{184305174} Fix delfiles to clear ufid-hash before deleting files

bdb/file.c

@@ -6554,6 +6554,7 @@ static int bdb_del_file(bdb_state_type *bdb_state, DB_TXN *tid, char *filename,
                         int *bdberr)
 {
     DB_ENV *dbenv;
+    DB *dbp = NULL;
     char transname[PATH_MAX];
     char *pname = bdb_trans(filename, transname);
     int rc = 0;
@@ -6564,7 +6565,31 @@ static int bdb_del_file(bdb_state_type *bdb_state, DB_TXN *tid, char *filename,
         dbenv = bdb_state->dbenv;
 
     if ((rc = access(pname, F_OK)) == 0) {
-        int rc = dbenv->dbremove(dbenv, tid, filename, NULL, 0);
+        extern int gbl_clear_ufid_on_db_close;
+
+        /* Clear any ufid-hash entries before deleting the file to prevent
+         * "deleted" file descriptors from being held open, especially on
+         * replicants that opened handles during recovery. */
+        if (gbl_clear_ufid_on_db_close) {
+            if ((rc = db_create(&dbp, dbenv, 0)) == 0 &&
+                (rc = dbp->open(dbp, NULL, pname, NULL, DB_BTREE, 0, 0666)) == 0) {
+                /* Clear ufid hash entry for this file */
+                int clear_rc = dbp->clear_ufid_hash(dbp, NULL, 0);
+                if (clear_rc != 0) {
+                    logmsg(LOGMSG_WARN, "bdb_del_file: clear_ufid_hash for %s failed: %d %s\n", filename, clear_rc,
+                           db_strerror(clear_rc));
+                    /* Continue anyway - not fatal */
+                }
+                dbp->close(dbp, DB_NOSYNC);
+                dbp = NULL;
+            } else if (rc != ENOENT) {
+                /* File exists but we couldn't open it - log but continue */
+                logmsg(LOGMSG_WARN, "bdb_del_file: failed to open %s for ufid clearing: %d %s\n", filename, rc,
+                       db_strerror(rc));
+            }
+        }
+
+        rc = dbenv->dbremove(dbenv, tid, filename, NULL, 0);
         if (rc) {
            logmsg(LOGMSG_ERROR, "bdb_del_file: dbremove %s failed: %d %s\n", filename, rc,
                    db_strerror(rc));

tests/delfiles_ufid_leak.test/Makefile

@@ -0,0 +1,5 @@
+ifeq ($(TESTSROOTDIR),)
+  include ../testcase.mk
+else
+  include $(TESTSROOTDIR)/testcase.mk
+endif

tests/delfiles_ufid_leak.test/lrl.options

@@ -0,0 +1,4 @@
+setattr CHECKPOINTTIME 9999999
+setattr MEMPTRICKLEPERCENT 10
+clear_ufid_on_db_close 1
+dtastripe 1

tests/delfiles_ufid_leak.test/runit

@@ -0,0 +1,91 @@
+#!/usr/bin/env bash
+bash -n "$0" | exit 1
+
+# This test validates that the delfiles command clears ufid-hash entries
+# BEFORE deleting files, preventing "deleted" file descriptors from being
+# held open on replicants.
+#
+# Unlike berkdb_file_leaks_rep.test which does TRUNCATE before delfiles,
+# this test runs delfiles WITHOUT any preceding schema change to reproduce
+# the issue where deleted files accumulate on replicants.
+
+source ${TESTSROOTDIR}/tools/cluster_utils.sh
+
+[ -z "${CLUSTER}" ] && { echo "Test requires a cluster"; exit 0; }
+
+dbnm=$1
+
+master=`cdb2sql --tabs ${CDB2_OPTIONS} $dbnm default "select host from comdb2_cluster where is_master='Y'"`
+replicant=`cdb2sql --tabs ${CDB2_OPTIONS} $dbnm default "select comdb2_host()"`
+
+echo "Creating table and inserting data..."
+cdb2sql ${CDB2_OPTIONS} $dbnm default "CREATE TABLE t1 (i integer, b blob)"
+sleep 2
+
+cdb2sql $dbnm --host $master "EXEC PROCEDURE sys.cmd.send('bdb checkpoint')"
+sleep 2
+
+# Insert data to create files
+cdb2sql ${CDB2_OPTIONS} $dbnm default "INSERT INTO t1 VALUES(1, x'DEADBEEF')"
+cdb2sql ${CDB2_OPTIONS} $dbnm default "INSERT INTO t1 VALUES(2, x'CAFEBABE')"
+sleep 2
+
+cdb2sql $dbnm --host $master "EXEC PROCEDURE sys.cmd.send('bdb checkpoint')"
+sleep 2
+
+# TRUNCATE to create an old file version
+echo "Truncating table to create old file version..."
+cdb2sql ${CDB2_OPTIONS} $dbnm default "TRUNCATE TABLE t1"
+sleep 2
+
+cdb2sql $dbnm --host $master "EXEC PROCEDURE sys.cmd.send('bdb checkpoint')"
+sleep 2
+
+# Insert new data in the new file version
+cdb2sql ${CDB2_OPTIONS} $dbnm default "INSERT INTO t1 VALUES(3, x'BAADF00D')"
+sleep 2
+
+cdb2sql $dbnm --host $master "EXEC PROCEDURE sys.cmd.send('bdb checkpoint')"
+sleep 2
+
+echo "Restarting replicant $replicant to force recovery..."
+# This causes the replicant to open DB handles via ufid-hash during recovery
+# These handles will be for BOTH the old and new file versions
+kill_restart_node $replicant 10 1
+cdb2sql $dbnm --host $replicant "SELECT 1"
+if [ $? -ne 0 ]; then
+    echo 'db not up?' >&2
+    exit 1
+fi
+
+sleep 2
+
+echo "Running delfiles WITHOUT any schema change..."
+# KEY: No TRUNCATE/ALTER/DROP before delfiles!
+# This is the scenario where delfiles is run standalone
+# Without the fix, ufid-hash holds open handles to old files
+cdb2sql -tabs $dbnm --host $master "EXEC PROCEDURE sys.cmd.send('delfiles t1')"
+
+sleep 2
+
+# Check all nodes for "deleted" files
+echo "Checking for deleted files held open..."
+for host in $CLUSTER; do
+    dbpid=`ssh $host 'pgrep -a comdb2' | grep $dbnm | awk '{print $1}' | head -1`
+    echo "Checking host $host (pid $dbpid)..."
+
+    deleted_count=`ssh $host "lsof -p $dbpid 2>/dev/null | grep -c deleted"`
+    rc=$?
+    [ $rc -ne 0 ] && deleted_count=0
+
+    if [ "$deleted_count" != "0" ]; then
+        echo "ERROR: Host $host has $deleted_count deleted files held open!" >&2
+        echo "Deleted files:" >&2
+        ssh $host "lsof -p $dbpid 2>/dev/null | grep deleted" >&2
+        exit 1
+    fi
+    echo "  No deleted files on $host"
+done
+
+echo "Success! delfiles properly cleared ufid-hash entries."
+exit 0