Merge #17897: init: Stop indexes on shutdown after ChainStateFlushed callback.

9dd58ca611f6f2b59c25d727a4e955333525d345 init: Stop indexes on shutdown after ChainStateFlushed callback. (Jim Posen)

Pull request description:

  Replaces https://github.com/bitcoin/bitcoin/pull/17852.

  Currently, the latest index state may not be committed to disk on shutdown. The state is committed on `ChainStateFlushed` callbacks and the current init order unregisters the indexes as validation interfaces before the final `ChainStateFlushed` callback is called on them.

  Issue identified by paulyc.

  For review: an alternative or supplemental solution would be to call `Commit` at the end of `BaseIndex::Stop`. I don't see any harm in doing so and it makes the less prone to user error. However, the destructor would have to be modified to not call `Stop` because `Commit` calls a virtual method, so I figured it wasn't worth it. But I'm curious how others feel.

ACKs for top commit:
  fjahr:
    tested ACK 9dd58ca611f6f2b59c25d727a4e955333525d345
  paulyc:
    > Code review ACK [9dd58ca](https://github.com/bitcoin/bitcoin/commit/9dd58ca611f6f2b59c25d727a4e955333525d345), but failed to test because I can't reproduce the original problem.
  kallewoof:
    Tested ACK 9dd58ca611f6f2b59c25d727a4e955333525d345
  promag:
    Code review ACK 9dd58ca611f6f2b59c25d727a4e955333525d345, but failed to test because I can't reproduce the original problem.

Tree-SHA512: 2918380b699833cb7eab07456d1667dbf8ebbe2d2b5988300a3cf5b6a6cfc818b6d9086e1936ffe7881f67e409306c4b91d61a08a169cfd0a301383479d4f3cb

1 files changed, 8 insertions, 4 deletions

diff --git a/src/init.cpp b/src/init.cpp
index 21b7a26cb9..539546117b 100644
--- a/src/init.cpp
+++ b/src/init.cpp
@@ -197,8 +197,6 @@ void Shutdown(NodeContext& node)
     // using the other before destroying them.
     if (node.peer_logic) UnregisterValidationInterface(node.peer_logic.get());
     if (node.connman) node.connman->Stop();
-    if (g_txindex) g_txindex->Stop();
-    ForEachBlockFilterIndex([](BlockFilterIndex& index) { index.Stop(); });
 
     StopTorControl();
 
@@ -212,8 +210,6 @@ void Shutdown(NodeContext& node)
     node.peer_logic.reset();
     node.connman.reset();
     node.banman.reset();
-    g_txindex.reset();
-    DestroyAllBlockFilterIndexes();
 
     if (::mempool.IsLoaded() && gArgs.GetArg("-persistmempool", DEFAULT_PERSIST_MEMPOOL)) {
         DumpMempool(::mempool);
@@ -246,6 +242,14 @@ void Shutdown(NodeContext& node)
     // CValidationInterface callbacks, flush them...
     GetMainSignals().FlushBackgroundCallbacks();
 
+    // Stop and delete all indexes only after flushing background callbacks.
+    if (g_txindex) {
+        g_txindex->Stop();
+        g_txindex.reset();
+    }
+    ForEachBlockFilterIndex([](BlockFilterIndex& index) { index.Stop(); });
+    DestroyAllBlockFilterIndexes();
+
     // Any future callbacks will be dropped. This should absolutely be safe - if
     // missing a callback results in an unrecoverable situation, unclean shutdown
     // would too. The only reason to do the above flushes is to let the wallet catch