Re: [Qemu-devel] [PATCH] iotests: test clearing unknown autoclear_features by qcow2

[Kevin Wolf <kwolf@redhat.com>]

Am 10.11.2017 um 18:54 hat Vladimir Sementsov-Ogievskiy geschrieben:
> Test clearing unknown autoclear_features by qcow2 on incoming
> migration.
> 
> Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
> ---
> 
> Hi all!
> 
> This patch shows degradation, added in 2.10 in commit
> 
> commit 9c5e6594f15b7364624a3ad40306c396c93a2145
> Author: Kevin Wolf <kwolf@redhat.com>
> Date:   Thu May 4 18:52:40 2017 +0200
> 
>     block: Fix write/resize permissions for inactive images
> 
> The problem:
> When on incoming migration with shared storage we reopen image to RW mode
> we call bdrv_invalidate_cache it firstly call drv->bdrv_invalidate_cache and
> only after it, through "parent->role->activate(parent, &local_err);" we set
> appropriate RW permission.
> 
> Because of this, if drv->bdrv_invalidate_cache wants to write something
> (image is RW and BDRV_O_INACTIVE is not set) it goes into
> "bdrv_aligned_pwritev: Assertion `child->perm & BLK_PERM_WRITE' failed"
> 
> One case, when qcow2_invalidate_cache wants to write
> something - when it wants to clear some unknown autoclear features. So,
> here is a test for it.
> 
> Another case is when we try to migrate persistent dirty bitmaps through
> shared storage, on invalidate_cache qcow2 will try to set DIRTY bit in
> all loaded bitmaps.
> 
> Kevin, what do you think? I understand that operation order should be changed
> somehow in bdrv_invalidate_cache, but I'm not sure about how "parent->role->.."
> things works and can we safely move this part from function end to the middle.

I don't think you need to move the parent->role->activate() callback,
but just the bdrv_set_perm() so that we request the correct permissions
for operation without the BDRV_O_INACTIVE flag.

The following seems to work for me (including a fix for the test case,
we don't seem to get a RESUME event). I'm not sure about the error paths
yet. We should probably try do undo the permission changes there. I can
have a closer look into this next week.

Kevin


diff --git a/block.c b/block.c
index edc5bb9f9b..f530b630b6 100644
--- a/block.c
+++ b/block.c
@@ -4178,7 +4178,17 @@ void bdrv_invalidate_cache(BlockDriverState *bs, Error **errp)
         }
     }
 
+    /* Update permissions, they may differ for inactive nodes */
     bs->open_flags &= ~BDRV_O_INACTIVE;
+    bdrv_get_cumulative_perm(bs, &perm, &shared_perm);
+    ret = bdrv_check_perm(bs, NULL, perm, shared_perm, NULL, &local_err);
+    if (ret < 0) {
+        bs->open_flags |= BDRV_O_INACTIVE;
+        error_propagate(errp, local_err);
+        return;
+    }
+    bdrv_set_perm(bs, perm, shared_perm);
+
     if (bs->drv->bdrv_invalidate_cache) {
         bs->drv->bdrv_invalidate_cache(bs, &local_err);
         if (local_err) {
@@ -4195,16 +4205,6 @@ void bdrv_invalidate_cache(BlockDriverState *bs, Error **errp)
         return;
     }
 
-    /* Update permissions, they may differ for inactive nodes */
-    bdrv_get_cumulative_perm(bs, &perm, &shared_perm);
-    ret = bdrv_check_perm(bs, NULL, perm, shared_perm, NULL, &local_err);
-    if (ret < 0) {
-        bs->open_flags |= BDRV_O_INACTIVE;
-        error_propagate(errp, local_err);
-        return;
-    }
-    bdrv_set_perm(bs, perm, shared_perm);
-
     QLIST_FOREACH(parent, &bs->parents, next_parent) {
         if (parent->role->activate) {
             parent->role->activate(parent, &local_err);
diff --git a/tests/qemu-iotests/196 b/tests/qemu-iotests/196
index 9ffbc723c2..4116ebc92b 100755
--- a/tests/qemu-iotests/196
+++ b/tests/qemu-iotests/196
@@ -53,7 +53,10 @@ class TestInvalidateAutoclear(iotests.QMPTestCase):
             f.write(b'\xff')
 
         self.vm_b.launch()
-        self.assertNotEqual(self.vm_b.event_wait("RESUME"), None)
+        while True:
+            result = self.vm_b.qmp('query-status')
+            if result['return']['status'] == 'running':
+                break
 
         with open(disk, 'rb') as f:
             f.seek(88)