2015-07-30 14:05:19

by David Vrabel

[permalink] [raw]
Subject: [PATCHv1] mm: always initialize pages as reserved to fix memory hotplug

Commit 92923ca3aacef63c92dc297a75ad0c6dfe4eab37 (mm: meminit: only set
page reserved in the memblock region) breaks memory hotplug because pages
within newly added sections are not marked as reserved as required by
the memory hotplug driver. If pages within an offline section are not
reserved, the secton cannot be onlined.

Re-add the SetPageReserved() call.

Signed-off-by: David Vrabel <[email protected]>
Cc: Robin Holt <[email protected]>
Cc: Nathan Zimmer <[email protected]>
Cc: Mel Gorman <[email protected]>
---
mm/page_alloc.c | 1 +
1 file changed, 1 insertion(+)

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index ef19f22..89492f6 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -842,6 +842,7 @@ static void __meminit __init_single_page(struct page *page, unsigned long pfn,
init_page_count(page);
page_mapcount_reset(page);
page_cpupid_reset_last(page);
+ SetPageReserved(page);

INIT_LIST_HEAD(&page->lru);
#ifdef WANT_PAGE_VIRTUAL
--
2.1.4


2015-07-30 14:46:01

by Mel Gorman

[permalink] [raw]
Subject: Re: [PATCHv1] mm: always initialize pages as reserved to fix memory hotplug

On Thu, Jul 30, 2015 at 03:04:43PM +0100, David Vrabel wrote:
> Commit 92923ca3aacef63c92dc297a75ad0c6dfe4eab37 (mm: meminit: only set
> page reserved in the memblock region) breaks memory hotplug because pages
> within newly added sections are not marked as reserved as required by
> the memory hotplug driver.

I don't have access to a large machine at the moment to verify and won't
have until Monday at the earliest but I think that will bust deferred
initialisation.

Why not either SetPageReserved from mem hotplug driver? It might be neater
to remove the PageReserved check from online_pages_range() but then care
would have to be taken to ensure that invalid PFNs within section that
have no memory backing them were properly reserved. This is an untested,
uncompiled version of the first suggestion

diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
index 26fbba7d888f..003dbe4b060d 100644
--- a/mm/memory_hotplug.c
+++ b/mm/memory_hotplug.c
@@ -446,7 +446,7 @@ static int __meminit __add_zone(struct zone *zone, unsigned long phys_start_pfn)
int nr_pages = PAGES_PER_SECTION;
int nid = pgdat->node_id;
int zone_type;
- unsigned long flags;
+ unsigned long flags, pfn;
int ret;

zone_type = zone - pgdat->node_zones;
@@ -461,6 +461,14 @@ static int __meminit __add_zone(struct zone *zone, unsigned long phys_start_pfn)
pgdat_resize_unlock(zone->zone_pgdat, &flags);
memmap_init_zone(nr_pages, nid, zone_type,
phys_start_pfn, MEMMAP_HOTPLUG);
+
+ /* online_page_range is called later and expects pages reserved */
+ for (pfn = phys_start_pfn; pfn < phys_start_pfn + nr_pages; pfn++) {
+ if (!pfn_valid(pfn))
+ continue;
+
+ SetPageReserved(pfn_to_page(pfn));
+ }
return 0;
}

--
Mel Gorman
SUSE Labs

2015-07-30 16:41:42

by David Vrabel

[permalink] [raw]
Subject: Re: [PATCHv1] mm: always initialize pages as reserved to fix memory hotplug

On 30/07/15 15:45, Mel Gorman wrote:
> On Thu, Jul 30, 2015 at 03:04:43PM +0100, David Vrabel wrote:
>> Commit 92923ca3aacef63c92dc297a75ad0c6dfe4eab37 (mm: meminit: only set
>> page reserved in the memblock region) breaks memory hotplug because pages
>> within newly added sections are not marked as reserved as required by
>> the memory hotplug driver.
>
> I don't have access to a large machine at the moment to verify and won't
> have until Monday at the earliest but I think that will bust deferred
> initialisation.
>
> Why not either SetPageReserved from mem hotplug driver? It might be neater
> to remove the PageReserved check from online_pages_range() but then care
> would have to be taken to ensure that invalid PFNs within section that
> have no memory backing them were properly reserved. This is an untested,
> uncompiled version of the first suggestion

Thanks.

Tested-by: David Vrabel <[email protected]>

8<------------------------------
>From 0c1acd3118d8bb911566c0af8f759fccbd90fb9e Mon Sep 17 00:00:00 2001
From: Mel Gorman <[email protected]>
Date: Thu, 30 Jul 2015 15:45:54 +0100
Subject: [PATCH] mm: initialize hotplugged pages as reserved

Commit 92923ca3aacef63c92dc297a75ad0c6dfe4eab37 (mm: meminit: only set
page reserved in the memblock region) breaks memory hotplug because pages
within newly added sections are not marked as reserved as required by
the memory hotplug driver.

Set hotplugged pages as reserved when adding new zones.

Tested-by: David Vrabel <[email protected]>
---
mm/memory_hotplug.c | 10 +++++++++-
1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
index 460d0fe..169770a 100644
--- a/mm/memory_hotplug.c
+++ b/mm/memory_hotplug.c
@@ -446,7 +446,7 @@ static int __meminit __add_zone(struct zone *zone, unsigned long phys_start_pfn)
int nr_pages = PAGES_PER_SECTION;
int nid = pgdat->node_id;
int zone_type;
- unsigned long flags;
+ unsigned long flags, pfn;
int ret;

zone_type = zone - pgdat->node_zones;
@@ -461,6 +461,14 @@ static int __meminit __add_zone(struct zone *zone, unsigned long phys_start_pfn)
pgdat_resize_unlock(zone->zone_pgdat, &flags);
memmap_init_zone(nr_pages, nid, zone_type,
phys_start_pfn, MEMMAP_HOTPLUG);
+
+ /* online_page_range is called later and expects pages reserved */
+ for (pfn = phys_start_pfn; pfn < phys_start_pfn + nr_pages; pfn++) {
+ if (!pfn_valid(pfn))
+ continue;
+
+ SetPageReserved(pfn_to_page(pfn));
+ }
return 0;
}

--
2.1.4