[PATCH 2.5.48-mm1] Break COW page tables on mmap

[Dave McCracken <dmccr@us.ibm.com>]

[-- Attachment #1: Type: text/plain, Size: 528 bytes --]


I found a fairly large hole in my unsharing logic.  Pte page COW behavior
breaks down when new objects are mapped.  This patch makes sure there
aren't any COW pte pages in the range of a new mapping at mmap time.

This should fix the KDE problem.  It fixed it on the test machine I've been
using.

Dave McCracken

======================================================================
Dave McCracken          IBM Linux Base Kernel Team      1-512-838-3059
dmccr@us.ibm.com                                        T/L   678-3059

[-- Attachment #2: shpte-2.5.48-mm1-2.diff --]
[-- Type: text/plain, Size: 2321 bytes --]

--- 2.5.48-mm1-shsent/./mm/memory.c	2002-11-21 11:26:32.000000000 -0600
+++ 2.5.48-mm1-shpte/./mm/memory.c	2002-11-22 10:27:39.000000000 -0600
@@ -1080,6 +1080,65 @@
 	tlb_finish_mmu(tlb, 0, TASK_SIZE);
 }
 
+#ifdef CONFIG_SHAREPTE
+void
+clear_share_range(struct mm_struct *mm, unsigned long address, unsigned long len)
+{
+	pgd_t		*pgd;
+	pmd_t		*pmd;
+	struct page	*ptepage;
+	unsigned long	end = address + len;
+
+	spin_lock(&mm->page_table_lock);
+
+	pgd = pgd_offset(mm, address);
+	if (pgd_none(*pgd) || pgd_bad(*pgd))
+		goto skip_start;
+
+	pmd = pmd_offset(pgd, address);
+	if (pmd_none(*pmd) || pmd_bad(*pmd))
+		goto skip_start;
+
+	ptepage = pmd_page(*pmd);
+	pte_page_lock(ptepage);
+	if (page_count(ptepage) > 1) {
+		pte_t *pte;
+
+		pte = pte_unshare(mm, pmd, address);
+		pte_unmap(pte);
+		ptepage = pmd_page(*pmd);
+	}
+	pte_page_unlock(ptepage);
+
+skip_start:
+	/* This range is contained in one pte page.  We're done. */
+	if ((address >> PMD_SHIFT) == (end >> PMD_SHIFT))
+		goto out;
+
+	pgd = pgd_offset(mm, end);
+	if (pgd_none(*pgd) || pgd_bad(*pgd))
+		goto out;
+
+	pmd = pmd_offset(pgd, end);
+	if (pmd_none(*pmd) || pmd_bad(*pmd))
+		goto out;
+
+	ptepage = pmd_page(*pmd);
+	pte_page_lock(ptepage);
+	if (page_count(ptepage) > 1) {
+		pte_t *pte;
+
+		pte = pte_unshare(mm, pmd, end);
+		pte_unmap(pte);
+		ptepage = pmd_page(*pmd);
+	}
+	pte_page_unlock(ptepage);
+
+out:
+	spin_unlock(&mm->page_table_lock);
+}
+#endif
+
 /*
  * Do a quick page-table lookup for a single page.
  * mm->page_table_lock must be held.
--- 2.5.48-mm1-shsent/./mm/mmap.c	2002-11-19 09:17:36.000000000 -0600
+++ 2.5.48-mm1-shpte/./mm/mmap.c	2002-11-22 10:26:11.000000000 -0600
@@ -57,6 +57,8 @@
 pgprot_t protection_pmd[8] = {
 	__PMD000, __PMD001, __PMD010, __PMD011, __PMD100, __PMD101, __PMD110, __PMD111
 };
+extern void clear_share_range(struct mm_struct *mm, unsigned long address,
+			      unsigned long len);
 #endif
 
 int sysctl_overcommit_memory = 0;	/* default is heuristic overcommit */
@@ -524,6 +526,9 @@
 			return -ENOMEM;
 		goto munmap_back;
 	}
+#ifdef CONFIG_SHAREPTE
+	clear_share_range(mm, addr, len);
+#endif
 
 	/* Check against address space limit. */
 	if ((mm->total_vm << PAGE_SHIFT) + len