cifs compounding

Big performance benefit (for this first narrow case) with Ronnie's
compounding patches (now merged into cifs-2.6.git for-next). This is
the first code path enabled for smb3 compounding but looks very promising, if we continue to get such dramatic perf improvements (as we go forward implementing compounding to help stat, readdir etc.). Note that we have seen that Macs are quite aggressive about compounding in SMB3 (and Windows benefits a lot as well) so there a quite a few future code paths that we can optimize. It obviously helps with more than network latency though, it can reduce server load as well. Additional improvements in handle caching and directory leases (currently only implemented for the root directory) will also be a big win in the future as they are extended in cifs.ko for SMB3.

I did some primitive tests with and without Ronnie's patch and it was
from 40% to 2000% slower before his patch (on the wire). I tried df
and stat -f. Note that stat -f shows the greater benefit (only one compounded request on the with his patch, instead of two requests only one of which is compounded with df)

Without his patch:
local (fast network, to Samba) ~2/100ths of a second remote (Azure, encrypted) - 16th/100ths to 17/100ths of a second

With his patch
local (fast network, to Samba) ~1/1000 of a second remote (Azure, encrypted) ~6/100ths to ~12/100ths (stat -f vs. df)

so basically "stat -f" on the root goes from 3 requests to 1 send on the wire, while df goes from 4 requests to 2.

Good job Ronnie!

--
Thanks,

Steve