Historically TCP has been limited to 64K buffers, but increasing
msize provides huge performance benefits especially as latency
increase so allow for bigger buffers.

Ideally further improvements could change the allocation from the
current contiguous chunk in slab (kmem_cache) to some scatter-gather
compatible API...

Note this only increases the max possible setting, not the default
value.

Link: http://lkml.kernel.org/r/YTQB5jCbvhmCWzNd@codewreck.orgSigned-off-by: Dominique Martinet <asmadeus@codewreck.org>