On Wed, Sep 09, 2020 at 07:54:44AM -0500, Eric W. Biederman wrote:
> Hao Lee <haolee.swjtu@xxxxxxxxx> writes:
>
> > On Tue, Sep 08, 2020 at 07:48:57PM +0100, Al Viro wrote:
> >> On Tue, Sep 08, 2020 at 01:06:56PM +0000, Hao Lee wrote:
> >> > ping
> >> >
> >> > On Wed, Jul 29, 2020 at 03:21:28PM +0000, Hao Lee wrote:
> >> > > The dentry local variable is introduced in 'commit 84d17192d2afd ("get
> >> > > rid of full-hash scan on detaching vfsmounts")' to reduce the length of
> >> > > some long statements for example
> >> > > mutex_lock(&path->dentry->d_inode->i_mutex). We have already used
> >> > > inode_lock(dentry->d_inode) to do the same thing now, and its length is
> >> > > acceptable. Furthermore, it seems not concise that assign path->dentry
> >> > > to local variable dentry in the statement before goto. So, this function
> >> > > would be more clear if we eliminate the local variable dentry.
> >>
> >> How does it make the function more clear? More specifically, what
> >> analysis of behaviour is simplified by that?
> >
> > When I first read this function, it takes me a few seconds to think
> > about if the local variable dentry is always equal to path->dentry and
> > want to know if it has special purpose. This local variable may confuse
> > other people too, so I think it would be better to eliminate it.
>
> I tend to have the opposite reaction. I read your patch and wonder
> why path->dentry needs to be reread what is changing path that I can not see.
> my back.
If I understand correctly, accessing path->dentry->d_inode needs one
more instruction than accessing dentry->d_inode, so the original code is
more efficient.
>
> Now for clarity it would probably help to do something like:
>
> diff --git a/fs/namespace.c b/fs/namespace.c
> index bae0e95b3713..430f3b4785e3 100644
> --- a/fs/namespace.c
> +++ b/fs/namespace.c
> @@ -2206,7 +2206,7 @@ static struct mountpoint *lock_mount(struct path *path)
> return mp;
> }
> namespace_unlock();
> - inode_unlock(path->dentry->d_inode);
> + inode_unlock(dentry->d_inode);
> path_put(path);
> path->mnt = mnt;
> dentry = path->dentry = dget(mnt->mnt_root);
>
>
> So at least the inode_lock and inode_unlock are properly paired.
>
> At first glance inode_unlock using path->dentry instead of dentry
> appears to be an oversight in 84d17192d2af ("get rid of full-hash scan
> on detaching vfsmounts").
I think I have understood why we use the local variable dentry. Thanks.
Regards,
Hao Lee
>
>
> Eric
