Re: NULL pointer dereference when access /proc/net

[haosdent <haosdent@gmail.com>]

> really should not assume ->d_inode stable

Hi, Alexander, sorry to disturb you again. Today I try to check what
`dentry->d_inode` and `nd->link_inode` looks like when `dentry` is
already been killed in `__dentry_kill`.

```
nd->last.name: net/sockstat, dentry->d_lockref.count: -128,
dentry->d_inode: (nil), nd->link_inode: 0xffffffffab299966
nd->last.name: net/sockstat, dentry->d_lockref.count: -128,
dentry->d_inode: (nil), nd->link_inode: 0xffffffffab299966
nd->last.name: net/sockstat, dentry->d_lockref.count: -128,
dentry->d_inode: (nil), nd->link_inode: 0xffffffffab299966
```

It looks like `dentry->d_inode` could be NULL while `nd->link_inode`
is always has value.
But this make me confuse, by right `nd->link_inode` is get from
`dentry->d_inode`, right?

For example, in `walk_component`, suppose we go into `lookup_slow`,

```
static int walk_component(struct nameidata *nd, int flags)
   if (unlikely(err <= 0)) {
    ...
    path.dentry = lookup_slow(&nd->last, nd->path.dentry, nd->flags);
    ...
    inode = d_backing_inode(path.dentry);     <=== get `inode` from
`dentry->d_inode`.
  }
  return step_into(nd, &path, flags, inode, seq);  <=== set `inode` to
`nd->link_inode`.
}

```

then in `step_into` -> `pick_link`

```
static int pick_link(struct nameidata *nd, struct path *link,
     struct inode *inode, unsigned seq)
{
  ...
  nd->link_inode = inode;  <=== set `inode` to `nd->link_inode`.
}
```

So for the mismatch of `nd->link_inode` and `dentry->d_inode` in
following output. Do it means in Thread 1,  `walk_component`
get a `dentry` from `d_lookup`, at the same time, in Thread 2,
`__dentry_kill` is run and set `dentry->d_inode` to NULL.

```
nd->last.name: net/sockstat, dentry->d_lockref.count: -128,
dentry->d_inode: (nil), nd->link_inode: 0xffffffffab299966
nd->last.name: net/sockstat, dentry->d_lockref.count: -128,
dentry->d_inode: (nil), nd->link_inode: 0xffffffffab299966
nd->last.name: net/sockstat, dentry->d_lockref.count: -128,
dentry->d_inode: (nil), nd->link_inode: 0xffffffffab299966
```

If these concurrent operations in `dentry->d_inode` could happen, how
we ensure `nd->link_inode = inode` and `d_backing_inode` are always
run before
`__dentry_kill`? I still could not find the questions for this from
dcache's code, sorry for the stupid question.

On Mon, Apr 26, 2021 at 1:22 AM Al Viro <viro@zeniv.linux.org.uk> wrote:
>
> On Mon, Apr 26, 2021 at 01:04:46AM +0800, haosdent wrote:
> > Hi, Alexander, thanks a lot for your quick reply.
> >
> > > Not really - the crucial part is ->d_count == -128, i.e. it's already past
> > > __dentry_kill().
> >
> > Thanks a lot for your information, we would check this.
> >
> > > Which tree is that?
> > > If you have some patches applied on top of that...
> >
> > We use Ubuntu Linux Kernel "4.15.0-42.45~16.04.1" from launchpad directly
> > without any modification,  the mapping Linux Kernel should be
> > "4.15.18" according
> > to https://people.canonical.com/~kernel/info/kernel-version-map.html
>
> Umm...  OK, I don't have it Ubuntu source at hand, but the thing to look into
> would be
>         * nd->flags contains LOOKUP_RCU
>         * in the mainline from that period (i.e. back when __atime_needs_update()
> used to exist) we had atime_needs_update_rcu() called in get_link() under those
> conditions, with
> static inline bool atime_needs_update_rcu(const struct path *path,
>                                           struct inode *inode)
> {
>         return __atime_needs_update(path, inode, true);
> }
> and __atime_needs_update() passing its last argument (rcu:true in this case) to
> relatime_need_update() in
>         if (!relatime_need_update(path, inode, now, rcu))
> relatime_need_update() hitting
>         update_ovl_inode_times(path->dentry, inode, rcu);
> and update_ovl_inode_times() starting with
>         if (rcu || likely(!(dentry->d_flags & DCACHE_OP_REAL)))
>                 return;
> with subsequent accesses to ->d_inode.  Those obviously are *NOT* supposed
> to be reached in rcu mode, due to that check.
>
> Your oops looks like something similar to that call chain had been involved and
> somehow had managed to get through to those ->d_inode uses.
>
> Again, in RCU mode we really, really should not assume ->d_inode stable.  That's
> why atime_needs_update() gets inode as a separate argument and does *NOT* look
> at path->dentry at all.  In the kernels of 4.8..4.18 period there it used to do
> so, but only in non-RCU mode (which is the reason for explicit rcu argument passed
> through that callchain).



-- 
Best Regards,
Haosdent Huang