1 /*-------------------------------------------------------------------------
4 * Asynchronous notification: NOTIFY, LISTEN, UNLISTEN
6 * Portions Copyright (c) 1996-2015, 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 * sets the process's latch, which triggers the event to be processed
83 * immediately if this backend is idle (i.e., it is waiting for a frontend
84 * command and is not within a transaction block. C.f.
85 * ProcessClientReadInterrupt()). Otherwise the handler may only set a
86 * flag, which will cause the processing to occur just before we next go
89 * Inbound-notify processing consists of reading all of the notifications
90 * that have arrived since scanning last time. We read every notification
91 * until we reach either a notification from an uncommitted transaction or
92 * the head pointer's position. Then we check if we were the laziest
93 * backend: if our pointer is set to the same position as the global tail
94 * pointer is set, then we move the global tail pointer ahead to where the
95 * second-laziest backend is (in general, we take the MIN of the current
96 * head position and all active backends' new tail pointers). Whenever we
97 * move the global tail pointer we also truncate now-unused pages (i.e.,
98 * delete files in pg_notify/ that are no longer used).
100 * An application that listens on the same channel it notifies will get
101 * NOTIFY messages for its own NOTIFYs. These can be ignored, if not useful,
102 * by comparing be_pid in the NOTIFY message to the application's own backend's
103 * PID. (As of FE/BE protocol 2.0, the backend's PID is provided to the
104 * frontend during startup.) The above design guarantees that notifies from
105 * other backends will never be missed by ignoring self-notifies.
107 * The amount of shared memory used for notify management (NUM_ASYNC_BUFFERS)
108 * can be varied without affecting anything but performance. The maximum
109 * amount of notification data that can be queued at one time is determined
110 * by slru.c's wraparound limit; see QUEUE_MAX_PAGE below.
111 *-------------------------------------------------------------------------
114 #include "postgres.h"
120 #include "access/slru.h"
121 #include "access/transam.h"
122 #include "access/xact.h"
123 #include "catalog/pg_database.h"
124 #include "commands/async.h"
126 #include "libpq/libpq.h"
127 #include "libpq/pqformat.h"
128 #include "miscadmin.h"
129 #include "storage/ipc.h"
130 #include "storage/lmgr.h"
131 #include "storage/proc.h"
132 #include "storage/procarray.h"
133 #include "storage/procsignal.h"
134 #include "storage/sinval.h"
135 #include "tcop/tcopprot.h"
136 #include "utils/builtins.h"
137 #include "utils/memutils.h"
138 #include "utils/ps_status.h"
139 #include "utils/timestamp.h"
143 * Maximum size of a NOTIFY payload, including terminating NULL. This
144 * must be kept small enough so that a notification message fits on one
145 * SLRU page. The magic fudge factor here is noncritical as long as it's
146 * more than AsyncQueueEntryEmptySize --- we make it significantly bigger
147 * than that, so changes in that data structure won't affect user-visible
150 #define NOTIFY_PAYLOAD_MAX_LENGTH (BLCKSZ - NAMEDATALEN - 128)
153 * Struct representing an entry in the global notify queue
155 * This struct declaration has the maximal length, but in a real queue entry
156 * the data area is only big enough for the actual channel and payload strings
157 * (each null-terminated). AsyncQueueEntryEmptySize is the minimum possible
158 * entry size, if both channel and payload strings are empty (but note it
159 * doesn't include alignment padding).
161 * The "length" field should always be rounded up to the next QUEUEALIGN
162 * multiple so that all fields are properly aligned.
164 typedef struct AsyncQueueEntry
166 int length; /* total allocated length of entry */
167 Oid dboid; /* sender's database OID */
168 TransactionId xid; /* sender's XID */
169 int32 srcPid; /* sender's PID */
170 char data[NAMEDATALEN + NOTIFY_PAYLOAD_MAX_LENGTH];
173 /* Currently, no field of AsyncQueueEntry requires more than int alignment */
174 #define QUEUEALIGN(len) INTALIGN(len)
176 #define AsyncQueueEntryEmptySize (offsetof(AsyncQueueEntry, data) + 2)
179 * Struct describing a queue position, and assorted macros for working with it
181 typedef struct QueuePosition
183 int page; /* SLRU page number */
184 int offset; /* byte offset within page */
187 #define QUEUE_POS_PAGE(x) ((x).page)
188 #define QUEUE_POS_OFFSET(x) ((x).offset)
190 #define SET_QUEUE_POS(x,y,z) \
196 #define QUEUE_POS_EQUAL(x,y) \
197 ((x).page == (y).page && (x).offset == (y).offset)
199 /* choose logically smaller QueuePosition */
200 #define QUEUE_POS_MIN(x,y) \
201 (asyncQueuePagePrecedes((x).page, (y).page) ? (x) : \
202 (x).page != (y).page ? (y) : \
203 (x).offset < (y).offset ? (x) : (y))
206 * Struct describing a listening backend's status
208 typedef struct QueueBackendStatus
210 int32 pid; /* either a PID or InvalidPid */
211 QueuePosition pos; /* backend has read queue up to here */
212 } QueueBackendStatus;
215 * Shared memory state for LISTEN/NOTIFY (excluding its SLRU stuff)
217 * The AsyncQueueControl structure is protected by the AsyncQueueLock.
219 * When holding the lock in SHARED mode, backends may only inspect their own
220 * entries as well as the head and tail pointers. Consequently we can allow a
221 * backend to update its own record while holding only SHARED lock (since no
222 * other backend will inspect it).
224 * When holding the lock in EXCLUSIVE mode, backends can inspect the entries
225 * of other backends and also change the head and tail pointers.
227 * In order to avoid deadlocks, whenever we need both locks, we always first
228 * get AsyncQueueLock and then AsyncCtlLock.
230 * Each backend uses the backend[] array entry with index equal to its
231 * BackendId (which can range from 1 to MaxBackends). We rely on this to make
232 * SendProcSignal fast.
234 typedef struct AsyncQueueControl
236 QueuePosition head; /* head points to the next free location */
237 QueuePosition tail; /* the global tail is equivalent to the tail
238 * of the "slowest" backend */
239 TimestampTz lastQueueFillWarn; /* time of last queue-full msg */
240 QueueBackendStatus backend[FLEXIBLE_ARRAY_MEMBER];
241 /* backend[0] is not used; used entries are from [1] to [MaxBackends] */
244 static AsyncQueueControl *asyncQueueControl;
246 #define QUEUE_HEAD (asyncQueueControl->head)
247 #define QUEUE_TAIL (asyncQueueControl->tail)
248 #define QUEUE_BACKEND_PID(i) (asyncQueueControl->backend[i].pid)
249 #define QUEUE_BACKEND_POS(i) (asyncQueueControl->backend[i].pos)
252 * The SLRU buffer area through which we access the notification queue
254 static SlruCtlData AsyncCtlData;
256 #define AsyncCtl (&AsyncCtlData)
257 #define QUEUE_PAGESIZE BLCKSZ
258 #define QUEUE_FULL_WARN_INTERVAL 5000 /* warn at most once every 5s */
261 * slru.c currently assumes that all filenames are four characters of hex
262 * digits. That means that we can use segments 0000 through FFFF.
263 * Each segment contains SLRU_PAGES_PER_SEGMENT pages which gives us
264 * the pages from 0 to SLRU_PAGES_PER_SEGMENT * 0x10000 - 1.
266 * It's of course possible to enhance slru.c, but this gives us so much
267 * space already that it doesn't seem worth the trouble.
269 * The most data we can have in the queue at a time is QUEUE_MAX_PAGE/2
270 * pages, because more than that would confuse slru.c into thinking there
271 * was a wraparound condition. With the default BLCKSZ this means there
272 * can be up to 8GB of queued-and-not-read data.
274 * Note: it's possible to redefine QUEUE_MAX_PAGE with a smaller multiple of
275 * SLRU_PAGES_PER_SEGMENT, for easier testing of queue-full behaviour.
277 #define QUEUE_MAX_PAGE (SLRU_PAGES_PER_SEGMENT * 0x10000 - 1)
280 * listenChannels identifies the channels we are actually listening to
281 * (ie, have committed a LISTEN on). It is a simple list of channel names,
282 * allocated in TopMemoryContext.
284 static List *listenChannels = NIL; /* list of C strings */
287 * State for pending LISTEN/UNLISTEN actions consists of an ordered list of
288 * all actions requested in the current transaction. As explained above,
289 * we don't actually change listenChannels until we reach transaction commit.
291 * The list is kept in CurTransactionContext. In subtransactions, each
292 * subtransaction has its own list in its own CurTransactionContext, but
293 * successful subtransactions attach their lists to their parent's list.
294 * Failed subtransactions simply discard their lists.
305 ListenActionKind action;
306 char channel[FLEXIBLE_ARRAY_MEMBER]; /* nul-terminated string */
309 static List *pendingActions = NIL; /* list of ListenAction */
311 static List *upperPendingActions = NIL; /* list of upper-xact lists */
314 * State for outbound notifies consists of a list of all channels+payloads
315 * NOTIFYed in the current transaction. We do not actually perform a NOTIFY
316 * until and unless the transaction commits. pendingNotifies is NIL if no
317 * NOTIFYs have been done in the current transaction.
319 * The list is kept in CurTransactionContext. In subtransactions, each
320 * subtransaction has its own list in its own CurTransactionContext, but
321 * successful subtransactions attach their lists to their parent's list.
322 * Failed subtransactions simply discard their lists.
324 * Note: the action and notify lists do not interact within a transaction.
325 * In particular, if a transaction does NOTIFY and then LISTEN on the same
326 * condition name, it will get a self-notify at commit. This is a bit odd
327 * but is consistent with our historical behavior.
329 typedef struct Notification
331 char *channel; /* channel name */
332 char *payload; /* payload string (can be empty) */
335 static List *pendingNotifies = NIL; /* list of Notifications */
337 static List *upperPendingNotifies = NIL; /* list of upper-xact lists */
340 * Inbound notifications are initially processed by HandleNotifyInterrupt(),
341 * called from inside a signal handler. That just sets the
342 * notifyInterruptPending flag and sets the process
343 * latch. ProcessNotifyInterrupt() will then be called whenever it's safe to
344 * actually deal with the interrupt.
346 volatile sig_atomic_t notifyInterruptPending = false;
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(volatile QueuePosition *position, int entryLength);
372 static void asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe);
373 static ListCell *asyncQueueAddEntries(ListCell *nextNotify);
374 static double asyncQueueUsage(void);
375 static void asyncQueueFillWarning(void);
376 static bool SignalBackends(void);
377 static void asyncQueueReadAllNotifications(void);
378 static bool asyncQueueProcessPageEntries(volatile QueuePosition *current,
381 static void asyncQueueAdvanceTail(void);
382 static void ProcessIncomingNotify(void);
383 static void NotifyMyFrontEnd(const char *channel,
386 static bool AsyncExistsPendingNotify(const char *channel, const char *payload);
387 static void ClearPendingActionsAndNotifies(void);
390 * We will work on the page range of 0..QUEUE_MAX_PAGE.
393 asyncQueuePagePrecedes(int p, int q)
398 * We have to compare modulo (QUEUE_MAX_PAGE+1)/2. Both inputs should be
399 * in the range 0..QUEUE_MAX_PAGE.
401 Assert(p >= 0 && p <= QUEUE_MAX_PAGE);
402 Assert(q >= 0 && q <= QUEUE_MAX_PAGE);
405 if (diff >= ((QUEUE_MAX_PAGE + 1) / 2))
406 diff -= QUEUE_MAX_PAGE + 1;
407 else if (diff < -((QUEUE_MAX_PAGE + 1) / 2))
408 diff += QUEUE_MAX_PAGE + 1;
413 * Report space needed for our shared memory area
420 /* This had better match AsyncShmemInit */
421 size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
422 size = add_size(size, offsetof(AsyncQueueControl, backend));
424 size = add_size(size, SimpleLruShmemSize(NUM_ASYNC_BUFFERS, 0));
430 * Initialize our shared memory area
440 * Create or attach to the AsyncQueueControl structure.
442 * The used entries in the backend[] array run from 1 to MaxBackends; the
443 * zero'th entry is unused but must be allocated.
445 size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
446 size = add_size(size, offsetof(AsyncQueueControl, backend));
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(offsetof(ListenAction, channel) +
609 strlen(channel) + 1);
610 actrec->action = action;
611 strcpy(actrec->channel, channel);
613 pendingActions = lappend(pendingActions, actrec);
615 MemoryContextSwitchTo(oldcontext);
621 * This is executed by the SQL listen command.
624 Async_Listen(const char *channel)
627 elog(DEBUG1, "Async_Listen(%s,%d)", channel, MyProcPid);
629 queue_listen(LISTEN_LISTEN, channel);
635 * This is executed by the SQL unlisten command.
638 Async_Unlisten(const char *channel)
641 elog(DEBUG1, "Async_Unlisten(%s,%d)", channel, MyProcPid);
643 /* If we couldn't possibly be listening, no need to queue anything */
644 if (pendingActions == NIL && !unlistenExitRegistered)
647 queue_listen(LISTEN_UNLISTEN, channel);
653 * This is invoked by UNLISTEN * command, and also at backend exit.
656 Async_UnlistenAll(void)
659 elog(DEBUG1, "Async_UnlistenAll(%d)", MyProcPid);
661 /* If we couldn't possibly be listening, no need to queue anything */
662 if (pendingActions == NIL && !unlistenExitRegistered)
665 queue_listen(LISTEN_UNLISTEN_ALL, "");
669 * SQL function: return a set of the channel names this backend is actively
672 * Note: this coding relies on the fact that the listenChannels list cannot
673 * change within a transaction.
676 pg_listening_channels(PG_FUNCTION_ARGS)
678 FuncCallContext *funcctx;
681 /* stuff done only on the first call of the function */
682 if (SRF_IS_FIRSTCALL())
684 MemoryContext oldcontext;
686 /* create a function context for cross-call persistence */
687 funcctx = SRF_FIRSTCALL_INIT();
689 /* switch to memory context appropriate for multiple function calls */
690 oldcontext = MemoryContextSwitchTo(funcctx->multi_call_memory_ctx);
692 /* allocate memory for user context */
693 lcp = (ListCell **) palloc(sizeof(ListCell *));
694 *lcp = list_head(listenChannels);
695 funcctx->user_fctx = (void *) lcp;
697 MemoryContextSwitchTo(oldcontext);
700 /* stuff done on every call of the function */
701 funcctx = SRF_PERCALL_SETUP();
702 lcp = (ListCell **) funcctx->user_fctx;
706 char *channel = (char *) lfirst(*lcp);
709 SRF_RETURN_NEXT(funcctx, CStringGetTextDatum(channel));
712 SRF_RETURN_DONE(funcctx);
716 * Async_UnlistenOnExit
718 * This is executed at backend exit if we have done any LISTENs in this
719 * backend. It might not be necessary anymore, if the user UNLISTENed
720 * everything, but we don't try to detect that case.
723 Async_UnlistenOnExit(int code, Datum arg)
725 Exec_UnlistenAllCommit();
726 asyncQueueUnregister();
732 * This is called at the prepare phase of a two-phase
733 * transaction. Save the state for possible commit later.
736 AtPrepare_Notify(void)
738 /* It's not allowed to have any pending LISTEN/UNLISTEN/NOTIFY actions */
739 if (pendingActions || pendingNotifies)
741 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
742 errmsg("cannot PREPARE a transaction that has executed LISTEN, UNLISTEN, or NOTIFY")));
748 * This is called at transaction commit, before actually committing to
751 * If there are pending LISTEN actions, make sure we are listed in the
752 * shared-memory listener array. This must happen before commit to
753 * ensure we don't miss any notifies from transactions that commit
756 * If there are outbound notify requests in the pendingNotifies list,
757 * add them to the global queue. We do that before commit so that
758 * we can still throw error if we run out of queue space.
761 PreCommit_Notify(void)
765 if (pendingActions == NIL && pendingNotifies == NIL)
766 return; /* no relevant statements in this xact */
769 elog(DEBUG1, "PreCommit_Notify");
771 /* Preflight for any pending listen/unlisten actions */
772 foreach(p, pendingActions)
774 ListenAction *actrec = (ListenAction *) lfirst(p);
776 switch (actrec->action)
779 Exec_ListenPreCommit();
781 case LISTEN_UNLISTEN:
782 /* there is no Exec_UnlistenPreCommit() */
784 case LISTEN_UNLISTEN_ALL:
785 /* there is no Exec_UnlistenAllPreCommit() */
790 /* Queue any pending notifies */
793 ListCell *nextNotify;
796 * Make sure that we have an XID assigned to the current transaction.
797 * GetCurrentTransactionId is cheap if we already have an XID, but not
798 * so cheap if we don't, and we'd prefer not to do that work while
799 * holding AsyncQueueLock.
801 (void) GetCurrentTransactionId();
804 * Serialize writers by acquiring a special lock that we hold till
805 * after commit. This ensures that queue entries appear in commit
806 * order, and in particular that there are never uncommitted queue
807 * entries ahead of committed ones, so an uncommitted transaction
808 * can't block delivery of deliverable notifications.
810 * We use a heavyweight lock so that it'll automatically be released
811 * after either commit or abort. This also allows deadlocks to be
812 * detected, though really a deadlock shouldn't be possible here.
814 * The lock is on "database 0", which is pretty ugly but it doesn't
815 * seem worth inventing a special locktag category just for this.
816 * (Historical note: before PG 9.0, a similar lock on "database 0" was
817 * used by the flatfiles mechanism.)
819 LockSharedObject(DatabaseRelationId, InvalidOid, 0,
820 AccessExclusiveLock);
822 /* Now push the notifications into the queue */
823 backendHasSentNotifications = true;
825 nextNotify = list_head(pendingNotifies);
826 while (nextNotify != NULL)
829 * Add the pending notifications to the queue. We acquire and
830 * release AsyncQueueLock once per page, which might be overkill
831 * but it does allow readers to get in while we're doing this.
833 * A full queue is very uncommon and should really not happen,
834 * given that we have so much space available in the SLRU pages.
835 * Nevertheless we need to deal with this possibility. Note that
836 * when we get here we are in the process of committing our
837 * transaction, but we have not yet committed to clog, so at this
838 * point in time we can still roll the transaction back.
840 LWLockAcquire(AsyncQueueLock, LW_EXCLUSIVE);
841 asyncQueueFillWarning();
842 if (asyncQueueIsFull())
844 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
845 errmsg("too many notifications in the NOTIFY queue")));
846 nextNotify = asyncQueueAddEntries(nextNotify);
847 LWLockRelease(AsyncQueueLock);
855 * This is called at transaction commit, after committing to clog.
857 * Update listenChannels and clear transaction-local state.
860 AtCommit_Notify(void)
865 * Allow transactions that have not executed LISTEN/UNLISTEN/NOTIFY to
866 * return as soon as possible
868 if (!pendingActions && !pendingNotifies)
872 elog(DEBUG1, "AtCommit_Notify");
874 /* Perform any pending listen/unlisten actions */
875 foreach(p, pendingActions)
877 ListenAction *actrec = (ListenAction *) lfirst(p);
879 switch (actrec->action)
882 Exec_ListenCommit(actrec->channel);
884 case LISTEN_UNLISTEN:
885 Exec_UnlistenCommit(actrec->channel);
887 case LISTEN_UNLISTEN_ALL:
888 Exec_UnlistenAllCommit();
893 /* If no longer listening to anything, get out of listener array */
894 if (amRegisteredListener && listenChannels == NIL)
895 asyncQueueUnregister();
898 ClearPendingActionsAndNotifies();
902 * Exec_ListenPreCommit --- subroutine for PreCommit_Notify
904 * This function must make sure we are ready to catch any incoming messages.
907 Exec_ListenPreCommit(void)
910 * Nothing to do if we are already listening to something, nor if we
911 * already ran this routine in this transaction.
913 if (amRegisteredListener)
917 elog(DEBUG1, "Exec_ListenPreCommit(%d)", MyProcPid);
920 * Before registering, make sure we will unlisten before dying. (Note:
921 * this action does not get undone if we abort later.)
923 if (!unlistenExitRegistered)
925 before_shmem_exit(Async_UnlistenOnExit, 0);
926 unlistenExitRegistered = true;
930 * This is our first LISTEN, so establish our pointer.
932 * We set our pointer to the global tail pointer and then move it forward
933 * over already-committed notifications. This ensures we cannot miss any
934 * not-yet-committed notifications. We might get a few more but that
937 LWLockAcquire(AsyncQueueLock, LW_SHARED);
938 QUEUE_BACKEND_POS(MyBackendId) = QUEUE_TAIL;
939 QUEUE_BACKEND_PID(MyBackendId) = MyProcPid;
940 LWLockRelease(AsyncQueueLock);
942 /* Now we are listed in the global array, so remember we're listening */
943 amRegisteredListener = true;
946 * Try to move our pointer forward as far as possible. This will skip over
947 * already-committed notifications. Still, we could get notifications that
948 * have already committed before we started to LISTEN.
950 * Note that we are not yet listening on anything, so we won't deliver any
951 * notification to the frontend.
953 * This will also advance the global tail pointer if possible.
955 asyncQueueReadAllNotifications();
959 * Exec_ListenCommit --- subroutine for AtCommit_Notify
961 * Add the channel to the list of channels we are listening on.
964 Exec_ListenCommit(const char *channel)
966 MemoryContext oldcontext;
968 /* Do nothing if we are already listening on this channel */
969 if (IsListeningOn(channel))
973 * Add the new channel name to listenChannels.
975 * XXX It is theoretically possible to get an out-of-memory failure here,
976 * which would be bad because we already committed. For the moment it
977 * doesn't seem worth trying to guard against that, but maybe improve this
980 oldcontext = MemoryContextSwitchTo(TopMemoryContext);
981 listenChannels = lappend(listenChannels, pstrdup(channel));
982 MemoryContextSwitchTo(oldcontext);
986 * Exec_UnlistenCommit --- subroutine for AtCommit_Notify
988 * Remove the specified channel name from listenChannels.
991 Exec_UnlistenCommit(const char *channel)
997 elog(DEBUG1, "Exec_UnlistenCommit(%s,%d)", channel, MyProcPid);
1000 foreach(q, listenChannels)
1002 char *lchan = (char *) lfirst(q);
1004 if (strcmp(lchan, channel) == 0)
1006 listenChannels = list_delete_cell(listenChannels, q, prev);
1014 * We do not complain about unlistening something not being listened;
1020 * Exec_UnlistenAllCommit --- subroutine for AtCommit_Notify
1022 * Unlisten on all channels for this backend.
1025 Exec_UnlistenAllCommit(void)
1028 elog(DEBUG1, "Exec_UnlistenAllCommit(%d)", MyProcPid);
1030 list_free_deep(listenChannels);
1031 listenChannels = NIL;
1035 * ProcessCompletedNotifies --- send out signals and self-notifies
1037 * This is called from postgres.c just before going idle at the completion
1038 * of a transaction. If we issued any notifications in the just-completed
1039 * transaction, send signals to other backends to process them, and also
1040 * process the queue ourselves to send messages to our own frontend.
1042 * The reason that this is not done in AtCommit_Notify is that there is
1043 * a nonzero chance of errors here (for example, encoding conversion errors
1044 * while trying to format messages to our frontend). An error during
1045 * AtCommit_Notify would be a PANIC condition. The timing is also arranged
1046 * to ensure that a transaction's self-notifies are delivered to the frontend
1047 * before it gets the terminating ReadyForQuery message.
1049 * Note that we send signals and process the queue even if the transaction
1050 * eventually aborted. This is because we need to clean out whatever got
1051 * added to the queue.
1053 * NOTE: we are outside of any transaction here.
1056 ProcessCompletedNotifies(void)
1058 MemoryContext caller_context;
1061 /* Nothing to do if we didn't send any notifications */
1062 if (!backendHasSentNotifications)
1066 * We reset the flag immediately; otherwise, if any sort of error occurs
1067 * below, we'd be locked up in an infinite loop, because control will come
1068 * right back here after error cleanup.
1070 backendHasSentNotifications = false;
1073 * We must preserve the caller's memory context (probably MessageContext)
1074 * across the transaction we do here.
1076 caller_context = CurrentMemoryContext;
1079 elog(DEBUG1, "ProcessCompletedNotifies");
1082 * We must run asyncQueueReadAllNotifications inside a transaction, else
1083 * bad things happen if it gets an error.
1085 StartTransactionCommand();
1087 /* Send signals to other backends */
1088 signalled = SignalBackends();
1090 if (listenChannels != NIL)
1092 /* Read the queue ourselves, and send relevant stuff to the frontend */
1093 asyncQueueReadAllNotifications();
1095 else if (!signalled)
1098 * If we found no other listening backends, and we aren't listening
1099 * ourselves, then we must execute asyncQueueAdvanceTail to flush the
1100 * queue, because ain't nobody else gonna do it. This prevents queue
1101 * overflow when we're sending useless notifies to nobody. (A new
1102 * listener could have joined since we looked, but if so this is
1105 asyncQueueAdvanceTail();
1108 CommitTransactionCommand();
1110 MemoryContextSwitchTo(caller_context);
1112 /* We don't need pq_flush() here since postgres.c will do one shortly */
1116 * Test whether we are actively listening on the given channel name.
1118 * Note: this function is executed for every notification found in the queue.
1119 * Perhaps it is worth further optimization, eg convert the list to a sorted
1120 * array so we can binary-search it. In practice the list is likely to be
1121 * fairly short, though.
1124 IsListeningOn(const char *channel)
1128 foreach(p, listenChannels)
1130 char *lchan = (char *) lfirst(p);
1132 if (strcmp(lchan, channel) == 0)
1139 * Remove our entry from the listeners array when we are no longer listening
1140 * on any channel. NB: must not fail if we're already not listening.
1143 asyncQueueUnregister(void)
1147 Assert(listenChannels == NIL); /* else caller error */
1149 if (!amRegisteredListener) /* nothing to do */
1152 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1153 /* check if entry is valid and oldest ... */
1154 advanceTail = (MyProcPid == QUEUE_BACKEND_PID(MyBackendId)) &&
1155 QUEUE_POS_EQUAL(QUEUE_BACKEND_POS(MyBackendId), QUEUE_TAIL);
1156 /* ... then mark it invalid */
1157 QUEUE_BACKEND_PID(MyBackendId) = InvalidPid;
1158 LWLockRelease(AsyncQueueLock);
1160 /* mark ourselves as no longer listed in the global array */
1161 amRegisteredListener = false;
1163 /* If we were the laziest backend, try to advance the tail pointer */
1165 asyncQueueAdvanceTail();
1169 * Test whether there is room to insert more notification messages.
1171 * Caller must hold at least shared AsyncQueueLock.
1174 asyncQueueIsFull(void)
1180 * The queue is full if creating a new head page would create a page that
1181 * logically precedes the current global tail pointer, ie, the head
1182 * pointer would wrap around compared to the tail. We cannot create such
1183 * a head page for fear of confusing slru.c. For safety we round the tail
1184 * pointer back to a segment boundary (compare the truncation logic in
1185 * asyncQueueAdvanceTail).
1187 * Note that this test is *not* dependent on how much space there is on
1188 * the current head page. This is necessary because asyncQueueAddEntries
1189 * might try to create the next head page in any case.
1191 nexthead = QUEUE_POS_PAGE(QUEUE_HEAD) + 1;
1192 if (nexthead > QUEUE_MAX_PAGE)
1193 nexthead = 0; /* wrap around */
1194 boundary = QUEUE_POS_PAGE(QUEUE_TAIL);
1195 boundary -= boundary % SLRU_PAGES_PER_SEGMENT;
1196 return asyncQueuePagePrecedes(nexthead, boundary);
1200 * Advance the QueuePosition to the next entry, assuming that the current
1201 * entry is of length entryLength. If we jump to a new page the function
1202 * returns true, else false.
1205 asyncQueueAdvance(volatile QueuePosition *position, int entryLength)
1207 int pageno = QUEUE_POS_PAGE(*position);
1208 int offset = QUEUE_POS_OFFSET(*position);
1209 bool pageJump = false;
1212 * Move to the next writing position: First jump over what we have just
1215 offset += entryLength;
1216 Assert(offset <= QUEUE_PAGESIZE);
1219 * In a second step check if another entry can possibly be written to the
1220 * page. If so, stay here, we have reached the next position. If not, then
1221 * we need to move on to the next page.
1223 if (offset + QUEUEALIGN(AsyncQueueEntryEmptySize) > QUEUE_PAGESIZE)
1226 if (pageno > QUEUE_MAX_PAGE)
1227 pageno = 0; /* wrap around */
1232 SET_QUEUE_POS(*position, pageno, offset);
1237 * Fill the AsyncQueueEntry at *qe with an outbound notification message.
1240 asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe)
1242 size_t channellen = strlen(n->channel);
1243 size_t payloadlen = strlen(n->payload);
1246 Assert(channellen < NAMEDATALEN);
1247 Assert(payloadlen < NOTIFY_PAYLOAD_MAX_LENGTH);
1249 /* The terminators are already included in AsyncQueueEntryEmptySize */
1250 entryLength = AsyncQueueEntryEmptySize + payloadlen + channellen;
1251 entryLength = QUEUEALIGN(entryLength);
1252 qe->length = entryLength;
1253 qe->dboid = MyDatabaseId;
1254 qe->xid = GetCurrentTransactionId();
1255 qe->srcPid = MyProcPid;
1256 memcpy(qe->data, n->channel, channellen + 1);
1257 memcpy(qe->data + channellen + 1, n->payload, payloadlen + 1);
1261 * Add pending notifications to the queue.
1263 * We go page by page here, i.e. we stop once we have to go to a new page but
1264 * we will be called again and then fill that next page. If an entry does not
1265 * fit into the current page, we write a dummy entry with an InvalidOid as the
1266 * database OID in order to fill the page. So every page is always used up to
1267 * the last byte which simplifies reading the page later.
1269 * We are passed the list cell containing the next notification to write
1270 * and return the first still-unwritten cell back. Eventually we will return
1271 * NULL indicating all is done.
1273 * We are holding AsyncQueueLock already from the caller and grab AsyncCtlLock
1274 * locally in this function.
1277 asyncQueueAddEntries(ListCell *nextNotify)
1280 QueuePosition queue_head;
1285 /* We hold both AsyncQueueLock and AsyncCtlLock during this operation */
1286 LWLockAcquire(AsyncCtlLock, LW_EXCLUSIVE);
1289 * We work with a local copy of QUEUE_HEAD, which we write back to shared
1290 * memory upon exiting. The reason for this is that if we have to advance
1291 * to a new page, SimpleLruZeroPage might fail (out of disk space, for
1292 * instance), and we must not advance QUEUE_HEAD if it does. (Otherwise,
1293 * subsequent insertions would try to put entries into a page that slru.c
1294 * thinks doesn't exist yet.) So, use a local position variable. Note
1295 * that if we do fail, any already-inserted queue entries are forgotten;
1296 * this is okay, since they'd be useless anyway after our transaction
1299 queue_head = QUEUE_HEAD;
1301 /* Fetch the current page */
1302 pageno = QUEUE_POS_PAGE(queue_head);
1303 slotno = SimpleLruReadPage(AsyncCtl, pageno, true, InvalidTransactionId);
1304 /* Note we mark the page dirty before writing in it */
1305 AsyncCtl->shared->page_dirty[slotno] = true;
1307 while (nextNotify != NULL)
1309 Notification *n = (Notification *) lfirst(nextNotify);
1311 /* Construct a valid queue entry in local variable qe */
1312 asyncQueueNotificationToEntry(n, &qe);
1314 offset = QUEUE_POS_OFFSET(queue_head);
1316 /* Check whether the entry really fits on the current page */
1317 if (offset + qe.length <= QUEUE_PAGESIZE)
1319 /* OK, so advance nextNotify past this item */
1320 nextNotify = lnext(nextNotify);
1325 * Write a dummy entry to fill up the page. Actually readers will
1326 * only check dboid and since it won't match any reader's database
1327 * OID, they will ignore this entry and move on.
1329 qe.length = QUEUE_PAGESIZE - offset;
1330 qe.dboid = InvalidOid;
1331 qe.data[0] = '\0'; /* empty channel */
1332 qe.data[1] = '\0'; /* empty payload */
1335 /* Now copy qe into the shared buffer page */
1336 memcpy(AsyncCtl->shared->page_buffer[slotno] + offset,
1340 /* Advance queue_head appropriately, and detect if page is full */
1341 if (asyncQueueAdvance(&(queue_head), qe.length))
1344 * Page is full, so we're done here, but first fill the next page
1345 * with zeroes. The reason to do this is to ensure that slru.c's
1346 * idea of the head page is always the same as ours, which avoids
1347 * boundary problems in SimpleLruTruncate. The test in
1348 * asyncQueueIsFull() ensured that there is room to create this
1349 * page without overrunning the queue.
1351 slotno = SimpleLruZeroPage(AsyncCtl, QUEUE_POS_PAGE(queue_head));
1352 /* And exit the loop */
1357 /* Success, so update the global QUEUE_HEAD */
1358 QUEUE_HEAD = queue_head;
1360 LWLockRelease(AsyncCtlLock);
1366 * SQL function to return the fraction of the notification queue currently
1370 pg_notification_queue_usage(PG_FUNCTION_ARGS)
1374 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1375 usage = asyncQueueUsage();
1376 LWLockRelease(AsyncQueueLock);
1378 PG_RETURN_FLOAT8(usage);
1382 * Return the fraction of the queue that is currently occupied.
1384 * The caller must hold AysncQueueLock in (at least) shared mode.
1387 asyncQueueUsage(void)
1389 int headPage = QUEUE_POS_PAGE(QUEUE_HEAD);
1390 int tailPage = QUEUE_POS_PAGE(QUEUE_TAIL);
1393 occupied = headPage - tailPage;
1396 return (double) 0; /* fast exit for common case */
1400 /* head has wrapped around, tail not yet */
1401 occupied += QUEUE_MAX_PAGE + 1;
1404 return (double) occupied / (double) ((QUEUE_MAX_PAGE + 1) / 2);
1408 * Check whether the queue is at least half full, and emit a warning if so.
1410 * This is unlikely given the size of the queue, but possible.
1411 * The warnings show up at most once every QUEUE_FULL_WARN_INTERVAL.
1413 * Caller must hold exclusive AsyncQueueLock.
1416 asyncQueueFillWarning(void)
1421 fillDegree = asyncQueueUsage();
1422 if (fillDegree < 0.5)
1425 t = GetCurrentTimestamp();
1427 if (TimestampDifferenceExceeds(asyncQueueControl->lastQueueFillWarn,
1428 t, QUEUE_FULL_WARN_INTERVAL))
1430 QueuePosition min = QUEUE_HEAD;
1431 int32 minPid = InvalidPid;
1434 for (i = 1; i <= MaxBackends; i++)
1436 if (QUEUE_BACKEND_PID(i) != InvalidPid)
1438 min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
1439 if (QUEUE_POS_EQUAL(min, QUEUE_BACKEND_POS(i)))
1440 minPid = QUEUE_BACKEND_PID(i);
1445 (errmsg("NOTIFY queue is %.0f%% full", fillDegree * 100),
1446 (minPid != InvalidPid ?
1447 errdetail("The server process with PID %d is among those with the oldest transactions.", minPid)
1449 (minPid != InvalidPid ?
1450 errhint("The NOTIFY queue cannot be emptied until that process ends its current transaction.")
1453 asyncQueueControl->lastQueueFillWarn = t;
1458 * Send signals to all listening backends (except our own).
1460 * Returns true if we sent at least one signal.
1462 * Since we need EXCLUSIVE lock anyway we also check the position of the other
1463 * backends and in case one is already up-to-date we don't signal it.
1464 * This can happen if concurrent notifying transactions have sent a signal and
1465 * the signaled backend has read the other notifications and ours in the same
1468 * Since we know the BackendId and the Pid the signalling is quite cheap.
1471 SignalBackends(void)
1473 bool signalled = false;
1481 * Identify all backends that are listening and not already up-to-date. We
1482 * don't want to send signals while holding the AsyncQueueLock, so we just
1483 * build a list of target PIDs.
1485 * XXX in principle these pallocs could fail, which would be bad. Maybe
1486 * preallocate the arrays? But in practice this is only run in trivial
1487 * transactions, so there should surely be space available.
1489 pids = (int32 *) palloc(MaxBackends * sizeof(int32));
1490 ids = (BackendId *) palloc(MaxBackends * sizeof(BackendId));
1493 LWLockAcquire(AsyncQueueLock, LW_EXCLUSIVE);
1494 for (i = 1; i <= MaxBackends; i++)
1496 pid = QUEUE_BACKEND_PID(i);
1497 if (pid != InvalidPid && pid != MyProcPid)
1499 QueuePosition pos = QUEUE_BACKEND_POS(i);
1501 if (!QUEUE_POS_EQUAL(pos, QUEUE_HEAD))
1509 LWLockRelease(AsyncQueueLock);
1511 /* Now send signals */
1512 for (i = 0; i < count; i++)
1517 * Note: assuming things aren't broken, a signal failure here could
1518 * only occur if the target backend exited since we released
1519 * AsyncQueueLock; which is unlikely but certainly possible. So we
1520 * just log a low-level debug message if it happens.
1522 if (SendProcSignal(pid, PROCSIG_NOTIFY_INTERRUPT, ids[i]) < 0)
1523 elog(DEBUG3, "could not signal backend with PID %d: %m", pid);
1537 * This is called at transaction abort.
1539 * Gets rid of pending actions and outbound notifies that we would have
1540 * executed if the transaction got committed.
1543 AtAbort_Notify(void)
1546 * If we LISTEN but then roll back the transaction after PreCommit_Notify,
1547 * we have registered as a listener but have not made any entry in
1548 * listenChannels. In that case, deregister again.
1550 if (amRegisteredListener && listenChannels == NIL)
1551 asyncQueueUnregister();
1554 ClearPendingActionsAndNotifies();
1558 * AtSubStart_Notify() --- Take care of subtransaction start.
1560 * Push empty state for the new subtransaction.
1563 AtSubStart_Notify(void)
1565 MemoryContext old_cxt;
1567 /* Keep the list-of-lists in TopTransactionContext for simplicity */
1568 old_cxt = MemoryContextSwitchTo(TopTransactionContext);
1570 upperPendingActions = lcons(pendingActions, upperPendingActions);
1572 Assert(list_length(upperPendingActions) ==
1573 GetCurrentTransactionNestLevel() - 1);
1575 pendingActions = NIL;
1577 upperPendingNotifies = lcons(pendingNotifies, upperPendingNotifies);
1579 Assert(list_length(upperPendingNotifies) ==
1580 GetCurrentTransactionNestLevel() - 1);
1582 pendingNotifies = NIL;
1584 MemoryContextSwitchTo(old_cxt);
1588 * AtSubCommit_Notify() --- Take care of subtransaction commit.
1590 * Reassign all items in the pending lists to the parent transaction.
1593 AtSubCommit_Notify(void)
1595 List *parentPendingActions;
1596 List *parentPendingNotifies;
1598 parentPendingActions = (List *) linitial(upperPendingActions);
1599 upperPendingActions = list_delete_first(upperPendingActions);
1601 Assert(list_length(upperPendingActions) ==
1602 GetCurrentTransactionNestLevel() - 2);
1605 * Mustn't try to eliminate duplicates here --- see queue_listen()
1607 pendingActions = list_concat(parentPendingActions, pendingActions);
1609 parentPendingNotifies = (List *) linitial(upperPendingNotifies);
1610 upperPendingNotifies = list_delete_first(upperPendingNotifies);
1612 Assert(list_length(upperPendingNotifies) ==
1613 GetCurrentTransactionNestLevel() - 2);
1616 * We could try to eliminate duplicates here, but it seems not worthwhile.
1618 pendingNotifies = list_concat(parentPendingNotifies, pendingNotifies);
1622 * AtSubAbort_Notify() --- Take care of subtransaction abort.
1625 AtSubAbort_Notify(void)
1627 int my_level = GetCurrentTransactionNestLevel();
1630 * All we have to do is pop the stack --- the actions/notifies made in
1631 * this subxact are no longer interesting, and the space will be freed
1632 * when CurTransactionContext is recycled.
1634 * This routine could be called more than once at a given nesting level if
1635 * there is trouble during subxact abort. Avoid dumping core by using
1636 * GetCurrentTransactionNestLevel as the indicator of how far we need to
1639 while (list_length(upperPendingActions) > my_level - 2)
1641 pendingActions = (List *) linitial(upperPendingActions);
1642 upperPendingActions = list_delete_first(upperPendingActions);
1645 while (list_length(upperPendingNotifies) > my_level - 2)
1647 pendingNotifies = (List *) linitial(upperPendingNotifies);
1648 upperPendingNotifies = list_delete_first(upperPendingNotifies);
1653 * HandleNotifyInterrupt
1655 * Signal handler portion of interrupt handling. Let the backend know
1656 * that there's a pending notify interrupt. If we're currently reading
1657 * from the client, this will interrupt the read and
1658 * ProcessClientReadInterrupt() will call ProcessNotifyInterrupt().
1661 HandleNotifyInterrupt(void)
1664 * Note: this is called by a SIGNAL HANDLER. You must be very wary what
1668 /* signal that work needs to be done */
1669 notifyInterruptPending = true;
1671 /* make sure the event is processed in due course */
1676 * ProcessNotifyInterrupt
1678 * This is called just after waiting for a frontend command. If a
1679 * interrupt arrives (via HandleNotifyInterrupt()) while reading, the
1680 * read will be interrupted via the process's latch, and this routine
1681 * will get called. If we are truly idle (ie, *not* inside a transaction
1682 * block), process the incoming notifies.
1685 ProcessNotifyInterrupt(void)
1687 if (IsTransactionOrTransactionBlock())
1688 return; /* not really idle */
1690 while (notifyInterruptPending)
1691 ProcessIncomingNotify();
1696 * Read all pending notifications from the queue, and deliver appropriate
1697 * ones to my frontend. Stop when we reach queue head or an uncommitted
1701 asyncQueueReadAllNotifications(void)
1703 volatile QueuePosition pos;
1704 QueuePosition oldpos;
1708 /* page_buffer must be adequately aligned, so use a union */
1711 char buf[QUEUE_PAGESIZE];
1712 AsyncQueueEntry align;
1715 /* Fetch current state */
1716 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1717 /* Assert checks that we have a valid state entry */
1718 Assert(MyProcPid == QUEUE_BACKEND_PID(MyBackendId));
1719 pos = oldpos = QUEUE_BACKEND_POS(MyBackendId);
1721 LWLockRelease(AsyncQueueLock);
1723 if (QUEUE_POS_EQUAL(pos, head))
1725 /* Nothing to do, we have read all notifications already. */
1730 * Note that we deliver everything that we see in the queue and that
1731 * matches our _current_ listening state.
1732 * Especially we do not take into account different commit times.
1733 * Consider the following example:
1735 * Backend 1: Backend 2:
1737 * transaction starts
1740 * transaction starts
1746 * It could happen that backend 2 sees the notification from backend 1 in
1747 * the queue. Even though the notifying transaction committed before
1748 * the listening transaction, we still deliver the notification.
1750 * The idea is that an additional notification does not do any harm, we
1751 * just need to make sure that we do not miss a notification.
1753 * It is possible that we fail while trying to send a message to our
1754 * frontend (for example, because of encoding conversion failure).
1755 * If that happens it is critical that we not try to send the same
1756 * message over and over again. Therefore, we place a PG_TRY block
1757 * here that will forcibly advance our backend position before we lose
1758 * control to an error. (We could alternatively retake AsyncQueueLock
1759 * and move the position before handling each individual message, but
1760 * that seems like too much lock traffic.)
1769 int curpage = QUEUE_POS_PAGE(pos);
1770 int curoffset = QUEUE_POS_OFFSET(pos);
1775 * We copy the data from SLRU into a local buffer, so as to avoid
1776 * holding the AsyncCtlLock while we are examining the entries and
1777 * possibly transmitting them to our frontend. Copy only the part
1778 * of the page we will actually inspect.
1780 slotno = SimpleLruReadPage_ReadOnly(AsyncCtl, curpage,
1781 InvalidTransactionId);
1782 if (curpage == QUEUE_POS_PAGE(head))
1784 /* we only want to read as far as head */
1785 copysize = QUEUE_POS_OFFSET(head) - curoffset;
1787 copysize = 0; /* just for safety */
1791 /* fetch all the rest of the page */
1792 copysize = QUEUE_PAGESIZE - curoffset;
1794 memcpy(page_buffer.buf + curoffset,
1795 AsyncCtl->shared->page_buffer[slotno] + curoffset,
1797 /* Release lock that we got from SimpleLruReadPage_ReadOnly() */
1798 LWLockRelease(AsyncCtlLock);
1801 * Process messages up to the stop position, end of page, or an
1802 * uncommitted message.
1804 * Our stop position is what we found to be the head's position
1805 * when we entered this function. It might have changed already.
1806 * But if it has, we will receive (or have already received and
1807 * queued) another signal and come here again.
1809 * We are not holding AsyncQueueLock here! The queue can only
1810 * extend beyond the head pointer (see above) and we leave our
1811 * backend's pointer where it is so nobody will truncate or
1812 * rewrite pages under us. Especially we don't want to hold a lock
1813 * while sending the notifications to the frontend.
1815 reachedStop = asyncQueueProcessPageEntries(&pos, head,
1817 } while (!reachedStop);
1821 /* Update shared state */
1822 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1823 QUEUE_BACKEND_POS(MyBackendId) = pos;
1824 advanceTail = QUEUE_POS_EQUAL(oldpos, QUEUE_TAIL);
1825 LWLockRelease(AsyncQueueLock);
1827 /* If we were the laziest backend, try to advance the tail pointer */
1829 asyncQueueAdvanceTail();
1835 /* Update shared state */
1836 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1837 QUEUE_BACKEND_POS(MyBackendId) = pos;
1838 advanceTail = QUEUE_POS_EQUAL(oldpos, QUEUE_TAIL);
1839 LWLockRelease(AsyncQueueLock);
1841 /* If we were the laziest backend, try to advance the tail pointer */
1843 asyncQueueAdvanceTail();
1847 * Fetch notifications from the shared queue, beginning at position current,
1848 * and deliver relevant ones to my frontend.
1850 * The current page must have been fetched into page_buffer from shared
1851 * memory. (We could access the page right in shared memory, but that
1852 * would imply holding the AsyncCtlLock throughout this routine.)
1854 * We stop if we reach the "stop" position, or reach a notification from an
1855 * uncommitted transaction, or reach the end of the page.
1857 * The function returns true once we have reached the stop position or an
1858 * uncommitted notification, and false if we have finished with the page.
1859 * In other words: once it returns true there is no need to look further.
1860 * The QueuePosition *current is advanced past all processed messages.
1863 asyncQueueProcessPageEntries(volatile QueuePosition *current,
1867 bool reachedStop = false;
1868 bool reachedEndOfPage;
1869 AsyncQueueEntry *qe;
1873 QueuePosition thisentry = *current;
1875 if (QUEUE_POS_EQUAL(thisentry, stop))
1878 qe = (AsyncQueueEntry *) (page_buffer + QUEUE_POS_OFFSET(thisentry));
1881 * Advance *current over this message, possibly to the next page. As
1882 * noted in the comments for asyncQueueReadAllNotifications, we must
1883 * do this before possibly failing while processing the message.
1885 reachedEndOfPage = asyncQueueAdvance(current, qe->length);
1887 /* Ignore messages destined for other databases */
1888 if (qe->dboid == MyDatabaseId)
1890 if (TransactionIdIsInProgress(qe->xid))
1893 * The source transaction is still in progress, so we can't
1894 * process this message yet. Break out of the loop, but first
1895 * back up *current so we will reprocess the message next
1896 * time. (Note: it is unlikely but not impossible for
1897 * TransactionIdDidCommit to fail, so we can't really avoid
1898 * this advance-then-back-up behavior when dealing with an
1899 * uncommitted message.)
1901 * Note that we must test TransactionIdIsInProgress before we
1902 * test TransactionIdDidCommit, else we might return a message
1903 * from a transaction that is not yet visible to snapshots;
1904 * compare the comments at the head of tqual.c.
1906 *current = thisentry;
1910 else if (TransactionIdDidCommit(qe->xid))
1912 /* qe->data is the null-terminated channel name */
1913 char *channel = qe->data;
1915 if (IsListeningOn(channel))
1917 /* payload follows channel name */
1918 char *payload = qe->data + strlen(channel) + 1;
1920 NotifyMyFrontEnd(channel, payload, qe->srcPid);
1926 * The source transaction aborted or crashed, so we just
1927 * ignore its notifications.
1932 /* Loop back if we're not at end of page */
1933 } while (!reachedEndOfPage);
1935 if (QUEUE_POS_EQUAL(*current, stop))
1942 * Advance the shared queue tail variable to the minimum of all the
1943 * per-backend tail pointers. Truncate pg_notify space if possible.
1946 asyncQueueAdvanceTail(void)
1954 LWLockAcquire(AsyncQueueLock, LW_EXCLUSIVE);
1956 for (i = 1; i <= MaxBackends; i++)
1958 if (QUEUE_BACKEND_PID(i) != InvalidPid)
1959 min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
1961 oldtailpage = QUEUE_POS_PAGE(QUEUE_TAIL);
1963 LWLockRelease(AsyncQueueLock);
1966 * We can truncate something if the global tail advanced across an SLRU
1969 * XXX it might be better to truncate only once every several segments, to
1970 * reduce the number of directory scans.
1972 newtailpage = QUEUE_POS_PAGE(min);
1973 boundary = newtailpage - (newtailpage % SLRU_PAGES_PER_SEGMENT);
1974 if (asyncQueuePagePrecedes(oldtailpage, boundary))
1977 * SimpleLruTruncate() will ask for AsyncCtlLock but will also release
1980 SimpleLruTruncate(AsyncCtl, newtailpage);
1985 * ProcessIncomingNotify
1987 * Deal with arriving NOTIFYs from other backends as soon as it's safe to
1988 * do so. This used to be called from the PROCSIG_NOTIFY_INTERRUPT
1989 * signal handler, but isn't anymore.
1991 * Scan the queue for arriving notifications and report them to my front
1994 * NOTE: since we are outside any transaction, we must create our own.
1997 ProcessIncomingNotify(void)
1999 /* We *must* reset the flag */
2000 notifyInterruptPending = false;
2002 /* Do nothing else if we aren't actively listening */
2003 if (listenChannels == NIL)
2007 elog(DEBUG1, "ProcessIncomingNotify");
2009 set_ps_display("notify interrupt", false);
2012 * We must run asyncQueueReadAllNotifications inside a transaction, else
2013 * bad things happen if it gets an error.
2015 StartTransactionCommand();
2017 asyncQueueReadAllNotifications();
2019 CommitTransactionCommand();
2022 * Must flush the notify messages to ensure frontend gets them promptly.
2026 set_ps_display("idle", false);
2029 elog(DEBUG1, "ProcessIncomingNotify: done");
2033 * Send NOTIFY message to my front end.
2036 NotifyMyFrontEnd(const char *channel, const char *payload, int32 srcPid)
2038 if (whereToSendOutput == DestRemote)
2042 pq_beginmessage(&buf, 'A');
2043 pq_sendint(&buf, srcPid, sizeof(int32));
2044 pq_sendstring(&buf, channel);
2045 if (PG_PROTOCOL_MAJOR(FrontendProtocol) >= 3)
2046 pq_sendstring(&buf, payload);
2047 pq_endmessage(&buf);
2050 * NOTE: we do not do pq_flush() here. For a self-notify, it will
2051 * happen at the end of the transaction, and for incoming notifies
2052 * ProcessIncomingNotify will do it after finding all the notifies.
2056 elog(INFO, "NOTIFY for \"%s\" payload \"%s\"", channel, payload);
2059 /* Does pendingNotifies include the given channel/payload? */
2061 AsyncExistsPendingNotify(const char *channel, const char *payload)
2066 if (pendingNotifies == NIL)
2069 if (payload == NULL)
2073 * We need to append new elements to the end of the list in order to keep
2074 * the order. However, on the other hand we'd like to check the list
2075 * backwards in order to make duplicate-elimination a tad faster when the
2076 * same condition is signaled many times in a row. So as a compromise we
2077 * check the tail element first which we can access directly. If this
2078 * doesn't match, we check the whole list.
2080 * As we are not checking our parents' lists, we can still get duplicates
2081 * in combination with subtransactions, like in:
2090 n = (Notification *) llast(pendingNotifies);
2091 if (strcmp(n->channel, channel) == 0 &&
2092 strcmp(n->payload, payload) == 0)
2095 foreach(p, pendingNotifies)
2097 n = (Notification *) lfirst(p);
2099 if (strcmp(n->channel, channel) == 0 &&
2100 strcmp(n->payload, payload) == 0)
2107 /* Clear the pendingActions and pendingNotifies lists. */
2109 ClearPendingActionsAndNotifies(void)
2112 * We used to have to explicitly deallocate the list members and nodes,
2113 * because they were malloc'd. Now, since we know they are palloc'd in
2114 * CurTransactionContext, we need not do that --- they'll go away
2115 * automatically at transaction exit. We need only reset the list head
2118 pendingActions = NIL;
2119 pendingNotifies = NIL;