patches/mm-shrink-the-page-frame-to-rt-size.patch

From: Peter Zijlstra <peterz@infradead.org>
Date: Fri, 3 Jul 2009 08:44:54 -0500
Subject: mm: shrink the page frame to !-rt size

He below is a boot-tested hack to shrink the page frame size back to
normal.

Should be a net win since there should be many less PTE-pages than
page-frames.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>

---
 include/linux/mm.h       |   46 +++++++++++++++++++++++++++++++++++++++-------
 include/linux/mm_types.h |    4 ++++
 mm/memory.c              |   32 ++++++++++++++++++++++++++++++++
 3 files changed, 75 insertions(+), 7 deletions(-)

--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -1259,27 +1259,59 @@ static inline pmd_t *pmd_alloc(struct mm
  * overflow into the next struct page (as it might with DEBUG_SPINLOCK).
  * When freeing, reset page->mapping so free_pages_check won't complain.
  */
+#ifndef CONFIG_PREEMPT_RT_FULL
+
 #define __pte_lockptr(page)	&((page)->ptl)
-#define pte_lock_init(_page)	do {					\
-	spin_lock_init(__pte_lockptr(_page));				\
-} while (0)
+
+static inline struct page *pte_lock_init(struct page *page)
+{
+	spin_lock_init(__pte_lockptr(page));
+	return page;
+}
+
 #define pte_lock_deinit(page)	((page)->mapping = NULL)
+
+#else /* !PREEMPT_RT_FULL */
+
+/*
+ * On PREEMPT_RT_FULL the spinlock_t's are too large to embed in the
+ * page frame, hence it only has a pointer and we need to dynamically
+ * allocate the lock when we allocate PTE-pages.
+ *
+ * This is an overall win, since only a small fraction of the pages
+ * will be PTE pages under normal circumstances.
+ */
+
+#define __pte_lockptr(page)	((page)->ptl)
+
+extern struct page *pte_lock_init(struct page *page);
+extern void pte_lock_deinit(struct page *page);
+
+#endif /* PREEMPT_RT_FULL */
+
 #define pte_lockptr(mm, pmd)	({(void)(mm); __pte_lockptr(pmd_page(*(pmd)));})
 #else	/* !USE_SPLIT_PTLOCKS */
 /*
  * We use mm->page_table_lock to guard all pagetable pages of the mm.
  */
-#define pte_lock_init(page)	do {} while (0)
+static inline struct page *pte_lock_init(struct page *page) { return page; }
 #define pte_lock_deinit(page)	do {} while (0)
 #define pte_lockptr(mm, pmd)	({(void)(pmd); &(mm)->page_table_lock;})
 #endif /* USE_SPLIT_PTLOCKS */
 
-static inline void pgtable_page_ctor(struct page *page)
+static inline struct page *__pgtable_page_ctor(struct page *page)
 {
-	pte_lock_init(page);
-	inc_zone_page_state(page, NR_PAGETABLE);
+	page = pte_lock_init(page);
+	if (page)
+		inc_zone_page_state(page, NR_PAGETABLE);
+	return page;
 }
 
+#define pgtable_page_ctor(page)				\
+do {							\
+	page = __pgtable_page_ctor(page);		\
+} while (0)
+
 static inline void pgtable_page_dtor(struct page *page)
 {
 	pte_lock_deinit(page);
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -141,7 +141,11 @@ struct page {
 						 * system if PG_buddy is set.
 						 */
 #if USE_SPLIT_PTLOCKS
+# ifndef CONFIG_PREEMPT_RT_FULL
 		spinlock_t ptl;
+# else
+		spinlock_t *ptl;
+# endif
 #endif
 		struct kmem_cache *slab_cache;	/* SL[AU]B: Pointer to slab */
 		struct page *first_page;	/* Compound tail pages */
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -4312,3 +4312,35 @@ void copy_user_huge_page(struct page *ds
 	}
 }
 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
+
+#if defined(CONFIG_PREEMPT_RT_FULL) && (USE_SPLIT_PTLOCKS > 0)
+/*
+ * Heinous hack, relies on the caller doing something like:
+ *
+ *   pte = alloc_pages(PGALLOC_GFP, 0);
+ *   if (pte)
+ *     pgtable_page_ctor(pte);
+ *   return pte;
+ *
+ * This ensures we release the page and return NULL when the
+ * lock allocation fails.
+ */
+struct page *pte_lock_init(struct page *page)
+{
+	page->ptl = kmalloc(sizeof(spinlock_t), GFP_KERNEL);
+	if (page->ptl) {
+		spin_lock_init(__pte_lockptr(page));
+	} else {
+		__free_page(page);
+		page = NULL;
+	}
+	return page;
+}
+
+void pte_lock_deinit(struct page *page)
+{
+	kfree(page->ptl);
+	page->mapping = NULL;
+}
+
+#endif