[Qemu-devel] [PULL 05/10] watchdog: fix deadlock with -watchdog-action pause

[Paolo Bonzini <pbonzini@redhat.com>]

qemu_clock_enable says:

/* Disabling the clock will wait for related timerlists to stop
 * executing qemu_run_timers.  Thus, this functions should not
 * be used from the callback of a timer that is based on @clock.
 * Doing so would cause a deadlock.
 */

and it indeed does: vm_stop uses qemu_clock_enable on QEMU_CLOCK_VIRTUAL
and watchdogs are based on QEMU_CLOCK_VIRTUAL, and we get a deadlock.

Use qemu_system_vmstop_request_prepare()/qemu_system_vmstop_request()
instead; yet another alternative could be a BH.

I checked other occurrences of vm_stop and they should not have this
problem.  RUN_STATE_IO_ERROR could in principle (it depends on the
code in the drivers) but it has been fixed by commit 2bd3bce, "block:
asynchronously stop the VM on I/O errors", 2014-06-05.

Tested-by: Luiz Capitulino <lcapitulino@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
 hw/watchdog/watchdog.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/hw/watchdog/watchdog.c b/hw/watchdog/watchdog.c
index 9f607d4..c307f9b 100644
--- a/hw/watchdog/watchdog.c
+++ b/hw/watchdog/watchdog.c
@@ -122,8 +122,12 @@ void watchdog_perform_action(void)
         exit(0);
 
     case WDT_PAUSE:             /* same as 'stop' command in monitor */
+        /* In a timer callback, when vm_stop calls qemu_clock_enable
+         * you would get a deadlock.  Bypass the problem.
+         */
+        qemu_system_vmstop_request_prepare();
         qapi_event_send_watchdog(WATCHDOG_EXPIRATION_ACTION_PAUSE, &error_abort);
-        vm_stop(RUN_STATE_WATCHDOG);
+        qemu_system_vmstop_request(RUN_STATE_WATCHDOG);
         break;
 
     case WDT_DEBUG:
-- 
1.8.3.1