Socket hashing patches, 5th and final installment

David S. Miller (
Mon, 3 Mar 1997 01:54:31 -0500

[ Linus please include this in 2.1.28, thanks... ]

It is getting to the point where I am fixing bugs that were in the
networking code before I made any changes ;-) This does:

1) Final refinements to tcp_good_socknum(), it is not worth improving
any further. (me)

2) linger option processing uses err before even setting it (eddie dost)

3) Someone buggered ip_setsockopt() big time, traceroute and IP socket
options work once again in 2.1.x (me)

4) "sock destroy delayed" killed for good, wrong skb charged to sk in
syn/ack transmissions (me)

5) Missing symbols added so ipx/appletalk/ipv6 have all their
references, this was due to all the sock_no_* stuff that recently
got added, shame on you alan ;-) (me)

Here is my current plan. I want this code to become tested heavily.
I will be running some high load benchmarks, I have high hopes for
things like WebSTONE etc. I have verified that none of our networking
latencies/bandwidth have suffered at _all_ due to my changes. In fact
connection latencies went down by 100usec on my 133MHZ hypersparc test
machine. Beat that.

Should no super major bugs be reported due to my code, these socket
hashing changes are to be back ported to 2.0.x for a special 2.0.x
release which Alan and myself have coined "Linux 2.0.30 ISS" ISS
stands for "Internet Server Supplement", it comes from what Sun calls
their server performance patch set for Solaris-2.1.5 ;-)

The 2.0.30 ISS will go through an initial testing period, probably
very short in fact. The ISS patches will also include Finn Gangstad's
scalable timers. Once the testing period is over (given Alan and
Linus's blessings) all those running large scale, heavily loaded Linux
internet servers will be encouraged to run this special patch set. It
improves performance on large internet servers by an alarmingly
considerable amount (ie. it's much faster than FreeBSD ;-)

Dig in kids, we scale...

Yow! 11.26 MB/s remote host TCP bandwidth & ////
199 usec remote TCP latency over 100Mb/s ////
ethernet. Beat that! ////
-----------------------------------------////__________ o
David S. Miller, /_____________/ / // /_/ ><