Re: [PATCH 4/4] hugetlb: add support for gigantic page allocation at runtime

Yasuaki Ishimatsu <isimatu.yasuaki@xxxxxxxxxxxxxx> · Tue, 8 Apr 2014 10:58:47 +0900

(2014/04/04 22:30), Luiz Capitulino wrote:
On Fri, 4 Apr 2014 12:05:17 +0900
Yasuaki Ishimatsu <isimatu.yasuaki@xxxxxxxxxxxxxx> wrote:

(2014/04/03 3:08), Luiz Capitulino wrote:
HugeTLB is limited to allocating hugepages whose size are less than
MAX_ORDER order. This is so because HugeTLB allocates hugepages via
the buddy allocator. Gigantic pages (that is, pages whose size is
greater than MAX_ORDER order) have to be allocated at boottime.

However, boottime allocation has at least two serious problems. First,
it doesn't support NUMA and second, gigantic pages allocated at
boottime can't be freed.

This commit solves both issues by adding support for allocating gigantic
pages during runtime. It works just like regular sized hugepages,
meaning that the interface in sysfs is the same, it supports NUMA,
and gigantic pages can be freed.

For example, on x86_64 gigantic pages are 1GB big. To allocate two 1G
gigantic pages on node 1, one can do:

   # echo 2 > \
     /sys/devices/system/node/node1/hugepages/hugepages-1048576kB/nr_hugepages

And to free them later:

   # echo 0 > \
     /sys/devices/system/node/node1/hugepages/hugepages-1048576kB/nr_hugepages

The one problem with gigantic page allocation at runtime is that it
can't be serviced by the buddy allocator. To overcome that problem, this
series scans all zones from a node looking for a large enough contiguous
region. When one is found, it's allocated by using CMA, that is, we call
alloc_contig_range() to do the actual allocation. For example, on x86_64
we scan all zones looking for a 1GB contiguous region. When one is found
it's allocated by alloc_contig_range().

One expected issue with that approach is that such gigantic contiguous
regions tend to vanish as time goes by. The best way to avoid this for
now is to make gigantic page allocations very early during boot, say
from a init script. Other possible optimization include using compaction,
which is supported by CMA but is not explicitly used by this commit.

It's also important to note the following:

   1. My target systems are x86_64 machines, so I have only tested 1GB
      pages allocation/release. I did try to make this arch indepedent
      and expect it to work on other archs but didn't try it myself

   2. I didn't add support for hugepage overcommit, that is allocating
      a gigantic page on demand when
     /proc/sys/vm/nr_overcommit_hugepages > 0. The reason is that I don't
     think it's reasonable to do the hard and long work required for
     allocating a gigantic page at fault time. But it should be simple
     to add this if wanted

Signed-off-by: Luiz Capitulino <lcapitulino@xxxxxxxxxx>
---
   arch/x86/include/asm/hugetlb.h |  10 +++
   mm/hugetlb.c                   | 177 ++++++++++++++++++++++++++++++++++++++---
   2 files changed, 176 insertions(+), 11 deletions(-)

diff --git a/arch/x86/include/asm/hugetlb.h b/arch/x86/include/asm/hugetlb.h
index a809121..2b262f7 100644
--- a/arch/x86/include/asm/hugetlb.h
+++ b/arch/x86/include/asm/hugetlb.h
@@ -91,6 +91,16 @@ static inline void arch_release_hugepage(struct page *page)

<snip>


+		start_pfn = z->zone_start_pfn; /* to silence gcc */
+		for (i = z->zone_start_pfn; i < zone_end_pfn(z); i++) {

This loop is not smart. On our system, one node has serveral TBytes.
So the maximum loop count is "TBytes/Page size".

Interesting. Would you be willing to test this series on such a
machine?

First page of gigantic page must be aligned.
So how about it:

		start_pfn = zone_start_pfn aligned gigantic page
		for (i = start_pfn; i < zone_end_pfn; i += size of gigantic page) {
			if (!pfn_valid_gigantic(i)) {
				count = 0;
				continue;
			}
			
			...
		}

I'm not sure that very loop will work because pfn_valid_gigantic() checks
a single PFN today, but we do have to scan every single PFN on a gigantic
page range.


On the other hand, I think got what you're suggesting. When an unsuitable
PFN is found, we should just skip to the next aligned PFN instead of
keep scanning for nothing (which is what my loop does today). Maybe you're
suggesting pfn_valid_gigantic() should do that?

That's right.

Thanks,
Yasuaki Ishimatsu


Anyway, I'll make that change, thank you very much for you review!


Thanks,
Yasuaki Ishimatsu

+			if (!pfn_valid_gigantic(i)) {
+				count = 0;
+				continue;
+			}
+			if (!count) {
+				if (!pfn_aligned_gigantic(i, order))
+					continue;
+				start_pfn = i;
+			}
+			if (++count == nr_pages) {
+				/*
+				 * We release the zone lock here because
+				 * alloc_contig_range() will also lock the zone
+				 * at some point. If there's an allocation
+				 * spinning on this lock, it may win the race
+				 * and cause alloc_contig_range() to fail...
+				 */
+				spin_unlock_irqrestore(&z->lock, flags);
+				ret = __alloc_gigantic_page(start_pfn, count);
+				if (!ret)
+					return pfn_to_page(start_pfn);
+				count = 0;
+				spin_lock_irqsave(&z->lock, flags);
+			}
+		}
+
+		spin_unlock_irqrestore(&z->lock, flags);
+	}
+
+	return NULL;
+}
+
+static void prep_new_huge_page(struct hstate *h, struct page *page, int nid);
+static void prep_compound_gigantic_page(struct page *page, unsigned long order);
+
+static struct page *alloc_fresh_gigantic_page_node(struct hstate *h, int nid)
+{
+	struct page *page;
+
+	page = alloc_gigantic_page(nid, huge_page_order(h));
+	if (page) {
+		if (arch_prepare_gigantic_page(page)) {
+			free_gigantic_page(page, huge_page_order(h));
+			return NULL;
+		}
+		prep_compound_gigantic_page(page, huge_page_order(h));
+		prep_new_huge_page(h, page, nid);
+	}
+
+	return page;
+}
+
+static int alloc_fresh_gigantic_page(struct hstate *h,
+				nodemask_t *nodes_allowed)
+{
+	struct page *page = NULL;
+	int nr_nodes, node;
+
+	for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
+		page = alloc_fresh_gigantic_page_node(h, node);
+		if (page)
+			return 1;
+	}
+
+	return 0;
+}
+
+static inline bool gigantic_page_supported(void) { return true; }
+#else /* !CONFIG_CMA */
+static inline bool gigantic_page_supported(void) { return false; }
+static inline void free_gigantic_page(struct page *page, unsigned order) { }
+static inline void destroy_compound_gigantic_page(struct page *page,
+						unsigned long order) { }
+static inline int alloc_fresh_gigantic_page(struct hstate *h,
+					nodemask_t *nodes_allowed) { return 0; }
+#endif /* CONFIG_CMA */
+
   static void update_and_free_page(struct hstate *h, struct page *page)
   {
   	int i;

-	VM_BUG_ON(hstate_is_gigantic(h));
+	if (hstate_is_gigantic(h) && !gigantic_page_supported())
+		return;

   	h->nr_huge_pages--;
   	h->nr_huge_pages_node[page_to_nid(page)]--;
@@ -661,8 +809,14 @@ static void update_and_free_page(struct hstate *h, struct page *page)
   	VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page), page);
   	set_compound_page_dtor(page, NULL);
   	set_page_refcounted(page);
-	arch_release_hugepage(page);
-	__free_pages(page, huge_page_order(h));
+	if (hstate_is_gigantic(h)) {
+		arch_release_gigantic_page(page);
+		destroy_compound_gigantic_page(page, huge_page_order(h));
+		free_gigantic_page(page, huge_page_order(h));
+	} else {
+		arch_release_hugepage(page);
+		__free_pages(page, huge_page_order(h));
+	}
   }

   struct hstate *size_to_hstate(unsigned long size)
@@ -701,7 +855,7 @@ static void free_huge_page(struct page *page)
   	if (restore_reserve)
   		h->resv_huge_pages++;

-	if (h->surplus_huge_pages_node[nid] && !hstate_is_gigantic(h)) {
+	if (h->surplus_huge_pages_node[nid]) {
   		/* remove the page from active list */
   		list_del(&page->lru);
   		update_and_free_page(h, page);
@@ -805,9 +959,6 @@ static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
   {
   	struct page *page;

-	if (hstate_is_gigantic(h))
-		return NULL;
-
   	page = alloc_pages_exact_node(nid,
   		htlb_alloc_mask(h)|__GFP_COMP|__GFP_THISNODE|
   						__GFP_REPEAT|__GFP_NOWARN,
@@ -1452,7 +1603,7 @@ static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
   {
   	unsigned long min_count, ret;

-	if (hstate_is_gigantic(h))
+	if (hstate_is_gigantic(h) && !gigantic_page_supported())
   		return h->max_huge_pages;

   	/*
@@ -1479,7 +1630,11 @@ static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
   		 * and reducing the surplus.
   		 */
   		spin_unlock(&hugetlb_lock);
-		ret = alloc_fresh_huge_page(h, nodes_allowed);
+		if (hstate_is_gigantic(h)) {
+			ret = alloc_fresh_gigantic_page(h, nodes_allowed);
+		} else {
+			ret = alloc_fresh_huge_page(h, nodes_allowed);
+		}
   		spin_lock(&hugetlb_lock);
   		if (!ret)
   			goto out;
@@ -1578,7 +1733,7 @@ static ssize_t nr_hugepages_store_common(bool obey_mempolicy,
   		goto out;

   	h = kobj_to_hstate(kobj, &nid);
-	if (hstate_is_gigantic(h)) {
+	if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
   		err = -EINVAL;
   		goto out;
   	}
@@ -2072,7 +2227,7 @@ static int hugetlb_sysctl_handler_common(bool obey_mempolicy,

   	tmp = h->max_huge_pages;

-	if (write && hstate_is_gigantic(h))
+	if (write && hstate_is_gigantic(h) && !gigantic_page_supported())
   		return -EINVAL;

   	table->data = &tmp;






--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>







