[PATCH] __alloc_pages_limit pages_min

[PATCH] __alloc_pages_limit pages_min

[Roger Larsson]

Hi all,

If __alloc_pages_limit is called with PAGES_HIGH and direct_reclaim.
It might alloc the last free page of a zone....

How:
* inactive_clean_pages >= pages_high

first if gives

if (free_pages + inactive_clean >= pages_high) => true

No direct_reclaim gives no attempt to call reclaim_page.
Instead rmqueue will be called...
now we have one page less - run it again...
free_pages will be decremented until zero...

Last page gone in a at PAGES_HIGH alloc!!!

Note: reclaim_page will fix this situation direct it is allowed to
run since it is kicked in __alloc_pages. But since we cannot
guarantee that this will never happen...

/RogerL

this patch is against 2.4.8-pre3 as it was my current kernel.
(searching for a USB storage bug...)
*******************************************
Patch prepared by: roger.larsson@norran.net

--- linux/mm/page_alloc.c.orig	Wed Aug 22 13:36:57 2001
+++ linux/mm/page_alloc.c	Wed Aug 22 13:50:31 2001
@@ -256,8 +256,9 @@
 			/* If possible, reclaim a page directly. */
 			if (direct_reclaim)
 				page = reclaim_page(z);
-			/* If that fails, fall back to rmqueue. */
-			if (!page)
+			/* If that fails, fall back to rmqueue, but do never
+			*  go below free_pages for any zone*/
+			if (!page && z->free_pages >= z->pages_min)
 				page = rmqueue(z, order);
 			if (page)
 				return page;
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/

Re: [PATCH] __alloc_pages_limit pages_min

[Rik van Riel]

On Wed, 22 Aug 2001, Roger Larsson wrote:

> Note: reclaim_page will fix this situation direct it is allowed to
> run since it is kicked in __alloc_pages. But since we cannot
> guarantee that this will never happen...

In this case kreclaimd will be woken up and the free pages
will be refilled.

Rik
--
IA64: a worthy successor to the i860.

		http://www.surriel.com/
http://www.conectiva.com/	http://distro.conectiva.com/

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/

Re: [PATCH] __alloc_pages_limit pages_min

[Roger Larsson]

On Wednesdayen den 22 August 2001 19:01, Rik van Riel wrote:
> On Wed, 22 Aug 2001, Roger Larsson wrote:
> > Note: reclaim_page will fix this situation direct it is allowed to
> > run since it is kicked in __alloc_pages. But since we cannot
> > guarantee that this will never happen...
>
> In this case kreclaimd will be woken up and the free pages
> will be refilled.
>
> Rik

Yes it will be woken up - but when will it actually do something?
(we might have lots of stuff before in the run queue)

Suppose we are at low for all zones on free pages - the natural.
And we get requsts for DMA pages, non direct_reclaim.
(with inactive_clean greater than pages_high)

An alloc will then take let you allocate at the first __alloc_pages_limit
until zero pages left. You might have got it anyway but not until a
later test in __alloc_pages but it is not guaranteed.

We should not accept to let free_pages go (more than one) under
pages_min, at least not as a undocumented feature of __alloc_pages_limit...

It might be worse then first tought of since all allocations with order != 0
are non direct_reclaim... you might eat the free pages fast...

And this limit at the end of alloc_pages
		if (z->free_pages < z->pages_min / 4 &&
				!(current->flags & PF_MEMALLOC))
is not enforced earlier in the same code...

/RogerL

-- 
Roger Larsson
Skelleftea
Sweden
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/

Re: [PATCH] __alloc_pages_limit pages_min

[Rik van Riel]

On Wed, 22 Aug 2001, Roger Larsson wrote:
> On Wednesdayen den 22 August 2001 19:01, Rik van Riel wrote:
> > On Wed, 22 Aug 2001, Roger Larsson wrote:
> > > Note: reclaim_page will fix this situation direct it is allowed to
> > > run since it is kicked in __alloc_pages. But since we cannot
> > > guarantee that this will never happen...
> >
> > In this case kreclaimd will be woken up and the free pages
> > will be refilled.
>
> Yes it will be woken up - but when will it actually do something?

> And this limit at the end of alloc_pages
> 		if (z->free_pages < z->pages_min / 4 &&
> 				!(current->flags & PF_MEMALLOC))
> is not enforced earlier in the same code...

Please read the code.  The first loop in __alloc_pages(),
before we even call __alloc_pages_limit() will wake up
kreclaimd as soon as 'z->free_pages < z->pages_min'.

If you have any more questions about the source code,
don't hesitate to ask ;)

Rik
--
IA64: a worthy successor to the i860.

		http://www.surriel.com/
http://www.conectiva.com/	http://distro.conectiva.com/

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/