[PATCH qemu-server v2] fix #7705: blockdev: detach backing chain on disk move/hot-unplug

[PATCH qemu-server v2] fix #7705: blockdev: detach backing chain on disk move/hot-unplug

[Arthur Bied-Charreton]

When snapshots-as-volume-chains is enabled, moving or hot-unplugging a
disk that holds snapshots does not detach the snapshot volumes.
`Blockdev::detach` is called in both paths and only follows the `file`
child, so the snapshot nodes are left behind. Since they were added
explicitly via blockdev-add, they are also not GCed by QEMU and, as a
result, keep their source volumes open.

Deleting one of those snapshots after the move does not detach the
orphaned block node because the disk is no longer attached to the
running VM. QEMU keeps the volume open for as long as the VM runs, and a
later attempt to remove the source fails with lvremove reporting the LV
as still in use.

Add 'follow-backing' option to `Blockdev::detach`. When enabled, the
'backing' children are detached in addition to the 'file' children,
meaning the snapshot blockdevs are released.

Enable 'follow-backing' at the following callsites:

* qemu_handle_concluded_blockjob (on complete the source is detached
  along with the backing snapshot chain, on cancel/error it detaches the
  mirror target, which is flat so follow-backing is a no-op)

* qemu_drivedel

Signed-off-by: Arthur Bied-Charreton <a.bied-charreton@proxmox.com>
---
Changes since v1:

- Update `Blockdev::detach`'s POD.

 src/PVE/QemuServer.pm          |  6 +++++-
 src/PVE/QemuServer/BlockJob.pm |  4 +++-
 src/PVE/QemuServer/Blockdev.pm | 26 ++++++++++++++++++++------
 3 files changed, 28 insertions(+), 8 deletions(-)

diff --git a/src/PVE/QemuServer.pm b/src/PVE/QemuServer.pm
index 55e9f520..472bee7e 100644
--- a/src/PVE/QemuServer.pm
+++ b/src/PVE/QemuServer.pm
@@ -4082,7 +4082,11 @@ sub qemu_drivedel {
 
     # for the switch to -blockdev
     if (PVE::QemuServer::Machine::is_machine_version_at_least($machine_type, 10, 0)) {
-        PVE::QemuServer::Blockdev::detach(vm_qmp_peer($vmid), "drive-$deviceid");
+        PVE::QemuServer::Blockdev::detach(
+            vm_qmp_peer($vmid),
+            "drive-$deviceid",
+            { 'follow-backing' => 1 },
+        );
         return 1;
     } else {
         my $ret = PVE::QemuServer::Monitor::hmp_cmd($vmid, "drive_del drive-$deviceid", 10 * 60);
diff --git a/src/PVE/QemuServer/BlockJob.pm b/src/PVE/QemuServer/BlockJob.pm
index 921f046c..493dda33 100644
--- a/src/PVE/QemuServer/BlockJob.pm
+++ b/src/PVE/QemuServer/BlockJob.pm
@@ -34,7 +34,9 @@ sub qemu_handle_concluded_blockjob {
     $job->{'detach-node-name'} = $job->{'target-node-name'} if $qmp_info->{error} || $job->{cancel};
 
     if (my $node_name = $job->{'detach-node-name'}) {
-        eval { PVE::QemuServer::Blockdev::detach($qmp_peer, $node_name); };
+        eval {
+            PVE::QemuServer::Blockdev::detach($qmp_peer, $node_name, { 'follow-backing' => 1 });
+        };
         log_warn($@) if $@;
     }
 
diff --git a/src/PVE/QemuServer/Blockdev.pm b/src/PVE/QemuServer/Blockdev.pm
index 101c747c..8eafb9fb 100644
--- a/src/PVE/QemuServer/Blockdev.pm
+++ b/src/PVE/QemuServer/Blockdev.pm
@@ -637,7 +637,7 @@ sub attach {
 
 =head3 detach
 
-    detach($qmp_peer, $node_name);
+    detach($qmp_peer, $node_name, $opts);
 
 Detach the block device C<$node_name> from the QMP peer C<$qmp_peer>. Also removes associated child
 block nodes.
@@ -650,15 +650,25 @@ Parameters:
 
 =item C<$node_name>: The node name identifying the block node in QEMU.
 
+=item C<$opts> Additional options. Possible values:
+
+=over
+
+=item C<'follow-backing'>: also detach the 'backing' children
+
+=back
+
 =back
 
 =cut
 
 sub detach {
-    my ($qmp_peer, $node_name) = @_;
+    my ($qmp_peer, $node_name, $opts) = @_;
 
     die "Blockdev::detach - no node name\n" if !$node_name;
 
+    my $follow_backing = $opts->{'follow-backing'};
+
     my $block_info = qmp_cmd($qmp_peer, "query-named-block-nodes");
     $block_info = { map { $_->{'node-name'} => $_ } $block_info->@* };
 
@@ -667,12 +677,13 @@ sub detach {
         $remove_throttle_group_id = throttle_group_id($drive_id);
     }
 
-    while ($node_name) {
-        last if !$block_info->{$node_name}; # already gone
+    my @queue = ($node_name);
+    while (defined(my $node_name = shift @queue)) {
+        next if !$block_info->{$node_name}; # already gone
 
         my $res = qmp_cmd($qmp_peer, 'blockdev-del', 'node-name' => "$node_name", noerr => 1);
         if (my $err = $res->{error}) {
-            last if $err =~ m/Failed to find node with node-name/; # already gone
+            next if $err =~ m/Failed to find node with node-name/; # already gone
             die "deleting blockdev '$node_name' failed : $err\n";
         }
 
@@ -680,7 +691,10 @@ sub detach {
         # Recursively remove 'file' child nodes. QEMU will auto-remove implicitly added child nodes,
         # but e.g. the child of the top throttle node might have been explicitly added as a mirror
         # target, and needs to be removed manually.
-        $node_name = $children->{file}->{'node-name'};
+        # 'backing' is only followed when requested, because callers tearing down a single chain node
+        # rely on the nodes below it surviving.
+        push @queue, $children->{file}->{'node-name'} if $children->{file};
+        push @queue, $children->{backing}->{'node-name'} if $follow_backing && $children->{backing};
     }
 
     if ($remove_throttle_group_id) {
-- 
2.47.3