Re: [RFC PATCH v5 07/10] ovl: cache dirty overlayfs' inode

[Chengguang Xu]

在 2021/10/7 19:09, Miklos Szeredi 写道:
> On Thu, 23 Sept 2021 at 15:08, Chengguang Xu <cgxu519@xxxxxxxxxxxx> wrote:
> > Now drop overlayfs' inode will sync dirty data,
> > so we change to only drop clean inode.
> >
> > The purpose of doing this is to keep compatible
> > behavior with before because without this change
> > dropping overlayfs inode will not trigger syncing
> > of underlying dirty inode.
> >
> > Signed-off-by: Chengguang Xu <cgxu519@xxxxxxxxxxxx>
> > ---
> >   fs/overlayfs/super.c | 16 +++++++++++++++-
> >   1 file changed, 15 insertions(+), 1 deletion(-)
> >
> > diff --git a/fs/overlayfs/super.c b/fs/overlayfs/super.c
> > index cddae3ca2fa5..bf4000eb9be8 100644
> > --- a/fs/overlayfs/super.c
> > +++ b/fs/overlayfs/super.c
> > @@ -441,11 +441,25 @@ static int ovl_write_inode(struct inode *inode,
> >          return ret;
> >   }
> >
> > +/*
> > + * In iput_final(), clean inode will drop directly and dirty inode will
> > + * keep in the cache until write back to sync dirty data then add to lru
> > + * list to wait reclaim.
> > + */
> > +static int ovl_drop_inode(struct inode *inode)
> > +{
> > +       struct inode *upper = ovl_inode_upper(inode);
> > +
> > +       if (!upper || !(inode->i_state & I_DIRTY_ALL))
> Could we check upper dirtyness here? That would give a more precise result.

We keep tracking mmapped-file(shared mode) by explicitely marking 
overlay inode dirty,

so if we drop overlay inode by checking upper dirtyness, we may lose 
control on those mmapped upper inodes.

> Alternatively don't set .drop_inode (i.e. use generic_drop_inode())
> and set I_DONTCACHE on overlay inodes.  That would cause the upper
> inode to be always written back before eviction.
>
> The latter would result in simpler logic, and I think performance-wise
> it wouldn't matter.  But I may be missing something.

I think we may seperate mmapped-file(shared) inode and other inode by

clear/set I_DONTCACHE flag on overlay inode if you prefer this approach.


Thanks,

Chengguang