watchdog/diag288: avoid race condition on expired watchdog

When configured to inject an NMI, watchdog_perform_action() may cause
the BQL to be temporarily relinquished (inject_nmi() → ... →
s390_nmi() → s390_cpu_restart() → run_on_cpu()). When the guest issues
diag 288 again in response to the NMI, the diag 288 operation will
race against wdt_diag288_reset(). Depending on scheduler behaviour,
wdt_diag288_reset() may be run after the guest issued a diag 288
Init. As a result, we will cancel the timer the guest just set up. The
effect observed by the guest is that a second expiry does not trigger
the watchdog action and diag 288 Change operations fail.

Fix this by resetting the timer _before_ invoking the action.

Signed-off-by: Sascha Silbe <silbe@linux.vnet.ibm.com>
Acked-by: David Hildenbrand <dahi@linux.vnet.ibm.com>
Signed-off-by: Cornelia Huck <cornelia.huck@de.ibm.com>

1 files changed, 8 insertions, 4 deletions

diff --git a/hw/watchdog/wdt_diag288.c b/hw/watchdog/wdt_diag288.c
index 5eb5b94..1c3658e 100644
--- a/hw/watchdog/wdt_diag288.c
+++ b/hw/watchdog/wdt_diag288.c
@@ -51,15 +51,19 @@ static void diag288_reset(void *opaque)
 static void diag288_timer_expired(void *dev)
 {
     qemu_log_mask(CPU_LOG_RESET, "Watchdog timer expired.\n");
-    watchdog_perform_action();
-    /* Reset the watchdog only if the guest was notified about expiry. */
+    /* Reset the watchdog only if the guest gets notified about
+     * expiry. watchdog_perform_action() may temporarily relinquish
+     * the BQL; reset before triggering the action to avoid races with
+     * diag288 instructions. */
     switch (get_watchdog_action()) {
     case WDT_DEBUG:
     case WDT_NONE:
     case WDT_PAUSE:
-         return;
+        break;
+    default:
+        wdt_diag288_reset(dev);
     }
-    wdt_diag288_reset(dev);
+    watchdog_perform_action();
 }
 
 static int wdt_diag288_handle_timer(DIAG288State *diag288,