On 02/14/2011 02:51 PM, Daniel P. Berrange wrote:
This approach feels like a nasty hack to me and potentially still leaves us with a problem in netcf which is also using netlink sockets. I think we need to get a clearer picture of what the root cause is before going for this kind of patch
Correct, I am 'fixing' this in the wrong place. The issues is in the call sequence nl_handle = nl_handle_alloc() nl_connect(nl_handle, NETLINK_ROUTE) with the second one failing taking merely input from the 1st one. These are obviously two libnl calls. Something is either not using libn or not using it correctly. Thanks for pointing out netcf. I looked at libnetcf code and found this sequence here: [...] int netlink_init(struct netcf *ncf) { ncf->driver->nl_sock = nl_handle_alloc(); if (ncf->driver->nl_sock == NULL) goto error; if (nl_connect(ncf->driver->nl_sock, NETLINK_ROUTE) < 0) { goto error; } This seems to be doing the same as I do. Maybe there is yet 'something else' that's using netlink sockets. What's also strange is that the first 'virsh start' still works, but the subsequent 'virsh destroy' then does not. Stefan