fix mapping_writably_mapped()
authorHugh Dickins <hugh@veritas.com>
Wed, 10 Dec 2008 20:48:52 +0000 (20:48 +0000)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 10 Dec 2008 22:40:45 +0000 (14:40 -0800)
Lee Schermerhorn noticed yesterday that I broke the mapping_writably_mapped
test in 2.6.7!  Bad bad bug, good good find.

The i_mmap_writable count must be incremented for VM_SHARED (just as
i_writecount is for VM_DENYWRITE, but while holding the i_mmap_lock)
when dup_mmap() copies the vma for fork: it has its own more optimal
version of __vma_link_file(), and I missed this out.  So the count
was later going down to 0 (dangerous) when one end unmapped, then
wrapping negative (inefficient) when the other end unmapped.

The only impact on x86 would have been that setting a mandatory lock on
a file which has at some time been opened O_RDWR and mapped MAP_SHARED
(but not necessarily PROT_WRITE) across a fork, might fail with -EAGAIN
when it should succeed, or succeed when it should fail.

But those architectures which rely on flush_dcache_page() to flush
userspace modifications back into the page before the kernel reads it,
may in some cases have skipped the flush after such a fork - though any
repetitive test will soon wrap the count negative, in which case it will
flush_dcache_page() unnecessarily.

Fix would be a two-liner, but mapping variable added, and comment moved.

Reported-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
kernel/fork.c

index 8d6a7dd9282b8308f041d20b18dab139b85c18b3..495da2e9a8b4d92a1b464b98196228f818c892e7 100644 (file)
@@ -315,17 +315,20 @@ static int dup_mmap(struct mm_struct *mm, struct mm_struct *oldmm)
                file = tmp->vm_file;
                if (file) {
                        struct inode *inode = file->f_path.dentry->d_inode;
+                       struct address_space *mapping = file->f_mapping;
+
                        get_file(file);
                        if (tmp->vm_flags & VM_DENYWRITE)
                                atomic_dec(&inode->i_writecount);
-
-                       /* insert tmp into the share list, just after mpnt */
-                       spin_lock(&file->f_mapping->i_mmap_lock);
+                       spin_lock(&mapping->i_mmap_lock);
+                       if (tmp->vm_flags & VM_SHARED)
+                               mapping->i_mmap_writable++;
                        tmp->vm_truncate_count = mpnt->vm_truncate_count;
-                       flush_dcache_mmap_lock(file->f_mapping);
+                       flush_dcache_mmap_lock(mapping);
+                       /* insert tmp into the share list, just after mpnt */
                        vma_prio_tree_add(tmp, mpnt);
-                       flush_dcache_mmap_unlock(file->f_mapping);
-                       spin_unlock(&file->f_mapping->i_mmap_lock);
+                       flush_dcache_mmap_unlock(mapping);
+                       spin_unlock(&mapping->i_mmap_lock);
                }
 
                /*