migration: regain control of images when migration fails to complete
authorGreg Kurz <gkurz@linux.vnet.ibm.com>
Wed, 18 May 2016 13:44:36 +0000 (15:44 +0200)
committerAmit Shah <amit.shah@redhat.com>
Mon, 23 May 2016 16:49:36 +0000 (22:19 +0530)
We currently have an error path during migration that can cause
the source QEMU to abort:

migration_thread()
  migration_completion()
    runstate_is_running() ----------------> true if guest is running
    bdrv_inactivate_all() ----------------> inactivate images
    qemu_savevm_state_complete_precopy()
     ... qemu_fflush()
           socket_writev_buffer() --------> error because destination fails
         qemu_fflush() -------------------> set error on migration stream
  migration_completion() -----------------> set migrate state to FAILED
migration_thread() -----------------------> break migration loop
  vm_start() -----------------------------> restart guest with inactive
                                            images

and you get:

qemu-system-ppc64: socket_writev_buffer: Got err=104 for (32768/18446744073709551615)
qemu-system-ppc64: /home/greg/Work/qemu/qemu-master/block/io.c:1342:bdrv_co_do_pwritev: Assertion `!(bs->open_flags & 0x0800)' failed.
Aborted (core dumped)

If we try postcopy with a similar scenario, we also get the writev error
message but QEMU leaves the guest paused because entered_postcopy is true.

We could possibly do the same with precopy and leave the guest paused.
But since the historical default for migration errors is to restart the
source, this patch adds a call to bdrv_invalidate_cache_all() instead.

Signed-off-by: Greg Kurz <gkurz@linux.vnet.ibm.com>
Message-Id: <146357896785.6003.11983081732454362715.stgit@bahia.huguette.org>
Signed-off-by: Amit Shah <amit.shah@redhat.com>
migration/migration.c

index 721d010fde7d2db2ee85fa8cc6f73ef798d89ad5..f5327e8c0a9531cb429ae7f66af16bb915c5f404 100644 (file)
@@ -1606,19 +1606,32 @@ static void migration_completion(MigrationState *s, int current_active_state,
         rp_error = await_return_path_close_on_source(s);
         trace_migration_completion_postcopy_end_after_rp(rp_error);
         if (rp_error) {
-            goto fail;
+            goto fail_invalidate;
         }
     }
 
     if (qemu_file_get_error(s->to_dst_file)) {
         trace_migration_completion_file_err();
-        goto fail;
+        goto fail_invalidate;
     }
 
     migrate_set_state(&s->state, current_active_state,
                       MIGRATION_STATUS_COMPLETED);
     return;
 
+fail_invalidate:
+    /* If not doing postcopy, vm_start() will be called: let's regain
+     * control on images.
+     */
+    if (s->state == MIGRATION_STATUS_ACTIVE) {
+        Error *local_err = NULL;
+
+        bdrv_invalidate_cache_all(&local_err);
+        if (local_err) {
+            error_report_err(local_err);
+        }
+    }
+
 fail:
     migrate_set_state(&s->state, current_active_state,
                       MIGRATION_STATUS_FAILED);