DEFAULT_RCVBUF can be (again) small with more than 127 VFs

[elajkat]

In our lab we had NICs with more and more VFs with recent increase of DEFAULT_RCVBUF to 65536, we can have more than 127 VFs. The previous attempts to push out this limit (#813 or #751 ) just pushed out the issue but not solved it.

One of our experts checked the problem, and suggested a solution to detect the real buffer size for receiving netlink, something like this in NetlinkSocketBase:
from socket import MSG_DONTWAIT, MSG_PEEK, MSG_TRUNC
def peek_bufsize(socket_descriptor):
data = bytearray()
bufsize, _ = socket_descriptor.recvfrom_into(data, 0, MSG_DONTWAIT | MSG_PEEK | MSG_TRUNC)
return bufsize

def recv(self, *argv, **kwarg):
....
return self._sock.recv(self._peek_bufsize(self._sock), **kwarg)
# return self._sock.recv(*argv, **kwarg)

This can be easily check in a dummy environment (I used a VM without sriov or any extra HW):
import sys
from pyroute2 import IPRoute

device_idx = int(sys.argv[1])
with IPRoute() as ip:
print(ip.link('get', index=device_idx, ext_mask=1))

And changing DEFAULT_RCVBUF to 1024.
I got NLMSG_ERROR with the original code, and with peek_bufsize I just got the listing of links as I expected.