Any thoughts on removing the dependency on srv_mutex in deleting
completed requests (and freeing its memory) ? Otherwise it can cause
problems with long running socket writes (on sending new requests)
which don't get the benefit of finishing up response processing due to
the blocking of DeleteMidQEntry on server->srv_mutex

This should improve performance as well.

(Long Li was noticing this looking at RDMA with cifs.ko)

See attached patch

-- 
Thanks,

Steve