Re: [PATCH v2 1/2] ovl: ovl_permission: Fix null pointer dereference at realinode in rcu-walk mode

[Zhihao Cheng <chengzhihao1@huawei.com>]

在 2023/5/15 23:36, Amir Goldstein 写道:
> On Mon, May 15, 2023 at 6:16 PM Christian Brauner <brauner@kernel.org> wrote:
>>
>> On Mon, May 15, 2023 at 05:58:55PM +0300, Amir Goldstein wrote:
>>> On Mon, May 15, 2023 at 4:58 PM Christian Brauner <brauner@kernel.org> wrote:
>>>>
>>>> On Mon, May 15, 2023 at 09:36:28PM +0800, Zhihao Cheng wrote:
>>>>> Following process:
>>>>>            P1                     P2
>>>>>   path_lookupat
>>>>>    link_path_walk
>>>>>     inode_permission
>>>>>      ovl_permission
>>>>>        ovl_i_path_real(inode, &realpath)
>>>>>          path->dentry = ovl_i_dentry_upper(inode)
>>>>>                            drop_cache
>>>>>                           __dentry_kill(ovl_dentry)
>>>>>                            iput(ovl_inode)
>>>>>                             ovl_destroy_inode(ovl_inode)
>>>>>                              dput(oi->__upperdentry)
>>>>>                               dentry_kill(upperdentry)
>>>>>                                dentry_unlink_inode
>>>>>                                 upperdentry->d_inode = NULL
>>>>>        realinode = d_inode(realpath.dentry) // return NULL
>>>>>        inode_permission(realinode)
>>>>>         inode->i_sb  // NULL pointer dereference
>>>>> , will trigger an null pointer dereference at realinode:
>>>>>    [  335.664979] BUG: kernel NULL pointer dereference,
>>>>>                   address: 0000000000000002
>>>>>    [  335.668032] CPU: 0 PID: 2592 Comm: ls Not tainted 6.3.0
>>>>>    [  335.669956] RIP: 0010:inode_permission+0x33/0x2c0
>>>>>    [  335.678939] Call Trace:
>>>>>    [  335.679165]  <TASK>
>>>>>    [  335.679371]  ovl_permission+0xde/0x320
>>>>>    [  335.679723]  inode_permission+0x15e/0x2c0
>>>>>    [  335.680090]  link_path_walk+0x115/0x550
>>>>>    [  335.680771]  path_lookupat.isra.0+0xb2/0x200
>>>>>    [  335.681170]  filename_lookup+0xda/0x240
>>>>>    [  335.681922]  vfs_statx+0xa6/0x1f0
>>>>>    [  335.682233]  vfs_fstatat+0x7b/0xb0
>>>>>
>>>>> Fetch a reproducer in [Link].
>>>>>
>>>>> Add a new helper ovl_i_path_realinode() to get realpath and real inode
>>>>> after non-nullptr checking, use the helper to replace the current realpath
>>>>> getting logic.
>>>>>
>>>>> Link: https://bugzilla.kernel.org/show_bug.cgi?id=217405
>>>>> Fixes: 4b7791b2e958 ("ovl: handle idmappings in ovl_permission()")
>>>>> Signed-off-by: Zhihao Cheng <chengzhihao1@huawei.com>
>>>>> Suggested-by: Christian Brauner <brauner@kernel.org>
>>>>> ---
>>>>>   fs/overlayfs/inode.c | 31 ++++++++++++++++++++++++-------
>>>>>   1 file changed, 24 insertions(+), 7 deletions(-)
>>>>>
>>>>> diff --git a/fs/overlayfs/inode.c b/fs/overlayfs/inode.c
>>>>> index 541cf3717fc2..cc3ef5a6666a 100644
>>>>> --- a/fs/overlayfs/inode.c
>>>>> +++ b/fs/overlayfs/inode.c
>>>>> @@ -278,6 +278,26 @@ int ovl_getattr(struct mnt_idmap *idmap, const struct path *path,
>>>>>        return err;
>>>>>   }
>>>>>
>>>>> +static inline int ovl_i_path_realinode(struct inode *inode,
>>>>> +                                    struct path *realpath,
>>>>> +                                    struct inode **realinode, int rcu)
>>>>> +{
>>>>> +     /* Careful in RCU walk mode */
>>>>> +     ovl_i_path_real(inode, realpath);
>>>>> +     if (!realpath->dentry) {
>>>>> +             WARN_ON(!rcu);
>>>>> +             return -ECHILD;
>>>>> +     }
>>>>> +
>>>>> +     *realinode = d_inode(realpath->dentry);
>>>>> +     if (!*realinode) {
>>>>> +             WARN_ON(!rcu);
>>>>> +             return -ECHILD;
>>>>> +     }
>>>>> +
>>>>> +     return 0;
>>>>> +}
>>>>
>>>> If you want to return the inode wouldn't it possibly make more sense to
>>>> return the inode from the function directly? But not fuzzed. Maybe Amir
>>>> has a preference. As I said, I'm even fine with the original approach.
>>>
>>> Sorry for not reviewing v1, I was traveling, even though it is hard to use
>>> this excuse when I was traveling with Christian who did review v1 :)
>>
>> Well, I did only do it this morning. :)
>>
>>>
>>>>
>>>> static inline struct inode *ovl_i_path_realinode(struct inode *inode,
>>>>                                                   struct path *realpath,
>>>>                                                   int rcu)
>>>> {
>>>>          struct inode *realinode;
>>>>
>>>>          /* Careful in RCU walk mode */
>>>>          ovl_i_path_real(inode, realpath);
>>>>          if (!realpath->dentry) {
>>>>                  WARN_ON(!rcu);
>>>>                  return ERR_PTR(-ECHILD);
>>>>          }
>>>>
>>>>          realinode = d_inode(realpath->dentry);
>>>>          if (!realinode) {
>>>>                  WARN_ON(!rcu);
>>>>                  return ERR_PTR(-ECHILD);
>>>>          }
>>>>
>>>>          return realinode;
>>>> }
>>>>
>>>
>>> I think this helper is over engineered ;-)
>>
>> Yes. As mentioned before, I would've been happy even with v1 that didn't
>> have any helper.
>>
>>> The idea for a helper that returns inode is good,
>>> but I see no reason to mix RCU walk in this helper
>>> and don't even need a new helper (see untested patch below).
>>
>> Looks good to me too.
>>
>>>
>>> Thanks,
>>> Amir.
>>>
>>> ---
>>> -void ovl_i_path_real(struct inode *inode, struct path *path)
>>> +struct inode *ovl_i_path_real(struct inode *inode, struct path *path)
>>>   {
>>>          struct ovl_path *lowerpath = ovl_lowerpath(OVL_I_E(inode));
>>>
>>> @@ -342,6 +342,8 @@ void ovl_i_path_real(struct inode *inode, struct path *path)
>>>          } else {
>>>                  path->mnt = ovl_upper_mnt(OVL_FS(inode->i_sb));
>>>          }
>>> +
>>> +       return path->dentry ? d_inode(path->dentry) : NULL;
>>>   }
>>>
>>> @@ -295,8 +295,8 @@ int ovl_permission(struct mnt_idmap *idmap,
>>>          int err;
>>>
>>>          /* Careful in RCU walk mode */
>>> -       ovl_i_path_real(inode, &realpath);
>>> -       if (!realpath.dentry) {
>>> +       realinode = ovl_i_path_real(inode, &realpath);
>>> +       if (!realpath.dentry || !realinode) {
>>>                  WARN_ON(!(mask & MAY_NOT_BLOCK));
>>>                  return -ECHILD;
>>>          }
>>> @@ -309,7 +309,6 @@ int ovl_permission(struct mnt_idmap *idmap,
>>>
>>>          if (err)
>>>                  return err;
>>>
>>> -       realinode = d_inode(realpath.dentry);
>>>          old_cred = ovl_override_creds(inode->i_sb);
>>

Thanks for reviewings and helpful discussion from Amir and Christian.

>> Btw, if the reproducer that Zhihao has posted in the bugzilla link:
>>
>> #!/bin/bash
>>
>> mkdir -p /root/tmp/merge /root/tmp/upper/dir /root/tmp/lower /root/tmp/work
>> touch /root/tmp/upper/dir/file
>> chown freg:freg -R /root/tmp/upper/dir
>> mount -t overlay none -oupperdir=/root/tmp/upper,lowerdir=/root/tmp/lower,workdir=/root/tmp/work /root/tmp/merge
>> ls /root/tmp/merge/dir/file &
>> echo 3 > /proc/sys/vm/drop_caches
>>
>> is reliable we should add it to xfstests...
> 
> If only it was that easy to trigger this race.
> Look at the debug kernel patch named 'diff' in bugzilla...

Yes, I can hardly reproduce the problem without the time delay in kernel.

> 
> Thanks,
> Amir.
> .
>