Repair two places where SIGTERM exit could leave shared memory state
authorTom Lane <tgl@sss.pgh.pa.us>
Wed, 16 Apr 2008 23:59:51 +0000 (23:59 +0000)
committerTom Lane <tgl@sss.pgh.pa.us>
Wed, 16 Apr 2008 23:59:51 +0000 (23:59 +0000)
corrupted.  (Neither is very important if SIGTERM is used to shut down the
whole database cluster together, but there's a problem if someone tries to
SIGTERM individual backends.)  To do this, introduce new infrastructure
macros PG_ENSURE_ERROR_CLEANUP/PG_END_ENSURE_ERROR_CLEANUP that take care
of transiently pushing an on_shmem_exit cleanup hook.  Also use this method
for createdb cleanup --- that wasn't a shared-memory-corruption problem,
but SIGTERM abort of createdb could leave orphaned files lying around.

Backpatch as far as 8.2.  The shmem corruption cases don't exist in 8.1,
and the createdb usage doesn't seem important enough to risk backpatching
further.

src/backend/access/nbtree/nbtree.c
src/backend/access/nbtree/nbtutils.c
src/backend/access/transam/xlog.c
src/backend/commands/dbcommands.c
src/backend/port/ipc_test.c
src/backend/storage/ipc/ipc.c
src/include/access/nbtree.h
src/include/storage/ipc.h
src/include/utils/elog.h

index 5fd2b0f771e8244c1c4d349fb25b50b940941572..52e0e2ed61c1aba3773bd7695b2f5f3c996e2015 100644 (file)
@@ -23,6 +23,7 @@
 #include "catalog/index.h"
 #include "commands/vacuum.h"
 #include "storage/freespace.h"
+#include "storage/ipc.h"
 #include "storage/lmgr.h"
 #include "utils/memutils.h"
 
@@ -538,21 +539,15 @@ btbulkdelete(PG_FUNCTION_ARGS)
                stats = (IndexBulkDeleteResult *) palloc0(sizeof(IndexBulkDeleteResult));
 
        /* Establish the vacuum cycle ID to use for this scan */
-       PG_TRY();
+       /* The ENSURE stuff ensures we clean up shared memory on failure */
+       PG_ENSURE_ERROR_CLEANUP(_bt_end_vacuum_callback, PointerGetDatum(rel));
        {
                cycleid = _bt_start_vacuum(rel);
 
                btvacuumscan(info, stats, callback, callback_state, cycleid);
-
-               _bt_end_vacuum(rel);
-       }
-       PG_CATCH();
-       {
-               /* Make sure shared memory gets cleaned up */
-               _bt_end_vacuum(rel);
-               PG_RE_THROW();
        }
-       PG_END_TRY();
+       PG_END_ENSURE_ERROR_CLEANUP(_bt_end_vacuum_callback, PointerGetDatum(rel));
+       _bt_end_vacuum(rel);
 
        PG_RETURN_POINTER(stats);
 }
index 80ccf3c3db5c262c32223c2a70e9b3a33ec5238b..b477bdc0216b9461ce32e6548b931d61405a93d9 100644 (file)
@@ -1252,8 +1252,11 @@ _bt_vacuum_cycleid(Relation rel)
 /*
  * _bt_start_vacuum --- assign a cycle ID to a just-starting VACUUM operation
  *
- * Note: the caller must guarantee (via PG_TRY) that it will eventually call
- * _bt_end_vacuum, else we'll permanently leak an array slot.
+ * Note: the caller must guarantee that it will eventually call
+ * _bt_end_vacuum, else we'll permanently leak an array slot.  To ensure
+ * that this happens even in elog(FATAL) scenarios, the appropriate coding
+ * is not just a PG_TRY, but
+ *             PG_ENSURE_ERROR_CLEANUP(_bt_end_vacuum_callback, PointerGetDatum(rel))
  */
 BTCycleId
 _bt_start_vacuum(Relation rel)
@@ -1337,6 +1340,15 @@ _bt_end_vacuum(Relation rel)
        LWLockRelease(BtreeVacuumLock);
 }
 
+/*
+ * _bt_end_vacuum wrapped as an on_shmem_exit callback function
+ */
+void
+_bt_end_vacuum_callback(int code, Datum arg)
+{
+       _bt_end_vacuum((Relation) DatumGetPointer(arg));
+}
+
 /*
  * BTreeShmemSize --- report amount of shared memory space needed
  */
index ba8a3b661620a3254ae72312a9a7c125419b2b20..170c55c5b21b53d52680f206f9df9f9294cde9c0 100644 (file)
@@ -43,6 +43,7 @@
 #include "postmaster/bgwriter.h"
 #include "storage/bufpage.h"
 #include "storage/fd.h"
+#include "storage/ipc.h"
 #include "storage/pmsignal.h"
 #include "storage/procarray.h"
 #include "storage/smgr.h"
@@ -420,11 +421,11 @@ static void writeTimeLineHistory(TimeLineID newTLI, TimeLineID parentTLI,
 static void WriteControlFile(void);
 static void ReadControlFile(void);
 static char *str_time(pg_time_t tnow);
-static void issue_xlog_fsync(void);
-
 #ifdef WAL_DEBUG
 static void xlog_outrec(StringInfo buf, XLogRecord *record);
 #endif
+static void issue_xlog_fsync(void);
+static void pg_start_backup_callback(int code, Datum arg);
 static bool read_backup_label(XLogRecPtr *checkPointLoc,
                                  XLogRecPtr *minRecoveryLoc);
 static void rm_redo_error_callback(void *arg);
@@ -6445,8 +6446,8 @@ pg_start_backup(PG_FUNCTION_ARGS)
        XLogCtl->Insert.forcePageWrites = true;
        LWLockRelease(WALInsertLock);
 
-       /* Use a TRY block to ensure we release forcePageWrites if fail below */
-       PG_TRY();
+       /* Ensure we release forcePageWrites if fail below */
+       PG_ENSURE_ERROR_CLEANUP(pg_start_backup_callback, (Datum) 0);
        {
                /*
                 * Force a CHECKPOINT.  Aside from being necessary to prevent torn
@@ -6518,16 +6519,7 @@ pg_start_backup(PG_FUNCTION_ARGS)
                                         errmsg("could not write file \"%s\": %m",
                                                        BACKUP_LABEL_FILE)));
        }
-       PG_CATCH();
-       {
-               /* Turn off forcePageWrites on failure */
-               LWLockAcquire(WALInsertLock, LW_EXCLUSIVE);
-               XLogCtl->Insert.forcePageWrites = false;
-               LWLockRelease(WALInsertLock);
-
-               PG_RE_THROW();
-       }
-       PG_END_TRY();
+       PG_END_ENSURE_ERROR_CLEANUP(pg_start_backup_callback, (Datum) 0);
 
        /*
         * We're done.  As a convenience, return the starting WAL location.
@@ -6539,6 +6531,16 @@ pg_start_backup(PG_FUNCTION_ARGS)
        PG_RETURN_TEXT_P(result);
 }
 
+/* Error cleanup callback for pg_start_backup */
+static void
+pg_start_backup_callback(int code, Datum arg)
+{
+       /* Turn off forcePageWrites on failure */
+       LWLockAcquire(WALInsertLock, LW_EXCLUSIVE);
+       XLogCtl->Insert.forcePageWrites = false;
+       LWLockRelease(WALInsertLock);
+}
+
 /*
  * pg_stop_backup: finish taking an on-line backup dump
  *
index 0e8b18ddf298541d007d634cfb6106d65e9f4c5c..9e05aba0cbeda465a941786462b740acd3478406 100644 (file)
@@ -41,6 +41,7 @@
 #include "pgstat.h"
 #include "postmaster/bgwriter.h"
 #include "storage/freespace.h"
+#include "storage/ipc.h"
 #include "storage/procarray.h"
 #include "storage/smgr.h"
 #include "utils/acl.h"
 #include "utils/syscache.h"
 
 
+typedef struct
+{
+       Oid                     src_dboid;              /* source (template) DB */
+       Oid                     dest_dboid;             /* DB we are trying to create */
+} createdb_failure_params;
+
 /* non-export function prototypes */
+static void createdb_failure_callback(int code, Datum arg);
 static bool get_db_info(const char *name, LOCKMODE lockmode,
                        Oid *dbIdP, Oid *ownerIdP,
                        int *encodingP, bool *dbIsTemplateP, bool *dbAllowConnP,
@@ -98,6 +106,7 @@ createdb(const CreatedbStmt *stmt)
        int                     encoding = -1;
        int                     dbconnlimit = -1;
        int                     ctype_encoding;
+       createdb_failure_params fparms;
 
        /* Extract options from the statement node tree */
        foreach(option, stmt->options)
@@ -448,12 +457,15 @@ createdb(const CreatedbStmt *stmt)
 
        /*
         * Once we start copying subdirectories, we need to be able to clean 'em
-        * up if we fail.  Establish a TRY block to make sure this happens. (This
+        * up if we fail.  Use an ENSURE block to make sure this happens.  (This
         * is not a 100% solution, because of the possibility of failure during
         * transaction commit after we leave this routine, but it should handle
         * most scenarios.)
         */
-       PG_TRY();
+       fparms.src_dboid = src_dboid;
+       fparms.dest_dboid = dboid;
+       PG_ENSURE_ERROR_CLEANUP(createdb_failure_callback,
+                                                       PointerGetDatum(&fparms));
        {
                /*
                 * Iterate through all tablespaces of the template database, and copy
@@ -564,18 +576,25 @@ createdb(const CreatedbStmt *stmt)
                 */
                database_file_update_needed();
        }
-       PG_CATCH();
-       {
-               /* Release lock on source database before doing recursive remove */
-               UnlockSharedObject(DatabaseRelationId, src_dboid, 0,
-                                                  ShareLock);
+       PG_END_ENSURE_ERROR_CLEANUP(createdb_failure_callback,
+                                                               PointerGetDatum(&fparms));
+}
+
+/* Error cleanup callback for createdb */
+static void
+createdb_failure_callback(int code, Datum arg)
+{
+       createdb_failure_params *fparms = (createdb_failure_params *) DatumGetPointer(arg);
 
-               /* Throw away any successfully copied subdirectories */
-               remove_dbtablespaces(dboid);
+       /*
+        * Release lock on source database before doing recursive remove.
+        * This is not essential but it seems desirable to release the lock
+        * as soon as possible.
+        */
+       UnlockSharedObject(DatabaseRelationId, fparms->src_dboid, 0, ShareLock);
 
-               PG_RE_THROW();
-       }
-       PG_END_TRY();
+       /* Throw away any successfully copied subdirectories */
+       remove_dbtablespaces(fparms->dest_dboid);
 }
 
 
index d7320195a5f01ce6e1e24d387196e564b3682b61..e799e86242f88670a386a5c9b70eb36c77bb3c00 100644 (file)
@@ -58,7 +58,7 @@ char     *DataDir = ".";
 
 static struct ONEXIT
 {
-       void            (*function) (int code, Datum arg);
+       pg_on_exit_callback function;
        Datum           arg;
 }      on_proc_exit_list[MAX_ON_EXITS], on_shmem_exit_list[MAX_ON_EXITS];
 
@@ -85,7 +85,7 @@ shmem_exit(int code)
 }
 
 void
-                       on_shmem_exit(void (*function) (int code, Datum arg), Datum arg)
+on_shmem_exit(pg_on_exit_callback function, Datum arg)
 {
        if (on_shmem_exit_index >= MAX_ON_EXITS)
                elog(FATAL, "out of on_shmem_exit slots");
@@ -114,9 +114,9 @@ ProcessInterrupts(void)
 }
 
 int
-ExceptionalCondition(char *conditionName,
-                                        char *errorType,
-                                        char *fileName,
+ExceptionalCondition(const char *conditionName,
+                                        const char *errorType,
+                                        const char *fileName,
                                         int lineNumber)
 {
        fprintf(stderr, "TRAP: %s(\"%s\", File: \"%s\", Line: %d)\n",
index 998c37bc27508a90bb27a0435d71db8c1891483c..a7c5c72ef61449f43b3948acd80cf5ae3ba3d213 100644 (file)
@@ -56,7 +56,7 @@ bool          proc_exit_inprogress = false;
 
 static struct ONEXIT
 {
-       void            (*function) (int code, Datum arg);
+       pg_on_exit_callback function;
        Datum           arg;
 }      on_proc_exit_list[MAX_ON_EXITS], on_shmem_exit_list[MAX_ON_EXITS];
 
@@ -184,7 +184,7 @@ shmem_exit(int code)
  * ----------------------------------------------------------------
  */
 void
-                       on_proc_exit(void (*function) (int code, Datum arg), Datum arg)
+on_proc_exit(pg_on_exit_callback function, Datum arg)
 {
        if (on_proc_exit_index >= MAX_ON_EXITS)
                ereport(FATAL,
@@ -205,7 +205,7 @@ void
  * ----------------------------------------------------------------
  */
 void
-                       on_shmem_exit(void (*function) (int code, Datum arg), Datum arg)
+on_shmem_exit(pg_on_exit_callback function, Datum arg)
 {
        if (on_shmem_exit_index >= MAX_ON_EXITS)
                ereport(FATAL,
@@ -218,6 +218,24 @@ void
        ++on_shmem_exit_index;
 }
 
+/* ----------------------------------------------------------------
+ *             cancel_shmem_exit
+ *
+ *             this function removes an entry, if present, from the list of
+ *             functions to be invoked by shmem_exit().  For simplicity,
+ *             only the latest entry can be removed.  (We could work harder
+ *             but there is no need for current uses.)
+ * ----------------------------------------------------------------
+ */
+void
+cancel_shmem_exit(pg_on_exit_callback function, Datum arg)
+{
+       if (on_shmem_exit_index > 0 &&
+               on_shmem_exit_list[on_shmem_exit_index - 1].function == function &&
+               on_shmem_exit_list[on_shmem_exit_index - 1].arg == arg)
+               --on_shmem_exit_index;
+}
+
 /* ----------------------------------------------------------------
  *             on_exit_reset
  *
index 26d254f36a1fb77f8f7675935393c729685776d5..de1e51e4fcea0a89d8e9a7192bae691fae77b264 100644 (file)
@@ -574,6 +574,7 @@ extern void _bt_killitems(IndexScanDesc scan, bool haveLock);
 extern BTCycleId _bt_vacuum_cycleid(Relation rel);
 extern BTCycleId _bt_start_vacuum(Relation rel);
 extern void _bt_end_vacuum(Relation rel);
+extern void _bt_end_vacuum_callback(int code, Datum arg);
 extern Size BTreeShmemSize(void);
 extern void BTreeShmemInit(void);
 
index e8ffb713048b4ad3c8598128b01f72e44bb025ba..243c141bfa4fe65c38c1e0818e90a5639a4c1783 100644 (file)
 #ifndef IPC_H
 #define IPC_H
 
+typedef void (*pg_on_exit_callback) (int code, Datum arg);
+
+/*----------
+ * API for handling cleanup that must occur during either ereport(ERROR)
+ * or ereport(FATAL) exits from a block of code.  (Typical examples are
+ * undoing transient changes to shared-memory state.)
+ *
+ *             PG_ENSURE_ERROR_CLEANUP(cleanup_function, arg);
+ *             {
+ *                     ... code that might throw ereport(ERROR) or ereport(FATAL) ...
+ *             }
+ *             PG_END_ENSURE_ERROR_CLEANUP(cleanup_function, arg);
+ *
+ * where the cleanup code is in a function declared per pg_on_exit_callback.
+ * The Datum value "arg" can carry any information the cleanup function
+ * needs.
+ *
+ * This construct ensures that cleanup_function() will be called during
+ * either ERROR or FATAL exits.  It will not be called on successful
+ * exit from the controlled code.  (If you want it to happen then too,
+ * call the function yourself from just after the construct.)
+ *
+ * Note: the macro arguments are multiply evaluated, so avoid side-effects.
+ *----------
+ */
+#define PG_ENSURE_ERROR_CLEANUP(cleanup_function, arg)  \
+       do { \
+               on_shmem_exit(cleanup_function, arg); \
+               PG_TRY()
+
+#define PG_END_ENSURE_ERROR_CLEANUP(cleanup_function, arg)  \
+               cancel_shmem_exit(cleanup_function, arg); \
+               PG_CATCH(); \
+               { \
+                       cancel_shmem_exit(cleanup_function, arg); \
+                       cleanup_function (0, arg); \
+                       PG_RE_THROW(); \
+               } \
+               PG_END_TRY(); \
+       } while (0)
+
 
 /* ipc.c */
 extern bool proc_exit_inprogress;
 
 extern void proc_exit(int code);
 extern void shmem_exit(int code);
-extern void on_proc_exit(void (*function) (int code, Datum arg), Datum arg);
-extern void on_shmem_exit(void (*function) (int code, Datum arg), Datum arg);
+extern void on_proc_exit(pg_on_exit_callback function, Datum arg);
+extern void on_shmem_exit(pg_on_exit_callback function, Datum arg);
+extern void cancel_shmem_exit(pg_on_exit_callback function, Datum arg);
 extern void on_exit_reset(void);
 
 /* ipci.c */
index 7e71618ca93edc8f840a72efd5862a292e86466b..3995d011329bde6203287caae617f1fbab53bfcc 100644 (file)
@@ -198,6 +198,13 @@ extern PGDLLIMPORT ErrorContextCallback *error_context_stack;
  * of levels this will work for.  It's best to keep the error recovery
  * section simple enough that it can't generate any new errors, at least
  * not before popping the error stack.
+ *
+ * Note: an ereport(FATAL) will not be caught by this construct; control will
+ * exit straight through proc_exit().  Therefore, do NOT put any cleanup
+ * of non-process-local resources into the error recovery section, at least
+ * not without taking thought for what will happen during ereport(FATAL).
+ * The PG_ENSURE_ERROR_CLEANUP macros provided by storage/ipc.h may be
+ * helpful in such cases.
  *----------
  */
 #define PG_TRY()  \