Ok. The tcp_input.c patch should be totally innocent, as people have
verified that X over TCP is as fast in 24 as it was in 23 (easy to do:
do a
export DISPLAY=localhost:0
rather than just the normal ":0" to make the program use TCP rather than UNIX
domain sockets)
However, the net/sock.h change does two separate things:
- the TCP sequence number addition for ka9q-based stacks
- a 16->32 bit counter change
Can you test if the detrimental effect comes from _only_ the short->int
changes? In that case it seems like the 24 behaviour is just due to
better counters, and maybe the slowdown is just due to a limit check that
never triggered due to 16-bit overflows or something like that..
Linus