Re: [RFC PATCH v2 14/14] dcache: Implement object migration

From: Al Viro
Date: Wed Apr 03 2019 - 13:49:25 EST

On Wed, Apr 03, 2019 at 06:19:21PM +0100, Al Viro wrote:
> On Wed, Apr 03, 2019 at 06:08:11PM +0100, Al Viro wrote:
> > Oh, *brilliant*
> >
> > Let's do d_invalidate() on random dentries and hope they go away.
> > With convoluted and brittle logics for deciding which ones to
> > spare, which is actually wrong. This will pick mountpoints
> > and tear them out, to start with.
> >
> > NAKed-by: Al Viro <viro@xxxxxxxxxxxxxxxxxx>
> >
> > And this is a NAK for the entire approach; if it has a positive refcount,
> > LEAVE IT ALONE. Period. Don't play this kind of games, they are wrong.
> > d_invalidate() is not something that can be done to an arbitrary dentry.
> PS: "try to evict what can be evicted out of this set" can be done, but
> you want something like
> start with empty list
> go through your array of references
> grab dentry->d_lock
> if dentry->d_lockref.count is not zero
> unlock and continue
> if dentry->d_flags & DCACHE_SHRINK_LIST
> ditto, it's not for us to play with
> if (dentry->d_flags & DCACHE_LRU_LIST)
> d_lru_del(dentry);
> d_shrink_add(dentry, &list);
> unlock
> on the collection phase and
> if the list is not empty by the end of that loop
> shrink_dentry_list(&list);
> on the disposal.

Note, BTW, that your constructor is wrong - all it really needs to do
is spin_lock_init() and setting ->d_lockref.count same as lockref_mark_dead()
does, to match the state of dentries being torn down.

__d_alloc() is not holding ->d_lock, since the object is not visible to
anybody else yet; with your changes it *is* visible. However, if the
assignment to ->d_lockref.count in __d_alloc() is guaranteed to be
non-zero to non-zero, the above should be safe.