Fix get_user_pages() race for write access
authorLinus Torvalds <torvalds@g5.osdl.org>
Mon, 1 Aug 2005 18:14:49 +0000 (11:14 -0700)
committerLinus Torvalds <torvalds@g5.osdl.org>
Mon, 1 Aug 2005 18:14:49 +0000 (11:14 -0700)
There's no real guarantee that handle_mm_fault() will always be able to
break a COW situation - if an update from another thread ends up
modifying the page table some way, handle_mm_fault() may end up
requiring us to re-try the operation.

That's normally fine, but get_user_pages() ended up re-trying it as a
read, and thus a write access could in theory end up losing the dirty
bit or be done on a page that had not been properly COW'ed.

This makes get_user_pages() always retry write accesses as write
accesses by making "follow_page()" require that a writable follow has
the dirty bit set.  That simplifies the code and solves the race: if the
COW break fails for some reason, we'll just loop around and try again.

Signed-off-by: Linus Torvalds <torvalds@osdl.org>
mm/memory.c

index 6fe77ac..4e1c673 100644 (file)
@@ -811,18 +811,15 @@ static struct page *__follow_page(struct mm_struct *mm, unsigned long address,
        pte = *ptep;
        pte_unmap(ptep);
        if (pte_present(pte)) {
-               if (write && !pte_write(pte))
+               if (write && !pte_dirty(pte))
                        goto out;
                if (read && !pte_read(pte))
                        goto out;
                pfn = pte_pfn(pte);
                if (pfn_valid(pfn)) {
                        page = pfn_to_page(pfn);
-                       if (accessed) {
-                               if (write && !pte_dirty(pte) &&!PageDirty(page))
-                                       set_page_dirty(page);
+                       if (accessed)
                                mark_page_accessed(page);
-                       }
                        return page;
                }
        }
@@ -941,10 +938,9 @@ int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
                spin_lock(&mm->page_table_lock);
                do {
                        struct page *page;
-                       int lookup_write = write;
 
                        cond_resched_lock(&mm->page_table_lock);
-                       while (!(page = follow_page(mm, start, lookup_write))) {
+                       while (!(page = follow_page(mm, start, write))) {
                                /*
                                 * Shortcut for anonymous pages. We don't want
                                 * to force the creation of pages tables for
@@ -952,8 +948,7 @@ int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
                                 * nobody touched so far. This is important
                                 * for doing a core dump for these mappings.
                                 */
-                               if (!lookup_write &&
-                                   untouched_anonymous_page(mm,vma,start)) {
+                               if (!write && untouched_anonymous_page(mm,vma,start)) {
                                        page = ZERO_PAGE(start);
                                        break;
                                }
@@ -972,14 +967,6 @@ int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
                                default:
                                        BUG();
                                }
-                               /*
-                                * Now that we have performed a write fault
-                                * and surely no longer have a shared page we
-                                * shouldn't write, we shouldn't ignore an
-                                * unwritable page in the page table if
-                                * we are forcing write access.
-                                */
-                               lookup_write = write && !force;
                                spin_lock(&mm->page_table_lock);
                        }
                        if (pages) {