qemu-system-m68k -M q800 has an old bug that causes the kernel to
occasionally complain about a soft lockup:
watchdog: BUG: soft lockup - CPU#0 stuck for 5107s!
There isn't any actual lockup. The via1 clocksource produced a large
jump in jiffies, causing the watchdog to detect a stale timestamp.
The 32-bit clocksource counter runs at 783360 Hz and its period is
about 5482 seconds. Applying the "nanosecond" approximation used in
get_timestamp() in kernel/watchdog.c then yields the duration reported
in the log message above (always 5107 or 5108 in my tests):
0xffffffff / VIA_CLOCK_FREQ * 10**9 / 2**30 = 5106.209 seconds
It is notoriously difficult to correctly emulate a MOS6522 VIA chip. So
it seems wise to document the VIA clocksource driver better, especially
those hardware behaviours which the kernel relies upon.
Cc: Joshua Thompson <funaho@jurai.org>
Signed-off-by: Finn Thain <fthain@linux-m68k.org>
---
arch/m68k/mac/via.c | 16 ++++++++++++++++
1 file changed, 16 insertions(+)
diff --git a/arch/m68k/mac/via.c b/arch/m68k/mac/via.c
index 01e6b0e37f8d..142c2ed77c84 100644
--- a/arch/m68k/mac/via.c
+++ b/arch/m68k/mac/via.c
@@ -621,6 +621,22 @@ static u64 mac_read_clk(struct clocksource *cs)
* These problems are avoided by ignoring the low byte. Clock accuracy
* is 256 times worse (error can reach 0.327 ms) but CPU overhead is
* reduced by avoiding slow VIA register accesses.
+ *
+ * The VIA timer counter observably decrements to 0xFFFF before the
+ * counter reload interrupt gets raised. That complicates things a bit.
+ *
+ * State | vT1CH | VIA_TIMER_1_INT | inference drawn
+ * ------+------------+-----------------+-----------------------------
+ * A | FE thru 00 | false | counter is decrementing
+ * B | FF | false | counter wrapped
+ * C | FF | true | wrapped, interrupt raised
+ * D | FF | false | wrapped, interrupt handled
+ * E | FE thru 00 | true | wrapped, interrupt unhandled
+ *
+ * State D is never observed because handling the interrupt involves
+ * a 6522 register access and every access consumes a "phi 2" clock
+ * cycle. So 0xFF implies either state B or C, depending on the value
+ * of the VIA_TIMER_1_INT bit.
*/
local_irq_save(flags);
--
2.45.3