1 /*-------------------------------------------------------------------------
4 * Asynchronous notification: NOTIFY, LISTEN, UNLISTEN
6 * Portions Copyright (c) 1996-2013, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
10 * src/backend/commands/async.c
12 *-------------------------------------------------------------------------
15 /*-------------------------------------------------------------------------
16 * Async Notification Model as of 9.0:
18 * 1. Multiple backends on same machine. Multiple backends listening on
19 * several channels. (Channels are also called "conditions" in other
22 * 2. There is one central queue in disk-based storage (directory pg_notify/),
23 * with actively-used pages mapped into shared memory by the slru.c module.
24 * All notification messages are placed in the queue and later read out
25 * by listening backends.
27 * There is no central knowledge of which backend listens on which channel;
28 * every backend has its own list of interesting channels.
30 * Although there is only one queue, notifications are treated as being
31 * database-local; this is done by including the sender's database OID
32 * in each notification message. Listening backends ignore messages
33 * that don't match their database OID. This is important because it
34 * ensures senders and receivers have the same database encoding and won't
35 * misinterpret non-ASCII text in the channel name or payload string.
37 * Since notifications are not expected to survive database crashes,
38 * we can simply clean out the pg_notify data at any reboot, and there
39 * is no need for WAL support or fsync'ing.
41 * 3. Every backend that is listening on at least one channel registers by
42 * entering its PID into the array in AsyncQueueControl. It then scans all
43 * incoming notifications in the central queue and first compares the
44 * database OID of the notification with its own database OID and then
45 * compares the notified channel with the list of channels that it listens
46 * to. In case there is a match it delivers the notification event to its
47 * frontend. Non-matching events are simply skipped.
49 * 4. The NOTIFY statement (routine Async_Notify) stores the notification in
50 * a backend-local list which will not be processed until transaction end.
52 * Duplicate notifications from the same transaction are sent out as one
53 * notification only. This is done to save work when for example a trigger
54 * on a 2 million row table fires a notification for each row that has been
55 * changed. If the application needs to receive every single notification
56 * that has been sent, it can easily add some unique string into the extra
59 * When the transaction is ready to commit, PreCommit_Notify() adds the
60 * pending notifications to the head of the queue. The head pointer of the
61 * queue always points to the next free position and a position is just a
62 * page number and the offset in that page. This is done before marking the
63 * transaction as committed in clog. If we run into problems writing the
64 * notifications, we can still call elog(ERROR, ...) and the transaction
67 * Once we have put all of the notifications into the queue, we return to
68 * CommitTransaction() which will then do the actual transaction commit.
70 * After commit we are called another time (AtCommit_Notify()). Here we
71 * make the actual updates to the effective listen state (listenChannels).
73 * Finally, after we are out of the transaction altogether, we check if
74 * we need to signal listening backends. In SignalBackends() we scan the
75 * list of listening backends and send a PROCSIG_NOTIFY_INTERRUPT signal
76 * to every listening backend (we don't know which backend is listening on
77 * which channel so we must signal them all). We can exclude backends that
78 * are already up to date, though. We don't bother with a self-signal
79 * either, but just process the queue directly.
81 * 5. Upon receipt of a PROCSIG_NOTIFY_INTERRUPT signal, the signal handler
82 * can call inbound-notify processing immediately if this backend is idle
83 * (ie, it is waiting for a frontend command and is not within a transaction
84 * block). Otherwise the handler may only set a flag, which will cause the
85 * processing to occur just before we next go idle.
87 * Inbound-notify processing consists of reading all of the notifications
88 * that have arrived since scanning last time. We read every notification
89 * until we reach either a notification from an uncommitted transaction or
90 * the head pointer's position. Then we check if we were the laziest
91 * backend: if our pointer is set to the same position as the global tail
92 * pointer is set, then we move the global tail pointer ahead to where the
93 * second-laziest backend is (in general, we take the MIN of the current
94 * head position and all active backends' new tail pointers). Whenever we
95 * move the global tail pointer we also truncate now-unused pages (i.e.,
96 * delete files in pg_notify/ that are no longer used).
98 * An application that listens on the same channel it notifies will get
99 * NOTIFY messages for its own NOTIFYs. These can be ignored, if not useful,
100 * by comparing be_pid in the NOTIFY message to the application's own backend's
101 * PID. (As of FE/BE protocol 2.0, the backend's PID is provided to the
102 * frontend during startup.) The above design guarantees that notifies from
103 * other backends will never be missed by ignoring self-notifies.
105 * The amount of shared memory used for notify management (NUM_ASYNC_BUFFERS)
106 * can be varied without affecting anything but performance. The maximum
107 * amount of notification data that can be queued at one time is determined
108 * by slru.c's wraparound limit; see QUEUE_MAX_PAGE below.
109 *-------------------------------------------------------------------------
112 #include "postgres.h"
118 #include "access/slru.h"
119 #include "access/transam.h"
120 #include "access/xact.h"
121 #include "catalog/pg_database.h"
122 #include "commands/async.h"
124 #include "libpq/libpq.h"
125 #include "libpq/pqformat.h"
126 #include "miscadmin.h"
127 #include "storage/ipc.h"
128 #include "storage/lmgr.h"
129 #include "storage/procsignal.h"
130 #include "storage/sinval.h"
131 #include "tcop/tcopprot.h"
132 #include "utils/builtins.h"
133 #include "utils/memutils.h"
134 #include "utils/ps_status.h"
135 #include "utils/timestamp.h"
139 * Maximum size of a NOTIFY payload, including terminating NULL. This
140 * must be kept small enough so that a notification message fits on one
141 * SLRU page. The magic fudge factor here is noncritical as long as it's
142 * more than AsyncQueueEntryEmptySize --- we make it significantly bigger
143 * than that, so changes in that data structure won't affect user-visible
146 #define NOTIFY_PAYLOAD_MAX_LENGTH (BLCKSZ - NAMEDATALEN - 128)
149 * Struct representing an entry in the global notify queue
151 * This struct declaration has the maximal length, but in a real queue entry
152 * the data area is only big enough for the actual channel and payload strings
153 * (each null-terminated). AsyncQueueEntryEmptySize is the minimum possible
154 * entry size, if both channel and payload strings are empty (but note it
155 * doesn't include alignment padding).
157 * The "length" field should always be rounded up to the next QUEUEALIGN
158 * multiple so that all fields are properly aligned.
160 typedef struct AsyncQueueEntry
162 int length; /* total allocated length of entry */
163 Oid dboid; /* sender's database OID */
164 TransactionId xid; /* sender's XID */
165 int32 srcPid; /* sender's PID */
166 char data[NAMEDATALEN + NOTIFY_PAYLOAD_MAX_LENGTH];
169 /* Currently, no field of AsyncQueueEntry requires more than int alignment */
170 #define QUEUEALIGN(len) INTALIGN(len)
172 #define AsyncQueueEntryEmptySize (offsetof(AsyncQueueEntry, data) + 2)
175 * Struct describing a queue position, and assorted macros for working with it
177 typedef struct QueuePosition
179 int page; /* SLRU page number */
180 int offset; /* byte offset within page */
183 #define QUEUE_POS_PAGE(x) ((x).page)
184 #define QUEUE_POS_OFFSET(x) ((x).offset)
186 #define SET_QUEUE_POS(x,y,z) \
192 #define QUEUE_POS_EQUAL(x,y) \
193 ((x).page == (y).page && (x).offset == (y).offset)
195 /* choose logically smaller QueuePosition */
196 #define QUEUE_POS_MIN(x,y) \
197 (asyncQueuePagePrecedes((x).page, (y).page) ? (x) : \
198 (x).page != (y).page ? (y) : \
199 (x).offset < (y).offset ? (x) : (y))
202 * Struct describing a listening backend's status
204 typedef struct QueueBackendStatus
206 int32 pid; /* either a PID or InvalidPid */
207 QueuePosition pos; /* backend has read queue up to here */
208 } QueueBackendStatus;
211 * Shared memory state for LISTEN/NOTIFY (excluding its SLRU stuff)
213 * The AsyncQueueControl structure is protected by the AsyncQueueLock.
215 * When holding the lock in SHARED mode, backends may only inspect their own
216 * entries as well as the head and tail pointers. Consequently we can allow a
217 * backend to update its own record while holding only SHARED lock (since no
218 * other backend will inspect it).
220 * When holding the lock in EXCLUSIVE mode, backends can inspect the entries
221 * of other backends and also change the head and tail pointers.
223 * In order to avoid deadlocks, whenever we need both locks, we always first
224 * get AsyncQueueLock and then AsyncCtlLock.
226 * Each backend uses the backend[] array entry with index equal to its
227 * BackendId (which can range from 1 to MaxBackends). We rely on this to make
228 * SendProcSignal fast.
230 typedef struct AsyncQueueControl
232 QueuePosition head; /* head points to the next free location */
233 QueuePosition tail; /* the global tail is equivalent to the tail
234 * of the "slowest" backend */
235 TimestampTz lastQueueFillWarn; /* time of last queue-full msg */
236 QueueBackendStatus backend[1]; /* actually of length MaxBackends+1 */
237 /* DO NOT ADD FURTHER STRUCT MEMBERS HERE */
240 static AsyncQueueControl *asyncQueueControl;
242 #define QUEUE_HEAD (asyncQueueControl->head)
243 #define QUEUE_TAIL (asyncQueueControl->tail)
244 #define QUEUE_BACKEND_PID(i) (asyncQueueControl->backend[i].pid)
245 #define QUEUE_BACKEND_POS(i) (asyncQueueControl->backend[i].pos)
248 * The SLRU buffer area through which we access the notification queue
250 static SlruCtlData AsyncCtlData;
252 #define AsyncCtl (&AsyncCtlData)
253 #define QUEUE_PAGESIZE BLCKSZ
254 #define QUEUE_FULL_WARN_INTERVAL 5000 /* warn at most once every 5s */
257 * slru.c currently assumes that all filenames are four characters of hex
258 * digits. That means that we can use segments 0000 through FFFF.
259 * Each segment contains SLRU_PAGES_PER_SEGMENT pages which gives us
260 * the pages from 0 to SLRU_PAGES_PER_SEGMENT * 0x10000 - 1.
262 * It's of course possible to enhance slru.c, but this gives us so much
263 * space already that it doesn't seem worth the trouble.
265 * The most data we can have in the queue at a time is QUEUE_MAX_PAGE/2
266 * pages, because more than that would confuse slru.c into thinking there
267 * was a wraparound condition. With the default BLCKSZ this means there
268 * can be up to 8GB of queued-and-not-read data.
270 * Note: it's possible to redefine QUEUE_MAX_PAGE with a smaller multiple of
271 * SLRU_PAGES_PER_SEGMENT, for easier testing of queue-full behaviour.
273 #define QUEUE_MAX_PAGE (SLRU_PAGES_PER_SEGMENT * 0x10000 - 1)
276 * listenChannels identifies the channels we are actually listening to
277 * (ie, have committed a LISTEN on). It is a simple list of channel names,
278 * allocated in TopMemoryContext.
280 static List *listenChannels = NIL; /* list of C strings */
283 * State for pending LISTEN/UNLISTEN actions consists of an ordered list of
284 * all actions requested in the current transaction. As explained above,
285 * we don't actually change listenChannels until we reach transaction commit.
287 * The list is kept in CurTransactionContext. In subtransactions, each
288 * subtransaction has its own list in its own CurTransactionContext, but
289 * successful subtransactions attach their lists to their parent's list.
290 * Failed subtransactions simply discard their lists.
301 ListenActionKind action;
302 char channel[1]; /* actually, as long as needed */
305 static List *pendingActions = NIL; /* list of ListenAction */
307 static List *upperPendingActions = NIL; /* list of upper-xact lists */
310 * State for outbound notifies consists of a list of all channels+payloads
311 * NOTIFYed in the current transaction. We do not actually perform a NOTIFY
312 * until and unless the transaction commits. pendingNotifies is NIL if no
313 * NOTIFYs have been done in the current transaction.
315 * The list is kept in CurTransactionContext. In subtransactions, each
316 * subtransaction has its own list in its own CurTransactionContext, but
317 * successful subtransactions attach their lists to their parent's list.
318 * Failed subtransactions simply discard their lists.
320 * Note: the action and notify lists do not interact within a transaction.
321 * In particular, if a transaction does NOTIFY and then LISTEN on the same
322 * condition name, it will get a self-notify at commit. This is a bit odd
323 * but is consistent with our historical behavior.
325 typedef struct Notification
327 char *channel; /* channel name */
328 char *payload; /* payload string (can be empty) */
331 static List *pendingNotifies = NIL; /* list of Notifications */
333 static List *upperPendingNotifies = NIL; /* list of upper-xact lists */
336 * State for inbound notifications consists of two flags: one saying whether
337 * the signal handler is currently allowed to call ProcessIncomingNotify
338 * directly, and one saying whether the signal has occurred but the handler
339 * was not allowed to call ProcessIncomingNotify at the time.
341 * NB: the "volatile" on these declarations is critical! If your compiler
342 * does not grok "volatile", you'd be best advised to compile this file
343 * with all optimization turned off.
345 static volatile sig_atomic_t notifyInterruptEnabled = 0;
346 static volatile sig_atomic_t notifyInterruptOccurred = 0;
348 /* True if we've registered an on_shmem_exit cleanup */
349 static bool unlistenExitRegistered = false;
351 /* True if we're currently registered as a listener in asyncQueueControl */
352 static bool amRegisteredListener = false;
354 /* has this backend sent notifications in the current transaction? */
355 static bool backendHasSentNotifications = false;
358 bool Trace_notify = false;
360 /* local function prototypes */
361 static bool asyncQueuePagePrecedes(int p, int q);
362 static void queue_listen(ListenActionKind action, const char *channel);
363 static void Async_UnlistenOnExit(int code, Datum arg);
364 static void Exec_ListenPreCommit(void);
365 static void Exec_ListenCommit(const char *channel);
366 static void Exec_UnlistenCommit(const char *channel);
367 static void Exec_UnlistenAllCommit(void);
368 static bool IsListeningOn(const char *channel);
369 static void asyncQueueUnregister(void);
370 static bool asyncQueueIsFull(void);
371 static bool asyncQueueAdvance(QueuePosition *position, int entryLength);
372 static void asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe);
373 static ListCell *asyncQueueAddEntries(ListCell *nextNotify);
374 static void asyncQueueFillWarning(void);
375 static bool SignalBackends(void);
376 static void asyncQueueReadAllNotifications(void);
377 static bool asyncQueueProcessPageEntries(QueuePosition *current,
380 static void asyncQueueAdvanceTail(void);
381 static void ProcessIncomingNotify(void);
382 static void NotifyMyFrontEnd(const char *channel,
385 static bool AsyncExistsPendingNotify(const char *channel, const char *payload);
386 static void ClearPendingActionsAndNotifies(void);
389 * We will work on the page range of 0..QUEUE_MAX_PAGE.
392 asyncQueuePagePrecedes(int p, int q)
397 * We have to compare modulo (QUEUE_MAX_PAGE+1)/2. Both inputs should be
398 * in the range 0..QUEUE_MAX_PAGE.
400 Assert(p >= 0 && p <= QUEUE_MAX_PAGE);
401 Assert(q >= 0 && q <= QUEUE_MAX_PAGE);
404 if (diff >= ((QUEUE_MAX_PAGE + 1) / 2))
405 diff -= QUEUE_MAX_PAGE + 1;
406 else if (diff < -((QUEUE_MAX_PAGE + 1) / 2))
407 diff += QUEUE_MAX_PAGE + 1;
412 * Report space needed for our shared memory area
419 /* This had better match AsyncShmemInit */
420 size = mul_size(MaxBackends, sizeof(QueueBackendStatus));
421 size = add_size(size, sizeof(AsyncQueueControl));
423 size = add_size(size, SimpleLruShmemSize(NUM_ASYNC_BUFFERS, 0));
429 * Initialize our shared memory area
439 * Create or attach to the AsyncQueueControl structure.
441 * The used entries in the backend[] array run from 1 to MaxBackends.
442 * sizeof(AsyncQueueControl) already includes space for the unused zero'th
443 * entry, but we need to add on space for the used entries.
445 size = mul_size(MaxBackends, sizeof(QueueBackendStatus));
446 size = add_size(size, sizeof(AsyncQueueControl));
448 asyncQueueControl = (AsyncQueueControl *)
449 ShmemInitStruct("Async Queue Control", size, &found);
453 /* First time through, so initialize it */
456 SET_QUEUE_POS(QUEUE_HEAD, 0, 0);
457 SET_QUEUE_POS(QUEUE_TAIL, 0, 0);
458 asyncQueueControl->lastQueueFillWarn = 0;
459 /* zero'th entry won't be used, but let's initialize it anyway */
460 for (i = 0; i <= MaxBackends; i++)
462 QUEUE_BACKEND_PID(i) = InvalidPid;
463 SET_QUEUE_POS(QUEUE_BACKEND_POS(i), 0, 0);
468 * Set up SLRU management of the pg_notify data.
470 AsyncCtl->PagePrecedes = asyncQueuePagePrecedes;
471 SimpleLruInit(AsyncCtl, "Async Ctl", NUM_ASYNC_BUFFERS, 0,
472 AsyncCtlLock, "pg_notify");
473 /* Override default assumption that writes should be fsync'd */
474 AsyncCtl->do_fsync = false;
479 * During start or reboot, clean out the pg_notify directory.
481 (void) SlruScanDirectory(AsyncCtl, SlruScanDirCbDeleteAll, NULL);
483 /* Now initialize page zero to empty */
484 LWLockAcquire(AsyncCtlLock, LW_EXCLUSIVE);
485 slotno = SimpleLruZeroPage(AsyncCtl, QUEUE_POS_PAGE(QUEUE_HEAD));
486 /* This write is just to verify that pg_notify/ is writable */
487 SimpleLruWritePage(AsyncCtl, slotno);
488 LWLockRelease(AsyncCtlLock);
495 * SQL function to send a notification event
498 pg_notify(PG_FUNCTION_ARGS)
506 channel = text_to_cstring(PG_GETARG_TEXT_PP(0));
511 payload = text_to_cstring(PG_GETARG_TEXT_PP(1));
513 /* For NOTIFY as a statement, this is checked in ProcessUtility */
514 PreventCommandDuringRecovery("NOTIFY");
516 Async_Notify(channel, payload);
525 * This is executed by the SQL notify command.
527 * Adds the message to the list of pending notifies.
528 * Actual notification happens during transaction commit.
529 * ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
532 Async_Notify(const char *channel, const char *payload)
535 MemoryContext oldcontext;
538 elog(DEBUG1, "Async_Notify(%s)", channel);
540 /* a channel name must be specified */
541 if (!channel || !strlen(channel))
543 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
544 errmsg("channel name cannot be empty")));
546 if (strlen(channel) >= NAMEDATALEN)
548 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
549 errmsg("channel name too long")));
553 if (strlen(payload) >= NOTIFY_PAYLOAD_MAX_LENGTH)
555 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
556 errmsg("payload string too long")));
559 /* no point in making duplicate entries in the list ... */
560 if (AsyncExistsPendingNotify(channel, payload))
564 * The notification list needs to live until end of transaction, so store
565 * it in the transaction context.
567 oldcontext = MemoryContextSwitchTo(CurTransactionContext);
569 n = (Notification *) palloc(sizeof(Notification));
570 n->channel = pstrdup(channel);
572 n->payload = pstrdup(payload);
577 * We want to preserve the order so we need to append every notification.
578 * See comments at AsyncExistsPendingNotify().
580 pendingNotifies = lappend(pendingNotifies, n);
582 MemoryContextSwitchTo(oldcontext);
587 * Common code for listen, unlisten, unlisten all commands.
589 * Adds the request to the list of pending actions.
590 * Actual update of the listenChannels list happens during transaction
594 queue_listen(ListenActionKind action, const char *channel)
596 MemoryContext oldcontext;
597 ListenAction *actrec;
600 * Unlike Async_Notify, we don't try to collapse out duplicates. It would
601 * be too complicated to ensure we get the right interactions of
602 * conflicting LISTEN/UNLISTEN/UNLISTEN_ALL, and it's unlikely that there
603 * would be any performance benefit anyway in sane applications.
605 oldcontext = MemoryContextSwitchTo(CurTransactionContext);
607 /* space for terminating null is included in sizeof(ListenAction) */
608 actrec = (ListenAction *) palloc(sizeof(ListenAction) + strlen(channel));
609 actrec->action = action;
610 strcpy(actrec->channel, channel);
612 pendingActions = lappend(pendingActions, actrec);
614 MemoryContextSwitchTo(oldcontext);
620 * This is executed by the SQL listen command.
623 Async_Listen(const char *channel)
626 elog(DEBUG1, "Async_Listen(%s,%d)", channel, MyProcPid);
628 queue_listen(LISTEN_LISTEN, channel);
634 * This is executed by the SQL unlisten command.
637 Async_Unlisten(const char *channel)
640 elog(DEBUG1, "Async_Unlisten(%s,%d)", channel, MyProcPid);
642 /* If we couldn't possibly be listening, no need to queue anything */
643 if (pendingActions == NIL && !unlistenExitRegistered)
646 queue_listen(LISTEN_UNLISTEN, channel);
652 * This is invoked by UNLISTEN * command, and also at backend exit.
655 Async_UnlistenAll(void)
658 elog(DEBUG1, "Async_UnlistenAll(%d)", MyProcPid);
660 /* If we couldn't possibly be listening, no need to queue anything */
661 if (pendingActions == NIL && !unlistenExitRegistered)
664 queue_listen(LISTEN_UNLISTEN_ALL, "");
668 * SQL function: return a set of the channel names this backend is actively
671 * Note: this coding relies on the fact that the listenChannels list cannot
672 * change within a transaction.
675 pg_listening_channels(PG_FUNCTION_ARGS)
677 FuncCallContext *funcctx;
680 /* stuff done only on the first call of the function */
681 if (SRF_IS_FIRSTCALL())
683 MemoryContext oldcontext;
685 /* create a function context for cross-call persistence */
686 funcctx = SRF_FIRSTCALL_INIT();
688 /* switch to memory context appropriate for multiple function calls */
689 oldcontext = MemoryContextSwitchTo(funcctx->multi_call_memory_ctx);
691 /* allocate memory for user context */
692 lcp = (ListCell **) palloc(sizeof(ListCell *));
693 *lcp = list_head(listenChannels);
694 funcctx->user_fctx = (void *) lcp;
696 MemoryContextSwitchTo(oldcontext);
699 /* stuff done on every call of the function */
700 funcctx = SRF_PERCALL_SETUP();
701 lcp = (ListCell **) funcctx->user_fctx;
705 char *channel = (char *) lfirst(*lcp);
708 SRF_RETURN_NEXT(funcctx, CStringGetTextDatum(channel));
711 SRF_RETURN_DONE(funcctx);
715 * Async_UnlistenOnExit
717 * This is executed at backend exit if we have done any LISTENs in this
718 * backend. It might not be necessary anymore, if the user UNLISTENed
719 * everything, but we don't try to detect that case.
722 Async_UnlistenOnExit(int code, Datum arg)
724 Exec_UnlistenAllCommit();
725 asyncQueueUnregister();
731 * This is called at the prepare phase of a two-phase
732 * transaction. Save the state for possible commit later.
735 AtPrepare_Notify(void)
737 /* It's not allowed to have any pending LISTEN/UNLISTEN/NOTIFY actions */
738 if (pendingActions || pendingNotifies)
740 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
741 errmsg("cannot PREPARE a transaction that has executed LISTEN, UNLISTEN, or NOTIFY")));
747 * This is called at transaction commit, before actually committing to
750 * If there are pending LISTEN actions, make sure we are listed in the
751 * shared-memory listener array. This must happen before commit to
752 * ensure we don't miss any notifies from transactions that commit
755 * If there are outbound notify requests in the pendingNotifies list,
756 * add them to the global queue. We do that before commit so that
757 * we can still throw error if we run out of queue space.
760 PreCommit_Notify(void)
764 if (pendingActions == NIL && pendingNotifies == NIL)
765 return; /* no relevant statements in this xact */
768 elog(DEBUG1, "PreCommit_Notify");
770 /* Preflight for any pending listen/unlisten actions */
771 foreach(p, pendingActions)
773 ListenAction *actrec = (ListenAction *) lfirst(p);
775 switch (actrec->action)
778 Exec_ListenPreCommit();
780 case LISTEN_UNLISTEN:
781 /* there is no Exec_UnlistenPreCommit() */
783 case LISTEN_UNLISTEN_ALL:
784 /* there is no Exec_UnlistenAllPreCommit() */
789 /* Queue any pending notifies */
792 ListCell *nextNotify;
795 * Make sure that we have an XID assigned to the current transaction.
796 * GetCurrentTransactionId is cheap if we already have an XID, but not
797 * so cheap if we don't, and we'd prefer not to do that work while
798 * holding AsyncQueueLock.
800 (void) GetCurrentTransactionId();
803 * Serialize writers by acquiring a special lock that we hold till
804 * after commit. This ensures that queue entries appear in commit
805 * order, and in particular that there are never uncommitted queue
806 * entries ahead of committed ones, so an uncommitted transaction
807 * can't block delivery of deliverable notifications.
809 * We use a heavyweight lock so that it'll automatically be released
810 * after either commit or abort. This also allows deadlocks to be
811 * detected, though really a deadlock shouldn't be possible here.
813 * The lock is on "database 0", which is pretty ugly but it doesn't
814 * seem worth inventing a special locktag category just for this.
815 * (Historical note: before PG 9.0, a similar lock on "database 0" was
816 * used by the flatfiles mechanism.)
818 LockSharedObject(DatabaseRelationId, InvalidOid, 0,
819 AccessExclusiveLock);
821 /* Now push the notifications into the queue */
822 backendHasSentNotifications = true;
824 nextNotify = list_head(pendingNotifies);
825 while (nextNotify != NULL)
828 * Add the pending notifications to the queue. We acquire and
829 * release AsyncQueueLock once per page, which might be overkill
830 * but it does allow readers to get in while we're doing this.
832 * A full queue is very uncommon and should really not happen,
833 * given that we have so much space available in the SLRU pages.
834 * Nevertheless we need to deal with this possibility. Note that
835 * when we get here we are in the process of committing our
836 * transaction, but we have not yet committed to clog, so at this
837 * point in time we can still roll the transaction back.
839 LWLockAcquire(AsyncQueueLock, LW_EXCLUSIVE);
840 asyncQueueFillWarning();
841 if (asyncQueueIsFull())
843 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
844 errmsg("too many notifications in the NOTIFY queue")));
845 nextNotify = asyncQueueAddEntries(nextNotify);
846 LWLockRelease(AsyncQueueLock);
854 * This is called at transaction commit, after committing to clog.
856 * Update listenChannels and clear transaction-local state.
859 AtCommit_Notify(void)
864 * Allow transactions that have not executed LISTEN/UNLISTEN/NOTIFY to
865 * return as soon as possible
867 if (!pendingActions && !pendingNotifies)
871 elog(DEBUG1, "AtCommit_Notify");
873 /* Perform any pending listen/unlisten actions */
874 foreach(p, pendingActions)
876 ListenAction *actrec = (ListenAction *) lfirst(p);
878 switch (actrec->action)
881 Exec_ListenCommit(actrec->channel);
883 case LISTEN_UNLISTEN:
884 Exec_UnlistenCommit(actrec->channel);
886 case LISTEN_UNLISTEN_ALL:
887 Exec_UnlistenAllCommit();
892 /* If no longer listening to anything, get out of listener array */
893 if (amRegisteredListener && listenChannels == NIL)
894 asyncQueueUnregister();
897 ClearPendingActionsAndNotifies();
901 * Exec_ListenPreCommit --- subroutine for PreCommit_Notify
903 * This function must make sure we are ready to catch any incoming messages.
906 Exec_ListenPreCommit(void)
909 * Nothing to do if we are already listening to something, nor if we
910 * already ran this routine in this transaction.
912 if (amRegisteredListener)
916 elog(DEBUG1, "Exec_ListenPreCommit(%d)", MyProcPid);
919 * Before registering, make sure we will unlisten before dying. (Note:
920 * this action does not get undone if we abort later.)
922 if (!unlistenExitRegistered)
924 before_shmem_exit(Async_UnlistenOnExit, 0);
925 unlistenExitRegistered = true;
929 * This is our first LISTEN, so establish our pointer.
931 * We set our pointer to the global tail pointer and then move it forward
932 * over already-committed notifications. This ensures we cannot miss any
933 * not-yet-committed notifications. We might get a few more but that
936 LWLockAcquire(AsyncQueueLock, LW_SHARED);
937 QUEUE_BACKEND_POS(MyBackendId) = QUEUE_TAIL;
938 QUEUE_BACKEND_PID(MyBackendId) = MyProcPid;
939 LWLockRelease(AsyncQueueLock);
941 /* Now we are listed in the global array, so remember we're listening */
942 amRegisteredListener = true;
945 * Try to move our pointer forward as far as possible. This will skip over
946 * already-committed notifications. Still, we could get notifications that
947 * have already committed before we started to LISTEN.
949 * Note that we are not yet listening on anything, so we won't deliver any
950 * notification to the frontend.
952 * This will also advance the global tail pointer if possible.
954 asyncQueueReadAllNotifications();
958 * Exec_ListenCommit --- subroutine for AtCommit_Notify
960 * Add the channel to the list of channels we are listening on.
963 Exec_ListenCommit(const char *channel)
965 MemoryContext oldcontext;
967 /* Do nothing if we are already listening on this channel */
968 if (IsListeningOn(channel))
972 * Add the new channel name to listenChannels.
974 * XXX It is theoretically possible to get an out-of-memory failure here,
975 * which would be bad because we already committed. For the moment it
976 * doesn't seem worth trying to guard against that, but maybe improve this
979 oldcontext = MemoryContextSwitchTo(TopMemoryContext);
980 listenChannels = lappend(listenChannels, pstrdup(channel));
981 MemoryContextSwitchTo(oldcontext);
985 * Exec_UnlistenCommit --- subroutine for AtCommit_Notify
987 * Remove the specified channel name from listenChannels.
990 Exec_UnlistenCommit(const char *channel)
996 elog(DEBUG1, "Exec_UnlistenCommit(%s,%d)", channel, MyProcPid);
999 foreach(q, listenChannels)
1001 char *lchan = (char *) lfirst(q);
1003 if (strcmp(lchan, channel) == 0)
1005 listenChannels = list_delete_cell(listenChannels, q, prev);
1013 * We do not complain about unlistening something not being listened;
1019 * Exec_UnlistenAllCommit --- subroutine for AtCommit_Notify
1021 * Unlisten on all channels for this backend.
1024 Exec_UnlistenAllCommit(void)
1027 elog(DEBUG1, "Exec_UnlistenAllCommit(%d)", MyProcPid);
1029 list_free_deep(listenChannels);
1030 listenChannels = NIL;
1034 * ProcessCompletedNotifies --- send out signals and self-notifies
1036 * This is called from postgres.c just before going idle at the completion
1037 * of a transaction. If we issued any notifications in the just-completed
1038 * transaction, send signals to other backends to process them, and also
1039 * process the queue ourselves to send messages to our own frontend.
1041 * The reason that this is not done in AtCommit_Notify is that there is
1042 * a nonzero chance of errors here (for example, encoding conversion errors
1043 * while trying to format messages to our frontend). An error during
1044 * AtCommit_Notify would be a PANIC condition. The timing is also arranged
1045 * to ensure that a transaction's self-notifies are delivered to the frontend
1046 * before it gets the terminating ReadyForQuery message.
1048 * Note that we send signals and process the queue even if the transaction
1049 * eventually aborted. This is because we need to clean out whatever got
1050 * added to the queue.
1052 * NOTE: we are outside of any transaction here.
1055 ProcessCompletedNotifies(void)
1057 MemoryContext caller_context;
1060 /* Nothing to do if we didn't send any notifications */
1061 if (!backendHasSentNotifications)
1065 * We reset the flag immediately; otherwise, if any sort of error occurs
1066 * below, we'd be locked up in an infinite loop, because control will come
1067 * right back here after error cleanup.
1069 backendHasSentNotifications = false;
1072 * We must preserve the caller's memory context (probably MessageContext)
1073 * across the transaction we do here.
1075 caller_context = CurrentMemoryContext;
1078 elog(DEBUG1, "ProcessCompletedNotifies");
1081 * We must run asyncQueueReadAllNotifications inside a transaction, else
1082 * bad things happen if it gets an error.
1084 StartTransactionCommand();
1086 /* Send signals to other backends */
1087 signalled = SignalBackends();
1089 if (listenChannels != NIL)
1091 /* Read the queue ourselves, and send relevant stuff to the frontend */
1092 asyncQueueReadAllNotifications();
1094 else if (!signalled)
1097 * If we found no other listening backends, and we aren't listening
1098 * ourselves, then we must execute asyncQueueAdvanceTail to flush the
1099 * queue, because ain't nobody else gonna do it. This prevents queue
1100 * overflow when we're sending useless notifies to nobody. (A new
1101 * listener could have joined since we looked, but if so this is
1104 asyncQueueAdvanceTail();
1107 CommitTransactionCommand();
1109 MemoryContextSwitchTo(caller_context);
1111 /* We don't need pq_flush() here since postgres.c will do one shortly */
1115 * Test whether we are actively listening on the given channel name.
1117 * Note: this function is executed for every notification found in the queue.
1118 * Perhaps it is worth further optimization, eg convert the list to a sorted
1119 * array so we can binary-search it. In practice the list is likely to be
1120 * fairly short, though.
1123 IsListeningOn(const char *channel)
1127 foreach(p, listenChannels)
1129 char *lchan = (char *) lfirst(p);
1131 if (strcmp(lchan, channel) == 0)
1138 * Remove our entry from the listeners array when we are no longer listening
1139 * on any channel. NB: must not fail if we're already not listening.
1142 asyncQueueUnregister(void)
1146 Assert(listenChannels == NIL); /* else caller error */
1148 if (!amRegisteredListener) /* nothing to do */
1151 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1152 /* check if entry is valid and oldest ... */
1153 advanceTail = (MyProcPid == QUEUE_BACKEND_PID(MyBackendId)) &&
1154 QUEUE_POS_EQUAL(QUEUE_BACKEND_POS(MyBackendId), QUEUE_TAIL);
1155 /* ... then mark it invalid */
1156 QUEUE_BACKEND_PID(MyBackendId) = InvalidPid;
1157 LWLockRelease(AsyncQueueLock);
1159 /* mark ourselves as no longer listed in the global array */
1160 amRegisteredListener = false;
1162 /* If we were the laziest backend, try to advance the tail pointer */
1164 asyncQueueAdvanceTail();
1168 * Test whether there is room to insert more notification messages.
1170 * Caller must hold at least shared AsyncQueueLock.
1173 asyncQueueIsFull(void)
1179 * The queue is full if creating a new head page would create a page that
1180 * logically precedes the current global tail pointer, ie, the head
1181 * pointer would wrap around compared to the tail. We cannot create such
1182 * a head page for fear of confusing slru.c. For safety we round the tail
1183 * pointer back to a segment boundary (compare the truncation logic in
1184 * asyncQueueAdvanceTail).
1186 * Note that this test is *not* dependent on how much space there is on
1187 * the current head page. This is necessary because asyncQueueAddEntries
1188 * might try to create the next head page in any case.
1190 nexthead = QUEUE_POS_PAGE(QUEUE_HEAD) + 1;
1191 if (nexthead > QUEUE_MAX_PAGE)
1192 nexthead = 0; /* wrap around */
1193 boundary = QUEUE_POS_PAGE(QUEUE_TAIL);
1194 boundary -= boundary % SLRU_PAGES_PER_SEGMENT;
1195 return asyncQueuePagePrecedes(nexthead, boundary);
1199 * Advance the QueuePosition to the next entry, assuming that the current
1200 * entry is of length entryLength. If we jump to a new page the function
1201 * returns true, else false.
1204 asyncQueueAdvance(QueuePosition *position, int entryLength)
1206 int pageno = QUEUE_POS_PAGE(*position);
1207 int offset = QUEUE_POS_OFFSET(*position);
1208 bool pageJump = false;
1211 * Move to the next writing position: First jump over what we have just
1214 offset += entryLength;
1215 Assert(offset <= QUEUE_PAGESIZE);
1218 * In a second step check if another entry can possibly be written to the
1219 * page. If so, stay here, we have reached the next position. If not, then
1220 * we need to move on to the next page.
1222 if (offset + QUEUEALIGN(AsyncQueueEntryEmptySize) > QUEUE_PAGESIZE)
1225 if (pageno > QUEUE_MAX_PAGE)
1226 pageno = 0; /* wrap around */
1231 SET_QUEUE_POS(*position, pageno, offset);
1236 * Fill the AsyncQueueEntry at *qe with an outbound notification message.
1239 asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe)
1241 size_t channellen = strlen(n->channel);
1242 size_t payloadlen = strlen(n->payload);
1245 Assert(channellen < NAMEDATALEN);
1246 Assert(payloadlen < NOTIFY_PAYLOAD_MAX_LENGTH);
1248 /* The terminators are already included in AsyncQueueEntryEmptySize */
1249 entryLength = AsyncQueueEntryEmptySize + payloadlen + channellen;
1250 entryLength = QUEUEALIGN(entryLength);
1251 qe->length = entryLength;
1252 qe->dboid = MyDatabaseId;
1253 qe->xid = GetCurrentTransactionId();
1254 qe->srcPid = MyProcPid;
1255 memcpy(qe->data, n->channel, channellen + 1);
1256 memcpy(qe->data + channellen + 1, n->payload, payloadlen + 1);
1260 * Add pending notifications to the queue.
1262 * We go page by page here, i.e. we stop once we have to go to a new page but
1263 * we will be called again and then fill that next page. If an entry does not
1264 * fit into the current page, we write a dummy entry with an InvalidOid as the
1265 * database OID in order to fill the page. So every page is always used up to
1266 * the last byte which simplifies reading the page later.
1268 * We are passed the list cell containing the next notification to write
1269 * and return the first still-unwritten cell back. Eventually we will return
1270 * NULL indicating all is done.
1272 * We are holding AsyncQueueLock already from the caller and grab AsyncCtlLock
1273 * locally in this function.
1276 asyncQueueAddEntries(ListCell *nextNotify)
1279 QueuePosition queue_head;
1284 /* We hold both AsyncQueueLock and AsyncCtlLock during this operation */
1285 LWLockAcquire(AsyncCtlLock, LW_EXCLUSIVE);
1288 * We work with a local copy of QUEUE_HEAD, which we write back to shared
1289 * memory upon exiting. The reason for this is that if we have to advance
1290 * to a new page, SimpleLruZeroPage might fail (out of disk space, for
1291 * instance), and we must not advance QUEUE_HEAD if it does. (Otherwise,
1292 * subsequent insertions would try to put entries into a page that slru.c
1293 * thinks doesn't exist yet.) So, use a local position variable. Note
1294 * that if we do fail, any already-inserted queue entries are forgotten;
1295 * this is okay, since they'd be useless anyway after our transaction
1298 queue_head = QUEUE_HEAD;
1300 /* Fetch the current page */
1301 pageno = QUEUE_POS_PAGE(queue_head);
1302 slotno = SimpleLruReadPage(AsyncCtl, pageno, true, InvalidTransactionId);
1303 /* Note we mark the page dirty before writing in it */
1304 AsyncCtl->shared->page_dirty[slotno] = true;
1306 while (nextNotify != NULL)
1308 Notification *n = (Notification *) lfirst(nextNotify);
1310 /* Construct a valid queue entry in local variable qe */
1311 asyncQueueNotificationToEntry(n, &qe);
1313 offset = QUEUE_POS_OFFSET(queue_head);
1315 /* Check whether the entry really fits on the current page */
1316 if (offset + qe.length <= QUEUE_PAGESIZE)
1318 /* OK, so advance nextNotify past this item */
1319 nextNotify = lnext(nextNotify);
1324 * Write a dummy entry to fill up the page. Actually readers will
1325 * only check dboid and since it won't match any reader's database
1326 * OID, they will ignore this entry and move on.
1328 qe.length = QUEUE_PAGESIZE - offset;
1329 qe.dboid = InvalidOid;
1330 qe.data[0] = '\0'; /* empty channel */
1331 qe.data[1] = '\0'; /* empty payload */
1334 /* Now copy qe into the shared buffer page */
1335 memcpy(AsyncCtl->shared->page_buffer[slotno] + offset,
1339 /* Advance queue_head appropriately, and detect if page is full */
1340 if (asyncQueueAdvance(&(queue_head), qe.length))
1343 * Page is full, so we're done here, but first fill the next page
1344 * with zeroes. The reason to do this is to ensure that slru.c's
1345 * idea of the head page is always the same as ours, which avoids
1346 * boundary problems in SimpleLruTruncate. The test in
1347 * asyncQueueIsFull() ensured that there is room to create this
1348 * page without overrunning the queue.
1350 slotno = SimpleLruZeroPage(AsyncCtl, QUEUE_POS_PAGE(queue_head));
1351 /* And exit the loop */
1356 /* Success, so update the global QUEUE_HEAD */
1357 QUEUE_HEAD = queue_head;
1359 LWLockRelease(AsyncCtlLock);
1365 * Check whether the queue is at least half full, and emit a warning if so.
1367 * This is unlikely given the size of the queue, but possible.
1368 * The warnings show up at most once every QUEUE_FULL_WARN_INTERVAL.
1370 * Caller must hold exclusive AsyncQueueLock.
1373 asyncQueueFillWarning(void)
1375 int headPage = QUEUE_POS_PAGE(QUEUE_HEAD);
1376 int tailPage = QUEUE_POS_PAGE(QUEUE_TAIL);
1381 occupied = headPage - tailPage;
1384 return; /* fast exit for common case */
1388 /* head has wrapped around, tail not yet */
1389 occupied += QUEUE_MAX_PAGE + 1;
1392 fillDegree = (double) occupied / (double) ((QUEUE_MAX_PAGE + 1) / 2);
1394 if (fillDegree < 0.5)
1397 t = GetCurrentTimestamp();
1399 if (TimestampDifferenceExceeds(asyncQueueControl->lastQueueFillWarn,
1400 t, QUEUE_FULL_WARN_INTERVAL))
1402 QueuePosition min = QUEUE_HEAD;
1403 int32 minPid = InvalidPid;
1406 for (i = 1; i <= MaxBackends; i++)
1408 if (QUEUE_BACKEND_PID(i) != InvalidPid)
1410 min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
1411 if (QUEUE_POS_EQUAL(min, QUEUE_BACKEND_POS(i)))
1412 minPid = QUEUE_BACKEND_PID(i);
1417 (errmsg("NOTIFY queue is %.0f%% full", fillDegree * 100),
1418 (minPid != InvalidPid ?
1419 errdetail("The server process with PID %d is among those with the oldest transactions.", minPid)
1421 (minPid != InvalidPid ?
1422 errhint("The NOTIFY queue cannot be emptied until that process ends its current transaction.")
1425 asyncQueueControl->lastQueueFillWarn = t;
1430 * Send signals to all listening backends (except our own).
1432 * Returns true if we sent at least one signal.
1434 * Since we need EXCLUSIVE lock anyway we also check the position of the other
1435 * backends and in case one is already up-to-date we don't signal it.
1436 * This can happen if concurrent notifying transactions have sent a signal and
1437 * the signaled backend has read the other notifications and ours in the same
1440 * Since we know the BackendId and the Pid the signalling is quite cheap.
1443 SignalBackends(void)
1445 bool signalled = false;
1453 * Identify all backends that are listening and not already up-to-date. We
1454 * don't want to send signals while holding the AsyncQueueLock, so we just
1455 * build a list of target PIDs.
1457 * XXX in principle these pallocs could fail, which would be bad. Maybe
1458 * preallocate the arrays? But in practice this is only run in trivial
1459 * transactions, so there should surely be space available.
1461 pids = (int32 *) palloc(MaxBackends * sizeof(int32));
1462 ids = (BackendId *) palloc(MaxBackends * sizeof(BackendId));
1465 LWLockAcquire(AsyncQueueLock, LW_EXCLUSIVE);
1466 for (i = 1; i <= MaxBackends; i++)
1468 pid = QUEUE_BACKEND_PID(i);
1469 if (pid != InvalidPid && pid != MyProcPid)
1471 QueuePosition pos = QUEUE_BACKEND_POS(i);
1473 if (!QUEUE_POS_EQUAL(pos, QUEUE_HEAD))
1481 LWLockRelease(AsyncQueueLock);
1483 /* Now send signals */
1484 for (i = 0; i < count; i++)
1489 * Note: assuming things aren't broken, a signal failure here could
1490 * only occur if the target backend exited since we released
1491 * AsyncQueueLock; which is unlikely but certainly possible. So we
1492 * just log a low-level debug message if it happens.
1494 if (SendProcSignal(pid, PROCSIG_NOTIFY_INTERRUPT, ids[i]) < 0)
1495 elog(DEBUG3, "could not signal backend with PID %d: %m", pid);
1509 * This is called at transaction abort.
1511 * Gets rid of pending actions and outbound notifies that we would have
1512 * executed if the transaction got committed.
1515 AtAbort_Notify(void)
1518 * If we LISTEN but then roll back the transaction after PreCommit_Notify,
1519 * we have registered as a listener but have not made any entry in
1520 * listenChannels. In that case, deregister again.
1522 if (amRegisteredListener && listenChannels == NIL)
1523 asyncQueueUnregister();
1526 ClearPendingActionsAndNotifies();
1530 * AtSubStart_Notify() --- Take care of subtransaction start.
1532 * Push empty state for the new subtransaction.
1535 AtSubStart_Notify(void)
1537 MemoryContext old_cxt;
1539 /* Keep the list-of-lists in TopTransactionContext for simplicity */
1540 old_cxt = MemoryContextSwitchTo(TopTransactionContext);
1542 upperPendingActions = lcons(pendingActions, upperPendingActions);
1544 Assert(list_length(upperPendingActions) ==
1545 GetCurrentTransactionNestLevel() - 1);
1547 pendingActions = NIL;
1549 upperPendingNotifies = lcons(pendingNotifies, upperPendingNotifies);
1551 Assert(list_length(upperPendingNotifies) ==
1552 GetCurrentTransactionNestLevel() - 1);
1554 pendingNotifies = NIL;
1556 MemoryContextSwitchTo(old_cxt);
1560 * AtSubCommit_Notify() --- Take care of subtransaction commit.
1562 * Reassign all items in the pending lists to the parent transaction.
1565 AtSubCommit_Notify(void)
1567 List *parentPendingActions;
1568 List *parentPendingNotifies;
1570 parentPendingActions = (List *) linitial(upperPendingActions);
1571 upperPendingActions = list_delete_first(upperPendingActions);
1573 Assert(list_length(upperPendingActions) ==
1574 GetCurrentTransactionNestLevel() - 2);
1577 * Mustn't try to eliminate duplicates here --- see queue_listen()
1579 pendingActions = list_concat(parentPendingActions, pendingActions);
1581 parentPendingNotifies = (List *) linitial(upperPendingNotifies);
1582 upperPendingNotifies = list_delete_first(upperPendingNotifies);
1584 Assert(list_length(upperPendingNotifies) ==
1585 GetCurrentTransactionNestLevel() - 2);
1588 * We could try to eliminate duplicates here, but it seems not worthwhile.
1590 pendingNotifies = list_concat(parentPendingNotifies, pendingNotifies);
1594 * AtSubAbort_Notify() --- Take care of subtransaction abort.
1597 AtSubAbort_Notify(void)
1599 int my_level = GetCurrentTransactionNestLevel();
1602 * All we have to do is pop the stack --- the actions/notifies made in
1603 * this subxact are no longer interesting, and the space will be freed
1604 * when CurTransactionContext is recycled.
1606 * This routine could be called more than once at a given nesting level if
1607 * there is trouble during subxact abort. Avoid dumping core by using
1608 * GetCurrentTransactionNestLevel as the indicator of how far we need to
1611 while (list_length(upperPendingActions) > my_level - 2)
1613 pendingActions = (List *) linitial(upperPendingActions);
1614 upperPendingActions = list_delete_first(upperPendingActions);
1617 while (list_length(upperPendingNotifies) > my_level - 2)
1619 pendingNotifies = (List *) linitial(upperPendingNotifies);
1620 upperPendingNotifies = list_delete_first(upperPendingNotifies);
1625 * HandleNotifyInterrupt
1627 * This is called when PROCSIG_NOTIFY_INTERRUPT is received.
1629 * If we are idle (notifyInterruptEnabled is set), we can safely invoke
1630 * ProcessIncomingNotify directly. Otherwise, just set a flag
1634 HandleNotifyInterrupt(void)
1637 * Note: this is called by a SIGNAL HANDLER. You must be very wary what
1638 * you do here. Some helpful soul had this routine sprinkled with
1639 * TPRINTFs, which would likely lead to corruption of stdio buffers if
1640 * they were ever turned on.
1643 /* Don't joggle the elbow of proc_exit */
1644 if (proc_exit_inprogress)
1647 if (notifyInterruptEnabled)
1649 bool save_ImmediateInterruptOK = ImmediateInterruptOK;
1652 * We may be called while ImmediateInterruptOK is true; turn it off
1653 * while messing with the NOTIFY state. This prevents problems if
1654 * SIGINT or similar arrives while we're working. Just to be real
1655 * sure, bump the interrupt holdoff counter as well. That way, even
1656 * if something inside ProcessIncomingNotify() transiently sets
1657 * ImmediateInterruptOK (eg while waiting on a lock), we won't get
1658 * interrupted until we're done with the notify interrupt.
1660 ImmediateInterruptOK = false;
1664 * I'm not sure whether some flavors of Unix might allow another
1665 * SIGUSR1 occurrence to recursively interrupt this routine. To cope
1666 * with the possibility, we do the same sort of dance that
1667 * EnableNotifyInterrupt must do --- see that routine for comments.
1669 notifyInterruptEnabled = 0; /* disable any recursive signal */
1670 notifyInterruptOccurred = 1; /* do at least one iteration */
1673 notifyInterruptEnabled = 1;
1674 if (!notifyInterruptOccurred)
1676 notifyInterruptEnabled = 0;
1677 if (notifyInterruptOccurred)
1679 /* Here, it is finally safe to do stuff. */
1681 elog(DEBUG1, "HandleNotifyInterrupt: perform async notify");
1683 ProcessIncomingNotify();
1686 elog(DEBUG1, "HandleNotifyInterrupt: done");
1691 * Restore the holdoff level and ImmediateInterruptOK, and check for
1692 * interrupts if needed.
1694 RESUME_INTERRUPTS();
1695 ImmediateInterruptOK = save_ImmediateInterruptOK;
1696 if (save_ImmediateInterruptOK)
1697 CHECK_FOR_INTERRUPTS();
1702 * In this path it is NOT SAFE to do much of anything, except this:
1704 notifyInterruptOccurred = 1;
1709 * EnableNotifyInterrupt
1711 * This is called by the PostgresMain main loop just before waiting
1712 * for a frontend command. If we are truly idle (ie, *not* inside
1713 * a transaction block), then process any pending inbound notifies,
1714 * and enable the signal handler to process future notifies directly.
1716 * NOTE: the signal handler starts out disabled, and stays so until
1717 * PostgresMain calls this the first time.
1720 EnableNotifyInterrupt(void)
1722 if (IsTransactionOrTransactionBlock())
1723 return; /* not really idle */
1726 * This code is tricky because we are communicating with a signal handler
1727 * that could interrupt us at any point. If we just checked
1728 * notifyInterruptOccurred and then set notifyInterruptEnabled, we could
1729 * fail to respond promptly to a signal that happens in between those two
1730 * steps. (A very small time window, perhaps, but Murphy's Law says you
1731 * can hit it...) Instead, we first set the enable flag, then test the
1732 * occurred flag. If we see an unserviced interrupt has occurred, we
1733 * re-clear the enable flag before going off to do the service work. (That
1734 * prevents re-entrant invocation of ProcessIncomingNotify() if another
1735 * interrupt occurs.) If an interrupt comes in between the setting and
1736 * clearing of notifyInterruptEnabled, then it will have done the service
1737 * work and left notifyInterruptOccurred zero, so we have to check again
1738 * after clearing enable. The whole thing has to be in a loop in case
1739 * another interrupt occurs while we're servicing the first. Once we get
1740 * out of the loop, enable is set and we know there is no unserviced
1743 * NB: an overenthusiastic optimizing compiler could easily break this
1744 * code. Hopefully, they all understand what "volatile" means these days.
1748 notifyInterruptEnabled = 1;
1749 if (!notifyInterruptOccurred)
1751 notifyInterruptEnabled = 0;
1752 if (notifyInterruptOccurred)
1755 elog(DEBUG1, "EnableNotifyInterrupt: perform async notify");
1757 ProcessIncomingNotify();
1760 elog(DEBUG1, "EnableNotifyInterrupt: done");
1766 * DisableNotifyInterrupt
1768 * This is called by the PostgresMain main loop just after receiving
1769 * a frontend command. Signal handler execution of inbound notifies
1770 * is disabled until the next EnableNotifyInterrupt call.
1772 * The PROCSIG_CATCHUP_INTERRUPT signal handler also needs to call this,
1773 * so as to prevent conflicts if one signal interrupts the other. So we
1774 * must return the previous state of the flag.
1777 DisableNotifyInterrupt(void)
1779 bool result = (notifyInterruptEnabled != 0);
1781 notifyInterruptEnabled = 0;
1787 * Read all pending notifications from the queue, and deliver appropriate
1788 * ones to my frontend. Stop when we reach queue head or an uncommitted
1792 asyncQueueReadAllNotifications(void)
1795 QueuePosition oldpos;
1799 /* page_buffer must be adequately aligned, so use a union */
1802 char buf[QUEUE_PAGESIZE];
1803 AsyncQueueEntry align;
1806 /* Fetch current state */
1807 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1808 /* Assert checks that we have a valid state entry */
1809 Assert(MyProcPid == QUEUE_BACKEND_PID(MyBackendId));
1810 pos = oldpos = QUEUE_BACKEND_POS(MyBackendId);
1812 LWLockRelease(AsyncQueueLock);
1814 if (QUEUE_POS_EQUAL(pos, head))
1816 /* Nothing to do, we have read all notifications already. */
1821 * Note that we deliver everything that we see in the queue and that
1822 * matches our _current_ listening state.
1823 * Especially we do not take into account different commit times.
1824 * Consider the following example:
1826 * Backend 1: Backend 2:
1828 * transaction starts
1831 * transaction starts
1837 * It could happen that backend 2 sees the notification from backend 1 in
1838 * the queue. Even though the notifying transaction committed before
1839 * the listening transaction, we still deliver the notification.
1841 * The idea is that an additional notification does not do any harm, we
1842 * just need to make sure that we do not miss a notification.
1844 * It is possible that we fail while trying to send a message to our
1845 * frontend (for example, because of encoding conversion failure).
1846 * If that happens it is critical that we not try to send the same
1847 * message over and over again. Therefore, we place a PG_TRY block
1848 * here that will forcibly advance our backend position before we lose
1849 * control to an error. (We could alternatively retake AsyncQueueLock
1850 * and move the position before handling each individual message, but
1851 * that seems like too much lock traffic.)
1860 int curpage = QUEUE_POS_PAGE(pos);
1861 int curoffset = QUEUE_POS_OFFSET(pos);
1866 * We copy the data from SLRU into a local buffer, so as to avoid
1867 * holding the AsyncCtlLock while we are examining the entries and
1868 * possibly transmitting them to our frontend. Copy only the part
1869 * of the page we will actually inspect.
1871 slotno = SimpleLruReadPage_ReadOnly(AsyncCtl, curpage,
1872 InvalidTransactionId);
1873 if (curpage == QUEUE_POS_PAGE(head))
1875 /* we only want to read as far as head */
1876 copysize = QUEUE_POS_OFFSET(head) - curoffset;
1878 copysize = 0; /* just for safety */
1882 /* fetch all the rest of the page */
1883 copysize = QUEUE_PAGESIZE - curoffset;
1885 memcpy(page_buffer.buf + curoffset,
1886 AsyncCtl->shared->page_buffer[slotno] + curoffset,
1888 /* Release lock that we got from SimpleLruReadPage_ReadOnly() */
1889 LWLockRelease(AsyncCtlLock);
1892 * Process messages up to the stop position, end of page, or an
1893 * uncommitted message.
1895 * Our stop position is what we found to be the head's position
1896 * when we entered this function. It might have changed already.
1897 * But if it has, we will receive (or have already received and
1898 * queued) another signal and come here again.
1900 * We are not holding AsyncQueueLock here! The queue can only
1901 * extend beyond the head pointer (see above) and we leave our
1902 * backend's pointer where it is so nobody will truncate or
1903 * rewrite pages under us. Especially we don't want to hold a lock
1904 * while sending the notifications to the frontend.
1906 reachedStop = asyncQueueProcessPageEntries(&pos, head,
1908 } while (!reachedStop);
1912 /* Update shared state */
1913 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1914 QUEUE_BACKEND_POS(MyBackendId) = pos;
1915 advanceTail = QUEUE_POS_EQUAL(oldpos, QUEUE_TAIL);
1916 LWLockRelease(AsyncQueueLock);
1918 /* If we were the laziest backend, try to advance the tail pointer */
1920 asyncQueueAdvanceTail();
1926 /* Update shared state */
1927 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1928 QUEUE_BACKEND_POS(MyBackendId) = pos;
1929 advanceTail = QUEUE_POS_EQUAL(oldpos, QUEUE_TAIL);
1930 LWLockRelease(AsyncQueueLock);
1932 /* If we were the laziest backend, try to advance the tail pointer */
1934 asyncQueueAdvanceTail();
1938 * Fetch notifications from the shared queue, beginning at position current,
1939 * and deliver relevant ones to my frontend.
1941 * The current page must have been fetched into page_buffer from shared
1942 * memory. (We could access the page right in shared memory, but that
1943 * would imply holding the AsyncCtlLock throughout this routine.)
1945 * We stop if we reach the "stop" position, or reach a notification from an
1946 * uncommitted transaction, or reach the end of the page.
1948 * The function returns true once we have reached the stop position or an
1949 * uncommitted notification, and false if we have finished with the page.
1950 * In other words: once it returns true there is no need to look further.
1951 * The QueuePosition *current is advanced past all processed messages.
1954 asyncQueueProcessPageEntries(QueuePosition *current,
1958 bool reachedStop = false;
1959 bool reachedEndOfPage;
1960 AsyncQueueEntry *qe;
1964 QueuePosition thisentry = *current;
1966 if (QUEUE_POS_EQUAL(thisentry, stop))
1969 qe = (AsyncQueueEntry *) (page_buffer + QUEUE_POS_OFFSET(thisentry));
1972 * Advance *current over this message, possibly to the next page. As
1973 * noted in the comments for asyncQueueReadAllNotifications, we must
1974 * do this before possibly failing while processing the message.
1976 reachedEndOfPage = asyncQueueAdvance(current, qe->length);
1978 /* Ignore messages destined for other databases */
1979 if (qe->dboid == MyDatabaseId)
1981 if (TransactionIdDidCommit(qe->xid))
1983 /* qe->data is the null-terminated channel name */
1984 char *channel = qe->data;
1986 if (IsListeningOn(channel))
1988 /* payload follows channel name */
1989 char *payload = qe->data + strlen(channel) + 1;
1991 NotifyMyFrontEnd(channel, payload, qe->srcPid);
1994 else if (TransactionIdDidAbort(qe->xid))
1997 * If the source transaction aborted, we just ignore its
2004 * The transaction has neither committed nor aborted so far,
2005 * so we can't process its message yet. Break out of the
2006 * loop, but first back up *current so we will reprocess the
2007 * message next time. (Note: it is unlikely but not
2008 * impossible for TransactionIdDidCommit to fail, so we can't
2009 * really avoid this advance-then-back-up behavior when
2010 * dealing with an uncommitted message.)
2012 *current = thisentry;
2018 /* Loop back if we're not at end of page */
2019 } while (!reachedEndOfPage);
2021 if (QUEUE_POS_EQUAL(*current, stop))
2028 * Advance the shared queue tail variable to the minimum of all the
2029 * per-backend tail pointers. Truncate pg_notify space if possible.
2032 asyncQueueAdvanceTail(void)
2040 LWLockAcquire(AsyncQueueLock, LW_EXCLUSIVE);
2042 for (i = 1; i <= MaxBackends; i++)
2044 if (QUEUE_BACKEND_PID(i) != InvalidPid)
2045 min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
2047 oldtailpage = QUEUE_POS_PAGE(QUEUE_TAIL);
2049 LWLockRelease(AsyncQueueLock);
2052 * We can truncate something if the global tail advanced across an SLRU
2055 * XXX it might be better to truncate only once every several segments, to
2056 * reduce the number of directory scans.
2058 newtailpage = QUEUE_POS_PAGE(min);
2059 boundary = newtailpage - (newtailpage % SLRU_PAGES_PER_SEGMENT);
2060 if (asyncQueuePagePrecedes(oldtailpage, boundary))
2063 * SimpleLruTruncate() will ask for AsyncCtlLock but will also release
2066 SimpleLruTruncate(AsyncCtl, newtailpage);
2071 * ProcessIncomingNotify
2073 * Deal with arriving NOTIFYs from other backends.
2074 * This is called either directly from the PROCSIG_NOTIFY_INTERRUPT
2075 * signal handler, or the next time control reaches the outer idle loop.
2076 * Scan the queue for arriving notifications and report them to my front
2079 * NOTE: since we are outside any transaction, we must create our own.
2082 ProcessIncomingNotify(void)
2084 bool catchup_enabled;
2086 /* We *must* reset the flag */
2087 notifyInterruptOccurred = 0;
2089 /* Do nothing else if we aren't actively listening */
2090 if (listenChannels == NIL)
2093 /* Must prevent catchup interrupt while I am running */
2094 catchup_enabled = DisableCatchupInterrupt();
2097 elog(DEBUG1, "ProcessIncomingNotify");
2099 set_ps_display("notify interrupt", false);
2102 * We must run asyncQueueReadAllNotifications inside a transaction, else
2103 * bad things happen if it gets an error.
2105 StartTransactionCommand();
2107 asyncQueueReadAllNotifications();
2109 CommitTransactionCommand();
2112 * Must flush the notify messages to ensure frontend gets them promptly.
2116 set_ps_display("idle", false);
2119 elog(DEBUG1, "ProcessIncomingNotify: done");
2121 if (catchup_enabled)
2122 EnableCatchupInterrupt();
2126 * Send NOTIFY message to my front end.
2129 NotifyMyFrontEnd(const char *channel, const char *payload, int32 srcPid)
2131 if (whereToSendOutput == DestRemote)
2135 pq_beginmessage(&buf, 'A');
2136 pq_sendint(&buf, srcPid, sizeof(int32));
2137 pq_sendstring(&buf, channel);
2138 if (PG_PROTOCOL_MAJOR(FrontendProtocol) >= 3)
2139 pq_sendstring(&buf, payload);
2140 pq_endmessage(&buf);
2143 * NOTE: we do not do pq_flush() here. For a self-notify, it will
2144 * happen at the end of the transaction, and for incoming notifies
2145 * ProcessIncomingNotify will do it after finding all the notifies.
2149 elog(INFO, "NOTIFY for \"%s\" payload \"%s\"", channel, payload);
2152 /* Does pendingNotifies include the given channel/payload? */
2154 AsyncExistsPendingNotify(const char *channel, const char *payload)
2159 if (pendingNotifies == NIL)
2162 if (payload == NULL)
2166 * We need to append new elements to the end of the list in order to keep
2167 * the order. However, on the other hand we'd like to check the list
2168 * backwards in order to make duplicate-elimination a tad faster when the
2169 * same condition is signaled many times in a row. So as a compromise we
2170 * check the tail element first which we can access directly. If this
2171 * doesn't match, we check the whole list.
2173 * As we are not checking our parents' lists, we can still get duplicates
2174 * in combination with subtransactions, like in:
2183 n = (Notification *) llast(pendingNotifies);
2184 if (strcmp(n->channel, channel) == 0 &&
2185 strcmp(n->payload, payload) == 0)
2188 foreach(p, pendingNotifies)
2190 n = (Notification *) lfirst(p);
2192 if (strcmp(n->channel, channel) == 0 &&
2193 strcmp(n->payload, payload) == 0)
2200 /* Clear the pendingActions and pendingNotifies lists. */
2202 ClearPendingActionsAndNotifies(void)
2205 * We used to have to explicitly deallocate the list members and nodes,
2206 * because they were malloc'd. Now, since we know they are palloc'd in
2207 * CurTransactionContext, we need not do that --- they'll go away
2208 * automatically at transaction exit. We need only reset the list head
2211 pendingActions = NIL;
2212 pendingNotifies = NIL;