Re: [PATCH RFC] __bd_forget should wait for inodes using the mapping

[Andrew Morton <akpm@osdl.org>]

Chris Mason <mason@suse.com> wrote:
>
> On Fri, 2004-06-18 at 17:27, Andrew Morton wrote:
> > Chris Mason <mason@suse.com> wrote:
> > >
> > > [ skip writing block-special inodes ]
> > > 
> > > Hmmm, any risk in missing data integrity syncs because of this?
> > 
> > Need to think about that.  sys_fsync(), sys_fdatasync() and sys_msync() go
> > direct to file->f_mapping and sys_sync() will sync the blockdev via its
> > kernel-internal inode.  What does that leave?
> 
> I was worried about O_SYNC, That actually looks safe though,
> generic_osync_inode will first write the mapping via filemap_fdatawrite
> (the mapping comes from f_mapping).
> 
> It doesn't really give me warm fuzzies, but looks safe enough.  Al had a
> slightly different plan, maybe with your patch we can push his larger
> changes off a bit?

>From a design POV the patch I sent isn't very nice, and does add code to a
warmpath.  If there's some way in which we can defer the i_mapping switch
until all references have gone away, that would be better?