From 72ef9e1eeb6ab60a22e5b1c99cf5545a70f7aa25 Mon Sep 17 00:00:00 2001
From: cdecker <cdecker@blockstream.com>
Date: Tue, 13 Oct 2026 17:07:33 +0200
Subject: [PATCH] plugin: Add FORGET_CHANNEL to stuck HSM types

Type 34 (WIRE_HSMD_FORGET_CHANNEL) is sent synchronously by
lightningd's delete_channel() once onchaind completes for a closed
channel. In bgsync mode no signer is attached and the StagingHsmServer
only answers the canned startup messages, so the request sits in the
stager forever and blocks lightningd's main loop.

Because the type was not sticky the node did not report itself as
stuck, so update_state called getinfo on the blocked lightningd and
the state poll stalled: no BlockSync self-preempt, no 60s stuck-grace
kill and no stuck webhook to summon the signer. The session then
idled until the 1h session wall. On EU bgsync this accounted for
~95% of the ~2,029 cap-hit sessions/day, with 1,936 "forgetting
peer" events/day repeating on every visit to the same nodes (Loki,
2026-10-12).

Marking the type stuck makes update_state skip the hanging getinfo,
fires the stuck webhook so the signer gets summoned, and kills the
session after ~60s instead of ~3,540s. The forget then commits for
good once a user session attaches the signer.
---
 libs/gl-plugin/src/stager.rs | 37 ++++++++++++++++++++++++++++++++++++
 1 file changed, 37 insertions(+)

diff --git a/libs/gl-plugin/src/stager.rs b/libs/gl-plugin/src/stager.rs
index 05ecbc697..b4a858bf7 100644
--- a/libs/gl-plugin/src/stager.rs
+++ b/libs/gl-plugin/src/stager.rs
@@ -109,6 +109,7 @@ impl Stage {
             20,  // WIRE_HSMD_SIGN_REMOTE_HTLC_TX
             21,  // WIRE_HSMD_SIGN_MUTUAL_CLOSE_TX
             28,  // WIRE_HSMD_CHECK_PUBKEY
+            34,  // WIRE_HSMD_FORGET_CHANNEL (lightningd delete_channel, sync)
             142, // WIRE_HSMD_SIGN_ANY_DELAYED_PAYMENT_TO_US
             143, // WIRE_HSMD_SIGN_ANY_REMOTE_HTLC_TO_US
             144, // WIRE_HSMD_SIGN_ANY_PENALTY_TO_US
@@ -260,4 +261,40 @@ mod test {
         f1.await.unwrap();
         f2.await.unwrap();
     }
+
+    #[tokio::test]
+    async fn test_stuck_request_types() {
+        // WIRE_HSMD_FORGET_CHANNEL (34) is queued but never answered: this is
+        // the bgsync deadlock, lightningd's delete_channel() blocks on it
+        // synchronously, so the node must report itself as stuck.
+        let stage = Stage::new();
+        stage
+            .send(pb::HsmRequest {
+                request_id: 0,
+                context: None,
+                raw: vec![0, 34, 0, 0, 0, 0, 0, 0, 0, 1],
+                signer_state: vec![],
+                requests: vec![],
+            })
+            .await
+            .unwrap();
+        assert_eq!(stage.stuck_request_types().await, vec![34]);
+        assert!(stage.is_stuck().await);
+
+        // WIRE_HSMD_DERIVE_SECRET (27) is not on the sticky list, a queued
+        // request of this type alone must not mark the node as stuck.
+        let stage = Stage::new();
+        stage
+            .send(pb::HsmRequest {
+                request_id: 0,
+                context: None,
+                raw: vec![0, 27, 0, 0, 0, 0, 0, 0, 0, 1],
+                signer_state: vec![],
+                requests: vec![],
+            })
+            .await
+            .unwrap();
+        assert_eq!(stage.stuck_request_types().await, Vec::<u16>::new());
+        assert!(!stage.is_stuck().await);
+    }
 }