On Wed, Aug 07, 2019 at 12:42:08PM -0400, Olga Kornievskaia wrote:
> On Wed, Aug 7, 2019 at 12:09 PM J. Bruce Fields <bfields@xxxxxxxxxxxx> wrote:
> >
> > On Wed, Aug 07, 2019 at 12:02:40PM -0400, Olga Kornievskaia wrote:
> > > On Thu, Aug 1, 2019 at 3:36 PM J. Bruce Fields <bfields@xxxxxxxxxx> wrote:
> > > >
> > > > On Thu, Aug 01, 2019 at 02:24:04PM -0400, Olga Kornievskaia wrote:
> > > > > i was just looking at close_lru and delegation_lru but I guess that's
> > > > > not a list of delegation or open stateids but rather some complex of
> > > > > not deleting the stateid right away but moving it to nfs4_ol_stateid
> > > > > and the list on the nfsd_net. Are you looking for something similar
> > > > > for the copy_notify state or can I just keep a global list of the
> > > > > nfs4_client and add and delete of that (not move to the delete later)?
> > > >
> > > > A global list seems like it should work if the locking's OK.
> > >
> > > I'm having issues taking a reference on a parent stateid and being
> > > able to clean it. Let me try to explain.
> >
> > With other stateid parent relationships I believe what we do is: instead
> > of the child taking a reference on the parent, we ensure that the child
> > is destroyed, and that nobody can be holding a pointer to it, before we
> > destroy the parent.
>
> I don't think we can get away from not taking a reference on the
> parent. When a READ comes with the copy_notify stateid, it's used to
> lookup the parent state because the nfs4_preprocess_stateid_op() that
> checks the validity of the stateid for a given operation needs to
> check validity of that parent stateid). Otherwise, we'd have to
> special case the READ calling nfs4_preprocess_stateid_op() and special
> call that function to when called from READ and finding a copy_notify
> stateid will forego the other checks. Do you want me to that instead
> of what I proposed below?
Um, honestly I'm not sure I understand your code below yet. I'll take
another look....
> > > Since I take a reference on the stateid, then during what would have
> > > been the last put (due to say a close operation), stateid isn't
> > > released. Now that stateid is sticking around. I personally would have
> > > liked on what would have been a close and release of the stateid to
> > > release the copy notify state(s)
That's OK with me as long as it works. Did I complain about it? The
only real requirement is that we've got *some* way to assure that we
aren't going to find a copy_notify stateid and try to follow it to its
parent, after the parent's been freed.
--b.
> > > (which was being done before but
> > > having a reference makes it hard? i want to count number of copy
> > > notify states and if then somehow if the num_copies-1 is going to make
> > > it 0, then decrement by num_copies (and the normal -1) but if it's not
> > > the last reference then it shouldn't be decremented.
> > >
> > > Now say no fancy logic happens on close so we have these stateids left
> > > over . What to do on unmount? It will error with err_client_busy since
> > > there are non-zero copy notify states and only after a lease period it
> > > will release the resources (when the close of the file should have
> > > removed any copy notify state)?
> > >
> > > Question: would it be acceptable to do something like this on freeing
> > > of the parent stateid?
> > >
> > > @@ -896,8 +931,12 @@ static void block_delegations(struct knfsd_fh *fh)
> > > might_lock(&clp->cl_lock);
> > >
> > > if (!refcount_dec_and_lock(&s->sc_count, &clp->cl_lock)) {
> > > - wake_up_all(&close_wq);
> > > - return;
> > > + if (!refcount_sub_and_test_checked(s->sc_cp_list_size,
> > > + &s->sc_count)) {
> > > + refcount_add_checked(s->sc_cp_list_size, &s->sc_count);
> > > + wake_up_all(&close_wq);
> > > + return;
> > > + }
> > > }
> > > idr_remove(&clp->cl_stateids, s->sc_stateid.si_opaque.so_id);
> > > spin_unlock(&clp->cl_lock);
> > >
> > > then free the copy notify stateids associated with stateid.
> > >
> > > Laundromat would still be checking the copy_notify stateids for
> > > anything that's been not active for a while (but not closed).
> > >
> > >
> > >
> > >
> > >
> > > >
> > > > --b.
