block: fix wrong order in live block migration setup

The function init_blk_migration is better to be called before
set_dirty_tracking as the reasons below.

If we want to track dirty blocks via dirty_maps on a BlockDriverState
when doing live block-migration, its correspoding 'BlkMigDevState' should be
added to block_mig_state.bmds_list first for subsequent processing.
Otherwise set_dirty_tracking will do nothing on an empty list than allocating
dirty_bitmaps for them. And bdrv_get_dirty_count will access the
bmds->dirty_maps directly, then there would be a segfault triggered.

If the set_dirty_tracking fails, qemu_savevm_state_cancel will handle
the cleanup of init_blk_migration automatically.

Reviewed-by: Fam Zheng <famz@redhat.com>
Signed-off-by: chai wen <chaiw.fnst@cn.fujitsu.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>

diff --git a/block-migration.c b/block-migration.c
index 16562709c853e60d86b3977c947f4e81318a1dca..25a03889f471c0a81dec66d59519137c9f7fd13d 100644 (file)
--- a/block-migration.c
+++ b/block-migration.c
@@ -629,6 +629,7 @@ static int block_save_setup(QEMUFile *f, void *opaque)
             block_mig_state.submitted, block_mig_state.transferred);
 
     qemu_mutex_lock_iothread();
+    init_blk_migration(f);
 
     /* start track dirty blocks */
     ret = set_dirty_tracking();
@@ -638,8 +639,6 @@ static int block_save_setup(QEMUFile *f, void *opaque)
         return ret;
     }
 
-    init_blk_migration(f);
-
     qemu_mutex_unlock_iothread();
 
     ret = flush_blks(f);