1 /*-------------------------------------------------------------------------
4 * Asynchronous notification: NOTIFY, LISTEN, UNLISTEN
6 * Portions Copyright (c) 1996-2000, PostgreSQL, Inc
7 * Portions Copyright (c) 1994, Regents of the University of California
10 * $Header: /cvsroot/pgsql/src/backend/commands/async.c,v 1.69 2000/10/02 19:42:45 petere Exp $
12 *-------------------------------------------------------------------------
15 /*-------------------------------------------------------------------------
16 * New Async Notification Model:
17 * 1. Multiple backends on same machine. Multiple backends listening on
18 * one relation. (Note: "listening on a relation" is not really the
19 * right way to think about it, since the notify names need not have
20 * anything to do with the names of relations actually in the database.
21 * But this terminology is all over the code and docs, and I don't feel
22 * like trying to replace it.)
24 * 2. There is a tuple in relation "pg_listener" for each active LISTEN,
25 * ie, each relname/listenerPID pair. The "notification" field of the
26 * tuple is zero when no NOTIFY is pending for that listener, or the PID
27 * of the originating backend when a cross-backend NOTIFY is pending.
28 * (We skip writing to pg_listener when doing a self-NOTIFY, so the
29 * notification field should never be equal to the listenerPID field.)
31 * 3. The NOTIFY statement itself (routine Async_Notify) just adds the target
32 * relname to a list of outstanding NOTIFY requests. Actual processing
33 * happens if and only if we reach transaction commit. At that time (in
34 * routine AtCommit_Notify) we scan pg_listener for matching relnames.
35 * If the listenerPID in a matching tuple is ours, we just send a notify
36 * message to our own front end. If it is not ours, and "notification"
37 * is not already nonzero, we set notification to our own PID and send a
38 * SIGUSR2 signal to the receiving process (indicated by listenerPID).
39 * BTW: if the signal operation fails, we presume that the listener backend
40 * crashed without removing this tuple, and remove the tuple for it.
42 * 4. Upon receipt of a SIGUSR2 signal, the signal handler can call inbound-
43 * notify processing immediately if this backend is idle (ie, it is
44 * waiting for a frontend command and is not within a transaction block).
45 * Otherwise the handler may only set a flag, which will cause the
46 * processing to occur just before we next go idle.
48 * 5. Inbound-notify processing consists of scanning pg_listener for tuples
49 * matching our own listenerPID and having nonzero notification fields.
50 * For each such tuple, we send a message to our frontend and clear the
51 * notification field. BTW: this routine has to start/commit its own
52 * transaction, since by assumption it is only called from outside any
55 * Although we grab AccessExclusiveLock on pg_listener for any operation,
56 * the lock is never held very long, so it shouldn't cause too much of
57 * a performance problem.
59 * An application that listens on the same relname it notifies will get
60 * NOTIFY messages for its own NOTIFYs. These can be ignored, if not useful,
61 * by comparing be_pid in the NOTIFY message to the application's own backend's
62 * PID. (As of FE/BE protocol 2.0, the backend's PID is provided to the
63 * frontend during startup.) The above design guarantees that notifies from
64 * other backends will never be missed by ignoring self-notifies. Note,
65 * however, that we do *not* guarantee that a separate frontend message will
66 * be sent for every outside NOTIFY. Since there is only room for one
67 * originating PID in pg_listener, outside notifies occurring at about the
68 * same time may be collapsed into a single message bearing the PID of the
69 * first outside backend to perform the NOTIFY.
70 *-------------------------------------------------------------------------
78 #include <sys/types.h>
79 #include <netinet/in.h>
81 #include "access/heapam.h"
82 #include "catalog/catname.h"
83 #include "catalog/indexing.h"
84 #include "catalog/pg_listener.h"
85 #include "commands/async.h"
86 #include "lib/dllist.h"
87 #include "libpq/libpq.h"
88 #include "libpq/pqformat.h"
89 #include "miscadmin.h"
90 #include "tcop/dest.h"
91 #include "utils/fmgroids.h"
92 #include "utils/ps_status.h"
93 #include "utils/syscache.h"
96 /* stuff that we really ought not be touching directly :-( */
97 extern TransactionState CurrentTransactionState;
98 extern CommandDest whereToSendOutput;
101 * State for outbound notifies consists of a list of all relnames NOTIFYed
102 * in the current transaction. We do not actually perform a NOTIFY until
103 * and unless the transaction commits. pendingNotifies is NULL if no
104 * NOTIFYs have been done in the current transaction.
106 static Dllist *pendingNotifies = NULL;
109 * State for inbound notifies consists of two flags: one saying whether
110 * the signal handler is currently allowed to call ProcessIncomingNotify
111 * directly, and one saying whether the signal has occurred but the handler
112 * was not allowed to call ProcessIncomingNotify at the time.
114 * NB: the "volatile" on these declarations is critical! If your compiler
115 * does not grok "volatile", you'd be best advised to compile this file
116 * with all optimization turned off.
118 static volatile int notifyInterruptEnabled = 0;
119 static volatile int notifyInterruptOccurred = 0;
121 /* True if we've registered an on_shmem_exit cleanup (or at least tried to). */
122 static int unlistenExitRegistered = 0;
125 static void Async_UnlistenAll(void);
126 static void Async_UnlistenOnExit(void);
127 static void ProcessIncomingNotify(void);
128 static void NotifyMyFrontEnd(char *relname, int32 listenerPID);
129 static int AsyncExistsPendingNotify(char *relname);
130 static void ClearPendingNotifies(void);
132 bool Trace_notify = false;
136 *--------------------------------------------------------------
139 * This is executed by the SQL notify command.
141 * Adds the relation to the list of pending notifies.
142 * Actual notification happens during transaction commit.
143 * ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
148 *--------------------------------------------------------------
151 Async_Notify(char *relname)
156 elog(DEBUG, "Async_Notify: %s", relname);
158 if (!pendingNotifies)
159 pendingNotifies = DLNewList();
160 /* no point in making duplicate entries in the list ... */
161 if (!AsyncExistsPendingNotify(relname))
164 * We allocate list memory from the global malloc pool to ensure
165 * that it will live until we want to use it. This is probably
166 * not necessary any longer, since we will use it before the end
167 * of the transaction. DLList only knows how to use malloc()
168 * anyway, but we could probably palloc() the strings...
170 notifyName = strdup(relname);
171 DLAddHead(pendingNotifies, DLNewElem(notifyName));
176 *--------------------------------------------------------------
179 * This is executed by the SQL listen command.
181 * Register a backend (identified by its Unix PID) as listening
182 * on the specified relation.
188 * pg_listener is updated.
190 *--------------------------------------------------------------
193 Async_Listen(char *relname, int pid)
199 Datum values[Natts_pg_listener];
200 char nulls[Natts_pg_listener];
205 elog(DEBUG, "Async_Listen: %s", relname);
207 lRel = heap_openr(ListenerRelationName, AccessExclusiveLock);
208 tdesc = RelationGetDescr(lRel);
210 /* Detect whether we are already listening on this relname */
211 tuple = SearchSysCacheTuple(LISTENREL, Int32GetDatum(pid),
212 PointerGetDatum(relname),
216 /* No need to scan the rest of the table */
217 heap_close(lRel, AccessExclusiveLock);
218 elog(NOTICE, "Async_Listen: We are already listening on %s", relname);
223 * OK to insert a new tuple
226 for (i = 0; i < Natts_pg_listener; i++)
229 values[i] = PointerGetDatum(NULL);
233 values[i++] = (Datum) relname;
234 values[i++] = (Datum) pid;
235 values[i++] = (Datum) 0; /* no notifies pending */
237 tupDesc = lRel->rd_att;
238 newtup = heap_formtuple(tupDesc, values, nulls);
239 heap_insert(lRel, newtup);
240 if (RelationGetForm(lRel)->relhasindex)
242 Relation idescs[Num_pg_listener_indices];
244 CatalogOpenIndices(Num_pg_listener_indices, Name_pg_listener_indices, idescs);
245 CatalogIndexInsert(idescs, Num_pg_listener_indices, lRel, newtup);
246 CatalogCloseIndices(Num_pg_listener_indices, idescs);
249 heap_freetuple(newtup);
251 heap_close(lRel, AccessExclusiveLock);
254 * now that we are listening, make sure we will unlisten before dying.
256 if (!unlistenExitRegistered)
258 if (on_shmem_exit(Async_UnlistenOnExit, 0) < 0)
259 elog(NOTICE, "Async_Listen: out of shmem_exit slots");
260 unlistenExitRegistered = 1;
265 *--------------------------------------------------------------
268 * This is executed by the SQL unlisten command.
270 * Remove the backend from the list of listening backends
271 * for the specified relation.
277 * pg_listener is updated.
279 *--------------------------------------------------------------
282 Async_Unlisten(char *relname, int pid)
287 /* Handle specially the `unlisten "*"' command */
288 if ((!relname) || (*relname == '\0') || (strcmp(relname, "*") == 0))
295 elog(DEBUG, "Async_Unlisten %s", relname);
297 lRel = heap_openr(ListenerRelationName, AccessExclusiveLock);
298 /* Note we assume there can be only one matching tuple. */
299 lTuple = SearchSysCacheTuple(LISTENREL, Int32GetDatum(pid),
300 PointerGetDatum(relname),
303 heap_delete(lRel, &lTuple->t_self, NULL);
304 heap_close(lRel, AccessExclusiveLock);
307 * We do not complain about unlistening something not being listened;
313 *--------------------------------------------------------------
316 * Unlisten all relations for this backend.
318 * This is invoked by UNLISTEN "*" command, and also at backend exit.
324 * pg_listener is updated.
326 *--------------------------------------------------------------
338 elog(DEBUG, "Async_UnlistenAll");
340 lRel = heap_openr(ListenerRelationName, AccessExclusiveLock);
341 tdesc = RelationGetDescr(lRel);
343 /* Find and delete all entries with my listenerPID */
344 ScanKeyEntryInitialize(&key[0], 0,
345 Anum_pg_listener_pid,
347 Int32GetDatum(MyProcPid));
348 sRel = heap_beginscan(lRel, 0, SnapshotNow, 1, key);
350 while (HeapTupleIsValid(lTuple = heap_getnext(sRel, 0)))
351 heap_delete(lRel, &lTuple->t_self, NULL);
354 heap_close(lRel, AccessExclusiveLock);
358 *--------------------------------------------------------------
359 * Async_UnlistenOnExit
361 * Clean up the pg_listener table at backend exit.
363 * This is executed if we have done any LISTENs in this backend.
364 * It might not be necessary anymore, if the user UNLISTENed everything,
365 * but we don't try to detect that case.
371 * pg_listener is updated if necessary.
373 *--------------------------------------------------------------
376 Async_UnlistenOnExit(void)
380 * We need to start/commit a transaction for the unlisten, but if
381 * there is already an active transaction we had better abort that one
382 * first. Otherwise we'd end up committing changes that probably
383 * ought to be discarded.
385 AbortOutOfAnyTransaction();
386 /* Now we can do the unlisten */
387 StartTransactionCommand();
389 CommitTransactionCommand();
393 *--------------------------------------------------------------
396 * This is called at transaction commit.
398 * If there are outbound notify requests in the pendingNotifies list,
399 * scan pg_listener for matching tuples, and either signal the other
400 * backend or send a message to our own frontend.
402 * NOTE: we are still inside the current transaction, therefore can
403 * piggyback on its committing of changes.
409 * Tuples in pg_listener that have matching relnames and other peoples'
410 * listenerPIDs are updated with a nonzero notification field.
412 *--------------------------------------------------------------
423 value[Natts_pg_listener];
424 char repl[Natts_pg_listener],
425 nulls[Natts_pg_listener];
430 if (!pendingNotifies)
431 return; /* no NOTIFY statements in this
435 * NOTIFY is disabled if not normal processing mode. This test used to
436 * be in xact.c, but it seems cleaner to do it here.
438 if (!IsNormalProcessingMode())
440 ClearPendingNotifies();
445 elog(DEBUG, "AtCommit_Notify");
447 lRel = heap_openr(ListenerRelationName, AccessExclusiveLock);
448 tdesc = RelationGetDescr(lRel);
449 sRel = heap_beginscan(lRel, 0, SnapshotNow, 0, (ScanKey) NULL);
451 /* preset data to update notify column to MyProcPid */
452 nulls[0] = nulls[1] = nulls[2] = ' ';
453 repl[0] = repl[1] = repl[2] = ' ';
454 repl[Anum_pg_listener_notify - 1] = 'r';
455 value[0] = value[1] = value[2] = (Datum) 0;
456 value[Anum_pg_listener_notify - 1] = Int32GetDatum(MyProcPid);
458 while (HeapTupleIsValid(lTuple = heap_getnext(sRel, 0)))
460 d = heap_getattr(lTuple, Anum_pg_listener_relname, tdesc, &isnull);
461 relname = (char *) DatumGetPointer(d);
463 if (AsyncExistsPendingNotify(relname))
465 d = heap_getattr(lTuple, Anum_pg_listener_pid, tdesc, &isnull);
466 listenerPID = DatumGetInt32(d);
468 if (listenerPID == MyProcPid)
472 * Self-notify: no need to bother with table update.
473 * Indeed, we *must not* clear the notification field in
474 * this path, or we could lose an outside notify, which'd
475 * be bad for applications that ignore self-notify
480 elog(DEBUG, "AtCommit_Notify: notifying self");
482 NotifyMyFrontEnd(relname, listenerPID);
487 elog(DEBUG, "AtCommit_Notify: notifying pid %d", listenerPID);
490 * If someone has already notified this listener, we don't
491 * bother modifying the table, but we do still send a
492 * SIGUSR2 signal, just in case that backend missed the
493 * earlier signal for some reason. It's OK to send the
494 * signal first, because the other guy can't read
495 * pg_listener until we unlock it.
497 if (kill(listenerPID, SIGUSR2) < 0)
501 * Get rid of pg_listener entry if it refers to a PID
502 * that no longer exists. Presumably, that backend
503 * crashed without deleting its pg_listener entries.
504 * This code used to only delete the entry if
505 * errno==ESRCH, but as far as I can see we should
506 * just do it for any failure (certainly at least for
509 heap_delete(lRel, &lTuple->t_self, NULL);
513 d = heap_getattr(lTuple, Anum_pg_listener_notify,
515 if (DatumGetInt32(d) == 0)
517 rTuple = heap_modifytuple(lTuple, lRel,
519 heap_update(lRel, &lTuple->t_self, rTuple, NULL);
520 if (RelationGetForm(lRel)->relhasindex)
522 Relation idescs[Num_pg_listener_indices];
524 CatalogOpenIndices(Num_pg_listener_indices, Name_pg_listener_indices, idescs);
525 CatalogIndexInsert(idescs, Num_pg_listener_indices, lRel, rTuple);
526 CatalogCloseIndices(Num_pg_listener_indices, idescs);
537 * We do NOT release the lock on pg_listener here; we need to hold it
538 * until end of transaction (which is about to happen, anyway) to
539 * ensure that notified backends see our tuple updates when they look.
540 * Else they might disregard the signal, which would make the
541 * application programmer very unhappy.
543 heap_close(lRel, NoLock);
545 ClearPendingNotifies();
548 elog(DEBUG, "AtCommit_Notify: done");
552 *--------------------------------------------------------------
555 * This is called at transaction abort.
557 * Gets rid of pending outbound notifies that we would have executed
558 * if the transaction got committed.
563 *--------------------------------------------------------------
568 ClearPendingNotifies();
572 *--------------------------------------------------------------
573 * Async_NotifyHandler
575 * This is the signal handler for SIGUSR2.
577 * If we are idle (notifyInterruptEnabled is set), we can safely invoke
578 * ProcessIncomingNotify directly. Otherwise, just set a flag
586 *--------------------------------------------------------------
590 Async_NotifyHandler(SIGNAL_ARGS)
594 * Note: this is a SIGNAL HANDLER. You must be very wary what you do
595 * here. Some helpful soul had this routine sprinkled with TPRINTFs,
596 * which would likely lead to corruption of stdio buffers if they were
600 if (notifyInterruptEnabled)
604 * I'm not sure whether some flavors of Unix might allow another
605 * SIGUSR2 occurrence to recursively interrupt this routine. To
606 * cope with the possibility, we do the same sort of dance that
607 * EnableNotifyInterrupt must do --- see that routine for
610 notifyInterruptEnabled = 0; /* disable any recursive signal */
611 notifyInterruptOccurred = 1; /* do at least one iteration */
614 notifyInterruptEnabled = 1;
615 if (!notifyInterruptOccurred)
617 notifyInterruptEnabled = 0;
618 if (notifyInterruptOccurred)
620 /* Here, it is finally safe to do stuff. */
622 elog(DEBUG, "Async_NotifyHandler: perform async notify");
624 ProcessIncomingNotify();
627 elog(DEBUG, "Async_NotifyHandler: done");
635 * In this path it is NOT SAFE to do much of anything, except
638 notifyInterruptOccurred = 1;
643 * --------------------------------------------------------------
644 * EnableNotifyInterrupt
646 * This is called by the PostgresMain main loop just before waiting
647 * for a frontend command. If we are truly idle (ie, *not* inside
648 * a transaction block), then process any pending inbound notifies,
649 * and enable the signal handler to process future notifies directly.
651 * NOTE: the signal handler starts out disabled, and stays so until
652 * PostgresMain calls this the first time.
653 * --------------------------------------------------------------
657 EnableNotifyInterrupt(void)
659 if (CurrentTransactionState->blockState != TRANS_DEFAULT)
660 return; /* not really idle */
663 * This code is tricky because we are communicating with a signal
664 * handler that could interrupt us at any point. If we just checked
665 * notifyInterruptOccurred and then set notifyInterruptEnabled, we
666 * could fail to respond promptly to a signal that happens in between
667 * those two steps. (A very small time window, perhaps, but Murphy's
668 * Law says you can hit it...) Instead, we first set the enable flag,
669 * then test the occurred flag. If we see an unserviced interrupt has
670 * occurred, we re-clear the enable flag before going off to do the
671 * service work. (That prevents re-entrant invocation of
672 * ProcessIncomingNotify() if another interrupt occurs.) If an
673 * interrupt comes in between the setting and clearing of
674 * notifyInterruptEnabled, then it will have done the service work and
675 * left notifyInterruptOccurred zero, so we have to check again after
676 * clearing enable. The whole thing has to be in a loop in case
677 * another interrupt occurs while we're servicing the first. Once we
678 * get out of the loop, enable is set and we know there is no
679 * unserviced interrupt.
681 * NB: an overenthusiastic optimizing compiler could easily break this
682 * code. Hopefully, they all understand what "volatile" means these
687 notifyInterruptEnabled = 1;
688 if (!notifyInterruptOccurred)
690 notifyInterruptEnabled = 0;
691 if (notifyInterruptOccurred)
694 elog(DEBUG, "EnableNotifyInterrupt: perform async notify");
696 ProcessIncomingNotify();
699 elog(DEBUG, "EnableNotifyInterrupt: done");
705 * --------------------------------------------------------------
706 * DisableNotifyInterrupt
708 * This is called by the PostgresMain main loop just after receiving
709 * a frontend command. Signal handler execution of inbound notifies
710 * is disabled until the next EnableNotifyInterrupt call.
711 * --------------------------------------------------------------
715 DisableNotifyInterrupt(void)
717 notifyInterruptEnabled = 0;
721 * --------------------------------------------------------------
722 * ProcessIncomingNotify
724 * Deal with arriving NOTIFYs from other backends.
725 * This is called either directly from the SIGUSR2 signal handler,
726 * or the next time control reaches the outer idle loop.
727 * Scan pg_listener for arriving notifies, report them to my front end,
728 * and clear the notification field in pg_listener until next time.
730 * NOTE: since we are outside any transaction, we must create our own.
735 * --------------------------------------------------------------
738 ProcessIncomingNotify(void)
747 value[Natts_pg_listener];
748 char repl[Natts_pg_listener],
749 nulls[Natts_pg_listener];
755 elog(DEBUG, "ProcessIncomingNotify");
757 set_ps_display("async_notify");
759 notifyInterruptOccurred = 0;
761 StartTransactionCommand();
763 lRel = heap_openr(ListenerRelationName, AccessExclusiveLock);
764 tdesc = RelationGetDescr(lRel);
766 /* Scan only entries with my listenerPID */
767 ScanKeyEntryInitialize(&key[0], 0,
768 Anum_pg_listener_pid,
770 Int32GetDatum(MyProcPid));
771 sRel = heap_beginscan(lRel, 0, SnapshotNow, 1, key);
773 /* Prepare data for rewriting 0 into notification field */
774 nulls[0] = nulls[1] = nulls[2] = ' ';
775 repl[0] = repl[1] = repl[2] = ' ';
776 repl[Anum_pg_listener_notify - 1] = 'r';
777 value[0] = value[1] = value[2] = (Datum) 0;
778 value[Anum_pg_listener_notify - 1] = Int32GetDatum(0);
780 while (HeapTupleIsValid(lTuple = heap_getnext(sRel, 0)))
782 d = heap_getattr(lTuple, Anum_pg_listener_notify, tdesc, &isnull);
783 sourcePID = DatumGetInt32(d);
786 d = heap_getattr(lTuple, Anum_pg_listener_relname, tdesc, &isnull);
787 relname = (char *) DatumGetPointer(d);
788 /* Notify the frontend */
791 elog(DEBUG, "ProcessIncomingNotify: received %s from %d",
792 relname, (int) sourcePID);
794 NotifyMyFrontEnd(relname, sourcePID);
795 /* Rewrite the tuple with 0 in notification column */
796 rTuple = heap_modifytuple(lTuple, lRel, value, nulls, repl);
797 heap_update(lRel, &lTuple->t_self, rTuple, NULL);
798 if (RelationGetForm(lRel)->relhasindex)
800 Relation idescs[Num_pg_listener_indices];
802 CatalogOpenIndices(Num_pg_listener_indices, Name_pg_listener_indices, idescs);
803 CatalogIndexInsert(idescs, Num_pg_listener_indices, lRel, rTuple);
804 CatalogCloseIndices(Num_pg_listener_indices, idescs);
811 * We do NOT release the lock on pg_listener here; we need to hold it
812 * until end of transaction (which is about to happen, anyway) to
813 * ensure that other backends see our tuple updates when they look.
814 * Otherwise, a transaction started after this one might mistakenly
815 * think it doesn't need to send this backend a new NOTIFY.
817 heap_close(lRel, NoLock);
819 CommitTransactionCommand();
822 * Must flush the notify messages to ensure frontend gets them
827 set_ps_display("idle");
830 elog(DEBUG, "ProcessIncomingNotify: done");
833 /* Send NOTIFY message to my front end. */
836 NotifyMyFrontEnd(char *relname, int32 listenerPID)
838 if (whereToSendOutput == Remote)
842 pq_beginmessage(&buf);
843 pq_sendbyte(&buf, 'A');
844 pq_sendint(&buf, listenerPID, sizeof(int32));
845 pq_sendstring(&buf, relname);
849 * NOTE: we do not do pq_flush() here. For a self-notify, it will
850 * happen at the end of the transaction, and for incoming notifies
851 * ProcessIncomingNotify will do it after finding all the
856 elog(NOTICE, "NOTIFY for %s", relname);
859 /* Does pendingNotifies include the given relname?
861 * NB: not called unless pendingNotifies != NULL.
865 AsyncExistsPendingNotify(char *relname)
869 for (p = DLGetHead(pendingNotifies);
873 /* Use NAMEDATALEN for relname comparison. DZ - 26-08-1996 */
874 if (!strncmp((const char *) DLE_VAL(p), relname, NAMEDATALEN))
881 /* Clear the pendingNotifies list. */
884 ClearPendingNotifies()
892 * Since the referenced strings are malloc'd, we have to scan the
893 * list and delete them individually. If we used palloc for the
894 * strings then we could just do DLFreeList to get rid of both the
895 * list nodes and the list base...
897 while ((p = DLRemHead(pendingNotifies)) != NULL)
902 DLFreeList(pendingNotifies);
903 pendingNotifies = NULL;