Re: [PATCH] Set MS_ACTIVE in isofs_fill_super()

[Andrew Morton <akpm@osdl.org>]

Russ Weight <rweight@us.ibm.com> wrote:
>
> This patch sets the MS_ACTIVE bit in isofs_fill_super() prior to calling
> iget() or iput(). This eliminates a race condition between mount
> (for isofs) and kswapd that results in a system panic.
> 
> Signed-off-by: Russ Weight <rweight@us.ibm.com>
> 
> --- linux-2.6.12-rc1/fs/isofs/inode.c	2005-03-17 17:34:36.000000000
> -0800
> +++ linux-2.6.12-rc1-isofsfix/fs/isofs/inode.c	2005-03-22
> 15:29:51.945607217 -0800
> @@ -820,6 +820,7 @@
>  	 * the s_rock flag. Once we have the final s_rock value,
>  	 * we then decide whether to use the Joliet descriptor.
>  	 */
> +	s->s_flags |= MS_ACTIVE;
>  	inode = isofs_iget(s, sbi->s_firstdatazone, 0);
>  
>  	/*
> @@ -909,6 +910,7 @@
>  		kfree(opt.iocharset);
>  	kfree(sbi);
>  	s->s_fs_info = NULL;
> +	s->s_flags &= ~MS_ACTIVE;
>  	return -EINVAL;
>  }
>  

The patch is obviously safe enough, but seems a bit kludgy.

The basic problem here appears to be that isofs is doing iget/iput in
->fill_super before MS_ACTIVE is set and the inode freeing code
(generic_forget_inode) doesn't expect that to happen, yes?

I wonder if it would make more sense for all the ->fill_super callers to
set MS_ACTIVE prior to calling ->fill_super(), and clear MS_ACTIVE if
fill_super() failed?