[PATCH 1/3] nfsd: fix CB_NOTIFY workqueue loop when queue overflows

[Jeff Layton <jlayton@kernel.org>]

When NOTIFY4_CHANGE_DIR_ATTRS is requested, nfsd4_cb_notify_prepare()
lowers its limit to NOTIFY4_EVENT_QUEUE_SIZE - 1 to reserve a slot for
the dir attribute update. The producer in nfsd_handle_dir_event() caps
ncn_evt_cnt at NOTIFY4_EVENT_QUEUE_SIZE, so the queue can fill to one
more than that limit.

In that case prepare() took the "we can't keep up!" branch and jumped to
out_recall without draining the queue, leaving ncn_evt_cnt nonzero.
Returning false from prepare causes nfsd41_destroy_cb() to run the
release op, and nfsd4_cb_notify_release() requeues the callback whenever
ncn_evt_cnt > 0. The requeued callback hits the same overflow check and
recalls again, spinning a workqueue thread forever. Because each cycle
holds an stid reference across the requeue, sc_count never reaches zero,
so the delegation is never freed even after the client returns it -
exhausting CPU and leaking the delegation.

Drain the queued events (dropping their references) and reset
ncn_evt_cnt before recalling, so the release op won't requeue. This also
fixes an event reference leak that previously existed on the recall path.

Also guard the release-side requeue against a revoked delegation: there
is no point notifying a client that no longer holds the delegation, and
skipping the requeue avoids pinning the stid and spinning the workqueue.

Fixes: fd0d6dde2a57 ("nfsd: add support to CB_NOTIFY for dir attribute changes")
Reported-by: Sashiko AI <https://sashiko.dev>
Signed-off-by: Jeff Layton <jlayton@kernel.org>
---
 fs/nfsd/nfs4state.c | 26 ++++++++++++++++++--------
 1 file changed, 18 insertions(+), 8 deletions(-)

diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
index 2f7210accdf1..b830aed7ae39 100644
--- a/fs/nfsd/nfs4state.c
+++ b/fs/nfsd/nfs4state.c
@@ -3546,16 +3546,21 @@ nfsd4_cb_notify_prepare(struct nfsd4_callback *cb)
 		return false;
 	}
 
-	/* we can't keep up! */
-	if (count > limit) {
-		spin_unlock(&ncn->ncn_lock);
-		goto out_recall;
-	}
-
 	memcpy(events, ncn->ncn_evt, sizeof(*events) * count);
 	ncn->ncn_evt_cnt = 0;
 	spin_unlock(&ncn->ncn_lock);
 
+	/*
+	 * We can't keep up! Drop the queued events and recall. The queue must
+	 * be drained here: out_recall leaves ncn_evt_cnt at 0, so the release
+	 * op won't see leftover events and requeue this callback forever.
+	 */
+	if (count > limit) {
+		for (i = 0; i < count; ++i)
+			nfsd_notify_event_put(events[i]);
+		goto out_recall;
+	}
+
 	rcu_read_lock();
 	nf = nfsd_file_get(rcu_dereference(dp->dl_stid.sc_file->fi_deleg_file));
 	rcu_read_unlock();
@@ -3663,8 +3668,13 @@ nfsd4_cb_notify_release(struct nfsd4_callback *cb)
 	struct nfs4_delegation *dp =
 			container_of(ncn, struct nfs4_delegation, dl_cb_notify);
 
-	/* Drain events that arrived while this callback was in flight */
-	if (READ_ONCE(ncn->ncn_evt_cnt) > 0)
+	/*
+	 * Drain events that arrived while this callback was in flight, but
+	 * don't requeue against a revoked delegation: there's no point in
+	 * notifying a client that no longer holds it, and doing so can pin the
+	 * stid and spin the workqueue.
+	 */
+	if (!dp->dl_stid.sc_status && READ_ONCE(ncn->ncn_evt_cnt) > 0)
 		nfsd4_run_cb_notify(ncn);
 	nfs4_put_stid(&dp->dl_stid);
 }

-- 
2.54.0