Roger Larsson wrote:
> 
> Hi,
> 
> The attached output shows that when we hit swap - there are
> code lines with latency problems :-(
> [the actual code tested is test3-pre2 with my latency modifications
>  (improvement and profiling) but has one modification relative
> test3-pre4
>  kswapd in the tested version always sleeps => problems accounted
>  to the process causing it]

I'm losing sleep over kswapd.

> see the 293ms in generic_make_request...

Try adding the attached patch to your existing tree.  It fixes a lot of stuff.  sys_close() and sys_exit() still need attention.

> and the 704ms used to busy loop in modprobe...
> (SB16 non PnP)

Don't worry about it.

> These are worse then the previously found aux_write_dev :-(

Or this.