Description of problem:
when sending fragmented packets over multicast sockets, if the
IP_MULTICAST_LOOP option is enabled, each skbuff is cloned and
fragmented.  The accounting for the sockets SNDBUF is preformed on
each fragment, but after the frame is transmitted the kfree_skb call
fails to preform the reverse accounting, since the cloned skbuffs are
not freed, only the top level skbuff representing the entire frame. 
Only after the socket is read from, clearing the received skbuff from
the receive queue, is the reverse accounting preformed.  This can lead
to inacurate sndbuf values, and if the socket is not read from often
enough, the user process using the socket can block indefinately on a
send(to/msg) call.

Version-Release number of selected component (if applicable):
all


How reproducible:
always

Steps to Reproduce:
1) Remove iptables modules from the system (iptables can prematurely
free buffers to prevent queue backlog).

2) build and run the attached testcase (multicast_test.C).  Specify
when running a large packet size (I've been using 10000 bytes) so as
to incur fragmenting.  Also specify that IP_MUTLICAST_LOOP should be
set, and that the reader thread should never read.



Actual results:
After a few packets, the call to sendto will block forever, waiting
for available memory in the SNDBUF buffer.  This will not be made
available until the socket is read from

Expected results:
Socket sends only block until frames are transmitted, at which point
memory should be available in the send buffer.

Additional info: