[PATCH] Timer bug fix ported to 2.4.20

[PATCH] Timer bug fix ported to 2.4.20

[Robert Wruck]

I recently upgraded from 2.2.19 to 2.4.20 and noted that the via timer
bugfix has disappeared.

I'm not sure if this is a VIA problem, because it happens on my PIII
i440BX-based board, Gigabyte GA-6BXE, which does not seem to have any
chips made by VIA. (82443BX, 32371EB PIIX4, ITE 8671 SuperIO)

However, whenever i start X, the system clock slows down significantly
(it takes about 10 real seconds for 1 system second to elapse).

The only patch I could find for this was the one by Neale Banks:
http://www.uwsg.iu.edu/hypermail/linux/kernel/0204.1/0005.html

I ported this to 2.4.20 and it works (at least for me..)
It can still be disabled with "timer=no-via686a" (from the original patch)

Any comments? (I'm not subscribed to linux-kernel)

Here is some dmesg output generated by the patch:

timer.c: VIA bug check triggered. Value read 62758 [0xf526], re-read 62758 [0xf526]
timer.c VIA bug really present. <4>Resetting PIT timer.
Resetting PIT timer.
Resetting PIT timer.
timer.c: VIA bug check triggered. Value read 62762 [0xf52a], re-read 62762 [0xf52a]
timer.c VIA bug really present. <4>Resetting PIT timer.
Resetting PIT timer.
Resetting PIT timer.
timer.c: VIA bug check triggered. Value read 62756 [0xf524], re-read 62756 [0xf524]
timer.c VIA bug really present. <4>Resetting PIT timer.
Resetting PIT timer.
Resetting PIT timer.
Resetting PIT timer.
timer.c: VIA bug check triggered. Value read 62760 [0xf528], re-read 62760 [0xf528]
timer.c VIA bug really present. <4>Resetting PIT timer.
Resetting PIT timer.
Resetting PIT timer.


Cheers,
Robert


--- linux-2.4.20/arch/i386/kernel/time.c.orig	2003-01-15 11:33:04.000000000 +0100
+++ linux-2.4.20/arch/i386/kernel/time.c	2003-01-16 16:48:32.000000000 +0100
@@ -84,6 +84,8 @@
 
 spinlock_t rtc_lock = SPIN_LOCK_UNLOCKED;
 
+static int via686a_hacks = 1; /* default to enabled */
+
 static inline unsigned long do_fast_gettimeoffset(void)
 {
 	register unsigned long eax, edx;
@@ -113,6 +115,54 @@
 	return delay_at_last_interrupt + edx;
 }
 
+/*
+ * VIA hardware bug workaround with check if it is really needed and
+ * a printk that could tell us what's exactly happening on machines which
+ * trigger the check, but are not VIA-based.
+ *
+ * Must be called with the i8253_spinlock held.
+ */
+
+static void via_reset_and_whine(int *count)
+{
+	static unsigned long last_whine = 0;
+	unsigned long new_whine;
+	int count2;
+	
+	new_whine = last_whine;
+	
+	outb_p(0x00, 0x43); /* Re-read the timer */
+	count2 = inb_p(0x40);
+	count2 |= inb(0x40) << 8;
+	
+	if (time_after(jiffies, last_whine)) {
+		printk(KERN_WARNING "timer.c: VIA bug check triggered. "
+		"Value read %d [%#x], re-read %d [%#x]\n",
+		*count, *count, count2, count2);
+		new_whine = jiffies + HZ;
+	}
+
+	*count = count2;
+	
+	if (count2 > LATCH) { /* Still bad */
+		if (time_after(jiffies, last_whine)) {
+			printk(KERN_WARNING "timer.c VIA bug really present. ");
+			new_whine = jiffies + HZ;
+		}
+		if (via686a_hacks) {
+			printk(KERN_WARNING "Resetting PIT timer.\n");
+			outb_p(0x34, 0x43);
+			outb_p(LATCH & 0xff, 0x40);
+			outb(LATCH >> 8, 0x40);
+		} else {
+			printk(KERN_WARNING "But VIA hacks disabled.\n");
+		}
+		*count = LATCH - 1;
+	}
+	
+	last_whine = new_whine;
+}
+
 #define TICK_SIZE tick
 
 spinlock_t i8253_lock = SPIN_LOCK_UNLOCKED;
@@ -184,10 +234,7 @@
 	
         /* VIA686a test code... reset the latch if count > max + 1 */
         if (count > LATCH) {
-                outb_p(0x34, 0x43);
-                outb_p(LATCH & 0xff, 0x40);
-                outb(LATCH >> 8, 0x40);
-                count = LATCH - 1;
+		via_reset_and_whine(&count);
         }
 	
 	spin_unlock(&i8253_lock);
@@ -504,6 +551,8 @@
 
 		count = inb_p(0x40);    /* read the latched count */
 		count |= inb(0x40) << 8;
+		if (count > LATCH)
+			via_reset_and_whine(&count);
 		spin_unlock(&i8253_lock);
 
 		count = ((LATCH-1) - count) * TICK_SIZE;
@@ -717,3 +766,25 @@
 	setup_irq(0, &irq0);
 #endif
 }
+
+static int __init timer_setup(char *str)
+{
+	int invert;
+	
+	while ((str != NULL) && (*str != '\0')) {
+		invert = (strncmp(str, "no-", 3) == 0);
+		if (invert)
+			str += 3;
+		if (strncmp(str, "via686a", 7) == 0) {
+			via686a_hacks = !invert;
+			if (invert)
+				printk(KERN_INFO "timer: VIA686a workaround disabled.\n");
+		}
+		str = strchr(str, ',');
+		if (str != NULL)
+			str += strspn(str, ", \t");
+	}
+	return 1;
+}
+
+__setup("timer=", timer_setup);

Re: [PATCH] Timer bug fix ported to 2.4.20

[Neale Banks]

On Thu, 16 Jan 2003, Robert Wruck wrote:

> I recently upgraded from 2.2.19 to 2.4.20 and noted that the via timer
> bugfix has disappeared.
> 
> I'm not sure if this is a VIA problem, because it happens on my PIII
> i440BX-based board, Gigabyte GA-6BXE, which does not seem to have any
> chips made by VIA. (82443BX, 32371EB PIIX4, ITE 8671 SuperIO)

Different people have reported this check triggered on various hardware
which doesn't have VIA686-anything.  In my case on both an old AcerNote
and a Toshiba-1800.

> However, whenever i start X, the system clock slows down significantly
> (it takes about 10 real seconds for 1 system second to elapse).

Ouch! Interesting to see that your system triggers the "VIA bug really
present.", whereas mine doesn't and I haven't found any gross time-keeping
problems.  Also, while we both report "read" and "re-read" values over a
narrow range they are quite different ranges.

I wonder if we are seeing different bugs which exhibit similar symptoms?

> The only patch I could find for this was the one by Neale Banks:
> http://www.uwsg.iu.edu/hypermail/linux/kernel/0204.1/0005.html

OK, just to set the record straight, my involvement has been merely to
port Vojtech's patch (for the original see:
http://marc.theaimsgroup.com/?l=linux-kernel&m=100560251020988&w=2 ), add
the ''disable with "timer=no-via686a"'' bit (which people more
knowledgeable than I say is quite unnecessary) and to report some
"interesting" results (see:
http://marc.theaimsgroup.com/?l=linux-kernel&m=102621884412934&w=2 ).

> I ported this to 2.4.20 and it works (at least for me..)
> It can still be disabled with "timer=no-via686a" (from the original patch)
> 
> Any comments? (I'm not subscribed to linux-kernel)

HTH,
Neale.

Re: [PATCH] Timer bug fix ported to 2.4.20

[Robert Wruck]

> > I recently upgraded from 2.2.19 to 2.4.20 and noted that the via timer
> > bugfix has disappeared.
> > 
> > I'm not sure if this is a VIA problem, because it happens on my PIII
> > i440BX-based board, Gigabyte GA-6BXE, which does not seem to have any
> > chips made by VIA. (82443BX, 32371EB PIIX4, ITE 8671 SuperIO)
> 
> Different people have reported this check triggered on various hardware
> which doesn't have VIA686-anything.  In my case on both an old AcerNote
> and a Toshiba-1800.
> 

OK, it seems in my case it's triggered by the initialization of the X
"s3virge" driver (for my s3 virge/dx pci card).
It doesn't occur when using "vga" nor when switching ttys.
Starting X with s3virge seems to be the only way to trigger it on my
system.

i'm not very familiar with those problems, i only know that resetting the
timer keeps my system clock from going mad.


robert