> 
> 
> 
> the tsleep() on lbolt appears twice in vm_output.c (both with VM_PAGER_AGAIN).
                                         ^^^^^^^^^^^ vm_pageout.c ? :-).
> 
> does your suggested change need to be applied at both places ?
> 

Sure, sounds good!!!  I have initially gotten positive feedback
on the change, but it still isn't really blocking on the precisely
correct resource.  This will get you'all through the worst of it,
but the only reason why this is working is because the rundown code
for the swap_pager.c wakes up on the pages needed variable.  The way
that it is being used it is a side-effect that it works (and is 
not "invalid"), but could be restructured to afford a significantly
greater performance improvement.

Since I don't have a NetBSD system (and FreeBSD's VM is almost
totally different now), I could only guess at how to restructure it.

John