block: asynchronously stop the VM on I/O errors

With virtio-blk dataplane, I/O errors might occur while QEMU is
not in the main I/O thread.  However, it's invalid to call vm_stop
when we're neither in a VCPU thread nor in the main I/O thread,
even if we were to take the iothread mutex around it.

To avoid this problem, we can raise a request to the main I/O thread,
similar to what QEMU does when vm_stop is called from a CPU thread.
We know that bdrv_error_action is called from an AIO callback, and
the moment at which the callback will fire is not well-defined; it
depends on the moment at which the disk or OS finishes the operation,
which can happen at any time.  Note that QEMU is certainly not in a CPU
thread and we do not need to call cpu_stop_current() like vm_stop() does.

However, we need to ensure that any action taken by management will
result in correct detection of the error _and_ a running VM.  In particular:

- the event must be raised after the iostatus has been set, so that
"info block" will return an iostatus that matches the event.

- the VM must be stopped after the iostatus has been set, so that
"info block" will return an iostatus that matches the runstate.

The ordering between the STOP and BLOCK_IO_ERROR events is preserved;
BLOCK_IO_ERROR is documented to come first.

This makes bdrv_error_action() thread safe (assuming QMP events are,
which is attacked by a separate series).

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>

3 files changed, 26 insertions, 4 deletions

diff --git a/block.c b/block.c
index 43abe96db6..ff44e76c87 100644
--- a/block.c
+++ b/block.c
@@ -3626,10 +3626,27 @@ void bdrv_error_action(BlockDriverState *bs, BlockErrorAction action,
                        bool is_read, int error)
 {
     assert(error >= 0);
-    bdrv_emit_qmp_error_event(bs, QEVENT_BLOCK_IO_ERROR, action, is_read);
+
     if (action == BDRV_ACTION_STOP) {
-        vm_stop(RUN_STATE_IO_ERROR);
+        /* First set the iostatus, so that "info block" returns an iostatus
+         * that matches the events raised so far (an additional error iostatus
+         * is fine, but not a lost one).
+         */
         bdrv_iostatus_set_err(bs, error);
+
+        /* Then raise the request to stop the VM and the event.
+         * qemu_system_vmstop_request_prepare has two effects.  First,
+         * it ensures that the STOP event always comes after the
+         * BLOCK_IO_ERROR event.  Second, it ensures that even if management
+         * can observe the STOP event and do a "cont" before the STOP
+         * event is issued, the VM will not stop.  In this case, vm_start()
+         * also ensures that the STOP/RESUME pair of events is emitted.
+         */
+        qemu_system_vmstop_request_prepare();
+        bdrv_emit_qmp_error_event(bs, QEVENT_BLOCK_IO_ERROR, action, is_read);
+        qemu_system_vmstop_request(RUN_STATE_IO_ERROR);
+    } else {
+        bdrv_emit_qmp_error_event(bs, QEVENT_BLOCK_IO_ERROR, action, is_read);
     }
 }
 
diff --git a/docs/qmp/qmp-events.txt b/docs/qmp/qmp-events.txt
index 019db53ec8..22fea580a9 100644
--- a/docs/qmp/qmp-events.txt
+++ b/docs/qmp/qmp-events.txt
@@ -62,7 +62,7 @@ Data:
 - "action": action that has been taken, it's one of the following (json-string):
     "ignore": error has been ignored
     "report": error has been reported to the device
-    "stop": error caused VM to be stopped
+    "stop": the VM is going to stop because of the error
 
 Example:
 
diff --git a/stubs/vm-stop.c b/stubs/vm-stop.c
index f82c897dfe..69fd86b2e8 100644
--- a/stubs/vm-stop.c
+++ b/stubs/vm-stop.c
@@ -1,7 +1,12 @@
 #include "qemu-common.h"
 #include "sysemu/sysemu.h"
 
-int vm_stop(RunState state)
+void qemu_system_vmstop_request_prepare(void)
+{
+    abort();
+}
+
+void qemu_system_vmstop_request(RunState state)
 {
     abort();
 }