hi,

>  Only two files changed in this time interval: uvm/uvm_page.c and uvm/ 
>  uvm_pglist.c, each changed in one line aside from CVS keywords.   
>  Version 1.100 of uvm_page.c and version 1.32 of uvm_pglist.c have  
>  this log message:
>  
>  date: 2004/09/17 20:46:03;  author: yamt;  state: Exp;  lines: +3 -3
>  make free page queue filo rather than fifo.
>  data in pages freed more recently are more likely on cpu cache.

i hardly see how it causes the symptom.
i think it discovered a bug in somewhere else as you said.

>  The uvm_page.c change itself seems logical.  Assuming it's actually  
>  correct, I would guess that my problem means that some page is being  
>  put onto the free list, and probably allocated again, while some  
>  other part of the kernel (or a DMA device) isn't done with it yet,  
>  and the FIFO version of the queue happens to give the extra time  
>  needed.  Or maybe it's bad memory and during the boot process one  
>  pattern of usage trips over it consistently and the other pattern (as  
>  well as running NetBSD 2.0 and doing nightly builds of some code I  
>  work on) does not in any noticeable way.  But I think I'm done for  
>  tonight....

because uvm_page_physload also uses uvm_pagefree,
boot process likely uses very different set of pages after the change.

YAMAMOTO Takashi