Skip to contentSkip to conversation

net: delay freeing peer host device

PATCH Superseded Michael S. Tsirkin 2 participants 3 messages Sep 27, 2010

v1 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 net: delay freeing peer host device

With -netdev, virtio devices present offload features to guest, depending on the backend used. Thus, removing host netdev peer while guest is active leads to guest-visible inconsistency and/or crashes.

As a solution, while guest (NIC) peer device exists, we prevent the host peer from being deleted.
This patch does this by adding peer_deleted flag in nic state:
if host device is going away while guest device
is around, set this flag and keep a shell of
the host device around for as long as guest device exists.

The link is put down so all packets will get discarded.

At the moment, management can detect that device deletion is delayed by doing info net. As a next step, we shall add commands that control hotplug/unplug without removing the device, and an event to report that guest has responded to the hotplug event.

---
 net.c |   40 +++++++++++++++++++++++++++++++++-------
 net.h |    1 +
 2 files changed, 34 insertions(+), 7 deletions(-)
net.c +33 -7
index 3d0fde7..edb87af 100644
@@ -281,29 +281,55 @@ NICState *qemu_new_nic(NetClientInfo *info,
     return nic;
 }
 
-void qemu_del_vlan_client(VLANClientState *vc)
+static void qemu_free_vlan_client(VLANClientState *vc)
 {
     if (vc->vlan) {
         QTAILQ_REMOVE(&vc->vlan->clients, vc, next);
     } else {
+        QTAILQ_REMOVE(&non_vlan_clients, vc, next);
         if (vc->send_queue) {
             qemu_del_net_queue(vc->send_queue);
         }
-        QTAILQ_REMOVE(&non_vlan_clients, vc, next);
         if (vc->peer) {
             vc->peer->peer = NULL;
         }
     }
-
-    if (vc->info->cleanup) {
-        vc->info->cleanup(vc);
-    }
-
     qemu_free(vc->name);
     qemu_free(vc->model);
     qemu_free(vc);
 }
 
+void qemu_del_vlan_client(VLANClientState *vc)
+{
+    /* If there is a peer NIC, delete and cleanup client, but do not free. */
+    if (!vc->vlan && vc->peer && vc->peer->info->type == NET_CLIENT_TYPE_NIC) {
+        NICState *nic = DO_UPCAST(NICState, nc, vc->peer);
+        if (nic->peer_deleted) {
+            return;
+        }
+        nic->peer_deleted = true;
+        /* Let NIC know peer is gone. */
+        vc->peer->link_down = true;
+        if (vc->peer->info->link_status_changed) {
+            vc->peer->info->link_status_changed(vc->peer);
+        }
+        if (vc->info->cleanup) {
+            vc->info->cleanup(vc);
+        }
+        return;
+    }
+
+    /* If this is a peer NIC and peer has already been deleted, free it now. */
+    if (!vc->vlan && vc->peer && vc->info->type == NET_CLIENT_TYPE_NIC) {
+        NICState *nic = DO_UPCAST(NICState, nc, vc);
+        if (nic->peer_deleted) {
+            qemu_free_vlan_client(vc->peer);
+        }
+    }
+
+    qemu_free_vlan_client(vc);
+}
+
 VLANClientState *
 qemu_find_vlan_client_by_name(Monitor *mon, int vlan_id,
                               const char *client_str)
net.h +1 -0
index 518cf9c..44c31a9 100644
@@ -72,6 +72,7 @@ typedef struct NICState {
     VLANClientState nc;
     NICConf *conf;
     void *opaque;
+    bool peer_deleted;
 } NICState;
 
 struct VLANState {
Signature · 1 line
1.7.3-rc1
#1 lore raw
Quoted 62 lines · Michael S. Tsirkin

With -netdev, virtio devices present offload features to guest, depending on the backend used. Thus, removing host netdev peer while guest is active leads to guest-visible inconsistency and/or crashes.

As a solution, while guest (NIC) peer device exists, we prevent the host peer from being deleted.
This patch does this by adding peer_deleted flag in nic state:
if host device is going away while guest device
is around, set this flag and keep a shell of
the host device around for as long as guest device exists.

The link is put down so all packets will get discarded.

At the moment, management can detect that device deletion is delayed by doing info net. As a next step, we shall add commands that control hotplug/unplug without removing the device, and an event to report that guest has responded to the hotplug event.

---
 net.c |   40 +++++++++++++++++++++++++++++++++-------
 net.h |    1 +
 2 files changed, 34 insertions(+), 7 deletions(-)
net.c +5 -7
index 3d0fde7..edb87af 100644
@@ -281,29 +281,55 @@ NICState *qemu_new_nic(NetClientInfo *info,
     return nic;
 }
 
-void qemu_del_vlan_client(VLANClientState *vc)
+static void qemu_free_vlan_client(VLANClientState *vc)
 {
     if (vc->vlan) {
         QTAILQ_REMOVE(&vc->vlan->clients, vc, next);
     } else {
+        QTAILQ_REMOVE(&non_vlan_clients, vc, next);
         if (vc->send_queue) {
             qemu_del_net_queue(vc->send_queue);
         }
-        QTAILQ_REMOVE(&non_vlan_clients, vc, next);
         if (vc->peer) {
             vc->peer->peer = NULL;
         }
     }
-
-    if (vc->info->cleanup) {
-        vc->info->cleanup(vc);
-    }
-
     qemu_free(vc->name);
     qemu_free(vc->model);
     qemu_free(vc);
 }
 
+void qemu_del_vlan_client(VLANClientState *vc)
+{
+    /* If there is a peer NIC, delete and cleanup client, but do not free. */

net.c NICState *qemu_new_nic(NetClientInfo *info, · line 314

Quoted 10 lines
net.c +10 -0
+    if (!vc->vlan && vc->peer && vc->peer->info->type == NET_CLIENT_TYPE_NIC) {
+        NICState *nic = DO_UPCAST(NICState, nc, vc->peer);
+        if (nic->peer_deleted) {
+            return;
+        }
+        nic->peer_deleted = true;
+        /* Let NIC know peer is gone. */
+        vc->peer->link_down = true;
+        if (vc->peer->info->link_status_changed) {
+            vc->peer->info->link_status_changed(vc->peer);

Are we potentially introducing more races here? Now we have the tap trying to call link change on a peer device that's potentially in the middle of going away. Thanks,

Alex

Quoted 33 lines
net.c +18 -0
+        }
+        if (vc->info->cleanup) {
+            vc->info->cleanup(vc);
+        }
+        return;
+    }
+
+    /* If this is a peer NIC and peer has already been deleted, free it now. */
+    if (!vc->vlan && vc->peer && vc->info->type == NET_CLIENT_TYPE_NIC) {
+        NICState *nic = DO_UPCAST(NICState, nc, vc);
+        if (nic->peer_deleted) {
+            qemu_free_vlan_client(vc->peer);
+        }
+    }
+
+    qemu_free_vlan_client(vc);
+}
+
 VLANClientState *
 qemu_find_vlan_client_by_name(Monitor *mon, int vlan_id,
                               const char *client_str)
net.h +1 -0
index 518cf9c..44c31a9 100644
@@ -72,6 +72,7 @@ typedef struct NICState {
     VLANClientState nc;
     NICConf *conf;
     void *opaque;
+    bool peer_deleted;
 } NICState;
 
 struct VLANState {
#2 lore raw
Quoted 78 lines · Alex Williamson
Quoted 62 lines · Michael S. Tsirkin

With -netdev, virtio devices present offload features to guest, depending on the backend used. Thus, removing host netdev peer while guest is active leads to guest-visible inconsistency and/or crashes.

As a solution, while guest (NIC) peer device exists, we prevent the host peer from being deleted.
This patch does this by adding peer_deleted flag in nic state:
if host device is going away while guest device
is around, set this flag and keep a shell of
the host device around for as long as guest device exists.

The link is put down so all packets will get discarded.

At the moment, management can detect that device deletion is delayed by doing info net. As a next step, we shall add commands that control hotplug/unplug without removing the device, and an event to report that guest has responded to the hotplug event.

---
 net.c |   40 +++++++++++++++++++++++++++++++++-------
 net.h |    1 +
 2 files changed, 34 insertions(+), 7 deletions(-)
net.c +5 -7
index 3d0fde7..edb87af 100644
@@ -281,29 +281,55 @@ NICState *qemu_new_nic(NetClientInfo *info,
     return nic;
 }
 
-void qemu_del_vlan_client(VLANClientState *vc)
+static void qemu_free_vlan_client(VLANClientState *vc)
 {
     if (vc->vlan) {
         QTAILQ_REMOVE(&vc->vlan->clients, vc, next);
     } else {
+        QTAILQ_REMOVE(&non_vlan_clients, vc, next);
         if (vc->send_queue) {
             qemu_del_net_queue(vc->send_queue);
         }
-        QTAILQ_REMOVE(&non_vlan_clients, vc, next);
         if (vc->peer) {
             vc->peer->peer = NULL;
         }
     }
-
-    if (vc->info->cleanup) {
-        vc->info->cleanup(vc);
-    }
-
     qemu_free(vc->name);
     qemu_free(vc->model);
     qemu_free(vc);
 }
 
+void qemu_del_vlan_client(VLANClientState *vc)
+{
+    /* If there is a peer NIC, delete and cleanup client, but do not free. */

net.c NICState *qemu_new_nic(NetClientInfo *info, · line 314

Quoted 10 lines
net.c +10 -0
+    if (!vc->vlan && vc->peer && vc->peer->info->type == NET_CLIENT_TYPE_NIC) {
+        NICState *nic = DO_UPCAST(NICState, nc, vc->peer);
+        if (nic->peer_deleted) {
+            return;
+        }
+        nic->peer_deleted = true;
+        /* Let NIC know peer is gone. */
+        vc->peer->link_down = true;
+        if (vc->peer->info->link_status_changed) {
+            vc->peer->info->link_status_changed(vc->peer);

Are we potentially introducing more races here? Now we have the tap trying to call link change on a peer device that's potentially in the middle of going away. Thanks,

Quoted 1 line

Alex

You mean in guest?
This could always happen anyway: user can call link state change command. So drivers must be ready for this.
There's no race in qemu because everything is singlethreaded.

Quoted 35 lines
Quoted 33 lines
net.c +18 -0
+        }
+        if (vc->info->cleanup) {
+            vc->info->cleanup(vc);
+        }
+        return;
+    }
+
+    /* If this is a peer NIC and peer has already been deleted, free it now. */
+    if (!vc->vlan && vc->peer && vc->info->type == NET_CLIENT_TYPE_NIC) {
+        NICState *nic = DO_UPCAST(NICState, nc, vc);
+        if (nic->peer_deleted) {
+            qemu_free_vlan_client(vc->peer);
+        }
+    }
+
+    qemu_free_vlan_client(vc);
+}
+
 VLANClientState *
 qemu_find_vlan_client_by_name(Monitor *mon, int vlan_id,
                               const char *client_str)
net.h +1 -0
index 518cf9c..44c31a9 100644
@@ -72,6 +72,7 @@ typedef struct NICState {
     VLANClientState nc;
     NICConf *conf;
     void *opaque;
+    bool peer_deleted;
 } NICState;
 
 struct VLANState {