1 /*-------------------------------------------------------------------------
4 * Asynchronous notification: NOTIFY, LISTEN, UNLISTEN
6 * Portions Copyright (c) 1996-2011, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
10 * src/backend/commands/async.c
12 *-------------------------------------------------------------------------
15 /*-------------------------------------------------------------------------
16 * Async Notification Model as of 9.0:
18 * 1. Multiple backends on same machine. Multiple backends listening on
19 * several channels. (Channels are also called "conditions" in other
22 * 2. There is one central queue in disk-based storage (directory pg_notify/),
23 * with actively-used pages mapped into shared memory by the slru.c module.
24 * All notification messages are placed in the queue and later read out
25 * by listening backends.
27 * There is no central knowledge of which backend listens on which channel;
28 * every backend has its own list of interesting channels.
30 * Although there is only one queue, notifications are treated as being
31 * database-local; this is done by including the sender's database OID
32 * in each notification message. Listening backends ignore messages
33 * that don't match their database OID. This is important because it
34 * ensures senders and receivers have the same database encoding and won't
35 * misinterpret non-ASCII text in the channel name or payload string.
37 * Since notifications are not expected to survive database crashes,
38 * we can simply clean out the pg_notify data at any reboot, and there
39 * is no need for WAL support or fsync'ing.
41 * 3. Every backend that is listening on at least one channel registers by
42 * entering its PID into the array in AsyncQueueControl. It then scans all
43 * incoming notifications in the central queue and first compares the
44 * database OID of the notification with its own database OID and then
45 * compares the notified channel with the list of channels that it listens
46 * to. In case there is a match it delivers the notification event to its
47 * frontend. Non-matching events are simply skipped.
49 * 4. The NOTIFY statement (routine Async_Notify) stores the notification in
50 * a backend-local list which will not be processed until transaction end.
52 * Duplicate notifications from the same transaction are sent out as one
53 * notification only. This is done to save work when for example a trigger
54 * on a 2 million row table fires a notification for each row that has been
55 * changed. If the application needs to receive every single notification
56 * that has been sent, it can easily add some unique string into the extra
59 * When the transaction is ready to commit, PreCommit_Notify() adds the
60 * pending notifications to the head of the queue. The head pointer of the
61 * queue always points to the next free position and a position is just a
62 * page number and the offset in that page. This is done before marking the
63 * transaction as committed in clog. If we run into problems writing the
64 * notifications, we can still call elog(ERROR, ...) and the transaction
67 * Once we have put all of the notifications into the queue, we return to
68 * CommitTransaction() which will then do the actual transaction commit.
70 * After commit we are called another time (AtCommit_Notify()). Here we
71 * make the actual updates to the effective listen state (listenChannels).
73 * Finally, after we are out of the transaction altogether, we check if
74 * we need to signal listening backends. In SignalBackends() we scan the
75 * list of listening backends and send a PROCSIG_NOTIFY_INTERRUPT signal
76 * to every listening backend (we don't know which backend is listening on
77 * which channel so we must signal them all). We can exclude backends that
78 * are already up to date, though. We don't bother with a self-signal
79 * either, but just process the queue directly.
81 * 5. Upon receipt of a PROCSIG_NOTIFY_INTERRUPT signal, the signal handler
82 * can call inbound-notify processing immediately if this backend is idle
83 * (ie, it is waiting for a frontend command and is not within a transaction
84 * block). Otherwise the handler may only set a flag, which will cause the
85 * processing to occur just before we next go idle.
87 * Inbound-notify processing consists of reading all of the notifications
88 * that have arrived since scanning last time. We read every notification
89 * until we reach either a notification from an uncommitted transaction or
90 * the head pointer's position. Then we check if we were the laziest
91 * backend: if our pointer is set to the same position as the global tail
92 * pointer is set, then we move the global tail pointer ahead to where the
93 * second-laziest backend is (in general, we take the MIN of the current
94 * head position and all active backends' new tail pointers). Whenever we
95 * move the global tail pointer we also truncate now-unused pages (i.e.,
96 * delete files in pg_notify/ that are no longer used).
98 * An application that listens on the same channel it notifies will get
99 * NOTIFY messages for its own NOTIFYs. These can be ignored, if not useful,
100 * by comparing be_pid in the NOTIFY message to the application's own backend's
101 * PID. (As of FE/BE protocol 2.0, the backend's PID is provided to the
102 * frontend during startup.) The above design guarantees that notifies from
103 * other backends will never be missed by ignoring self-notifies.
105 * The amount of shared memory used for notify management (NUM_ASYNC_BUFFERS)
106 * can be varied without affecting anything but performance. The maximum
107 * amount of notification data that can be queued at one time is determined
108 * by slru.c's wraparound limit; see QUEUE_MAX_PAGE below.
109 *-------------------------------------------------------------------------
112 #include "postgres.h"
118 #include "access/slru.h"
119 #include "access/transam.h"
120 #include "access/xact.h"
121 #include "catalog/pg_database.h"
122 #include "commands/async.h"
124 #include "libpq/libpq.h"
125 #include "libpq/pqformat.h"
126 #include "miscadmin.h"
127 #include "storage/ipc.h"
128 #include "storage/lmgr.h"
129 #include "storage/procsignal.h"
130 #include "storage/sinval.h"
131 #include "tcop/tcopprot.h"
132 #include "utils/builtins.h"
133 #include "utils/memutils.h"
134 #include "utils/ps_status.h"
135 #include "utils/timestamp.h"
139 * Maximum size of a NOTIFY payload, including terminating NULL. This
140 * must be kept small enough so that a notification message fits on one
141 * SLRU page. The magic fudge factor here is noncritical as long as it's
142 * more than AsyncQueueEntryEmptySize --- we make it significantly bigger
143 * than that, so changes in that data structure won't affect user-visible
146 #define NOTIFY_PAYLOAD_MAX_LENGTH (BLCKSZ - NAMEDATALEN - 128)
149 * Struct representing an entry in the global notify queue
151 * This struct declaration has the maximal length, but in a real queue entry
152 * the data area is only big enough for the actual channel and payload strings
153 * (each null-terminated). AsyncQueueEntryEmptySize is the minimum possible
154 * entry size, if both channel and payload strings are empty (but note it
155 * doesn't include alignment padding).
157 * The "length" field should always be rounded up to the next QUEUEALIGN
158 * multiple so that all fields are properly aligned.
160 typedef struct AsyncQueueEntry
162 int length; /* total allocated length of entry */
163 Oid dboid; /* sender's database OID */
164 TransactionId xid; /* sender's XID */
165 int32 srcPid; /* sender's PID */
166 char data[NAMEDATALEN + NOTIFY_PAYLOAD_MAX_LENGTH];
169 /* Currently, no field of AsyncQueueEntry requires more than int alignment */
170 #define QUEUEALIGN(len) INTALIGN(len)
172 #define AsyncQueueEntryEmptySize (offsetof(AsyncQueueEntry, data) + 2)
175 * Struct describing a queue position, and assorted macros for working with it
177 typedef struct QueuePosition
179 int page; /* SLRU page number */
180 int offset; /* byte offset within page */
183 #define QUEUE_POS_PAGE(x) ((x).page)
184 #define QUEUE_POS_OFFSET(x) ((x).offset)
186 #define SET_QUEUE_POS(x,y,z) \
192 #define QUEUE_POS_EQUAL(x,y) \
193 ((x).page == (y).page && (x).offset == (y).offset)
195 /* choose logically smaller QueuePosition */
196 #define QUEUE_POS_MIN(x,y) \
197 (asyncQueuePagePrecedesLogically((x).page, (y).page) ? (x) : \
198 (x).page != (y).page ? (y) : \
199 (x).offset < (y).offset ? (x) : (y))
202 * Struct describing a listening backend's status
204 typedef struct QueueBackendStatus
206 int32 pid; /* either a PID or InvalidPid */
207 QueuePosition pos; /* backend has read queue up to here */
208 } QueueBackendStatus;
210 #define InvalidPid (-1)
213 * Shared memory state for LISTEN/NOTIFY (excluding its SLRU stuff)
215 * The AsyncQueueControl structure is protected by the AsyncQueueLock.
217 * When holding the lock in SHARED mode, backends may only inspect their own
218 * entries as well as the head and tail pointers. Consequently we can allow a
219 * backend to update its own record while holding only SHARED lock (since no
220 * other backend will inspect it).
222 * When holding the lock in EXCLUSIVE mode, backends can inspect the entries
223 * of other backends and also change the head and tail pointers.
225 * In order to avoid deadlocks, whenever we need both locks, we always first
226 * get AsyncQueueLock and then AsyncCtlLock.
228 * Each backend uses the backend[] array entry with index equal to its
229 * BackendId (which can range from 1 to MaxBackends). We rely on this to make
230 * SendProcSignal fast.
232 typedef struct AsyncQueueControl
234 QueuePosition head; /* head points to the next free location */
235 QueuePosition tail; /* the global tail is equivalent to the tail
236 * of the "slowest" backend */
237 TimestampTz lastQueueFillWarn; /* time of last queue-full msg */
238 QueueBackendStatus backend[1]; /* actually of length MaxBackends+1 */
239 /* DO NOT ADD FURTHER STRUCT MEMBERS HERE */
242 static AsyncQueueControl *asyncQueueControl;
244 #define QUEUE_HEAD (asyncQueueControl->head)
245 #define QUEUE_TAIL (asyncQueueControl->tail)
246 #define QUEUE_BACKEND_PID(i) (asyncQueueControl->backend[i].pid)
247 #define QUEUE_BACKEND_POS(i) (asyncQueueControl->backend[i].pos)
250 * The SLRU buffer area through which we access the notification queue
252 static SlruCtlData AsyncCtlData;
254 #define AsyncCtl (&AsyncCtlData)
255 #define QUEUE_PAGESIZE BLCKSZ
256 #define QUEUE_FULL_WARN_INTERVAL 5000 /* warn at most once every 5s */
259 * slru.c currently assumes that all filenames are four characters of hex
260 * digits. That means that we can use segments 0000 through FFFF.
261 * Each segment contains SLRU_PAGES_PER_SEGMENT pages which gives us
262 * the pages from 0 to SLRU_PAGES_PER_SEGMENT * 0x10000 - 1.
264 * It's of course possible to enhance slru.c, but this gives us so much
265 * space already that it doesn't seem worth the trouble.
267 * The most data we can have in the queue at a time is QUEUE_MAX_PAGE/2
268 * pages, because more than that would confuse slru.c into thinking there
269 * was a wraparound condition. With the default BLCKSZ this means there
270 * can be up to 8GB of queued-and-not-read data.
272 * Note: it's possible to redefine QUEUE_MAX_PAGE with a smaller multiple of
273 * SLRU_PAGES_PER_SEGMENT, for easier testing of queue-full behaviour.
275 #define QUEUE_MAX_PAGE (SLRU_PAGES_PER_SEGMENT * 0x10000 - 1)
278 * listenChannels identifies the channels we are actually listening to
279 * (ie, have committed a LISTEN on). It is a simple list of channel names,
280 * allocated in TopMemoryContext.
282 static List *listenChannels = NIL; /* list of C strings */
285 * State for pending LISTEN/UNLISTEN actions consists of an ordered list of
286 * all actions requested in the current transaction. As explained above,
287 * we don't actually change listenChannels until we reach transaction commit.
289 * The list is kept in CurTransactionContext. In subtransactions, each
290 * subtransaction has its own list in its own CurTransactionContext, but
291 * successful subtransactions attach their lists to their parent's list.
292 * Failed subtransactions simply discard their lists.
303 ListenActionKind action;
304 char channel[1]; /* actually, as long as needed */
307 static List *pendingActions = NIL; /* list of ListenAction */
309 static List *upperPendingActions = NIL; /* list of upper-xact lists */
312 * State for outbound notifies consists of a list of all channels+payloads
313 * NOTIFYed in the current transaction. We do not actually perform a NOTIFY
314 * until and unless the transaction commits. pendingNotifies is NIL if no
315 * NOTIFYs have been done in the current transaction.
317 * The list is kept in CurTransactionContext. In subtransactions, each
318 * subtransaction has its own list in its own CurTransactionContext, but
319 * successful subtransactions attach their lists to their parent's list.
320 * Failed subtransactions simply discard their lists.
322 * Note: the action and notify lists do not interact within a transaction.
323 * In particular, if a transaction does NOTIFY and then LISTEN on the same
324 * condition name, it will get a self-notify at commit. This is a bit odd
325 * but is consistent with our historical behavior.
327 typedef struct Notification
329 char *channel; /* channel name */
330 char *payload; /* payload string (can be empty) */
333 static List *pendingNotifies = NIL; /* list of Notifications */
335 static List *upperPendingNotifies = NIL; /* list of upper-xact lists */
338 * State for inbound notifications consists of two flags: one saying whether
339 * the signal handler is currently allowed to call ProcessIncomingNotify
340 * directly, and one saying whether the signal has occurred but the handler
341 * was not allowed to call ProcessIncomingNotify at the time.
343 * NB: the "volatile" on these declarations is critical! If your compiler
344 * does not grok "volatile", you'd be best advised to compile this file
345 * with all optimization turned off.
347 static volatile sig_atomic_t notifyInterruptEnabled = 0;
348 static volatile sig_atomic_t notifyInterruptOccurred = 0;
350 /* True if we've registered an on_shmem_exit cleanup */
351 static bool unlistenExitRegistered = false;
353 /* has this backend sent notifications in the current transaction? */
354 static bool backendHasSentNotifications = false;
356 /* has this backend executed its first LISTEN in the current transaction? */
357 static bool backendHasExecutedInitialListen = false;
360 bool Trace_notify = false;
362 /* local function prototypes */
363 static bool asyncQueuePagePrecedesPhysically(int p, int q);
364 static bool asyncQueuePagePrecedesLogically(int p, int q);
365 static void queue_listen(ListenActionKind action, const char *channel);
366 static void Async_UnlistenOnExit(int code, Datum arg);
367 static void Exec_ListenPreCommit(void);
368 static void Exec_ListenCommit(const char *channel);
369 static void Exec_UnlistenCommit(const char *channel);
370 static void Exec_UnlistenAllCommit(void);
371 static bool IsListeningOn(const char *channel);
372 static void asyncQueueUnregister(void);
373 static bool asyncQueueIsFull(void);
374 static bool asyncQueueAdvance(QueuePosition *position, int entryLength);
375 static void asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe);
376 static ListCell *asyncQueueAddEntries(ListCell *nextNotify);
377 static void asyncQueueFillWarning(void);
378 static bool SignalBackends(void);
379 static void asyncQueueReadAllNotifications(void);
380 static bool asyncQueueProcessPageEntries(QueuePosition *current,
383 static void asyncQueueAdvanceTail(void);
384 static void ProcessIncomingNotify(void);
385 static void NotifyMyFrontEnd(const char *channel,
388 static bool AsyncExistsPendingNotify(const char *channel, const char *payload);
389 static void ClearPendingActionsAndNotifies(void);
393 * We will work on the page range of 0..QUEUE_MAX_PAGE.
395 * asyncQueuePagePrecedesPhysically just checks numerically without any magic
396 * if one page precedes another one. This is wrong for normal operation but
397 * is helpful when clearing pg_notify/ during startup.
399 * asyncQueuePagePrecedesLogically compares using wraparound logic, as is
400 * required by slru.c.
403 asyncQueuePagePrecedesPhysically(int p, int q)
409 asyncQueuePagePrecedesLogically(int p, int q)
414 * We have to compare modulo (QUEUE_MAX_PAGE+1)/2. Both inputs should be
415 * in the range 0..QUEUE_MAX_PAGE.
417 Assert(p >= 0 && p <= QUEUE_MAX_PAGE);
418 Assert(q >= 0 && q <= QUEUE_MAX_PAGE);
421 if (diff >= ((QUEUE_MAX_PAGE + 1) / 2))
422 diff -= QUEUE_MAX_PAGE + 1;
423 else if (diff < -((QUEUE_MAX_PAGE + 1) / 2))
424 diff += QUEUE_MAX_PAGE + 1;
429 * Report space needed for our shared memory area
436 /* This had better match AsyncShmemInit */
437 size = mul_size(MaxBackends, sizeof(QueueBackendStatus));
438 size = add_size(size, sizeof(AsyncQueueControl));
440 size = add_size(size, SimpleLruShmemSize(NUM_ASYNC_BUFFERS, 0));
446 * Initialize our shared memory area
456 * Create or attach to the AsyncQueueControl structure.
458 * The used entries in the backend[] array run from 1 to MaxBackends.
459 * sizeof(AsyncQueueControl) already includes space for the unused zero'th
460 * entry, but we need to add on space for the used entries.
462 size = mul_size(MaxBackends, sizeof(QueueBackendStatus));
463 size = add_size(size, sizeof(AsyncQueueControl));
465 asyncQueueControl = (AsyncQueueControl *)
466 ShmemInitStruct("Async Queue Control", size, &found);
470 /* First time through, so initialize it */
473 SET_QUEUE_POS(QUEUE_HEAD, 0, 0);
474 SET_QUEUE_POS(QUEUE_TAIL, 0, 0);
475 asyncQueueControl->lastQueueFillWarn = 0;
476 /* zero'th entry won't be used, but let's initialize it anyway */
477 for (i = 0; i <= MaxBackends; i++)
479 QUEUE_BACKEND_PID(i) = InvalidPid;
480 SET_QUEUE_POS(QUEUE_BACKEND_POS(i), 0, 0);
485 * Set up SLRU management of the pg_notify data.
487 AsyncCtl->PagePrecedes = asyncQueuePagePrecedesLogically;
488 SimpleLruInit(AsyncCtl, "Async Ctl", NUM_ASYNC_BUFFERS, 0,
489 AsyncCtlLock, "pg_notify");
490 /* Override default assumption that writes should be fsync'd */
491 AsyncCtl->do_fsync = false;
496 * During start or reboot, clean out the pg_notify directory.
498 * Since we want to remove every file, we temporarily use
499 * asyncQueuePagePrecedesPhysically() and pass INT_MAX as the
500 * comparison value; every file in the directory should therefore
501 * appear to be less than that.
503 AsyncCtl->PagePrecedes = asyncQueuePagePrecedesPhysically;
504 (void) SlruScanDirectory(AsyncCtl, INT_MAX, true);
505 AsyncCtl->PagePrecedes = asyncQueuePagePrecedesLogically;
507 /* Now initialize page zero to empty */
508 LWLockAcquire(AsyncCtlLock, LW_EXCLUSIVE);
509 slotno = SimpleLruZeroPage(AsyncCtl, QUEUE_POS_PAGE(QUEUE_HEAD));
510 /* This write is just to verify that pg_notify/ is writable */
511 SimpleLruWritePage(AsyncCtl, slotno);
512 LWLockRelease(AsyncCtlLock);
519 * SQL function to send a notification event
522 pg_notify(PG_FUNCTION_ARGS)
530 channel = text_to_cstring(PG_GETARG_TEXT_PP(0));
535 payload = text_to_cstring(PG_GETARG_TEXT_PP(1));
537 /* For NOTIFY as a statement, this is checked in ProcessUtility */
538 PreventCommandDuringRecovery("NOTIFY");
540 Async_Notify(channel, payload);
549 * This is executed by the SQL notify command.
551 * Adds the message to the list of pending notifies.
552 * Actual notification happens during transaction commit.
553 * ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
556 Async_Notify(const char *channel, const char *payload)
559 MemoryContext oldcontext;
562 elog(DEBUG1, "Async_Notify(%s)", channel);
564 /* a channel name must be specified */
565 if (!channel || !strlen(channel))
567 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
568 errmsg("channel name cannot be empty")));
570 if (strlen(channel) >= NAMEDATALEN)
572 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
573 errmsg("channel name too long")));
577 if (strlen(payload) >= NOTIFY_PAYLOAD_MAX_LENGTH)
579 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
580 errmsg("payload string too long")));
583 /* no point in making duplicate entries in the list ... */
584 if (AsyncExistsPendingNotify(channel, payload))
588 * The notification list needs to live until end of transaction, so store
589 * it in the transaction context.
591 oldcontext = MemoryContextSwitchTo(CurTransactionContext);
593 n = (Notification *) palloc(sizeof(Notification));
594 n->channel = pstrdup(channel);
596 n->payload = pstrdup(payload);
601 * We want to preserve the order so we need to append every notification.
602 * See comments at AsyncExistsPendingNotify().
604 pendingNotifies = lappend(pendingNotifies, n);
606 MemoryContextSwitchTo(oldcontext);
611 * Common code for listen, unlisten, unlisten all commands.
613 * Adds the request to the list of pending actions.
614 * Actual update of the listenChannels list happens during transaction
618 queue_listen(ListenActionKind action, const char *channel)
620 MemoryContext oldcontext;
621 ListenAction *actrec;
624 * Unlike Async_Notify, we don't try to collapse out duplicates. It would
625 * be too complicated to ensure we get the right interactions of
626 * conflicting LISTEN/UNLISTEN/UNLISTEN_ALL, and it's unlikely that there
627 * would be any performance benefit anyway in sane applications.
629 oldcontext = MemoryContextSwitchTo(CurTransactionContext);
631 /* space for terminating null is included in sizeof(ListenAction) */
632 actrec = (ListenAction *) palloc(sizeof(ListenAction) + strlen(channel));
633 actrec->action = action;
634 strcpy(actrec->channel, channel);
636 pendingActions = lappend(pendingActions, actrec);
638 MemoryContextSwitchTo(oldcontext);
644 * This is executed by the SQL listen command.
647 Async_Listen(const char *channel)
650 elog(DEBUG1, "Async_Listen(%s,%d)", channel, MyProcPid);
652 queue_listen(LISTEN_LISTEN, channel);
658 * This is executed by the SQL unlisten command.
661 Async_Unlisten(const char *channel)
664 elog(DEBUG1, "Async_Unlisten(%s,%d)", channel, MyProcPid);
666 /* If we couldn't possibly be listening, no need to queue anything */
667 if (pendingActions == NIL && !unlistenExitRegistered)
670 queue_listen(LISTEN_UNLISTEN, channel);
676 * This is invoked by UNLISTEN * command, and also at backend exit.
679 Async_UnlistenAll(void)
682 elog(DEBUG1, "Async_UnlistenAll(%d)", MyProcPid);
684 /* If we couldn't possibly be listening, no need to queue anything */
685 if (pendingActions == NIL && !unlistenExitRegistered)
688 queue_listen(LISTEN_UNLISTEN_ALL, "");
692 * SQL function: return a set of the channel names this backend is actively
695 * Note: this coding relies on the fact that the listenChannels list cannot
696 * change within a transaction.
699 pg_listening_channels(PG_FUNCTION_ARGS)
701 FuncCallContext *funcctx;
704 /* stuff done only on the first call of the function */
705 if (SRF_IS_FIRSTCALL())
707 MemoryContext oldcontext;
709 /* create a function context for cross-call persistence */
710 funcctx = SRF_FIRSTCALL_INIT();
712 /* switch to memory context appropriate for multiple function calls */
713 oldcontext = MemoryContextSwitchTo(funcctx->multi_call_memory_ctx);
715 /* allocate memory for user context */
716 lcp = (ListCell **) palloc(sizeof(ListCell *));
717 *lcp = list_head(listenChannels);
718 funcctx->user_fctx = (void *) lcp;
720 MemoryContextSwitchTo(oldcontext);
723 /* stuff done on every call of the function */
724 funcctx = SRF_PERCALL_SETUP();
725 lcp = (ListCell **) funcctx->user_fctx;
729 char *channel = (char *) lfirst(*lcp);
732 SRF_RETURN_NEXT(funcctx, CStringGetTextDatum(channel));
735 SRF_RETURN_DONE(funcctx);
739 * Async_UnlistenOnExit
741 * This is executed at backend exit if we have done any LISTENs in this
742 * backend. It might not be necessary anymore, if the user UNLISTENed
743 * everything, but we don't try to detect that case.
746 Async_UnlistenOnExit(int code, Datum arg)
748 Exec_UnlistenAllCommit();
754 * This is called at the prepare phase of a two-phase
755 * transaction. Save the state for possible commit later.
758 AtPrepare_Notify(void)
760 /* It's not allowed to have any pending LISTEN/UNLISTEN/NOTIFY actions */
761 if (pendingActions || pendingNotifies)
763 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
764 errmsg("cannot PREPARE a transaction that has executed LISTEN, UNLISTEN or NOTIFY")));
770 * This is called at transaction commit, before actually committing to
773 * If there are pending LISTEN actions, make sure we are listed in the
774 * shared-memory listener array. This must happen before commit to
775 * ensure we don't miss any notifies from transactions that commit
778 * If there are outbound notify requests in the pendingNotifies list,
779 * add them to the global queue. We do that before commit so that
780 * we can still throw error if we run out of queue space.
783 PreCommit_Notify(void)
787 if (pendingActions == NIL && pendingNotifies == NIL)
788 return; /* no relevant statements in this xact */
791 elog(DEBUG1, "PreCommit_Notify");
793 Assert(backendHasExecutedInitialListen == false);
795 /* Preflight for any pending listen/unlisten actions */
796 foreach(p, pendingActions)
798 ListenAction *actrec = (ListenAction *) lfirst(p);
800 switch (actrec->action)
803 Exec_ListenPreCommit();
805 case LISTEN_UNLISTEN:
806 /* there is no Exec_UnlistenPreCommit() */
808 case LISTEN_UNLISTEN_ALL:
809 /* there is no Exec_UnlistenAllPreCommit() */
814 /* Queue any pending notifies */
817 ListCell *nextNotify;
820 * Make sure that we have an XID assigned to the current transaction.
821 * GetCurrentTransactionId is cheap if we already have an XID, but not
822 * so cheap if we don't, and we'd prefer not to do that work while
823 * holding AsyncQueueLock.
825 (void) GetCurrentTransactionId();
828 * Serialize writers by acquiring a special lock that we hold till
829 * after commit. This ensures that queue entries appear in commit
830 * order, and in particular that there are never uncommitted queue
831 * entries ahead of committed ones, so an uncommitted transaction
832 * can't block delivery of deliverable notifications.
834 * We use a heavyweight lock so that it'll automatically be released
835 * after either commit or abort. This also allows deadlocks to be
836 * detected, though really a deadlock shouldn't be possible here.
838 * The lock is on "database 0", which is pretty ugly but it doesn't
839 * seem worth inventing a special locktag category just for this.
840 * (Historical note: before PG 9.0, a similar lock on "database 0" was
841 * used by the flatfiles mechanism.)
843 LockSharedObject(DatabaseRelationId, InvalidOid, 0,
844 AccessExclusiveLock);
846 /* Now push the notifications into the queue */
847 backendHasSentNotifications = true;
849 nextNotify = list_head(pendingNotifies);
850 while (nextNotify != NULL)
853 * Add the pending notifications to the queue. We acquire and
854 * release AsyncQueueLock once per page, which might be overkill
855 * but it does allow readers to get in while we're doing this.
857 * A full queue is very uncommon and should really not happen,
858 * given that we have so much space available in the SLRU pages.
859 * Nevertheless we need to deal with this possibility. Note that
860 * when we get here we are in the process of committing our
861 * transaction, but we have not yet committed to clog, so at this
862 * point in time we can still roll the transaction back.
864 LWLockAcquire(AsyncQueueLock, LW_EXCLUSIVE);
865 asyncQueueFillWarning();
866 if (asyncQueueIsFull())
868 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
869 errmsg("too many notifications in the NOTIFY queue")));
870 nextNotify = asyncQueueAddEntries(nextNotify);
871 LWLockRelease(AsyncQueueLock);
879 * This is called at transaction commit, after committing to clog.
881 * Update listenChannels and clear transaction-local state.
884 AtCommit_Notify(void)
889 * Allow transactions that have not executed LISTEN/UNLISTEN/NOTIFY to
890 * return as soon as possible
892 if (!pendingActions && !pendingNotifies)
896 elog(DEBUG1, "AtCommit_Notify");
898 /* Perform any pending listen/unlisten actions */
899 foreach(p, pendingActions)
901 ListenAction *actrec = (ListenAction *) lfirst(p);
903 switch (actrec->action)
906 Exec_ListenCommit(actrec->channel);
908 case LISTEN_UNLISTEN:
909 Exec_UnlistenCommit(actrec->channel);
911 case LISTEN_UNLISTEN_ALL:
912 Exec_UnlistenAllCommit();
918 * If we did an initial LISTEN, listenChannels now has the entry, so we no
919 * longer need or want the flag to be set.
921 backendHasExecutedInitialListen = false;
924 ClearPendingActionsAndNotifies();
928 * Exec_ListenPreCommit --- subroutine for PreCommit_Notify
930 * This function must make sure we are ready to catch any incoming messages.
933 Exec_ListenPreCommit(void)
936 * Nothing to do if we are already listening to something, nor if we
937 * already ran this routine in this transaction.
939 if (listenChannels != NIL || backendHasExecutedInitialListen)
943 elog(DEBUG1, "Exec_ListenPreCommit(%d)", MyProcPid);
946 * We need this variable to detect an aborted initial LISTEN. In that case
947 * we would set up our pointer but not listen on any channel. This flag
948 * gets cleared in AtCommit_Notify or AtAbort_Notify().
950 backendHasExecutedInitialListen = true;
953 * Before registering, make sure we will unlisten before dying. (Note:
954 * this action does not get undone if we abort later.)
956 if (!unlistenExitRegistered)
958 on_shmem_exit(Async_UnlistenOnExit, 0);
959 unlistenExitRegistered = true;
963 * This is our first LISTEN, so establish our pointer.
965 * We set our pointer to the global tail pointer and then move it forward
966 * over already-committed notifications. This ensures we cannot miss any
967 * not-yet-committed notifications. We might get a few more but that
970 LWLockAcquire(AsyncQueueLock, LW_SHARED);
971 QUEUE_BACKEND_POS(MyBackendId) = QUEUE_TAIL;
972 QUEUE_BACKEND_PID(MyBackendId) = MyProcPid;
973 LWLockRelease(AsyncQueueLock);
976 * Try to move our pointer forward as far as possible. This will skip over
977 * already-committed notifications. Still, we could get notifications that
978 * have already committed before we started to LISTEN.
980 * Note that we are not yet listening on anything, so we won't deliver any
981 * notification to the frontend.
983 * This will also advance the global tail pointer if possible.
985 asyncQueueReadAllNotifications();
989 * Exec_ListenCommit --- subroutine for AtCommit_Notify
991 * Add the channel to the list of channels we are listening on.
994 Exec_ListenCommit(const char *channel)
996 MemoryContext oldcontext;
998 /* Do nothing if we are already listening on this channel */
999 if (IsListeningOn(channel))
1003 * Add the new channel name to listenChannels.
1005 * XXX It is theoretically possible to get an out-of-memory failure here,
1006 * which would be bad because we already committed. For the moment it
1007 * doesn't seem worth trying to guard against that, but maybe improve this
1010 oldcontext = MemoryContextSwitchTo(TopMemoryContext);
1011 listenChannels = lappend(listenChannels, pstrdup(channel));
1012 MemoryContextSwitchTo(oldcontext);
1016 * Exec_UnlistenCommit --- subroutine for AtCommit_Notify
1018 * Remove the specified channel name from listenChannels.
1021 Exec_UnlistenCommit(const char *channel)
1027 elog(DEBUG1, "Exec_UnlistenCommit(%s,%d)", channel, MyProcPid);
1030 foreach(q, listenChannels)
1032 char *lchan = (char *) lfirst(q);
1034 if (strcmp(lchan, channel) == 0)
1036 listenChannels = list_delete_cell(listenChannels, q, prev);
1044 * We do not complain about unlistening something not being listened;
1048 /* If no longer listening to anything, get out of listener array */
1049 if (listenChannels == NIL)
1050 asyncQueueUnregister();
1054 * Exec_UnlistenAllCommit --- subroutine for AtCommit_Notify
1056 * Unlisten on all channels for this backend.
1059 Exec_UnlistenAllCommit(void)
1062 elog(DEBUG1, "Exec_UnlistenAllCommit(%d)", MyProcPid);
1064 list_free_deep(listenChannels);
1065 listenChannels = NIL;
1067 asyncQueueUnregister();
1071 * ProcessCompletedNotifies --- send out signals and self-notifies
1073 * This is called from postgres.c just before going idle at the completion
1074 * of a transaction. If we issued any notifications in the just-completed
1075 * transaction, send signals to other backends to process them, and also
1076 * process the queue ourselves to send messages to our own frontend.
1078 * The reason that this is not done in AtCommit_Notify is that there is
1079 * a nonzero chance of errors here (for example, encoding conversion errors
1080 * while trying to format messages to our frontend). An error during
1081 * AtCommit_Notify would be a PANIC condition. The timing is also arranged
1082 * to ensure that a transaction's self-notifies are delivered to the frontend
1083 * before it gets the terminating ReadyForQuery message.
1085 * Note that we send signals and process the queue even if the transaction
1086 * eventually aborted. This is because we need to clean out whatever got
1087 * added to the queue.
1089 * NOTE: we are outside of any transaction here.
1092 ProcessCompletedNotifies(void)
1094 MemoryContext caller_context;
1097 /* Nothing to do if we didn't send any notifications */
1098 if (!backendHasSentNotifications)
1102 * We reset the flag immediately; otherwise, if any sort of error occurs
1103 * below, we'd be locked up in an infinite loop, because control will come
1104 * right back here after error cleanup.
1106 backendHasSentNotifications = false;
1109 * We must preserve the caller's memory context (probably MessageContext)
1110 * across the transaction we do here.
1112 caller_context = CurrentMemoryContext;
1115 elog(DEBUG1, "ProcessCompletedNotifies");
1118 * We must run asyncQueueReadAllNotifications inside a transaction, else
1119 * bad things happen if it gets an error.
1121 StartTransactionCommand();
1123 /* Send signals to other backends */
1124 signalled = SignalBackends();
1126 if (listenChannels != NIL)
1128 /* Read the queue ourselves, and send relevant stuff to the frontend */
1129 asyncQueueReadAllNotifications();
1131 else if (!signalled)
1134 * If we found no other listening backends, and we aren't listening
1135 * ourselves, then we must execute asyncQueueAdvanceTail to flush the
1136 * queue, because ain't nobody else gonna do it. This prevents queue
1137 * overflow when we're sending useless notifies to nobody. (A new
1138 * listener could have joined since we looked, but if so this is
1141 asyncQueueAdvanceTail();
1144 CommitTransactionCommand();
1146 MemoryContextSwitchTo(caller_context);
1148 /* We don't need pq_flush() here since postgres.c will do one shortly */
1152 * Test whether we are actively listening on the given channel name.
1154 * Note: this function is executed for every notification found in the queue.
1155 * Perhaps it is worth further optimization, eg convert the list to a sorted
1156 * array so we can binary-search it. In practice the list is likely to be
1157 * fairly short, though.
1160 IsListeningOn(const char *channel)
1164 foreach(p, listenChannels)
1166 char *lchan = (char *) lfirst(p);
1168 if (strcmp(lchan, channel) == 0)
1175 * Remove our entry from the listeners array when we are no longer listening
1176 * on any channel. NB: must not fail if we're already not listening.
1179 asyncQueueUnregister(void)
1183 Assert(listenChannels == NIL); /* else caller error */
1185 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1186 /* check if entry is valid and oldest ... */
1187 advanceTail = (MyProcPid == QUEUE_BACKEND_PID(MyBackendId)) &&
1188 QUEUE_POS_EQUAL(QUEUE_BACKEND_POS(MyBackendId), QUEUE_TAIL);
1189 /* ... then mark it invalid */
1190 QUEUE_BACKEND_PID(MyBackendId) = InvalidPid;
1191 LWLockRelease(AsyncQueueLock);
1193 /* If we were the laziest backend, try to advance the tail pointer */
1195 asyncQueueAdvanceTail();
1199 * Test whether there is room to insert more notification messages.
1201 * Caller must hold at least shared AsyncQueueLock.
1204 asyncQueueIsFull(void)
1210 * The queue is full if creating a new head page would create a page that
1211 * logically precedes the current global tail pointer, ie, the head
1212 * pointer would wrap around compared to the tail. We cannot create such
1213 * a head page for fear of confusing slru.c. For safety we round the tail
1214 * pointer back to a segment boundary (compare the truncation logic in
1215 * asyncQueueAdvanceTail).
1217 * Note that this test is *not* dependent on how much space there is on
1218 * the current head page. This is necessary because asyncQueueAddEntries
1219 * might try to create the next head page in any case.
1221 nexthead = QUEUE_POS_PAGE(QUEUE_HEAD) + 1;
1222 if (nexthead > QUEUE_MAX_PAGE)
1223 nexthead = 0; /* wrap around */
1224 boundary = QUEUE_POS_PAGE(QUEUE_TAIL);
1225 boundary -= boundary % SLRU_PAGES_PER_SEGMENT;
1226 return asyncQueuePagePrecedesLogically(nexthead, boundary);
1230 * Advance the QueuePosition to the next entry, assuming that the current
1231 * entry is of length entryLength. If we jump to a new page the function
1232 * returns true, else false.
1235 asyncQueueAdvance(QueuePosition *position, int entryLength)
1237 int pageno = QUEUE_POS_PAGE(*position);
1238 int offset = QUEUE_POS_OFFSET(*position);
1239 bool pageJump = false;
1242 * Move to the next writing position: First jump over what we have just
1245 offset += entryLength;
1246 Assert(offset <= QUEUE_PAGESIZE);
1249 * In a second step check if another entry can possibly be written to the
1250 * page. If so, stay here, we have reached the next position. If not, then
1251 * we need to move on to the next page.
1253 if (offset + QUEUEALIGN(AsyncQueueEntryEmptySize) > QUEUE_PAGESIZE)
1256 if (pageno > QUEUE_MAX_PAGE)
1257 pageno = 0; /* wrap around */
1262 SET_QUEUE_POS(*position, pageno, offset);
1267 * Fill the AsyncQueueEntry at *qe with an outbound notification message.
1270 asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe)
1272 size_t channellen = strlen(n->channel);
1273 size_t payloadlen = strlen(n->payload);
1276 Assert(channellen < NAMEDATALEN);
1277 Assert(payloadlen < NOTIFY_PAYLOAD_MAX_LENGTH);
1279 /* The terminators are already included in AsyncQueueEntryEmptySize */
1280 entryLength = AsyncQueueEntryEmptySize + payloadlen + channellen;
1281 entryLength = QUEUEALIGN(entryLength);
1282 qe->length = entryLength;
1283 qe->dboid = MyDatabaseId;
1284 qe->xid = GetCurrentTransactionId();
1285 qe->srcPid = MyProcPid;
1286 memcpy(qe->data, n->channel, channellen + 1);
1287 memcpy(qe->data + channellen + 1, n->payload, payloadlen + 1);
1291 * Add pending notifications to the queue.
1293 * We go page by page here, i.e. we stop once we have to go to a new page but
1294 * we will be called again and then fill that next page. If an entry does not
1295 * fit into the current page, we write a dummy entry with an InvalidOid as the
1296 * database OID in order to fill the page. So every page is always used up to
1297 * the last byte which simplifies reading the page later.
1299 * We are passed the list cell containing the next notification to write
1300 * and return the first still-unwritten cell back. Eventually we will return
1301 * NULL indicating all is done.
1303 * We are holding AsyncQueueLock already from the caller and grab AsyncCtlLock
1304 * locally in this function.
1307 asyncQueueAddEntries(ListCell *nextNotify)
1314 /* We hold both AsyncQueueLock and AsyncCtlLock during this operation */
1315 LWLockAcquire(AsyncCtlLock, LW_EXCLUSIVE);
1317 /* Fetch the current page */
1318 pageno = QUEUE_POS_PAGE(QUEUE_HEAD);
1319 slotno = SimpleLruReadPage(AsyncCtl, pageno, true, InvalidTransactionId);
1320 /* Note we mark the page dirty before writing in it */
1321 AsyncCtl->shared->page_dirty[slotno] = true;
1323 while (nextNotify != NULL)
1325 Notification *n = (Notification *) lfirst(nextNotify);
1327 /* Construct a valid queue entry in local variable qe */
1328 asyncQueueNotificationToEntry(n, &qe);
1330 offset = QUEUE_POS_OFFSET(QUEUE_HEAD);
1332 /* Check whether the entry really fits on the current page */
1333 if (offset + qe.length <= QUEUE_PAGESIZE)
1335 /* OK, so advance nextNotify past this item */
1336 nextNotify = lnext(nextNotify);
1341 * Write a dummy entry to fill up the page. Actually readers will
1342 * only check dboid and since it won't match any reader's database
1343 * OID, they will ignore this entry and move on.
1345 qe.length = QUEUE_PAGESIZE - offset;
1346 qe.dboid = InvalidOid;
1347 qe.data[0] = '\0'; /* empty channel */
1348 qe.data[1] = '\0'; /* empty payload */
1351 /* Now copy qe into the shared buffer page */
1352 memcpy(AsyncCtl->shared->page_buffer[slotno] + offset,
1356 /* Advance QUEUE_HEAD appropriately, and note if page is full */
1357 if (asyncQueueAdvance(&(QUEUE_HEAD), qe.length))
1360 * Page is full, so we're done here, but first fill the next page
1361 * with zeroes. The reason to do this is to ensure that slru.c's
1362 * idea of the head page is always the same as ours, which avoids
1363 * boundary problems in SimpleLruTruncate. The test in
1364 * asyncQueueIsFull() ensured that there is room to create this
1365 * page without overrunning the queue.
1367 slotno = SimpleLruZeroPage(AsyncCtl, QUEUE_POS_PAGE(QUEUE_HEAD));
1368 /* And exit the loop */
1373 LWLockRelease(AsyncCtlLock);
1379 * Check whether the queue is at least half full, and emit a warning if so.
1381 * This is unlikely given the size of the queue, but possible.
1382 * The warnings show up at most once every QUEUE_FULL_WARN_INTERVAL.
1384 * Caller must hold exclusive AsyncQueueLock.
1387 asyncQueueFillWarning(void)
1389 int headPage = QUEUE_POS_PAGE(QUEUE_HEAD);
1390 int tailPage = QUEUE_POS_PAGE(QUEUE_TAIL);
1395 occupied = headPage - tailPage;
1398 return; /* fast exit for common case */
1402 /* head has wrapped around, tail not yet */
1403 occupied += QUEUE_MAX_PAGE + 1;
1406 fillDegree = (double) occupied / (double) ((QUEUE_MAX_PAGE + 1) / 2);
1408 if (fillDegree < 0.5)
1411 t = GetCurrentTimestamp();
1413 if (TimestampDifferenceExceeds(asyncQueueControl->lastQueueFillWarn,
1414 t, QUEUE_FULL_WARN_INTERVAL))
1416 QueuePosition min = QUEUE_HEAD;
1417 int32 minPid = InvalidPid;
1420 for (i = 1; i <= MaxBackends; i++)
1422 if (QUEUE_BACKEND_PID(i) != InvalidPid)
1424 min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
1425 if (QUEUE_POS_EQUAL(min, QUEUE_BACKEND_POS(i)))
1426 minPid = QUEUE_BACKEND_PID(i);
1431 (errmsg("NOTIFY queue is %.0f%% full", fillDegree * 100),
1432 (minPid != InvalidPid ?
1433 errdetail("The server process with PID %d is among those with the oldest transactions.", minPid)
1435 (minPid != InvalidPid ?
1436 errhint("The NOTIFY queue cannot be emptied until that process ends its current transaction.")
1439 asyncQueueControl->lastQueueFillWarn = t;
1444 * Send signals to all listening backends (except our own).
1446 * Returns true if we sent at least one signal.
1448 * Since we need EXCLUSIVE lock anyway we also check the position of the other
1449 * backends and in case one is already up-to-date we don't signal it.
1450 * This can happen if concurrent notifying transactions have sent a signal and
1451 * the signaled backend has read the other notifications and ours in the same
1454 * Since we know the BackendId and the Pid the signalling is quite cheap.
1457 SignalBackends(void)
1459 bool signalled = false;
1467 * Identify all backends that are listening and not already up-to-date. We
1468 * don't want to send signals while holding the AsyncQueueLock, so we just
1469 * build a list of target PIDs.
1471 * XXX in principle these pallocs could fail, which would be bad. Maybe
1472 * preallocate the arrays? But in practice this is only run in trivial
1473 * transactions, so there should surely be space available.
1475 pids = (int32 *) palloc(MaxBackends * sizeof(int32));
1476 ids = (BackendId *) palloc(MaxBackends * sizeof(BackendId));
1479 LWLockAcquire(AsyncQueueLock, LW_EXCLUSIVE);
1480 for (i = 1; i <= MaxBackends; i++)
1482 pid = QUEUE_BACKEND_PID(i);
1483 if (pid != InvalidPid && pid != MyProcPid)
1485 QueuePosition pos = QUEUE_BACKEND_POS(i);
1487 if (!QUEUE_POS_EQUAL(pos, QUEUE_HEAD))
1495 LWLockRelease(AsyncQueueLock);
1497 /* Now send signals */
1498 for (i = 0; i < count; i++)
1503 * Note: assuming things aren't broken, a signal failure here could
1504 * only occur if the target backend exited since we released
1505 * AsyncQueueLock; which is unlikely but certainly possible. So we
1506 * just log a low-level debug message if it happens.
1508 if (SendProcSignal(pid, PROCSIG_NOTIFY_INTERRUPT, ids[i]) < 0)
1509 elog(DEBUG3, "could not signal backend with PID %d: %m", pid);
1523 * This is called at transaction abort.
1525 * Gets rid of pending actions and outbound notifies that we would have
1526 * executed if the transaction got committed.
1529 AtAbort_Notify(void)
1532 * If we LISTEN but then roll back the transaction we have set our pointer
1533 * but have not made any entry in listenChannels. In that case, remove our
1536 if (backendHasExecutedInitialListen)
1539 * Checking listenChannels should be redundant but it can't hurt doing
1540 * it for safety reasons.
1542 if (listenChannels == NIL)
1543 asyncQueueUnregister();
1545 backendHasExecutedInitialListen = false;
1549 ClearPendingActionsAndNotifies();
1553 * AtSubStart_Notify() --- Take care of subtransaction start.
1555 * Push empty state for the new subtransaction.
1558 AtSubStart_Notify(void)
1560 MemoryContext old_cxt;
1562 /* Keep the list-of-lists in TopTransactionContext for simplicity */
1563 old_cxt = MemoryContextSwitchTo(TopTransactionContext);
1565 upperPendingActions = lcons(pendingActions, upperPendingActions);
1567 Assert(list_length(upperPendingActions) ==
1568 GetCurrentTransactionNestLevel() - 1);
1570 pendingActions = NIL;
1572 upperPendingNotifies = lcons(pendingNotifies, upperPendingNotifies);
1574 Assert(list_length(upperPendingNotifies) ==
1575 GetCurrentTransactionNestLevel() - 1);
1577 pendingNotifies = NIL;
1579 MemoryContextSwitchTo(old_cxt);
1583 * AtSubCommit_Notify() --- Take care of subtransaction commit.
1585 * Reassign all items in the pending lists to the parent transaction.
1588 AtSubCommit_Notify(void)
1590 List *parentPendingActions;
1591 List *parentPendingNotifies;
1593 parentPendingActions = (List *) linitial(upperPendingActions);
1594 upperPendingActions = list_delete_first(upperPendingActions);
1596 Assert(list_length(upperPendingActions) ==
1597 GetCurrentTransactionNestLevel() - 2);
1600 * Mustn't try to eliminate duplicates here --- see queue_listen()
1602 pendingActions = list_concat(parentPendingActions, pendingActions);
1604 parentPendingNotifies = (List *) linitial(upperPendingNotifies);
1605 upperPendingNotifies = list_delete_first(upperPendingNotifies);
1607 Assert(list_length(upperPendingNotifies) ==
1608 GetCurrentTransactionNestLevel() - 2);
1611 * We could try to eliminate duplicates here, but it seems not worthwhile.
1613 pendingNotifies = list_concat(parentPendingNotifies, pendingNotifies);
1617 * AtSubAbort_Notify() --- Take care of subtransaction abort.
1620 AtSubAbort_Notify(void)
1622 int my_level = GetCurrentTransactionNestLevel();
1625 * All we have to do is pop the stack --- the actions/notifies made in
1626 * this subxact are no longer interesting, and the space will be freed
1627 * when CurTransactionContext is recycled.
1629 * This routine could be called more than once at a given nesting level if
1630 * there is trouble during subxact abort. Avoid dumping core by using
1631 * GetCurrentTransactionNestLevel as the indicator of how far we need to
1634 while (list_length(upperPendingActions) > my_level - 2)
1636 pendingActions = (List *) linitial(upperPendingActions);
1637 upperPendingActions = list_delete_first(upperPendingActions);
1640 while (list_length(upperPendingNotifies) > my_level - 2)
1642 pendingNotifies = (List *) linitial(upperPendingNotifies);
1643 upperPendingNotifies = list_delete_first(upperPendingNotifies);
1648 * HandleNotifyInterrupt
1650 * This is called when PROCSIG_NOTIFY_INTERRUPT is received.
1652 * If we are idle (notifyInterruptEnabled is set), we can safely invoke
1653 * ProcessIncomingNotify directly. Otherwise, just set a flag
1657 HandleNotifyInterrupt(void)
1660 * Note: this is called by a SIGNAL HANDLER. You must be very wary what
1661 * you do here. Some helpful soul had this routine sprinkled with
1662 * TPRINTFs, which would likely lead to corruption of stdio buffers if
1663 * they were ever turned on.
1666 /* Don't joggle the elbow of proc_exit */
1667 if (proc_exit_inprogress)
1670 if (notifyInterruptEnabled)
1672 bool save_ImmediateInterruptOK = ImmediateInterruptOK;
1675 * We may be called while ImmediateInterruptOK is true; turn it off
1676 * while messing with the NOTIFY state. (We would have to save and
1677 * restore it anyway, because PGSemaphore operations inside
1678 * ProcessIncomingNotify() might reset it.)
1680 ImmediateInterruptOK = false;
1683 * I'm not sure whether some flavors of Unix might allow another
1684 * SIGUSR1 occurrence to recursively interrupt this routine. To cope
1685 * with the possibility, we do the same sort of dance that
1686 * EnableNotifyInterrupt must do --- see that routine for comments.
1688 notifyInterruptEnabled = 0; /* disable any recursive signal */
1689 notifyInterruptOccurred = 1; /* do at least one iteration */
1692 notifyInterruptEnabled = 1;
1693 if (!notifyInterruptOccurred)
1695 notifyInterruptEnabled = 0;
1696 if (notifyInterruptOccurred)
1698 /* Here, it is finally safe to do stuff. */
1700 elog(DEBUG1, "HandleNotifyInterrupt: perform async notify");
1702 ProcessIncomingNotify();
1705 elog(DEBUG1, "HandleNotifyInterrupt: done");
1710 * Restore ImmediateInterruptOK, and check for interrupts if needed.
1712 ImmediateInterruptOK = save_ImmediateInterruptOK;
1713 if (save_ImmediateInterruptOK)
1714 CHECK_FOR_INTERRUPTS();
1719 * In this path it is NOT SAFE to do much of anything, except this:
1721 notifyInterruptOccurred = 1;
1726 * EnableNotifyInterrupt
1728 * This is called by the PostgresMain main loop just before waiting
1729 * for a frontend command. If we are truly idle (ie, *not* inside
1730 * a transaction block), then process any pending inbound notifies,
1731 * and enable the signal handler to process future notifies directly.
1733 * NOTE: the signal handler starts out disabled, and stays so until
1734 * PostgresMain calls this the first time.
1737 EnableNotifyInterrupt(void)
1739 if (IsTransactionOrTransactionBlock())
1740 return; /* not really idle */
1743 * This code is tricky because we are communicating with a signal handler
1744 * that could interrupt us at any point. If we just checked
1745 * notifyInterruptOccurred and then set notifyInterruptEnabled, we could
1746 * fail to respond promptly to a signal that happens in between those two
1747 * steps. (A very small time window, perhaps, but Murphy's Law says you
1748 * can hit it...) Instead, we first set the enable flag, then test the
1749 * occurred flag. If we see an unserviced interrupt has occurred, we
1750 * re-clear the enable flag before going off to do the service work. (That
1751 * prevents re-entrant invocation of ProcessIncomingNotify() if another
1752 * interrupt occurs.) If an interrupt comes in between the setting and
1753 * clearing of notifyInterruptEnabled, then it will have done the service
1754 * work and left notifyInterruptOccurred zero, so we have to check again
1755 * after clearing enable. The whole thing has to be in a loop in case
1756 * another interrupt occurs while we're servicing the first. Once we get
1757 * out of the loop, enable is set and we know there is no unserviced
1760 * NB: an overenthusiastic optimizing compiler could easily break this
1761 * code. Hopefully, they all understand what "volatile" means these days.
1765 notifyInterruptEnabled = 1;
1766 if (!notifyInterruptOccurred)
1768 notifyInterruptEnabled = 0;
1769 if (notifyInterruptOccurred)
1772 elog(DEBUG1, "EnableNotifyInterrupt: perform async notify");
1774 ProcessIncomingNotify();
1777 elog(DEBUG1, "EnableNotifyInterrupt: done");
1783 * DisableNotifyInterrupt
1785 * This is called by the PostgresMain main loop just after receiving
1786 * a frontend command. Signal handler execution of inbound notifies
1787 * is disabled until the next EnableNotifyInterrupt call.
1789 * The PROCSIG_CATCHUP_INTERRUPT signal handler also needs to call this,
1790 * so as to prevent conflicts if one signal interrupts the other. So we
1791 * must return the previous state of the flag.
1794 DisableNotifyInterrupt(void)
1796 bool result = (notifyInterruptEnabled != 0);
1798 notifyInterruptEnabled = 0;
1804 * Read all pending notifications from the queue, and deliver appropriate
1805 * ones to my frontend. Stop when we reach queue head or an uncommitted
1809 asyncQueueReadAllNotifications(void)
1812 QueuePosition oldpos;
1816 /* page_buffer must be adequately aligned, so use a union */
1819 char buf[QUEUE_PAGESIZE];
1820 AsyncQueueEntry align;
1823 /* Fetch current state */
1824 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1825 /* Assert checks that we have a valid state entry */
1826 Assert(MyProcPid == QUEUE_BACKEND_PID(MyBackendId));
1827 pos = oldpos = QUEUE_BACKEND_POS(MyBackendId);
1829 LWLockRelease(AsyncQueueLock);
1831 if (QUEUE_POS_EQUAL(pos, head))
1833 /* Nothing to do, we have read all notifications already. */
1838 * Note that we deliver everything that we see in the queue and that
1839 * matches our _current_ listening state.
1840 * Especially we do not take into account different commit times.
1841 * Consider the following example:
1843 * Backend 1: Backend 2:
1845 * transaction starts
1848 * transaction starts
1854 * It could happen that backend 2 sees the notification from backend 1 in
1855 * the queue. Even though the notifying transaction committed before
1856 * the listening transaction, we still deliver the notification.
1858 * The idea is that an additional notification does not do any harm, we
1859 * just need to make sure that we do not miss a notification.
1861 * It is possible that we fail while trying to send a message to our
1862 * frontend (for example, because of encoding conversion failure).
1863 * If that happens it is critical that we not try to send the same
1864 * message over and over again. Therefore, we place a PG_TRY block
1865 * here that will forcibly advance our backend position before we lose
1866 * control to an error. (We could alternatively retake AsyncQueueLock
1867 * and move the position before handling each individual message, but
1868 * that seems like too much lock traffic.)
1877 int curpage = QUEUE_POS_PAGE(pos);
1878 int curoffset = QUEUE_POS_OFFSET(pos);
1883 * We copy the data from SLRU into a local buffer, so as to avoid
1884 * holding the AsyncCtlLock while we are examining the entries and
1885 * possibly transmitting them to our frontend. Copy only the part
1886 * of the page we will actually inspect.
1888 slotno = SimpleLruReadPage_ReadOnly(AsyncCtl, curpage,
1889 InvalidTransactionId);
1890 if (curpage == QUEUE_POS_PAGE(head))
1892 /* we only want to read as far as head */
1893 copysize = QUEUE_POS_OFFSET(head) - curoffset;
1895 copysize = 0; /* just for safety */
1899 /* fetch all the rest of the page */
1900 copysize = QUEUE_PAGESIZE - curoffset;
1902 memcpy(page_buffer.buf + curoffset,
1903 AsyncCtl->shared->page_buffer[slotno] + curoffset,
1905 /* Release lock that we got from SimpleLruReadPage_ReadOnly() */
1906 LWLockRelease(AsyncCtlLock);
1909 * Process messages up to the stop position, end of page, or an
1910 * uncommitted message.
1912 * Our stop position is what we found to be the head's position
1913 * when we entered this function. It might have changed already.
1914 * But if it has, we will receive (or have already received and
1915 * queued) another signal and come here again.
1917 * We are not holding AsyncQueueLock here! The queue can only
1918 * extend beyond the head pointer (see above) and we leave our
1919 * backend's pointer where it is so nobody will truncate or
1920 * rewrite pages under us. Especially we don't want to hold a lock
1921 * while sending the notifications to the frontend.
1923 reachedStop = asyncQueueProcessPageEntries(&pos, head,
1925 } while (!reachedStop);
1929 /* Update shared state */
1930 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1931 QUEUE_BACKEND_POS(MyBackendId) = pos;
1932 advanceTail = QUEUE_POS_EQUAL(oldpos, QUEUE_TAIL);
1933 LWLockRelease(AsyncQueueLock);
1935 /* If we were the laziest backend, try to advance the tail pointer */
1937 asyncQueueAdvanceTail();
1943 /* Update shared state */
1944 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1945 QUEUE_BACKEND_POS(MyBackendId) = pos;
1946 advanceTail = QUEUE_POS_EQUAL(oldpos, QUEUE_TAIL);
1947 LWLockRelease(AsyncQueueLock);
1949 /* If we were the laziest backend, try to advance the tail pointer */
1951 asyncQueueAdvanceTail();
1955 * Fetch notifications from the shared queue, beginning at position current,
1956 * and deliver relevant ones to my frontend.
1958 * The current page must have been fetched into page_buffer from shared
1959 * memory. (We could access the page right in shared memory, but that
1960 * would imply holding the AsyncCtlLock throughout this routine.)
1962 * We stop if we reach the "stop" position, or reach a notification from an
1963 * uncommitted transaction, or reach the end of the page.
1965 * The function returns true once we have reached the stop position or an
1966 * uncommitted notification, and false if we have finished with the page.
1967 * In other words: once it returns true there is no need to look further.
1968 * The QueuePosition *current is advanced past all processed messages.
1971 asyncQueueProcessPageEntries(QueuePosition *current,
1975 bool reachedStop = false;
1976 bool reachedEndOfPage;
1977 AsyncQueueEntry *qe;
1981 QueuePosition thisentry = *current;
1983 if (QUEUE_POS_EQUAL(thisentry, stop))
1986 qe = (AsyncQueueEntry *) (page_buffer + QUEUE_POS_OFFSET(thisentry));
1989 * Advance *current over this message, possibly to the next page. As
1990 * noted in the comments for asyncQueueReadAllNotifications, we must
1991 * do this before possibly failing while processing the message.
1993 reachedEndOfPage = asyncQueueAdvance(current, qe->length);
1995 /* Ignore messages destined for other databases */
1996 if (qe->dboid == MyDatabaseId)
1998 if (TransactionIdDidCommit(qe->xid))
2000 /* qe->data is the null-terminated channel name */
2001 char *channel = qe->data;
2003 if (IsListeningOn(channel))
2005 /* payload follows channel name */
2006 char *payload = qe->data + strlen(channel) + 1;
2008 NotifyMyFrontEnd(channel, payload, qe->srcPid);
2011 else if (TransactionIdDidAbort(qe->xid))
2014 * If the source transaction aborted, we just ignore its
2021 * The transaction has neither committed nor aborted so far,
2022 * so we can't process its message yet. Break out of the
2023 * loop, but first back up *current so we will reprocess the
2024 * message next time. (Note: it is unlikely but not
2025 * impossible for TransactionIdDidCommit to fail, so we can't
2026 * really avoid this advance-then-back-up behavior when
2027 * dealing with an uncommitted message.)
2029 *current = thisentry;
2035 /* Loop back if we're not at end of page */
2036 } while (!reachedEndOfPage);
2038 if (QUEUE_POS_EQUAL(*current, stop))
2045 * Advance the shared queue tail variable to the minimum of all the
2046 * per-backend tail pointers. Truncate pg_notify space if possible.
2049 asyncQueueAdvanceTail(void)
2057 LWLockAcquire(AsyncQueueLock, LW_EXCLUSIVE);
2059 for (i = 1; i <= MaxBackends; i++)
2061 if (QUEUE_BACKEND_PID(i) != InvalidPid)
2062 min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
2064 oldtailpage = QUEUE_POS_PAGE(QUEUE_TAIL);
2066 LWLockRelease(AsyncQueueLock);
2069 * We can truncate something if the global tail advanced across an SLRU
2072 * XXX it might be better to truncate only once every several segments, to
2073 * reduce the number of directory scans.
2075 newtailpage = QUEUE_POS_PAGE(min);
2076 boundary = newtailpage - (newtailpage % SLRU_PAGES_PER_SEGMENT);
2077 if (asyncQueuePagePrecedesLogically(oldtailpage, boundary))
2080 * SimpleLruTruncate() will ask for AsyncCtlLock but will also release
2083 SimpleLruTruncate(AsyncCtl, newtailpage);
2088 * ProcessIncomingNotify
2090 * Deal with arriving NOTIFYs from other backends.
2091 * This is called either directly from the PROCSIG_NOTIFY_INTERRUPT
2092 * signal handler, or the next time control reaches the outer idle loop.
2093 * Scan the queue for arriving notifications and report them to my front
2096 * NOTE: since we are outside any transaction, we must create our own.
2099 ProcessIncomingNotify(void)
2101 bool catchup_enabled;
2103 /* We *must* reset the flag */
2104 notifyInterruptOccurred = 0;
2106 /* Do nothing else if we aren't actively listening */
2107 if (listenChannels == NIL)
2110 /* Must prevent catchup interrupt while I am running */
2111 catchup_enabled = DisableCatchupInterrupt();
2114 elog(DEBUG1, "ProcessIncomingNotify");
2116 set_ps_display("notify interrupt", false);
2119 * We must run asyncQueueReadAllNotifications inside a transaction, else
2120 * bad things happen if it gets an error.
2122 StartTransactionCommand();
2124 asyncQueueReadAllNotifications();
2126 CommitTransactionCommand();
2129 * Must flush the notify messages to ensure frontend gets them promptly.
2133 set_ps_display("idle", false);
2136 elog(DEBUG1, "ProcessIncomingNotify: done");
2138 if (catchup_enabled)
2139 EnableCatchupInterrupt();
2143 * Send NOTIFY message to my front end.
2146 NotifyMyFrontEnd(const char *channel, const char *payload, int32 srcPid)
2148 if (whereToSendOutput == DestRemote)
2152 pq_beginmessage(&buf, 'A');
2153 pq_sendint(&buf, srcPid, sizeof(int32));
2154 pq_sendstring(&buf, channel);
2155 if (PG_PROTOCOL_MAJOR(FrontendProtocol) >= 3)
2156 pq_sendstring(&buf, payload);
2157 pq_endmessage(&buf);
2160 * NOTE: we do not do pq_flush() here. For a self-notify, it will
2161 * happen at the end of the transaction, and for incoming notifies
2162 * ProcessIncomingNotify will do it after finding all the notifies.
2166 elog(INFO, "NOTIFY for \"%s\" payload \"%s\"", channel, payload);
2169 /* Does pendingNotifies include the given channel/payload? */
2171 AsyncExistsPendingNotify(const char *channel, const char *payload)
2176 if (pendingNotifies == NIL)
2179 if (payload == NULL)
2183 * We need to append new elements to the end of the list in order to keep
2184 * the order. However, on the other hand we'd like to check the list
2185 * backwards in order to make duplicate-elimination a tad faster when the
2186 * same condition is signaled many times in a row. So as a compromise we
2187 * check the tail element first which we can access directly. If this
2188 * doesn't match, we check the whole list.
2190 * As we are not checking our parents' lists, we can still get duplicates
2191 * in combination with subtransactions, like in:
2200 n = (Notification *) llast(pendingNotifies);
2201 if (strcmp(n->channel, channel) == 0 &&
2202 strcmp(n->payload, payload) == 0)
2205 foreach(p, pendingNotifies)
2207 n = (Notification *) lfirst(p);
2209 if (strcmp(n->channel, channel) == 0 &&
2210 strcmp(n->payload, payload) == 0)
2217 /* Clear the pendingActions and pendingNotifies lists. */
2219 ClearPendingActionsAndNotifies(void)
2222 * We used to have to explicitly deallocate the list members and nodes,
2223 * because they were malloc'd. Now, since we know they are palloc'd in
2224 * CurTransactionContext, we need not do that --- they'll go away
2225 * automatically at transaction exit. We need only reset the list head
2228 pendingActions = NIL;
2229 pendingNotifies = NIL;