Re: swapoff() behaviour

[Alexander Viro]

On Fri, 21 Sep 2001, Alexander Viro wrote:

> 	Guys, do we want
> 
> swapon /dev/sda6
> cp -a /dev/sda6 /tmp
> swapoff /tmp/sda6
> 
> to work?  Code in sys_swapoff() looks like an attempt to make it work.
> However, it doesn't -
>                 if ((p->flags & SWP_WRITEOK) == SWP_WRITEOK) {
>                         if (p->swap_file) {
>                                 if (p->swap_file == nd.dentry)
>                                   break;
>                         } else {
>                                 if (S_ISBLK(nd.dentry->d_inode->i_mode)
>                                     && (p->swap_device == nd.dentry->d_inode->i_
> rdev))
>                                   break;
>                         }
> 		}
> 
> _never_ hits the second break since we have ->swap_file set for all
> active swap components, be they swap files or swap partitions.
> 
> Comments?  The same goes for 2.2, BTW.

After a look through the CVS, it seems that prior to 2.1.45 we actually
had that behaviour - for swap partitions swapoff() cared only about
major:minor.  Since 2.1.45 we really need the same dentry - not even
the same inode (i.e.
swapon /dev/sda6 && ln /dev/sda6 /dev/foo && swapoff /dev/foo
is guaranteed to fail; prior to 2.1.45 it used to work).

In other words, during the last 4 years quite a few pieces of mm/swapfile.c
had been a dead code.  Looks like we either need to restore old behaviour
or perform the amputation.  Snippet above is not the only place of that
kind.

Linus, it's your call.