Re: [2.4.19-rc4 and 2.4.19-rc4-mm2] super block list corruption following fill_super returns fail

[Andrew Morton <akpm@osdl.org>]

On Fri, 10 Nov 2006 20:33:42 +0800
Ian Kent <raven@themaw.net> wrote:

> I'm seeing an oops after returning a fail status from the autofs and
> autofs4 fill_super methods. The scenario is a little contrived but does
> demonstrate the mount fail case.
> 
> get_super+0x78 corresponds to:
> 
>                         down_read(&sb->s_umount);
> ---->                   if (sb->s_root)
>                                 return sb;
>                         up_read(&sb->s_umount);
> 
> So I believe that, following the fill_super call in get_sb_nodev the
> super block is freed during the call to deactivate_super but not removed
> from the supers list.
> 
> As far as I can tell I've done the appropriate housekeeping in the
> autofs[4] fill_super function. In particular, sb->s_root is not set upon
> mount fail.
> 

Yup, sget() adds the superblock to super_blocks and deactivate_super()
doesn't take it off.

> Does anyone have any suggestions as to what I might not be doing that I
> should be doing that is preventing this removal? 

Well afacit the only piece of code which knows how to remove a superblock
from the global list is generic_shutdown_super().  So perhaps your
->fill_super() implementation is supposed to run generic_shutdown_super()
if it's about to return an error.

Seems like an odd API though.