Without any of HEAD^ or HEAD^^ applied, qemu will most likely crash on
the qemu-io invocation, for a variety of immediate reasons.  The
underlying problem is generally a use-after-free access into
backup-top's BlockCopyState.

With only HEAD^ applied, qemu-io will run into an EIO (which is not
capture by the output, but you can see that the qemu-io invocation will
be accepted (i.e., qemu-io will run) in contrast to the reference
output, where the node name cannot be found), and qemu will then crash
in query-named-block-nodes: bdrv_get_allocated_file_size() detects
backup-top to be a filter and passes the request through to its child.
However, after bdrv_backup_top_drop(), that child is NULL, so the
recursive call crashes.

With HEAD^^ applied, this test should pass.

Signed-off-by: Max Reitz <mreitz@redhat.com>
---
 tests/qemu-iotests/283     | 55 ++++++++++++++++++++++++++++++++++++++
 tests/qemu-iotests/283.out | 15 +++++++++++
 2 files changed, 70 insertions(+)

diff --git a/tests/qemu-iotests/283 b/tests/qemu-iotests/283
index 79643e375b..509dcbbcf4 100755
--- a/tests/qemu-iotests/283
+++ b/tests/qemu-iotests/283
@@ -97,3 +97,58 @@ vm.qmp_log('blockdev-add', **{
 vm.qmp_log('blockdev-backup', sync='full', device='source', target='target')
 
 vm.shutdown()
+
+
+"""
+Check that the backup-top node is gone after job-finalize.
+
+During finalization, the node becomes inactive and can no longer
+function.  If it is still present, new parents might be attached, and
+there would be no meaningful way to handle their I/O requests.
+"""
+
+print('\n=== backup-top should be gone after job-finalize ===\n')
+
+vm = iotests.VM()
+vm.launch()
+
+vm.qmp_log('blockdev-add', **{
+    'node-name': 'source',
+    'driver': 'null-co',
+})
+
+vm.qmp_log('blockdev-add', **{
+    'node-name': 'target',
+    'driver': 'null-co',
+})
+
+vm.qmp_log('blockdev-backup',
+           job_id='backup',
+           device='source',
+           target='target',
+           sync='full',
+           filter_node_name='backup-filter',
+           auto_finalize=False,
+           auto_dismiss=False)
+
+vm.event_wait('BLOCK_JOB_PENDING', 5.0)
+
+# The backup-top filter should still be present prior to finalization
+assert vm.node_info('backup-filter') is not None
+
+vm.qmp_log('job-finalize', id='backup')
+vm.event_wait('BLOCK_JOB_COMPLETED', 5.0)
+
+# The filter should be gone now.  Check that by trying to access it
+# with qemu-io (which will most likely crash qemu if it is still
+# there.).
+vm.qmp_log('human-monitor-command',
+           command_line='qemu-io backup-filter "write 0 1M"')
+
+# (Also, do an explicit check.)
+assert vm.node_info('backup-filter') is None
+
+vm.qmp_log('job-dismiss', id='backup')
+vm.event_wait('JOB_STATUS_CHANGE', 5.0, {'data': {'status': 'null'}})
+
+vm.shutdown()
diff --git a/tests/qemu-iotests/283.out b/tests/qemu-iotests/283.out
index d8cff22cc1..7e9cd9a7d4 100644
--- a/tests/qemu-iotests/283.out
+++ b/tests/qemu-iotests/283.out
@@ -6,3 +6,18 @@
 {"return": {}}
 {"execute": "blockdev-backup", "arguments": {"device": "source", "sync": "full", "target": "target"}}
 {"error": {"class": "GenericError", "desc": "Cannot set permissions for backup-top filter: Conflicts with use by other as 'image', which uses 'write' on base"}}
+
+=== backup-top should be gone after job-finalize ===
+
+{"execute": "blockdev-add", "arguments": {"driver": "null-co", "node-name": "source"}}
+{"return": {}}
+{"execute": "blockdev-add", "arguments": {"driver": "null-co", "node-name": "target"}}
+{"return": {}}
+{"execute": "blockdev-backup", "arguments": {"auto-dismiss": false, "auto-finalize": false, "device": "source", "filter-node-name": "backup-filter", "job-id": "backup", "sync": "full", "target": "target"}}
+{"return": {}}
+{"execute": "job-finalize", "arguments": {"id": "backup"}}
+{"return": {}}
+{"execute": "human-monitor-command", "arguments": {"command-line": "qemu-io backup-filter \"write 0 1M\""}}
+{"return": "Error: Cannot find device= nor node_name=backup-filter\r\n"}
+{"execute": "job-dismiss", "arguments": {"id": "backup"}}
+{"return": {}}
-- 
2.29.2

On 19.02.21 16:33, Max Reitz wrote:
> Without any of HEAD^ or HEAD^^ applied, qemu will most likely crash on
> the qemu-io invocation, for a variety of immediate reasons.  The
> underlying problem is generally a use-after-free access into
> backup-top's BlockCopyState.
> 
> With only HEAD^ applied, qemu-io will run into an EIO (which is not
> capture by the output, but you can see that the qemu-io invocation will
> be accepted (i.e., qemu-io will run) in contrast to the reference
> output, where the node name cannot be found), and qemu will then crash
> in query-named-block-nodes: bdrv_get_allocated_file_size() detects
> backup-top to be a filter and passes the request through to its child.
> However, after bdrv_backup_top_drop(), that child is NULL, so the
> recursive call crashes.
> 
> With HEAD^^ applied, this test should pass.
> 
> Signed-off-by: Max Reitz <mreitz@redhat.com>
> ---
>   tests/qemu-iotests/283     | 55 ++++++++++++++++++++++++++++++++++++++
>   tests/qemu-iotests/283.out | 15 +++++++++++
>   2 files changed, 70 insertions(+)
> 
> diff --git a/tests/qemu-iotests/283 b/tests/qemu-iotests/283
> index 79643e375b..509dcbbcf4 100755
> --- a/tests/qemu-iotests/283
> +++ b/tests/qemu-iotests/283
> @@ -97,3 +97,58 @@ vm.qmp_log('blockdev-add', **{
>   vm.qmp_log('blockdev-backup', sync='full', device='source', target='target')
>   
>   vm.shutdown()
> +
> +
> +"""
> +Check that the backup-top node is gone after job-finalize.
> +
> +During finalization, the node becomes inactive and can no longer
> +function.  If it is still present, new parents might be attached, and
> +there would be no meaningful way to handle their I/O requests.
> +"""

Oh no, 297/pylint complains that this “string statement has no effect”. 
  Guess it should be a normal comment under the following print() then...

Max

> +print('\n=== backup-top should be gone after job-finalize ===\n')

Am 19.02.2021 um 16:59 hat Max Reitz geschrieben:
> On 19.02.21 16:33, Max Reitz wrote:
> > Without any of HEAD^ or HEAD^^ applied, qemu will most likely crash on
> > the qemu-io invocation, for a variety of immediate reasons.  The
> > underlying problem is generally a use-after-free access into
> > backup-top's BlockCopyState.
> > 
> > With only HEAD^ applied, qemu-io will run into an EIO (which is not
> > capture by the output, but you can see that the qemu-io invocation will
> > be accepted (i.e., qemu-io will run) in contrast to the reference
> > output, where the node name cannot be found), and qemu will then crash
> > in query-named-block-nodes: bdrv_get_allocated_file_size() detects
> > backup-top to be a filter and passes the request through to its child.
> > However, after bdrv_backup_top_drop(), that child is NULL, so the
> > recursive call crashes.
> > 
> > With HEAD^^ applied, this test should pass.
> > 
> > Signed-off-by: Max Reitz <mreitz@redhat.com>
> > ---
> >   tests/qemu-iotests/283     | 55 ++++++++++++++++++++++++++++++++++++++
> >   tests/qemu-iotests/283.out | 15 +++++++++++
> >   2 files changed, 70 insertions(+)
> > 
> > diff --git a/tests/qemu-iotests/283 b/tests/qemu-iotests/283
> > index 79643e375b..509dcbbcf4 100755
> > --- a/tests/qemu-iotests/283
> > +++ b/tests/qemu-iotests/283
> > @@ -97,3 +97,58 @@ vm.qmp_log('blockdev-add', **{
> >   vm.qmp_log('blockdev-backup', sync='full', device='source', target='target')
> >   vm.shutdown()
> > +
> > +
> > +"""
> > +Check that the backup-top node is gone after job-finalize.
> > +
> > +During finalization, the node becomes inactive and can no longer
> > +function.  If it is still present, new parents might be attached, and
> > +there would be no meaningful way to handle their I/O requests.
> > +"""
> 
> Oh no, 297/pylint complains that this “string statement has no effect”.
> Guess it should be a normal comment under the following print() then...

Thanks, fixed up the comment as you suggest and applied to the block
branch.

Kevin