pageout question

[Andriy Gapon]

on 25/07/2010 16:41 RW said the following:
> On Sun, 25 Jul 2010 13:07:21 +0300
> Andriy Gapon <avg at freebsd.org> wrote:
> 
>> on 25/07/2010 02:31 RW said the following:
> 
>>> As I understand it the hysteresis is done inside vm_pageout_scan,
>>> and the expectation is that one pass will typically satisfy this
>>> because the design aims to keep enough clean pages in the inactive
>>> queue.  
> 
>> But I am not sure about "clean pages in the inactive queue" ... But I
>> do not see any code ensuring level of _clean_ inactive pages. 
> 
> In FreeBSD the inactive queue contains disk cache pages which normally
> provide most of the clean pages needed. In addition pages are dribbled
> out to swap, and the resulting clean pages are placed at the back of
> the inactive queue to make another pass. 

Well, "normally" and "most" are not quite quantitative.
Personally, I do not see any guarantees that inactive queue would contain enough
clean pages to reach paging target on a single pass.

>>> I'm not sure if  the vm_paging_needed() call is correct or not, but
>>> it may be that that the intent is to avoid immediately going back
>>> to a depleted inactive queue when cache+free is within normal
>>> bounds, because it could result in avoidable paging to swap. 
>> Well, OTOH, if the current pass results in many pages being
>> re-activated and many pages still left on the inactive queue because
>> they are dirty (see maxlaunder in vm_pageout_scan), 
> 
> Dirty-pages  make three passes through the inactive queue: twice dirty,
> once clean. They are paged-out at the end of the second paass, so it's
> unlike that they reactivated except under very heavy thrashing. 

I didn't mean to say that dirty pages would get re-activated.
Clean pages can perfectly be re-activated if they were referenced since their
de-activation time.

>> then it is
>> premature to quit paging when we only reached bare minimum of
>> available pages (see pass and maxlaunder again).  IMHO, of course.
> 
> It's not the bare minimum, that's another level that vm_page_count_min()
> tests for.

I meant bare minimum to stop paging, that is, going above lower watermark of the
paging hysteresis.

-- 
Andriy Gapon