Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * async.c
4 : * Asynchronous notification: NOTIFY, LISTEN, UNLISTEN
5 : *
6 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : * IDENTIFICATION
10 : * src/backend/commands/async.c
11 : *
12 : *-------------------------------------------------------------------------
13 : */
14 :
15 : /*-------------------------------------------------------------------------
16 : * Async Notification Model as of 9.0:
17 : *
18 : * 1. Multiple backends on same machine. Multiple backends listening on
19 : * several channels. (Channels are also called "conditions" in other
20 : * parts of the code.)
21 : *
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.
26 : *
27 : * There is no central knowledge of which backend listens on which channel;
28 : * every backend has its own list of interesting channels.
29 : *
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.
36 : *
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.
40 : *
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.
48 : *
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.
51 : *
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
57 : * payload parameter.
58 : *
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
65 : * will roll back.
66 : *
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.
69 : *
70 : * After commit we are called another time (AtCommit_Notify()). Here we
71 : * make any actual updates to the effective listen state (listenChannels).
72 : * Then we signal any backends that may be interested in our messages
73 : * (including our own backend, if listening). This is done by
74 : * SignalBackends(), which scans the list of listening backends and sends a
75 : * PROCSIG_NOTIFY_INTERRUPT signal to every listening backend (we don't
76 : * know which backend is listening on which channel so we must signal them
77 : * all). We can exclude backends that are already up to date, though, and
78 : * we can also exclude backends that are in other databases (unless they
79 : * are way behind and should be kicked to make them advance their
80 : * pointers).
81 : *
82 : * Finally, after we are out of the transaction altogether and about to go
83 : * idle, we scan the queue for messages that need to be sent to our
84 : * frontend (which might be notifies from other backends, or self-notifies
85 : * from our own). This step is not part of the CommitTransaction sequence
86 : * for two important reasons. First, we could get errors while sending
87 : * data to our frontend, and it's really bad for errors to happen in
88 : * post-commit cleanup. Second, in cases where a procedure issues commits
89 : * within a single frontend command, we don't want to send notifies to our
90 : * frontend until the command is done; but notifies to other backends
91 : * should go out immediately after each commit.
92 : *
93 : * 5. Upon receipt of a PROCSIG_NOTIFY_INTERRUPT signal, the signal handler
94 : * sets the process's latch, which triggers the event to be processed
95 : * immediately if this backend is idle (i.e., it is waiting for a frontend
96 : * command and is not within a transaction block. C.f.
97 : * ProcessClientReadInterrupt()). Otherwise the handler may only set a
98 : * flag, which will cause the processing to occur just before we next go
99 : * idle.
100 : *
101 : * Inbound-notify processing consists of reading all of the notifications
102 : * that have arrived since scanning last time. We read every notification
103 : * until we reach either a notification from an uncommitted transaction or
104 : * the head pointer's position.
105 : *
106 : * 6. To limit disk space consumption, the tail pointer needs to be advanced
107 : * so that old pages can be truncated. This is relatively expensive
108 : * (notably, it requires an exclusive lock), so we don't want to do it
109 : * often. We make sending backends do this work if they advanced the queue
110 : * head into a new page, but only once every QUEUE_CLEANUP_DELAY pages.
111 : *
112 : * An application that listens on the same channel it notifies will get
113 : * NOTIFY messages for its own NOTIFYs. These can be ignored, if not useful,
114 : * by comparing be_pid in the NOTIFY message to the application's own backend's
115 : * PID. (As of FE/BE protocol 2.0, the backend's PID is provided to the
116 : * frontend during startup.) The above design guarantees that notifies from
117 : * other backends will never be missed by ignoring self-notifies.
118 : *
119 : * The amount of shared memory used for notify management (notify_buffers)
120 : * can be varied without affecting anything but performance. The maximum
121 : * amount of notification data that can be queued at one time is determined
122 : * by max_notify_queue_pages GUC.
123 : *-------------------------------------------------------------------------
124 : */
125 :
126 : #include "postgres.h"
127 :
128 : #include <limits.h>
129 : #include <unistd.h>
130 : #include <signal.h>
131 :
132 : #include "access/parallel.h"
133 : #include "access/slru.h"
134 : #include "access/transam.h"
135 : #include "access/xact.h"
136 : #include "catalog/pg_database.h"
137 : #include "commands/async.h"
138 : #include "common/hashfn.h"
139 : #include "funcapi.h"
140 : #include "libpq/libpq.h"
141 : #include "libpq/pqformat.h"
142 : #include "miscadmin.h"
143 : #include "storage/ipc.h"
144 : #include "storage/lmgr.h"
145 : #include "storage/procsignal.h"
146 : #include "tcop/tcopprot.h"
147 : #include "utils/builtins.h"
148 : #include "utils/guc_hooks.h"
149 : #include "utils/memutils.h"
150 : #include "utils/ps_status.h"
151 : #include "utils/snapmgr.h"
152 : #include "utils/timestamp.h"
153 :
154 :
155 : /*
156 : * Maximum size of a NOTIFY payload, including terminating NULL. This
157 : * must be kept small enough so that a notification message fits on one
158 : * SLRU page. The magic fudge factor here is noncritical as long as it's
159 : * more than AsyncQueueEntryEmptySize --- we make it significantly bigger
160 : * than that, so changes in that data structure won't affect user-visible
161 : * restrictions.
162 : */
163 : #define NOTIFY_PAYLOAD_MAX_LENGTH (BLCKSZ - NAMEDATALEN - 128)
164 :
165 : /*
166 : * Struct representing an entry in the global notify queue
167 : *
168 : * This struct declaration has the maximal length, but in a real queue entry
169 : * the data area is only big enough for the actual channel and payload strings
170 : * (each null-terminated). AsyncQueueEntryEmptySize is the minimum possible
171 : * entry size, if both channel and payload strings are empty (but note it
172 : * doesn't include alignment padding).
173 : *
174 : * The "length" field should always be rounded up to the next QUEUEALIGN
175 : * multiple so that all fields are properly aligned.
176 : */
177 : typedef struct AsyncQueueEntry
178 : {
179 : int length; /* total allocated length of entry */
180 : Oid dboid; /* sender's database OID */
181 : TransactionId xid; /* sender's XID */
182 : int32 srcPid; /* sender's PID */
183 : char data[NAMEDATALEN + NOTIFY_PAYLOAD_MAX_LENGTH];
184 : } AsyncQueueEntry;
185 :
186 : /* Currently, no field of AsyncQueueEntry requires more than int alignment */
187 : #define QUEUEALIGN(len) INTALIGN(len)
188 :
189 : #define AsyncQueueEntryEmptySize (offsetof(AsyncQueueEntry, data) + 2)
190 :
191 : /*
192 : * Struct describing a queue position, and assorted macros for working with it
193 : */
194 : typedef struct QueuePosition
195 : {
196 : int64 page; /* SLRU page number */
197 : int offset; /* byte offset within page */
198 : } QueuePosition;
199 :
200 : #define QUEUE_POS_PAGE(x) ((x).page)
201 : #define QUEUE_POS_OFFSET(x) ((x).offset)
202 :
203 : #define SET_QUEUE_POS(x,y,z) \
204 : do { \
205 : (x).page = (y); \
206 : (x).offset = (z); \
207 : } while (0)
208 :
209 : #define QUEUE_POS_EQUAL(x,y) \
210 : ((x).page == (y).page && (x).offset == (y).offset)
211 :
212 : #define QUEUE_POS_IS_ZERO(x) \
213 : ((x).page == 0 && (x).offset == 0)
214 :
215 : /* choose logically smaller QueuePosition */
216 : #define QUEUE_POS_MIN(x,y) \
217 : (asyncQueuePagePrecedes((x).page, (y).page) ? (x) : \
218 : (x).page != (y).page ? (y) : \
219 : (x).offset < (y).offset ? (x) : (y))
220 :
221 : /* choose logically larger QueuePosition */
222 : #define QUEUE_POS_MAX(x,y) \
223 : (asyncQueuePagePrecedes((x).page, (y).page) ? (y) : \
224 : (x).page != (y).page ? (x) : \
225 : (x).offset > (y).offset ? (x) : (y))
226 :
227 : /*
228 : * Parameter determining how often we try to advance the tail pointer:
229 : * we do that after every QUEUE_CLEANUP_DELAY pages of NOTIFY data. This is
230 : * also the distance by which a backend in another database needs to be
231 : * behind before we'll decide we need to wake it up to advance its pointer.
232 : *
233 : * Resist the temptation to make this really large. While that would save
234 : * work in some places, it would add cost in others. In particular, this
235 : * should likely be less than notify_buffers, to ensure that backends
236 : * catch up before the pages they'll need to read fall out of SLRU cache.
237 : */
238 : #define QUEUE_CLEANUP_DELAY 4
239 :
240 : /*
241 : * Struct describing a listening backend's status
242 : */
243 : typedef struct QueueBackendStatus
244 : {
245 : int32 pid; /* either a PID or InvalidPid */
246 : Oid dboid; /* backend's database OID, or InvalidOid */
247 : ProcNumber nextListener; /* id of next listener, or INVALID_PROC_NUMBER */
248 : QueuePosition pos; /* backend has read queue up to here */
249 : } QueueBackendStatus;
250 :
251 : /*
252 : * Shared memory state for LISTEN/NOTIFY (excluding its SLRU stuff)
253 : *
254 : * The AsyncQueueControl structure is protected by the NotifyQueueLock and
255 : * NotifyQueueTailLock.
256 : *
257 : * When holding NotifyQueueLock in SHARED mode, backends may only inspect
258 : * their own entries as well as the head and tail pointers. Consequently we
259 : * can allow a backend to update its own record while holding only SHARED lock
260 : * (since no other backend will inspect it).
261 : *
262 : * When holding NotifyQueueLock in EXCLUSIVE mode, backends can inspect the
263 : * entries of other backends and also change the head pointer. When holding
264 : * both NotifyQueueLock and NotifyQueueTailLock in EXCLUSIVE mode, backends
265 : * can change the tail pointers.
266 : *
267 : * SLRU buffer pool is divided in banks and bank wise SLRU lock is used as
268 : * the control lock for the pg_notify SLRU buffers.
269 : * In order to avoid deadlocks, whenever we need multiple locks, we first get
270 : * NotifyQueueTailLock, then NotifyQueueLock, and lastly SLRU bank lock.
271 : *
272 : * Each backend uses the backend[] array entry with index equal to its
273 : * ProcNumber. We rely on this to make SendProcSignal fast.
274 : *
275 : * The backend[] array entries for actively-listening backends are threaded
276 : * together using firstListener and the nextListener links, so that we can
277 : * scan them without having to iterate over inactive entries. We keep this
278 : * list in order by ProcNumber so that the scan is cache-friendly when there
279 : * are many active entries.
280 : */
281 : typedef struct AsyncQueueControl
282 : {
283 : QueuePosition head; /* head points to the next free location */
284 : QueuePosition tail; /* tail must be <= the queue position of every
285 : * listening backend */
286 : int64 stopPage; /* oldest unrecycled page; must be <=
287 : * tail.page */
288 : ProcNumber firstListener; /* id of first listener, or
289 : * INVALID_PROC_NUMBER */
290 : TimestampTz lastQueueFillWarn; /* time of last queue-full msg */
291 : QueueBackendStatus backend[FLEXIBLE_ARRAY_MEMBER];
292 : } AsyncQueueControl;
293 :
294 : static AsyncQueueControl *asyncQueueControl;
295 :
296 : #define QUEUE_HEAD (asyncQueueControl->head)
297 : #define QUEUE_TAIL (asyncQueueControl->tail)
298 : #define QUEUE_STOP_PAGE (asyncQueueControl->stopPage)
299 : #define QUEUE_FIRST_LISTENER (asyncQueueControl->firstListener)
300 : #define QUEUE_BACKEND_PID(i) (asyncQueueControl->backend[i].pid)
301 : #define QUEUE_BACKEND_DBOID(i) (asyncQueueControl->backend[i].dboid)
302 : #define QUEUE_NEXT_LISTENER(i) (asyncQueueControl->backend[i].nextListener)
303 : #define QUEUE_BACKEND_POS(i) (asyncQueueControl->backend[i].pos)
304 :
305 : /*
306 : * The SLRU buffer area through which we access the notification queue
307 : */
308 : static SlruCtlData NotifyCtlData;
309 :
310 : #define NotifyCtl (&NotifyCtlData)
311 : #define QUEUE_PAGESIZE BLCKSZ
312 :
313 : #define QUEUE_FULL_WARN_INTERVAL 5000 /* warn at most once every 5s */
314 :
315 : /*
316 : * listenChannels identifies the channels we are actually listening to
317 : * (ie, have committed a LISTEN on). It is a simple list of channel names,
318 : * allocated in TopMemoryContext.
319 : */
320 : static List *listenChannels = NIL; /* list of C strings */
321 :
322 : /*
323 : * State for pending LISTEN/UNLISTEN actions consists of an ordered list of
324 : * all actions requested in the current transaction. As explained above,
325 : * we don't actually change listenChannels until we reach transaction commit.
326 : *
327 : * The list is kept in CurTransactionContext. In subtransactions, each
328 : * subtransaction has its own list in its own CurTransactionContext, but
329 : * successful subtransactions attach their lists to their parent's list.
330 : * Failed subtransactions simply discard their lists.
331 : */
332 : typedef enum
333 : {
334 : LISTEN_LISTEN,
335 : LISTEN_UNLISTEN,
336 : LISTEN_UNLISTEN_ALL,
337 : } ListenActionKind;
338 :
339 : typedef struct
340 : {
341 : ListenActionKind action;
342 : char channel[FLEXIBLE_ARRAY_MEMBER]; /* nul-terminated string */
343 : } ListenAction;
344 :
345 : typedef struct ActionList
346 : {
347 : int nestingLevel; /* current transaction nesting depth */
348 : List *actions; /* list of ListenAction structs */
349 : struct ActionList *upper; /* details for upper transaction levels */
350 : } ActionList;
351 :
352 : static ActionList *pendingActions = NULL;
353 :
354 : /*
355 : * State for outbound notifies consists of a list of all channels+payloads
356 : * NOTIFYed in the current transaction. We do not actually perform a NOTIFY
357 : * until and unless the transaction commits. pendingNotifies is NULL if no
358 : * NOTIFYs have been done in the current (sub) transaction.
359 : *
360 : * We discard duplicate notify events issued in the same transaction.
361 : * Hence, in addition to the list proper (which we need to track the order
362 : * of the events, since we guarantee to deliver them in order), we build a
363 : * hash table which we can probe to detect duplicates. Since building the
364 : * hash table is somewhat expensive, we do so only once we have at least
365 : * MIN_HASHABLE_NOTIFIES events queued in the current (sub) transaction;
366 : * before that we just scan the events linearly.
367 : *
368 : * The list is kept in CurTransactionContext. In subtransactions, each
369 : * subtransaction has its own list in its own CurTransactionContext, but
370 : * successful subtransactions add their entries to their parent's list.
371 : * Failed subtransactions simply discard their lists. Since these lists
372 : * are independent, there may be notify events in a subtransaction's list
373 : * that duplicate events in some ancestor (sub) transaction; we get rid of
374 : * the dups when merging the subtransaction's list into its parent's.
375 : *
376 : * Note: the action and notify lists do not interact within a transaction.
377 : * In particular, if a transaction does NOTIFY and then LISTEN on the same
378 : * condition name, it will get a self-notify at commit. This is a bit odd
379 : * but is consistent with our historical behavior.
380 : */
381 : typedef struct Notification
382 : {
383 : uint16 channel_len; /* length of channel-name string */
384 : uint16 payload_len; /* length of payload string */
385 : /* null-terminated channel name, then null-terminated payload follow */
386 : char data[FLEXIBLE_ARRAY_MEMBER];
387 : } Notification;
388 :
389 : typedef struct NotificationList
390 : {
391 : int nestingLevel; /* current transaction nesting depth */
392 : List *events; /* list of Notification structs */
393 : HTAB *hashtab; /* hash of NotificationHash structs, or NULL */
394 : struct NotificationList *upper; /* details for upper transaction levels */
395 : } NotificationList;
396 :
397 : #define MIN_HASHABLE_NOTIFIES 16 /* threshold to build hashtab */
398 :
399 : struct NotificationHash
400 : {
401 : Notification *event; /* => the actual Notification struct */
402 : };
403 :
404 : static NotificationList *pendingNotifies = NULL;
405 :
406 : /*
407 : * Inbound notifications are initially processed by HandleNotifyInterrupt(),
408 : * called from inside a signal handler. That just sets the
409 : * notifyInterruptPending flag and sets the process
410 : * latch. ProcessNotifyInterrupt() will then be called whenever it's safe to
411 : * actually deal with the interrupt.
412 : */
413 : volatile sig_atomic_t notifyInterruptPending = false;
414 :
415 : /* True if we've registered an on_shmem_exit cleanup */
416 : static bool unlistenExitRegistered = false;
417 :
418 : /* True if we're currently registered as a listener in asyncQueueControl */
419 : static bool amRegisteredListener = false;
420 :
421 : /* have we advanced to a page that's a multiple of QUEUE_CLEANUP_DELAY? */
422 : static bool tryAdvanceTail = false;
423 :
424 : /* GUC parameters */
425 : bool Trace_notify = false;
426 :
427 : /* For 8 KB pages this gives 8 GB of disk space */
428 : int max_notify_queue_pages = 1048576;
429 :
430 : /* local function prototypes */
431 : static inline int64 asyncQueuePageDiff(int64 p, int64 q);
432 : static inline bool asyncQueuePagePrecedes(int64 p, int64 q);
433 : static void queue_listen(ListenActionKind action, const char *channel);
434 : static void Async_UnlistenOnExit(int code, Datum arg);
435 : static void Exec_ListenPreCommit(void);
436 : static void Exec_ListenCommit(const char *channel);
437 : static void Exec_UnlistenCommit(const char *channel);
438 : static void Exec_UnlistenAllCommit(void);
439 : static bool IsListeningOn(const char *channel);
440 : static void asyncQueueUnregister(void);
441 : static bool asyncQueueIsFull(void);
442 : static bool asyncQueueAdvance(volatile QueuePosition *position, int entryLength);
443 : static void asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe);
444 : static ListCell *asyncQueueAddEntries(ListCell *nextNotify);
445 : static double asyncQueueUsage(void);
446 : static void asyncQueueFillWarning(void);
447 : static void SignalBackends(void);
448 : static void asyncQueueReadAllNotifications(void);
449 : static bool asyncQueueProcessPageEntries(volatile QueuePosition *current,
450 : QueuePosition stop,
451 : char *page_buffer,
452 : Snapshot snapshot);
453 : static void asyncQueueAdvanceTail(void);
454 : static void ProcessIncomingNotify(bool flush);
455 : static bool AsyncExistsPendingNotify(Notification *n);
456 : static void AddEventToPendingNotifies(Notification *n);
457 : static uint32 notification_hash(const void *key, Size keysize);
458 : static int notification_match(const void *key1, const void *key2, Size keysize);
459 : static void ClearPendingActionsAndNotifies(void);
460 :
461 : /*
462 : * Compute the difference between two queue page numbers.
463 : * Previously this function accounted for a wraparound.
464 : */
465 : static inline int64
466 0 : asyncQueuePageDiff(int64 p, int64 q)
467 : {
468 0 : return p - q;
469 : }
470 :
471 : /*
472 : * Determines whether p precedes q.
473 : * Previously this function accounted for a wraparound.
474 : */
475 : static inline bool
476 50 : asyncQueuePagePrecedes(int64 p, int64 q)
477 : {
478 50 : return p < q;
479 : }
480 :
481 : /*
482 : * Report space needed for our shared memory area
483 : */
484 : Size
485 3906 : AsyncShmemSize(void)
486 : {
487 : Size size;
488 :
489 : /* This had better match AsyncShmemInit */
490 3906 : size = mul_size(MaxBackends, sizeof(QueueBackendStatus));
491 3906 : size = add_size(size, offsetof(AsyncQueueControl, backend));
492 :
493 3906 : size = add_size(size, SimpleLruShmemSize(notify_buffers, 0));
494 :
495 3906 : return size;
496 : }
497 :
498 : /*
499 : * Initialize our shared memory area
500 : */
501 : void
502 2100 : AsyncShmemInit(void)
503 : {
504 : bool found;
505 : Size size;
506 :
507 : /*
508 : * Create or attach to the AsyncQueueControl structure.
509 : */
510 2100 : size = mul_size(MaxBackends, sizeof(QueueBackendStatus));
511 2100 : size = add_size(size, offsetof(AsyncQueueControl, backend));
512 :
513 2100 : asyncQueueControl = (AsyncQueueControl *)
514 2100 : ShmemInitStruct("Async Queue Control", size, &found);
515 :
516 2100 : if (!found)
517 : {
518 : /* First time through, so initialize it */
519 2100 : SET_QUEUE_POS(QUEUE_HEAD, 0, 0);
520 2100 : SET_QUEUE_POS(QUEUE_TAIL, 0, 0);
521 2100 : QUEUE_STOP_PAGE = 0;
522 2100 : QUEUE_FIRST_LISTENER = INVALID_PROC_NUMBER;
523 2100 : asyncQueueControl->lastQueueFillWarn = 0;
524 199686 : for (int i = 0; i < MaxBackends; i++)
525 : {
526 197586 : QUEUE_BACKEND_PID(i) = InvalidPid;
527 197586 : QUEUE_BACKEND_DBOID(i) = InvalidOid;
528 197586 : QUEUE_NEXT_LISTENER(i) = INVALID_PROC_NUMBER;
529 197586 : SET_QUEUE_POS(QUEUE_BACKEND_POS(i), 0, 0);
530 : }
531 : }
532 :
533 : /*
534 : * Set up SLRU management of the pg_notify data. Note that long segment
535 : * names are used in order to avoid wraparound.
536 : */
537 2100 : NotifyCtl->PagePrecedes = asyncQueuePagePrecedes;
538 2100 : SimpleLruInit(NotifyCtl, "notify", notify_buffers, 0,
539 : "pg_notify", LWTRANCHE_NOTIFY_BUFFER, LWTRANCHE_NOTIFY_SLRU,
540 : SYNC_HANDLER_NONE, true);
541 :
542 2100 : if (!found)
543 : {
544 : /*
545 : * During start or reboot, clean out the pg_notify directory.
546 : */
547 2100 : (void) SlruScanDirectory(NotifyCtl, SlruScanDirCbDeleteAll, NULL);
548 : }
549 2100 : }
550 :
551 :
552 : /*
553 : * pg_notify -
554 : * SQL function to send a notification event
555 : */
556 : Datum
557 2108 : pg_notify(PG_FUNCTION_ARGS)
558 : {
559 : const char *channel;
560 : const char *payload;
561 :
562 2108 : if (PG_ARGISNULL(0))
563 6 : channel = "";
564 : else
565 2102 : channel = text_to_cstring(PG_GETARG_TEXT_PP(0));
566 :
567 2108 : if (PG_ARGISNULL(1))
568 12 : payload = "";
569 : else
570 2096 : payload = text_to_cstring(PG_GETARG_TEXT_PP(1));
571 :
572 : /* For NOTIFY as a statement, this is checked in ProcessUtility */
573 2108 : PreventCommandDuringRecovery("NOTIFY");
574 :
575 2108 : Async_Notify(channel, payload);
576 :
577 2090 : PG_RETURN_VOID();
578 : }
579 :
580 :
581 : /*
582 : * Async_Notify
583 : *
584 : * This is executed by the SQL notify command.
585 : *
586 : * Adds the message to the list of pending notifies.
587 : * Actual notification happens during transaction commit.
588 : * ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
589 : */
590 : void
591 2206 : Async_Notify(const char *channel, const char *payload)
592 : {
593 2206 : int my_level = GetCurrentTransactionNestLevel();
594 : size_t channel_len;
595 : size_t payload_len;
596 : Notification *n;
597 : MemoryContext oldcontext;
598 :
599 2206 : if (IsParallelWorker())
600 0 : elog(ERROR, "cannot send notifications from a parallel worker");
601 :
602 2206 : if (Trace_notify)
603 0 : elog(DEBUG1, "Async_Notify(%s)", channel);
604 :
605 2206 : channel_len = channel ? strlen(channel) : 0;
606 2206 : payload_len = payload ? strlen(payload) : 0;
607 :
608 : /* a channel name must be specified */
609 2206 : if (channel_len == 0)
610 12 : ereport(ERROR,
611 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
612 : errmsg("channel name cannot be empty")));
613 :
614 : /* enforce length limits */
615 2194 : if (channel_len >= NAMEDATALEN)
616 6 : ereport(ERROR,
617 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
618 : errmsg("channel name too long")));
619 :
620 2188 : if (payload_len >= NOTIFY_PAYLOAD_MAX_LENGTH)
621 0 : ereport(ERROR,
622 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
623 : errmsg("payload string too long")));
624 :
625 : /*
626 : * We must construct the Notification entry, even if we end up not using
627 : * it, in order to compare it cheaply to existing list entries.
628 : *
629 : * The notification list needs to live until end of transaction, so store
630 : * it in the transaction context.
631 : */
632 2188 : oldcontext = MemoryContextSwitchTo(CurTransactionContext);
633 :
634 2188 : n = (Notification *) palloc(offsetof(Notification, data) +
635 2188 : channel_len + payload_len + 2);
636 2188 : n->channel_len = channel_len;
637 2188 : n->payload_len = payload_len;
638 2188 : strcpy(n->data, channel);
639 2188 : if (payload)
640 2162 : strcpy(n->data + channel_len + 1, payload);
641 : else
642 26 : n->data[channel_len + 1] = '\0';
643 :
644 2188 : if (pendingNotifies == NULL || my_level > pendingNotifies->nestingLevel)
645 100 : {
646 : NotificationList *notifies;
647 :
648 : /*
649 : * First notify event in current (sub)xact. Note that we allocate the
650 : * NotificationList in TopTransactionContext; the nestingLevel might
651 : * get changed later by AtSubCommit_Notify.
652 : */
653 : notifies = (NotificationList *)
654 100 : MemoryContextAlloc(TopTransactionContext,
655 : sizeof(NotificationList));
656 100 : notifies->nestingLevel = my_level;
657 100 : notifies->events = list_make1(n);
658 : /* We certainly don't need a hashtable yet */
659 100 : notifies->hashtab = NULL;
660 100 : notifies->upper = pendingNotifies;
661 100 : pendingNotifies = notifies;
662 : }
663 : else
664 : {
665 : /* Now check for duplicates */
666 2088 : if (AsyncExistsPendingNotify(n))
667 : {
668 : /* It's a dup, so forget it */
669 24 : pfree(n);
670 24 : MemoryContextSwitchTo(oldcontext);
671 24 : return;
672 : }
673 :
674 : /* Append more events to existing list */
675 2064 : AddEventToPendingNotifies(n);
676 : }
677 :
678 2164 : MemoryContextSwitchTo(oldcontext);
679 : }
680 :
681 : /*
682 : * queue_listen
683 : * Common code for listen, unlisten, unlisten all commands.
684 : *
685 : * Adds the request to the list of pending actions.
686 : * Actual update of the listenChannels list happens during transaction
687 : * commit.
688 : */
689 : static void
690 114 : queue_listen(ListenActionKind action, const char *channel)
691 : {
692 : MemoryContext oldcontext;
693 : ListenAction *actrec;
694 114 : int my_level = GetCurrentTransactionNestLevel();
695 :
696 : /*
697 : * Unlike Async_Notify, we don't try to collapse out duplicates. It would
698 : * be too complicated to ensure we get the right interactions of
699 : * conflicting LISTEN/UNLISTEN/UNLISTEN_ALL, and it's unlikely that there
700 : * would be any performance benefit anyway in sane applications.
701 : */
702 114 : oldcontext = MemoryContextSwitchTo(CurTransactionContext);
703 :
704 : /* space for terminating null is included in sizeof(ListenAction) */
705 114 : actrec = (ListenAction *) palloc(offsetof(ListenAction, channel) +
706 114 : strlen(channel) + 1);
707 114 : actrec->action = action;
708 114 : strcpy(actrec->channel, channel);
709 :
710 114 : if (pendingActions == NULL || my_level > pendingActions->nestingLevel)
711 100 : {
712 : ActionList *actions;
713 :
714 : /*
715 : * First action in current sub(xact). Note that we allocate the
716 : * ActionList in TopTransactionContext; the nestingLevel might get
717 : * changed later by AtSubCommit_Notify.
718 : */
719 : actions = (ActionList *)
720 100 : MemoryContextAlloc(TopTransactionContext, sizeof(ActionList));
721 100 : actions->nestingLevel = my_level;
722 100 : actions->actions = list_make1(actrec);
723 100 : actions->upper = pendingActions;
724 100 : pendingActions = actions;
725 : }
726 : else
727 14 : pendingActions->actions = lappend(pendingActions->actions, actrec);
728 :
729 114 : MemoryContextSwitchTo(oldcontext);
730 114 : }
731 :
732 : /*
733 : * Async_Listen
734 : *
735 : * This is executed by the SQL listen command.
736 : */
737 : void
738 74 : Async_Listen(const char *channel)
739 : {
740 74 : if (Trace_notify)
741 0 : elog(DEBUG1, "Async_Listen(%s,%d)", channel, MyProcPid);
742 :
743 74 : queue_listen(LISTEN_LISTEN, channel);
744 74 : }
745 :
746 : /*
747 : * Async_Unlisten
748 : *
749 : * This is executed by the SQL unlisten command.
750 : */
751 : void
752 6 : Async_Unlisten(const char *channel)
753 : {
754 6 : if (Trace_notify)
755 0 : elog(DEBUG1, "Async_Unlisten(%s,%d)", channel, MyProcPid);
756 :
757 : /* If we couldn't possibly be listening, no need to queue anything */
758 6 : if (pendingActions == NULL && !unlistenExitRegistered)
759 0 : return;
760 :
761 6 : queue_listen(LISTEN_UNLISTEN, channel);
762 : }
763 :
764 : /*
765 : * Async_UnlistenAll
766 : *
767 : * This is invoked by UNLISTEN * command, and also at backend exit.
768 : */
769 : void
770 38 : Async_UnlistenAll(void)
771 : {
772 38 : if (Trace_notify)
773 0 : elog(DEBUG1, "Async_UnlistenAll(%d)", MyProcPid);
774 :
775 : /* If we couldn't possibly be listening, no need to queue anything */
776 38 : if (pendingActions == NULL && !unlistenExitRegistered)
777 4 : return;
778 :
779 34 : queue_listen(LISTEN_UNLISTEN_ALL, "");
780 : }
781 :
782 : /*
783 : * SQL function: return a set of the channel names this backend is actively
784 : * listening to.
785 : *
786 : * Note: this coding relies on the fact that the listenChannels list cannot
787 : * change within a transaction.
788 : */
789 : Datum
790 18 : pg_listening_channels(PG_FUNCTION_ARGS)
791 : {
792 : FuncCallContext *funcctx;
793 :
794 : /* stuff done only on the first call of the function */
795 18 : if (SRF_IS_FIRSTCALL())
796 : {
797 : /* create a function context for cross-call persistence */
798 12 : funcctx = SRF_FIRSTCALL_INIT();
799 : }
800 :
801 : /* stuff done on every call of the function */
802 18 : funcctx = SRF_PERCALL_SETUP();
803 :
804 18 : if (funcctx->call_cntr < list_length(listenChannels))
805 : {
806 6 : char *channel = (char *) list_nth(listenChannels,
807 6 : funcctx->call_cntr);
808 :
809 6 : SRF_RETURN_NEXT(funcctx, CStringGetTextDatum(channel));
810 : }
811 :
812 12 : SRF_RETURN_DONE(funcctx);
813 : }
814 :
815 : /*
816 : * Async_UnlistenOnExit
817 : *
818 : * This is executed at backend exit if we have done any LISTENs in this
819 : * backend. It might not be necessary anymore, if the user UNLISTENed
820 : * everything, but we don't try to detect that case.
821 : */
822 : static void
823 28 : Async_UnlistenOnExit(int code, Datum arg)
824 : {
825 28 : Exec_UnlistenAllCommit();
826 28 : asyncQueueUnregister();
827 28 : }
828 :
829 : /*
830 : * AtPrepare_Notify
831 : *
832 : * This is called at the prepare phase of a two-phase
833 : * transaction. Save the state for possible commit later.
834 : */
835 : void
836 568 : AtPrepare_Notify(void)
837 : {
838 : /* It's not allowed to have any pending LISTEN/UNLISTEN/NOTIFY actions */
839 568 : if (pendingActions || pendingNotifies)
840 0 : ereport(ERROR,
841 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
842 : errmsg("cannot PREPARE a transaction that has executed LISTEN, UNLISTEN, or NOTIFY")));
843 568 : }
844 :
845 : /*
846 : * PreCommit_Notify
847 : *
848 : * This is called at transaction commit, before actually committing to
849 : * clog.
850 : *
851 : * If there are pending LISTEN actions, make sure we are listed in the
852 : * shared-memory listener array. This must happen before commit to
853 : * ensure we don't miss any notifies from transactions that commit
854 : * just after ours.
855 : *
856 : * If there are outbound notify requests in the pendingNotifies list,
857 : * add them to the global queue. We do that before commit so that
858 : * we can still throw error if we run out of queue space.
859 : */
860 : void
861 813106 : PreCommit_Notify(void)
862 : {
863 : ListCell *p;
864 :
865 813106 : if (!pendingActions && !pendingNotifies)
866 812912 : return; /* no relevant statements in this xact */
867 :
868 194 : if (Trace_notify)
869 0 : elog(DEBUG1, "PreCommit_Notify");
870 :
871 : /* Preflight for any pending listen/unlisten actions */
872 194 : if (pendingActions != NULL)
873 : {
874 210 : foreach(p, pendingActions->actions)
875 : {
876 112 : ListenAction *actrec = (ListenAction *) lfirst(p);
877 :
878 112 : switch (actrec->action)
879 : {
880 74 : case LISTEN_LISTEN:
881 74 : Exec_ListenPreCommit();
882 74 : break;
883 6 : case LISTEN_UNLISTEN:
884 : /* there is no Exec_UnlistenPreCommit() */
885 6 : break;
886 32 : case LISTEN_UNLISTEN_ALL:
887 : /* there is no Exec_UnlistenAllPreCommit() */
888 32 : break;
889 : }
890 112 : }
891 : }
892 :
893 : /* Queue any pending notifies (must happen after the above) */
894 194 : if (pendingNotifies)
895 : {
896 : ListCell *nextNotify;
897 :
898 : /*
899 : * Make sure that we have an XID assigned to the current transaction.
900 : * GetCurrentTransactionId is cheap if we already have an XID, but not
901 : * so cheap if we don't, and we'd prefer not to do that work while
902 : * holding NotifyQueueLock.
903 : */
904 96 : (void) GetCurrentTransactionId();
905 :
906 : /*
907 : * Serialize writers by acquiring a special lock that we hold till
908 : * after commit. This ensures that queue entries appear in commit
909 : * order, and in particular that there are never uncommitted queue
910 : * entries ahead of committed ones, so an uncommitted transaction
911 : * can't block delivery of deliverable notifications.
912 : *
913 : * We use a heavyweight lock so that it'll automatically be released
914 : * after either commit or abort. This also allows deadlocks to be
915 : * detected, though really a deadlock shouldn't be possible here.
916 : *
917 : * The lock is on "database 0", which is pretty ugly but it doesn't
918 : * seem worth inventing a special locktag category just for this.
919 : * (Historical note: before PG 9.0, a similar lock on "database 0" was
920 : * used by the flatfiles mechanism.)
921 : */
922 96 : LockSharedObject(DatabaseRelationId, InvalidOid, 0,
923 : AccessExclusiveLock);
924 :
925 : /* Now push the notifications into the queue */
926 96 : nextNotify = list_head(pendingNotifies->events);
927 262 : while (nextNotify != NULL)
928 : {
929 : /*
930 : * Add the pending notifications to the queue. We acquire and
931 : * release NotifyQueueLock once per page, which might be overkill
932 : * but it does allow readers to get in while we're doing this.
933 : *
934 : * A full queue is very uncommon and should really not happen,
935 : * given that we have so much space available in the SLRU pages.
936 : * Nevertheless we need to deal with this possibility. Note that
937 : * when we get here we are in the process of committing our
938 : * transaction, but we have not yet committed to clog, so at this
939 : * point in time we can still roll the transaction back.
940 : */
941 166 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
942 166 : asyncQueueFillWarning();
943 166 : if (asyncQueueIsFull())
944 0 : ereport(ERROR,
945 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
946 : errmsg("too many notifications in the NOTIFY queue")));
947 166 : nextNotify = asyncQueueAddEntries(nextNotify);
948 166 : LWLockRelease(NotifyQueueLock);
949 : }
950 :
951 : /* Note that we don't clear pendingNotifies; AtCommit_Notify will. */
952 : }
953 : }
954 :
955 : /*
956 : * AtCommit_Notify
957 : *
958 : * This is called at transaction commit, after committing to clog.
959 : *
960 : * Update listenChannels and clear transaction-local state.
961 : *
962 : * If we issued any notifications in the transaction, send signals to
963 : * listening backends (possibly including ourselves) to process them.
964 : * Also, if we filled enough queue pages with new notifies, try to
965 : * advance the queue tail pointer.
966 : */
967 : void
968 812796 : AtCommit_Notify(void)
969 : {
970 : ListCell *p;
971 :
972 : /*
973 : * Allow transactions that have not executed LISTEN/UNLISTEN/NOTIFY to
974 : * return as soon as possible
975 : */
976 812796 : if (!pendingActions && !pendingNotifies)
977 812602 : return;
978 :
979 194 : if (Trace_notify)
980 0 : elog(DEBUG1, "AtCommit_Notify");
981 :
982 : /* Perform any pending listen/unlisten actions */
983 194 : if (pendingActions != NULL)
984 : {
985 210 : foreach(p, pendingActions->actions)
986 : {
987 112 : ListenAction *actrec = (ListenAction *) lfirst(p);
988 :
989 112 : switch (actrec->action)
990 : {
991 74 : case LISTEN_LISTEN:
992 74 : Exec_ListenCommit(actrec->channel);
993 74 : break;
994 6 : case LISTEN_UNLISTEN:
995 6 : Exec_UnlistenCommit(actrec->channel);
996 6 : break;
997 32 : case LISTEN_UNLISTEN_ALL:
998 32 : Exec_UnlistenAllCommit();
999 32 : break;
1000 : }
1001 112 : }
1002 : }
1003 :
1004 : /* If no longer listening to anything, get out of listener array */
1005 194 : if (amRegisteredListener && listenChannels == NIL)
1006 26 : asyncQueueUnregister();
1007 :
1008 : /*
1009 : * Send signals to listening backends. We need do this only if there are
1010 : * pending notifies, which were previously added to the shared queue by
1011 : * PreCommit_Notify().
1012 : */
1013 194 : if (pendingNotifies != NULL)
1014 96 : SignalBackends();
1015 :
1016 : /*
1017 : * If it's time to try to advance the global tail pointer, do that.
1018 : *
1019 : * (It might seem odd to do this in the sender, when more than likely the
1020 : * listeners won't yet have read the messages we just sent. However,
1021 : * there's less contention if only the sender does it, and there is little
1022 : * need for urgency in advancing the global tail. So this typically will
1023 : * be clearing out messages that were sent some time ago.)
1024 : */
1025 194 : if (tryAdvanceTail)
1026 : {
1027 16 : tryAdvanceTail = false;
1028 16 : asyncQueueAdvanceTail();
1029 : }
1030 :
1031 : /* And clean up */
1032 194 : ClearPendingActionsAndNotifies();
1033 : }
1034 :
1035 : /*
1036 : * Exec_ListenPreCommit --- subroutine for PreCommit_Notify
1037 : *
1038 : * This function must make sure we are ready to catch any incoming messages.
1039 : */
1040 : static void
1041 74 : Exec_ListenPreCommit(void)
1042 : {
1043 : QueuePosition head;
1044 : QueuePosition max;
1045 : ProcNumber prevListener;
1046 :
1047 : /*
1048 : * Nothing to do if we are already listening to something, nor if we
1049 : * already ran this routine in this transaction.
1050 : */
1051 74 : if (amRegisteredListener)
1052 36 : return;
1053 :
1054 38 : if (Trace_notify)
1055 0 : elog(DEBUG1, "Exec_ListenPreCommit(%d)", MyProcPid);
1056 :
1057 : /*
1058 : * Before registering, make sure we will unlisten before dying. (Note:
1059 : * this action does not get undone if we abort later.)
1060 : */
1061 38 : if (!unlistenExitRegistered)
1062 : {
1063 28 : before_shmem_exit(Async_UnlistenOnExit, 0);
1064 28 : unlistenExitRegistered = true;
1065 : }
1066 :
1067 : /*
1068 : * This is our first LISTEN, so establish our pointer.
1069 : *
1070 : * We set our pointer to the global tail pointer and then move it forward
1071 : * over already-committed notifications. This ensures we cannot miss any
1072 : * not-yet-committed notifications. We might get a few more but that
1073 : * doesn't hurt.
1074 : *
1075 : * In some scenarios there might be a lot of committed notifications that
1076 : * have not yet been pruned away (because some backend is being lazy about
1077 : * reading them). To reduce our startup time, we can look at other
1078 : * backends and adopt the maximum "pos" pointer of any backend that's in
1079 : * our database; any notifications it's already advanced over are surely
1080 : * committed and need not be re-examined by us. (We must consider only
1081 : * backends connected to our DB, because others will not have bothered to
1082 : * check committed-ness of notifications in our DB.)
1083 : *
1084 : * We need exclusive lock here so we can look at other backends' entries
1085 : * and manipulate the list links.
1086 : */
1087 38 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
1088 38 : head = QUEUE_HEAD;
1089 38 : max = QUEUE_TAIL;
1090 38 : prevListener = INVALID_PROC_NUMBER;
1091 42 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
1092 : {
1093 4 : if (QUEUE_BACKEND_DBOID(i) == MyDatabaseId)
1094 4 : max = QUEUE_POS_MAX(max, QUEUE_BACKEND_POS(i));
1095 : /* Also find last listening backend before this one */
1096 4 : if (i < MyProcNumber)
1097 0 : prevListener = i;
1098 : }
1099 38 : QUEUE_BACKEND_POS(MyProcNumber) = max;
1100 38 : QUEUE_BACKEND_PID(MyProcNumber) = MyProcPid;
1101 38 : QUEUE_BACKEND_DBOID(MyProcNumber) = MyDatabaseId;
1102 : /* Insert backend into list of listeners at correct position */
1103 38 : if (prevListener != INVALID_PROC_NUMBER)
1104 : {
1105 0 : QUEUE_NEXT_LISTENER(MyProcNumber) = QUEUE_NEXT_LISTENER(prevListener);
1106 0 : QUEUE_NEXT_LISTENER(prevListener) = MyProcNumber;
1107 : }
1108 : else
1109 : {
1110 38 : QUEUE_NEXT_LISTENER(MyProcNumber) = QUEUE_FIRST_LISTENER;
1111 38 : QUEUE_FIRST_LISTENER = MyProcNumber;
1112 : }
1113 38 : LWLockRelease(NotifyQueueLock);
1114 :
1115 : /* Now we are listed in the global array, so remember we're listening */
1116 38 : amRegisteredListener = true;
1117 :
1118 : /*
1119 : * Try to move our pointer forward as far as possible. This will skip
1120 : * over already-committed notifications, which we want to do because they
1121 : * might be quite stale. Note that we are not yet listening on anything,
1122 : * so we won't deliver such notifications to our frontend. Also, although
1123 : * our transaction might have executed NOTIFY, those message(s) aren't
1124 : * queued yet so we won't skip them here.
1125 : */
1126 38 : if (!QUEUE_POS_EQUAL(max, head))
1127 22 : asyncQueueReadAllNotifications();
1128 : }
1129 :
1130 : /*
1131 : * Exec_ListenCommit --- subroutine for AtCommit_Notify
1132 : *
1133 : * Add the channel to the list of channels we are listening on.
1134 : */
1135 : static void
1136 74 : Exec_ListenCommit(const char *channel)
1137 : {
1138 : MemoryContext oldcontext;
1139 :
1140 : /* Do nothing if we are already listening on this channel */
1141 74 : if (IsListeningOn(channel))
1142 20 : return;
1143 :
1144 : /*
1145 : * Add the new channel name to listenChannels.
1146 : *
1147 : * XXX It is theoretically possible to get an out-of-memory failure here,
1148 : * which would be bad because we already committed. For the moment it
1149 : * doesn't seem worth trying to guard against that, but maybe improve this
1150 : * later.
1151 : */
1152 54 : oldcontext = MemoryContextSwitchTo(TopMemoryContext);
1153 54 : listenChannels = lappend(listenChannels, pstrdup(channel));
1154 54 : MemoryContextSwitchTo(oldcontext);
1155 : }
1156 :
1157 : /*
1158 : * Exec_UnlistenCommit --- subroutine for AtCommit_Notify
1159 : *
1160 : * Remove the specified channel name from listenChannels.
1161 : */
1162 : static void
1163 6 : Exec_UnlistenCommit(const char *channel)
1164 : {
1165 : ListCell *q;
1166 :
1167 6 : if (Trace_notify)
1168 0 : elog(DEBUG1, "Exec_UnlistenCommit(%s,%d)", channel, MyProcPid);
1169 :
1170 6 : foreach(q, listenChannels)
1171 : {
1172 6 : char *lchan = (char *) lfirst(q);
1173 :
1174 6 : if (strcmp(lchan, channel) == 0)
1175 : {
1176 6 : listenChannels = foreach_delete_current(listenChannels, q);
1177 6 : pfree(lchan);
1178 6 : break;
1179 : }
1180 : }
1181 :
1182 : /*
1183 : * We do not complain about unlistening something not being listened;
1184 : * should we?
1185 : */
1186 6 : }
1187 :
1188 : /*
1189 : * Exec_UnlistenAllCommit --- subroutine for AtCommit_Notify
1190 : *
1191 : * Unlisten on all channels for this backend.
1192 : */
1193 : static void
1194 60 : Exec_UnlistenAllCommit(void)
1195 : {
1196 60 : if (Trace_notify)
1197 0 : elog(DEBUG1, "Exec_UnlistenAllCommit(%d)", MyProcPid);
1198 :
1199 60 : list_free_deep(listenChannels);
1200 60 : listenChannels = NIL;
1201 60 : }
1202 :
1203 : /*
1204 : * Test whether we are actively listening on the given channel name.
1205 : *
1206 : * Note: this function is executed for every notification found in the queue.
1207 : * Perhaps it is worth further optimization, eg convert the list to a sorted
1208 : * array so we can binary-search it. In practice the list is likely to be
1209 : * fairly short, though.
1210 : */
1211 : static bool
1212 2392 : IsListeningOn(const char *channel)
1213 : {
1214 : ListCell *p;
1215 :
1216 2514 : foreach(p, listenChannels)
1217 : {
1218 204 : char *lchan = (char *) lfirst(p);
1219 :
1220 204 : if (strcmp(lchan, channel) == 0)
1221 82 : return true;
1222 : }
1223 2310 : return false;
1224 : }
1225 :
1226 : /*
1227 : * Remove our entry from the listeners array when we are no longer listening
1228 : * on any channel. NB: must not fail if we're already not listening.
1229 : */
1230 : static void
1231 54 : asyncQueueUnregister(void)
1232 : {
1233 : Assert(listenChannels == NIL); /* else caller error */
1234 :
1235 54 : if (!amRegisteredListener) /* nothing to do */
1236 16 : return;
1237 :
1238 : /*
1239 : * Need exclusive lock here to manipulate list links.
1240 : */
1241 38 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
1242 : /* Mark our entry as invalid */
1243 38 : QUEUE_BACKEND_PID(MyProcNumber) = InvalidPid;
1244 38 : QUEUE_BACKEND_DBOID(MyProcNumber) = InvalidOid;
1245 : /* and remove it from the list */
1246 38 : if (QUEUE_FIRST_LISTENER == MyProcNumber)
1247 34 : QUEUE_FIRST_LISTENER = QUEUE_NEXT_LISTENER(MyProcNumber);
1248 : else
1249 : {
1250 4 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
1251 : {
1252 4 : if (QUEUE_NEXT_LISTENER(i) == MyProcNumber)
1253 : {
1254 4 : QUEUE_NEXT_LISTENER(i) = QUEUE_NEXT_LISTENER(MyProcNumber);
1255 4 : break;
1256 : }
1257 : }
1258 : }
1259 38 : QUEUE_NEXT_LISTENER(MyProcNumber) = INVALID_PROC_NUMBER;
1260 38 : LWLockRelease(NotifyQueueLock);
1261 :
1262 : /* mark ourselves as no longer listed in the global array */
1263 38 : amRegisteredListener = false;
1264 : }
1265 :
1266 : /*
1267 : * Test whether there is room to insert more notification messages.
1268 : *
1269 : * Caller must hold at least shared NotifyQueueLock.
1270 : */
1271 : static bool
1272 166 : asyncQueueIsFull(void)
1273 : {
1274 166 : int64 headPage = QUEUE_POS_PAGE(QUEUE_HEAD);
1275 166 : int64 tailPage = QUEUE_POS_PAGE(QUEUE_TAIL);
1276 166 : int64 occupied = headPage - tailPage;
1277 :
1278 166 : return occupied >= max_notify_queue_pages;
1279 : }
1280 :
1281 : /*
1282 : * Advance the QueuePosition to the next entry, assuming that the current
1283 : * entry is of length entryLength. If we jump to a new page the function
1284 : * returns true, else false.
1285 : */
1286 : static bool
1287 4602 : asyncQueueAdvance(volatile QueuePosition *position, int entryLength)
1288 : {
1289 4602 : int64 pageno = QUEUE_POS_PAGE(*position);
1290 4602 : int offset = QUEUE_POS_OFFSET(*position);
1291 4602 : bool pageJump = false;
1292 :
1293 : /*
1294 : * Move to the next writing position: First jump over what we have just
1295 : * written or read.
1296 : */
1297 4602 : offset += entryLength;
1298 : Assert(offset <= QUEUE_PAGESIZE);
1299 :
1300 : /*
1301 : * In a second step check if another entry can possibly be written to the
1302 : * page. If so, stay here, we have reached the next position. If not, then
1303 : * we need to move on to the next page.
1304 : */
1305 4602 : if (offset + QUEUEALIGN(AsyncQueueEntryEmptySize) > QUEUE_PAGESIZE)
1306 : {
1307 140 : pageno++;
1308 140 : offset = 0;
1309 140 : pageJump = true;
1310 : }
1311 :
1312 4602 : SET_QUEUE_POS(*position, pageno, offset);
1313 4602 : return pageJump;
1314 : }
1315 :
1316 : /*
1317 : * Fill the AsyncQueueEntry at *qe with an outbound notification message.
1318 : */
1319 : static void
1320 2218 : asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe)
1321 : {
1322 2218 : size_t channellen = n->channel_len;
1323 2218 : size_t payloadlen = n->payload_len;
1324 : int entryLength;
1325 :
1326 : Assert(channellen < NAMEDATALEN);
1327 : Assert(payloadlen < NOTIFY_PAYLOAD_MAX_LENGTH);
1328 :
1329 : /* The terminators are already included in AsyncQueueEntryEmptySize */
1330 2218 : entryLength = AsyncQueueEntryEmptySize + payloadlen + channellen;
1331 2218 : entryLength = QUEUEALIGN(entryLength);
1332 2218 : qe->length = entryLength;
1333 2218 : qe->dboid = MyDatabaseId;
1334 2218 : qe->xid = GetCurrentTransactionId();
1335 2218 : qe->srcPid = MyProcPid;
1336 2218 : memcpy(qe->data, n->data, channellen + payloadlen + 2);
1337 2218 : }
1338 :
1339 : /*
1340 : * Add pending notifications to the queue.
1341 : *
1342 : * We go page by page here, i.e. we stop once we have to go to a new page but
1343 : * we will be called again and then fill that next page. If an entry does not
1344 : * fit into the current page, we write a dummy entry with an InvalidOid as the
1345 : * database OID in order to fill the page. So every page is always used up to
1346 : * the last byte which simplifies reading the page later.
1347 : *
1348 : * We are passed the list cell (in pendingNotifies->events) containing the next
1349 : * notification to write and return the first still-unwritten cell back.
1350 : * Eventually we will return NULL indicating all is done.
1351 : *
1352 : * We are holding NotifyQueueLock already from the caller and grab
1353 : * page specific SLRU bank lock locally in this function.
1354 : */
1355 : static ListCell *
1356 166 : asyncQueueAddEntries(ListCell *nextNotify)
1357 : {
1358 : AsyncQueueEntry qe;
1359 : QueuePosition queue_head;
1360 : int64 pageno;
1361 : int offset;
1362 : int slotno;
1363 : LWLock *prevlock;
1364 :
1365 : /*
1366 : * We work with a local copy of QUEUE_HEAD, which we write back to shared
1367 : * memory upon exiting. The reason for this is that if we have to advance
1368 : * to a new page, SimpleLruZeroPage might fail (out of disk space, for
1369 : * instance), and we must not advance QUEUE_HEAD if it does. (Otherwise,
1370 : * subsequent insertions would try to put entries into a page that slru.c
1371 : * thinks doesn't exist yet.) So, use a local position variable. Note
1372 : * that if we do fail, any already-inserted queue entries are forgotten;
1373 : * this is okay, since they'd be useless anyway after our transaction
1374 : * rolls back.
1375 : */
1376 166 : queue_head = QUEUE_HEAD;
1377 :
1378 : /*
1379 : * If this is the first write since the postmaster started, we need to
1380 : * initialize the first page of the async SLRU. Otherwise, the current
1381 : * page should be initialized already, so just fetch it.
1382 : */
1383 166 : pageno = QUEUE_POS_PAGE(queue_head);
1384 166 : prevlock = SimpleLruGetBankLock(NotifyCtl, pageno);
1385 :
1386 : /* We hold both NotifyQueueLock and SLRU bank lock during this operation */
1387 166 : LWLockAcquire(prevlock, LW_EXCLUSIVE);
1388 :
1389 166 : if (QUEUE_POS_IS_ZERO(queue_head))
1390 14 : slotno = SimpleLruZeroPage(NotifyCtl, pageno);
1391 : else
1392 152 : slotno = SimpleLruReadPage(NotifyCtl, pageno, true,
1393 : InvalidTransactionId);
1394 :
1395 : /* Note we mark the page dirty before writing in it */
1396 166 : NotifyCtl->shared->page_dirty[slotno] = true;
1397 :
1398 2314 : while (nextNotify != NULL)
1399 : {
1400 2218 : Notification *n = (Notification *) lfirst(nextNotify);
1401 :
1402 : /* Construct a valid queue entry in local variable qe */
1403 2218 : asyncQueueNotificationToEntry(n, &qe);
1404 :
1405 2218 : offset = QUEUE_POS_OFFSET(queue_head);
1406 :
1407 : /* Check whether the entry really fits on the current page */
1408 2218 : if (offset + qe.length <= QUEUE_PAGESIZE)
1409 : {
1410 : /* OK, so advance nextNotify past this item */
1411 2152 : nextNotify = lnext(pendingNotifies->events, nextNotify);
1412 : }
1413 : else
1414 : {
1415 : /*
1416 : * Write a dummy entry to fill up the page. Actually readers will
1417 : * only check dboid and since it won't match any reader's database
1418 : * OID, they will ignore this entry and move on.
1419 : */
1420 66 : qe.length = QUEUE_PAGESIZE - offset;
1421 66 : qe.dboid = InvalidOid;
1422 66 : qe.data[0] = '\0'; /* empty channel */
1423 66 : qe.data[1] = '\0'; /* empty payload */
1424 : }
1425 :
1426 : /* Now copy qe into the shared buffer page */
1427 2218 : memcpy(NotifyCtl->shared->page_buffer[slotno] + offset,
1428 : &qe,
1429 2218 : qe.length);
1430 :
1431 : /* Advance queue_head appropriately, and detect if page is full */
1432 2218 : if (asyncQueueAdvance(&(queue_head), qe.length))
1433 : {
1434 : LWLock *lock;
1435 :
1436 70 : pageno = QUEUE_POS_PAGE(queue_head);
1437 70 : lock = SimpleLruGetBankLock(NotifyCtl, pageno);
1438 70 : if (lock != prevlock)
1439 : {
1440 0 : LWLockRelease(prevlock);
1441 0 : LWLockAcquire(lock, LW_EXCLUSIVE);
1442 0 : prevlock = lock;
1443 : }
1444 :
1445 : /*
1446 : * Page is full, so we're done here, but first fill the next page
1447 : * with zeroes. The reason to do this is to ensure that slru.c's
1448 : * idea of the head page is always the same as ours, which avoids
1449 : * boundary problems in SimpleLruTruncate. The test in
1450 : * asyncQueueIsFull() ensured that there is room to create this
1451 : * page without overrunning the queue.
1452 : */
1453 70 : slotno = SimpleLruZeroPage(NotifyCtl, QUEUE_POS_PAGE(queue_head));
1454 :
1455 : /*
1456 : * If the new page address is a multiple of QUEUE_CLEANUP_DELAY,
1457 : * set flag to remember that we should try to advance the tail
1458 : * pointer (we don't want to actually do that right here).
1459 : */
1460 70 : if (QUEUE_POS_PAGE(queue_head) % QUEUE_CLEANUP_DELAY == 0)
1461 16 : tryAdvanceTail = true;
1462 :
1463 : /* And exit the loop */
1464 70 : break;
1465 : }
1466 : }
1467 :
1468 : /* Success, so update the global QUEUE_HEAD */
1469 166 : QUEUE_HEAD = queue_head;
1470 :
1471 166 : LWLockRelease(prevlock);
1472 :
1473 166 : return nextNotify;
1474 : }
1475 :
1476 : /*
1477 : * SQL function to return the fraction of the notification queue currently
1478 : * occupied.
1479 : */
1480 : Datum
1481 10 : pg_notification_queue_usage(PG_FUNCTION_ARGS)
1482 : {
1483 : double usage;
1484 :
1485 : /* Advance the queue tail so we don't report a too-large result */
1486 10 : asyncQueueAdvanceTail();
1487 :
1488 10 : LWLockAcquire(NotifyQueueLock, LW_SHARED);
1489 10 : usage = asyncQueueUsage();
1490 10 : LWLockRelease(NotifyQueueLock);
1491 :
1492 10 : PG_RETURN_FLOAT8(usage);
1493 : }
1494 :
1495 : /*
1496 : * Return the fraction of the queue that is currently occupied.
1497 : *
1498 : * The caller must hold NotifyQueueLock in (at least) shared mode.
1499 : *
1500 : * Note: we measure the distance to the logical tail page, not the physical
1501 : * tail page. In some sense that's wrong, but the relative position of the
1502 : * physical tail is affected by details such as SLRU segment boundaries,
1503 : * so that a result based on that is unpleasantly unstable.
1504 : */
1505 : static double
1506 176 : asyncQueueUsage(void)
1507 : {
1508 176 : int64 headPage = QUEUE_POS_PAGE(QUEUE_HEAD);
1509 176 : int64 tailPage = QUEUE_POS_PAGE(QUEUE_TAIL);
1510 176 : int64 occupied = headPage - tailPage;
1511 :
1512 176 : if (occupied == 0)
1513 82 : return (double) 0; /* fast exit for common case */
1514 :
1515 94 : return (double) occupied / (double) max_notify_queue_pages;
1516 : }
1517 :
1518 : /*
1519 : * Check whether the queue is at least half full, and emit a warning if so.
1520 : *
1521 : * This is unlikely given the size of the queue, but possible.
1522 : * The warnings show up at most once every QUEUE_FULL_WARN_INTERVAL.
1523 : *
1524 : * Caller must hold exclusive NotifyQueueLock.
1525 : */
1526 : static void
1527 166 : asyncQueueFillWarning(void)
1528 : {
1529 : double fillDegree;
1530 : TimestampTz t;
1531 :
1532 166 : fillDegree = asyncQueueUsage();
1533 166 : if (fillDegree < 0.5)
1534 166 : return;
1535 :
1536 0 : t = GetCurrentTimestamp();
1537 :
1538 0 : if (TimestampDifferenceExceeds(asyncQueueControl->lastQueueFillWarn,
1539 : t, QUEUE_FULL_WARN_INTERVAL))
1540 : {
1541 0 : QueuePosition min = QUEUE_HEAD;
1542 0 : int32 minPid = InvalidPid;
1543 :
1544 0 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
1545 : {
1546 : Assert(QUEUE_BACKEND_PID(i) != InvalidPid);
1547 0 : min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
1548 0 : if (QUEUE_POS_EQUAL(min, QUEUE_BACKEND_POS(i)))
1549 0 : minPid = QUEUE_BACKEND_PID(i);
1550 : }
1551 :
1552 0 : ereport(WARNING,
1553 : (errmsg("NOTIFY queue is %.0f%% full", fillDegree * 100),
1554 : (minPid != InvalidPid ?
1555 : errdetail("The server process with PID %d is among those with the oldest transactions.", minPid)
1556 : : 0),
1557 : (minPid != InvalidPid ?
1558 : errhint("The NOTIFY queue cannot be emptied until that process ends its current transaction.")
1559 : : 0)));
1560 :
1561 0 : asyncQueueControl->lastQueueFillWarn = t;
1562 : }
1563 : }
1564 :
1565 : /*
1566 : * Send signals to listening backends.
1567 : *
1568 : * Normally we signal only backends in our own database, since only those
1569 : * backends could be interested in notifies we send. However, if there's
1570 : * notify traffic in our database but no traffic in another database that
1571 : * does have listener(s), those listeners will fall further and further
1572 : * behind. Waken them anyway if they're far enough behind, so that they'll
1573 : * advance their queue position pointers, allowing the global tail to advance.
1574 : *
1575 : * Since we know the ProcNumber and the Pid the signaling is quite cheap.
1576 : *
1577 : * This is called during CommitTransaction(), so it's important for it
1578 : * to have very low probability of failure.
1579 : */
1580 : static void
1581 96 : SignalBackends(void)
1582 : {
1583 : int32 *pids;
1584 : ProcNumber *procnos;
1585 : int count;
1586 :
1587 : /*
1588 : * Identify backends that we need to signal. We don't want to send
1589 : * signals while holding the NotifyQueueLock, so this loop just builds a
1590 : * list of target PIDs.
1591 : *
1592 : * XXX in principle these pallocs could fail, which would be bad. Maybe
1593 : * preallocate the arrays? They're not that large, though.
1594 : */
1595 96 : pids = (int32 *) palloc(MaxBackends * sizeof(int32));
1596 96 : procnos = (ProcNumber *) palloc(MaxBackends * sizeof(ProcNumber));
1597 96 : count = 0;
1598 :
1599 96 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
1600 176 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
1601 : {
1602 80 : int32 pid = QUEUE_BACKEND_PID(i);
1603 : QueuePosition pos;
1604 :
1605 : Assert(pid != InvalidPid);
1606 80 : pos = QUEUE_BACKEND_POS(i);
1607 80 : if (QUEUE_BACKEND_DBOID(i) == MyDatabaseId)
1608 : {
1609 : /*
1610 : * Always signal listeners in our own database, unless they're
1611 : * already caught up (unlikely, but possible).
1612 : */
1613 80 : if (QUEUE_POS_EQUAL(pos, QUEUE_HEAD))
1614 0 : continue;
1615 : }
1616 : else
1617 : {
1618 : /*
1619 : * Listeners in other databases should be signaled only if they
1620 : * are far behind.
1621 : */
1622 0 : if (asyncQueuePageDiff(QUEUE_POS_PAGE(QUEUE_HEAD),
1623 : QUEUE_POS_PAGE(pos)) < QUEUE_CLEANUP_DELAY)
1624 0 : continue;
1625 : }
1626 : /* OK, need to signal this one */
1627 80 : pids[count] = pid;
1628 80 : procnos[count] = i;
1629 80 : count++;
1630 : }
1631 96 : LWLockRelease(NotifyQueueLock);
1632 :
1633 : /* Now send signals */
1634 176 : for (int i = 0; i < count; i++)
1635 : {
1636 80 : int32 pid = pids[i];
1637 :
1638 : /*
1639 : * If we are signaling our own process, no need to involve the kernel;
1640 : * just set the flag directly.
1641 : */
1642 80 : if (pid == MyProcPid)
1643 : {
1644 40 : notifyInterruptPending = true;
1645 40 : continue;
1646 : }
1647 :
1648 : /*
1649 : * Note: assuming things aren't broken, a signal failure here could
1650 : * only occur if the target backend exited since we released
1651 : * NotifyQueueLock; which is unlikely but certainly possible. So we
1652 : * just log a low-level debug message if it happens.
1653 : */
1654 40 : if (SendProcSignal(pid, PROCSIG_NOTIFY_INTERRUPT, procnos[i]) < 0)
1655 0 : elog(DEBUG3, "could not signal backend with PID %d: %m", pid);
1656 : }
1657 :
1658 96 : pfree(pids);
1659 96 : pfree(procnos);
1660 96 : }
1661 :
1662 : /*
1663 : * AtAbort_Notify
1664 : *
1665 : * This is called at transaction abort.
1666 : *
1667 : * Gets rid of pending actions and outbound notifies that we would have
1668 : * executed if the transaction got committed.
1669 : */
1670 : void
1671 48974 : AtAbort_Notify(void)
1672 : {
1673 : /*
1674 : * If we LISTEN but then roll back the transaction after PreCommit_Notify,
1675 : * we have registered as a listener but have not made any entry in
1676 : * listenChannels. In that case, deregister again.
1677 : */
1678 48974 : if (amRegisteredListener && listenChannels == NIL)
1679 0 : asyncQueueUnregister();
1680 :
1681 : /* And clean up */
1682 48974 : ClearPendingActionsAndNotifies();
1683 48974 : }
1684 :
1685 : /*
1686 : * AtSubCommit_Notify() --- Take care of subtransaction commit.
1687 : *
1688 : * Reassign all items in the pending lists to the parent transaction.
1689 : */
1690 : void
1691 10672 : AtSubCommit_Notify(void)
1692 : {
1693 10672 : int my_level = GetCurrentTransactionNestLevel();
1694 :
1695 : /* If there are actions at our nesting level, we must reparent them. */
1696 10672 : if (pendingActions != NULL &&
1697 0 : pendingActions->nestingLevel >= my_level)
1698 : {
1699 0 : if (pendingActions->upper == NULL ||
1700 0 : pendingActions->upper->nestingLevel < my_level - 1)
1701 : {
1702 : /* nothing to merge; give the whole thing to the parent */
1703 0 : --pendingActions->nestingLevel;
1704 : }
1705 : else
1706 : {
1707 0 : ActionList *childPendingActions = pendingActions;
1708 :
1709 0 : pendingActions = pendingActions->upper;
1710 :
1711 : /*
1712 : * Mustn't try to eliminate duplicates here --- see queue_listen()
1713 : */
1714 0 : pendingActions->actions =
1715 0 : list_concat(pendingActions->actions,
1716 0 : childPendingActions->actions);
1717 0 : pfree(childPendingActions);
1718 : }
1719 : }
1720 :
1721 : /* If there are notifies at our nesting level, we must reparent them. */
1722 10672 : if (pendingNotifies != NULL &&
1723 4 : pendingNotifies->nestingLevel >= my_level)
1724 : {
1725 : Assert(pendingNotifies->nestingLevel == my_level);
1726 :
1727 2 : if (pendingNotifies->upper == NULL ||
1728 2 : pendingNotifies->upper->nestingLevel < my_level - 1)
1729 : {
1730 : /* nothing to merge; give the whole thing to the parent */
1731 0 : --pendingNotifies->nestingLevel;
1732 : }
1733 : else
1734 : {
1735 : /*
1736 : * Formerly, we didn't bother to eliminate duplicates here, but
1737 : * now we must, else we fall foul of "Assert(!found)", either here
1738 : * or during a later attempt to build the parent-level hashtable.
1739 : */
1740 2 : NotificationList *childPendingNotifies = pendingNotifies;
1741 : ListCell *l;
1742 :
1743 2 : pendingNotifies = pendingNotifies->upper;
1744 : /* Insert all the subxact's events into parent, except for dups */
1745 10 : foreach(l, childPendingNotifies->events)
1746 : {
1747 8 : Notification *childn = (Notification *) lfirst(l);
1748 :
1749 8 : if (!AsyncExistsPendingNotify(childn))
1750 4 : AddEventToPendingNotifies(childn);
1751 : }
1752 2 : pfree(childPendingNotifies);
1753 : }
1754 : }
1755 10672 : }
1756 :
1757 : /*
1758 : * AtSubAbort_Notify() --- Take care of subtransaction abort.
1759 : */
1760 : void
1761 9362 : AtSubAbort_Notify(void)
1762 : {
1763 9362 : int my_level = GetCurrentTransactionNestLevel();
1764 :
1765 : /*
1766 : * All we have to do is pop the stack --- the actions/notifies made in
1767 : * this subxact are no longer interesting, and the space will be freed
1768 : * when CurTransactionContext is recycled. We still have to free the
1769 : * ActionList and NotificationList objects themselves, though, because
1770 : * those are allocated in TopTransactionContext.
1771 : *
1772 : * Note that there might be no entries at all, or no entries for the
1773 : * current subtransaction level, either because none were ever created, or
1774 : * because we reentered this routine due to trouble during subxact abort.
1775 : */
1776 9362 : while (pendingActions != NULL &&
1777 0 : pendingActions->nestingLevel >= my_level)
1778 : {
1779 0 : ActionList *childPendingActions = pendingActions;
1780 :
1781 0 : pendingActions = pendingActions->upper;
1782 0 : pfree(childPendingActions);
1783 : }
1784 :
1785 9364 : while (pendingNotifies != NULL &&
1786 4 : pendingNotifies->nestingLevel >= my_level)
1787 : {
1788 2 : NotificationList *childPendingNotifies = pendingNotifies;
1789 :
1790 2 : pendingNotifies = pendingNotifies->upper;
1791 2 : pfree(childPendingNotifies);
1792 : }
1793 9362 : }
1794 :
1795 : /*
1796 : * HandleNotifyInterrupt
1797 : *
1798 : * Signal handler portion of interrupt handling. Let the backend know
1799 : * that there's a pending notify interrupt. If we're currently reading
1800 : * from the client, this will interrupt the read and
1801 : * ProcessClientReadInterrupt() will call ProcessNotifyInterrupt().
1802 : */
1803 : void
1804 40 : HandleNotifyInterrupt(void)
1805 : {
1806 : /*
1807 : * Note: this is called by a SIGNAL HANDLER. You must be very wary what
1808 : * you do here.
1809 : */
1810 :
1811 : /* signal that work needs to be done */
1812 40 : notifyInterruptPending = true;
1813 :
1814 : /* make sure the event is processed in due course */
1815 40 : SetLatch(MyLatch);
1816 40 : }
1817 :
1818 : /*
1819 : * ProcessNotifyInterrupt
1820 : *
1821 : * This is called if we see notifyInterruptPending set, just before
1822 : * transmitting ReadyForQuery at the end of a frontend command, and
1823 : * also if a notify signal occurs while reading from the frontend.
1824 : * HandleNotifyInterrupt() will cause the read to be interrupted
1825 : * via the process's latch, and this routine will get called.
1826 : * If we are truly idle (ie, *not* inside a transaction block),
1827 : * process the incoming notifies.
1828 : *
1829 : * If "flush" is true, force any frontend messages out immediately.
1830 : * This can be false when being called at the end of a frontend command,
1831 : * since we'll flush after sending ReadyForQuery.
1832 : */
1833 : void
1834 188 : ProcessNotifyInterrupt(bool flush)
1835 : {
1836 188 : if (IsTransactionOrTransactionBlock())
1837 110 : return; /* not really idle */
1838 :
1839 : /* Loop in case another signal arrives while sending messages */
1840 156 : while (notifyInterruptPending)
1841 78 : ProcessIncomingNotify(flush);
1842 : }
1843 :
1844 :
1845 : /*
1846 : * Read all pending notifications from the queue, and deliver appropriate
1847 : * ones to my frontend. Stop when we reach queue head or an uncommitted
1848 : * notification.
1849 : */
1850 : static void
1851 100 : asyncQueueReadAllNotifications(void)
1852 : {
1853 : volatile QueuePosition pos;
1854 : QueuePosition head;
1855 : Snapshot snapshot;
1856 :
1857 : /* page_buffer must be adequately aligned, so use a union */
1858 : union
1859 : {
1860 : char buf[QUEUE_PAGESIZE];
1861 : AsyncQueueEntry align;
1862 : } page_buffer;
1863 :
1864 : /* Fetch current state */
1865 100 : LWLockAcquire(NotifyQueueLock, LW_SHARED);
1866 : /* Assert checks that we have a valid state entry */
1867 : Assert(MyProcPid == QUEUE_BACKEND_PID(MyProcNumber));
1868 100 : pos = QUEUE_BACKEND_POS(MyProcNumber);
1869 100 : head = QUEUE_HEAD;
1870 100 : LWLockRelease(NotifyQueueLock);
1871 :
1872 100 : if (QUEUE_POS_EQUAL(pos, head))
1873 : {
1874 : /* Nothing to do, we have read all notifications already. */
1875 0 : return;
1876 : }
1877 :
1878 : /*----------
1879 : * Get snapshot we'll use to decide which xacts are still in progress.
1880 : * This is trickier than it might seem, because of race conditions.
1881 : * Consider the following example:
1882 : *
1883 : * Backend 1: Backend 2:
1884 : *
1885 : * transaction starts
1886 : * UPDATE foo SET ...;
1887 : * NOTIFY foo;
1888 : * commit starts
1889 : * queue the notify message
1890 : * transaction starts
1891 : * LISTEN foo; -- first LISTEN in session
1892 : * SELECT * FROM foo WHERE ...;
1893 : * commit to clog
1894 : * commit starts
1895 : * add backend 2 to array of listeners
1896 : * advance to queue head (this code)
1897 : * commit to clog
1898 : *
1899 : * Transaction 2's SELECT has not seen the UPDATE's effects, since that
1900 : * wasn't committed yet. Ideally we'd ensure that client 2 would
1901 : * eventually get transaction 1's notify message, but there's no way
1902 : * to do that; until we're in the listener array, there's no guarantee
1903 : * that the notify message doesn't get removed from the queue.
1904 : *
1905 : * Therefore the coding technique transaction 2 is using is unsafe:
1906 : * applications must commit a LISTEN before inspecting database state,
1907 : * if they want to ensure they will see notifications about subsequent
1908 : * changes to that state.
1909 : *
1910 : * What we do guarantee is that we'll see all notifications from
1911 : * transactions committing after the snapshot we take here.
1912 : * Exec_ListenPreCommit has already added us to the listener array,
1913 : * so no not-yet-committed messages can be removed from the queue
1914 : * before we see them.
1915 : *----------
1916 : */
1917 100 : snapshot = RegisterSnapshot(GetLatestSnapshot());
1918 :
1919 : /*
1920 : * It is possible that we fail while trying to send a message to our
1921 : * frontend (for example, because of encoding conversion failure). If
1922 : * that happens it is critical that we not try to send the same message
1923 : * over and over again. Therefore, we place a PG_TRY block here that will
1924 : * forcibly advance our queue position before we lose control to an error.
1925 : * (We could alternatively retake NotifyQueueLock and move the position
1926 : * before handling each individual message, but that seems like too much
1927 : * lock traffic.)
1928 : */
1929 100 : PG_TRY();
1930 : {
1931 : bool reachedStop;
1932 :
1933 : do
1934 : {
1935 170 : int64 curpage = QUEUE_POS_PAGE(pos);
1936 170 : int curoffset = QUEUE_POS_OFFSET(pos);
1937 : int slotno;
1938 : int copysize;
1939 :
1940 : /*
1941 : * We copy the data from SLRU into a local buffer, so as to avoid
1942 : * holding the SLRU lock while we are examining the entries and
1943 : * possibly transmitting them to our frontend. Copy only the part
1944 : * of the page we will actually inspect.
1945 : */
1946 170 : slotno = SimpleLruReadPage_ReadOnly(NotifyCtl, curpage,
1947 : InvalidTransactionId);
1948 170 : if (curpage == QUEUE_POS_PAGE(head))
1949 : {
1950 : /* we only want to read as far as head */
1951 100 : copysize = QUEUE_POS_OFFSET(head) - curoffset;
1952 100 : if (copysize < 0)
1953 0 : copysize = 0; /* just for safety */
1954 : }
1955 : else
1956 : {
1957 : /* fetch all the rest of the page */
1958 70 : copysize = QUEUE_PAGESIZE - curoffset;
1959 : }
1960 170 : memcpy(page_buffer.buf + curoffset,
1961 170 : NotifyCtl->shared->page_buffer[slotno] + curoffset,
1962 : copysize);
1963 : /* Release lock that we got from SimpleLruReadPage_ReadOnly() */
1964 170 : LWLockRelease(SimpleLruGetBankLock(NotifyCtl, curpage));
1965 :
1966 : /*
1967 : * Process messages up to the stop position, end of page, or an
1968 : * uncommitted message.
1969 : *
1970 : * Our stop position is what we found to be the head's position
1971 : * when we entered this function. It might have changed already.
1972 : * But if it has, we will receive (or have already received and
1973 : * queued) another signal and come here again.
1974 : *
1975 : * We are not holding NotifyQueueLock here! The queue can only
1976 : * extend beyond the head pointer (see above) and we leave our
1977 : * backend's pointer where it is so nobody will truncate or
1978 : * rewrite pages under us. Especially we don't want to hold a lock
1979 : * while sending the notifications to the frontend.
1980 : */
1981 170 : reachedStop = asyncQueueProcessPageEntries(&pos, head,
1982 : page_buffer.buf,
1983 : snapshot);
1984 170 : } while (!reachedStop);
1985 : }
1986 0 : PG_FINALLY();
1987 : {
1988 : /* Update shared state */
1989 100 : LWLockAcquire(NotifyQueueLock, LW_SHARED);
1990 100 : QUEUE_BACKEND_POS(MyProcNumber) = pos;
1991 100 : LWLockRelease(NotifyQueueLock);
1992 : }
1993 100 : PG_END_TRY();
1994 :
1995 : /* Done with snapshot */
1996 100 : UnregisterSnapshot(snapshot);
1997 : }
1998 :
1999 : /*
2000 : * Fetch notifications from the shared queue, beginning at position current,
2001 : * and deliver relevant ones to my frontend.
2002 : *
2003 : * The current page must have been fetched into page_buffer from shared
2004 : * memory. (We could access the page right in shared memory, but that
2005 : * would imply holding the SLRU bank lock throughout this routine.)
2006 : *
2007 : * We stop if we reach the "stop" position, or reach a notification from an
2008 : * uncommitted transaction, or reach the end of the page.
2009 : *
2010 : * The function returns true once we have reached the stop position or an
2011 : * uncommitted notification, and false if we have finished with the page.
2012 : * In other words: once it returns true there is no need to look further.
2013 : * The QueuePosition *current is advanced past all processed messages.
2014 : */
2015 : static bool
2016 170 : asyncQueueProcessPageEntries(volatile QueuePosition *current,
2017 : QueuePosition stop,
2018 : char *page_buffer,
2019 : Snapshot snapshot)
2020 : {
2021 170 : bool reachedStop = false;
2022 : bool reachedEndOfPage;
2023 : AsyncQueueEntry *qe;
2024 :
2025 : do
2026 : {
2027 2484 : QueuePosition thisentry = *current;
2028 :
2029 2484 : if (QUEUE_POS_EQUAL(thisentry, stop))
2030 100 : break;
2031 :
2032 2384 : qe = (AsyncQueueEntry *) (page_buffer + QUEUE_POS_OFFSET(thisentry));
2033 :
2034 : /*
2035 : * Advance *current over this message, possibly to the next page. As
2036 : * noted in the comments for asyncQueueReadAllNotifications, we must
2037 : * do this before possibly failing while processing the message.
2038 : */
2039 2384 : reachedEndOfPage = asyncQueueAdvance(current, qe->length);
2040 :
2041 : /* Ignore messages destined for other databases */
2042 2384 : if (qe->dboid == MyDatabaseId)
2043 : {
2044 2318 : if (XidInMVCCSnapshot(qe->xid, snapshot))
2045 : {
2046 : /*
2047 : * The source transaction is still in progress, so we can't
2048 : * process this message yet. Break out of the loop, but first
2049 : * back up *current so we will reprocess the message next
2050 : * time. (Note: it is unlikely but not impossible for
2051 : * TransactionIdDidCommit to fail, so we can't really avoid
2052 : * this advance-then-back-up behavior when dealing with an
2053 : * uncommitted message.)
2054 : *
2055 : * Note that we must test XidInMVCCSnapshot before we test
2056 : * TransactionIdDidCommit, else we might return a message from
2057 : * a transaction that is not yet visible to snapshots; compare
2058 : * the comments at the head of heapam_visibility.c.
2059 : *
2060 : * Also, while our own xact won't be listed in the snapshot,
2061 : * we need not check for TransactionIdIsCurrentTransactionId
2062 : * because our transaction cannot (yet) have queued any
2063 : * messages.
2064 : */
2065 0 : *current = thisentry;
2066 0 : reachedStop = true;
2067 0 : break;
2068 : }
2069 2318 : else if (TransactionIdDidCommit(qe->xid))
2070 : {
2071 : /* qe->data is the null-terminated channel name */
2072 2318 : char *channel = qe->data;
2073 :
2074 2318 : if (IsListeningOn(channel))
2075 : {
2076 : /* payload follows channel name */
2077 62 : char *payload = qe->data + strlen(channel) + 1;
2078 :
2079 62 : NotifyMyFrontEnd(channel, payload, qe->srcPid);
2080 : }
2081 : }
2082 : else
2083 : {
2084 : /*
2085 : * The source transaction aborted or crashed, so we just
2086 : * ignore its notifications.
2087 : */
2088 : }
2089 : }
2090 :
2091 : /* Loop back if we're not at end of page */
2092 2384 : } while (!reachedEndOfPage);
2093 :
2094 170 : if (QUEUE_POS_EQUAL(*current, stop))
2095 100 : reachedStop = true;
2096 :
2097 170 : return reachedStop;
2098 : }
2099 :
2100 : /*
2101 : * Advance the shared queue tail variable to the minimum of all the
2102 : * per-backend tail pointers. Truncate pg_notify space if possible.
2103 : *
2104 : * This is (usually) called during CommitTransaction(), so it's important for
2105 : * it to have very low probability of failure.
2106 : */
2107 : static void
2108 26 : asyncQueueAdvanceTail(void)
2109 : {
2110 : QueuePosition min;
2111 : int64 oldtailpage;
2112 : int64 newtailpage;
2113 : int64 boundary;
2114 :
2115 : /* Restrict task to one backend per cluster; see SimpleLruTruncate(). */
2116 26 : LWLockAcquire(NotifyQueueTailLock, LW_EXCLUSIVE);
2117 :
2118 : /*
2119 : * Compute the new tail. Pre-v13, it's essential that QUEUE_TAIL be exact
2120 : * (ie, exactly match at least one backend's queue position), so it must
2121 : * be updated atomically with the actual computation. Since v13, we could
2122 : * get away with not doing it like that, but it seems prudent to keep it
2123 : * so.
2124 : *
2125 : * Also, because incoming backends will scan forward from QUEUE_TAIL, that
2126 : * must be advanced before we can truncate any data. Thus, QUEUE_TAIL is
2127 : * the logical tail, while QUEUE_STOP_PAGE is the physical tail, or oldest
2128 : * un-truncated page. When QUEUE_STOP_PAGE != QUEUE_POS_PAGE(QUEUE_TAIL),
2129 : * there are pages we can truncate but haven't yet finished doing so.
2130 : *
2131 : * For concurrency's sake, we don't want to hold NotifyQueueLock while
2132 : * performing SimpleLruTruncate. This is OK because no backend will try
2133 : * to access the pages we are in the midst of truncating.
2134 : */
2135 26 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
2136 26 : min = QUEUE_HEAD;
2137 46 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
2138 : {
2139 : Assert(QUEUE_BACKEND_PID(i) != InvalidPid);
2140 20 : min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
2141 : }
2142 26 : QUEUE_TAIL = min;
2143 26 : oldtailpage = QUEUE_STOP_PAGE;
2144 26 : LWLockRelease(NotifyQueueLock);
2145 :
2146 : /*
2147 : * We can truncate something if the global tail advanced across an SLRU
2148 : * segment boundary.
2149 : *
2150 : * XXX it might be better to truncate only once every several segments, to
2151 : * reduce the number of directory scans.
2152 : */
2153 26 : newtailpage = QUEUE_POS_PAGE(min);
2154 26 : boundary = newtailpage - (newtailpage % SLRU_PAGES_PER_SEGMENT);
2155 26 : if (asyncQueuePagePrecedes(oldtailpage, boundary))
2156 : {
2157 : /*
2158 : * SimpleLruTruncate() will ask for SLRU bank locks but will also
2159 : * release the lock again.
2160 : */
2161 0 : SimpleLruTruncate(NotifyCtl, newtailpage);
2162 :
2163 0 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
2164 0 : QUEUE_STOP_PAGE = newtailpage;
2165 0 : LWLockRelease(NotifyQueueLock);
2166 : }
2167 :
2168 26 : LWLockRelease(NotifyQueueTailLock);
2169 26 : }
2170 :
2171 : /*
2172 : * ProcessIncomingNotify
2173 : *
2174 : * Scan the queue for arriving notifications and report them to the front
2175 : * end. The notifications might be from other sessions, or our own;
2176 : * there's no need to distinguish here.
2177 : *
2178 : * If "flush" is true, force any frontend messages out immediately.
2179 : *
2180 : * NOTE: since we are outside any transaction, we must create our own.
2181 : */
2182 : static void
2183 78 : ProcessIncomingNotify(bool flush)
2184 : {
2185 : /* We *must* reset the flag */
2186 78 : notifyInterruptPending = false;
2187 :
2188 : /* Do nothing else if we aren't actively listening */
2189 78 : if (listenChannels == NIL)
2190 0 : return;
2191 :
2192 78 : if (Trace_notify)
2193 0 : elog(DEBUG1, "ProcessIncomingNotify");
2194 :
2195 78 : set_ps_display("notify interrupt");
2196 :
2197 : /*
2198 : * We must run asyncQueueReadAllNotifications inside a transaction, else
2199 : * bad things happen if it gets an error.
2200 : */
2201 78 : StartTransactionCommand();
2202 :
2203 78 : asyncQueueReadAllNotifications();
2204 :
2205 78 : CommitTransactionCommand();
2206 :
2207 : /*
2208 : * If this isn't an end-of-command case, we must flush the notify messages
2209 : * to ensure frontend gets them promptly.
2210 : */
2211 78 : if (flush)
2212 20 : pq_flush();
2213 :
2214 78 : set_ps_display("idle");
2215 :
2216 78 : if (Trace_notify)
2217 0 : elog(DEBUG1, "ProcessIncomingNotify: done");
2218 : }
2219 :
2220 : /*
2221 : * Send NOTIFY message to my front end.
2222 : */
2223 : void
2224 62 : NotifyMyFrontEnd(const char *channel, const char *payload, int32 srcPid)
2225 : {
2226 62 : if (whereToSendOutput == DestRemote)
2227 : {
2228 : StringInfoData buf;
2229 :
2230 62 : pq_beginmessage(&buf, PqMsg_NotificationResponse);
2231 62 : pq_sendint32(&buf, srcPid);
2232 62 : pq_sendstring(&buf, channel);
2233 62 : pq_sendstring(&buf, payload);
2234 62 : pq_endmessage(&buf);
2235 :
2236 : /*
2237 : * NOTE: we do not do pq_flush() here. Some level of caller will
2238 : * handle it later, allowing this message to be combined into a packet
2239 : * with other ones.
2240 : */
2241 : }
2242 : else
2243 0 : elog(INFO, "NOTIFY for \"%s\" payload \"%s\"", channel, payload);
2244 62 : }
2245 :
2246 : /* Does pendingNotifies include a match for the given event? */
2247 : static bool
2248 2096 : AsyncExistsPendingNotify(Notification *n)
2249 : {
2250 2096 : if (pendingNotifies == NULL)
2251 0 : return false;
2252 :
2253 2096 : if (pendingNotifies->hashtab != NULL)
2254 : {
2255 : /* Use the hash table to probe for a match */
2256 1966 : if (hash_search(pendingNotifies->hashtab,
2257 : &n,
2258 : HASH_FIND,
2259 : NULL))
2260 0 : return true;
2261 : }
2262 : else
2263 : {
2264 : /* Must scan the event list */
2265 : ListCell *l;
2266 :
2267 542 : foreach(l, pendingNotifies->events)
2268 : {
2269 440 : Notification *oldn = (Notification *) lfirst(l);
2270 :
2271 440 : if (n->channel_len == oldn->channel_len &&
2272 440 : n->payload_len == oldn->payload_len &&
2273 250 : memcmp(n->data, oldn->data,
2274 250 : n->channel_len + n->payload_len + 2) == 0)
2275 28 : return true;
2276 : }
2277 : }
2278 :
2279 2068 : return false;
2280 : }
2281 :
2282 : /*
2283 : * Add a notification event to a pre-existing pendingNotifies list.
2284 : *
2285 : * Because pendingNotifies->events is already nonempty, this works
2286 : * correctly no matter what CurrentMemoryContext is.
2287 : */
2288 : static void
2289 2068 : AddEventToPendingNotifies(Notification *n)
2290 : {
2291 : Assert(pendingNotifies->events != NIL);
2292 :
2293 : /* Create the hash table if it's time to */
2294 2068 : if (list_length(pendingNotifies->events) >= MIN_HASHABLE_NOTIFIES &&
2295 1968 : pendingNotifies->hashtab == NULL)
2296 : {
2297 : HASHCTL hash_ctl;
2298 : ListCell *l;
2299 :
2300 : /* Create the hash table */
2301 2 : hash_ctl.keysize = sizeof(Notification *);
2302 2 : hash_ctl.entrysize = sizeof(struct NotificationHash);
2303 2 : hash_ctl.hash = notification_hash;
2304 2 : hash_ctl.match = notification_match;
2305 2 : hash_ctl.hcxt = CurTransactionContext;
2306 4 : pendingNotifies->hashtab =
2307 2 : hash_create("Pending Notifies",
2308 : 256L,
2309 : &hash_ctl,
2310 : HASH_ELEM | HASH_FUNCTION | HASH_COMPARE | HASH_CONTEXT);
2311 :
2312 : /* Insert all the already-existing events */
2313 34 : foreach(l, pendingNotifies->events)
2314 : {
2315 32 : Notification *oldn = (Notification *) lfirst(l);
2316 : bool found;
2317 :
2318 32 : (void) hash_search(pendingNotifies->hashtab,
2319 : &oldn,
2320 : HASH_ENTER,
2321 : &found);
2322 : Assert(!found);
2323 : }
2324 : }
2325 :
2326 : /* Add new event to the list, in order */
2327 2068 : pendingNotifies->events = lappend(pendingNotifies->events, n);
2328 :
2329 : /* Add event to the hash table if needed */
2330 2068 : if (pendingNotifies->hashtab != NULL)
2331 : {
2332 : bool found;
2333 :
2334 1968 : (void) hash_search(pendingNotifies->hashtab,
2335 : &n,
2336 : HASH_ENTER,
2337 : &found);
2338 : Assert(!found);
2339 : }
2340 2068 : }
2341 :
2342 : /*
2343 : * notification_hash: hash function for notification hash table
2344 : *
2345 : * The hash "keys" are pointers to Notification structs.
2346 : */
2347 : static uint32
2348 3966 : notification_hash(const void *key, Size keysize)
2349 : {
2350 3966 : const Notification *k = *(const Notification *const *) key;
2351 :
2352 : Assert(keysize == sizeof(Notification *));
2353 : /* We don't bother to include the payload's trailing null in the hash */
2354 3966 : return DatumGetUInt32(hash_any((const unsigned char *) k->data,
2355 3966 : k->channel_len + k->payload_len + 1));
2356 : }
2357 :
2358 : /*
2359 : * notification_match: match function to use with notification_hash
2360 : */
2361 : static int
2362 0 : notification_match(const void *key1, const void *key2, Size keysize)
2363 : {
2364 0 : const Notification *k1 = *(const Notification *const *) key1;
2365 0 : const Notification *k2 = *(const Notification *const *) key2;
2366 :
2367 : Assert(keysize == sizeof(Notification *));
2368 0 : if (k1->channel_len == k2->channel_len &&
2369 0 : k1->payload_len == k2->payload_len &&
2370 0 : memcmp(k1->data, k2->data,
2371 0 : k1->channel_len + k1->payload_len + 2) == 0)
2372 0 : return 0; /* equal */
2373 0 : return 1; /* not equal */
2374 : }
2375 :
2376 : /* Clear the pendingActions and pendingNotifies lists. */
2377 : static void
2378 49168 : ClearPendingActionsAndNotifies(void)
2379 : {
2380 : /*
2381 : * Everything's allocated in either TopTransactionContext or the context
2382 : * for the subtransaction to which it corresponds. So, there's nothing to
2383 : * do here except reset the pointers; the space will be reclaimed when the
2384 : * contexts are deleted.
2385 : */
2386 49168 : pendingActions = NULL;
2387 49168 : pendingNotifies = NULL;
2388 49168 : }
2389 :
2390 : /*
2391 : * GUC check_hook for notify_buffers
2392 : */
2393 : bool
2394 2168 : check_notify_buffers(int *newval, void **extra, GucSource source)
2395 : {
2396 2168 : return check_slru_buffers("notify_buffers", newval);
2397 : }
|
