Re: [PATCH v3 00/24] Convert vfs.txt to vfs.rst

[Ian Kent]

On Tue, 2019-04-02 at 20:08 +0100, Al Viro wrote:
> On Tue, Apr 02, 2019 at 06:54:01PM +0100, Al Viro wrote:
> > static void autofs_dentry_release(struct dentry *de)
> > {
> >         struct autofs_info *ino = autofs_dentry_ino(de);
> >         struct autofs_sb_info *sbi = autofs_sbi(de->d_sb);
> > 
> >         pr_debug("releasing %p\n", de);
> > 
> >         if (!ino)
> >                 return;
> > ...
> >         autofs_free_ino(ino);
> > }
> > with autofs_free_ino() being straight kfree().  Which means
> > that the lockless case of autofs_d_manage() can run into
> > autofs_dentry_ino(dentry) getting freed right under it.
> > 
> > And there we do have this reachable:
> > int autofs_expire_wait(const struct path *path, int rcu_walk)
> > {
> >         struct dentry *dentry = path->dentry;
> >         struct autofs_sb_info *sbi = autofs_sbi(dentry->d_sb);
> >         struct autofs_info *ino = autofs_dentry_ino(dentry);
> >         int status;
> >         int state;
> > 
> >         /* Block on any pending expire */
> >         if (!(ino->flags & AUTOFS_INF_WANT_EXPIRE))

Oh yes, this is saying the dentry hasn't been selected
for expire on the first pass, there's a second pass at
expire selection so there's a delay there and both flags
(this one and the expiring flag) are kept throughout the
expire operation if dentry is selected.

That might be partly why an oops has never been seen but
path walks can occur at any time so it's a bit puzzling.

LOL, and Neil probably can't remember the deeper detail
on what he did there now either.

> >                 return 0;
> >         if (rcu_walk)
> >                 return -ECHILD;
> > 
> > the second check buggers off in lockless mode; the first one
> > can be done in lockless mode just fine, so AFAICS we do have
> > a problem there.  Smells like we ought to make that kfree
> > in autofs_free_ino() RCU-delayed...  Ian, have you, by any
> > chance, run into reports like that?  Use-after-free or
> > oopsen in autofs_expire_wait() and friends, that is...
> 
> Alternatively, we could clear ->d_fsdata in autofs_d_release()
> under ->d_lock and have all potentially lockless users of
> autofs_dentry_ino() take ->d_lock around messing with that.
> I'd still prefer to do it as below, though.  Ian, do you have
> any objections against the following and, if you are OK with
> it, which tree would you prefer it to go through?
> 
> autofs: fix use-after-free in lockless ->d_manage()
> 
> autofs_d_release() can overlap with lockless ->d_manage(),
> ending up with autofs_dentry_ino() freed under the latter.
> Make freeing autofs_info instances RCU-delayed...
> 
> Signed-off-by: Al Viro <viro@xxxxxxxxxxxxxxxxxx>
> ---
> diff --git a/fs/autofs/autofs_i.h b/fs/autofs/autofs_i.h
> index 70c132acdab1..e1091312abe1 100644
> --- a/fs/autofs/autofs_i.h
> +++ b/fs/autofs/autofs_i.h
> @@ -71,6 +71,7 @@ struct autofs_info {
>  
>  	kuid_t uid;
>  	kgid_t gid;
> +	struct rcu_head rcu;
>  };
>  
>  #define AUTOFS_INF_EXPIRING	(1<<0) /* dentry in the process of expiring */
> diff --git a/fs/autofs/inode.c b/fs/autofs/inode.c
> index 80597b88718b..fb0225f21c12 100644
> --- a/fs/autofs/inode.c
> +++ b/fs/autofs/inode.c
> @@ -36,7 +36,7 @@ void autofs_clean_ino(struct autofs_info *ino)
>  
>  void autofs_free_ino(struct autofs_info *ino)
>  {
> -	kfree(ino);
> +	kfree_rcu(ino, rcu);
>  }
>  
>  void autofs_kill_sb(struct super_block *sb)