hugetlb: Try to grow hugetlb pool for MAP_PRIVATE mappings
authorAdam Litke <agl@us.ibm.com>
Tue, 16 Oct 2007 08:26:18 +0000 (01:26 -0700)
committerLinus Torvalds <torvalds@woody.linux-foundation.org>
Tue, 16 Oct 2007 16:43:02 +0000 (09:43 -0700)
Because we overcommit hugepages for MAP_PRIVATE mappings, it is possible that
the hugetlb pool will be exhausted or completely reserved when a hugepage is
needed to satisfy a page fault.  Before killing the process in this situation,
try to allocate a hugepage directly from the buddy allocator.

The explicitly configured pool size becomes a low watermark.  When dynamically
grown, the allocated huge pages are accounted as a surplus over the watermark.
 As huge pages are freed on a node, surplus pages are released to the buddy
allocator so that the pool will shrink back to the watermark.

Surplus accounting also allows for friendlier explicit pool resizing.  When
shrinking a pool that is fully in-use, increase the surplus so pages will be
returned to the buddy allocator as soon as they are freed.  When growing a
pool that has a surplus, consume the surplus first and then allocate new
pages.

Signed-off-by: Adam Litke <agl@us.ibm.com>
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Andy Whitcroft <apw@shadowen.org>
Acked-by: Dave McCracken <dave.mccracken@oracle.com>
Cc: William Irwin <bill.irwin@oracle.com>
Cc: David Gibson <david@gibson.dropbear.id.au>
Cc: Ken Chen <kenchen@google.com>
Cc: Badari Pulavarty <pbadari@us.ibm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/hugetlb.c

index ba029d6..8768e52 100644 (file)
 
 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
 static unsigned long nr_huge_pages, free_huge_pages, resv_huge_pages;
+static unsigned long surplus_huge_pages;
 unsigned long max_huge_pages;
 static struct list_head hugepage_freelists[MAX_NUMNODES];
 static unsigned int nr_huge_pages_node[MAX_NUMNODES];
 static unsigned int free_huge_pages_node[MAX_NUMNODES];
+static unsigned int surplus_huge_pages_node[MAX_NUMNODES];
 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
 unsigned long hugepages_treat_as_movable;
 
@@ -109,15 +111,57 @@ static void update_and_free_page(struct page *page)
 
 static void free_huge_page(struct page *page)
 {
-       BUG_ON(page_count(page));
+       int nid = page_to_nid(page);
 
+       BUG_ON(page_count(page));
        INIT_LIST_HEAD(&page->lru);
 
        spin_lock(&hugetlb_lock);
-       enqueue_huge_page(page);
+       if (surplus_huge_pages_node[nid]) {
+               update_and_free_page(page);
+               surplus_huge_pages--;
+               surplus_huge_pages_node[nid]--;
+       } else {
+               enqueue_huge_page(page);
+       }
        spin_unlock(&hugetlb_lock);
 }
 
+/*
+ * Increment or decrement surplus_huge_pages.  Keep node-specific counters
+ * balanced by operating on them in a round-robin fashion.
+ * Returns 1 if an adjustment was made.
+ */
+static int adjust_pool_surplus(int delta)
+{
+       static int prev_nid;
+       int nid = prev_nid;
+       int ret = 0;
+
+       VM_BUG_ON(delta != -1 && delta != 1);
+       do {
+               nid = next_node(nid, node_online_map);
+               if (nid == MAX_NUMNODES)
+                       nid = first_node(node_online_map);
+
+               /* To shrink on this node, there must be a surplus page */
+               if (delta < 0 && !surplus_huge_pages_node[nid])
+                       continue;
+               /* Surplus cannot exceed the total number of pages */
+               if (delta > 0 && surplus_huge_pages_node[nid] >=
+                                               nr_huge_pages_node[nid])
+                       continue;
+
+               surplus_huge_pages += delta;
+               surplus_huge_pages_node[nid] += delta;
+               ret = 1;
+               break;
+       } while (nid != prev_nid);
+
+       prev_nid = nid;
+       return ret;
+}
+
 static int alloc_fresh_huge_page(void)
 {
        static int prev_nid;
@@ -150,10 +194,30 @@ static int alloc_fresh_huge_page(void)
        return 0;
 }
 
+static struct page *alloc_buddy_huge_page(struct vm_area_struct *vma,
+                                               unsigned long address)
+{
+       struct page *page;
+
+       page = alloc_pages(htlb_alloc_mask|__GFP_COMP|__GFP_NOWARN,
+                                       HUGETLB_PAGE_ORDER);
+       if (page) {
+               set_compound_page_dtor(page, free_huge_page);
+               spin_lock(&hugetlb_lock);
+               nr_huge_pages++;
+               nr_huge_pages_node[page_to_nid(page)]++;
+               surplus_huge_pages++;
+               surplus_huge_pages_node[page_to_nid(page)]++;
+               spin_unlock(&hugetlb_lock);
+       }
+
+       return page;
+}
+
 static struct page *alloc_huge_page(struct vm_area_struct *vma,
                                    unsigned long addr)
 {
-       struct page *page;
+       struct page *page = NULL;
 
        spin_lock(&hugetlb_lock);
        if (vma->vm_flags & VM_MAYSHARE)
@@ -173,7 +237,16 @@ fail:
        if (vma->vm_flags & VM_MAYSHARE)
                resv_huge_pages++;
        spin_unlock(&hugetlb_lock);
-       return NULL;
+
+       /*
+        * Private mappings do not use reserved huge pages so the allocation
+        * may have failed due to an undersized hugetlb pool.  Try to grab a
+        * surplus huge page from the buddy allocator.
+        */
+       if (!(vma->vm_flags & VM_MAYSHARE))
+               page = alloc_buddy_huge_page(vma, addr);
+
+       return page;
 }
 
 static int __init hugetlb_init(void)
@@ -241,26 +314,62 @@ static inline void try_to_free_low(unsigned long count)
 }
 #endif
 
+#define persistent_huge_pages (nr_huge_pages - surplus_huge_pages)
 static unsigned long set_max_huge_pages(unsigned long count)
 {
-       while (count > nr_huge_pages) {
-               if (!alloc_fresh_huge_page())
-                       return nr_huge_pages;
-       }
-       if (count >= nr_huge_pages)
-               return nr_huge_pages;
+       unsigned long min_count, ret;
 
+       /*
+        * Increase the pool size
+        * First take pages out of surplus state.  Then make up the
+        * remaining difference by allocating fresh huge pages.
+        */
        spin_lock(&hugetlb_lock);
-       count = max(count, resv_huge_pages);
-       try_to_free_low(count);
-       while (count < nr_huge_pages) {
+       while (surplus_huge_pages && count > persistent_huge_pages) {
+               if (!adjust_pool_surplus(-1))
+                       break;
+       }
+
+       while (count > persistent_huge_pages) {
+               int ret;
+               /*
+                * If this allocation races such that we no longer need the
+                * page, free_huge_page will handle it by freeing the page
+                * and reducing the surplus.
+                */
+               spin_unlock(&hugetlb_lock);
+               ret = alloc_fresh_huge_page();
+               spin_lock(&hugetlb_lock);
+               if (!ret)
+                       goto out;
+
+       }
+       if (count >= persistent_huge_pages)
+               goto out;
+
+       /*
+        * Decrease the pool size
+        * First return free pages to the buddy allocator (being careful
+        * to keep enough around to satisfy reservations).  Then place
+        * pages into surplus state as needed so the pool will shrink
+        * to the desired size as pages become free.
+        */
+       min_count = max(count, resv_huge_pages);
+       try_to_free_low(min_count);
+       while (min_count < persistent_huge_pages) {
                struct page *page = dequeue_huge_page(NULL, 0);
                if (!page)
                        break;
                update_and_free_page(page);
        }
+       while (count < persistent_huge_pages) {
+               if (!adjust_pool_surplus(1))
+                       break;
+       }
+out:
+       ret = persistent_huge_pages;
        spin_unlock(&hugetlb_lock);
-       return nr_huge_pages;
+       return ret;
 }
 
 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
@@ -292,10 +401,12 @@ int hugetlb_report_meminfo(char *buf)
                        "HugePages_Total: %5lu\n"
                        "HugePages_Free:  %5lu\n"
                        "HugePages_Rsvd:  %5lu\n"
+                       "HugePages_Surp:  %5lu\n"
                        "Hugepagesize:    %5lu kB\n",
                        nr_huge_pages,
                        free_huge_pages,
                        resv_huge_pages,
+                       surplus_huge_pages,
                        HPAGE_SIZE/1024);
 }