net: do not delete nics in net_cleanup()

In net_cleanup() we only need to delete the netdevs, as those may have
state which outlives Qemu when it exits, and thus may actually need to
be cleaned up on exit.

The nics, on the other hand, are owned by the device which created them.
Most devices don't bother to clean up on exit because they don't have
any state which will outlive Qemu... but XenBus devices do need to clean
up their nodes in XenStore, and do have an exit handler to delete them.

When the XenBus exit handler destroys the xen-net-device, it attempts
to delete its nic after net_cleanup() had already done so. And crashes.

Fix this by only deleting netdevs as we walk the list. As the comment
notes, we can't use QTAILQ_FOREACH_SAFE() as each deletion may remove
*multiple* entries, including the "safely" saved 'next' pointer. But
we can store the *previous* entry, since nics are safe.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>
Reviewed-by: Paul Durrant <paul@xen.org>
Signed-off-by: Jason Wang <jasowang@redhat.com>

1 files changed, 22 insertions, 6 deletions

diff --git a/net/net.c b/net/net.c
index 4d1ff7a..0520bc1 100644
--- a/net/net.c
+++ b/net/net.c
@@ -1514,18 +1514,34 @@ static void net_vm_change_state_handler(void *opaque, bool running,
 
 void net_cleanup(void)
 {
-    NetClientState *nc;
+    NetClientState *nc, **p = &QTAILQ_FIRST(&net_clients);
 
     /*cleanup colo compare module for COLO*/
     colo_compare_cleanup();
 
-    /* We may del multiple entries during qemu_del_net_client(),
-     * so QTAILQ_FOREACH_SAFE() is also not safe here.
+    /*
+     * Walk the net_clients list and remove the netdevs but *not* any
+     * NET_CLIENT_DRIVER_NIC entries. The latter are owned by the device
+     * model which created them, and in some cases (e.g. xen-net-device)
+     * the device itself may do cleanup at exit and will be upset if we
+     * just delete its NIC from underneath it.
+     *
+     * Since qemu_del_net_client() may delete multiple entries, using
+     * QTAILQ_FOREACH_SAFE() is not safe here. The only safe pointer
+     * to keep as a bookmark is a NET_CLIENT_DRIVER_NIC entry, so keep
+     * 'p' pointing to either the head of the list, or the 'next' field
+     * of the latest NET_CLIENT_DRIVER_NIC, and operate on *p as we walk
+     * the list.
+     *
+     * The 'nc' variable isn't part of the list traversal; it's purely
+     * for convenience as too much '(*p)->' has a tendency to make the
+     * readers' eyes bleed.
      */
-    while (!QTAILQ_EMPTY(&net_clients)) {
-        nc = QTAILQ_FIRST(&net_clients);
+    while (*p) {
+        nc = *p;
         if (nc->info->type == NET_CLIENT_DRIVER_NIC) {
-            qemu_del_nic(qemu_get_nic(nc));
+            /* Skip NET_CLIENT_DRIVER_NIC entries */
+            p = &QTAILQ_NEXT(nc, next);
         } else {
             qemu_del_net_client(nc);
         }