[PATCH rc5-mm1 1/3] mm-have-zonelist: fix memcg ooms

[PATCH rc5-mm1 1/3] mm-have-zonelist: fix memcg ooms

[Hugh Dickins]

Memcgroups up against their limit were OOMing unjustifiably on x86_32
with highmem: pages ripe for freeing were not being found.

It's because the zonelist changes in vmscan.c now derive zone from gfp_mask,
and the gfp mask being passed into try_to_free_mem_cgroup_pages is atuned
to allocating a struct page_cgroup (__GFP_HIGHMEM masked off), whereas here
we need to be scanning highmem zones.

Frankenstein a gfp_mask from GFP_HIGHUSER_MOVABLE and the mask passed in.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
---
Same error was in -rc3-mm1, sorry for not sending fix in time for -rc5-mm1.
The error comes 1 or 4 patches earlier, but this patch atop 2.6.25-rc5-mm1
fits best just after
mm-have-zonelist-contains-structs-with-both-a-zone-pointer-and-zone_idx.patch

I've other GFP masking fixes or cleanups to come another day: this fixes
just one clear bug, peculiar to the -mm tree, orthogonal to the rest.
But two little do_try_to_free_pages cleanups follow on from this...

 mm/vmscan.c |    6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

--- 2.6.25-rc5-mm1/mm/vmscan.c	2008-03-11 14:19:36.000000000 +0000
+++ linux/mm/vmscan.c	2008-03-11 20:09:05.000000000 +0000
@@ -1444,7 +1444,6 @@ unsigned long try_to_free_mem_cgroup_pag
 						gfp_t gfp_mask)
 {
 	struct scan_control sc = {
-		.gfp_mask = gfp_mask,
 		.may_writepage = !laptop_mode,
 		.may_swap = 1,
 		.swap_cluster_max = SWAP_CLUSTER_MAX,
@@ -1454,9 +1453,10 @@ unsigned long try_to_free_mem_cgroup_pag
 		.isolate_pages = mem_cgroup_isolate_pages,
 	};
 	struct zonelist *zonelist;
-	int target_zone = gfp_zonelist(GFP_HIGHUSER_MOVABLE);
 
-	zonelist = &NODE_DATA(numa_node_id())->node_zonelists[target_zone];
+	sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
+			(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
+	zonelist = NODE_DATA(numa_node_id())->node_zonelists;
 	if (do_try_to_free_pages(zonelist, sc.gfp_mask, &sc))
 		return 1;
 	return 0;

[PATCH rc5-mm1 2/3] just return do_try_to_free_pages

[Hugh Dickins]

do_try_to_free_pages returns 0 or 1, and try_to_free_mem_cgroup_pages
returns 0 or 1 accordingly: let's accomplish that in the obvious way.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
---

 mm/vmscan.c |    4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

--- 2.6.25-rc5-mm1+/mm/vmscan.c	2008-03-11 20:09:05.000000000 +0000
+++ linux/mm/vmscan.c	2008-03-11 20:15:47.000000000 +0000
@@ -1457,9 +1457,7 @@ unsigned long try_to_free_mem_cgroup_pag
 	sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
 			(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
 	zonelist = NODE_DATA(numa_node_id())->node_zonelists;
-	if (do_try_to_free_pages(zonelist, sc.gfp_mask, &sc))
-		return 1;
-	return 0;
+	return do_try_to_free_pages(zonelist, sc.gfp_mask, &sc);
 }
 #endif
 

Re: [PATCH rc5-mm1 2/3] just return do_try_to_free_pages

[Mel Gorman]

On (11/03/08 21:14), Hugh Dickins didst pronounce:
> do_try_to_free_pages returns 0 or 1, and try_to_free_mem_cgroup_pages
> returns 0 or 1 accordingly: let's accomplish that in the obvious way.
> 
> Signed-off-by: Hugh Dickins <hugh@veritas.com>

Looks right, it means it's behaviour closer matches try_to_free_pages()
which is a bonus.

Acked-by: Mel Gorman <mel@csn.ul.ie>

> ---
> 
>  mm/vmscan.c |    4 +---
>  1 file changed, 1 insertion(+), 3 deletions(-)
> 
> --- 2.6.25-rc5-mm1+/mm/vmscan.c	2008-03-11 20:09:05.000000000 +0000
> +++ linux/mm/vmscan.c	2008-03-11 20:15:47.000000000 +0000
> @@ -1457,9 +1457,7 @@ unsigned long try_to_free_mem_cgroup_pag
>  	sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
>  			(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
>  	zonelist = NODE_DATA(numa_node_id())->node_zonelists;
> -	if (do_try_to_free_pages(zonelist, sc.gfp_mask, &sc))
> -		return 1;
> -	return 0;
> +	return do_try_to_free_pages(zonelist, sc.gfp_mask, &sc);
>  }
>  #endif
>  
> 

-- 
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab

[PATCH rc5-mm1 3/3] do_try_to_free_pages gfp_mask redundant

[Hugh Dickins]

do_try_to_free_pages is taking both a gfp_mask and an sc->gfp_mask,
sometimes one used and sometimes the other; at present they're always
identical, but such duplication begs for trouble: remove gfp_mask arg.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
---

 mm/vmscan.c |   10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

--- 2.6.25-rc5-mm1++/mm/vmscan.c	2008-03-11 20:15:47.000000000 +0000
+++ linux/mm/vmscan.c	2008-03-11 20:21:14.000000000 +0000
@@ -1324,7 +1324,7 @@ static unsigned long shrink_zones(int pr
  * allocation attempt will fail.
  */
 static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
-					gfp_t gfp_mask, struct scan_control *sc)
+					struct scan_control *sc)
 {
 	int priority;
 	int ret = 0;
@@ -1334,7 +1334,7 @@ static unsigned long do_try_to_free_page
 	unsigned long lru_pages = 0;
 	struct zoneref *z;
 	struct zone *zone;
-	enum zone_type high_zoneidx = gfp_zone(gfp_mask);
+	enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
 
 	if (scan_global_lru(sc))
 		count_vm_event(ALLOCSTALL);
@@ -1363,7 +1363,7 @@ static unsigned long do_try_to_free_page
 		 * over limit cgroups
 		 */
 		if (scan_global_lru(sc)) {
-			shrink_slab(sc->nr_scanned, gfp_mask, lru_pages);
+			shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
 			if (reclaim_state) {
 				nr_reclaimed += reclaim_state->reclaimed_slab;
 				reclaim_state->reclaimed_slab = 0;
@@ -1435,7 +1435,7 @@ unsigned long try_to_free_pages(struct z
 		.isolate_pages = isolate_pages_global,
 	};
 
-	return do_try_to_free_pages(zonelist, gfp_mask, &sc);
+	return do_try_to_free_pages(zonelist, &sc);
 }
 
 #ifdef CONFIG_CGROUP_MEM_RES_CTLR
@@ -1457,7 +1457,7 @@ unsigned long try_to_free_mem_cgroup_pag
 	sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
 			(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
 	zonelist = NODE_DATA(numa_node_id())->node_zonelists;
-	return do_try_to_free_pages(zonelist, sc.gfp_mask, &sc);
+	return do_try_to_free_pages(zonelist, &sc);
 }
 #endif
 

Re: [PATCH rc5-mm1 3/3] do_try_to_free_pages gfp_mask redundant

[Mel Gorman]

On (11/03/08 21:16), Hugh Dickins didst pronounce:
> do_try_to_free_pages is taking both a gfp_mask and an sc->gfp_mask,
> sometimes one used and sometimes the other; at present they're always
> identical, but such duplication begs for trouble: remove gfp_mask arg.
> 
> Signed-off-by: Hugh Dickins <hugh@veritas.com>

Acked-by: Mel Gorman <mel@csn.ul.ie>

> ---
> 
>  mm/vmscan.c |   10 +++++-----
>  1 file changed, 5 insertions(+), 5 deletions(-)
> 
> --- 2.6.25-rc5-mm1++/mm/vmscan.c	2008-03-11 20:15:47.000000000 +0000
> +++ linux/mm/vmscan.c	2008-03-11 20:21:14.000000000 +0000
> @@ -1324,7 +1324,7 @@ static unsigned long shrink_zones(int pr
>   * allocation attempt will fail.
>   */
>  static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
> -					gfp_t gfp_mask, struct scan_control *sc)
> +					struct scan_control *sc)
>  {
>  	int priority;
>  	int ret = 0;
> @@ -1334,7 +1334,7 @@ static unsigned long do_try_to_free_page
>  	unsigned long lru_pages = 0;
>  	struct zoneref *z;
>  	struct zone *zone;
> -	enum zone_type high_zoneidx = gfp_zone(gfp_mask);
> +	enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
>  
>  	if (scan_global_lru(sc))
>  		count_vm_event(ALLOCSTALL);
> @@ -1363,7 +1363,7 @@ static unsigned long do_try_to_free_page
>  		 * over limit cgroups
>  		 */
>  		if (scan_global_lru(sc)) {
> -			shrink_slab(sc->nr_scanned, gfp_mask, lru_pages);
> +			shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
>  			if (reclaim_state) {
>  				nr_reclaimed += reclaim_state->reclaimed_slab;
>  				reclaim_state->reclaimed_slab = 0;
> @@ -1435,7 +1435,7 @@ unsigned long try_to_free_pages(struct z
>  		.isolate_pages = isolate_pages_global,
>  	};
>  
> -	return do_try_to_free_pages(zonelist, gfp_mask, &sc);
> +	return do_try_to_free_pages(zonelist, &sc);
>  }
>  
>  #ifdef CONFIG_CGROUP_MEM_RES_CTLR
> @@ -1457,7 +1457,7 @@ unsigned long try_to_free_mem_cgroup_pag
>  	sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
>  			(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
>  	zonelist = NODE_DATA(numa_node_id())->node_zonelists;
> -	return do_try_to_free_pages(zonelist, sc.gfp_mask, &sc);
> +	return do_try_to_free_pages(zonelist, &sc);
>  }
>  #endif
>  
> 

-- 
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab

Re: [PATCH rc5-mm1 1/3] mm-have-zonelist: fix memcg ooms

[Balbir Singh]

Hugh Dickins wrote:
> Memcgroups up against their limit were OOMing unjustifiably on x86_32
> with highmem: pages ripe for freeing were not being found.
> 
> It's because the zonelist changes in vmscan.c now derive zone from gfp_mask,
> and the gfp mask being passed into try_to_free_mem_cgroup_pages is atuned
> to allocating a struct page_cgroup (__GFP_HIGHMEM masked off), whereas here
> we need to be scanning highmem zones.
> 

Very good catch. I see a high_zoneidx field being derived from gfp_mask -mm.
I've been looking at mainline code too often.


-- 
	Warm Regards,
	Balbir Singh
	Linux Technology Center
	IBM, ISTL

Re: [PATCH rc5-mm1 1/3] mm-have-zonelist: fix memcg ooms

[Mel Gorman]

On (11/03/08 21:12), Hugh Dickins didst pronounce:
> Memcgroups up against their limit were OOMing unjustifiably on x86_32
> with highmem: pages ripe for freeing were not being found.
> 
> It's because the zonelist changes in vmscan.c now derive zone from gfp_mask,
> and the gfp mask being passed into try_to_free_mem_cgroup_pages is atuned
> to allocating a struct page_cgroup (__GFP_HIGHMEM masked off), whereas here
> we need to be scanning highmem zones.
> 
> Frankenstein a gfp_mask from GFP_HIGHUSER_MOVABLE and the mask passed in.
> 

Well spotted, but I have one query below.

> Signed-off-by: Hugh Dickins <hugh@veritas.com>
> ---
> Same error was in -rc3-mm1, sorry for not sending fix in time for -rc5-mm1.
> The error comes 1 or 4 patches earlier, but this patch atop 2.6.25-rc5-mm1
> fits best just after
> mm-have-zonelist-contains-structs-with-both-a-zone-pointer-and-zone_idx.patch
> 
> I've other GFP masking fixes or cleanups to come another day: this fixes
> just one clear bug, peculiar to the -mm tree, orthogonal to the rest.
> But two little do_try_to_free_pages cleanups follow on from this...
> 
>  mm/vmscan.c |    6 +++---
>  1 file changed, 3 insertions(+), 3 deletions(-)
> 
> --- 2.6.25-rc5-mm1/mm/vmscan.c	2008-03-11 14:19:36.000000000 +0000
> +++ linux/mm/vmscan.c	2008-03-11 20:09:05.000000000 +0000
> @@ -1444,7 +1444,6 @@ unsigned long try_to_free_mem_cgroup_pag
>  						gfp_t gfp_mask)
>  {
>  	struct scan_control sc = {
> -		.gfp_mask = gfp_mask,
>  		.may_writepage = !laptop_mode,
>  		.may_swap = 1,
>  		.swap_cluster_max = SWAP_CLUSTER_MAX,
> @@ -1454,9 +1453,10 @@ unsigned long try_to_free_mem_cgroup_pag
>  		.isolate_pages = mem_cgroup_isolate_pages,
>  	};
>  	struct zonelist *zonelist;
> -	int target_zone = gfp_zonelist(GFP_HIGHUSER_MOVABLE);
>  
> -	zonelist = &NODE_DATA(numa_node_id())->node_zonelists[target_zone];
> +	sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
> +			(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
> +	zonelist = NODE_DATA(numa_node_id())->node_zonelists;

While it is clear you are setting the mask to include HIGHMEM-related flags,
it's not as clear to me why you alter the zonelist as well. target_zone was
already based on HIGHMEM so what are you fixing there?

It should still work as ->node_zonelists[0] is a zonelist suitable for node
fallback as opposed to node_zonelists[1] which is for GFP_THISNODE but
maybe this was not quite what you intended?

Probably something obvious that will hit me the second I push send :)

>  	if (do_try_to_free_pages(zonelist, sc.gfp_mask, &sc))
>  		return 1;
>  	return 0;
> 

-- 
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab

Re: [PATCH rc5-mm1 1/3] mm-have-zonelist: fix memcg ooms

[Hugh Dickins]

On Wed, 12 Mar 2008, Mel Gorman wrote:
> On (11/03/08 21:12), Hugh Dickins didst pronounce:
> > @@ -1454,9 +1453,10 @@ unsigned long try_to_free_mem_cgroup_pag
> >  		.isolate_pages = mem_cgroup_isolate_pages,
> >  	};
> >  	struct zonelist *zonelist;
> > -	int target_zone = gfp_zonelist(GFP_HIGHUSER_MOVABLE);
> >  
> > -	zonelist = &NODE_DATA(numa_node_id())->node_zonelists[target_zone];
> > +	sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
> > +			(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
> > +	zonelist = NODE_DATA(numa_node_id())->node_zonelists;
> 
> While it is clear you are setting the mask to include HIGHMEM-related flags,
> it's not as clear to me why you alter the zonelist as well. target_zone was
> already based on HIGHMEM so what are you fixing there?
> 
> It should still work as ->node_zonelists[0] is a zonelist suitable for node
> fallback as opposed to node_zonelists[1] which is for GFP_THISNODE but
> maybe this was not quite what you intended?
> 
> Probably something obvious that will hit me the second I push send :)

That bit wasn't a fix as such, it just came from not wanting to repeat
GFP_HIGHUSER_MOVABLE in there: after I'd looked at gfp_zonelist, it
appeared to be redundant in this context, so I preferred to cut out
the target_zone, and let sc.gfp_mask handle it all.  That worries you?

Hugh

Re: [PATCH rc5-mm1 1/3] mm-have-zonelist: fix memcg ooms

[Mel Gorman]

On (12/03/08 13:34), Hugh Dickins didst pronounce:
> On Wed, 12 Mar 2008, Mel Gorman wrote:
> > On (11/03/08 21:12), Hugh Dickins didst pronounce:
> > > @@ -1454,9 +1453,10 @@ unsigned long try_to_free_mem_cgroup_pag
> > >  		.isolate_pages = mem_cgroup_isolate_pages,
> > >  	};
> > >  	struct zonelist *zonelist;
> > > -	int target_zone = gfp_zonelist(GFP_HIGHUSER_MOVABLE);
> > >  
> > > -	zonelist = &NODE_DATA(numa_node_id())->node_zonelists[target_zone];
> > > +	sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
> > > +			(GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
> > > +	zonelist = NODE_DATA(numa_node_id())->node_zonelists;
> > 
> > While it is clear you are setting the mask to include HIGHMEM-related flags,
> > it's not as clear to me why you alter the zonelist as well. target_zone was
> > already based on HIGHMEM so what are you fixing there?
> > 
> > It should still work as ->node_zonelists[0] is a zonelist suitable for node
> > fallback as opposed to node_zonelists[1] which is for GFP_THISNODE but
> > maybe this was not quite what you intended?
> > 
> > Probably something obvious that will hit me the second I push send :)
> 
> That bit wasn't a fix as such, it just came from not wanting to repeat
> GFP_HIGHUSER_MOVABLE in there: after I'd looked at gfp_zonelist, it
> appeared to be redundant in this context, so I preferred to cut out
> the target_zone, and let sc.gfp_mask handle it all.  That worries you?
> 

Only a little. If the layout of node_zonelists[] changes so that [0] does not
contain the node fallback list, it may cause a problem but I can't imagine why
such a situation would occur either. My initial concern was because I couldn't
see what difference the change made so assumed I must be missing something.

-- 
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab