On Mon, Nov 9, 2020 at 10:34 AM Chengguang Xu <cgxu519@xxxxxxxxxxxx> wrote:
>
> ---- 在 星期一, 2020-11-09 15:07:18 Amir Goldstein <amir73il@xxxxxxxxx> 撰写 ----
> > On Mon, Nov 9, 2020 at 5:34 AM Chengguang Xu <cgxu519@xxxxxxxxxxxx> wrote:
> > >
> > > ---- 在 星期日, 2020-11-08 22:03:06 Chengguang Xu <cgxu519@xxxxxxxxxxxx> 撰写 ----
> > > > Introduce a helper ovl_wait_wb_inodes() to wait until all
> > > > target upper inodes finish writeback.
> > > >
> > > > Signed-off-by: Chengguang Xu <cgxu519@xxxxxxxxxxxx>
> > > > ---
> > > > fs/overlayfs/super.c | 30 ++++++++++++++++++++++++++++++
> > > > 1 file changed, 30 insertions(+)
> > > >
> > > > diff --git a/fs/overlayfs/super.c b/fs/overlayfs/super.c
> > > > index e5607a908d82..9a535fc11221 100644
> > > > --- a/fs/overlayfs/super.c
> > > > +++ b/fs/overlayfs/super.c
> > > > @@ -255,6 +255,36 @@ static void ovl_put_super(struct super_block *sb)
> > > > ovl_free_fs(ofs);
> > > > }
> > > >
> > > > +void ovl_wait_wb_inodes(struct ovl_fs *ofs)
> > > > +{
> > > > + LIST_HEAD(tmp_list);
> > > > + struct ovl_inode *oi;
> > > > + struct inode *upper;
> > > > +
> > > > + spin_lock(&ofs->syncfs_wait_list_lock);
> > > > + list_splice_init(&ofs->syncfs_wait_list, &tmp_list);
> > > > +
> > > > + while (!list_empty(&tmp_list)) {
> > > > + oi = list_first_entry(&tmp_list, struct ovl_inode, wait_list);
> > > > + list_del_init(&oi->wait_list);
> > > > + ihold(&oi->vfs_inode);
> > >
> > > Maybe I overlooked race condition with inode eviction, so still need to introduce
> > > OVL_EVICT_PENDING flag just like we did in old syncfs efficiency patch series.
> > >
> >
> > I am not sure why you added the ovl wait list.
> >
> > I think you misunderstood Jan's suggestion.
> > I think what Jan meant is that ovl_sync_fs() should call
> > wait_sb_inodes(upper_sb)
> > to wait for writeback of ALL upper inodes after sync_filesystem()
> > started writeback
> > only on this ovl instance upper inodes.
> >
>
>
> Maybe you are right, the wait list is just for accuracy that can completely
> avoid interferes between ovl instances, otherwise we may need to face
> waiting interferes in high density environment.
>
>
> > I am not sure if this is acceptable or not - it is certainly an improvement over
> > current situation, but I have a feeling that on a large scale (many
> > containers) it
> > won't be enough.
> >
>
> The same as your thought.
>
>
> > The idea was to keep it simple without over optimizing, since anyway
> > you are going for the "correct" solution long term (ovl inode aops),
> > so I wouldn't
> > add the wait list.
> >
>
> Maybe, I think it depends on how to implement ovl page-cache, so at current
> stage I have no idea for the wait list.
>
>
> > As long as the upper inode is still dirty, we can keep the ovl inode in cache,
> > so the worst outcome is that drop_caches needs to get called twice before the
> > ovl inode can be evicted, no?
> >
>
> IIUC, since currently ovl does not have it's own page-cache, so there is no affect to page-cache reclaim,
> also there is no ovl shrinker to reclaim slab because we drop ovl inode directly after final iput.
> So should we add a shrinker in this series?
>
Would that add a lot of complexity?
Thinking out loud: maybe we follow Jan's suggestion and fix remaining
performance with followup series?
Thanks,
Amir.
