On Thu, 6 Nov 2008 09:32:31 -0500
Chuck Lever <chuck.lever@xxxxxxxxxx> wrote:
> On Nov 6, 2008, at Nov 6, 2008, 7:17 AM, Jeff Layton wrote:
> > On Wed, 5 Nov 2008 15:06:51 -0500
> > "J. Bruce Fields" <bfields@xxxxxxxxxxxxxx> wrote:
> >
> >> These clients will all be shut down by nlm_destroy_host() when we do
> >> garbage collection a little later, so this is redundant.
> >>
> >> XXX: Ask Jeff Layton why he added this again?
> >>
> >> Signed-off-by: J. Bruce Fields <bfields@xxxxxxxxxxxxxx>
> >> ---
> >> fs/lockd/host.c | 8 +-------
> >> 1 files changed, 1 insertions(+), 7 deletions(-)
> >>
> >> diff --git a/fs/lockd/host.c b/fs/lockd/host.c
> >> index 73c2be2..0387c6b 100644
> >> --- a/fs/lockd/host.c
> >> +++ b/fs/lockd/host.c
> >> @@ -602,14 +602,8 @@ static void expire_hosts(struct host_table
> >> *table)
> >> struct hlist_node *pos;
> >> struct nlm_host *host;
> >>
> >> - dprintk("lockd: nuking all hosts...\n");
> >> - for_each_host(host, pos, chain, table) {
> >> + for_each_host(host, pos, chain, table)
> >> host->h_expires = jiffies - 1;
> >> - if (host->h_rpcclnt) {
> >> - rpc_shutdown_client(host->h_rpcclnt);
> >> - host->h_rpcclnt = NULL;
> >> - }
> >> - }
> >> }
> >>
> >> /*
> >
> > Thank goodness for my OC commenting in the BZ I was using to track
> > this!
> >
> > https://bugzilla.redhat.com/show_bug.cgi?id=254195#c4
> >
> > IIRC, there is a chicken and egg problem with refcounting -- at
> > least in the code at the time that I did this. If there is still
> > an active grant callback in queue at the time that nlm_shutdown_hosts
> > is called, the h_count will stay high and nlm_destroy_host won't be
> > called. This can happen if we try to bring down lockd while trying to
> > do a grant callback to an unresponsive client.
>
> If that's the case, it seems to me what we want is, during lockd
> shutdown, a way to forcibly retire all outstanding grant callbacks
> before walking the host cache. Isn't that what
> nlmsvc_invalidate_all() does?
>
No, I don't think it does. It tears down the nlm_blocks and such, but
in-flight grant callbacks are "disconnected" from the NLM subsystem to
a large degree. The NLM subsystem doesn't really have any connection to
these RPC's other than via the rpc_client.
Note that I'm going from memory here, so recent changes may mean that
this is wrong. From a quick glance at the code though, I think it's
still correct.
--
Jeff Layton <jlayton@xxxxxxxxxx>
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
