I've attached my latest lockmeter data from 2.5.8-pre1.  This is an 
8-way PIII Xeon machine, running dbench 16.  The clearest win that I saw 
for the BKL between the 2.5.7 and the current 2.5.8-pre1 is in do_exit. 
   The BKL was removed from do_exit, and pushed down into 2 other 
functions.  Granted, disassociate_ctty and sem_exit weren't called loads 
of times, and the tasks probably weren't holding a lot of semaphores, 
but that is part of the point.

2.5.7:               avg hold(   max)
do_exit+0xd8           5003us(  19ms)

reduced to this in 2.5.8-pre1:
disassociate_ctty+0x28  1.8us( 1.8us)
sem_exit+0x24           1.5us( 3.0us)

As for other locks, kmap_lock, lru_list_lock, and pagecache_lock are the 
ones that stick out most.  But, things are looking better.  Are there 
any workloads which people are really interested in?  Name it, and I'll 
run it on the 8-way Xeon and a 16-way NUMA-Q.

I also ported lockmeter to 2.5.8-pre1.  If you want it, I've got it.

-- 
Dave Hansen
haveblue@us.ibm.com