1 /*-------------------------------------------------------------------------
4 * Asynchronous notification: NOTIFY, LISTEN, UNLISTEN
6 * Portions Copyright (c) 1996-2015, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
10 * src/backend/commands/async.c
12 *-------------------------------------------------------------------------
15 /*-------------------------------------------------------------------------
16 * Async Notification Model as of 9.0:
18 * 1. Multiple backends on same machine. Multiple backends listening on
19 * several channels. (Channels are also called "conditions" in other
22 * 2. There is one central queue in disk-based storage (directory pg_notify/),
23 * with actively-used pages mapped into shared memory by the slru.c module.
24 * All notification messages are placed in the queue and later read out
25 * by listening backends.
27 * There is no central knowledge of which backend listens on which channel;
28 * every backend has its own list of interesting channels.
30 * Although there is only one queue, notifications are treated as being
31 * database-local; this is done by including the sender's database OID
32 * in each notification message. Listening backends ignore messages
33 * that don't match their database OID. This is important because it
34 * ensures senders and receivers have the same database encoding and won't
35 * misinterpret non-ASCII text in the channel name or payload string.
37 * Since notifications are not expected to survive database crashes,
38 * we can simply clean out the pg_notify data at any reboot, and there
39 * is no need for WAL support or fsync'ing.
41 * 3. Every backend that is listening on at least one channel registers by
42 * entering its PID into the array in AsyncQueueControl. It then scans all
43 * incoming notifications in the central queue and first compares the
44 * database OID of the notification with its own database OID and then
45 * compares the notified channel with the list of channels that it listens
46 * to. In case there is a match it delivers the notification event to its
47 * frontend. Non-matching events are simply skipped.
49 * 4. The NOTIFY statement (routine Async_Notify) stores the notification in
50 * a backend-local list which will not be processed until transaction end.
52 * Duplicate notifications from the same transaction are sent out as one
53 * notification only. This is done to save work when for example a trigger
54 * on a 2 million row table fires a notification for each row that has been
55 * changed. If the application needs to receive every single notification
56 * that has been sent, it can easily add some unique string into the extra
59 * When the transaction is ready to commit, PreCommit_Notify() adds the
60 * pending notifications to the head of the queue. The head pointer of the
61 * queue always points to the next free position and a position is just a
62 * page number and the offset in that page. This is done before marking the
63 * transaction as committed in clog. If we run into problems writing the
64 * notifications, we can still call elog(ERROR, ...) and the transaction
67 * Once we have put all of the notifications into the queue, we return to
68 * CommitTransaction() which will then do the actual transaction commit.
70 * After commit we are called another time (AtCommit_Notify()). Here we
71 * make the actual updates to the effective listen state (listenChannels).
73 * Finally, after we are out of the transaction altogether, we check if
74 * we need to signal listening backends. In SignalBackends() we scan the
75 * list of listening backends and send a PROCSIG_NOTIFY_INTERRUPT signal
76 * to every listening backend (we don't know which backend is listening on
77 * which channel so we must signal them all). We can exclude backends that
78 * are already up to date, though. We don't bother with a self-signal
79 * either, but just process the queue directly.
81 * 5. Upon receipt of a PROCSIG_NOTIFY_INTERRUPT signal, the signal handler
82 * sets the process's latch, which triggers the event to be processed
83 * immediately if this backend is idle (i.e., it is waiting for a frontend
84 * command and is not within a transaction block. C.f.
85 * ProcessClientReadInterrupt()). Otherwise the handler may only set a
86 * flag, which will cause the processing to occur just before we next go
89 * Inbound-notify processing consists of reading all of the notifications
90 * that have arrived since scanning last time. We read every notification
91 * until we reach either a notification from an uncommitted transaction or
92 * the head pointer's position. Then we check if we were the laziest
93 * backend: if our pointer is set to the same position as the global tail
94 * pointer is set, then we move the global tail pointer ahead to where the
95 * second-laziest backend is (in general, we take the MIN of the current
96 * head position and all active backends' new tail pointers). Whenever we
97 * move the global tail pointer we also truncate now-unused pages (i.e.,
98 * delete files in pg_notify/ that are no longer used).
100 * An application that listens on the same channel it notifies will get
101 * NOTIFY messages for its own NOTIFYs. These can be ignored, if not useful,
102 * by comparing be_pid in the NOTIFY message to the application's own backend's
103 * PID. (As of FE/BE protocol 2.0, the backend's PID is provided to the
104 * frontend during startup.) The above design guarantees that notifies from
105 * other backends will never be missed by ignoring self-notifies.
107 * The amount of shared memory used for notify management (NUM_ASYNC_BUFFERS)
108 * can be varied without affecting anything but performance. The maximum
109 * amount of notification data that can be queued at one time is determined
110 * by slru.c's wraparound limit; see QUEUE_MAX_PAGE below.
111 *-------------------------------------------------------------------------
114 #include "postgres.h"
120 #include "access/slru.h"
121 #include "access/transam.h"
122 #include "access/xact.h"
123 #include "catalog/pg_database.h"
124 #include "commands/async.h"
126 #include "libpq/libpq.h"
127 #include "libpq/pqformat.h"
128 #include "miscadmin.h"
129 #include "storage/ipc.h"
130 #include "storage/lmgr.h"
131 #include "storage/proc.h"
132 #include "storage/procarray.h"
133 #include "storage/procsignal.h"
134 #include "storage/sinval.h"
135 #include "tcop/tcopprot.h"
136 #include "utils/builtins.h"
137 #include "utils/memutils.h"
138 #include "utils/ps_status.h"
139 #include "utils/timestamp.h"
143 * Maximum size of a NOTIFY payload, including terminating NULL. This
144 * must be kept small enough so that a notification message fits on one
145 * SLRU page. The magic fudge factor here is noncritical as long as it's
146 * more than AsyncQueueEntryEmptySize --- we make it significantly bigger
147 * than that, so changes in that data structure won't affect user-visible
150 #define NOTIFY_PAYLOAD_MAX_LENGTH (BLCKSZ - NAMEDATALEN - 128)
153 * Struct representing an entry in the global notify queue
155 * This struct declaration has the maximal length, but in a real queue entry
156 * the data area is only big enough for the actual channel and payload strings
157 * (each null-terminated). AsyncQueueEntryEmptySize is the minimum possible
158 * entry size, if both channel and payload strings are empty (but note it
159 * doesn't include alignment padding).
161 * The "length" field should always be rounded up to the next QUEUEALIGN
162 * multiple so that all fields are properly aligned.
164 typedef struct AsyncQueueEntry
166 int length; /* total allocated length of entry */
167 Oid dboid; /* sender's database OID */
168 TransactionId xid; /* sender's XID */
169 int32 srcPid; /* sender's PID */
170 char data[NAMEDATALEN + NOTIFY_PAYLOAD_MAX_LENGTH];
173 /* Currently, no field of AsyncQueueEntry requires more than int alignment */
174 #define QUEUEALIGN(len) INTALIGN(len)
176 #define AsyncQueueEntryEmptySize (offsetof(AsyncQueueEntry, data) + 2)
179 * Struct describing a queue position, and assorted macros for working with it
181 typedef struct QueuePosition
183 int page; /* SLRU page number */
184 int offset; /* byte offset within page */
187 #define QUEUE_POS_PAGE(x) ((x).page)
188 #define QUEUE_POS_OFFSET(x) ((x).offset)
190 #define SET_QUEUE_POS(x,y,z) \
196 #define QUEUE_POS_EQUAL(x,y) \
197 ((x).page == (y).page && (x).offset == (y).offset)
199 /* choose logically smaller QueuePosition */
200 #define QUEUE_POS_MIN(x,y) \
201 (asyncQueuePagePrecedes((x).page, (y).page) ? (x) : \
202 (x).page != (y).page ? (y) : \
203 (x).offset < (y).offset ? (x) : (y))
205 /* choose logically larger QueuePosition */
206 #define QUEUE_POS_MAX(x,y) \
207 (asyncQueuePagePrecedes((x).page, (y).page) ? (y) : \
208 (x).page != (y).page ? (x) : \
209 (x).offset > (y).offset ? (x) : (y))
212 * Struct describing a listening backend's status
214 typedef struct QueueBackendStatus
216 int32 pid; /* either a PID or InvalidPid */
217 Oid dboid; /* backend's database OID, or InvalidOid */
218 QueuePosition pos; /* backend has read queue up to here */
219 } QueueBackendStatus;
222 * Shared memory state for LISTEN/NOTIFY (excluding its SLRU stuff)
224 * The AsyncQueueControl structure is protected by the AsyncQueueLock.
226 * When holding the lock in SHARED mode, backends may only inspect their own
227 * entries as well as the head and tail pointers. Consequently we can allow a
228 * backend to update its own record while holding only SHARED lock (since no
229 * other backend will inspect it).
231 * When holding the lock in EXCLUSIVE mode, backends can inspect the entries
232 * of other backends and also change the head and tail pointers.
234 * AsyncCtlLock is used as the control lock for the pg_notify SLRU buffers.
235 * In order to avoid deadlocks, whenever we need both locks, we always first
236 * get AsyncQueueLock and then AsyncCtlLock.
238 * Each backend uses the backend[] array entry with index equal to its
239 * BackendId (which can range from 1 to MaxBackends). We rely on this to make
240 * SendProcSignal fast.
242 typedef struct AsyncQueueControl
244 QueuePosition head; /* head points to the next free location */
245 QueuePosition tail; /* the global tail is equivalent to the pos of
246 * the "slowest" backend */
247 TimestampTz lastQueueFillWarn; /* time of last queue-full msg */
248 QueueBackendStatus backend[FLEXIBLE_ARRAY_MEMBER];
249 /* backend[0] is not used; used entries are from [1] to [MaxBackends] */
252 static AsyncQueueControl *asyncQueueControl;
254 #define QUEUE_HEAD (asyncQueueControl->head)
255 #define QUEUE_TAIL (asyncQueueControl->tail)
256 #define QUEUE_BACKEND_PID(i) (asyncQueueControl->backend[i].pid)
257 #define QUEUE_BACKEND_DBOID(i) (asyncQueueControl->backend[i].dboid)
258 #define QUEUE_BACKEND_POS(i) (asyncQueueControl->backend[i].pos)
261 * The SLRU buffer area through which we access the notification queue
263 static SlruCtlData AsyncCtlData;
265 #define AsyncCtl (&AsyncCtlData)
266 #define QUEUE_PAGESIZE BLCKSZ
267 #define QUEUE_FULL_WARN_INTERVAL 5000 /* warn at most once every 5s */
270 * slru.c currently assumes that all filenames are four characters of hex
271 * digits. That means that we can use segments 0000 through FFFF.
272 * Each segment contains SLRU_PAGES_PER_SEGMENT pages which gives us
273 * the pages from 0 to SLRU_PAGES_PER_SEGMENT * 0x10000 - 1.
275 * It's of course possible to enhance slru.c, but this gives us so much
276 * space already that it doesn't seem worth the trouble.
278 * The most data we can have in the queue at a time is QUEUE_MAX_PAGE/2
279 * pages, because more than that would confuse slru.c into thinking there
280 * was a wraparound condition. With the default BLCKSZ this means there
281 * can be up to 8GB of queued-and-not-read data.
283 * Note: it's possible to redefine QUEUE_MAX_PAGE with a smaller multiple of
284 * SLRU_PAGES_PER_SEGMENT, for easier testing of queue-full behaviour.
286 #define QUEUE_MAX_PAGE (SLRU_PAGES_PER_SEGMENT * 0x10000 - 1)
289 * listenChannels identifies the channels we are actually listening to
290 * (ie, have committed a LISTEN on). It is a simple list of channel names,
291 * allocated in TopMemoryContext.
293 static List *listenChannels = NIL; /* list of C strings */
296 * State for pending LISTEN/UNLISTEN actions consists of an ordered list of
297 * all actions requested in the current transaction. As explained above,
298 * we don't actually change listenChannels until we reach transaction commit.
300 * The list is kept in CurTransactionContext. In subtransactions, each
301 * subtransaction has its own list in its own CurTransactionContext, but
302 * successful subtransactions attach their lists to their parent's list.
303 * Failed subtransactions simply discard their lists.
314 ListenActionKind action;
315 char channel[FLEXIBLE_ARRAY_MEMBER]; /* nul-terminated string */
318 static List *pendingActions = NIL; /* list of ListenAction */
320 static List *upperPendingActions = NIL; /* list of upper-xact lists */
323 * State for outbound notifies consists of a list of all channels+payloads
324 * NOTIFYed in the current transaction. We do not actually perform a NOTIFY
325 * until and unless the transaction commits. pendingNotifies is NIL if no
326 * NOTIFYs have been done in the current transaction.
328 * The list is kept in CurTransactionContext. In subtransactions, each
329 * subtransaction has its own list in its own CurTransactionContext, but
330 * successful subtransactions attach their lists to their parent's list.
331 * Failed subtransactions simply discard their lists.
333 * Note: the action and notify lists do not interact within a transaction.
334 * In particular, if a transaction does NOTIFY and then LISTEN on the same
335 * condition name, it will get a self-notify at commit. This is a bit odd
336 * but is consistent with our historical behavior.
338 typedef struct Notification
340 char *channel; /* channel name */
341 char *payload; /* payload string (can be empty) */
344 static List *pendingNotifies = NIL; /* list of Notifications */
346 static List *upperPendingNotifies = NIL; /* list of upper-xact lists */
349 * Inbound notifications are initially processed by HandleNotifyInterrupt(),
350 * called from inside a signal handler. That just sets the
351 * notifyInterruptPending flag and sets the process
352 * latch. ProcessNotifyInterrupt() will then be called whenever it's safe to
353 * actually deal with the interrupt.
355 volatile sig_atomic_t notifyInterruptPending = false;
357 /* True if we've registered an on_shmem_exit cleanup */
358 static bool unlistenExitRegistered = false;
360 /* True if we're currently registered as a listener in asyncQueueControl */
361 static bool amRegisteredListener = false;
363 /* has this backend sent notifications in the current transaction? */
364 static bool backendHasSentNotifications = false;
367 bool Trace_notify = false;
369 /* local function prototypes */
370 static bool asyncQueuePagePrecedes(int p, int q);
371 static void queue_listen(ListenActionKind action, const char *channel);
372 static void Async_UnlistenOnExit(int code, Datum arg);
373 static void Exec_ListenPreCommit(void);
374 static void Exec_ListenCommit(const char *channel);
375 static void Exec_UnlistenCommit(const char *channel);
376 static void Exec_UnlistenAllCommit(void);
377 static bool IsListeningOn(const char *channel);
378 static void asyncQueueUnregister(void);
379 static bool asyncQueueIsFull(void);
380 static bool asyncQueueAdvance(volatile QueuePosition *position, int entryLength);
381 static void asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe);
382 static ListCell *asyncQueueAddEntries(ListCell *nextNotify);
383 static double asyncQueueUsage(void);
384 static void asyncQueueFillWarning(void);
385 static bool SignalBackends(void);
386 static void asyncQueueReadAllNotifications(void);
387 static bool asyncQueueProcessPageEntries(volatile QueuePosition *current,
390 static void asyncQueueAdvanceTail(void);
391 static void ProcessIncomingNotify(void);
392 static void NotifyMyFrontEnd(const char *channel,
395 static bool AsyncExistsPendingNotify(const char *channel, const char *payload);
396 static void ClearPendingActionsAndNotifies(void);
399 * We will work on the page range of 0..QUEUE_MAX_PAGE.
402 asyncQueuePagePrecedes(int p, int q)
407 * We have to compare modulo (QUEUE_MAX_PAGE+1)/2. Both inputs should be
408 * in the range 0..QUEUE_MAX_PAGE.
410 Assert(p >= 0 && p <= QUEUE_MAX_PAGE);
411 Assert(q >= 0 && q <= QUEUE_MAX_PAGE);
414 if (diff >= ((QUEUE_MAX_PAGE + 1) / 2))
415 diff -= QUEUE_MAX_PAGE + 1;
416 else if (diff < -((QUEUE_MAX_PAGE + 1) / 2))
417 diff += QUEUE_MAX_PAGE + 1;
422 * Report space needed for our shared memory area
429 /* This had better match AsyncShmemInit */
430 size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
431 size = add_size(size, offsetof(AsyncQueueControl, backend));
433 size = add_size(size, SimpleLruShmemSize(NUM_ASYNC_BUFFERS, 0));
439 * Initialize our shared memory area
449 * Create or attach to the AsyncQueueControl structure.
451 * The used entries in the backend[] array run from 1 to MaxBackends; the
452 * zero'th entry is unused but must be allocated.
454 size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
455 size = add_size(size, offsetof(AsyncQueueControl, backend));
457 asyncQueueControl = (AsyncQueueControl *)
458 ShmemInitStruct("Async Queue Control", size, &found);
462 /* First time through, so initialize it */
465 SET_QUEUE_POS(QUEUE_HEAD, 0, 0);
466 SET_QUEUE_POS(QUEUE_TAIL, 0, 0);
467 asyncQueueControl->lastQueueFillWarn = 0;
468 /* zero'th entry won't be used, but let's initialize it anyway */
469 for (i = 0; i <= MaxBackends; i++)
471 QUEUE_BACKEND_PID(i) = InvalidPid;
472 QUEUE_BACKEND_DBOID(i) = InvalidOid;
473 SET_QUEUE_POS(QUEUE_BACKEND_POS(i), 0, 0);
478 * Set up SLRU management of the pg_notify data.
480 AsyncCtl->PagePrecedes = asyncQueuePagePrecedes;
481 SimpleLruInit(AsyncCtl, "Async Ctl", NUM_ASYNC_BUFFERS, 0,
482 AsyncCtlLock, "pg_notify");
483 /* Override default assumption that writes should be fsync'd */
484 AsyncCtl->do_fsync = false;
489 * During start or reboot, clean out the pg_notify directory.
491 (void) SlruScanDirectory(AsyncCtl, SlruScanDirCbDeleteAll, NULL);
493 /* Now initialize page zero to empty */
494 LWLockAcquire(AsyncCtlLock, LW_EXCLUSIVE);
495 slotno = SimpleLruZeroPage(AsyncCtl, QUEUE_POS_PAGE(QUEUE_HEAD));
496 /* This write is just to verify that pg_notify/ is writable */
497 SimpleLruWritePage(AsyncCtl, slotno);
498 LWLockRelease(AsyncCtlLock);
505 * SQL function to send a notification event
508 pg_notify(PG_FUNCTION_ARGS)
516 channel = text_to_cstring(PG_GETARG_TEXT_PP(0));
521 payload = text_to_cstring(PG_GETARG_TEXT_PP(1));
523 /* For NOTIFY as a statement, this is checked in ProcessUtility */
524 PreventCommandDuringRecovery("NOTIFY");
526 Async_Notify(channel, payload);
535 * This is executed by the SQL notify command.
537 * Adds the message to the list of pending notifies.
538 * Actual notification happens during transaction commit.
539 * ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
542 Async_Notify(const char *channel, const char *payload)
545 MemoryContext oldcontext;
548 elog(DEBUG1, "Async_Notify(%s)", channel);
550 /* a channel name must be specified */
551 if (!channel || !strlen(channel))
553 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
554 errmsg("channel name cannot be empty")));
556 if (strlen(channel) >= NAMEDATALEN)
558 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
559 errmsg("channel name too long")));
563 if (strlen(payload) >= NOTIFY_PAYLOAD_MAX_LENGTH)
565 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
566 errmsg("payload string too long")));
569 /* no point in making duplicate entries in the list ... */
570 if (AsyncExistsPendingNotify(channel, payload))
574 * The notification list needs to live until end of transaction, so store
575 * it in the transaction context.
577 oldcontext = MemoryContextSwitchTo(CurTransactionContext);
579 n = (Notification *) palloc(sizeof(Notification));
580 n->channel = pstrdup(channel);
582 n->payload = pstrdup(payload);
587 * We want to preserve the order so we need to append every notification.
588 * See comments at AsyncExistsPendingNotify().
590 pendingNotifies = lappend(pendingNotifies, n);
592 MemoryContextSwitchTo(oldcontext);
597 * Common code for listen, unlisten, unlisten all commands.
599 * Adds the request to the list of pending actions.
600 * Actual update of the listenChannels list happens during transaction
604 queue_listen(ListenActionKind action, const char *channel)
606 MemoryContext oldcontext;
607 ListenAction *actrec;
610 * Unlike Async_Notify, we don't try to collapse out duplicates. It would
611 * be too complicated to ensure we get the right interactions of
612 * conflicting LISTEN/UNLISTEN/UNLISTEN_ALL, and it's unlikely that there
613 * would be any performance benefit anyway in sane applications.
615 oldcontext = MemoryContextSwitchTo(CurTransactionContext);
617 /* space for terminating null is included in sizeof(ListenAction) */
618 actrec = (ListenAction *) palloc(offsetof(ListenAction, channel) +
619 strlen(channel) + 1);
620 actrec->action = action;
621 strcpy(actrec->channel, channel);
623 pendingActions = lappend(pendingActions, actrec);
625 MemoryContextSwitchTo(oldcontext);
631 * This is executed by the SQL listen command.
634 Async_Listen(const char *channel)
637 elog(DEBUG1, "Async_Listen(%s,%d)", channel, MyProcPid);
639 queue_listen(LISTEN_LISTEN, channel);
645 * This is executed by the SQL unlisten command.
648 Async_Unlisten(const char *channel)
651 elog(DEBUG1, "Async_Unlisten(%s,%d)", channel, MyProcPid);
653 /* If we couldn't possibly be listening, no need to queue anything */
654 if (pendingActions == NIL && !unlistenExitRegistered)
657 queue_listen(LISTEN_UNLISTEN, channel);
663 * This is invoked by UNLISTEN * command, and also at backend exit.
666 Async_UnlistenAll(void)
669 elog(DEBUG1, "Async_UnlistenAll(%d)", MyProcPid);
671 /* If we couldn't possibly be listening, no need to queue anything */
672 if (pendingActions == NIL && !unlistenExitRegistered)
675 queue_listen(LISTEN_UNLISTEN_ALL, "");
679 * SQL function: return a set of the channel names this backend is actively
682 * Note: this coding relies on the fact that the listenChannels list cannot
683 * change within a transaction.
686 pg_listening_channels(PG_FUNCTION_ARGS)
688 FuncCallContext *funcctx;
691 /* stuff done only on the first call of the function */
692 if (SRF_IS_FIRSTCALL())
694 MemoryContext oldcontext;
696 /* create a function context for cross-call persistence */
697 funcctx = SRF_FIRSTCALL_INIT();
699 /* switch to memory context appropriate for multiple function calls */
700 oldcontext = MemoryContextSwitchTo(funcctx->multi_call_memory_ctx);
702 /* allocate memory for user context */
703 lcp = (ListCell **) palloc(sizeof(ListCell *));
704 *lcp = list_head(listenChannels);
705 funcctx->user_fctx = (void *) lcp;
707 MemoryContextSwitchTo(oldcontext);
710 /* stuff done on every call of the function */
711 funcctx = SRF_PERCALL_SETUP();
712 lcp = (ListCell **) funcctx->user_fctx;
716 char *channel = (char *) lfirst(*lcp);
719 SRF_RETURN_NEXT(funcctx, CStringGetTextDatum(channel));
722 SRF_RETURN_DONE(funcctx);
726 * Async_UnlistenOnExit
728 * This is executed at backend exit if we have done any LISTENs in this
729 * backend. It might not be necessary anymore, if the user UNLISTENed
730 * everything, but we don't try to detect that case.
733 Async_UnlistenOnExit(int code, Datum arg)
735 Exec_UnlistenAllCommit();
736 asyncQueueUnregister();
742 * This is called at the prepare phase of a two-phase
743 * transaction. Save the state for possible commit later.
746 AtPrepare_Notify(void)
748 /* It's not allowed to have any pending LISTEN/UNLISTEN/NOTIFY actions */
749 if (pendingActions || pendingNotifies)
751 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
752 errmsg("cannot PREPARE a transaction that has executed LISTEN, UNLISTEN, or NOTIFY")));
758 * This is called at transaction commit, before actually committing to
761 * If there are pending LISTEN actions, make sure we are listed in the
762 * shared-memory listener array. This must happen before commit to
763 * ensure we don't miss any notifies from transactions that commit
766 * If there are outbound notify requests in the pendingNotifies list,
767 * add them to the global queue. We do that before commit so that
768 * we can still throw error if we run out of queue space.
771 PreCommit_Notify(void)
775 if (pendingActions == NIL && pendingNotifies == NIL)
776 return; /* no relevant statements in this xact */
779 elog(DEBUG1, "PreCommit_Notify");
781 /* Preflight for any pending listen/unlisten actions */
782 foreach(p, pendingActions)
784 ListenAction *actrec = (ListenAction *) lfirst(p);
786 switch (actrec->action)
789 Exec_ListenPreCommit();
791 case LISTEN_UNLISTEN:
792 /* there is no Exec_UnlistenPreCommit() */
794 case LISTEN_UNLISTEN_ALL:
795 /* there is no Exec_UnlistenAllPreCommit() */
800 /* Queue any pending notifies */
803 ListCell *nextNotify;
806 * Make sure that we have an XID assigned to the current transaction.
807 * GetCurrentTransactionId is cheap if we already have an XID, but not
808 * so cheap if we don't, and we'd prefer not to do that work while
809 * holding AsyncQueueLock.
811 (void) GetCurrentTransactionId();
814 * Serialize writers by acquiring a special lock that we hold till
815 * after commit. This ensures that queue entries appear in commit
816 * order, and in particular that there are never uncommitted queue
817 * entries ahead of committed ones, so an uncommitted transaction
818 * can't block delivery of deliverable notifications.
820 * We use a heavyweight lock so that it'll automatically be released
821 * after either commit or abort. This also allows deadlocks to be
822 * detected, though really a deadlock shouldn't be possible here.
824 * The lock is on "database 0", which is pretty ugly but it doesn't
825 * seem worth inventing a special locktag category just for this.
826 * (Historical note: before PG 9.0, a similar lock on "database 0" was
827 * used by the flatfiles mechanism.)
829 LockSharedObject(DatabaseRelationId, InvalidOid, 0,
830 AccessExclusiveLock);
832 /* Now push the notifications into the queue */
833 backendHasSentNotifications = true;
835 nextNotify = list_head(pendingNotifies);
836 while (nextNotify != NULL)
839 * Add the pending notifications to the queue. We acquire and
840 * release AsyncQueueLock once per page, which might be overkill
841 * but it does allow readers to get in while we're doing this.
843 * A full queue is very uncommon and should really not happen,
844 * given that we have so much space available in the SLRU pages.
845 * Nevertheless we need to deal with this possibility. Note that
846 * when we get here we are in the process of committing our
847 * transaction, but we have not yet committed to clog, so at this
848 * point in time we can still roll the transaction back.
850 LWLockAcquire(AsyncQueueLock, LW_EXCLUSIVE);
851 asyncQueueFillWarning();
852 if (asyncQueueIsFull())
854 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
855 errmsg("too many notifications in the NOTIFY queue")));
856 nextNotify = asyncQueueAddEntries(nextNotify);
857 LWLockRelease(AsyncQueueLock);
865 * This is called at transaction commit, after committing to clog.
867 * Update listenChannels and clear transaction-local state.
870 AtCommit_Notify(void)
875 * Allow transactions that have not executed LISTEN/UNLISTEN/NOTIFY to
876 * return as soon as possible
878 if (!pendingActions && !pendingNotifies)
882 elog(DEBUG1, "AtCommit_Notify");
884 /* Perform any pending listen/unlisten actions */
885 foreach(p, pendingActions)
887 ListenAction *actrec = (ListenAction *) lfirst(p);
889 switch (actrec->action)
892 Exec_ListenCommit(actrec->channel);
894 case LISTEN_UNLISTEN:
895 Exec_UnlistenCommit(actrec->channel);
897 case LISTEN_UNLISTEN_ALL:
898 Exec_UnlistenAllCommit();
903 /* If no longer listening to anything, get out of listener array */
904 if (amRegisteredListener && listenChannels == NIL)
905 asyncQueueUnregister();
908 ClearPendingActionsAndNotifies();
912 * Exec_ListenPreCommit --- subroutine for PreCommit_Notify
914 * This function must make sure we are ready to catch any incoming messages.
917 Exec_ListenPreCommit(void)
924 * Nothing to do if we are already listening to something, nor if we
925 * already ran this routine in this transaction.
927 if (amRegisteredListener)
931 elog(DEBUG1, "Exec_ListenPreCommit(%d)", MyProcPid);
934 * Before registering, make sure we will unlisten before dying. (Note:
935 * this action does not get undone if we abort later.)
937 if (!unlistenExitRegistered)
939 before_shmem_exit(Async_UnlistenOnExit, 0);
940 unlistenExitRegistered = true;
944 * This is our first LISTEN, so establish our pointer.
946 * We set our pointer to the global tail pointer and then move it forward
947 * over already-committed notifications. This ensures we cannot miss any
948 * not-yet-committed notifications. We might get a few more but that
951 * In some scenarios there might be a lot of committed notifications that
952 * have not yet been pruned away (because some backend is being lazy about
953 * reading them). To reduce our startup time, we can look at other
954 * backends and adopt the maximum "pos" pointer of any backend that's in
955 * our database; any notifications it's already advanced over are surely
956 * committed and need not be re-examined by us. (We must consider only
957 * backends connected to our DB, because others will not have bothered to
958 * check committed-ness of notifications in our DB.) But we only bother
959 * with that if there's more than a page worth of notifications
960 * outstanding, otherwise scanning all the other backends isn't worth it.
962 * We need exclusive lock here so we can look at other backends' entries.
964 LWLockAcquire(AsyncQueueLock, LW_EXCLUSIVE);
967 if (QUEUE_POS_PAGE(max) != QUEUE_POS_PAGE(head))
969 for (i = 1; i <= MaxBackends; i++)
971 if (QUEUE_BACKEND_DBOID(i) == MyDatabaseId)
972 max = QUEUE_POS_MAX(max, QUEUE_BACKEND_POS(i));
975 QUEUE_BACKEND_POS(MyBackendId) = max;
976 QUEUE_BACKEND_PID(MyBackendId) = MyProcPid;
977 QUEUE_BACKEND_DBOID(MyBackendId) = MyDatabaseId;
978 LWLockRelease(AsyncQueueLock);
980 /* Now we are listed in the global array, so remember we're listening */
981 amRegisteredListener = true;
984 * Try to move our pointer forward as far as possible. This will skip over
985 * already-committed notifications. Still, we could get notifications that
986 * have already committed before we started to LISTEN.
988 * Note that we are not yet listening on anything, so we won't deliver any
989 * notification to the frontend.
991 * This will also advance the global tail pointer if possible.
993 if (!QUEUE_POS_EQUAL(max, head))
994 asyncQueueReadAllNotifications();
998 * Exec_ListenCommit --- subroutine for AtCommit_Notify
1000 * Add the channel to the list of channels we are listening on.
1003 Exec_ListenCommit(const char *channel)
1005 MemoryContext oldcontext;
1007 /* Do nothing if we are already listening on this channel */
1008 if (IsListeningOn(channel))
1012 * Add the new channel name to listenChannels.
1014 * XXX It is theoretically possible to get an out-of-memory failure here,
1015 * which would be bad because we already committed. For the moment it
1016 * doesn't seem worth trying to guard against that, but maybe improve this
1019 oldcontext = MemoryContextSwitchTo(TopMemoryContext);
1020 listenChannels = lappend(listenChannels, pstrdup(channel));
1021 MemoryContextSwitchTo(oldcontext);
1025 * Exec_UnlistenCommit --- subroutine for AtCommit_Notify
1027 * Remove the specified channel name from listenChannels.
1030 Exec_UnlistenCommit(const char *channel)
1036 elog(DEBUG1, "Exec_UnlistenCommit(%s,%d)", channel, MyProcPid);
1039 foreach(q, listenChannels)
1041 char *lchan = (char *) lfirst(q);
1043 if (strcmp(lchan, channel) == 0)
1045 listenChannels = list_delete_cell(listenChannels, q, prev);
1053 * We do not complain about unlistening something not being listened;
1059 * Exec_UnlistenAllCommit --- subroutine for AtCommit_Notify
1061 * Unlisten on all channels for this backend.
1064 Exec_UnlistenAllCommit(void)
1067 elog(DEBUG1, "Exec_UnlistenAllCommit(%d)", MyProcPid);
1069 list_free_deep(listenChannels);
1070 listenChannels = NIL;
1074 * ProcessCompletedNotifies --- send out signals and self-notifies
1076 * This is called from postgres.c just before going idle at the completion
1077 * of a transaction. If we issued any notifications in the just-completed
1078 * transaction, send signals to other backends to process them, and also
1079 * process the queue ourselves to send messages to our own frontend.
1081 * The reason that this is not done in AtCommit_Notify is that there is
1082 * a nonzero chance of errors here (for example, encoding conversion errors
1083 * while trying to format messages to our frontend). An error during
1084 * AtCommit_Notify would be a PANIC condition. The timing is also arranged
1085 * to ensure that a transaction's self-notifies are delivered to the frontend
1086 * before it gets the terminating ReadyForQuery message.
1088 * Note that we send signals and process the queue even if the transaction
1089 * eventually aborted. This is because we need to clean out whatever got
1090 * added to the queue.
1092 * NOTE: we are outside of any transaction here.
1095 ProcessCompletedNotifies(void)
1097 MemoryContext caller_context;
1100 /* Nothing to do if we didn't send any notifications */
1101 if (!backendHasSentNotifications)
1105 * We reset the flag immediately; otherwise, if any sort of error occurs
1106 * below, we'd be locked up in an infinite loop, because control will come
1107 * right back here after error cleanup.
1109 backendHasSentNotifications = false;
1112 * We must preserve the caller's memory context (probably MessageContext)
1113 * across the transaction we do here.
1115 caller_context = CurrentMemoryContext;
1118 elog(DEBUG1, "ProcessCompletedNotifies");
1121 * We must run asyncQueueReadAllNotifications inside a transaction, else
1122 * bad things happen if it gets an error.
1124 StartTransactionCommand();
1126 /* Send signals to other backends */
1127 signalled = SignalBackends();
1129 if (listenChannels != NIL)
1131 /* Read the queue ourselves, and send relevant stuff to the frontend */
1132 asyncQueueReadAllNotifications();
1134 else if (!signalled)
1137 * If we found no other listening backends, and we aren't listening
1138 * ourselves, then we must execute asyncQueueAdvanceTail to flush the
1139 * queue, because ain't nobody else gonna do it. This prevents queue
1140 * overflow when we're sending useless notifies to nobody. (A new
1141 * listener could have joined since we looked, but if so this is
1144 asyncQueueAdvanceTail();
1147 CommitTransactionCommand();
1149 MemoryContextSwitchTo(caller_context);
1151 /* We don't need pq_flush() here since postgres.c will do one shortly */
1155 * Test whether we are actively listening on the given channel name.
1157 * Note: this function is executed for every notification found in the queue.
1158 * Perhaps it is worth further optimization, eg convert the list to a sorted
1159 * array so we can binary-search it. In practice the list is likely to be
1160 * fairly short, though.
1163 IsListeningOn(const char *channel)
1167 foreach(p, listenChannels)
1169 char *lchan = (char *) lfirst(p);
1171 if (strcmp(lchan, channel) == 0)
1178 * Remove our entry from the listeners array when we are no longer listening
1179 * on any channel. NB: must not fail if we're already not listening.
1182 asyncQueueUnregister(void)
1186 Assert(listenChannels == NIL); /* else caller error */
1188 if (!amRegisteredListener) /* nothing to do */
1191 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1192 /* check if entry is valid and oldest ... */
1193 advanceTail = (MyProcPid == QUEUE_BACKEND_PID(MyBackendId)) &&
1194 QUEUE_POS_EQUAL(QUEUE_BACKEND_POS(MyBackendId), QUEUE_TAIL);
1195 /* ... then mark it invalid */
1196 QUEUE_BACKEND_PID(MyBackendId) = InvalidPid;
1197 QUEUE_BACKEND_DBOID(MyBackendId) = InvalidOid;
1198 LWLockRelease(AsyncQueueLock);
1200 /* mark ourselves as no longer listed in the global array */
1201 amRegisteredListener = false;
1203 /* If we were the laziest backend, try to advance the tail pointer */
1205 asyncQueueAdvanceTail();
1209 * Test whether there is room to insert more notification messages.
1211 * Caller must hold at least shared AsyncQueueLock.
1214 asyncQueueIsFull(void)
1220 * The queue is full if creating a new head page would create a page that
1221 * logically precedes the current global tail pointer, ie, the head
1222 * pointer would wrap around compared to the tail. We cannot create such
1223 * a head page for fear of confusing slru.c. For safety we round the tail
1224 * pointer back to a segment boundary (compare the truncation logic in
1225 * asyncQueueAdvanceTail).
1227 * Note that this test is *not* dependent on how much space there is on
1228 * the current head page. This is necessary because asyncQueueAddEntries
1229 * might try to create the next head page in any case.
1231 nexthead = QUEUE_POS_PAGE(QUEUE_HEAD) + 1;
1232 if (nexthead > QUEUE_MAX_PAGE)
1233 nexthead = 0; /* wrap around */
1234 boundary = QUEUE_POS_PAGE(QUEUE_TAIL);
1235 boundary -= boundary % SLRU_PAGES_PER_SEGMENT;
1236 return asyncQueuePagePrecedes(nexthead, boundary);
1240 * Advance the QueuePosition to the next entry, assuming that the current
1241 * entry is of length entryLength. If we jump to a new page the function
1242 * returns true, else false.
1245 asyncQueueAdvance(volatile QueuePosition *position, int entryLength)
1247 int pageno = QUEUE_POS_PAGE(*position);
1248 int offset = QUEUE_POS_OFFSET(*position);
1249 bool pageJump = false;
1252 * Move to the next writing position: First jump over what we have just
1255 offset += entryLength;
1256 Assert(offset <= QUEUE_PAGESIZE);
1259 * In a second step check if another entry can possibly be written to the
1260 * page. If so, stay here, we have reached the next position. If not, then
1261 * we need to move on to the next page.
1263 if (offset + QUEUEALIGN(AsyncQueueEntryEmptySize) > QUEUE_PAGESIZE)
1266 if (pageno > QUEUE_MAX_PAGE)
1267 pageno = 0; /* wrap around */
1272 SET_QUEUE_POS(*position, pageno, offset);
1277 * Fill the AsyncQueueEntry at *qe with an outbound notification message.
1280 asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe)
1282 size_t channellen = strlen(n->channel);
1283 size_t payloadlen = strlen(n->payload);
1286 Assert(channellen < NAMEDATALEN);
1287 Assert(payloadlen < NOTIFY_PAYLOAD_MAX_LENGTH);
1289 /* The terminators are already included in AsyncQueueEntryEmptySize */
1290 entryLength = AsyncQueueEntryEmptySize + payloadlen + channellen;
1291 entryLength = QUEUEALIGN(entryLength);
1292 qe->length = entryLength;
1293 qe->dboid = MyDatabaseId;
1294 qe->xid = GetCurrentTransactionId();
1295 qe->srcPid = MyProcPid;
1296 memcpy(qe->data, n->channel, channellen + 1);
1297 memcpy(qe->data + channellen + 1, n->payload, payloadlen + 1);
1301 * Add pending notifications to the queue.
1303 * We go page by page here, i.e. we stop once we have to go to a new page but
1304 * we will be called again and then fill that next page. If an entry does not
1305 * fit into the current page, we write a dummy entry with an InvalidOid as the
1306 * database OID in order to fill the page. So every page is always used up to
1307 * the last byte which simplifies reading the page later.
1309 * We are passed the list cell containing the next notification to write
1310 * and return the first still-unwritten cell back. Eventually we will return
1311 * NULL indicating all is done.
1313 * We are holding AsyncQueueLock already from the caller and grab AsyncCtlLock
1314 * locally in this function.
1317 asyncQueueAddEntries(ListCell *nextNotify)
1320 QueuePosition queue_head;
1325 /* We hold both AsyncQueueLock and AsyncCtlLock during this operation */
1326 LWLockAcquire(AsyncCtlLock, LW_EXCLUSIVE);
1329 * We work with a local copy of QUEUE_HEAD, which we write back to shared
1330 * memory upon exiting. The reason for this is that if we have to advance
1331 * to a new page, SimpleLruZeroPage might fail (out of disk space, for
1332 * instance), and we must not advance QUEUE_HEAD if it does. (Otherwise,
1333 * subsequent insertions would try to put entries into a page that slru.c
1334 * thinks doesn't exist yet.) So, use a local position variable. Note
1335 * that if we do fail, any already-inserted queue entries are forgotten;
1336 * this is okay, since they'd be useless anyway after our transaction
1339 queue_head = QUEUE_HEAD;
1341 /* Fetch the current page */
1342 pageno = QUEUE_POS_PAGE(queue_head);
1343 slotno = SimpleLruReadPage(AsyncCtl, pageno, true, InvalidTransactionId);
1344 /* Note we mark the page dirty before writing in it */
1345 AsyncCtl->shared->page_dirty[slotno] = true;
1347 while (nextNotify != NULL)
1349 Notification *n = (Notification *) lfirst(nextNotify);
1351 /* Construct a valid queue entry in local variable qe */
1352 asyncQueueNotificationToEntry(n, &qe);
1354 offset = QUEUE_POS_OFFSET(queue_head);
1356 /* Check whether the entry really fits on the current page */
1357 if (offset + qe.length <= QUEUE_PAGESIZE)
1359 /* OK, so advance nextNotify past this item */
1360 nextNotify = lnext(nextNotify);
1365 * Write a dummy entry to fill up the page. Actually readers will
1366 * only check dboid and since it won't match any reader's database
1367 * OID, they will ignore this entry and move on.
1369 qe.length = QUEUE_PAGESIZE - offset;
1370 qe.dboid = InvalidOid;
1371 qe.data[0] = '\0'; /* empty channel */
1372 qe.data[1] = '\0'; /* empty payload */
1375 /* Now copy qe into the shared buffer page */
1376 memcpy(AsyncCtl->shared->page_buffer[slotno] + offset,
1380 /* Advance queue_head appropriately, and detect if page is full */
1381 if (asyncQueueAdvance(&(queue_head), qe.length))
1384 * Page is full, so we're done here, but first fill the next page
1385 * with zeroes. The reason to do this is to ensure that slru.c's
1386 * idea of the head page is always the same as ours, which avoids
1387 * boundary problems in SimpleLruTruncate. The test in
1388 * asyncQueueIsFull() ensured that there is room to create this
1389 * page without overrunning the queue.
1391 slotno = SimpleLruZeroPage(AsyncCtl, QUEUE_POS_PAGE(queue_head));
1392 /* And exit the loop */
1397 /* Success, so update the global QUEUE_HEAD */
1398 QUEUE_HEAD = queue_head;
1400 LWLockRelease(AsyncCtlLock);
1406 * SQL function to return the fraction of the notification queue currently
1410 pg_notification_queue_usage(PG_FUNCTION_ARGS)
1414 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1415 usage = asyncQueueUsage();
1416 LWLockRelease(AsyncQueueLock);
1418 PG_RETURN_FLOAT8(usage);
1422 * Return the fraction of the queue that is currently occupied.
1424 * The caller must hold AysncQueueLock in (at least) shared mode.
1427 asyncQueueUsage(void)
1429 int headPage = QUEUE_POS_PAGE(QUEUE_HEAD);
1430 int tailPage = QUEUE_POS_PAGE(QUEUE_TAIL);
1433 occupied = headPage - tailPage;
1436 return (double) 0; /* fast exit for common case */
1440 /* head has wrapped around, tail not yet */
1441 occupied += QUEUE_MAX_PAGE + 1;
1444 return (double) occupied / (double) ((QUEUE_MAX_PAGE + 1) / 2);
1448 * Check whether the queue is at least half full, and emit a warning if so.
1450 * This is unlikely given the size of the queue, but possible.
1451 * The warnings show up at most once every QUEUE_FULL_WARN_INTERVAL.
1453 * Caller must hold exclusive AsyncQueueLock.
1456 asyncQueueFillWarning(void)
1461 fillDegree = asyncQueueUsage();
1462 if (fillDegree < 0.5)
1465 t = GetCurrentTimestamp();
1467 if (TimestampDifferenceExceeds(asyncQueueControl->lastQueueFillWarn,
1468 t, QUEUE_FULL_WARN_INTERVAL))
1470 QueuePosition min = QUEUE_HEAD;
1471 int32 minPid = InvalidPid;
1474 for (i = 1; i <= MaxBackends; i++)
1476 if (QUEUE_BACKEND_PID(i) != InvalidPid)
1478 min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
1479 if (QUEUE_POS_EQUAL(min, QUEUE_BACKEND_POS(i)))
1480 minPid = QUEUE_BACKEND_PID(i);
1485 (errmsg("NOTIFY queue is %.0f%% full", fillDegree * 100),
1486 (minPid != InvalidPid ?
1487 errdetail("The server process with PID %d is among those with the oldest transactions.", minPid)
1489 (minPid != InvalidPid ?
1490 errhint("The NOTIFY queue cannot be emptied until that process ends its current transaction.")
1493 asyncQueueControl->lastQueueFillWarn = t;
1498 * Send signals to all listening backends (except our own).
1500 * Returns true if we sent at least one signal.
1502 * Since we need EXCLUSIVE lock anyway we also check the position of the other
1503 * backends and in case one is already up-to-date we don't signal it.
1504 * This can happen if concurrent notifying transactions have sent a signal and
1505 * the signaled backend has read the other notifications and ours in the same
1508 * Since we know the BackendId and the Pid the signalling is quite cheap.
1511 SignalBackends(void)
1513 bool signalled = false;
1521 * Identify all backends that are listening and not already up-to-date. We
1522 * don't want to send signals while holding the AsyncQueueLock, so we just
1523 * build a list of target PIDs.
1525 * XXX in principle these pallocs could fail, which would be bad. Maybe
1526 * preallocate the arrays? But in practice this is only run in trivial
1527 * transactions, so there should surely be space available.
1529 pids = (int32 *) palloc(MaxBackends * sizeof(int32));
1530 ids = (BackendId *) palloc(MaxBackends * sizeof(BackendId));
1533 LWLockAcquire(AsyncQueueLock, LW_EXCLUSIVE);
1534 for (i = 1; i <= MaxBackends; i++)
1536 pid = QUEUE_BACKEND_PID(i);
1537 if (pid != InvalidPid && pid != MyProcPid)
1539 QueuePosition pos = QUEUE_BACKEND_POS(i);
1541 if (!QUEUE_POS_EQUAL(pos, QUEUE_HEAD))
1549 LWLockRelease(AsyncQueueLock);
1551 /* Now send signals */
1552 for (i = 0; i < count; i++)
1557 * Note: assuming things aren't broken, a signal failure here could
1558 * only occur if the target backend exited since we released
1559 * AsyncQueueLock; which is unlikely but certainly possible. So we
1560 * just log a low-level debug message if it happens.
1562 if (SendProcSignal(pid, PROCSIG_NOTIFY_INTERRUPT, ids[i]) < 0)
1563 elog(DEBUG3, "could not signal backend with PID %d: %m", pid);
1577 * This is called at transaction abort.
1579 * Gets rid of pending actions and outbound notifies that we would have
1580 * executed if the transaction got committed.
1583 AtAbort_Notify(void)
1586 * If we LISTEN but then roll back the transaction after PreCommit_Notify,
1587 * we have registered as a listener but have not made any entry in
1588 * listenChannels. In that case, deregister again.
1590 if (amRegisteredListener && listenChannels == NIL)
1591 asyncQueueUnregister();
1594 ClearPendingActionsAndNotifies();
1598 * AtSubStart_Notify() --- Take care of subtransaction start.
1600 * Push empty state for the new subtransaction.
1603 AtSubStart_Notify(void)
1605 MemoryContext old_cxt;
1607 /* Keep the list-of-lists in TopTransactionContext for simplicity */
1608 old_cxt = MemoryContextSwitchTo(TopTransactionContext);
1610 upperPendingActions = lcons(pendingActions, upperPendingActions);
1612 Assert(list_length(upperPendingActions) ==
1613 GetCurrentTransactionNestLevel() - 1);
1615 pendingActions = NIL;
1617 upperPendingNotifies = lcons(pendingNotifies, upperPendingNotifies);
1619 Assert(list_length(upperPendingNotifies) ==
1620 GetCurrentTransactionNestLevel() - 1);
1622 pendingNotifies = NIL;
1624 MemoryContextSwitchTo(old_cxt);
1628 * AtSubCommit_Notify() --- Take care of subtransaction commit.
1630 * Reassign all items in the pending lists to the parent transaction.
1633 AtSubCommit_Notify(void)
1635 List *parentPendingActions;
1636 List *parentPendingNotifies;
1638 parentPendingActions = (List *) linitial(upperPendingActions);
1639 upperPendingActions = list_delete_first(upperPendingActions);
1641 Assert(list_length(upperPendingActions) ==
1642 GetCurrentTransactionNestLevel() - 2);
1645 * Mustn't try to eliminate duplicates here --- see queue_listen()
1647 pendingActions = list_concat(parentPendingActions, pendingActions);
1649 parentPendingNotifies = (List *) linitial(upperPendingNotifies);
1650 upperPendingNotifies = list_delete_first(upperPendingNotifies);
1652 Assert(list_length(upperPendingNotifies) ==
1653 GetCurrentTransactionNestLevel() - 2);
1656 * We could try to eliminate duplicates here, but it seems not worthwhile.
1658 pendingNotifies = list_concat(parentPendingNotifies, pendingNotifies);
1662 * AtSubAbort_Notify() --- Take care of subtransaction abort.
1665 AtSubAbort_Notify(void)
1667 int my_level = GetCurrentTransactionNestLevel();
1670 * All we have to do is pop the stack --- the actions/notifies made in
1671 * this subxact are no longer interesting, and the space will be freed
1672 * when CurTransactionContext is recycled.
1674 * This routine could be called more than once at a given nesting level if
1675 * there is trouble during subxact abort. Avoid dumping core by using
1676 * GetCurrentTransactionNestLevel as the indicator of how far we need to
1679 while (list_length(upperPendingActions) > my_level - 2)
1681 pendingActions = (List *) linitial(upperPendingActions);
1682 upperPendingActions = list_delete_first(upperPendingActions);
1685 while (list_length(upperPendingNotifies) > my_level - 2)
1687 pendingNotifies = (List *) linitial(upperPendingNotifies);
1688 upperPendingNotifies = list_delete_first(upperPendingNotifies);
1693 * HandleNotifyInterrupt
1695 * Signal handler portion of interrupt handling. Let the backend know
1696 * that there's a pending notify interrupt. If we're currently reading
1697 * from the client, this will interrupt the read and
1698 * ProcessClientReadInterrupt() will call ProcessNotifyInterrupt().
1701 HandleNotifyInterrupt(void)
1704 * Note: this is called by a SIGNAL HANDLER. You must be very wary what
1708 /* signal that work needs to be done */
1709 notifyInterruptPending = true;
1711 /* make sure the event is processed in due course */
1716 * ProcessNotifyInterrupt
1718 * This is called just after waiting for a frontend command. If a
1719 * interrupt arrives (via HandleNotifyInterrupt()) while reading, the
1720 * read will be interrupted via the process's latch, and this routine
1721 * will get called. If we are truly idle (ie, *not* inside a transaction
1722 * block), process the incoming notifies.
1725 ProcessNotifyInterrupt(void)
1727 if (IsTransactionOrTransactionBlock())
1728 return; /* not really idle */
1730 while (notifyInterruptPending)
1731 ProcessIncomingNotify();
1736 * Read all pending notifications from the queue, and deliver appropriate
1737 * ones to my frontend. Stop when we reach queue head or an uncommitted
1741 asyncQueueReadAllNotifications(void)
1743 volatile QueuePosition pos;
1744 QueuePosition oldpos;
1748 /* page_buffer must be adequately aligned, so use a union */
1751 char buf[QUEUE_PAGESIZE];
1752 AsyncQueueEntry align;
1755 /* Fetch current state */
1756 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1757 /* Assert checks that we have a valid state entry */
1758 Assert(MyProcPid == QUEUE_BACKEND_PID(MyBackendId));
1759 pos = oldpos = QUEUE_BACKEND_POS(MyBackendId);
1761 LWLockRelease(AsyncQueueLock);
1763 if (QUEUE_POS_EQUAL(pos, head))
1765 /* Nothing to do, we have read all notifications already. */
1770 * Note that we deliver everything that we see in the queue and that
1771 * matches our _current_ listening state.
1772 * Especially we do not take into account different commit times.
1773 * Consider the following example:
1775 * Backend 1: Backend 2:
1777 * transaction starts
1780 * transaction starts
1786 * It could happen that backend 2 sees the notification from backend 1 in
1787 * the queue. Even though the notifying transaction committed before
1788 * the listening transaction, we still deliver the notification.
1790 * The idea is that an additional notification does not do any harm, we
1791 * just need to make sure that we do not miss a notification.
1793 * It is possible that we fail while trying to send a message to our
1794 * frontend (for example, because of encoding conversion failure).
1795 * If that happens it is critical that we not try to send the same
1796 * message over and over again. Therefore, we place a PG_TRY block
1797 * here that will forcibly advance our backend position before we lose
1798 * control to an error. (We could alternatively retake AsyncQueueLock
1799 * and move the position before handling each individual message, but
1800 * that seems like too much lock traffic.)
1809 int curpage = QUEUE_POS_PAGE(pos);
1810 int curoffset = QUEUE_POS_OFFSET(pos);
1815 * We copy the data from SLRU into a local buffer, so as to avoid
1816 * holding the AsyncCtlLock while we are examining the entries and
1817 * possibly transmitting them to our frontend. Copy only the part
1818 * of the page we will actually inspect.
1820 slotno = SimpleLruReadPage_ReadOnly(AsyncCtl, curpage,
1821 InvalidTransactionId);
1822 if (curpage == QUEUE_POS_PAGE(head))
1824 /* we only want to read as far as head */
1825 copysize = QUEUE_POS_OFFSET(head) - curoffset;
1827 copysize = 0; /* just for safety */
1831 /* fetch all the rest of the page */
1832 copysize = QUEUE_PAGESIZE - curoffset;
1834 memcpy(page_buffer.buf + curoffset,
1835 AsyncCtl->shared->page_buffer[slotno] + curoffset,
1837 /* Release lock that we got from SimpleLruReadPage_ReadOnly() */
1838 LWLockRelease(AsyncCtlLock);
1841 * Process messages up to the stop position, end of page, or an
1842 * uncommitted message.
1844 * Our stop position is what we found to be the head's position
1845 * when we entered this function. It might have changed already.
1846 * But if it has, we will receive (or have already received and
1847 * queued) another signal and come here again.
1849 * We are not holding AsyncQueueLock here! The queue can only
1850 * extend beyond the head pointer (see above) and we leave our
1851 * backend's pointer where it is so nobody will truncate or
1852 * rewrite pages under us. Especially we don't want to hold a lock
1853 * while sending the notifications to the frontend.
1855 reachedStop = asyncQueueProcessPageEntries(&pos, head,
1857 } while (!reachedStop);
1861 /* Update shared state */
1862 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1863 QUEUE_BACKEND_POS(MyBackendId) = pos;
1864 advanceTail = QUEUE_POS_EQUAL(oldpos, QUEUE_TAIL);
1865 LWLockRelease(AsyncQueueLock);
1867 /* If we were the laziest backend, try to advance the tail pointer */
1869 asyncQueueAdvanceTail();
1875 /* Update shared state */
1876 LWLockAcquire(AsyncQueueLock, LW_SHARED);
1877 QUEUE_BACKEND_POS(MyBackendId) = pos;
1878 advanceTail = QUEUE_POS_EQUAL(oldpos, QUEUE_TAIL);
1879 LWLockRelease(AsyncQueueLock);
1881 /* If we were the laziest backend, try to advance the tail pointer */
1883 asyncQueueAdvanceTail();
1887 * Fetch notifications from the shared queue, beginning at position current,
1888 * and deliver relevant ones to my frontend.
1890 * The current page must have been fetched into page_buffer from shared
1891 * memory. (We could access the page right in shared memory, but that
1892 * would imply holding the AsyncCtlLock throughout this routine.)
1894 * We stop if we reach the "stop" position, or reach a notification from an
1895 * uncommitted transaction, or reach the end of the page.
1897 * The function returns true once we have reached the stop position or an
1898 * uncommitted notification, and false if we have finished with the page.
1899 * In other words: once it returns true there is no need to look further.
1900 * The QueuePosition *current is advanced past all processed messages.
1903 asyncQueueProcessPageEntries(volatile QueuePosition *current,
1907 bool reachedStop = false;
1908 bool reachedEndOfPage;
1909 AsyncQueueEntry *qe;
1913 QueuePosition thisentry = *current;
1915 if (QUEUE_POS_EQUAL(thisentry, stop))
1918 qe = (AsyncQueueEntry *) (page_buffer + QUEUE_POS_OFFSET(thisentry));
1921 * Advance *current over this message, possibly to the next page. As
1922 * noted in the comments for asyncQueueReadAllNotifications, we must
1923 * do this before possibly failing while processing the message.
1925 reachedEndOfPage = asyncQueueAdvance(current, qe->length);
1927 /* Ignore messages destined for other databases */
1928 if (qe->dboid == MyDatabaseId)
1930 if (TransactionIdIsInProgress(qe->xid))
1933 * The source transaction is still in progress, so we can't
1934 * process this message yet. Break out of the loop, but first
1935 * back up *current so we will reprocess the message next
1936 * time. (Note: it is unlikely but not impossible for
1937 * TransactionIdDidCommit to fail, so we can't really avoid
1938 * this advance-then-back-up behavior when dealing with an
1939 * uncommitted message.)
1941 * Note that we must test TransactionIdIsInProgress before we
1942 * test TransactionIdDidCommit, else we might return a message
1943 * from a transaction that is not yet visible to snapshots;
1944 * compare the comments at the head of tqual.c.
1946 *current = thisentry;
1950 else if (TransactionIdDidCommit(qe->xid))
1952 /* qe->data is the null-terminated channel name */
1953 char *channel = qe->data;
1955 if (IsListeningOn(channel))
1957 /* payload follows channel name */
1958 char *payload = qe->data + strlen(channel) + 1;
1960 NotifyMyFrontEnd(channel, payload, qe->srcPid);
1966 * The source transaction aborted or crashed, so we just
1967 * ignore its notifications.
1972 /* Loop back if we're not at end of page */
1973 } while (!reachedEndOfPage);
1975 if (QUEUE_POS_EQUAL(*current, stop))
1982 * Advance the shared queue tail variable to the minimum of all the
1983 * per-backend tail pointers. Truncate pg_notify space if possible.
1986 asyncQueueAdvanceTail(void)
1994 LWLockAcquire(AsyncQueueLock, LW_EXCLUSIVE);
1996 for (i = 1; i <= MaxBackends; i++)
1998 if (QUEUE_BACKEND_PID(i) != InvalidPid)
1999 min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
2001 oldtailpage = QUEUE_POS_PAGE(QUEUE_TAIL);
2003 LWLockRelease(AsyncQueueLock);
2006 * We can truncate something if the global tail advanced across an SLRU
2009 * XXX it might be better to truncate only once every several segments, to
2010 * reduce the number of directory scans.
2012 newtailpage = QUEUE_POS_PAGE(min);
2013 boundary = newtailpage - (newtailpage % SLRU_PAGES_PER_SEGMENT);
2014 if (asyncQueuePagePrecedes(oldtailpage, boundary))
2017 * SimpleLruTruncate() will ask for AsyncCtlLock but will also release
2020 SimpleLruTruncate(AsyncCtl, newtailpage);
2025 * ProcessIncomingNotify
2027 * Deal with arriving NOTIFYs from other backends as soon as it's safe to
2028 * do so. This used to be called from the PROCSIG_NOTIFY_INTERRUPT
2029 * signal handler, but isn't anymore.
2031 * Scan the queue for arriving notifications and report them to my front
2034 * NOTE: since we are outside any transaction, we must create our own.
2037 ProcessIncomingNotify(void)
2039 /* We *must* reset the flag */
2040 notifyInterruptPending = false;
2042 /* Do nothing else if we aren't actively listening */
2043 if (listenChannels == NIL)
2047 elog(DEBUG1, "ProcessIncomingNotify");
2049 set_ps_display("notify interrupt", false);
2052 * We must run asyncQueueReadAllNotifications inside a transaction, else
2053 * bad things happen if it gets an error.
2055 StartTransactionCommand();
2057 asyncQueueReadAllNotifications();
2059 CommitTransactionCommand();
2062 * Must flush the notify messages to ensure frontend gets them promptly.
2066 set_ps_display("idle", false);
2069 elog(DEBUG1, "ProcessIncomingNotify: done");
2073 * Send NOTIFY message to my front end.
2076 NotifyMyFrontEnd(const char *channel, const char *payload, int32 srcPid)
2078 if (whereToSendOutput == DestRemote)
2082 pq_beginmessage(&buf, 'A');
2083 pq_sendint(&buf, srcPid, sizeof(int32));
2084 pq_sendstring(&buf, channel);
2085 if (PG_PROTOCOL_MAJOR(FrontendProtocol) >= 3)
2086 pq_sendstring(&buf, payload);
2087 pq_endmessage(&buf);
2090 * NOTE: we do not do pq_flush() here. For a self-notify, it will
2091 * happen at the end of the transaction, and for incoming notifies
2092 * ProcessIncomingNotify will do it after finding all the notifies.
2096 elog(INFO, "NOTIFY for \"%s\" payload \"%s\"", channel, payload);
2099 /* Does pendingNotifies include the given channel/payload? */
2101 AsyncExistsPendingNotify(const char *channel, const char *payload)
2106 if (pendingNotifies == NIL)
2109 if (payload == NULL)
2113 * We need to append new elements to the end of the list in order to keep
2114 * the order. However, on the other hand we'd like to check the list
2115 * backwards in order to make duplicate-elimination a tad faster when the
2116 * same condition is signaled many times in a row. So as a compromise we
2117 * check the tail element first which we can access directly. If this
2118 * doesn't match, we check the whole list.
2120 * As we are not checking our parents' lists, we can still get duplicates
2121 * in combination with subtransactions, like in:
2130 n = (Notification *) llast(pendingNotifies);
2131 if (strcmp(n->channel, channel) == 0 &&
2132 strcmp(n->payload, payload) == 0)
2135 foreach(p, pendingNotifies)
2137 n = (Notification *) lfirst(p);
2139 if (strcmp(n->channel, channel) == 0 &&
2140 strcmp(n->payload, payload) == 0)
2147 /* Clear the pendingActions and pendingNotifies lists. */
2149 ClearPendingActionsAndNotifies(void)
2152 * We used to have to explicitly deallocate the list members and nodes,
2153 * because they were malloc'd. Now, since we know they are palloc'd in
2154 * CurTransactionContext, we need not do that --- they'll go away
2155 * automatically at transaction exit. We need only reset the list head
2158 pendingActions = NIL;
2159 pendingNotifies = NIL;