x86: further cpa largepage-split cleanups
authorIngo Molnar <mingo@elte.hu>
Wed, 30 Jan 2008 12:33:56 +0000 (13:33 +0100)
committerIngo Molnar <mingo@elte.hu>
Wed, 30 Jan 2008 12:33:56 +0000 (13:33 +0100)
further cpa largepage-split cleanups: make the splitup isolated
functionality, without leaking details back into __change_page_attr().

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
arch/x86/mm/pageattr_32.c

index 1011b21f8db093f4165c59588614f018992591d7..14c923b3b07f272b03a8726216d036ab6768157a 100644 (file)
@@ -37,33 +37,6 @@ pte_t *lookup_address(unsigned long address, int *level)
        return pte_offset_kernel(pmd, address);
 }
 
-static struct page *
-split_large_page(unsigned long address, pgprot_t ref_prot)
-{
-       unsigned long addr;
-       struct page *base;
-       pte_t *pbase;
-       int i;
-
-       base = alloc_pages(GFP_KERNEL, 0);
-       if (!base)
-               return NULL;
-
-       /*
-        * page_private is used to track the number of entries in
-        * the page table page that have non standard attributes.
-        */
-       address = __pa(address);
-       addr = address & LARGE_PAGE_MASK;
-       pbase = (pte_t *)page_address(base);
-       paravirt_alloc_pt(&init_mm, page_to_pfn(base));
-
-       for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE)
-               set_pte(&pbase[i], pfn_pte(addr >> PAGE_SHIFT, ref_prot));
-
-       return base;
-}
-
 static void set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 {
        unsigned long flags;
@@ -88,14 +61,58 @@ static void set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
        spin_unlock_irqrestore(&pgd_lock, flags);
 }
 
+static int
+split_large_page(pte_t *kpte, unsigned long address, pgprot_t ref_prot)
+{
+       int i, level;
+       unsigned long addr;
+       pte_t *pbase, *tmp;
+       struct page *base;
+
+       base = alloc_pages(GFP_KERNEL, 0);
+       if (!base)
+               return -ENOMEM;
+
+       down_write(&init_mm.mmap_sem);
+       /*
+        * Check for races, another CPU might have split this page
+        * up for us already:
+        */
+       tmp = lookup_address(address, &level);
+       if (tmp != kpte)
+               goto out_unlock;
+
+       address = __pa(address);
+       addr = address & LARGE_PAGE_MASK;
+       pbase = (pte_t *)page_address(base);
+       paravirt_alloc_pt(&init_mm, page_to_pfn(base));
+
+       for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE)
+               set_pte(&pbase[i], pfn_pte(addr >> PAGE_SHIFT, ref_prot));
+
+       /*
+        * Install the new, split up pagetable:
+        */
+       set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
+       base = NULL;
+
+out_unlock:
+       up_write(&init_mm.mmap_sem);
+
+       if (base)
+               __free_pages(base, 0);
+
+       return 0;
+}
+
 static int __change_page_attr(struct page *page, pgprot_t prot)
 {
        pgprot_t ref_prot = PAGE_KERNEL;
        struct page *kpte_page;
        unsigned long address;
+       int level, err = 0;
        pgprot_t oldprot;
        pte_t *kpte;
-       int level;
 
        BUG_ON(PageHighMem(page));
        address = (unsigned long)page_address(page);
@@ -127,19 +144,11 @@ repeat:
        if (level == 3) {
                set_pte_atomic(kpte, mk_pte(page, prot));
        } else {
-               struct page *split;
-
-               split = split_large_page(address, ref_prot);
-               if (!split)
-                       return -ENOMEM;
-
-               /*
-                * There's a small window here to waste a bit of RAM:
-                */
-               set_pmd_pte(kpte, address, mk_pte(split, ref_prot));
-               goto repeat;
+               err = split_large_page(kpte, address, ref_prot);
+               if (!err)
+                       goto repeat;
        }
-       return 0;
+       return err;
 }
 
 /*