Re: [PATCH] adding per sb inode list to make invalidate_inodes() faster

[Andrew Morton <akpm@osdl.org>]

William Lee Irwin III <wli@holomorphy.com> wrote:
>
> On Thu, Sep 09, 2004 at 08:51:45AM -0700, Linus Torvalds wrote:
>  > Hmm.. I don't mind the approach per se, but I get very nervous about the 
>  > fact that I don't see any initialization of "inode->i_sb_list".
>  > Yes, you do a
>  > 	list_add(&inode->i_sb_list, &sb->s_inodes);
>  > in new_inode(), but there are a ton of users that allocate inodes other 
>  > ways, and more importantly, even if this was the only allocation function, 
>  > you do various "list_del(&inode->i_sb_list)" things which leaves the inode 
>  > around but with an invalid superblock list.
>  > So at the very _least_, you should document why all of this is safe very 
>  > carefully (I get nervous about fundamental FS infrastructure changes), and 
>  > it should be left to simmer in -mm for a longish time to make sure it 
>  > really works..
>  > Call me chicken.
> 
>  Some version of this patch has been in 2.6.x-mm for a long while.

One year.

> I've
>  not reviewed this version of the patch for differences with the -mm
>  code. It would probably be best to look at the -mm bits as they've had
>  sustained exposure for quite some time.

Yes.

I have not merged it up because it seems rather dopey to add eight bytes to
the inode to speed up something as rare as umount.

Is there a convincing reason for proceeding with the change?