Re: [PATCH v1] ovl: Fix leaked dentry

[Vivek Goyal]

On Mon, Mar 29, 2021 at 06:49:07PM +0200, Mickaël Salaün wrote:
> From: Mickaël Salaün <mic@xxxxxxxxxxxxxxxxxxx>
> 
> Since commit 6815f479ca90 ("ovl: use only uppermetacopy state in
> ovl_lookup()"), overlayfs doesn't put temporary dentry when there is a
> metacopy error, which leads to dentry leaks when shutting down the
> related superblock:
> 

Hi,

Thanks for finding and fixing this bug. Patch looks correct to me. We
need to drop that reference to this.

I am not able to trigger this warning on umount of overlayfs. I copied
up a file with metacopy enabled and then remounted overlay again with
metacopy disabled. That does hit this code and I see the warning.

refusing to follow metacopy origin for (/foo.txt)

This should have lead to leak of dentry pointed by "this".

But after that I unmounted, overlay and that succeeds. Is there any
additional step to be done to trigger this VFS warning.

Vivek

>   overlayfs: refusing to follow metacopy origin for (/file0)
>   ...
>   BUG: Dentry (____ptrval____){i=3f33,n=file3}  still in use (1) [unmount of overlay overlay]
>   ...
>   WARNING: CPU: 1 PID: 432 at umount_check.cold+0x107/0x14d
>   CPU: 1 PID: 432 Comm: unmount-overlay Not tainted 5.12.0-rc5 #1
>   ...
>   RIP: 0010:umount_check.cold+0x107/0x14d
>   ...
>   Call Trace:
>    d_walk+0x28c/0x950
>    ? dentry_lru_isolate+0x2b0/0x2b0
>    ? __kasan_slab_free+0x12/0x20
>    do_one_tree+0x33/0x60
>    shrink_dcache_for_umount+0x78/0x1d0
>    generic_shutdown_super+0x70/0x440
>    kill_anon_super+0x3e/0x70
>    deactivate_locked_super+0xc4/0x160
>    deactivate_super+0xfa/0x140
>    cleanup_mnt+0x22e/0x370
>    __cleanup_mnt+0x1a/0x30
>    task_work_run+0x139/0x210
>    do_exit+0xb0c/0x2820
>    ? __kasan_check_read+0x1d/0x30
>    ? find_held_lock+0x35/0x160
>    ? lock_release+0x1b6/0x660
>    ? mm_update_next_owner+0xa20/0xa20
>    ? reacquire_held_locks+0x3f0/0x3f0
>    ? __sanitizer_cov_trace_const_cmp4+0x22/0x30
>    do_group_exit+0x135/0x380
>    __do_sys_exit_group.isra.0+0x20/0x20
>    __x64_sys_exit_group+0x3c/0x50
>    do_syscall_64+0x45/0x70
>    entry_SYSCALL_64_after_hwframe+0x44/0xae
>   ...
>   VFS: Busy inodes after unmount of overlay. Self-destruct in 5 seconds.  Have a nice day...
> 
> This fix has been tested with a syzkaller reproducer.
> 
> Cc: Amir Goldstein <amir73il@xxxxxxxxx>
> Cc: Miklos Szeredi <miklos@xxxxxxxxxx>
> Cc: Vivek Goyal <vgoyal@xxxxxxxxxx>
> Cc: <stable@xxxxxxxxxxxxxxx> # v5.7+
> Reported-by: syzbot <syzkaller@xxxxxxxxxxxxxxxx>
> Fixes: 6815f479ca90 ("ovl: use only uppermetacopy state in ovl_lookup()")
> Signed-off-by: Mickaël Salaün <mic@xxxxxxxxxxxxxxxxxxx>
> Link: https://lore.kernel.org/r/20210329164907.2133175-1-mic@xxxxxxxxxxx
> ---
>  fs/overlayfs/namei.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/fs/overlayfs/namei.c b/fs/overlayfs/namei.c
> index 3fe05fb5d145..424c594afd79 100644
> --- a/fs/overlayfs/namei.c
> +++ b/fs/overlayfs/namei.c
> @@ -921,6 +921,7 @@ struct dentry *ovl_lookup(struct inode *dir, struct dentry *dentry,
>  		if ((uppermetacopy || d.metacopy) && !ofs->config.metacopy) {
>  			err = -EPERM;
>  			pr_warn_ratelimited("refusing to follow metacopy origin for (%pd2)\n", dentry);
> +			dput(this);
>  			goto out_put;
>  		}
>  
> 
> base-commit: a5e13c6df0e41702d2b2c77c8ad41677ebb065b3
> -- 
> 2.30.2
>