SERVER-58392 Unpin durable history on rollback to refetch

6 files changed, 41 insertions, 12 deletions

diff --git a/src/mongo/db/repl/SConscript b/src/mongo/db/repl/SConscript
index 88c49da1d7f..ed47b9c0440 100644
--- a/src/mongo/db/repl/SConscript
+++ b/src/mongo/db/repl/SConscript
@@ -273,6 +273,7 @@ env.Library(
     LIBDEPS_PRIVATE=[
         '$BUILD_DIR/mongo/base',
         '$BUILD_DIR/mongo/db/storage/journal_flusher',
+        '$BUILD_DIR/mongo/db/storage/storage_control',
         '$BUILD_DIR/mongo/db/storage/storage_options',
         'oplog',
         'oplog_application',
@@ -314,7 +315,7 @@ env.Library(
     source=[
         'scatter_gather_algorithm.cpp',
         'scatter_gather_runner.cpp',
-        ],
+    ],
     LIBDEPS=[
         '$BUILD_DIR/mongo/executor/task_executor_interface',
     ],
@@ -816,7 +817,7 @@ env.Library(
         'member_config.idl',
         'repl_set_config.idl',
         'repl_set_config_params.idl',
-            ],
+    ],
     LIBDEPS=[
         '$BUILD_DIR/mongo/base',
         '$BUILD_DIR/mongo/bson/util/bson_extract',
@@ -967,7 +968,7 @@ env.Library(
         'database_cloner.cpp',
         'initial_sync_base_cloner.cpp',
     ],
-    LIBDEPS = [
+    LIBDEPS=[
         '$BUILD_DIR/mongo/base',
         '$BUILD_DIR/mongo/client/clientdriver_network',
         '$BUILD_DIR/mongo/util/concurrency/thread_pool',
@@ -997,7 +998,7 @@ env.Library(
         'tenant_database_cloner.cpp',
         'tenant_base_cloner.cpp',
     ],
-    LIBDEPS = [
+    LIBDEPS=[
         '$BUILD_DIR/mongo/base',
         '$BUILD_DIR/mongo/db/dbdirectclient',
         'base_cloner',
@@ -1248,7 +1249,7 @@ env.Library(
 
 env.Library(
     target='tenant_migration_errors',
-    source= [
+    source=[
         'tenant_migration_conflict_info.cpp',
     ],
     LIBDEPS=[
@@ -1313,7 +1314,7 @@ env.Library(
 
 env.Library(
     target='tenant_migration_recipient_service',
-    source= [
+    source=[
         'tenant_migration_recipient_service.cpp',
     ],
     LIBDEPS=[
@@ -1607,6 +1608,7 @@ if wiredtiger:
             '$BUILD_DIR/mongo/db/service_context_test_fixture',
             '$BUILD_DIR/mongo/db/stats/counters',
             '$BUILD_DIR/mongo/db/storage/ephemeral_for_test/storage_ephemeral_for_test_core',
+            '$BUILD_DIR/mongo/db/storage/storage_control',
             '$BUILD_DIR/mongo/db/transaction',
             '$BUILD_DIR/mongo/db/update/update_test_helpers',
             '$BUILD_DIR/mongo/dbtests/mocklib',
@@ -1721,7 +1723,8 @@ env.CppUnitTest(
         'tenant_cloner_test_fixture.cpp',
     ],
     LIBDEPS=[
-        '$BUILD_DIR/mongo/db/auth/authmocks',  # Required for service context test fixture
+        # Required for service context test fixture
+        '$BUILD_DIR/mongo/db/auth/authmocks',
         '$BUILD_DIR/mongo/db/catalog/collection_options',
         '$BUILD_DIR/mongo/db/service_context_d_test_fixture',
         '$BUILD_DIR/mongo/dbtests/mocklib',
diff --git a/src/mongo/db/repl/replication_recovery.cpp b/src/mongo/db/repl/replication_recovery.cpp
index e4f49e5c132..2af2027aaec 100644
--- a/src/mongo/db/repl/replication_recovery.cpp
+++ b/src/mongo/db/repl/replication_recovery.cpp
@@ -51,6 +51,7 @@
 #include "mongo/db/server_recovery.h"
 #include "mongo/db/session.h"
 #include "mongo/db/storage/control/journal_flusher.h"
+#include "mongo/db/storage/durable_history_pin.h"
 #include "mongo/db/storage/storage_parameters_gen.h"
 #include "mongo/db/transaction_history_iterator.h"
 #include "mongo/db/transaction_participant.h"
@@ -551,13 +552,19 @@ void ReplicationRecoveryImpl::_recoverFromUnstableCheckpoint(OperationContext* o
         // When `recoverFromOplog` truncates the oplog, that also happens to set the "oldest
         // timestamp" to the truncation point[1]. `_applyToEndOfOplog` will then perform writes
         // before the truncation point. Doing so violates the constraint that all updates must be
-        // timestamped newer than the "oldest timestamp". This call will move the "oldest
+        // timestamped newer than the "oldest timestamp". So we will need to move the "oldest
         // timestamp" back to the `startPoint`.
         //
+        // Before doing so, we will remove any pins. Forcing the oldest timestamp backwards will
+        // error if there are pins in place, as those pin requests will no longer be satisfied.
+        // Recovering from an unstable checkpoint has no history in the first place. Thus, clearing
+        // pins has no real effect on history being held.
+        //
         // [1] This is arguably incorrect. On rollback for nodes that are not keeping history to
         // the "majority point", the "oldest timestamp" likely needs to go back in time. The
         // oplog's `cappedTruncateAfter` method was a convenient location for this logic, which,
         // unfortunately, conflicts with the usage above.
+        DurableHistoryRegistry::get(opCtx->getServiceContext())->clearPins(opCtx);
         opCtx->getServiceContext()->getStorageEngine()->setOldestTimestamp(
             appliedThrough.getTimestamp());
 
diff --git a/src/mongo/db/repl/replication_recovery_test.cpp b/src/mongo/db/repl/replication_recovery_test.cpp
index 742313ddae1..e3818a07c71 100644
--- a/src/mongo/db/repl/replication_recovery_test.cpp
+++ b/src/mongo/db/repl/replication_recovery_test.cpp
@@ -50,6 +50,7 @@
 #include "mongo/db/service_context_d_test_fixture.h"
 #include "mongo/db/session_catalog_mongod.h"
 #include "mongo/db/session_txn_record_gen.h"
+#include "mongo/db/storage/durable_history_pin.h"
 #include "mongo/db/storage/storage_options.h"
 #include "mongo/db/storage/storage_parameters_gen.h"
 #include "mongo/db/transaction_participant.h"
@@ -188,6 +189,7 @@ private:
         auto service = getServiceContext();
         StorageInterface::set(service, std::make_unique<StorageInterfaceRecovery>());
         _storageInterface = static_cast<StorageInterfaceRecovery*>(StorageInterface::get(service));
+        DurableHistoryRegistry::set(service, std::make_unique<DurableHistoryRegistry>());
 
         _createOpCtx();
         _consistencyMarkers = std::make_unique<ReplicationConsistencyMarkersMock>();
diff --git a/src/mongo/db/storage/durable_history_pin.cpp b/src/mongo/db/storage/durable_history_pin.cpp
index 09e4e97bcd4..cc1d0c10c9a 100644
--- a/src/mongo/db/storage/durable_history_pin.cpp
+++ b/src/mongo/db/storage/durable_history_pin.cpp
@@ -101,4 +101,15 @@ void DurableHistoryRegistry::reconcilePins(OperationContext* opCtx) {
     }
 }
 
+void DurableHistoryRegistry::clearPins(OperationContext* opCtx) {
+    StorageEngine* engine = opCtx->getServiceContext()->getStorageEngine();
+    if (!engine->supportsRecoveryTimestamp()) {
+        return;
+    }
+
+    for (auto& pin : _pins) {
+        engine->unpinOldestTimestamp(pin->getName());
+    }
+}
+
 }  // namespace mongo
diff --git a/src/mongo/db/storage/durable_history_pin.h b/src/mongo/db/storage/durable_history_pin.h
index 0a202772579..e50ed11944e 100644
--- a/src/mongo/db/storage/durable_history_pin.h
+++ b/src/mongo/db/storage/durable_history_pin.h
@@ -75,6 +75,11 @@ public:
      */
     void reconcilePins(OperationContext* opCtx);
 
+    /**
+     * Removes all registered pins.
+     */
+    void clearPins(OperationContext* opCtx);
+
 private:
     std::vector<std::unique_ptr<DurableHistoryPin>> _pins;
 };
diff --git a/src/mongo/db/storage/wiredtiger/wiredtiger_kv_engine.cpp b/src/mongo/db/storage/wiredtiger/wiredtiger_kv_engine.cpp
index bdc58ff901a..28a0212f967 100644
--- a/src/mongo/db/storage/wiredtiger/wiredtiger_kv_engine.cpp
+++ b/src/mongo/db/storage/wiredtiger/wiredtiger_kv_engine.cpp
@@ -2072,10 +2072,11 @@ void WiredTigerKVEngine::setOldestTimestamp(Timestamp newOldestTimestamp, bool f
     }
 
     if (force) {
-        // Forcing the oldest timestamp backwards (e.g: eMRC=off rollback) to a value of T
-        // invalidates all snapshots > T. Components that register a pinned timestamp must
-        // synchronize with events that invalidate their snapshots, unpin themselves and either
-        // fail themselves, or reacquire a new snapshot after the rollback event.
+        // The oldest timestamp should only be forced backwards during replication recovery in order
+        // to do rollback via refetch. This refetching process invalidates any timestamped snapshots
+        // until after it completes. Components that register a pinned timestamp must synchronize
+        // with events that invalidate their snapshots, unpin themselves and either fail themselves,
+        // or reacquire a new snapshot after the rollback event.
         //
         // Forcing the oldest timestamp forward -- potentially past a pin request raises the
         // question of whether the pin should be honored. For now we will invariant there is no pin,