x86: cpa, preserve large pages if possible
authorThomas Gleixner <tglx@linutronix.de>
Mon, 4 Feb 2008 15:48:07 +0000 (16:48 +0100)
committerIngo Molnar <mingo@elte.hu>
Mon, 4 Feb 2008 15:48:07 +0000 (16:48 +0100)
When CPA is called on a range which fits into a large page mapping,
avoid to split the page when:

1) There is no change of attributes
2) The range to change is a complete large mapping

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

index 79a9f1b42ddd60c5d0f24290b691eaba5021514d..40b7ac58e671ff8f39f70a2e051671f1564d4de5 100644 (file)
 
 struct cpa_data {
        unsigned long   vaddr;
-       int             numpages;
        pgprot_t        mask_set;
        pgprot_t        mask_clr;
+       int             numpages;
        int             flushtlb;
 };
 
+enum {
+       CPA_NO_SPLIT = 0,
+       CPA_SPLIT,
+};
+
 static inline int
 within(unsigned long addr, unsigned long start, unsigned long end)
 {
@@ -230,6 +235,86 @@ static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
 #endif
 }
 
+static int try_preserve_large_page(pte_t *kpte, unsigned long address,
+                                  struct cpa_data *cpa)
+{
+       unsigned long nextpage_addr, numpages, pmask, psize, flags;
+       pte_t new_pte, old_pte, *tmp;
+       pgprot_t old_prot, new_prot;
+       int level, res = CPA_SPLIT;
+
+       spin_lock_irqsave(&pgd_lock, flags);
+       /*
+        * Check for races, another CPU might have split this page
+        * up already:
+        */
+       tmp = lookup_address(address, &level);
+       if (tmp != kpte)
+               goto out_unlock;
+
+       switch (level) {
+       case PG_LEVEL_2M:
+               psize = LARGE_PAGE_SIZE;
+               pmask = LARGE_PAGE_MASK;
+               break;
+       case PG_LEVEL_1G:
+       default:
+               res = -EINVAL;
+               goto out_unlock;
+       }
+
+       /*
+        * Calculate the number of pages, which fit into this large
+        * page starting at address:
+        */
+       nextpage_addr = (address + psize) & pmask;
+       numpages = (nextpage_addr - address) >> PAGE_SHIFT;
+       if (numpages < cpa->numpages)
+               cpa->numpages = numpages;
+
+       /*
+        * We are safe now. Check whether the new pgprot is the same:
+        */
+       old_pte = *kpte;
+       old_prot = new_prot = pte_pgprot(old_pte);
+
+       pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
+       pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
+       new_prot = static_protections(new_prot, address);
+
+       /*
+        * If there are no changes, return. maxpages has been updated
+        * above:
+        */
+       if (pgprot_val(new_prot) == pgprot_val(old_prot)) {
+               res = CPA_NO_SPLIT;
+               goto out_unlock;
+       }
+
+       /*
+        * We need to change the attributes. Check, whether we can
+        * change the large page in one go. We request a split, when
+        * the address is not aligned and the number of pages is
+        * smaller than the number of pages in the large page. Note
+        * that we limited the number of possible pages already to
+        * the number of pages in the large page.
+        */
+       if (address == (nextpage_addr - psize) && cpa->numpages == numpages) {
+               /*
+                * The address is aligned and the number of pages
+                * covers the full page.
+                */
+               new_pte = pfn_pte(pte_pfn(old_pte), canon_pgprot(new_prot));
+               __set_pmd_pte(kpte, address, new_pte);
+               cpa->flushtlb = 1;
+               res = CPA_NO_SPLIT;
+       }
+
+out_unlock:
+       spin_unlock_irqrestore(&pgd_lock, flags);
+       return res;
+}
+
 static int split_large_page(pte_t *kpte, unsigned long address)
 {
        pgprot_t ref_prot = pte_pgprot(pte_clrhuge(*kpte));
@@ -295,7 +380,7 @@ out_unlock:
 static int __change_page_attr(unsigned long address, struct cpa_data *cpa)
 {
        struct page *kpte_page;
-       int level, err = 0;
+       int level, res;
        pte_t *kpte;
 
 repeat:
@@ -338,13 +423,34 @@ repeat:
                        set_pte_atomic(kpte, new_pte);
                        cpa->flushtlb = 1;
                }
-       } else {
-               err = split_large_page(kpte, address);
-               if (!err)
-                       goto repeat;
-               cpa->flushtlb = 1;
+               cpa->numpages = 1;
+               return 0;
        }
-       return err;
+
+       /*
+        * Check, whether we can keep the large page intact
+        * and just change the pte:
+        */
+       res = try_preserve_large_page(kpte, address, cpa);
+       if (res < 0)
+               return res;
+
+       /*
+        * When the range fits into the existing large page,
+        * return. cp->numpages and cpa->tlbflush have been updated in
+        * try_large_page:
+        */
+       if (res == CPA_NO_SPLIT)
+               return 0;
+
+       /*
+        * We have to split the large page:
+        */
+       res = split_large_page(kpte, address);
+       if (res)
+               return res;
+       cpa->flushtlb = 1;
+       goto repeat;
 }
 
 /**
@@ -410,15 +516,27 @@ static int change_page_attr_addr(struct cpa_data *cpa)
 
 static int __change_page_attr_set_clr(struct cpa_data *cpa)
 {
-       unsigned int i;
-       int ret;
+       int ret, numpages = cpa->numpages;
 
-       for (i = 0; i < cpa->numpages ; i++, cpa->vaddr += PAGE_SIZE) {
+       while (numpages) {
+               /*
+                * Store the remaining nr of pages for the large page
+                * preservation check.
+                */
+               cpa->numpages = numpages;
                ret = change_page_attr_addr(cpa);
                if (ret)
                        return ret;
-       }
 
+               /*
+                * Adjust the number of pages with the result of the
+                * CPA operation. Either a large page has been
+                * preserved or a single page update happened.
+                */
+               BUG_ON(cpa->numpages > numpages);
+               numpages -= cpa->numpages;
+               cpa->vaddr += cpa->numpages * PAGE_SIZE;
+       }
        return 0;
 }