1 /*-------------------------------------------------------------------------
4 * Asynchronous notification: NOTIFY, LISTEN, UNLISTEN
6 * Portions Copyright (c) 1996-2016, 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/parallel.h"
121 #include "access/slru.h"
122 #include "access/transam.h"
123 #include "access/xact.h"
124 #include "catalog/pg_database.h"
125 #include "commands/async.h"
127 #include "libpq/libpq.h"
128 #include "libpq/pqformat.h"
129 #include "miscadmin.h"
130 #include "storage/ipc.h"
131 #include "storage/lmgr.h"
132 #include "storage/proc.h"
133 #include "storage/procarray.h"
134 #include "storage/procsignal.h"
135 #include "storage/sinval.h"
136 #include "tcop/tcopprot.h"
137 #include "utils/builtins.h"
138 #include "utils/memutils.h"
139 #include "utils/ps_status.h"
140 #include "utils/timestamp.h"
144 * Maximum size of a NOTIFY payload, including terminating NULL. This
145 * must be kept small enough so that a notification message fits on one
146 * SLRU page. The magic fudge factor here is noncritical as long as it's
147 * more than AsyncQueueEntryEmptySize --- we make it significantly bigger
148 * than that, so changes in that data structure won't affect user-visible
151 #define NOTIFY_PAYLOAD_MAX_LENGTH (BLCKSZ - NAMEDATALEN - 128)
154 * Struct representing an entry in the global notify queue
156 * This struct declaration has the maximal length, but in a real queue entry
157 * the data area is only big enough for the actual channel and payload strings
158 * (each null-terminated). AsyncQueueEntryEmptySize is the minimum possible
159 * entry size, if both channel and payload strings are empty (but note it
160 * doesn't include alignment padding).
162 * The "length" field should always be rounded up to the next QUEUEALIGN
163 * multiple so that all fields are properly aligned.
165 typedef struct AsyncQueueEntry
167 int length; /* total allocated length of entry */
168 Oid dboid; /* sender's database OID */
169 TransactionId xid; /* sender's XID */
170 int32 srcPid; /* sender's PID */
171 char data[NAMEDATALEN + NOTIFY_PAYLOAD_MAX_LENGTH];
174 /* Currently, no field of AsyncQueueEntry requires more than int alignment */
175 #define QUEUEALIGN(len) INTALIGN(len)
177 #define AsyncQueueEntryEmptySize (offsetof(AsyncQueueEntry, data) + 2)
180 * Struct describing a queue position, and assorted macros for working with it
182 typedef struct QueuePosition
184 int page; /* SLRU page number */
185 int offset; /* byte offset within page */
188 #define QUEUE_POS_PAGE(x) ((x).page)
189 #define QUEUE_POS_OFFSET(x) ((x).offset)
191 #define SET_QUEUE_POS(x,y,z) \
197 #define QUEUE_POS_EQUAL(x,y) \
198 ((x).page == (y).page && (x).offset == (y).offset)
200 /* choose logically smaller QueuePosition */
201 #define QUEUE_POS_MIN(x,y) \
202 (asyncQueuePagePrecedes((x).page, (y).page) ? (x) : \
203 (x).page != (y).page ? (y) : \
204 (x).offset < (y).offset ? (x) : (y))
206 /* choose logically larger QueuePosition */
207 #define QUEUE_POS_MAX(x,y) \
208 (asyncQueuePagePrecedes((x).page, (y).page) ? (y) : \
209 (x).page != (y).page ? (x) : \
210 (x).offset > (y).offset ? (x) : (y))
213 * Struct describing a listening backend's status
215 typedef struct QueueBackendStatus
217 int32 pid; /* either a PID or InvalidPid */
218 Oid dboid; /* backend's database OID, or InvalidOid */
219 QueuePosition pos; /* backend has read queue up to here */
220 } QueueBackendStatus;
223 * Shared memory state for LISTEN/NOTIFY (excluding its SLRU stuff)
225 * The AsyncQueueControl structure is protected by the AsyncQueueLock.
227 * When holding the lock in SHARED mode, backends may only inspect their own
228 * entries as well as the head and tail pointers. Consequently we can allow a
229 * backend to update its own record while holding only SHARED lock (since no
230 * other backend will inspect it).
232 * When holding the lock in EXCLUSIVE mode, backends can inspect the entries
233 * of other backends and also change the head and tail pointers.
235 * AsyncCtlLock is used as the control lock for the pg_notify SLRU buffers.
236 * In order to avoid deadlocks, whenever we need both locks, we always first
237 * get AsyncQueueLock and then AsyncCtlLock.
239 * Each backend uses the backend[] array entry with index equal to its
240 * BackendId (which can range from 1 to MaxBackends). We rely on this to make
241 * SendProcSignal fast.
243 typedef struct AsyncQueueControl
245 QueuePosition head; /* head points to the next free location */
246 QueuePosition tail; /* the global tail is equivalent to the pos of
247 * the "slowest" backend */
248 TimestampTz lastQueueFillWarn; /* time of last queue-full msg */
249 QueueBackendStatus backend[FLEXIBLE_ARRAY_MEMBER];
250 /* backend[0] is not used; used entries are from [1] to [MaxBackends] */
253 static AsyncQueueControl *asyncQueueControl;
255 #define QUEUE_HEAD (asyncQueueControl->head)
256 #define QUEUE_TAIL (asyncQueueControl->tail)
257 #define QUEUE_BACKEND_PID(i) (asyncQueueControl->backend[i].pid)
258 #define QUEUE_BACKEND_DBOID(i) (asyncQueueControl->backend[i].dboid)
259 #define QUEUE_BACKEND_POS(i) (asyncQueueControl->backend[i].pos)
262 * The SLRU buffer area through which we access the notification queue
264 static SlruCtlData AsyncCtlData;
266 #define AsyncCtl (&AsyncCtlData)
267 #define QUEUE_PAGESIZE BLCKSZ
268 #define QUEUE_FULL_WARN_INTERVAL 5000 /* warn at most once every 5s */
271 * slru.c currently assumes that all filenames are four characters of hex
272 * digits. That means that we can use segments 0000 through FFFF.
273 * Each segment contains SLRU_PAGES_PER_SEGMENT pages which gives us
274 * the pages from 0 to SLRU_PAGES_PER_SEGMENT * 0x10000 - 1.
276 * It's of course possible to enhance slru.c, but this gives us so much
277 * space already that it doesn't seem worth the trouble.
279 * The most data we can have in the queue at a time is QUEUE_MAX_PAGE/2
280 * pages, because more than that would confuse slru.c into thinking there
281 * was a wraparound condition. With the default BLCKSZ this means there
282 * can be up to 8GB of queued-and-not-read data.
284 * Note: it's possible to redefine QUEUE_MAX_PAGE with a smaller multiple of
285 * SLRU_PAGES_PER_SEGMENT, for easier testing of queue-full behaviour.
287 #define QUEUE_MAX_PAGE (SLRU_PAGES_PER_SEGMENT * 0x10000 - 1)
290 * listenChannels identifies the channels we are actually listening to
291 * (ie, have committed a LISTEN on). It is a simple list of channel names,
292 * allocated in TopMemoryContext.
294 static List *listenChannels = NIL; /* list of C strings */
297 * State for pending LISTEN/UNLISTEN actions consists of an ordered list of
298 * all actions requested in the current transaction. As explained above,
299 * we don't actually change listenChannels until we reach transaction commit.
301 * The list is kept in CurTransactionContext. In subtransactions, each
302 * subtransaction has its own list in its own CurTransactionContext, but
303 * successful subtransactions attach their lists to their parent's list.
304 * Failed subtransactions simply discard their lists.
315 ListenActionKind action;
316 char channel[FLEXIBLE_ARRAY_MEMBER]; /* nul-terminated string */
319 static List *pendingActions = NIL; /* list of ListenAction */
321 static List *upperPendingActions = NIL; /* list of upper-xact lists */
324 * State for outbound notifies consists of a list of all channels+payloads
325 * NOTIFYed in the current transaction. We do not actually perform a NOTIFY
326 * until and unless the transaction commits. pendingNotifies is NIL if no
327 * NOTIFYs have been done in the current transaction.
329 * The list is kept in CurTransactionContext. In subtransactions, each
330 * subtransaction has its own list in its own CurTransactionContext, but
331 * successful subtransactions attach their lists to their parent's list.
332 * Failed subtransactions simply discard their lists.
334 * Note: the action and notify lists do not interact within a transaction.
335 * In particular, if a transaction does NOTIFY and then LISTEN on the same
336 * condition name, it will get a self-notify at commit. This is a bit odd
337 * but is consistent with our historical behavior.
339 typedef struct Notification
341 char *channel; /* channel name */
342 char *payload; /* payload string (can be empty) */
345 static List *pendingNotifies = NIL; /* list of Notifications */
347 static List *upperPendingNotifies = NIL; /* list of upper-xact lists */
350 * Inbound notifications are initially processed by HandleNotifyInterrupt(),
351 * called from inside a signal handler. That just sets the
352 * notifyInterruptPending flag and sets the process
353 * latch. ProcessNotifyInterrupt() will then be called whenever it's safe to
354 * actually deal with the interrupt.
356 volatile sig_atomic_t notifyInterruptPending = false;
358 /* True if we've registered an on_shmem_exit cleanup */
359 static bool unlistenExitRegistered = false;
361 /* True if we're currently registered as a listener in asyncQueueControl */
362 static bool amRegisteredListener = false;
364 /* has this backend sent notifications in the current transaction? */
365 static bool backendHasSentNotifications = false;
368 bool Trace_notify = false;
370 /* local function prototypes */
371 static bool asyncQueuePagePrecedes(int p, int q);
372 static void queue_listen(ListenActionKind action, const char *channel);
373 static void Async_UnlistenOnExit(int code, Datum arg);
374 static void Exec_ListenPreCommit(void);
375 static void Exec_ListenCommit(const char *channel);
376 static void Exec_UnlistenCommit(const char *channel);
377 static void Exec_UnlistenAllCommit(void);
378 static bool IsListeningOn(const char *channel);
379 static void asyncQueueUnregister(void);
380 static bool asyncQueueIsFull(void);
381 static bool asyncQueueAdvance(volatile QueuePosition *position, int entryLength);
382 static void asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe);
383 static ListCell *asyncQueueAddEntries(ListCell *nextNotify);
384 static double asyncQueueUsage(void);
385 static void asyncQueueFillWarning(void);
386 static bool SignalBackends(void);
387 static void asyncQueueReadAllNotifications(void);
388 static bool asyncQueueProcessPageEntries(volatile QueuePosition *current,
391 static void asyncQueueAdvanceTail(void);
392 static void ProcessIncomingNotify(void);
393 static void NotifyMyFrontEnd(const char *channel,
396 static bool AsyncExistsPendingNotify(const char *channel, const char *payload);
397 static void ClearPendingActionsAndNotifies(void);
400 * We will work on the page range of 0..QUEUE_MAX_PAGE.
403 asyncQueuePagePrecedes(int p, int q)
408 * We have to compare modulo (QUEUE_MAX_PAGE+1)/2. Both inputs should be
409 * in the range 0..QUEUE_MAX_PAGE.
411 Assert(p >= 0 && p <= QUEUE_MAX_PAGE);
412 Assert(q >= 0 && q <= QUEUE_MAX_PAGE);
415 if (diff >= ((QUEUE_MAX_PAGE + 1) / 2))
416 diff -= QUEUE_MAX_PAGE + 1;
417 else if (diff < -((QUEUE_MAX_PAGE + 1) / 2))
418 diff += QUEUE_MAX_PAGE + 1;
423 * Report space needed for our shared memory area
430 /* This had better match AsyncShmemInit */
431 size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
432 size = add_size(size, offsetof(AsyncQueueControl, backend));
434 size = add_size(size, SimpleLruShmemSize(NUM_ASYNC_BUFFERS, 0));
440 * Initialize our shared memory area
450 * Create or attach to the AsyncQueueControl structure.
452 * The used entries in the backend[] array run from 1 to MaxBackends; the
453 * zero'th entry is unused but must be allocated.
455 size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
456 size = add_size(size, offsetof(AsyncQueueControl, backend));
458 asyncQueueControl = (AsyncQueueControl *)
459 ShmemInitStruct("Async Queue Control", size, &found);
463 /* First time through, so initialize it */
466 SET_QUEUE_POS(QUEUE_HEAD, 0, 0);
467 SET_QUEUE_POS(QUEUE_TAIL, 0, 0);
468 asyncQueueControl->lastQueueFillWarn = 0;
469 /* zero'th entry won't be used, but let's initialize it anyway */
470 for (i = 0; i <= MaxBackends; i++)
472 QUEUE_BACKEND_PID(i) = InvalidPid;
473 QUEUE_BACKEND_DBOID(i) = InvalidOid;
474 SET_QUEUE_POS(QUEUE_BACKEND_POS(i), 0, 0);
479 * Set up SLRU management of the pg_notify data.
481 AsyncCtl->PagePrecedes = asyncQueuePagePrecedes;
482 SimpleLruInit(AsyncCtl, "async", NUM_ASYNC_BUFFERS, 0,
483 AsyncCtlLock, "pg_notify");
484 /* Override default assumption that writes should be fsync'd */
485 AsyncCtl->do_fsync = false;
490 * During start or reboot, clean out the pg_notify directory.
492 (void) SlruScanDirectory(AsyncCtl, SlruScanDirCbDeleteAll, NULL);
494 /* Now initialize page zero to empty */
495 LWLockAcquire(AsyncCtlLock, LW_EXCLUSIVE);
496 slotno = SimpleLruZeroPage(AsyncCtl, QUEUE_POS_PAGE(QUEUE_HEAD));
497 /* This write is just to verify that pg_notify/ is writable */
498 SimpleLruWritePage(AsyncCtl, slotno);
499 LWLockRelease(AsyncCtlLock);
506 * SQL function to send a notification event
509 pg_notify(PG_FUNCTION_ARGS)
517 channel = text_to_cstring(PG_GETARG_TEXT_PP(0));
522 payload = text_to_cstring(PG_GETARG_TEXT_PP(1));
524 /* For NOTIFY as a statement, this is checked in ProcessUtility */
525 PreventCommandDuringRecovery("NOTIFY");
527 Async_Notify(channel, payload);
536 * This is executed by the SQL notify command.
538 * Adds the message to the list of pending notifies.
539 * Actual notification happens during transaction commit.
540 * ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
543 Async_Notify(const char *channel, const char *payload)
546 MemoryContext oldcontext;
548 if (IsParallelWorker())
549 elog(ERROR, "cannot send notifications from a parallel worker");
552 elog(DEBUG1, "Async_Notify(%s)", channel);
554 /* a channel name must be specified */
555 if (!channel || !strlen(channel))
557 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
558 errmsg("channel name cannot be empty")));
560 if (strlen(channel) >= NAMEDATALEN)
562 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
563 errmsg("channel name too long")));
567 if (strlen(payload) >= NOTIFY_PAYLOAD_MAX_LENGTH)
569 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
570 errmsg("payload string too long")));
573 /* no point in making duplicate entries in the list ... */
574 if (AsyncExistsPendingNotify(channel, payload))
578 * The notification list needs to live until end of transaction, so store
579 * it in the transaction context.
581 oldcontext = MemoryContextSwitchTo(CurTransactionContext);
583 n = (Notification *) palloc(sizeof(Notification));
584 n->channel = pstrdup(channel);
586 n->payload = pstrdup(payload);
591 * We want to preserve the order so we need to append every notification.
592 * See comments at AsyncExistsPendingNotify().
594 pendingNotifies = lappend(pendingNotifies, n);
596 MemoryContextSwitchTo(oldcontext);
601 * Common code for listen, unlisten, unlisten all commands.
603 * Adds the request to the list of pending actions.
604 * Actual update of the listenChannels list happens during transaction
608 queue_listen(ListenActionKind action, const char *channel)
610 MemoryContext oldcontext;
611 ListenAction *actrec;
614 * Unlike Async_Notify, we don't try to collapse out duplicates. It would
615 * be too complicated to ensure we get the right interactions of
616 * conflicting LISTEN/UNLISTEN/UNLISTEN_ALL, and it's unlikely that there
617 * would be any performance benefit anyway in sane applications.
619 oldcontext = MemoryContextSwitchTo(CurTransactionContext);
621 /* space for terminating null is included in sizeof(ListenAction) */
622 actrec = (ListenAction *) palloc(offsetof(ListenAction, channel) +
623 strlen(channel) + 1);
624 actrec->action = action;
625 strcpy(actrec->channel, channel);
627 pendingActions = lappend(pendingActions, actrec);
629 MemoryContextSwitchTo(oldcontext);
635 * This is executed by the SQL listen command.
638 Async_Listen(const char *channel)
641 elog(DEBUG1, "Async_Listen(%s,%d)", channel, MyProcPid);
643 queue_listen(LISTEN_LISTEN, channel);
649 * This is executed by the SQL unlisten command.
652 Async_Unlisten(const char *channel)
655 elog(DEBUG1, "Async_Unlisten(%s,%d)", channel, MyProcPid);
657 /* If we couldn't possibly be listening, no need to queue anything */
658 if (pendingActions == NIL && !unlistenExitRegistered)
661 queue_listen(LISTEN_UNLISTEN, channel);
667 * This is invoked by UNLISTEN * command, and also at backend exit.
670 Async_UnlistenAll(void)
673 elog(DEBUG1, "Async_UnlistenAll(%d)", MyProcPid);
675 /* If we couldn't possibly be listening, no need to queue anything */
676 if (pendingActions == NIL && !unlistenExitRegistered)
679 queue_listen(LISTEN_UNLISTEN_ALL, "");
683 * SQL function: return a set of the channel names this backend is actively
686 * Note: this coding relies on the fact that the listenChannels list cannot
687 * change within a transaction.
690 pg_listening_channels(PG_FUNCTION_ARGS)
692 FuncCallContext *funcctx;
695 /* stuff done only on the first call of the function */
696 if (SRF_IS_FIRSTCALL())
698 MemoryContext oldcontext;
700 /* create a function context for cross-call persistence */
701 funcctx = SRF_FIRSTCALL_INIT();
703 /* switch to memory context appropriate for multiple function calls */
704 oldcontext = MemoryContextSwitchTo(funcctx->multi_call_memory_ctx);
706 /* allocate memory for user context */
707 lcp = (ListCell **) palloc(sizeof(ListCell *));
708 *lcp = list_head(listenChannels);
709 funcctx->user_fctx = (void *) lcp;
711 MemoryContextSwitchTo(oldcontext);
714 /* stuff done on every call of the function */
715 funcctx = SRF_PERCALL_SETUP();
716 lcp = (ListCell **) funcctx->user_fctx;
720 char *channel = (char *) lfirst(*lcp);
723 SRF_RETURN_NEXT(funcctx, CStringGetTextDatum(channel));
726 SRF_RETURN_DONE(funcctx);
730 * Async_UnlistenOnExit
732 * This is executed at backend exit if we have done any LISTENs in this
733 * backend. It might not be necessary anymore, if the user UNLISTENed
734 * everything, but we don't try to detect that case.
737 Async_UnlistenOnExit(int code, Datum arg)
739 Exec_UnlistenAllCommit();
740 asyncQueueUnregister();
746 * This is called at the prepare phase of a two-phase
747 * transaction. Save the state for possible commit later.
750 AtPrepare_Notify(void)
752 /* It's not allowed to have any pending LISTEN/UNLISTEN/NOTIFY actions */
753 if (pendingActions || pendingNotifies)
755 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
756 errmsg("cannot PREPARE a transaction that has executed LISTEN, UNLISTEN, or NOTIFY")));
762 * This is called at transaction commit, before actually committing to
765 * If there are pending LISTEN actions, make sure we are listed in the
766 * shared-memory listener array. This must happen before commit to
767 * ensure we don't miss any notifies from transactions that commit
770 * If there are outbound notify requests in the pendingNotifies list,
771 * add them to the global queue. We do that before commit so that
772 * we can still throw error if we run out of queue space.
775 PreCommit_Notify(void)
779 if (pendingActions == NIL && pendingNotifies == NIL)
780 return; /* no relevant statements in this xact */
783 elog(DEBUG1, "PreCommit_Notify");
785 /* Preflight for any pending listen/unlisten actions */
786 foreach(p, pendingActions)
788 ListenAction *actrec = (ListenAction *) lfirst(p);
790 switch (actrec->action)
793 Exec_ListenPreCommit();
795 case LISTEN_UNLISTEN:
796 /* there is no Exec_UnlistenPreCommit() */
798 case LISTEN_UNLISTEN_ALL:
799 /* there is no Exec_UnlistenAllPreCommit() */
804 /* Queue any pending notifies */
807 ListCell *nextNotify;
810 * Make sure that we have an XID assigned to the current transaction.
811 * GetCurrentTransactionId is cheap if we already have an XID, but not
812 * so cheap if we don't, and we'd prefer not to do that work while
813 * holding AsyncQueueLock.
815 (void) GetCurrentTransactionId();
818 * Serialize writers by acquiring a special lock that we hold till
819 * after commit. This ensures that queue entries appear in commit
820 * order, and in particular that there are never uncommitted queue
821 * entries ahead of committed ones, so an uncommitted transaction
822 * can't block delivery of deliverable notifications.
824 * We use a heavyweight lock so that it'll automatically be released
825 * after either commit or abort. This also allows deadlocks to be
826 * detected, though really a deadlock shouldn't be possible here.
828 * The lock is on "database 0", which is pretty ugly but it doesn't
829 * seem worth inventing a special locktag category just for this.
830 * (Historical note: before PG 9.0, a similar lock on "database 0" was
831 * used by the flatfiles mechanism.)
833 LockSharedObject(DatabaseRelationId, InvalidOid, 0,
834 AccessExclusiveLock);
836 /* Now push the notifications into the queue */
837 backendHasSentNotifications = true;
839 nextNotify = list_head(pendingNotifies);
840 while (nextNotify != NULL)
843 * Add the pending notifications to the queue. We acquire and
844 * release AsyncQueueLock once per page, which might be overkill
845 * but it does allow readers to get in while we're doing this.
847 * A full queue is very uncommon and should really not happen,
848 * given that we have so much space available in the SLRU pages.
849 * Nevertheless we need to deal with this possibility. Note that
850 * when we get here we are in the process of committing our
851 * transaction, but we have not yet committed to clog, so at this
852 * point in time we can still roll the transaction back.
854 LWLockAcquire(AsyncQueueLock, LW_EXCLUSIVE);
855 asyncQueueFillWarning();
856 if (asyncQueueIsFull())
858 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
859 errmsg("too many notifications in the NOTIFY queue")));
860 nextNotify = asyncQueueAddEntries(nextNotify);
861 LWLockRelease(AsyncQueueLock);
869 * This is called at transaction commit, after committing to clog.
871 * Update listenChannels and clear transaction-local state.
874 AtCommit_Notify(void)
879 * Allow transactions that have not executed LISTEN/UNLISTEN/NOTIFY to
880 * return as soon as possible
882 if (!pendingActions && !pendingNotifies)
886 elog(DEBUG1, "AtCommit_Notify");
888 /* Perform any pending listen/unlisten actions */
889 foreach(p, pendingActions)
891 ListenAction *actrec = (ListenAction *) lfirst(p);
893 switch (actrec->action)
896 Exec_ListenCommit(actrec->channel);
898 case LISTEN_UNLISTEN:
899 Exec_UnlistenCommit(actrec->channel);
901 case LISTEN_UNLISTEN_ALL:
902 Exec_UnlistenAllCommit();
907 /* If no longer listening to anything, get out of listener array */
908 if (amRegisteredListener && listenChannels == NIL)
909 asyncQueueUnregister();
912 ClearPendingActionsAndNotifies();
916 * Exec_ListenPreCommit --- subroutine for PreCommit_Notify
918 * This function must make sure we are ready to catch any incoming messages.
921 Exec_ListenPreCommit(void)
928 * Nothing to do if we are already listening to something, nor if we
929 * already ran this routine in this transaction.
931 if (amRegisteredListener)
935 elog(DEBUG1, "Exec_ListenPreCommit(%d)", MyProcPid);
938 * Before registering, make sure we will unlisten before dying. (Note:
939 * this action does not get undone if we abort later.)
941 if (!unlistenExitRegistered)
943 before_shmem_exit(Async_UnlistenOnExit, 0);
944 unlistenExitRegistered = true;
948 * This is our first LISTEN, so establish our pointer.
950 * We set our pointer to the global tail pointer and then move it forward
951 * over already-committed notifications. This ensures we cannot miss any
952 * not-yet-committed notifications. We might get a few more but that
955 * In some scenarios there might be a lot of committed notifications that
956 * have not yet been pruned away (because some backend is being lazy about
957 * reading them). To reduce our startup time, we can look at other
958 * backends and adopt the maximum "pos" pointer of any backend that's in
959 * our database; any notifications it's already advanced over are surely
960 * committed and need not be re-examined by us. (We must consider only
961 * backends connected to our DB, because others will not have bothered to
962 * check committed-ness of notifications in our DB.) But we only bother
963 * with that if there's more than a page worth of notifications
964 * outstanding, otherwise scanning all the other backends isn't worth it.
966 * We need exclusive lock here so we can look at other backends' entries.
968 LWLockAcquire(AsyncQueueLock, LW_EXCLUSIVE);
971 if (QUEUE_POS_PAGE(max) != QUEUE_POS_PAGE(head))
973 for (i = 1; i <= MaxBackends; i++)
975 if (QUEUE_BACKEND_DBOID(i) == MyDatabaseId)
976 max = QUEUE_POS_MAX(max, QUEUE_BACKEND_POS(i));
979 QUEUE_BACKEND_POS(MyBackendId) = max;
980 QUEUE_BACKEND_PID(MyBackendId) = MyProcPid;
981 QUEUE_BACKEND_DBOID(MyBackendId) = MyDatabaseId;
982 LWLockRelease(AsyncQueueLock);
984 /* Now we are listed in the global array, so remember we're listening */
985 amRegisteredListener = true;
988 * Try to move our pointer forward as far as possible. This will skip over
989 * already-committed notifications. Still, we could get notifications that
990 * have already committed before we started to LISTEN.
992 * Note that we are not yet listening on anything, so we won't deliver any
993 * notification to the frontend.
995 * This will also advance the global tail pointer if possible.
997 if (!QUEUE_POS_EQUAL(max, head))
998 asyncQueueReadAllNotifications();
1002 * Exec_ListenCommit --- subroutine for AtCommit_Notify
1004 * Add the channel to the list of channels we are listening on.
1007 Exec_ListenCommit(const char *channel)
1009 MemoryContext oldcontext;
1011 /* Do nothing if we are already listening on this channel */
1012 if (IsListeningOn(channel))
1016 * Add the new channel name to listenChannels.
1018 * XXX It is theoretically possible to get an out-of-memory failure here,
1019 * which would be bad because we already committed. For the moment it
1020 * doesn't seem worth trying to guard against that, but maybe improve this
1023 oldcontext = MemoryContextSwitchTo(TopMemoryContext);
1024 listenChannels = lappend(listenChannels, pstrdup(channel));
1025 MemoryContextSwitchTo(oldcontext);
1029 * Exec_UnlistenCommit --- subroutine for AtCommit_Notify
1031 * Remove the specified channel name from listenChannels.
1034 Exec_UnlistenCommit(const char *channel)
1040 elog(DEBUG1, "Exec_UnlistenCommit(%s,%d)", channel, MyProcPid);
1043 foreach(q, listenChannels)
1045 char *lchan = (char *) lfirst(q);
1047 if (strcmp(lchan, channel) == 0)
1049 listenChannels = list_delete_cell(listenChannels, q, prev);
1057 * We do not complain about unlistening something not being listened;
1063 * Exec_UnlistenAllCommit --- subroutine for AtCommit_Notify
1065 * Unlisten on all channels for this backend.
1068 Exec_UnlistenAllCommit(void)
1071 elog(DEBUG1, "Exec_UnlistenAllCommit(%d)", MyProcPid);
1073 list_free_deep(listenChannels);
1074 listenChannels = NIL;
1078 * ProcessCompletedNotifies --- send out signals and self-notifies
1080 * This is called from postgres.c just before going idle at the completion
1081 * of a transaction. If we issued any notifications in the just-completed
1082 * transaction, send signals to other backends to process them, and also
1083 * process the queue ourselves to send messages to our own frontend.
1085 * The reason that this is not done in AtCommit_Notify is that there is
1086 * a nonzero chance of errors here (for example, encoding conversion errors
1087 * while trying to format messages to our frontend). An error during
1088 * AtCommit_Notify would be a PANIC condition. The timing is also arranged
1089 * to ensure that a transaction's self-notifies are delivered to the frontend
1090 * before it gets the terminating ReadyForQuery message.
1092 * Note that we send signals and process the queue even if the transaction
1093 * eventually aborted. This is because we need to clean out whatever got
1094 * added to the queue.
1096 * NOTE: we are outside of any transaction here.
1099 ProcessCompletedNotifies(void)
1101 MemoryContext caller_context;
1104 /* Nothing to do if we didn't send any notifications */
1105 if (!backendHasSentNotifications)
1109 * We reset the flag immediately; otherwise, if any sort of error occurs
1110 * below, we'd be locked up in an infinite loop, because control will come
1111 * right back here after error cleanup.
1113 backendHasSentNotifications = false;
1116 * We must preserve the caller's memory context (probably MessageContext)
1117 * across the transaction we do here.
1119 caller_context = CurrentMemoryContext;
1122 elog(DEBUG1, "ProcessCompletedNotifies");
1125 * We must run asyncQueueReadAllNotifications inside a transaction, else
1126 * bad things happen if it gets an error.
1128 StartTransactionCommand();
1130 /* Send signals to other backends */
1131 signalled = SignalBackends();
1133 if (listenChannels != NIL)
1135 /* Read the queue ourselves, and send relevant stuff to the frontend */
1136 asyncQueueReadAllNotifications();
1138 else if (!signalled)
1141 * If we found no other listening backends, and we aren't listening
1142 * ourselves, then we must execute asyncQueueAdvanceTail to flush the
1143 * queue, because ain't nobody else gonna do it. This prevents queue
1144 * overflow when we're sending useless notifies to nobody. (A new
1145 * listener could have joined since we looked, but if so this is
1148 asyncQueueAdvanceTail();
1151 CommitTransactionCommand();
1153 MemoryContextSwitchTo(caller_context);
1155 /* We don't need pq_flush() here since postgres.c will do one shortly */
1159 * Test whether we are actively listening on the given channel name.
1161 * Note: this function is executed for every notification found in the queue.
1162 * Perhaps it is worth further optimization, eg convert the list to a sorted
1163 * array so we can binary-search it. In practice the list is likely to be
1164 * fairly short, though.
1167 IsListeningOn(const char *channel)
1171 foreach(p, listenChannels)
1173 char *lchan = (char *) lfirst(p);
1175 if (strcmp(lchan, channel) == 0)
1182 * Remove our entry from the listeners array when we are no longer listening
1183 * on any channel. NB: must not fail if we're already not listening.
1186 asyncQueueUnregister(void)
1190 Assert(listenChannels == NIL); /* else caller error */
1192 if (!amRegisteredListener) /* nothing to do */
1195 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1196 /* check if entry is valid and oldest ... */
1197 advanceTail = (MyProcPid == QUEUE_BACKEND_PID(MyBackendId)) &&
1198 QUEUE_POS_EQUAL(QUEUE_BACKEND_POS(MyBackendId), QUEUE_TAIL);
1199 /* ... then mark it invalid */
1200 QUEUE_BACKEND_PID(MyBackendId) = InvalidPid;
1201 QUEUE_BACKEND_DBOID(MyBackendId) = InvalidOid;
1202 LWLockRelease(AsyncQueueLock);
1204 /* mark ourselves as no longer listed in the global array */
1205 amRegisteredListener = false;
1207 /* If we were the laziest backend, try to advance the tail pointer */
1209 asyncQueueAdvanceTail();
1213 * Test whether there is room to insert more notification messages.
1215 * Caller must hold at least shared AsyncQueueLock.
1218 asyncQueueIsFull(void)
1224 * The queue is full if creating a new head page would create a page that
1225 * logically precedes the current global tail pointer, ie, the head
1226 * pointer would wrap around compared to the tail. We cannot create such
1227 * a head page for fear of confusing slru.c. For safety we round the tail
1228 * pointer back to a segment boundary (compare the truncation logic in
1229 * asyncQueueAdvanceTail).
1231 * Note that this test is *not* dependent on how much space there is on
1232 * the current head page. This is necessary because asyncQueueAddEntries
1233 * might try to create the next head page in any case.
1235 nexthead = QUEUE_POS_PAGE(QUEUE_HEAD) + 1;
1236 if (nexthead > QUEUE_MAX_PAGE)
1237 nexthead = 0; /* wrap around */
1238 boundary = QUEUE_POS_PAGE(QUEUE_TAIL);
1239 boundary -= boundary % SLRU_PAGES_PER_SEGMENT;
1240 return asyncQueuePagePrecedes(nexthead, boundary);
1244 * Advance the QueuePosition to the next entry, assuming that the current
1245 * entry is of length entryLength. If we jump to a new page the function
1246 * returns true, else false.
1249 asyncQueueAdvance(volatile QueuePosition *position, int entryLength)
1251 int pageno = QUEUE_POS_PAGE(*position);
1252 int offset = QUEUE_POS_OFFSET(*position);
1253 bool pageJump = false;
1256 * Move to the next writing position: First jump over what we have just
1259 offset += entryLength;
1260 Assert(offset <= QUEUE_PAGESIZE);
1263 * In a second step check if another entry can possibly be written to the
1264 * page. If so, stay here, we have reached the next position. If not, then
1265 * we need to move on to the next page.
1267 if (offset + QUEUEALIGN(AsyncQueueEntryEmptySize) > QUEUE_PAGESIZE)
1270 if (pageno > QUEUE_MAX_PAGE)
1271 pageno = 0; /* wrap around */
1276 SET_QUEUE_POS(*position, pageno, offset);
1281 * Fill the AsyncQueueEntry at *qe with an outbound notification message.
1284 asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe)
1286 size_t channellen = strlen(n->channel);
1287 size_t payloadlen = strlen(n->payload);
1290 Assert(channellen < NAMEDATALEN);
1291 Assert(payloadlen < NOTIFY_PAYLOAD_MAX_LENGTH);
1293 /* The terminators are already included in AsyncQueueEntryEmptySize */
1294 entryLength = AsyncQueueEntryEmptySize + payloadlen + channellen;
1295 entryLength = QUEUEALIGN(entryLength);
1296 qe->length = entryLength;
1297 qe->dboid = MyDatabaseId;
1298 qe->xid = GetCurrentTransactionId();
1299 qe->srcPid = MyProcPid;
1300 memcpy(qe->data, n->channel, channellen + 1);
1301 memcpy(qe->data + channellen + 1, n->payload, payloadlen + 1);
1305 * Add pending notifications to the queue.
1307 * We go page by page here, i.e. we stop once we have to go to a new page but
1308 * we will be called again and then fill that next page. If an entry does not
1309 * fit into the current page, we write a dummy entry with an InvalidOid as the
1310 * database OID in order to fill the page. So every page is always used up to
1311 * the last byte which simplifies reading the page later.
1313 * We are passed the list cell containing the next notification to write
1314 * and return the first still-unwritten cell back. Eventually we will return
1315 * NULL indicating all is done.
1317 * We are holding AsyncQueueLock already from the caller and grab AsyncCtlLock
1318 * locally in this function.
1321 asyncQueueAddEntries(ListCell *nextNotify)
1324 QueuePosition queue_head;
1329 /* We hold both AsyncQueueLock and AsyncCtlLock during this operation */
1330 LWLockAcquire(AsyncCtlLock, LW_EXCLUSIVE);
1333 * We work with a local copy of QUEUE_HEAD, which we write back to shared
1334 * memory upon exiting. The reason for this is that if we have to advance
1335 * to a new page, SimpleLruZeroPage might fail (out of disk space, for
1336 * instance), and we must not advance QUEUE_HEAD if it does. (Otherwise,
1337 * subsequent insertions would try to put entries into a page that slru.c
1338 * thinks doesn't exist yet.) So, use a local position variable. Note
1339 * that if we do fail, any already-inserted queue entries are forgotten;
1340 * this is okay, since they'd be useless anyway after our transaction
1343 queue_head = QUEUE_HEAD;
1345 /* Fetch the current page */
1346 pageno = QUEUE_POS_PAGE(queue_head);
1347 slotno = SimpleLruReadPage(AsyncCtl, pageno, true, InvalidTransactionId);
1348 /* Note we mark the page dirty before writing in it */
1349 AsyncCtl->shared->page_dirty[slotno] = true;
1351 while (nextNotify != NULL)
1353 Notification *n = (Notification *) lfirst(nextNotify);
1355 /* Construct a valid queue entry in local variable qe */
1356 asyncQueueNotificationToEntry(n, &qe);
1358 offset = QUEUE_POS_OFFSET(queue_head);
1360 /* Check whether the entry really fits on the current page */
1361 if (offset + qe.length <= QUEUE_PAGESIZE)
1363 /* OK, so advance nextNotify past this item */
1364 nextNotify = lnext(nextNotify);
1369 * Write a dummy entry to fill up the page. Actually readers will
1370 * only check dboid and since it won't match any reader's database
1371 * OID, they will ignore this entry and move on.
1373 qe.length = QUEUE_PAGESIZE - offset;
1374 qe.dboid = InvalidOid;
1375 qe.data[0] = '\0'; /* empty channel */
1376 qe.data[1] = '\0'; /* empty payload */
1379 /* Now copy qe into the shared buffer page */
1380 memcpy(AsyncCtl->shared->page_buffer[slotno] + offset,
1384 /* Advance queue_head appropriately, and detect if page is full */
1385 if (asyncQueueAdvance(&(queue_head), qe.length))
1388 * Page is full, so we're done here, but first fill the next page
1389 * with zeroes. The reason to do this is to ensure that slru.c's
1390 * idea of the head page is always the same as ours, which avoids
1391 * boundary problems in SimpleLruTruncate. The test in
1392 * asyncQueueIsFull() ensured that there is room to create this
1393 * page without overrunning the queue.
1395 slotno = SimpleLruZeroPage(AsyncCtl, QUEUE_POS_PAGE(queue_head));
1396 /* And exit the loop */
1401 /* Success, so update the global QUEUE_HEAD */
1402 QUEUE_HEAD = queue_head;
1404 LWLockRelease(AsyncCtlLock);
1410 * SQL function to return the fraction of the notification queue currently
1414 pg_notification_queue_usage(PG_FUNCTION_ARGS)
1418 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1419 usage = asyncQueueUsage();
1420 LWLockRelease(AsyncQueueLock);
1422 PG_RETURN_FLOAT8(usage);
1426 * Return the fraction of the queue that is currently occupied.
1428 * The caller must hold AsyncQueueLock in (at least) shared mode.
1431 asyncQueueUsage(void)
1433 int headPage = QUEUE_POS_PAGE(QUEUE_HEAD);
1434 int tailPage = QUEUE_POS_PAGE(QUEUE_TAIL);
1437 occupied = headPage - tailPage;
1440 return (double) 0; /* fast exit for common case */
1444 /* head has wrapped around, tail not yet */
1445 occupied += QUEUE_MAX_PAGE + 1;
1448 return (double) occupied / (double) ((QUEUE_MAX_PAGE + 1) / 2);
1452 * Check whether the queue is at least half full, and emit a warning if so.
1454 * This is unlikely given the size of the queue, but possible.
1455 * The warnings show up at most once every QUEUE_FULL_WARN_INTERVAL.
1457 * Caller must hold exclusive AsyncQueueLock.
1460 asyncQueueFillWarning(void)
1465 fillDegree = asyncQueueUsage();
1466 if (fillDegree < 0.5)
1469 t = GetCurrentTimestamp();
1471 if (TimestampDifferenceExceeds(asyncQueueControl->lastQueueFillWarn,
1472 t, QUEUE_FULL_WARN_INTERVAL))
1474 QueuePosition min = QUEUE_HEAD;
1475 int32 minPid = InvalidPid;
1478 for (i = 1; i <= MaxBackends; i++)
1480 if (QUEUE_BACKEND_PID(i) != InvalidPid)
1482 min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
1483 if (QUEUE_POS_EQUAL(min, QUEUE_BACKEND_POS(i)))
1484 minPid = QUEUE_BACKEND_PID(i);
1489 (errmsg("NOTIFY queue is %.0f%% full", fillDegree * 100),
1490 (minPid != InvalidPid ?
1491 errdetail("The server process with PID %d is among those with the oldest transactions.", minPid)
1493 (minPid != InvalidPid ?
1494 errhint("The NOTIFY queue cannot be emptied until that process ends its current transaction.")
1497 asyncQueueControl->lastQueueFillWarn = t;
1502 * Send signals to all listening backends (except our own).
1504 * Returns true if we sent at least one signal.
1506 * Since we need EXCLUSIVE lock anyway we also check the position of the other
1507 * backends and in case one is already up-to-date we don't signal it.
1508 * This can happen if concurrent notifying transactions have sent a signal and
1509 * the signaled backend has read the other notifications and ours in the same
1512 * Since we know the BackendId and the Pid the signalling is quite cheap.
1515 SignalBackends(void)
1517 bool signalled = false;
1525 * Identify all backends that are listening and not already up-to-date. We
1526 * don't want to send signals while holding the AsyncQueueLock, so we just
1527 * build a list of target PIDs.
1529 * XXX in principle these pallocs could fail, which would be bad. Maybe
1530 * preallocate the arrays? But in practice this is only run in trivial
1531 * transactions, so there should surely be space available.
1533 pids = (int32 *) palloc(MaxBackends * sizeof(int32));
1534 ids = (BackendId *) palloc(MaxBackends * sizeof(BackendId));
1537 LWLockAcquire(AsyncQueueLock, LW_EXCLUSIVE);
1538 for (i = 1; i <= MaxBackends; i++)
1540 pid = QUEUE_BACKEND_PID(i);
1541 if (pid != InvalidPid && pid != MyProcPid)
1543 QueuePosition pos = QUEUE_BACKEND_POS(i);
1545 if (!QUEUE_POS_EQUAL(pos, QUEUE_HEAD))
1553 LWLockRelease(AsyncQueueLock);
1555 /* Now send signals */
1556 for (i = 0; i < count; i++)
1561 * Note: assuming things aren't broken, a signal failure here could
1562 * only occur if the target backend exited since we released
1563 * AsyncQueueLock; which is unlikely but certainly possible. So we
1564 * just log a low-level debug message if it happens.
1566 if (SendProcSignal(pid, PROCSIG_NOTIFY_INTERRUPT, ids[i]) < 0)
1567 elog(DEBUG3, "could not signal backend with PID %d: %m", pid);
1581 * This is called at transaction abort.
1583 * Gets rid of pending actions and outbound notifies that we would have
1584 * executed if the transaction got committed.
1587 AtAbort_Notify(void)
1590 * If we LISTEN but then roll back the transaction after PreCommit_Notify,
1591 * we have registered as a listener but have not made any entry in
1592 * listenChannels. In that case, deregister again.
1594 if (amRegisteredListener && listenChannels == NIL)
1595 asyncQueueUnregister();
1598 ClearPendingActionsAndNotifies();
1602 * AtSubStart_Notify() --- Take care of subtransaction start.
1604 * Push empty state for the new subtransaction.
1607 AtSubStart_Notify(void)
1609 MemoryContext old_cxt;
1611 /* Keep the list-of-lists in TopTransactionContext for simplicity */
1612 old_cxt = MemoryContextSwitchTo(TopTransactionContext);
1614 upperPendingActions = lcons(pendingActions, upperPendingActions);
1616 Assert(list_length(upperPendingActions) ==
1617 GetCurrentTransactionNestLevel() - 1);
1619 pendingActions = NIL;
1621 upperPendingNotifies = lcons(pendingNotifies, upperPendingNotifies);
1623 Assert(list_length(upperPendingNotifies) ==
1624 GetCurrentTransactionNestLevel() - 1);
1626 pendingNotifies = NIL;
1628 MemoryContextSwitchTo(old_cxt);
1632 * AtSubCommit_Notify() --- Take care of subtransaction commit.
1634 * Reassign all items in the pending lists to the parent transaction.
1637 AtSubCommit_Notify(void)
1639 List *parentPendingActions;
1640 List *parentPendingNotifies;
1642 parentPendingActions = (List *) linitial(upperPendingActions);
1643 upperPendingActions = list_delete_first(upperPendingActions);
1645 Assert(list_length(upperPendingActions) ==
1646 GetCurrentTransactionNestLevel() - 2);
1649 * Mustn't try to eliminate duplicates here --- see queue_listen()
1651 pendingActions = list_concat(parentPendingActions, pendingActions);
1653 parentPendingNotifies = (List *) linitial(upperPendingNotifies);
1654 upperPendingNotifies = list_delete_first(upperPendingNotifies);
1656 Assert(list_length(upperPendingNotifies) ==
1657 GetCurrentTransactionNestLevel() - 2);
1660 * We could try to eliminate duplicates here, but it seems not worthwhile.
1662 pendingNotifies = list_concat(parentPendingNotifies, pendingNotifies);
1666 * AtSubAbort_Notify() --- Take care of subtransaction abort.
1669 AtSubAbort_Notify(void)
1671 int my_level = GetCurrentTransactionNestLevel();
1674 * All we have to do is pop the stack --- the actions/notifies made in
1675 * this subxact are no longer interesting, and the space will be freed
1676 * when CurTransactionContext is recycled.
1678 * This routine could be called more than once at a given nesting level if
1679 * there is trouble during subxact abort. Avoid dumping core by using
1680 * GetCurrentTransactionNestLevel as the indicator of how far we need to
1683 while (list_length(upperPendingActions) > my_level - 2)
1685 pendingActions = (List *) linitial(upperPendingActions);
1686 upperPendingActions = list_delete_first(upperPendingActions);
1689 while (list_length(upperPendingNotifies) > my_level - 2)
1691 pendingNotifies = (List *) linitial(upperPendingNotifies);
1692 upperPendingNotifies = list_delete_first(upperPendingNotifies);
1697 * HandleNotifyInterrupt
1699 * Signal handler portion of interrupt handling. Let the backend know
1700 * that there's a pending notify interrupt. If we're currently reading
1701 * from the client, this will interrupt the read and
1702 * ProcessClientReadInterrupt() will call ProcessNotifyInterrupt().
1705 HandleNotifyInterrupt(void)
1708 * Note: this is called by a SIGNAL HANDLER. You must be very wary what
1712 /* signal that work needs to be done */
1713 notifyInterruptPending = true;
1715 /* make sure the event is processed in due course */
1720 * ProcessNotifyInterrupt
1722 * This is called just after waiting for a frontend command. If a
1723 * interrupt arrives (via HandleNotifyInterrupt()) while reading, the
1724 * read will be interrupted via the process's latch, and this routine
1725 * will get called. If we are truly idle (ie, *not* inside a transaction
1726 * block), process the incoming notifies.
1729 ProcessNotifyInterrupt(void)
1731 if (IsTransactionOrTransactionBlock())
1732 return; /* not really idle */
1734 while (notifyInterruptPending)
1735 ProcessIncomingNotify();
1740 * Read all pending notifications from the queue, and deliver appropriate
1741 * ones to my frontend. Stop when we reach queue head or an uncommitted
1745 asyncQueueReadAllNotifications(void)
1747 volatile QueuePosition pos;
1748 QueuePosition oldpos;
1752 /* page_buffer must be adequately aligned, so use a union */
1755 char buf[QUEUE_PAGESIZE];
1756 AsyncQueueEntry align;
1759 /* Fetch current state */
1760 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1761 /* Assert checks that we have a valid state entry */
1762 Assert(MyProcPid == QUEUE_BACKEND_PID(MyBackendId));
1763 pos = oldpos = QUEUE_BACKEND_POS(MyBackendId);
1765 LWLockRelease(AsyncQueueLock);
1767 if (QUEUE_POS_EQUAL(pos, head))
1769 /* Nothing to do, we have read all notifications already. */
1774 * Note that we deliver everything that we see in the queue and that
1775 * matches our _current_ listening state.
1776 * Especially we do not take into account different commit times.
1777 * Consider the following example:
1779 * Backend 1: Backend 2:
1781 * transaction starts
1784 * transaction starts
1790 * It could happen that backend 2 sees the notification from backend 1 in
1791 * the queue. Even though the notifying transaction committed before
1792 * the listening transaction, we still deliver the notification.
1794 * The idea is that an additional notification does not do any harm, we
1795 * just need to make sure that we do not miss a notification.
1797 * It is possible that we fail while trying to send a message to our
1798 * frontend (for example, because of encoding conversion failure).
1799 * If that happens it is critical that we not try to send the same
1800 * message over and over again. Therefore, we place a PG_TRY block
1801 * here that will forcibly advance our backend position before we lose
1802 * control to an error. (We could alternatively retake AsyncQueueLock
1803 * and move the position before handling each individual message, but
1804 * that seems like too much lock traffic.)
1813 int curpage = QUEUE_POS_PAGE(pos);
1814 int curoffset = QUEUE_POS_OFFSET(pos);
1819 * We copy the data from SLRU into a local buffer, so as to avoid
1820 * holding the AsyncCtlLock while we are examining the entries and
1821 * possibly transmitting them to our frontend. Copy only the part
1822 * of the page we will actually inspect.
1824 slotno = SimpleLruReadPage_ReadOnly(AsyncCtl, curpage,
1825 InvalidTransactionId);
1826 if (curpage == QUEUE_POS_PAGE(head))
1828 /* we only want to read as far as head */
1829 copysize = QUEUE_POS_OFFSET(head) - curoffset;
1831 copysize = 0; /* just for safety */
1835 /* fetch all the rest of the page */
1836 copysize = QUEUE_PAGESIZE - curoffset;
1838 memcpy(page_buffer.buf + curoffset,
1839 AsyncCtl->shared->page_buffer[slotno] + curoffset,
1841 /* Release lock that we got from SimpleLruReadPage_ReadOnly() */
1842 LWLockRelease(AsyncCtlLock);
1845 * Process messages up to the stop position, end of page, or an
1846 * uncommitted message.
1848 * Our stop position is what we found to be the head's position
1849 * when we entered this function. It might have changed already.
1850 * But if it has, we will receive (or have already received and
1851 * queued) another signal and come here again.
1853 * We are not holding AsyncQueueLock here! The queue can only
1854 * extend beyond the head pointer (see above) and we leave our
1855 * backend's pointer where it is so nobody will truncate or
1856 * rewrite pages under us. Especially we don't want to hold a lock
1857 * while sending the notifications to the frontend.
1859 reachedStop = asyncQueueProcessPageEntries(&pos, head,
1861 } while (!reachedStop);
1865 /* Update shared state */
1866 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1867 QUEUE_BACKEND_POS(MyBackendId) = pos;
1868 advanceTail = QUEUE_POS_EQUAL(oldpos, QUEUE_TAIL);
1869 LWLockRelease(AsyncQueueLock);
1871 /* If we were the laziest backend, try to advance the tail pointer */
1873 asyncQueueAdvanceTail();
1879 /* Update shared state */
1880 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1881 QUEUE_BACKEND_POS(MyBackendId) = pos;
1882 advanceTail = QUEUE_POS_EQUAL(oldpos, QUEUE_TAIL);
1883 LWLockRelease(AsyncQueueLock);
1885 /* If we were the laziest backend, try to advance the tail pointer */
1887 asyncQueueAdvanceTail();
1891 * Fetch notifications from the shared queue, beginning at position current,
1892 * and deliver relevant ones to my frontend.
1894 * The current page must have been fetched into page_buffer from shared
1895 * memory. (We could access the page right in shared memory, but that
1896 * would imply holding the AsyncCtlLock throughout this routine.)
1898 * We stop if we reach the "stop" position, or reach a notification from an
1899 * uncommitted transaction, or reach the end of the page.
1901 * The function returns true once we have reached the stop position or an
1902 * uncommitted notification, and false if we have finished with the page.
1903 * In other words: once it returns true there is no need to look further.
1904 * The QueuePosition *current is advanced past all processed messages.
1907 asyncQueueProcessPageEntries(volatile QueuePosition *current,
1911 bool reachedStop = false;
1912 bool reachedEndOfPage;
1913 AsyncQueueEntry *qe;
1917 QueuePosition thisentry = *current;
1919 if (QUEUE_POS_EQUAL(thisentry, stop))
1922 qe = (AsyncQueueEntry *) (page_buffer + QUEUE_POS_OFFSET(thisentry));
1925 * Advance *current over this message, possibly to the next page. As
1926 * noted in the comments for asyncQueueReadAllNotifications, we must
1927 * do this before possibly failing while processing the message.
1929 reachedEndOfPage = asyncQueueAdvance(current, qe->length);
1931 /* Ignore messages destined for other databases */
1932 if (qe->dboid == MyDatabaseId)
1934 if (TransactionIdIsInProgress(qe->xid))
1937 * The source transaction is still in progress, so we can't
1938 * process this message yet. Break out of the loop, but first
1939 * back up *current so we will reprocess the message next
1940 * time. (Note: it is unlikely but not impossible for
1941 * TransactionIdDidCommit to fail, so we can't really avoid
1942 * this advance-then-back-up behavior when dealing with an
1943 * uncommitted message.)
1945 * Note that we must test TransactionIdIsInProgress before we
1946 * test TransactionIdDidCommit, else we might return a message
1947 * from a transaction that is not yet visible to snapshots;
1948 * compare the comments at the head of tqual.c.
1950 *current = thisentry;
1954 else if (TransactionIdDidCommit(qe->xid))
1956 /* qe->data is the null-terminated channel name */
1957 char *channel = qe->data;
1959 if (IsListeningOn(channel))
1961 /* payload follows channel name */
1962 char *payload = qe->data + strlen(channel) + 1;
1964 NotifyMyFrontEnd(channel, payload, qe->srcPid);
1970 * The source transaction aborted or crashed, so we just
1971 * ignore its notifications.
1976 /* Loop back if we're not at end of page */
1977 } while (!reachedEndOfPage);
1979 if (QUEUE_POS_EQUAL(*current, stop))
1986 * Advance the shared queue tail variable to the minimum of all the
1987 * per-backend tail pointers. Truncate pg_notify space if possible.
1990 asyncQueueAdvanceTail(void)
1998 LWLockAcquire(AsyncQueueLock, LW_EXCLUSIVE);
2000 for (i = 1; i <= MaxBackends; i++)
2002 if (QUEUE_BACKEND_PID(i) != InvalidPid)
2003 min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
2005 oldtailpage = QUEUE_POS_PAGE(QUEUE_TAIL);
2007 LWLockRelease(AsyncQueueLock);
2010 * We can truncate something if the global tail advanced across an SLRU
2013 * XXX it might be better to truncate only once every several segments, to
2014 * reduce the number of directory scans.
2016 newtailpage = QUEUE_POS_PAGE(min);
2017 boundary = newtailpage - (newtailpage % SLRU_PAGES_PER_SEGMENT);
2018 if (asyncQueuePagePrecedes(oldtailpage, boundary))
2021 * SimpleLruTruncate() will ask for AsyncCtlLock but will also release
2024 SimpleLruTruncate(AsyncCtl, newtailpage);
2029 * ProcessIncomingNotify
2031 * Deal with arriving NOTIFYs from other backends as soon as it's safe to
2032 * do so. This used to be called from the PROCSIG_NOTIFY_INTERRUPT
2033 * signal handler, but isn't anymore.
2035 * Scan the queue for arriving notifications and report them to my front
2038 * NOTE: since we are outside any transaction, we must create our own.
2041 ProcessIncomingNotify(void)
2043 /* We *must* reset the flag */
2044 notifyInterruptPending = false;
2046 /* Do nothing else if we aren't actively listening */
2047 if (listenChannels == NIL)
2051 elog(DEBUG1, "ProcessIncomingNotify");
2053 set_ps_display("notify interrupt", false);
2056 * We must run asyncQueueReadAllNotifications inside a transaction, else
2057 * bad things happen if it gets an error.
2059 StartTransactionCommand();
2061 asyncQueueReadAllNotifications();
2063 CommitTransactionCommand();
2066 * Must flush the notify messages to ensure frontend gets them promptly.
2070 set_ps_display("idle", false);
2073 elog(DEBUG1, "ProcessIncomingNotify: done");
2077 * Send NOTIFY message to my front end.
2080 NotifyMyFrontEnd(const char *channel, const char *payload, int32 srcPid)
2082 if (whereToSendOutput == DestRemote)
2086 pq_beginmessage(&buf, 'A');
2087 pq_sendint(&buf, srcPid, sizeof(int32));
2088 pq_sendstring(&buf, channel);
2089 if (PG_PROTOCOL_MAJOR(FrontendProtocol) >= 3)
2090 pq_sendstring(&buf, payload);
2091 pq_endmessage(&buf);
2094 * NOTE: we do not do pq_flush() here. For a self-notify, it will
2095 * happen at the end of the transaction, and for incoming notifies
2096 * ProcessIncomingNotify will do it after finding all the notifies.
2100 elog(INFO, "NOTIFY for \"%s\" payload \"%s\"", channel, payload);
2103 /* Does pendingNotifies include the given channel/payload? */
2105 AsyncExistsPendingNotify(const char *channel, const char *payload)
2110 if (pendingNotifies == NIL)
2113 if (payload == NULL)
2117 * We need to append new elements to the end of the list in order to keep
2118 * the order. However, on the other hand we'd like to check the list
2119 * backwards in order to make duplicate-elimination a tad faster when the
2120 * same condition is signaled many times in a row. So as a compromise we
2121 * check the tail element first which we can access directly. If this
2122 * doesn't match, we check the whole list.
2124 * As we are not checking our parents' lists, we can still get duplicates
2125 * in combination with subtransactions, like in:
2134 n = (Notification *) llast(pendingNotifies);
2135 if (strcmp(n->channel, channel) == 0 &&
2136 strcmp(n->payload, payload) == 0)
2139 foreach(p, pendingNotifies)
2141 n = (Notification *) lfirst(p);
2143 if (strcmp(n->channel, channel) == 0 &&
2144 strcmp(n->payload, payload) == 0)
2151 /* Clear the pendingActions and pendingNotifies lists. */
2153 ClearPendingActionsAndNotifies(void)
2156 * We used to have to explicitly deallocate the list members and nodes,
2157 * because they were malloc'd. Now, since we know they are palloc'd in
2158 * CurTransactionContext, we need not do that --- they'll go away
2159 * automatically at transaction exit. We need only reset the list head
2162 pendingActions = NIL;
2163 pendingNotifies = NIL;