[PATCH] hugepage allocator cleanup
Nick Piggin [Wed, 22 Mar 2006 08:08:08 +0000 (00:08 -0800)]
Insert "fresh" huge pages into the hugepage allocator by the same means as
they are freed back into it.  This reduces code size and allows
enqueue_huge_page to be inlined into the hugepage free fastpath.

Eliminate occurances of hugepages on the free list with non-zero refcount.
This can allow stricter refcount checks in future.  Also required for
lockless pagecache.

Signed-off-by: Nick Piggin <npiggin@suse.de>

"This patch also eliminates a leak "cleaned up" by re-clobbering the
refcount on every allocation from the hugepage freelists.  With respect to
the lockless pagecache, the crucial aspect is to eliminate unconditional
set_page_count() to 0 on pages with potentially nonzero refcounts, though
closer inspection suggests the assignments removed are entirely spurious."

Acked-by: William Irwin <wli@holomorphy.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

mm/hugetlb.c

index 5087077..39d49ec 100644 (file)
@@ -64,7 +64,7 @@ static struct page *dequeue_huge_page(struct vm_area_struct *vma,
        return page;
 }
 
-static struct page *alloc_fresh_huge_page(void)
+static int alloc_fresh_huge_page(void)
 {
        static int nid = 0;
        struct page *page;
@@ -72,12 +72,15 @@ static struct page *alloc_fresh_huge_page(void)
                                        HUGETLB_PAGE_ORDER);
        nid = (nid + 1) % num_online_nodes();
        if (page) {
+               page[1].lru.next = (void *)free_huge_page;      /* dtor */
                spin_lock(&hugetlb_lock);
                nr_huge_pages++;
                nr_huge_pages_node[page_to_nid(page)]++;
                spin_unlock(&hugetlb_lock);
+               put_page(page); /* free it into the hugepage allocator */
+               return 1;
        }
-       return page;
+       return 0;
 }
 
 void free_huge_page(struct page *page)
@@ -85,7 +88,6 @@ void free_huge_page(struct page *page)
        BUG_ON(page_count(page));
 
        INIT_LIST_HEAD(&page->lru);
-       page[1].lru.next = NULL;                        /* reset dtor */
 
        spin_lock(&hugetlb_lock);
        enqueue_huge_page(page);
@@ -105,7 +107,6 @@ struct page *alloc_huge_page(struct vm_area_struct *vma, unsigned long addr)
        }
        spin_unlock(&hugetlb_lock);
        set_page_count(page, 1);
-       page[1].lru.next = (void *)free_huge_page;      /* set dtor */
        for (i = 0; i < (HPAGE_SIZE/PAGE_SIZE); ++i)
                clear_user_highpage(&page[i], addr);
        return page;
@@ -114,7 +115,6 @@ struct page *alloc_huge_page(struct vm_area_struct *vma, unsigned long addr)
 static int __init hugetlb_init(void)
 {
        unsigned long i;
-       struct page *page;
 
        if (HPAGE_SHIFT == 0)
                return 0;
@@ -123,12 +123,8 @@ static int __init hugetlb_init(void)
                INIT_LIST_HEAD(&hugepage_freelists[i]);
 
        for (i = 0; i < max_huge_pages; ++i) {
-               page = alloc_fresh_huge_page();
-               if (!page)
+               if (!alloc_fresh_huge_page())
                        break;
-               spin_lock(&hugetlb_lock);
-               enqueue_huge_page(page);
-               spin_unlock(&hugetlb_lock);
        }
        max_huge_pages = free_huge_pages = nr_huge_pages = i;
        printk("Total HugeTLB memory allocated, %ld\n", free_huge_pages);
@@ -154,8 +150,8 @@ static void update_and_free_page(struct page *page)
                page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
                                1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
                                1 << PG_private | 1<< PG_writeback);
-               set_page_count(&page[i], 0);
        }
+       page[1].lru.next = NULL;
        set_page_count(page, 1);
        __free_pages(page, HUGETLB_PAGE_ORDER);
 }
@@ -188,12 +184,8 @@ static inline void try_to_free_low(unsigned long count)
 static unsigned long set_max_huge_pages(unsigned long count)
 {
        while (count > nr_huge_pages) {
-               struct page *page = alloc_fresh_huge_page();
-               if (!page)
+               if (!alloc_fresh_huge_page())
                        return nr_huge_pages;
-               spin_lock(&hugetlb_lock);
-               enqueue_huge_page(page);
-               spin_unlock(&hugetlb_lock);
        }
        if (count >= nr_huge_pages)
                return nr_huge_pages;