[PATCH v2 4/4] hugetlbfs: don't retry when pool page allocations start to fail

[Mike Kravetz]

When allocating hugetlbfs pool pages via /proc/sys/vm/nr_hugepages,
the pages will be interleaved between all nodes of the system.  If
nodes are not equal, it is quite possible for one node to fill up
before the others.  When this happens, the code still attempts to
allocate pages from the full node.  This results in calls to direct
reclaim and compaction which slow things down considerably.

When allocating pool pages, note the state of the previous allocation
for each node.  If previous allocation failed, do not use the
aggressive retry algorithm on successive attempts.  The allocation
will still succeed if there is memory available, but it will not try
as hard to free up memory.

Signed-off-by: Mike Kravetz <mike.kravetz@xxxxxxxxxx>
---
v2 - Removed __GFP_NORETRY from bit mask allocations and added more
     comments.  OK to pass NULL to NODEMASK_FREE.

 mm/hugetlb.c | 89 ++++++++++++++++++++++++++++++++++++++++++++++------
 1 file changed, 79 insertions(+), 10 deletions(-)

diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index ede7e7f5d1ab..2be0b055958a 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c
@@ -1405,12 +1405,25 @@ pgoff_t __basepage_index(struct page *page)
 }
 
 static struct page *alloc_buddy_huge_page(struct hstate *h,
-		gfp_t gfp_mask, int nid, nodemask_t *nmask)
+		gfp_t gfp_mask, int nid, nodemask_t *nmask,
+		nodemask_t *node_alloc_noretry)
 {
 	int order = huge_page_order(h);
 	struct page *page;
+	bool alloc_try_hard = true;
 
-	gfp_mask |= __GFP_COMP|__GFP_RETRY_MAYFAIL|__GFP_NOWARN;
+	/*
+	 * By default we always try hard to allocate the page with
+	 * __GFP_RETRY_MAYFAIL flag.  However, if we are allocating pages in
+	 * a loop (to adjust global huge page counts) and previous allocation
+	 * failed, do not continue to try hard on the same node.  Use the
+	 * node_alloc_noretry bitmap to manage this state information.
+	 */
+	if (node_alloc_noretry && node_isset(nid, *node_alloc_noretry))
+		alloc_try_hard = false;
+	gfp_mask |= __GFP_COMP|__GFP_NOWARN;
+	if (alloc_try_hard)
+		gfp_mask |= __GFP_RETRY_MAYFAIL;
 	if (nid == NUMA_NO_NODE)
 		nid = numa_mem_id();
 	page = __alloc_pages_nodemask(gfp_mask, order, nid, nmask);
@@ -1419,6 +1432,22 @@ static struct page *alloc_buddy_huge_page(struct hstate *h,
 	else
 		__count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
 
+	/*
+	 * If we did not specify __GFP_RETRY_MAYFAIL, but still got a page this
+	 * indicates an overall state change.  Clear bit so that we resume
+	 * normal 'try hard' allocations.
+	 */
+	if (node_alloc_noretry && page && !alloc_try_hard)
+		node_clear(nid, *node_alloc_noretry);
+
+	/*
+	 * If we tried hard to get a page but failed, set bit so that
+	 * subsequent attempts will not try as hard until there is an
+	 * overall state change.
+	 */
+	if (node_alloc_noretry && !page && alloc_try_hard)
+		node_set(nid, *node_alloc_noretry);
+
 	return page;
 }
 
@@ -1427,7 +1456,8 @@ static struct page *alloc_buddy_huge_page(struct hstate *h,
  * should use this function to get new hugetlb pages
  */
 static struct page *alloc_fresh_huge_page(struct hstate *h,
-		gfp_t gfp_mask, int nid, nodemask_t *nmask)
+		gfp_t gfp_mask, int nid, nodemask_t *nmask,
+		nodemask_t *node_alloc_noretry)
 {
 	struct page *page;
 
@@ -1435,7 +1465,7 @@ static struct page *alloc_fresh_huge_page(struct hstate *h,
 		page = alloc_gigantic_page(h, gfp_mask, nid, nmask);
 	else
 		page = alloc_buddy_huge_page(h, gfp_mask,
-				nid, nmask);
+				nid, nmask, node_alloc_noretry);
 	if (!page)
 		return NULL;
 
@@ -1450,14 +1480,16 @@ static struct page *alloc_fresh_huge_page(struct hstate *h,
  * Allocates a fresh page to the hugetlb allocator pool in the node interleaved
  * manner.
  */
-static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed)
+static int alloc_pool_huge_page(struct hstate *h, nodemask_t *nodes_allowed,
+				nodemask_t *node_alloc_noretry)
 {
 	struct page *page;
 	int nr_nodes, node;
 	gfp_t gfp_mask = htlb_alloc_mask(h) | __GFP_THISNODE;
 
 	for_each_node_mask_to_alloc(h, nr_nodes, node, nodes_allowed) {
-		page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed);
+		page = alloc_fresh_huge_page(h, gfp_mask, node, nodes_allowed,
+						node_alloc_noretry);
 		if (page)
 			break;
 	}
@@ -1601,7 +1633,7 @@ static struct page *alloc_surplus_huge_page(struct hstate *h, gfp_t gfp_mask,
 		goto out_unlock;
 	spin_unlock(&hugetlb_lock);
 
-	page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
+	page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
 	if (!page)
 		return NULL;
 
@@ -1637,7 +1669,7 @@ struct page *alloc_migrate_huge_page(struct hstate *h, gfp_t gfp_mask,
 	if (hstate_is_gigantic(h))
 		return NULL;
 
-	page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask);
+	page = alloc_fresh_huge_page(h, gfp_mask, nid, nmask, NULL);
 	if (!page)
 		return NULL;
 
@@ -2207,13 +2239,33 @@ static void __init gather_bootmem_prealloc(void)
 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
 {
 	unsigned long i;
+	nodemask_t *node_alloc_noretry;
+
+	if (!hstate_is_gigantic(h)) {
+		/*
+		 * Bit mask controlling how hard we retry per-node allocations.
+		 * Ignore errors as lower level routines can deal with
+		 * node_alloc_noretry == NULL.  If this kmalloc fails at boot
+		 * time, we are likely in bigger trouble.
+		 */
+		node_alloc_noretry = kmalloc(sizeof(*node_alloc_noretry),
+						GFP_KERNEL);
+	} else {
+		/* allocations done at boot time */
+		node_alloc_noretry = NULL;
+	}
+
+	/* bit mask controlling how hard we retry per-node allocations */
+	if (node_alloc_noretry)
+		nodes_clear(*node_alloc_noretry);
 
 	for (i = 0; i < h->max_huge_pages; ++i) {
 		if (hstate_is_gigantic(h)) {
 			if (!alloc_bootmem_huge_page(h))
 				break;
 		} else if (!alloc_pool_huge_page(h,
-					 &node_states[N_MEMORY]))
+					 &node_states[N_MEMORY],
+					 node_alloc_noretry))
 			break;
 		cond_resched();
 	}
@@ -2225,6 +2277,8 @@ static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
 			h->max_huge_pages, buf, i);
 		h->max_huge_pages = i;
 	}
+
+	kfree(node_alloc_noretry);
 }
 
 static void __init hugetlb_init_hstates(void)
@@ -2323,6 +2377,17 @@ static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
 			      nodemask_t *nodes_allowed)
 {
 	unsigned long min_count, ret;
+	NODEMASK_ALLOC(nodemask_t, node_alloc_noretry, GFP_KERNEL);
+
+	/*
+	 * Bit mask controlling how hard we retry per-node allocations.
+	 * If we can not allocate the bit mask, do not attempt to allocate
+	 * the requested huge pages.
+	 */
+	if (node_alloc_noretry)
+		nodes_clear(*node_alloc_noretry);
+	else
+		return -ENOMEM;
 
 	spin_lock(&hugetlb_lock);
 
@@ -2356,6 +2421,7 @@ static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
 	if (hstate_is_gigantic(h) && !IS_ENABLED(CONFIG_CONTIG_ALLOC)) {
 		if (count > persistent_huge_pages(h)) {
 			spin_unlock(&hugetlb_lock);
+			NODEMASK_FREE(node_alloc_noretry);
 			return -EINVAL;
 		}
 		/* Fall through to decrease pool */
@@ -2388,7 +2454,8 @@ static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
 		/* yield cpu to avoid soft lockup */
 		cond_resched();
 
-		ret = alloc_pool_huge_page(h, nodes_allowed);
+		ret = alloc_pool_huge_page(h, nodes_allowed,
+						node_alloc_noretry);
 		spin_lock(&hugetlb_lock);
 		if (!ret)
 			goto out;
@@ -2429,6 +2496,8 @@ static int set_max_huge_pages(struct hstate *h, unsigned long count, int nid,
 	h->max_huge_pages = persistent_huge_pages(h);
 	spin_unlock(&hugetlb_lock);
 
+	NODEMASK_FREE(node_alloc_noretry);
+
 	return 0;
 }
 
-- 
2.20.1