Skip to contentSkip to conversation

vfs: Don't copy mount bind mounts of /proc/<pid>/ns/mnt between namespaces

DISCUSSION Dan Carpenter 2 participants 5 messages Sep 2, 2013 VFS

lore
Share & tools
lore
#0 lore raw

Hello Eric W. Biederman,

This is a semi-automatic email about new static checker warnings.

The patch 4ce5d2b1a8fd: "vfs: Don't copy mount bind mounts of /proc/<pid>/ns/mnt between namespaces" from Mar 30, 2013, leads to the following Smatch complaint:

fs/namespace.c:2459 dup_mnt_ns()
warn: variable dereferenced before check 'p' (see line 2475)

fs/namespace.c

  2441          new = copy_tree(old, old->mnt.mnt_root, copy_flags);
                                     ^^^^^^^^^^^^^^^^^

Old is dereferenced.

  2442          if (IS_ERR(new)) {
  2443                  namespace_unlock();
  2444                  free_mnt_ns(new_ns);
  2445                  return ERR_CAST(new);
  2446          }
  2447          new_ns->root = new;
  2448          br_write_lock(&vfsmount_lock);
  2449          list_add_tail(&new_ns->list, &new->mnt_list);
  2450          br_write_unlock(&vfsmount_lock);
  2451  
  2452          /*
  2453           * Second pass: switch the tsk->fs->* elements and mark new vfsmounts
  2454           * as belonging to new namespace.  We have already acquired a private
  2455           * fs_struct, so tsk->fs->lock is not needed.
  2456           */
  2457          p = old;
                ^^^^^^^

P is a valid pointer.

2458 q = new;

  2459		while (p) {
                      ^^^

Existing check for "p".

2460 q->mnt_ns = new_ns;
2461 if (fs) {
2462 if (&p->mnt == fs->root.mnt) {
2463 fs->root.mnt = mntget(&q->mnt);
2464 rootmnt = &p->mnt;
2465 }
2466 if (&p->mnt == fs->pwd.mnt) {
2467 fs->pwd.mnt = mntget(&q->mnt);
2468 pwdmnt = &p->mnt;
2469 }
2470 }
2471 p = next_mnt(p, old);
2472 q = next_mnt(q, new);
2473 if (!q)
2474 break;
2475 while (p->mnt.mnt_root != q->mnt.mnt_root)

                               ^^^^^^^^^^^^^^^

We dereference "p".

  2476				p = next_mnt(p, old);
  2477		}

regards,
dan carpenter

#1 lore raw
Quoted 7 lines · Dan Carpenter

Hello Eric W. Biederman,

This is a semi-automatic email about new static checker warnings.

The patch 4ce5d2b1a8fd: "vfs: Don't copy mount bind mounts of /proc/<pid>/ns/mnt between namespaces" from Mar 30, 2013, leads to the following Smatch complaint:

Semi-autoautomatic reply. The test !q is enough to ensure p is valid until p->mnt.mnt_root == q->mnt.mnt_root.

This has been verified through both testing and reading of the code.

Eric

Quoted 57 lines

fs/namespace.c:2459 dup_mnt_ns()
warn: variable dereferenced before check 'p' (see line 2475)

fs/namespace.c

  2441          new = copy_tree(old, old->mnt.mnt_root, copy_flags);
                                     ^^^^^^^^^^^^^^^^^

Old is dereferenced.

  2442          if (IS_ERR(new)) {
  2443                  namespace_unlock();
  2444                  free_mnt_ns(new_ns);
  2445                  return ERR_CAST(new);
  2446          }
  2447          new_ns->root = new;
  2448          br_write_lock(&vfsmount_lock);
  2449          list_add_tail(&new_ns->list, &new->mnt_list);
  2450          br_write_unlock(&vfsmount_lock);
  2451  
  2452          /*
  2453           * Second pass: switch the tsk->fs->* elements and mark new vfsmounts
  2454           * as belonging to new namespace.  We have already acquired a private
  2455           * fs_struct, so tsk->fs->lock is not needed.
  2456           */
  2457          p = old;
                ^^^^^^^

P is a valid pointer.

2458 q = new;

  2459		while (p) {
                      ^^^

Existing check for "p".

2460 q->mnt_ns = new_ns;
2461 if (fs) {
2462 if (&p->mnt == fs->root.mnt) {
2463 fs->root.mnt = mntget(&q->mnt);
2464 rootmnt = &p->mnt;
2465 }
2466 if (&p->mnt == fs->pwd.mnt) {
2467 fs->pwd.mnt = mntget(&q->mnt);
2468 pwdmnt = &p->mnt;
2469 }
2470 }
2471 p = next_mnt(p, old);
2472 q = next_mnt(q, new);
2473 if (!q)
2474 break;
2475 while (p->mnt.mnt_root != q->mnt.mnt_root)

                               ^^^^^^^^^^^^^^^

We dereference "p".

  2476				p = next_mnt(p, old);
  2477		}

regards,
dan carpenter

#2 lore raw
Quoted 6 lines · Eric W. Biederman
Quoted 4 lines · Dan Carpenter

Hello Eric W. Biederman,

This is a semi-automatic email about new static checker warnings.

Quoted 9 lines
Quoted 3 lines

The patch 4ce5d2b1a8fd: "vfs: Don't copy mount bind mounts of /proc/<pid>/ns/mnt between namespaces" from Mar 30, 2013, leads to the following Smatch complaint:

Semi-autoautomatic reply. The test !q is enough to ensure p is valid until p->mnt.mnt_root == q->mnt.mnt_root.

This has been verified through both testing and reading of the code.

Why not make the test:

	while (1) { ...

regards,
dan carpenter

#3 lore raw
Quoted 14 lines · Dan Carpenter
Quoted 13 lines · Eric W. Biederman
Quoted 7 lines · Dan Carpenter

Hello Eric W. Biederman,

This is a semi-automatic email about new static checker warnings.

The patch 4ce5d2b1a8fd: "vfs: Don't copy mount bind mounts of /proc/<pid>/ns/mnt between namespaces" from Mar 30, 2013, leads to the following Smatch complaint:

Semi-autoautomatic reply. The test !q is enough to ensure p is valid until p->mnt.mnt_root == q->mnt.mnt_root.

Quoted 6 lines
Quoted 2 lines

This has been verified through both testing and reading of the code.

Why not make the test:

	while (1) { ...

I really don't understand what you are getting at, and it feels like this conversation is descending into inanity fast.

The way this work is.
p = old
q = new

Where new is a copy of old.

p == NULL means that we have completed the traversal.

The entire point of the loop is to find corresponding entities in new and old. Because in a few correr cases I slice off leaf mounts that are mounted on top of files I need an extra loop, and I made that with a very small code change.

So yes it does look like the case that except for the first iteration the test at the top of the while loop is now redundant, because of the if (!q) break; that I added.

I think any greater change to the loop would have obscured the change that I made a great deal more.

I think all of this talking about the syntactical elements of the code instead of brining this up to the level of human beings and talking about what is going on is extremely dangerous way to look at code. Especially given that C does not allow everything of importance about the code to be expressed.

Eric

#4 lore raw
Quoted 1 line · Eric W. Biederman

So yes it does look like the case that except for the first iteration

                                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Quoted 2 lines

the test at the top of the while loop is now redundant, because of the if (!q) break; that I added.

Ah... I see the confusion now.

The test is redundant on the first iteration as well. The test is entirely redundant. That's the point.

regards,
dan carpenter