See https://www.illumos.org/issues/1517
This change was previously reviewed back in Sept. 2011.  See:

http://www.listbox.com/member/archive/182179/2011/09/sort/time_rev/page/1/?search_for=1517

This code has been in production for a couple years.
We have also done some studies of how the server behaves as it approaches saturation of the I/O capacity of the disk subsystem.

With this change, the server gracefully degrades service.  As load increases, SMB clients see gradually increasing latency.

Without this change, the server basically runs itself out of memory as the SMB receiver threads try to "buffer" too much.