Skip to contentSkip to conversation

Unlock 1GB-hugetlb on x86_64

PATCH RESEND New Oscar Salvador 1 participant 3 messages Mar 20, 2019

Subsystems (2)
lore
Share & tools
lore

Oscar Salvador

author maintainer
#0 lore raw

cover Unlock 1GB-hugetlb on x86_64

From: Oscar Salvador <[email protected]>

RFC -> V1:
- Split up the patch
- Added Michal's Acked-by
- Rebased and added David's Reviewed-by

The RFC version of this patch was discussed here [1], and it did not find any objection.
I decided to split up the former patch because one of the changes enables offlining operation for 1GB-hugetlb pages, while the other change is a mere cleanup.

Patch1 contains all the information regarding 1GB-hugetlb pages change.

[1] https://lore.kernel.org/linux-mm/[email protected]/

Oscar Salvador (2):
mm,memory_hotplug: Unlock 1GB-hugetlb on x86_64 mm,memory_hotplug: Drop redundant hugepage_migration_supported check

 mm/memory_hotplug.c | 7 +------
 1 file changed, 1 insertion(+), 6 deletions(-)
Signature · 1 line
2.13.7
author maintainer
#1 lore raw

PATCH RESEND 1/2 mm,memory_hotplug: Unlock 1GB-hugetlb on x86_64

On x86_64, 1GB-hugetlb pages could never be offlined due to the fact that hugepage_migration_supported() returned false for PUD_SHIFT. So whenever we wanted to offline a memblock containing a gigantic hugetlb page, we never got beyond has_unmovable_pages() check. This changed with [1], where now we also return true for PUD_SHIFT.

After that patch, the check in has_unmovable_pages() and scan_movable_pages() returned true, but we still had a final barrier in do_migrate_range():

if (compound_order(head) > PFN_SECTION_SHIFT) {
	ret = -EBUSY;
	break;
}

This is not really nice, and we do not really need it.
It is perfectly possible to migrate a gigantic page as long as another node has a spare gigantic page for us.
In alloc_huge_page_nodemask(), we calculate the __real__ number of free pages, and if any, we try to dequeue one from another node.

This all works fine when we do have another node with a spare gigantic page, but if that is not the case, alloc_huge_page_nodemask() ends up calling alloc_migrate_huge_page() which bails out if the wanted page is gigantic. That is mainly because finding a 1GB (or even 16GB on powerpc) contiguous memory is quite unlikely when the system has been running for a while.

In that situation, we will keep looping forever because scan_movable_pages() will give us the same page and we will fail again because there is no node where we can dequeue a gigantic page from.
This is not nice, and it has been raised that we might want to treat -ENOMEM as a fatal error in do_migrate_range(), but this has to be checked further.

Anyway, I would tend say that this is the administrator's job, to make sure that the system can keep up with the memory to be offlined, so that would mean that if we want to use gigantic pages, make sure that the other nodes have at least enough gigantic pages to keep up in case we need to offline memory.

Just for the sake of completeness, this is one of the tests done:

 # echo 1 > /sys/devices/system/node/node1/hugepages/hugepages-1048576kB/nr_hugepages
 # echo 1 > /sys/devices/system/node/node2/hugepages/hugepages-1048576kB/nr_hugepages
 # cat /sys/devices/system/node/node1/hugepages/hugepages-1048576kB/nr_hugepages
   1
 # cat /sys/devices/system/node/node1/hugepages/hugepages-1048576kB/free_hugepages
   1
 # cat /sys/devices/system/node/node2/hugepages/hugepages-1048576kB/nr_hugepages
   1
 # cat /sys/devices/system/node/node2/hugepages/hugepages-1048576kB/free_hugepages
   1

(hugetlb1gb is a program that maps 1GB region using MAP_HUGE_1GB)

 # numactl -m 1 ./hugetlb1gb
 # cat /sys/devices/system/node/node1/hugepages/hugepages-1048576kB/free_hugepages
   0
 # cat /sys/devices/system/node/node2/hugepages/hugepages-1048576kB/free_hugepages
   1
 # offline node1 memory
 # cat /sys/devices/system/node/node2/hugepages/hugepages-1048576kB/free_hugepages
   0

[1] https://lore.kernel.org/patchwork/patch/998796/

---
 mm/memory_hotplug.c | 4 ----
 1 file changed, 4 deletions(-)
mm/memory_hotplug.c +0 -4
index f767582af4f8..02283732544e 100644
@@ -1381,10 +1381,6 @@ do_migrate_range(unsigned long start_pfn, unsigned long end_pfn)
 
 		if (PageHuge(page)) {
 			struct page *head = compound_head(page);
-			if (compound_order(head) > PFN_SECTION_SHIFT) {
-				ret = -EBUSY;
-				break;
-			}
 			pfn = page_to_pfn(head) + (1<<compound_order(head)) - 1;
 			isolate_huge_page(head, &source);
 			continue;
Signature · 1 line
2.13.7
author maintainer
#2 lore raw

PATCH RESEND 2/2 mm,memory_hotplug: Drop redundant hugepage_migration_supported check

In reply to Oscar Salvador

has_unmovable_pages() does alreay check whether the hugetlb page supports migration, so all non-migrateable hugetlb pages should have been caught there. Let us drop the check from scan_movable_pages() as is redundant.

---
 mm/memory_hotplug.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)
mm/memory_hotplug.c +1 -2
index 02283732544e..37a83a83e8f8 100644
@@ -1340,8 +1340,7 @@ static unsigned long scan_movable_pages(unsigned long start, unsigned long end)
 		if (!PageHuge(page))
 			continue;
 		head = compound_head(page);
-		if (hugepage_migration_supported(page_hstate(head)) &&
-		    page_huge_active(head))
+		if (page_huge_active(head))
 			return pfn;
 		skip = (1 << compound_order(head)) - (page - head);
 		pfn += skip - 1;
Signature · 1 line
2.13.7