console);
-		}
-	}
 
-	/*
-	 * We do this here because the controlling process above may have
-	 * gone, and so there is now a new vc_mode
-	 */
-	if (old_vc_mode != vt_cons[new_console]->vc_mode)
-	{
-		if (vt_cons[new_console]->vc_mode = KD_TEXT)
-			unblank_screen();
-		else
-			do_blank_screen(1);
+			if (old_vc_mode != vt_cons[new_console]->vc_mode)
+			{
+				if (vt_cons[new_console]->vc_mode = KD_TEXT)
+					unblank_screen();
+				else
+					do_blank_screen(1);
+			}
+		}
 	}
 
 	/*

Re: [Linux-ia64] [PATCH] Fix for the X->text->X console crash

[David Mosberger]

>>>>> On Fri, 26 Oct 2001 10:17:15 -0700, Chris Ahna <christopher.j.ahna@intel.com> said:

  Chris> A patch which should fix the famous X->text->X console switch
  Chris> crashes is appended.

This is excellent news.  Great job!

  Chris> It seems that the kernel was sending a 'console reacquired'
  Chris> signal to X _before_ doing a screen blank operation.  This
  Chris> allowed X to be configuring the video card from one processor
  Chris> while the kernel cleared parts of the VGA range from another.
  Chris> Interleaving these types of accesses apparently caused the
  Chris> system crashes (at least that's what the hardware guys with
  Chris> PCI bus analyzers saw and the problem goes away when running
  Chris> UP :-)

Argh, a generic kernel race condition.  Who would have thought of that.

Any idea why this is crashing on ia64 but not on MP x86s?

In any case, it sounds like we can finally take this bug off the list.
I'm sure lots of people will appreciate this.

Thanks,

	--david