CIFS: Fix NULL-pointer dereference in smb2_push_mandatory_locks

Currently when the client creates a cifsFileInfo structure for
a newly opened file, it allocates a list of byte-range locks
with a pointer to the new cfile and attaches this list to the
inode's lock list. The latter happens before initializing all
other fields, e.g. cfile->tlink. Thus a partially initialized
cifsFileInfo structure becomes available to other threads that
walk through the inode's lock list. One example of such a thread
may be an oplock break worker thread that tries to push all
cached byte-range locks. This causes NULL-pointer dereference
in smb2_push_mandatory_locks() when accessing cfile->tlink:

[598428.945633] BUG: kernel NULL pointer dereference, address: 0000000000000038
...
[598428.945749] Workqueue: cifsoplockd cifs_oplock_break [cifs]
[598428.945793] RIP: 0010:smb2_push_mandatory_locks+0xd6/0x5a0 [cifs]
...
[598428.945834] Call Trace:
[598428.945870]  ? cifs_revalidate_mapping+0x45/0x90 [cifs]
[598428.945901]  cifs_oplock_break+0x13d/0x450 [cifs]
[598428.945909]  process_one_work+0x1db/0x380
[598428.945914]  worker_thread+0x4d/0x400
[598428.945921]  kthread+0x104/0x140
[598428.945925]  ? process_one_work+0x380/0x380
[598428.945931]  ? kthread_park+0x80/0x80
[598428.945937]  ret_from_fork+0x35/0x40

Fix this by reordering initialization steps of the cifsFileInfo
structure: initialize all the fields first and then add the new
byte-range lock list to the inode's lock list.

Cc: Stable <stable@vger.kernel.org>
Signed-off-by: Pavel Shilovsky <pshilov@microsoft.com>
---
 fs/cifs/file.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

Merged into cifs-2.6.git for-next

On Wed, Nov 27, 2019 at 6:18 PM Pavel Shilovsky <piastryyy@gmail.com> wrote:
>
> Currently when the client creates a cifsFileInfo structure for
> a newly opened file, it allocates a list of byte-range locks
> with a pointer to the new cfile and attaches this list to the
> inode's lock list. The latter happens before initializing all
> other fields, e.g. cfile->tlink. Thus a partially initialized
> cifsFileInfo structure becomes available to other threads that
> walk through the inode's lock list. One example of such a thread
> may be an oplock break worker thread that tries to push all
> cached byte-range locks. This causes NULL-pointer dereference
> in smb2_push_mandatory_locks() when accessing cfile->tlink:
>
> [598428.945633] BUG: kernel NULL pointer dereference, address: 0000000000000038
> ...
> [598428.945749] Workqueue: cifsoplockd cifs_oplock_break [cifs]
> [598428.945793] RIP: 0010:smb2_push_mandatory_locks+0xd6/0x5a0 [cifs]
> ...
> [598428.945834] Call Trace:
> [598428.945870]  ? cifs_revalidate_mapping+0x45/0x90 [cifs]
> [598428.945901]  cifs_oplock_break+0x13d/0x450 [cifs]
> [598428.945909]  process_one_work+0x1db/0x380
> [598428.945914]  worker_thread+0x4d/0x400
> [598428.945921]  kthread+0x104/0x140
> [598428.945925]  ? process_one_work+0x380/0x380
> [598428.945931]  ? kthread_park+0x80/0x80
> [598428.945937]  ret_from_fork+0x35/0x40
>
> Fix this by reordering initialization steps of the cifsFileInfo
> structure: initialize all the fields first and then add the new
> byte-range lock list to the inode's lock list.
>
> Cc: Stable <stable@vger.kernel.org>
> Signed-off-by: Pavel Shilovsky <pshilov@microsoft.com>
> ---
>  fs/cifs/file.c | 7 ++++---
>  1 file changed, 4 insertions(+), 3 deletions(-)
>
> diff --git a/fs/cifs/file.c b/fs/cifs/file.c
> index 520fbe4d42b9..069635ec9d94 100644
> --- a/fs/cifs/file.c
> +++ b/fs/cifs/file.c
> @@ -313,9 +313,6 @@ cifs_new_fileinfo(struct cifs_fid *fid, struct file *file,
>         INIT_LIST_HEAD(&fdlocks->locks);
>         fdlocks->cfile = cfile;
>         cfile->llist = fdlocks;
> -       cifs_down_write(&cinode->lock_sem);
> -       list_add(&fdlocks->llist, &cinode->llist);
> -       up_write(&cinode->lock_sem);
>
>         cfile->count = 1;
>         cfile->pid = current->tgid;
> @@ -339,6 +336,10 @@ cifs_new_fileinfo(struct cifs_fid *fid, struct file *file,
>                 oplock = 0;
>         }
>
> +       cifs_down_write(&cinode->lock_sem);
> +       list_add(&fdlocks->llist, &cinode->llist);
> +       up_write(&cinode->lock_sem);
> +
>         spin_lock(&tcon->open_file_lock);
>         if (fid->pending_open->oplock != CIFS_OPLOCK_NO_CHANGE && oplock)
>                 oplock = fid->pending_open->oplock;
> --
> 2.17.1
>

Pavel Shilovsky <piastryyy@gmail.com> writes:
> Currently when the client creates a cifsFileInfo structure for
> a newly opened file, it allocates a list of byte-range locks
> with a pointer to the new cfile and attaches this list to the
> inode's lock list. The latter happens before initializing all
> other fields, e.g. cfile->tlink. Thus a partially initialized
> cifsFileInfo structure becomes available to other threads that
> walk through the inode's lock list. One example of such a thread
> may be an oplock break worker thread that tries to push all
> cached byte-range locks. This causes NULL-pointer dereference
> in smb2_push_mandatory_locks() when accessing cfile->tlink:

reviewing late but this makes sense.

Reviewed-by: Aurelien Aptel <aaptel@suse.com>