Age Owner TLA Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * async.c
4 : * Asynchronous notification: NOTIFY, LISTEN, UNLISTEN
5 : *
6 : * Portions Copyright (c) 1996-2023, 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 avoid SLRU wraparound and limit disk space consumption, the tail
107 : * pointer needs to be advanced so that old pages can be truncated.
108 : * This is relatively expensive (notably, it requires an exclusive lock),
109 : * so we don't want to do it often. We make sending backends do this work
110 : * if they advanced the queue head into a new page, but only once every
111 : * QUEUE_CLEANUP_DELAY pages.
112 : *
113 : * An application that listens on the same channel it notifies will get
114 : * NOTIFY messages for its own NOTIFYs. These can be ignored, if not useful,
115 : * by comparing be_pid in the NOTIFY message to the application's own backend's
116 : * PID. (As of FE/BE protocol 2.0, the backend's PID is provided to the
117 : * frontend during startup.) The above design guarantees that notifies from
118 : * other backends will never be missed by ignoring self-notifies.
119 : *
120 : * The amount of shared memory used for notify management (NUM_NOTIFY_BUFFERS)
121 : * can be varied without affecting anything but performance. The maximum
122 : * amount of notification data that can be queued at one time is determined
123 : * by slru.c's wraparound limit; see QUEUE_MAX_PAGE below.
124 : *-------------------------------------------------------------------------
125 : */
126 :
127 : #include "postgres.h"
128 :
129 : #include <limits.h>
130 : #include <unistd.h>
131 : #include <signal.h>
132 :
133 : #include "access/parallel.h"
134 : #include "access/slru.h"
135 : #include "access/transam.h"
136 : #include "access/xact.h"
137 : #include "catalog/pg_database.h"
138 : #include "commands/async.h"
139 : #include "common/hashfn.h"
140 : #include "funcapi.h"
141 : #include "libpq/libpq.h"
142 : #include "libpq/pqformat.h"
143 : #include "miscadmin.h"
144 : #include "storage/ipc.h"
145 : #include "storage/lmgr.h"
146 : #include "storage/proc.h"
147 : #include "storage/procarray.h"
148 : #include "storage/procsignal.h"
149 : #include "storage/sinval.h"
150 : #include "tcop/tcopprot.h"
151 : #include "utils/builtins.h"
152 : #include "utils/memutils.h"
153 : #include "utils/ps_status.h"
154 : #include "utils/snapmgr.h"
155 : #include "utils/timestamp.h"
156 :
157 :
158 : /*
159 : * Maximum size of a NOTIFY payload, including terminating NULL. This
160 : * must be kept small enough so that a notification message fits on one
161 : * SLRU page. The magic fudge factor here is noncritical as long as it's
162 : * more than AsyncQueueEntryEmptySize --- we make it significantly bigger
163 : * than that, so changes in that data structure won't affect user-visible
164 : * restrictions.
165 : */
166 : #define NOTIFY_PAYLOAD_MAX_LENGTH (BLCKSZ - NAMEDATALEN - 128)
167 :
168 : /*
169 : * Struct representing an entry in the global notify queue
170 : *
171 : * This struct declaration has the maximal length, but in a real queue entry
172 : * the data area is only big enough for the actual channel and payload strings
173 : * (each null-terminated). AsyncQueueEntryEmptySize is the minimum possible
174 : * entry size, if both channel and payload strings are empty (but note it
175 : * doesn't include alignment padding).
176 : *
177 : * The "length" field should always be rounded up to the next QUEUEALIGN
178 : * multiple so that all fields are properly aligned.
179 : */
180 : typedef struct AsyncQueueEntry
181 : {
182 : int length; /* total allocated length of entry */
183 : Oid dboid; /* sender's database OID */
184 : TransactionId xid; /* sender's XID */
185 : int32 srcPid; /* sender's PID */
186 : char data[NAMEDATALEN + NOTIFY_PAYLOAD_MAX_LENGTH];
187 : } AsyncQueueEntry;
188 :
189 : /* Currently, no field of AsyncQueueEntry requires more than int alignment */
190 : #define QUEUEALIGN(len) INTALIGN(len)
191 :
192 : #define AsyncQueueEntryEmptySize (offsetof(AsyncQueueEntry, data) + 2)
193 :
194 : /*
195 : * Struct describing a queue position, and assorted macros for working with it
196 : */
197 : typedef struct QueuePosition
198 : {
199 : int page; /* SLRU page number */
200 : int offset; /* byte offset within page */
201 : } QueuePosition;
202 :
203 : #define QUEUE_POS_PAGE(x) ((x).page)
204 : #define QUEUE_POS_OFFSET(x) ((x).offset)
205 :
206 : #define SET_QUEUE_POS(x,y,z) \
207 : do { \
208 : (x).page = (y); \
209 : (x).offset = (z); \
210 : } while (0)
211 :
212 : #define QUEUE_POS_EQUAL(x,y) \
213 : ((x).page == (y).page && (x).offset == (y).offset)
214 :
215 : #define QUEUE_POS_IS_ZERO(x) \
216 : ((x).page == 0 && (x).offset == 0)
217 :
218 : /* choose logically smaller QueuePosition */
219 : #define QUEUE_POS_MIN(x,y) \
220 : (asyncQueuePagePrecedes((x).page, (y).page) ? (x) : \
221 : (x).page != (y).page ? (y) : \
222 : (x).offset < (y).offset ? (x) : (y))
223 :
224 : /* choose logically larger QueuePosition */
225 : #define QUEUE_POS_MAX(x,y) \
226 : (asyncQueuePagePrecedes((x).page, (y).page) ? (y) : \
227 : (x).page != (y).page ? (x) : \
228 : (x).offset > (y).offset ? (x) : (y))
229 :
230 : /*
231 : * Parameter determining how often we try to advance the tail pointer:
232 : * we do that after every QUEUE_CLEANUP_DELAY pages of NOTIFY data. This is
233 : * also the distance by which a backend in another database needs to be
234 : * behind before we'll decide we need to wake it up to advance its pointer.
235 : *
236 : * Resist the temptation to make this really large. While that would save
237 : * work in some places, it would add cost in others. In particular, this
238 : * should likely be less than NUM_NOTIFY_BUFFERS, to ensure that backends
239 : * catch up before the pages they'll need to read fall out of SLRU cache.
240 : */
241 : #define QUEUE_CLEANUP_DELAY 4
242 :
243 : /*
244 : * Struct describing a listening backend's status
245 : */
246 : typedef struct QueueBackendStatus
247 : {
248 : int32 pid; /* either a PID or InvalidPid */
249 : Oid dboid; /* backend's database OID, or InvalidOid */
250 : BackendId nextListener; /* id of next listener, or InvalidBackendId */
251 : QueuePosition pos; /* backend has read queue up to here */
252 : } QueueBackendStatus;
253 :
254 : /*
255 : * Shared memory state for LISTEN/NOTIFY (excluding its SLRU stuff)
256 : *
257 : * The AsyncQueueControl structure is protected by the NotifyQueueLock and
258 : * NotifyQueueTailLock.
259 : *
260 : * When holding NotifyQueueLock in SHARED mode, backends may only inspect
261 : * their own entries as well as the head and tail pointers. Consequently we
262 : * can allow a backend to update its own record while holding only SHARED lock
263 : * (since no other backend will inspect it).
264 : *
265 : * When holding NotifyQueueLock in EXCLUSIVE mode, backends can inspect the
266 : * entries of other backends and also change the head pointer. When holding
267 : * both NotifyQueueLock and NotifyQueueTailLock in EXCLUSIVE mode, backends
268 : * can change the tail pointers.
269 : *
270 : * NotifySLRULock is used as the control lock for the pg_notify SLRU buffers.
271 : * In order to avoid deadlocks, whenever we need multiple locks, we first get
272 : * NotifyQueueTailLock, then NotifyQueueLock, and lastly NotifySLRULock.
273 : *
274 : * Each backend uses the backend[] array entry with index equal to its
275 : * BackendId (which can range from 1 to MaxBackends). We rely on this to make
276 : * SendProcSignal fast.
277 : *
278 : * The backend[] array entries for actively-listening backends are threaded
279 : * together using firstListener and the nextListener links, so that we can
280 : * scan them without having to iterate over inactive entries. We keep this
281 : * list in order by BackendId so that the scan is cache-friendly when there
282 : * are many active entries.
283 : */
284 : typedef struct AsyncQueueControl
285 : {
286 : QueuePosition head; /* head points to the next free location */
287 : QueuePosition tail; /* tail must be <= the queue position of every
288 : * listening backend */
289 : int stopPage; /* oldest unrecycled page; must be <=
290 : * tail.page */
291 : BackendId firstListener; /* id of first listener, or InvalidBackendId */
292 : TimestampTz lastQueueFillWarn; /* time of last queue-full msg */
293 : QueueBackendStatus backend[FLEXIBLE_ARRAY_MEMBER];
294 : /* backend[0] is not used; used entries are from [1] to [MaxBackends] */
295 : } AsyncQueueControl;
296 :
297 : static AsyncQueueControl *asyncQueueControl;
298 :
299 : #define QUEUE_HEAD (asyncQueueControl->head)
300 : #define QUEUE_TAIL (asyncQueueControl->tail)
301 : #define QUEUE_STOP_PAGE (asyncQueueControl->stopPage)
302 : #define QUEUE_FIRST_LISTENER (asyncQueueControl->firstListener)
303 : #define QUEUE_BACKEND_PID(i) (asyncQueueControl->backend[i].pid)
304 : #define QUEUE_BACKEND_DBOID(i) (asyncQueueControl->backend[i].dboid)
305 : #define QUEUE_NEXT_LISTENER(i) (asyncQueueControl->backend[i].nextListener)
306 : #define QUEUE_BACKEND_POS(i) (asyncQueueControl->backend[i].pos)
307 :
308 : /*
309 : * The SLRU buffer area through which we access the notification queue
310 : */
311 : static SlruCtlData NotifyCtlData;
312 :
313 : #define NotifyCtl (&NotifyCtlData)
314 : #define QUEUE_PAGESIZE BLCKSZ
315 : #define QUEUE_FULL_WARN_INTERVAL 5000 /* warn at most once every 5s */
316 :
317 : /*
318 : * Use segments 0000 through FFFF. Each contains SLRU_PAGES_PER_SEGMENT pages
319 : * which gives us the pages from 0 to SLRU_PAGES_PER_SEGMENT * 0x10000 - 1.
320 : * We could use as many segments as SlruScanDirectory() allows, but this gives
321 : * us so much space already that it doesn't seem worth the trouble.
322 : *
323 : * The most data we can have in the queue at a time is QUEUE_MAX_PAGE/2
324 : * pages, because more than that would confuse slru.c into thinking there
325 : * was a wraparound condition. With the default BLCKSZ this means there
326 : * can be up to 8GB of queued-and-not-read data.
327 : *
328 : * Note: it's possible to redefine QUEUE_MAX_PAGE with a smaller multiple of
329 : * SLRU_PAGES_PER_SEGMENT, for easier testing of queue-full behaviour.
330 : */
331 : #define QUEUE_MAX_PAGE (SLRU_PAGES_PER_SEGMENT * 0x10000 - 1)
332 :
333 : /*
334 : * listenChannels identifies the channels we are actually listening to
335 : * (ie, have committed a LISTEN on). It is a simple list of channel names,
336 : * allocated in TopMemoryContext.
337 : */
338 : static List *listenChannels = NIL; /* list of C strings */
339 :
340 : /*
341 : * State for pending LISTEN/UNLISTEN actions consists of an ordered list of
342 : * all actions requested in the current transaction. As explained above,
343 : * we don't actually change listenChannels until we reach transaction commit.
344 : *
345 : * The list is kept in CurTransactionContext. In subtransactions, each
346 : * subtransaction has its own list in its own CurTransactionContext, but
347 : * successful subtransactions attach their lists to their parent's list.
348 : * Failed subtransactions simply discard their lists.
349 : */
350 : typedef enum
351 : {
352 : LISTEN_LISTEN,
353 : LISTEN_UNLISTEN,
354 : LISTEN_UNLISTEN_ALL
355 : } ListenActionKind;
356 :
357 : typedef struct
358 : {
359 : ListenActionKind action;
360 : char channel[FLEXIBLE_ARRAY_MEMBER]; /* nul-terminated string */
361 : } ListenAction;
362 :
363 : typedef struct ActionList
364 : {
365 : int nestingLevel; /* current transaction nesting depth */
366 : List *actions; /* list of ListenAction structs */
367 : struct ActionList *upper; /* details for upper transaction levels */
368 : } ActionList;
369 :
370 : static ActionList *pendingActions = NULL;
371 :
372 : /*
373 : * State for outbound notifies consists of a list of all channels+payloads
374 : * NOTIFYed in the current transaction. We do not actually perform a NOTIFY
375 : * until and unless the transaction commits. pendingNotifies is NULL if no
376 : * NOTIFYs have been done in the current (sub) transaction.
377 : *
378 : * We discard duplicate notify events issued in the same transaction.
379 : * Hence, in addition to the list proper (which we need to track the order
380 : * of the events, since we guarantee to deliver them in order), we build a
381 : * hash table which we can probe to detect duplicates. Since building the
382 : * hash table is somewhat expensive, we do so only once we have at least
383 : * MIN_HASHABLE_NOTIFIES events queued in the current (sub) transaction;
384 : * before that we just scan the events linearly.
385 : *
386 : * The list is kept in CurTransactionContext. In subtransactions, each
387 : * subtransaction has its own list in its own CurTransactionContext, but
388 : * successful subtransactions add their entries to their parent's list.
389 : * Failed subtransactions simply discard their lists. Since these lists
390 : * are independent, there may be notify events in a subtransaction's list
391 : * that duplicate events in some ancestor (sub) transaction; we get rid of
392 : * the dups when merging the subtransaction's list into its parent's.
393 : *
394 : * Note: the action and notify lists do not interact within a transaction.
395 : * In particular, if a transaction does NOTIFY and then LISTEN on the same
396 : * condition name, it will get a self-notify at commit. This is a bit odd
397 : * but is consistent with our historical behavior.
398 : */
399 : typedef struct Notification
400 : {
401 : uint16 channel_len; /* length of channel-name string */
402 : uint16 payload_len; /* length of payload string */
403 : /* null-terminated channel name, then null-terminated payload follow */
404 : char data[FLEXIBLE_ARRAY_MEMBER];
405 : } Notification;
406 :
407 : typedef struct NotificationList
408 : {
409 : int nestingLevel; /* current transaction nesting depth */
410 : List *events; /* list of Notification structs */
411 : HTAB *hashtab; /* hash of NotificationHash structs, or NULL */
412 : struct NotificationList *upper; /* details for upper transaction levels */
413 : } NotificationList;
414 :
415 : #define MIN_HASHABLE_NOTIFIES 16 /* threshold to build hashtab */
416 :
417 : typedef struct NotificationHash
418 : {
419 : Notification *event; /* => the actual Notification struct */
420 : } NotificationHash;
421 :
422 : static NotificationList *pendingNotifies = NULL;
423 :
424 : /*
425 : * Inbound notifications are initially processed by HandleNotifyInterrupt(),
426 : * called from inside a signal handler. That just sets the
427 : * notifyInterruptPending flag and sets the process
428 : * latch. ProcessNotifyInterrupt() will then be called whenever it's safe to
429 : * actually deal with the interrupt.
430 : */
431 : volatile sig_atomic_t notifyInterruptPending = false;
432 :
433 : /* True if we've registered an on_shmem_exit cleanup */
434 : static bool unlistenExitRegistered = false;
435 :
436 : /* True if we're currently registered as a listener in asyncQueueControl */
437 : static bool amRegisteredListener = false;
438 :
439 : /* have we advanced to a page that's a multiple of QUEUE_CLEANUP_DELAY? */
440 : static bool tryAdvanceTail = false;
441 :
442 : /* GUC parameter */
443 : bool Trace_notify = false;
444 :
445 : /* local function prototypes */
446 : static int asyncQueuePageDiff(int p, int q);
447 : static bool asyncQueuePagePrecedes(int p, int q);
448 : static void queue_listen(ListenActionKind action, const char *channel);
449 : static void Async_UnlistenOnExit(int code, Datum arg);
450 : static void Exec_ListenPreCommit(void);
451 : static void Exec_ListenCommit(const char *channel);
452 : static void Exec_UnlistenCommit(const char *channel);
453 : static void Exec_UnlistenAllCommit(void);
454 : static bool IsListeningOn(const char *channel);
455 : static void asyncQueueUnregister(void);
456 : static bool asyncQueueIsFull(void);
457 : static bool asyncQueueAdvance(volatile QueuePosition *position, int entryLength);
458 : static void asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe);
459 : static ListCell *asyncQueueAddEntries(ListCell *nextNotify);
460 : static double asyncQueueUsage(void);
461 : static void asyncQueueFillWarning(void);
462 : static void SignalBackends(void);
463 : static void asyncQueueReadAllNotifications(void);
464 : static bool asyncQueueProcessPageEntries(volatile QueuePosition *current,
465 : QueuePosition stop,
466 : char *page_buffer,
467 : Snapshot snapshot);
468 : static void asyncQueueAdvanceTail(void);
469 : static void ProcessIncomingNotify(bool flush);
470 : static bool AsyncExistsPendingNotify(Notification *n);
471 : static void AddEventToPendingNotifies(Notification *n);
472 : static uint32 notification_hash(const void *key, Size keysize);
473 : static int notification_match(const void *key1, const void *key2, Size keysize);
474 : static void ClearPendingActionsAndNotifies(void);
475 :
476 : /*
477 : * Compute the difference between two queue page numbers (i.e., p - q),
478 : * accounting for wraparound.
479 : */
480 : static int
1295 tgl 481 CBC 108 : asyncQueuePageDiff(int p, int q)
482 : {
483 : int diff;
484 :
485 : /*
486 : * We have to compare modulo (QUEUE_MAX_PAGE+1)/2. Both inputs should be
487 : * in the range 0..QUEUE_MAX_PAGE.
488 : */
4800 489 108 : Assert(p >= 0 && p <= QUEUE_MAX_PAGE);
490 108 : Assert(q >= 0 && q <= QUEUE_MAX_PAGE);
491 :
492 108 : diff = p - q;
4790 bruce 493 108 : if (diff >= ((QUEUE_MAX_PAGE + 1) / 2))
4790 bruce 494 UBC 0 : diff -= QUEUE_MAX_PAGE + 1;
4790 bruce 495 CBC 108 : else if (diff < -((QUEUE_MAX_PAGE + 1) / 2))
4790 bruce 496 UBC 0 : diff += QUEUE_MAX_PAGE + 1;
1295 tgl 497 CBC 108 : return diff;
498 : }
499 :
500 : /*
501 : * Is p < q, accounting for wraparound?
502 : *
503 : * Since asyncQueueIsFull() blocks creation of a page that could precede any
504 : * extant page, we need not assess entries within a page.
505 : */
506 : static bool
507 108 : asyncQueuePagePrecedes(int p, int q)
508 : {
509 108 : return asyncQueuePageDiff(p, q) < 0;
510 : }
511 :
512 : /*
513 : * Report space needed for our shared memory area
514 : */
515 : Size
4800 516 2738 : AsyncShmemSize(void)
517 : {
518 : Size size;
519 :
520 : /* This had better match AsyncShmemInit */
362 rhaas 521 2738 : size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
2969 tgl 522 2738 : size = add_size(size, offsetof(AsyncQueueControl, backend));
523 :
1059 524 2738 : size = add_size(size, SimpleLruShmemSize(NUM_NOTIFY_BUFFERS, 0));
525 :
4800 526 2738 : return size;
527 : }
528 :
529 : /*
530 : * Initialize our shared memory area
531 : */
532 : void
533 1826 : AsyncShmemInit(void)
534 : {
535 : bool found;
536 : Size size;
537 :
538 : /*
539 : * Create or attach to the AsyncQueueControl structure.
540 : *
541 : * The used entries in the backend[] array run from 1 to MaxBackends; the
542 : * zero'th entry is unused but must be allocated.
543 : */
362 rhaas 544 1826 : size = mul_size(MaxBackends + 1, sizeof(QueueBackendStatus));
2969 tgl 545 1826 : size = add_size(size, offsetof(AsyncQueueControl, backend));
546 :
4800 547 1826 : asyncQueueControl = (AsyncQueueControl *)
548 1826 : ShmemInitStruct("Async Queue Control", size, &found);
549 :
550 1826 : if (!found)
551 : {
552 : /* First time through, so initialize it */
553 1826 : SET_QUEUE_POS(QUEUE_HEAD, 0, 0);
554 1826 : SET_QUEUE_POS(QUEUE_TAIL, 0, 0);
862 555 1826 : QUEUE_STOP_PAGE = 0;
1307 556 1826 : QUEUE_FIRST_LISTENER = InvalidBackendId;
4800 557 1826 : asyncQueueControl->lastQueueFillWarn = 0;
558 : /* zero'th entry won't be used, but let's initialize it anyway */
362 rhaas 559 195029 : for (int i = 0; i <= MaxBackends; i++)
560 : {
4800 tgl 561 193203 : QUEUE_BACKEND_PID(i) = InvalidPid;
2748 562 193203 : QUEUE_BACKEND_DBOID(i) = InvalidOid;
1307 563 193203 : QUEUE_NEXT_LISTENER(i) = InvalidBackendId;
4800 564 193203 : SET_QUEUE_POS(QUEUE_BACKEND_POS(i), 0, 0);
565 : }
566 : }
567 :
568 : /*
569 : * Set up SLRU management of the pg_notify data.
570 : */
1059 571 1826 : NotifyCtl->PagePrecedes = asyncQueuePagePrecedes;
572 1826 : SimpleLruInit(NotifyCtl, "Notify", NUM_NOTIFY_BUFFERS, 0,
926 tmunro 573 1826 : NotifySLRULock, "pg_notify", LWTRANCHE_NOTIFY_BUFFER,
574 : SYNC_HANDLER_NONE);
575 :
4800 tgl 576 1826 : if (!found)
577 : {
578 : /*
579 : * During start or reboot, clean out the pg_notify directory.
580 : */
1059 581 1826 : (void) SlruScanDirectory(NotifyCtl, SlruScanDirCbDeleteAll, NULL);
582 : }
4800 583 1826 : }
584 :
585 :
586 : /*
587 : * pg_notify -
588 : * SQL function to send a notification event
589 : */
590 : Datum
591 1054 : pg_notify(PG_FUNCTION_ARGS)
592 : {
593 : const char *channel;
594 : const char *payload;
595 :
596 1054 : if (PG_ARGISNULL(0))
597 3 : channel = "";
598 : else
599 1051 : channel = text_to_cstring(PG_GETARG_TEXT_PP(0));
600 :
601 1054 : if (PG_ARGISNULL(1))
602 6 : payload = "";
603 : else
604 1048 : payload = text_to_cstring(PG_GETARG_TEXT_PP(1));
605 :
606 : /* For NOTIFY as a statement, this is checked in ProcessUtility */
4796 607 1054 : PreventCommandDuringRecovery("NOTIFY");
608 :
4800 609 1054 : Async_Notify(channel, payload);
610 :
611 1045 : PG_RETURN_VOID();
612 : }
613 :
614 :
615 : /*
616 : * Async_Notify
617 : *
618 : * This is executed by the SQL notify command.
619 : *
620 : * Adds the message to the list of pending notifies.
621 : * Actual notification happens during transaction commit.
622 : * ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
623 : */
624 : void
625 1103 : Async_Notify(const char *channel, const char *payload)
626 : {
1283 rhaas 627 1103 : int my_level = GetCurrentTransactionNestLevel();
628 : size_t channel_len;
629 : size_t payload_len;
630 : Notification *n;
631 : MemoryContext oldcontext;
632 :
2726 633 1103 : if (IsParallelWorker())
2726 rhaas 634 UBC 0 : elog(ERROR, "cannot send notifications from a parallel worker");
635 :
8348 peter_e 636 CBC 1103 : if (Trace_notify)
4800 tgl 637 UBC 0 : elog(DEBUG1, "Async_Notify(%s)", channel);
638 :
1333 tgl 639 CBC 1103 : channel_len = channel ? strlen(channel) : 0;
640 1103 : payload_len = payload ? strlen(payload) : 0;
641 :
642 : /* a channel name must be specified */
643 1103 : if (channel_len == 0)
4800 644 6 : ereport(ERROR,
645 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
646 : errmsg("channel name cannot be empty")));
647 :
648 : /* enforce length limits */
1333 649 1097 : if (channel_len >= NAMEDATALEN)
4800 650 3 : ereport(ERROR,
651 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
652 : errmsg("channel name too long")));
653 :
1333 654 1094 : if (payload_len >= NOTIFY_PAYLOAD_MAX_LENGTH)
1333 tgl 655 UBC 0 : ereport(ERROR,
656 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
657 : errmsg("payload string too long")));
658 :
659 : /*
660 : * We must construct the Notification entry, even if we end up not using
661 : * it, in order to compare it cheaply to existing list entries.
662 : *
663 : * The notification list needs to live until end of transaction, so store
664 : * it in the transaction context.
665 : */
4800 tgl 666 CBC 1094 : oldcontext = MemoryContextSwitchTo(CurTransactionContext);
667 :
1333 668 1094 : n = (Notification *) palloc(offsetof(Notification, data) +
669 1094 : channel_len + payload_len + 2);
670 1094 : n->channel_len = channel_len;
671 1094 : n->payload_len = payload_len;
672 1094 : strcpy(n->data, channel);
4800 673 1094 : if (payload)
1333 674 1081 : strcpy(n->data + channel_len + 1, payload);
675 : else
676 13 : n->data[channel_len + 1] = '\0';
677 :
1283 rhaas 678 1094 : if (pendingNotifies == NULL || my_level > pendingNotifies->nestingLevel)
1333 tgl 679 50 : {
680 : NotificationList *notifies;
681 :
682 : /*
683 : * First notify event in current (sub)xact. Note that we allocate the
684 : * NotificationList in TopTransactionContext; the nestingLevel might
685 : * get changed later by AtSubCommit_Notify.
686 : */
687 : notifies = (NotificationList *)
1283 rhaas 688 50 : MemoryContextAlloc(TopTransactionContext,
689 : sizeof(NotificationList));
690 50 : notifies->nestingLevel = my_level;
691 50 : notifies->events = list_make1(n);
692 : /* We certainly don't need a hashtable yet */
693 50 : notifies->hashtab = NULL;
694 50 : notifies->upper = pendingNotifies;
695 50 : pendingNotifies = notifies;
696 : }
697 : else
698 : {
699 : /* Now check for duplicates */
700 1044 : if (AsyncExistsPendingNotify(n))
701 : {
702 : /* It's a dup, so forget it */
703 12 : pfree(n);
704 12 : MemoryContextSwitchTo(oldcontext);
705 12 : return;
706 : }
707 :
708 : /* Append more events to existing list */
1333 tgl 709 1032 : AddEventToPendingNotifies(n);
710 : }
711 :
4800 712 1082 : MemoryContextSwitchTo(oldcontext);
713 : }
714 :
715 : /*
716 : * queue_listen
717 : * Common code for listen, unlisten, unlisten all commands.
718 : *
719 : * Adds the request to the list of pending actions.
720 : * Actual update of the listenChannels list happens during transaction
721 : * commit.
722 : */
723 : static void
724 57 : queue_listen(ListenActionKind action, const char *channel)
725 : {
726 : MemoryContext oldcontext;
727 : ListenAction *actrec;
1283 rhaas 728 57 : int my_level = GetCurrentTransactionNestLevel();
729 :
730 : /*
731 : * Unlike Async_Notify, we don't try to collapse out duplicates. It would
732 : * be too complicated to ensure we get the right interactions of
733 : * conflicting LISTEN/UNLISTEN/UNLISTEN_ALL, and it's unlikely that there
734 : * would be any performance benefit anyway in sane applications.
735 : */
5506 tgl 736 57 : oldcontext = MemoryContextSwitchTo(CurTransactionContext);
737 :
738 : /* space for terminating null is included in sizeof(ListenAction) */
2969 739 57 : actrec = (ListenAction *) palloc(offsetof(ListenAction, channel) +
740 57 : strlen(channel) + 1);
5506 741 57 : actrec->action = action;
4800 742 57 : strcpy(actrec->channel, channel);
743 :
1283 rhaas 744 57 : if (pendingActions == NULL || my_level > pendingActions->nestingLevel)
745 50 : {
746 : ActionList *actions;
747 :
748 : /*
749 : * First action in current sub(xact). Note that we allocate the
750 : * ActionList in TopTransactionContext; the nestingLevel might get
751 : * changed later by AtSubCommit_Notify.
752 : */
753 : actions = (ActionList *)
754 50 : MemoryContextAlloc(TopTransactionContext, sizeof(ActionList));
755 50 : actions->nestingLevel = my_level;
756 50 : actions->actions = list_make1(actrec);
757 50 : actions->upper = pendingActions;
758 50 : pendingActions = actions;
759 : }
760 : else
761 7 : pendingActions->actions = lappend(pendingActions->actions, actrec);
762 :
5506 tgl 763 57 : MemoryContextSwitchTo(oldcontext);
764 57 : }
765 :
766 : /*
767 : * Async_Listen
768 : *
769 : * This is executed by the SQL listen command.
770 : */
771 : void
4800 772 37 : Async_Listen(const char *channel)
773 : {
5506 774 37 : if (Trace_notify)
4800 tgl 775 UBC 0 : elog(DEBUG1, "Async_Listen(%s,%d)", channel, MyProcPid);
776 :
4800 tgl 777 CBC 37 : queue_listen(LISTEN_LISTEN, channel);
5506 778 37 : }
779 :
780 : /*
781 : * Async_Unlisten
782 : *
783 : * This is executed by the SQL unlisten command.
784 : */
785 : void
4800 786 3 : Async_Unlisten(const char *channel)
787 : {
5335 788 3 : if (Trace_notify)
4800 tgl 789 UBC 0 : elog(DEBUG1, "Async_Unlisten(%s,%d)", channel, MyProcPid);
790 :
791 : /* If we couldn't possibly be listening, no need to queue anything */
1283 rhaas 792 CBC 3 : if (pendingActions == NULL && !unlistenExitRegistered)
5168 tgl 793 UBC 0 : return;
794 :
4800 tgl 795 CBC 3 : queue_listen(LISTEN_UNLISTEN, channel);
796 : }
797 :
798 : /*
799 : * Async_UnlistenAll
800 : *
801 : * This is invoked by UNLISTEN * command, and also at backend exit.
802 : */
803 : void
5506 804 19 : Async_UnlistenAll(void)
805 : {
806 19 : if (Trace_notify)
5506 tgl 807 UBC 0 : elog(DEBUG1, "Async_UnlistenAll(%d)", MyProcPid);
808 :
809 : /* If we couldn't possibly be listening, no need to queue anything */
1283 rhaas 810 CBC 19 : if (pendingActions == NULL && !unlistenExitRegistered)
5168 tgl 811 2 : return;
812 :
5506 813 17 : queue_listen(LISTEN_UNLISTEN_ALL, "");
814 : }
815 :
816 : /*
817 : * SQL function: return a set of the channel names this backend is actively
818 : * listening to.
819 : *
820 : * Note: this coding relies on the fact that the listenChannels list cannot
821 : * change within a transaction.
822 : */
823 : Datum
4800 824 9 : pg_listening_channels(PG_FUNCTION_ARGS)
825 : {
826 : FuncCallContext *funcctx;
827 :
828 : /* stuff done only on the first call of the function */
829 9 : if (SRF_IS_FIRSTCALL())
830 : {
831 : /* create a function context for cross-call persistence */
832 6 : funcctx = SRF_FIRSTCALL_INIT();
833 : }
834 :
835 : /* stuff done on every call of the function */
836 9 : funcctx = SRF_PERCALL_SETUP();
837 :
1364 838 9 : if (funcctx->call_cntr < list_length(listenChannels))
839 : {
840 3 : char *channel = (char *) list_nth(listenChannels,
841 3 : funcctx->call_cntr);
842 :
4800 843 3 : SRF_RETURN_NEXT(funcctx, CStringGetTextDatum(channel));
844 : }
845 :
846 6 : SRF_RETURN_DONE(funcctx);
847 : }
848 :
849 : /*
850 : * Async_UnlistenOnExit
851 : *
852 : * This is executed at backend exit if we have done any LISTENs in this
853 : * backend. It might not be necessary anymore, if the user UNLISTENed
854 : * everything, but we don't try to detect that case.
855 : */
856 : static void
5506 857 14 : Async_UnlistenOnExit(int code, Datum arg)
858 : {
4800 859 14 : Exec_UnlistenAllCommit();
3707 860 14 : asyncQueueUnregister();
5506 861 14 : }
862 :
863 : /*
864 : * AtPrepare_Notify
865 : *
866 : * This is called at the prepare phase of a two-phase
867 : * transaction. Save the state for possible commit later.
868 : */
869 : void
870 358 : AtPrepare_Notify(void)
871 : {
872 : /* It's not allowed to have any pending LISTEN/UNLISTEN/NOTIFY actions */
4800 873 358 : if (pendingActions || pendingNotifies)
5506 tgl 874 UBC 0 : ereport(ERROR,
875 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
876 : errmsg("cannot PREPARE a transaction that has executed LISTEN, UNLISTEN, or NOTIFY")));
4800 tgl 877 CBC 358 : }
878 :
879 : /*
880 : * PreCommit_Notify
881 : *
882 : * This is called at transaction commit, before actually committing to
883 : * clog.
884 : *
885 : * If there are pending LISTEN actions, make sure we are listed in the
886 : * shared-memory listener array. This must happen before commit to
887 : * ensure we don't miss any notifies from transactions that commit
888 : * just after ours.
889 : *
890 : * If there are outbound notify requests in the pendingNotifies list,
891 : * add them to the global queue. We do that before commit so that
892 : * we can still throw error if we run out of queue space.
893 : */
894 : void
895 465495 : PreCommit_Notify(void)
896 : {
897 : ListCell *p;
898 :
1333 899 465495 : if (!pendingActions && !pendingNotifies)
4800 900 465398 : return; /* no relevant statements in this xact */
901 :
902 97 : if (Trace_notify)
4800 tgl 903 UBC 0 : elog(DEBUG1, "PreCommit_Notify");
904 :
905 : /* Preflight for any pending listen/unlisten actions */
1283 rhaas 906 CBC 97 : if (pendingActions != NULL)
907 : {
908 105 : foreach(p, pendingActions->actions)
909 : {
910 56 : ListenAction *actrec = (ListenAction *) lfirst(p);
911 :
912 56 : switch (actrec->action)
913 : {
914 37 : case LISTEN_LISTEN:
915 37 : Exec_ListenPreCommit();
916 37 : break;
917 3 : case LISTEN_UNLISTEN:
918 : /* there is no Exec_UnlistenPreCommit() */
919 3 : break;
920 16 : case LISTEN_UNLISTEN_ALL:
921 : /* there is no Exec_UnlistenAllPreCommit() */
922 16 : break;
923 : }
924 : }
925 : }
926 :
927 : /* Queue any pending notifies (must happen after the above) */
4800 tgl 928 97 : if (pendingNotifies)
929 : {
930 : ListCell *nextNotify;
931 :
932 : /*
933 : * Make sure that we have an XID assigned to the current transaction.
934 : * GetCurrentTransactionId is cheap if we already have an XID, but not
935 : * so cheap if we don't, and we'd prefer not to do that work while
936 : * holding NotifyQueueLock.
937 : */
938 48 : (void) GetCurrentTransactionId();
939 :
940 : /*
941 : * Serialize writers by acquiring a special lock that we hold till
942 : * after commit. This ensures that queue entries appear in commit
943 : * order, and in particular that there are never uncommitted queue
944 : * entries ahead of committed ones, so an uncommitted transaction
945 : * can't block delivery of deliverable notifications.
946 : *
947 : * We use a heavyweight lock so that it'll automatically be released
948 : * after either commit or abort. This also allows deadlocks to be
949 : * detected, though really a deadlock shouldn't be possible here.
950 : *
951 : * The lock is on "database 0", which is pretty ugly but it doesn't
952 : * seem worth inventing a special locktag category just for this.
953 : * (Historical note: before PG 9.0, a similar lock on "database 0" was
954 : * used by the flatfiles mechanism.)
955 : */
956 48 : LockSharedObject(DatabaseRelationId, InvalidOid, 0,
957 : AccessExclusiveLock);
958 :
959 : /* Now push the notifications into the queue */
1333 960 48 : nextNotify = list_head(pendingNotifies->events);
4800 961 131 : while (nextNotify != NULL)
962 : {
963 : /*
964 : * Add the pending notifications to the queue. We acquire and
965 : * release NotifyQueueLock once per page, which might be overkill
966 : * but it does allow readers to get in while we're doing this.
967 : *
968 : * A full queue is very uncommon and should really not happen,
969 : * given that we have so much space available in the SLRU pages.
970 : * Nevertheless we need to deal with this possibility. Note that
971 : * when we get here we are in the process of committing our
972 : * transaction, but we have not yet committed to clog, so at this
973 : * point in time we can still roll the transaction back.
974 : */
1059 975 83 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
4800 976 83 : asyncQueueFillWarning();
977 83 : if (asyncQueueIsFull())
4800 tgl 978 UBC 0 : ereport(ERROR,
979 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
980 : errmsg("too many notifications in the NOTIFY queue")));
4800 tgl 981 CBC 83 : nextNotify = asyncQueueAddEntries(nextNotify);
1059 982 83 : LWLockRelease(NotifyQueueLock);
983 : }
984 :
985 : /* Note that we don't clear pendingNotifies; AtCommit_Notify will. */
986 : }
987 : }
988 :
989 : /*
990 : * AtCommit_Notify
991 : *
992 : * This is called at transaction commit, after committing to clog.
993 : *
994 : * Update listenChannels and clear transaction-local state.
995 : *
996 : * If we issued any notifications in the transaction, send signals to
997 : * listening backends (possibly including ourselves) to process them.
998 : * Also, if we filled enough queue pages with new notifies, try to
999 : * advance the queue tail pointer.
1000 : */
1001 : void
5506 1002 465340 : AtCommit_Notify(void)
1003 : {
1004 : ListCell *p;
1005 :
1006 : /*
1007 : * Allow transactions that have not executed LISTEN/UNLISTEN/NOTIFY to
1008 : * return as soon as possible
1009 : */
4800 1010 465340 : if (!pendingActions && !pendingNotifies)
5506 1011 465243 : return;
1012 :
1013 97 : if (Trace_notify)
5506 tgl 1014 UBC 0 : elog(DEBUG1, "AtCommit_Notify");
1015 :
1016 : /* Perform any pending listen/unlisten actions */
1283 rhaas 1017 CBC 97 : if (pendingActions != NULL)
1018 : {
1019 105 : foreach(p, pendingActions->actions)
1020 : {
1021 56 : ListenAction *actrec = (ListenAction *) lfirst(p);
1022 :
1023 56 : switch (actrec->action)
1024 : {
1025 37 : case LISTEN_LISTEN:
1026 37 : Exec_ListenCommit(actrec->channel);
1027 37 : break;
1028 3 : case LISTEN_UNLISTEN:
1029 3 : Exec_UnlistenCommit(actrec->channel);
1030 3 : break;
1031 16 : case LISTEN_UNLISTEN_ALL:
1032 16 : Exec_UnlistenAllCommit();
1033 16 : break;
1034 : }
1035 : }
1036 : }
1037 :
1038 : /* If no longer listening to anything, get out of listener array */
3707 tgl 1039 97 : if (amRegisteredListener && listenChannels == NIL)
1040 13 : asyncQueueUnregister();
1041 :
1042 : /*
1043 : * Send signals to listening backends. We need do this only if there are
1044 : * pending notifies, which were previously added to the shared queue by
1045 : * PreCommit_Notify().
1046 : */
572 1047 97 : if (pendingNotifies != NULL)
1048 48 : SignalBackends();
1049 :
1050 : /*
1051 : * If it's time to try to advance the global tail pointer, do that.
1052 : *
1053 : * (It might seem odd to do this in the sender, when more than likely the
1054 : * listeners won't yet have read the messages we just sent. However,
1055 : * there's less contention if only the sender does it, and there is little
1056 : * need for urgency in advancing the global tail. So this typically will
1057 : * be clearing out messages that were sent some time ago.)
1058 : */
1059 97 : if (tryAdvanceTail)
1060 : {
1061 8 : tryAdvanceTail = false;
1062 8 : asyncQueueAdvanceTail();
1063 : }
1064 :
1065 : /* And clean up */
4800 1066 97 : ClearPendingActionsAndNotifies();
1067 : }
1068 :
1069 : /*
1070 : * Exec_ListenPreCommit --- subroutine for PreCommit_Notify
1071 : *
1072 : * This function must make sure we are ready to catch any incoming messages.
1073 : */
1074 : static void
1075 37 : Exec_ListenPreCommit(void)
1076 : {
1077 : QueuePosition head;
1078 : QueuePosition max;
1079 : BackendId prevListener;
1080 :
1081 : /*
1082 : * Nothing to do if we are already listening to something, nor if we
1083 : * already ran this routine in this transaction.
1084 : */
3707 1085 37 : if (amRegisteredListener)
4800 1086 18 : return;
1087 :
1088 19 : if (Trace_notify)
4800 tgl 1089 UBC 0 : elog(DEBUG1, "Exec_ListenPreCommit(%d)", MyProcPid);
1090 :
1091 : /*
1092 : * Before registering, make sure we will unlisten before dying. (Note:
1093 : * this action does not get undone if we abort later.)
1094 : */
4800 tgl 1095 CBC 19 : if (!unlistenExitRegistered)
1096 : {
3399 rhaas 1097 14 : before_shmem_exit(Async_UnlistenOnExit, 0);
4800 tgl 1098 14 : unlistenExitRegistered = true;
1099 : }
1100 :
1101 : /*
1102 : * This is our first LISTEN, so establish our pointer.
1103 : *
1104 : * We set our pointer to the global tail pointer and then move it forward
1105 : * over already-committed notifications. This ensures we cannot miss any
1106 : * not-yet-committed notifications. We might get a few more but that
1107 : * doesn't hurt.
1108 : *
1109 : * In some scenarios there might be a lot of committed notifications that
1110 : * have not yet been pruned away (because some backend is being lazy about
1111 : * reading them). To reduce our startup time, we can look at other
1112 : * backends and adopt the maximum "pos" pointer of any backend that's in
1113 : * our database; any notifications it's already advanced over are surely
1114 : * committed and need not be re-examined by us. (We must consider only
1115 : * backends connected to our DB, because others will not have bothered to
1116 : * check committed-ness of notifications in our DB.)
1117 : *
1118 : * We need exclusive lock here so we can look at other backends' entries
1119 : * and manipulate the list links.
1120 : */
1059 1121 19 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
2748 1122 19 : head = QUEUE_HEAD;
1123 19 : max = QUEUE_TAIL;
1307 1124 19 : prevListener = InvalidBackendId;
1125 21 : for (BackendId i = QUEUE_FIRST_LISTENER; i > 0; i = QUEUE_NEXT_LISTENER(i))
1126 : {
1127 2 : if (QUEUE_BACKEND_DBOID(i) == MyDatabaseId)
1128 2 : max = QUEUE_POS_MAX(max, QUEUE_BACKEND_POS(i));
1129 : /* Also find last listening backend before this one */
1130 2 : if (i < MyBackendId)
1131 1 : prevListener = i;
1132 : }
2748 1133 19 : QUEUE_BACKEND_POS(MyBackendId) = max;
4800 1134 19 : QUEUE_BACKEND_PID(MyBackendId) = MyProcPid;
2748 1135 19 : QUEUE_BACKEND_DBOID(MyBackendId) = MyDatabaseId;
1136 : /* Insert backend into list of listeners at correct position */
1307 1137 19 : if (prevListener > 0)
1138 : {
1139 1 : QUEUE_NEXT_LISTENER(MyBackendId) = QUEUE_NEXT_LISTENER(prevListener);
1140 1 : QUEUE_NEXT_LISTENER(prevListener) = MyBackendId;
1141 : }
1142 : else
1143 : {
1144 18 : QUEUE_NEXT_LISTENER(MyBackendId) = QUEUE_FIRST_LISTENER;
1145 18 : QUEUE_FIRST_LISTENER = MyBackendId;
1146 : }
1059 1147 19 : LWLockRelease(NotifyQueueLock);
1148 :
1149 : /* Now we are listed in the global array, so remember we're listening */
3707 1150 19 : amRegisteredListener = true;
1151 :
1152 : /*
1153 : * Try to move our pointer forward as far as possible. This will skip
1154 : * over already-committed notifications, which we want to do because they
1155 : * might be quite stale. Note that we are not yet listening on anything,
1156 : * so we won't deliver such notifications to our frontend. Also, although
1157 : * our transaction might have executed NOTIFY, those message(s) aren't
1158 : * queued yet so we won't skip them here.
1159 : */
2748 1160 19 : if (!QUEUE_POS_EQUAL(max, head))
1161 11 : asyncQueueReadAllNotifications();
1162 : }
1163 :
1164 : /*
1165 : * Exec_ListenCommit --- subroutine for AtCommit_Notify
1166 : *
1167 : * Add the channel to the list of channels we are listening on.
1168 : */
1169 : static void
4800 1170 37 : Exec_ListenCommit(const char *channel)
1171 : {
1172 : MemoryContext oldcontext;
1173 :
1174 : /* Do nothing if we are already listening on this channel */
1175 37 : if (IsListeningOn(channel))
1176 10 : return;
1177 :
1178 : /*
1179 : * Add the new channel name to listenChannels.
1180 : *
1181 : * XXX It is theoretically possible to get an out-of-memory failure here,
1182 : * which would be bad because we already committed. For the moment it
1183 : * doesn't seem worth trying to guard against that, but maybe improve this
1184 : * later.
1185 : */
1186 27 : oldcontext = MemoryContextSwitchTo(TopMemoryContext);
1187 27 : listenChannels = lappend(listenChannels, pstrdup(channel));
1188 27 : MemoryContextSwitchTo(oldcontext);
1189 : }
1190 :
1191 : /*
1192 : * Exec_UnlistenCommit --- subroutine for AtCommit_Notify
1193 : *
1194 : * Remove the specified channel name from listenChannels.
1195 : */
1196 : static void
1197 3 : Exec_UnlistenCommit(const char *channel)
1198 : {
1199 : ListCell *q;
1200 :
1201 3 : if (Trace_notify)
4800 tgl 1202 UBC 0 : elog(DEBUG1, "Exec_UnlistenCommit(%s,%d)", channel, MyProcPid);
1203 :
4800 tgl 1204 CBC 3 : foreach(q, listenChannels)
1205 : {
4790 bruce 1206 3 : char *lchan = (char *) lfirst(q);
1207 :
4800 tgl 1208 3 : if (strcmp(lchan, channel) == 0)
1209 : {
1364 1210 3 : listenChannels = foreach_delete_current(listenChannels, q);
4800 1211 3 : pfree(lchan);
1212 3 : break;
1213 : }
1214 : }
1215 :
1216 : /*
1217 : * We do not complain about unlistening something not being listened;
1218 : * should we?
1219 : */
1220 3 : }
1221 :
1222 : /*
1223 : * Exec_UnlistenAllCommit --- subroutine for AtCommit_Notify
1224 : *
1225 : * Unlisten on all channels for this backend.
1226 : */
1227 : static void
1228 30 : Exec_UnlistenAllCommit(void)
1229 : {
1230 30 : if (Trace_notify)
4800 tgl 1231 UBC 0 : elog(DEBUG1, "Exec_UnlistenAllCommit(%d)", MyProcPid);
1232 :
4800 tgl 1233 CBC 30 : list_free_deep(listenChannels);
1234 30 : listenChannels = NIL;
1235 30 : }
1236 :
1237 : /*
1238 : * Test whether we are actively listening on the given channel name.
1239 : *
1240 : * Note: this function is executed for every notification found in the queue.
1241 : * Perhaps it is worth further optimization, eg convert the list to a sorted
1242 : * array so we can binary-search it. In practice the list is likely to be
1243 : * fairly short, though.
1244 : */
1245 : static bool
1246 1196 : IsListeningOn(const char *channel)
1247 : {
1248 : ListCell *p;
1249 :
1250 1257 : foreach(p, listenChannels)
1251 : {
1252 102 : char *lchan = (char *) lfirst(p);
1253 :
1254 102 : if (strcmp(lchan, channel) == 0)
1255 41 : return true;
1256 : }
1257 1155 : return false;
1258 : }
1259 :
1260 : /*
1261 : * Remove our entry from the listeners array when we are no longer listening
1262 : * on any channel. NB: must not fail if we're already not listening.
1263 : */
1264 : static void
1265 27 : asyncQueueUnregister(void)
1266 : {
2118 1267 27 : Assert(listenChannels == NIL); /* else caller error */
1268 :
3602 bruce 1269 27 : if (!amRegisteredListener) /* nothing to do */
3707 tgl 1270 8 : return;
1271 :
1272 : /*
1273 : * Need exclusive lock here to manipulate list links.
1274 : */
1059 1275 19 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
1276 : /* Mark our entry as invalid */
4800 1277 19 : QUEUE_BACKEND_PID(MyBackendId) = InvalidPid;
2748 1278 19 : QUEUE_BACKEND_DBOID(MyBackendId) = InvalidOid;
1279 : /* and remove it from the list */
1307 1280 19 : if (QUEUE_FIRST_LISTENER == MyBackendId)
1281 18 : QUEUE_FIRST_LISTENER = QUEUE_NEXT_LISTENER(MyBackendId);
1282 : else
1283 : {
1284 1 : for (BackendId i = QUEUE_FIRST_LISTENER; i > 0; i = QUEUE_NEXT_LISTENER(i))
1285 : {
1286 1 : if (QUEUE_NEXT_LISTENER(i) == MyBackendId)
1287 : {
1288 1 : QUEUE_NEXT_LISTENER(i) = QUEUE_NEXT_LISTENER(MyBackendId);
1289 1 : break;
1290 : }
1291 : }
1292 : }
1293 19 : QUEUE_NEXT_LISTENER(MyBackendId) = InvalidBackendId;
1059 1294 19 : LWLockRelease(NotifyQueueLock);
1295 :
1296 : /* mark ourselves as no longer listed in the global array */
3707 1297 19 : amRegisteredListener = false;
1298 : }
1299 :
1300 : /*
1301 : * Test whether there is room to insert more notification messages.
1302 : *
1303 : * Caller must hold at least shared NotifyQueueLock.
1304 : */
1305 : static bool
4800 1306 83 : asyncQueueIsFull(void)
1307 : {
1308 : int nexthead;
1309 : int boundary;
1310 :
1311 : /*
1312 : * The queue is full if creating a new head page would create a page that
1313 : * logically precedes the current global tail pointer, ie, the head
1314 : * pointer would wrap around compared to the tail. We cannot create such
1315 : * a head page for fear of confusing slru.c. For safety we round the tail
1316 : * pointer back to a segment boundary (truncation logic in
1317 : * asyncQueueAdvanceTail does not do this, so doing it here is optional).
1318 : *
1319 : * Note that this test is *not* dependent on how much space there is on
1320 : * the current head page. This is necessary because asyncQueueAddEntries
1321 : * might try to create the next head page in any case.
1322 : */
1323 83 : nexthead = QUEUE_POS_PAGE(QUEUE_HEAD) + 1;
1324 83 : if (nexthead > QUEUE_MAX_PAGE)
4800 tgl 1325 UBC 0 : nexthead = 0; /* wrap around */
862 tgl 1326 CBC 83 : boundary = QUEUE_STOP_PAGE;
4800 1327 83 : boundary -= boundary % SLRU_PAGES_PER_SEGMENT;
4211 alvherre 1328 83 : return asyncQueuePagePrecedes(nexthead, boundary);
1329 : }
1330 :
1331 : /*
1332 : * Advance the QueuePosition to the next entry, assuming that the current
1333 : * entry is of length entryLength. If we jump to a new page the function
1334 : * returns true, else false.
1335 : */
1336 : static bool
2995 tgl 1337 2301 : asyncQueueAdvance(volatile QueuePosition *position, int entryLength)
1338 : {
4790 bruce 1339 2301 : int pageno = QUEUE_POS_PAGE(*position);
1340 2301 : int offset = QUEUE_POS_OFFSET(*position);
1341 2301 : bool pageJump = false;
1342 :
1343 : /*
1344 : * Move to the next writing position: First jump over what we have just
1345 : * written or read.
1346 : */
4800 tgl 1347 2301 : offset += entryLength;
1348 2301 : Assert(offset <= QUEUE_PAGESIZE);
1349 :
1350 : /*
1351 : * In a second step check if another entry can possibly be written to the
1352 : * page. If so, stay here, we have reached the next position. If not, then
1353 : * we need to move on to the next page.
1354 : */
1355 2301 : if (offset + QUEUEALIGN(AsyncQueueEntryEmptySize) > QUEUE_PAGESIZE)
1356 : {
1357 70 : pageno++;
1358 70 : if (pageno > QUEUE_MAX_PAGE)
4790 bruce 1359 UBC 0 : pageno = 0; /* wrap around */
4800 tgl 1360 CBC 70 : offset = 0;
1361 70 : pageJump = true;
1362 : }
1363 :
1364 2301 : SET_QUEUE_POS(*position, pageno, offset);
1365 2301 : return pageJump;
1366 : }
1367 :
1368 : /*
1369 : * Fill the AsyncQueueEntry at *qe with an outbound notification message.
1370 : */
1371 : static void
1372 1109 : asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe)
1373 : {
1333 1374 1109 : size_t channellen = n->channel_len;
1375 1109 : size_t payloadlen = n->payload_len;
1376 : int entryLength;
1377 :
4800 1378 1109 : Assert(channellen < NAMEDATALEN);
1379 1109 : Assert(payloadlen < NOTIFY_PAYLOAD_MAX_LENGTH);
1380 :
1381 : /* The terminators are already included in AsyncQueueEntryEmptySize */
1382 1109 : entryLength = AsyncQueueEntryEmptySize + payloadlen + channellen;
1383 1109 : entryLength = QUEUEALIGN(entryLength);
1384 1109 : qe->length = entryLength;
1385 1109 : qe->dboid = MyDatabaseId;
1386 1109 : qe->xid = GetCurrentTransactionId();
1387 1109 : qe->srcPid = MyProcPid;
1333 1388 1109 : memcpy(qe->data, n->data, channellen + payloadlen + 2);
4800 1389 1109 : }
1390 :
1391 : /*
1392 : * Add pending notifications to the queue.
1393 : *
1394 : * We go page by page here, i.e. we stop once we have to go to a new page but
1395 : * we will be called again and then fill that next page. If an entry does not
1396 : * fit into the current page, we write a dummy entry with an InvalidOid as the
1397 : * database OID in order to fill the page. So every page is always used up to
1398 : * the last byte which simplifies reading the page later.
1399 : *
1400 : * We are passed the list cell (in pendingNotifies->events) containing the next
1401 : * notification to write and return the first still-unwritten cell back.
1402 : * Eventually we will return NULL indicating all is done.
1403 : *
1404 : * We are holding NotifyQueueLock already from the caller and grab
1405 : * NotifySLRULock locally in this function.
1406 : */
1407 : static ListCell *
1408 83 : asyncQueueAddEntries(ListCell *nextNotify)
1409 : {
1410 : AsyncQueueEntry qe;
1411 : QueuePosition queue_head;
1412 : int pageno;
1413 : int offset;
1414 : int slotno;
1415 :
1416 : /* We hold both NotifyQueueLock and NotifySLRULock during this operation */
1059 1417 83 : LWLockAcquire(NotifySLRULock, LW_EXCLUSIVE);
1418 :
1419 : /*
1420 : * We work with a local copy of QUEUE_HEAD, which we write back to shared
1421 : * memory upon exiting. The reason for this is that if we have to advance
1422 : * to a new page, SimpleLruZeroPage might fail (out of disk space, for
1423 : * instance), and we must not advance QUEUE_HEAD if it does. (Otherwise,
1424 : * subsequent insertions would try to put entries into a page that slru.c
1425 : * thinks doesn't exist yet.) So, use a local position variable. Note
1426 : * that if we do fail, any already-inserted queue entries are forgotten;
1427 : * this is okay, since they'd be useless anyway after our transaction
1428 : * rolls back.
1429 : */
3936 1430 83 : queue_head = QUEUE_HEAD;
1431 :
1432 : /*
1433 : * If this is the first write since the postmaster started, we need to
1434 : * initialize the first page of the async SLRU. Otherwise, the current
1435 : * page should be initialized already, so just fetch it.
1436 : *
1437 : * (We could also take the first path when the SLRU position has just
1438 : * wrapped around, but re-zeroing the page is harmless in that case.)
1439 : */
1440 83 : pageno = QUEUE_POS_PAGE(queue_head);
1061 1441 83 : if (QUEUE_POS_IS_ZERO(queue_head))
1059 1442 7 : slotno = SimpleLruZeroPage(NotifyCtl, pageno);
1443 : else
1444 76 : slotno = SimpleLruReadPage(NotifyCtl, pageno, true,
1445 : InvalidTransactionId);
1446 :
1447 : /* Note we mark the page dirty before writing in it */
1448 83 : NotifyCtl->shared->page_dirty[slotno] = true;
1449 :
4800 1450 1157 : while (nextNotify != NULL)
1451 : {
4790 bruce 1452 1109 : Notification *n = (Notification *) lfirst(nextNotify);
1453 :
1454 : /* Construct a valid queue entry in local variable qe */
4800 tgl 1455 1109 : asyncQueueNotificationToEntry(n, &qe);
1456 :
3936 1457 1109 : offset = QUEUE_POS_OFFSET(queue_head);
1458 :
1459 : /* Check whether the entry really fits on the current page */
4800 1460 1109 : if (offset + qe.length <= QUEUE_PAGESIZE)
1461 : {
1462 : /* OK, so advance nextNotify past this item */
1333 1463 1076 : nextNotify = lnext(pendingNotifies->events, nextNotify);
1464 : }
1465 : else
1466 : {
1467 : /*
1468 : * Write a dummy entry to fill up the page. Actually readers will
1469 : * only check dboid and since it won't match any reader's database
1470 : * OID, they will ignore this entry and move on.
1471 : */
4800 1472 33 : qe.length = QUEUE_PAGESIZE - offset;
1473 33 : qe.dboid = InvalidOid;
4790 bruce 1474 33 : qe.data[0] = '\0'; /* empty channel */
1475 33 : qe.data[1] = '\0'; /* empty payload */
1476 : }
1477 :
1478 : /* Now copy qe into the shared buffer page */
1059 tgl 1479 1109 : memcpy(NotifyCtl->shared->page_buffer[slotno] + offset,
1480 : &qe,
4800 1481 1109 : qe.length);
1482 :
1483 : /* Advance queue_head appropriately, and detect if page is full */
3936 1484 1109 : if (asyncQueueAdvance(&(queue_head), qe.length))
1485 : {
1486 : /*
1487 : * Page is full, so we're done here, but first fill the next page
1488 : * with zeroes. The reason to do this is to ensure that slru.c's
1489 : * idea of the head page is always the same as ours, which avoids
1490 : * boundary problems in SimpleLruTruncate. The test in
1491 : * asyncQueueIsFull() ensured that there is room to create this
1492 : * page without overrunning the queue.
1493 : */
1059 1494 35 : slotno = SimpleLruZeroPage(NotifyCtl, QUEUE_POS_PAGE(queue_head));
1495 :
1496 : /*
1497 : * If the new page address is a multiple of QUEUE_CLEANUP_DELAY,
1498 : * set flag to remember that we should try to advance the tail
1499 : * pointer (we don't want to actually do that right here).
1500 : */
1295 1501 35 : if (QUEUE_POS_PAGE(queue_head) % QUEUE_CLEANUP_DELAY == 0)
572 1502 8 : tryAdvanceTail = true;
1503 :
1504 : /* And exit the loop */
7971 1505 35 : break;
1506 : }
1507 : }
1508 :
1509 : /* Success, so update the global QUEUE_HEAD */
3936 1510 83 : QUEUE_HEAD = queue_head;
1511 :
1059 1512 83 : LWLockRelease(NotifySLRULock);
1513 :
4800 1514 83 : return nextNotify;
1515 : }
1516 :
1517 : /*
1518 : * SQL function to return the fraction of the notification queue currently
1519 : * occupied.
1520 : */
1521 : Datum
2823 rhaas 1522 5 : pg_notification_queue_usage(PG_FUNCTION_ARGS)
1523 : {
1524 : double usage;
1525 :
1526 : /* Advance the queue tail so we don't report a too-large result */
1232 tgl 1527 5 : asyncQueueAdvanceTail();
1528 :
1059 1529 5 : LWLockAcquire(NotifyQueueLock, LW_SHARED);
2823 rhaas 1530 5 : usage = asyncQueueUsage();
1059 tgl 1531 5 : LWLockRelease(NotifyQueueLock);
1532 :
2823 rhaas 1533 5 : PG_RETURN_FLOAT8(usage);
1534 : }
1535 :
1536 : /*
1537 : * Return the fraction of the queue that is currently occupied.
1538 : *
1539 : * The caller must hold NotifyQueueLock in (at least) shared mode.
1540 : *
1541 : * Note: we measure the distance to the logical tail page, not the physical
1542 : * tail page. In some sense that's wrong, but the relative position of the
1543 : * physical tail is affected by details such as SLRU segment boundaries,
1544 : * so that a result based on that is unpleasantly unstable.
1545 : */
1546 : static double
1547 88 : asyncQueueUsage(void)
1548 : {
4790 bruce 1549 88 : int headPage = QUEUE_POS_PAGE(QUEUE_HEAD);
1550 88 : int tailPage = QUEUE_POS_PAGE(QUEUE_TAIL);
1551 : int occupied;
1552 :
4800 tgl 1553 88 : occupied = headPage - tailPage;
1554 :
1555 88 : if (occupied == 0)
2823 rhaas 1556 41 : return (double) 0; /* fast exit for common case */
1557 :
4800 tgl 1558 47 : if (occupied < 0)
1559 : {
1560 : /* head has wrapped around, tail not yet */
4790 bruce 1561 UBC 0 : occupied += QUEUE_MAX_PAGE + 1;
1562 : }
1563 :
2823 rhaas 1564 CBC 47 : return (double) occupied / (double) ((QUEUE_MAX_PAGE + 1) / 2);
1565 : }
1566 :
1567 : /*
1568 : * Check whether the queue is at least half full, and emit a warning if so.
1569 : *
1570 : * This is unlikely given the size of the queue, but possible.
1571 : * The warnings show up at most once every QUEUE_FULL_WARN_INTERVAL.
1572 : *
1573 : * Caller must hold exclusive NotifyQueueLock.
1574 : */
1575 : static void
1576 83 : asyncQueueFillWarning(void)
1577 : {
1578 : double fillDegree;
1579 : TimestampTz t;
1580 :
1581 83 : fillDegree = asyncQueueUsage();
4800 tgl 1582 83 : if (fillDegree < 0.5)
1583 83 : return;
1584 :
4800 tgl 1585 UBC 0 : t = GetCurrentTimestamp();
1586 :
1587 0 : if (TimestampDifferenceExceeds(asyncQueueControl->lastQueueFillWarn,
1588 : t, QUEUE_FULL_WARN_INTERVAL))
1589 : {
4790 bruce 1590 0 : QueuePosition min = QUEUE_HEAD;
1591 0 : int32 minPid = InvalidPid;
1592 :
1307 tgl 1593 0 : for (BackendId i = QUEUE_FIRST_LISTENER; i > 0; i = QUEUE_NEXT_LISTENER(i))
1594 : {
1595 0 : Assert(QUEUE_BACKEND_PID(i) != InvalidPid);
1596 0 : min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
1597 0 : if (QUEUE_POS_EQUAL(min, QUEUE_BACKEND_POS(i)))
1598 0 : minPid = QUEUE_BACKEND_PID(i);
1599 : }
1600 :
4800 1601 0 : ereport(WARNING,
1602 : (errmsg("NOTIFY queue is %.0f%% full", fillDegree * 100),
1603 : (minPid != InvalidPid ?
1604 : errdetail("The server process with PID %d is among those with the oldest transactions.", minPid)
1605 : : 0),
1606 : (minPid != InvalidPid ?
1607 : errhint("The NOTIFY queue cannot be emptied until that process ends its current transaction.")
1608 : : 0)));
1609 :
1610 0 : asyncQueueControl->lastQueueFillWarn = t;
1611 : }
1612 : }
1613 :
1614 : /*
1615 : * Send signals to listening backends.
1616 : *
1617 : * Normally we signal only backends in our own database, since only those
1618 : * backends could be interested in notifies we send. However, if there's
1619 : * notify traffic in our database but no traffic in another database that
1620 : * does have listener(s), those listeners will fall further and further
1621 : * behind. Waken them anyway if they're far enough behind, so that they'll
1622 : * advance their queue position pointers, allowing the global tail to advance.
1623 : *
1624 : * Since we know the BackendId and the Pid the signaling is quite cheap.
1625 : *
1626 : * This is called during CommitTransaction(), so it's important for it
1627 : * to have very low probability of failure.
1628 : */
1629 : static void
4800 tgl 1630 CBC 48 : SignalBackends(void)
1631 : {
1632 : int32 *pids;
1633 : BackendId *ids;
1634 : int count;
1635 :
1636 : /*
1637 : * Identify backends that we need to signal. We don't want to send
1638 : * signals while holding the NotifyQueueLock, so this loop just builds a
1639 : * list of target PIDs.
1640 : *
1641 : * XXX in principle these pallocs could fail, which would be bad. Maybe
1642 : * preallocate the arrays? They're not that large, though.
1643 : */
362 rhaas 1644 48 : pids = (int32 *) palloc(MaxBackends * sizeof(int32));
1645 48 : ids = (BackendId *) palloc(MaxBackends * sizeof(BackendId));
4800 tgl 1646 48 : count = 0;
1647 :
1059 1648 48 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
1307 1649 88 : for (BackendId i = QUEUE_FIRST_LISTENER; i > 0; i = QUEUE_NEXT_LISTENER(i))
1650 : {
1295 1651 40 : int32 pid = QUEUE_BACKEND_PID(i);
1652 : QueuePosition pos;
1653 :
1307 1654 40 : Assert(pid != InvalidPid);
1295 1655 40 : pos = QUEUE_BACKEND_POS(i);
1656 40 : if (QUEUE_BACKEND_DBOID(i) == MyDatabaseId)
1657 : {
1658 : /*
1659 : * Always signal listeners in our own database, unless they're
1660 : * already caught up (unlikely, but possible).
1661 : */
1662 40 : if (QUEUE_POS_EQUAL(pos, QUEUE_HEAD))
1295 tgl 1663 UBC 0 : continue;
1664 : }
1665 : else
1666 : {
1667 : /*
1668 : * Listeners in other databases should be signaled only if they
1669 : * are far behind.
1670 : */
1671 0 : if (asyncQueuePageDiff(QUEUE_POS_PAGE(QUEUE_HEAD),
1672 : QUEUE_POS_PAGE(pos)) < QUEUE_CLEANUP_DELAY)
1673 0 : continue;
1674 : }
1675 : /* OK, need to signal this one */
1295 tgl 1676 CBC 40 : pids[count] = pid;
1677 40 : ids[count] = i;
1678 40 : count++;
1679 : }
1059 1680 48 : LWLockRelease(NotifyQueueLock);
1681 :
1682 : /* Now send signals */
1307 1683 88 : for (int i = 0; i < count; i++)
1684 : {
1295 1685 40 : int32 pid = pids[i];
1686 :
1687 : /*
1688 : * If we are signaling our own process, no need to involve the kernel;
1689 : * just set the flag directly.
1690 : */
572 1691 40 : if (pid == MyProcPid)
1692 : {
1693 20 : notifyInterruptPending = true;
1694 20 : continue;
1695 : }
1696 :
1697 : /*
1698 : * Note: assuming things aren't broken, a signal failure here could
1699 : * only occur if the target backend exited since we released
1700 : * NotifyQueueLock; which is unlikely but certainly possible. So we
1701 : * just log a low-level debug message if it happens.
1702 : */
4800 1703 20 : if (SendProcSignal(pid, PROCSIG_NOTIFY_INTERRUPT, ids[i]) < 0)
4800 tgl 1704 UBC 0 : elog(DEBUG3, "could not signal backend with PID %d: %m", pid);
1705 : }
1706 :
4800 tgl 1707 CBC 48 : pfree(pids);
1708 48 : pfree(ids);
8951 1709 48 : }
1710 :
1711 : /*
1712 : * AtAbort_Notify
1713 : *
1714 : * This is called at transaction abort.
1715 : *
1716 : * Gets rid of pending actions and outbound notifies that we would have
1717 : * executed if the transaction got committed.
1718 : */
1719 : void
7966 1720 20099 : AtAbort_Notify(void)
1721 : {
1722 : /*
1723 : * If we LISTEN but then roll back the transaction after PreCommit_Notify,
1724 : * we have registered as a listener but have not made any entry in
1725 : * listenChannels. In that case, deregister again.
1726 : */
3707 1727 20099 : if (amRegisteredListener && listenChannels == NIL)
3707 tgl 1728 UBC 0 : asyncQueueUnregister();
1729 :
1730 : /* And clean up */
5506 tgl 1731 CBC 20099 : ClearPendingActionsAndNotifies();
8951 1732 20099 : }
1733 :
1734 : /*
1735 : * AtSubCommit_Notify() --- Take care of subtransaction commit.
1736 : *
1737 : * Reassign all items in the pending lists to the parent transaction.
1738 : */
1739 : void
6856 1740 4317 : AtSubCommit_Notify(void)
1741 : {
1283 rhaas 1742 4317 : int my_level = GetCurrentTransactionNestLevel();
1743 :
1744 : /* If there are actions at our nesting level, we must reparent them. */
1745 4317 : if (pendingActions != NULL &&
1283 rhaas 1746 UBC 0 : pendingActions->nestingLevel >= my_level)
1747 : {
1748 0 : if (pendingActions->upper == NULL ||
1749 0 : pendingActions->upper->nestingLevel < my_level - 1)
1750 : {
1751 : /* nothing to merge; give the whole thing to the parent */
1752 0 : --pendingActions->nestingLevel;
1753 : }
1754 : else
1755 : {
1756 0 : ActionList *childPendingActions = pendingActions;
1757 :
1758 0 : pendingActions = pendingActions->upper;
1759 :
1760 : /*
1761 : * Mustn't try to eliminate duplicates here --- see queue_listen()
1762 : */
1763 0 : pendingActions->actions =
1764 0 : list_concat(pendingActions->actions,
1765 0 : childPendingActions->actions);
1766 0 : pfree(childPendingActions);
1767 : }
1768 : }
1769 :
1770 : /* If there are notifies at our nesting level, we must reparent them. */
1283 rhaas 1771 CBC 4317 : if (pendingNotifies != NULL &&
1772 2 : pendingNotifies->nestingLevel >= my_level)
1773 : {
1774 1 : Assert(pendingNotifies->nestingLevel == my_level);
1775 :
1776 1 : if (pendingNotifies->upper == NULL ||
1777 1 : pendingNotifies->upper->nestingLevel < my_level - 1)
1778 : {
1779 : /* nothing to merge; give the whole thing to the parent */
1283 rhaas 1780 UBC 0 : --pendingNotifies->nestingLevel;
1781 : }
1782 : else
1783 : {
1784 : /*
1785 : * Formerly, we didn't bother to eliminate duplicates here, but
1786 : * now we must, else we fall foul of "Assert(!found)", either here
1787 : * or during a later attempt to build the parent-level hashtable.
1788 : */
1283 rhaas 1789 CBC 1 : NotificationList *childPendingNotifies = pendingNotifies;
1790 : ListCell *l;
1791 :
1792 1 : pendingNotifies = pendingNotifies->upper;
1793 : /* Insert all the subxact's events into parent, except for dups */
1794 5 : foreach(l, childPendingNotifies->events)
1795 : {
1796 4 : Notification *childn = (Notification *) lfirst(l);
1797 :
1798 4 : if (!AsyncExistsPendingNotify(childn))
1799 2 : AddEventToPendingNotifies(childn);
1800 : }
1801 1 : pfree(childPendingNotifies);
1802 : }
1803 : }
6856 tgl 1804 4317 : }
1805 :
1806 : /*
1807 : * AtSubAbort_Notify() --- Take care of subtransaction abort.
1808 : */
1809 : void
1810 4468 : AtSubAbort_Notify(void)
1811 : {
6789 1812 4468 : int my_level = GetCurrentTransactionNestLevel();
1813 :
1814 : /*
1815 : * All we have to do is pop the stack --- the actions/notifies made in
1816 : * this subxact are no longer interesting, and the space will be freed
1817 : * when CurTransactionContext is recycled. We still have to free the
1818 : * ActionList and NotificationList objects themselves, though, because
1819 : * those are allocated in TopTransactionContext.
1820 : *
1821 : * Note that there might be no entries at all, or no entries for the
1822 : * current subtransaction level, either because none were ever created, or
1823 : * because we reentered this routine due to trouble during subxact abort.
1824 : */
1283 rhaas 1825 4468 : while (pendingActions != NULL &&
1283 rhaas 1826 UBC 0 : pendingActions->nestingLevel >= my_level)
1827 : {
1828 0 : ActionList *childPendingActions = pendingActions;
1829 :
1830 0 : pendingActions = pendingActions->upper;
1831 0 : pfree(childPendingActions);
1832 : }
1833 :
1283 rhaas 1834 CBC 4469 : while (pendingNotifies != NULL &&
1835 2 : pendingNotifies->nestingLevel >= my_level)
1836 : {
1837 1 : NotificationList *childPendingNotifies = pendingNotifies;
1838 :
1839 1 : pendingNotifies = pendingNotifies->upper;
1840 1 : pfree(childPendingNotifies);
1841 : }
6856 tgl 1842 4468 : }
1843 :
1844 : /*
1845 : * HandleNotifyInterrupt
1846 : *
1847 : * Signal handler portion of interrupt handling. Let the backend know
1848 : * that there's a pending notify interrupt. If we're currently reading
1849 : * from the client, this will interrupt the read and
1850 : * ProcessClientReadInterrupt() will call ProcessNotifyInterrupt().
1851 : */
1852 : void
5000 1853 20 : HandleNotifyInterrupt(void)
1854 : {
1855 : /*
1856 : * Note: this is called by a SIGNAL HANDLER. You must be very wary what
1857 : * you do here.
1858 : */
1859 :
1860 : /* signal that work needs to be done */
2987 andres 1861 20 : notifyInterruptPending = true;
1862 :
1863 : /* make sure the event is processed in due course */
1864 20 : SetLatch(MyLatch);
8951 tgl 1865 20 : }
1866 :
1867 : /*
1868 : * ProcessNotifyInterrupt
1869 : *
1870 : * This is called if we see notifyInterruptPending set, just before
1871 : * transmitting ReadyForQuery at the end of a frontend command, and
1872 : * also if a notify signal occurs while reading from the frontend.
1873 : * HandleNotifyInterrupt() will cause the read to be interrupted
1874 : * via the process's latch, and this routine will get called.
1875 : * If we are truly idle (ie, *not* inside a transaction block),
1876 : * process the incoming notifies.
1877 : *
1878 : * If "flush" is true, force any frontend messages out immediately.
1879 : * This can be false when being called at the end of a frontend command,
1880 : * since we'll flush after sending ReadyForQuery.
1881 : */
1882 : void
572 1883 95 : ProcessNotifyInterrupt(bool flush)
1884 : {
7115 1885 95 : if (IsTransactionOrTransactionBlock())
8951 1886 56 : return; /* not really idle */
1887 :
1888 : /* Loop in case another signal arrives while sending messages */
2987 andres 1889 78 : while (notifyInterruptPending)
572 tgl 1890 39 : ProcessIncomingNotify(flush);
1891 : }
1892 :
1893 :
1894 : /*
1895 : * Read all pending notifications from the queue, and deliver appropriate
1896 : * ones to my frontend. Stop when we reach queue head or an uncommitted
1897 : * notification.
1898 : */
1899 : static void
4800 1900 50 : asyncQueueReadAllNotifications(void)
1901 : {
1902 : volatile QueuePosition pos;
1903 : QueuePosition head;
1904 : Snapshot snapshot;
1905 :
1906 : /* page_buffer must be adequately aligned, so use a union */
1907 : union
1908 : {
1909 : char buf[QUEUE_PAGESIZE];
1910 : AsyncQueueEntry align;
1911 : } page_buffer;
1912 :
1913 : /* Fetch current state */
1059 1914 50 : LWLockAcquire(NotifyQueueLock, LW_SHARED);
1915 : /* Assert checks that we have a valid state entry */
4800 1916 50 : Assert(MyProcPid == QUEUE_BACKEND_PID(MyBackendId));
946 1917 50 : pos = QUEUE_BACKEND_POS(MyBackendId);
4800 1918 50 : head = QUEUE_HEAD;
1059 1919 50 : LWLockRelease(NotifyQueueLock);
1920 :
4800 1921 50 : if (QUEUE_POS_EQUAL(pos, head))
1922 : {
1923 : /* Nothing to do, we have read all notifications already. */
4800 tgl 1924 UBC 0 : return;
1925 : }
1926 :
1927 : /*----------
1928 : * Get snapshot we'll use to decide which xacts are still in progress.
1929 : * This is trickier than it might seem, because of race conditions.
1930 : * Consider the following example:
1931 : *
1932 : * Backend 1: Backend 2:
1933 : *
1934 : * transaction starts
1935 : * UPDATE foo SET ...;
1936 : * NOTIFY foo;
1937 : * commit starts
1938 : * queue the notify message
1939 : * transaction starts
1940 : * LISTEN foo; -- first LISTEN in session
1941 : * SELECT * FROM foo WHERE ...;
1942 : * commit to clog
1943 : * commit starts
1944 : * add backend 2 to array of listeners
1945 : * advance to queue head (this code)
1946 : * commit to clog
1947 : *
1948 : * Transaction 2's SELECT has not seen the UPDATE's effects, since that
1949 : * wasn't committed yet. Ideally we'd ensure that client 2 would
1950 : * eventually get transaction 1's notify message, but there's no way
1951 : * to do that; until we're in the listener array, there's no guarantee
1952 : * that the notify message doesn't get removed from the queue.
1953 : *
1954 : * Therefore the coding technique transaction 2 is using is unsafe:
1955 : * applications must commit a LISTEN before inspecting database state,
1956 : * if they want to ensure they will see notifications about subsequent
1957 : * changes to that state.
1958 : *
1959 : * What we do guarantee is that we'll see all notifications from
1960 : * transactions committing after the snapshot we take here.
1961 : * Exec_ListenPreCommit has already added us to the listener array,
1962 : * so no not-yet-committed messages can be removed from the queue
1963 : * before we see them.
1964 : *----------
1965 : */
1232 tgl 1966 CBC 50 : snapshot = RegisterSnapshot(GetLatestSnapshot());
1967 :
1968 : /*
1969 : * It is possible that we fail while trying to send a message to our
1970 : * frontend (for example, because of encoding conversion failure). If
1971 : * that happens it is critical that we not try to send the same message
1972 : * over and over again. Therefore, we place a PG_TRY block here that will
1973 : * forcibly advance our queue position before we lose control to an error.
1974 : * (We could alternatively retake NotifyQueueLock and move the position
1975 : * before handling each individual message, but that seems like too much
1976 : * lock traffic.)
1977 : */
4800 1978 50 : PG_TRY();
1979 : {
1980 : bool reachedStop;
1981 :
1982 : do
1983 : {
1984 85 : int curpage = QUEUE_POS_PAGE(pos);
1985 85 : int curoffset = QUEUE_POS_OFFSET(pos);
1986 : int slotno;
1987 : int copysize;
1988 :
1989 : /*
1990 : * We copy the data from SLRU into a local buffer, so as to avoid
1991 : * holding the NotifySLRULock while we are examining the entries
1992 : * and possibly transmitting them to our frontend. Copy only the
1993 : * part of the page we will actually inspect.
1994 : */
1059 1995 85 : slotno = SimpleLruReadPage_ReadOnly(NotifyCtl, curpage,
1996 : InvalidTransactionId);
4800 1997 85 : if (curpage == QUEUE_POS_PAGE(head))
1998 : {
1999 : /* we only want to read as far as head */
2000 50 : copysize = QUEUE_POS_OFFSET(head) - curoffset;
2001 50 : if (copysize < 0)
2118 tgl 2002 UBC 0 : copysize = 0; /* just for safety */
2003 : }
2004 : else
2005 : {
2006 : /* fetch all the rest of the page */
4800 tgl 2007 CBC 35 : copysize = QUEUE_PAGESIZE - curoffset;
2008 : }
2009 85 : memcpy(page_buffer.buf + curoffset,
1059 2010 85 : NotifyCtl->shared->page_buffer[slotno] + curoffset,
2011 : copysize);
2012 : /* Release lock that we got from SimpleLruReadPage_ReadOnly() */
2013 85 : LWLockRelease(NotifySLRULock);
2014 :
2015 : /*
2016 : * Process messages up to the stop position, end of page, or an
2017 : * uncommitted message.
2018 : *
2019 : * Our stop position is what we found to be the head's position
2020 : * when we entered this function. It might have changed already.
2021 : * But if it has, we will receive (or have already received and
2022 : * queued) another signal and come here again.
2023 : *
2024 : * We are not holding NotifyQueueLock here! The queue can only
2025 : * extend beyond the head pointer (see above) and we leave our
2026 : * backend's pointer where it is so nobody will truncate or
2027 : * rewrite pages under us. Especially we don't want to hold a lock
2028 : * while sending the notifications to the frontend.
2029 : */
4800 2030 85 : reachedStop = asyncQueueProcessPageEntries(&pos, head,
2031 : page_buffer.buf,
2032 : snapshot);
2033 85 : } while (!reachedStop);
2034 : }
1255 peter 2035 UBC 0 : PG_FINALLY();
2036 : {
2037 : /* Update shared state */
1059 tgl 2038 CBC 50 : LWLockAcquire(NotifyQueueLock, LW_SHARED);
4800 2039 50 : QUEUE_BACKEND_POS(MyBackendId) = pos;
1059 2040 50 : LWLockRelease(NotifyQueueLock);
2041 : }
4800 2042 50 : PG_END_TRY();
2043 :
2044 : /* Done with snapshot */
2006 2045 50 : UnregisterSnapshot(snapshot);
2046 : }
2047 :
2048 : /*
2049 : * Fetch notifications from the shared queue, beginning at position current,
2050 : * and deliver relevant ones to my frontend.
2051 : *
2052 : * The current page must have been fetched into page_buffer from shared
2053 : * memory. (We could access the page right in shared memory, but that
2054 : * would imply holding the NotifySLRULock throughout this routine.)
2055 : *
2056 : * We stop if we reach the "stop" position, or reach a notification from an
2057 : * uncommitted transaction, or reach the end of the page.
2058 : *
2059 : * The function returns true once we have reached the stop position or an
2060 : * uncommitted notification, and false if we have finished with the page.
2061 : * In other words: once it returns true there is no need to look further.
2062 : * The QueuePosition *current is advanced past all processed messages.
2063 : */
2064 : static bool
2995 2065 85 : asyncQueueProcessPageEntries(volatile QueuePosition *current,
2066 : QueuePosition stop,
2067 : char *page_buffer,
2068 : Snapshot snapshot)
2069 : {
4800 2070 85 : bool reachedStop = false;
2071 : bool reachedEndOfPage;
2072 : AsyncQueueEntry *qe;
2073 :
2074 : do
2075 : {
4790 bruce 2076 1242 : QueuePosition thisentry = *current;
2077 :
4799 tgl 2078 1242 : if (QUEUE_POS_EQUAL(thisentry, stop))
4800 2079 50 : break;
2080 :
4799 2081 1192 : qe = (AsyncQueueEntry *) (page_buffer + QUEUE_POS_OFFSET(thisentry));
2082 :
2083 : /*
2084 : * Advance *current over this message, possibly to the next page. As
2085 : * noted in the comments for asyncQueueReadAllNotifications, we must
2086 : * do this before possibly failing while processing the message.
2087 : */
4800 2088 1192 : reachedEndOfPage = asyncQueueAdvance(current, qe->length);
2089 :
2090 : /* Ignore messages destined for other databases */
2091 1192 : if (qe->dboid == MyDatabaseId)
2092 : {
2006 2093 1159 : if (XidInMVCCSnapshot(qe->xid, snapshot))
2094 : {
2095 : /*
2096 : * The source transaction is still in progress, so we can't
2097 : * process this message yet. Break out of the loop, but first
2098 : * back up *current so we will reprocess the message next
2099 : * time. (Note: it is unlikely but not impossible for
2100 : * TransactionIdDidCommit to fail, so we can't really avoid
2101 : * this advance-then-back-up behavior when dealing with an
2102 : * uncommitted message.)
2103 : *
2104 : * Note that we must test XidInMVCCSnapshot before we test
2105 : * TransactionIdDidCommit, else we might return a message from
2106 : * a transaction that is not yet visible to snapshots; compare
2107 : * the comments at the head of heapam_visibility.c.
2108 : *
2109 : * Also, while our own xact won't be listed in the snapshot,
2110 : * we need not check for TransactionIdIsCurrentTransactionId
2111 : * because our transaction cannot (yet) have queued any
2112 : * messages.
2113 : */
3314 tgl 2114 UBC 0 : *current = thisentry;
2115 0 : reachedStop = true;
2116 0 : break;
2117 : }
3314 tgl 2118 CBC 1159 : else if (TransactionIdDidCommit(qe->xid))
2119 : {
2120 : /* qe->data is the null-terminated channel name */
4790 bruce 2121 1159 : char *channel = qe->data;
2122 :
4800 tgl 2123 1159 : if (IsListeningOn(channel))
2124 : {
2125 : /* payload follows channel name */
4790 bruce 2126 31 : char *payload = qe->data + strlen(channel) + 1;
2127 :
4800 tgl 2128 31 : NotifyMyFrontEnd(channel, payload, qe->srcPid);
2129 : }
2130 : }
2131 : else
2132 : {
2133 : /*
2134 : * The source transaction aborted or crashed, so we just
2135 : * ignore its notifications.
2136 : */
2137 : }
2138 : }
2139 :
2140 : /* Loop back if we're not at end of page */
2141 1192 : } while (!reachedEndOfPage);
2142 :
2143 85 : if (QUEUE_POS_EQUAL(*current, stop))
2144 50 : reachedStop = true;
2145 :
2146 85 : return reachedStop;
2147 : }
2148 :
2149 : /*
2150 : * Advance the shared queue tail variable to the minimum of all the
2151 : * per-backend tail pointers. Truncate pg_notify space if possible.
2152 : *
2153 : * This is (usually) called during CommitTransaction(), so it's important for
2154 : * it to have very low probability of failure.
2155 : */
2156 : static void
2157 13 : asyncQueueAdvanceTail(void)
2158 : {
2159 : QueuePosition min;
2160 : int oldtailpage;
2161 : int newtailpage;
2162 : int boundary;
2163 :
2164 : /* Restrict task to one backend per cluster; see SimpleLruTruncate(). */
967 noah 2165 13 : LWLockAcquire(NotifyQueueTailLock, LW_EXCLUSIVE);
2166 :
2167 : /*
2168 : * Compute the new tail. Pre-v13, it's essential that QUEUE_TAIL be exact
2169 : * (ie, exactly match at least one backend's queue position), so it must
2170 : * be updated atomically with the actual computation. Since v13, we could
2171 : * get away with not doing it like that, but it seems prudent to keep it
2172 : * so.
2173 : *
2174 : * Also, because incoming backends will scan forward from QUEUE_TAIL, that
2175 : * must be advanced before we can truncate any data. Thus, QUEUE_TAIL is
2176 : * the logical tail, while QUEUE_STOP_PAGE is the physical tail, or oldest
2177 : * un-truncated page. When QUEUE_STOP_PAGE != QUEUE_POS_PAGE(QUEUE_TAIL),
2178 : * there are pages we can truncate but haven't yet finished doing so.
2179 : *
2180 : * For concurrency's sake, we don't want to hold NotifyQueueLock while
2181 : * performing SimpleLruTruncate. This is OK because no backend will try
2182 : * to access the pages we are in the midst of truncating.
2183 : */
1059 tgl 2184 13 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
4800 2185 13 : min = QUEUE_HEAD;
1307 2186 23 : for (BackendId i = QUEUE_FIRST_LISTENER; i > 0; i = QUEUE_NEXT_LISTENER(i))
2187 : {
2188 10 : Assert(QUEUE_BACKEND_PID(i) != InvalidPid);
2189 10 : min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
2190 : }
862 2191 13 : QUEUE_TAIL = min;
2192 13 : oldtailpage = QUEUE_STOP_PAGE;
1059 2193 13 : LWLockRelease(NotifyQueueLock);
2194 :
2195 : /*
2196 : * We can truncate something if the global tail advanced across an SLRU
2197 : * segment boundary.
2198 : *
2199 : * XXX it might be better to truncate only once every several segments, to
2200 : * reduce the number of directory scans.
2201 : */
4800 2202 13 : newtailpage = QUEUE_POS_PAGE(min);
2203 13 : boundary = newtailpage - (newtailpage % SLRU_PAGES_PER_SEGMENT);
4211 alvherre 2204 13 : if (asyncQueuePagePrecedes(oldtailpage, boundary))
2205 : {
2206 : /*
2207 : * SimpleLruTruncate() will ask for NotifySLRULock but will also
2208 : * release the lock again.
2209 : */
1059 tgl 2210 UBC 0 : SimpleLruTruncate(NotifyCtl, newtailpage);
2211 :
2212 : /*
2213 : * Update QUEUE_STOP_PAGE. This changes asyncQueueIsFull()'s verdict
2214 : * for the segment immediately prior to the old tail, allowing fresh
2215 : * data into that segment.
2216 : */
862 2217 0 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
2218 0 : QUEUE_STOP_PAGE = newtailpage;
2219 0 : LWLockRelease(NotifyQueueLock);
2220 : }
2221 :
967 noah 2222 CBC 13 : LWLockRelease(NotifyQueueTailLock);
4800 tgl 2223 13 : }
2224 :
2225 : /*
2226 : * ProcessIncomingNotify
2227 : *
2228 : * Scan the queue for arriving notifications and report them to the front
2229 : * end. The notifications might be from other sessions, or our own;
2230 : * there's no need to distinguish here.
2231 : *
2232 : * If "flush" is true, force any frontend messages out immediately.
2233 : *
2234 : * NOTE: since we are outside any transaction, we must create our own.
2235 : */
2236 : static void
572 2237 39 : ProcessIncomingNotify(bool flush)
2238 : {
2239 : /* We *must* reset the flag */
2987 andres 2240 39 : notifyInterruptPending = false;
2241 :
2242 : /* Do nothing else if we aren't actively listening */
4800 tgl 2243 39 : if (listenChannels == NIL)
4800 tgl 2244 UBC 0 : return;
2245 :
8348 peter_e 2246 CBC 39 : if (Trace_notify)
7257 bruce 2247 UBC 0 : elog(DEBUG1, "ProcessIncomingNotify");
2248 :
1124 peter 2249 CBC 39 : set_ps_display("notify interrupt");
2250 :
2251 : /*
2252 : * We must run asyncQueueReadAllNotifications inside a transaction, else
2253 : * bad things happen if it gets an error.
2254 : */
4800 tgl 2255 39 : StartTransactionCommand();
2256 :
2257 39 : asyncQueueReadAllNotifications();
2258 :
7270 2259 39 : CommitTransactionCommand();
2260 :
2261 : /*
2262 : * If this isn't an end-of-command case, we must flush the notify messages
2263 : * to ensure frontend gets them promptly.
2264 : */
572 2265 39 : if (flush)
2266 10 : pq_flush();
2267 :
1124 peter 2268 39 : set_ps_display("idle");
2269 :
8348 peter_e 2270 39 : if (Trace_notify)
7257 bruce 2271 UBC 0 : elog(DEBUG1, "ProcessIncomingNotify: done");
2272 : }
2273 :
2274 : /*
2275 : * Send NOTIFY message to my front end.
2276 : */
2277 : void
4800 tgl 2278 CBC 31 : NotifyMyFrontEnd(const char *channel, const char *payload, int32 srcPid)
2279 : {
6366 alvherre 2280 31 : if (whereToSendOutput == DestRemote)
2281 : {
2282 : StringInfoData buf;
2283 :
7292 tgl 2284 31 : pq_beginmessage(&buf, 'A');
2006 andres 2285 31 : pq_sendint32(&buf, srcPid);
4800 tgl 2286 31 : pq_sendstring(&buf, channel);
766 heikki.linnakangas 2287 31 : pq_sendstring(&buf, payload);
8750 tgl 2288 31 : pq_endmessage(&buf);
2289 :
2290 : /*
2291 : * NOTE: we do not do pq_flush() here. Some level of caller will
2292 : * handle it later, allowing this message to be combined into a packet
2293 : * with other ones.
2294 : */
2295 : }
2296 : else
4800 tgl 2297 UBC 0 : elog(INFO, "NOTIFY for \"%s\" payload \"%s\"", channel, payload);
8951 tgl 2298 CBC 31 : }
2299 :
2300 : /* Does pendingNotifies include a match for the given event? */
2301 : static bool
1333 2302 1048 : AsyncExistsPendingNotify(Notification *n)
2303 : {
2304 1048 : if (pendingNotifies == NULL)
4800 tgl 2305 UBC 0 : return false;
2306 :
1333 tgl 2307 CBC 1048 : if (pendingNotifies->hashtab != NULL)
2308 : {
2309 : /* Use the hash table to probe for a match */
2310 983 : if (hash_search(pendingNotifies->hashtab,
2311 : &n,
2312 : HASH_FIND,
2313 : NULL))
1333 tgl 2314 UBC 0 : return true;
2315 : }
2316 : else
2317 : {
2318 : /* Must scan the event list */
2319 : ListCell *l;
2320 :
1333 tgl 2321 CBC 271 : foreach(l, pendingNotifies->events)
2322 : {
2323 220 : Notification *oldn = (Notification *) lfirst(l);
2324 :
2325 220 : if (n->channel_len == oldn->channel_len &&
2326 220 : n->payload_len == oldn->payload_len &&
2327 125 : memcmp(n->data, oldn->data,
2328 125 : n->channel_len + n->payload_len + 2) == 0)
2329 14 : return true;
2330 : }
2331 : }
2332 :
2333 1034 : return false;
2334 : }
2335 :
2336 : /*
2337 : * Add a notification event to a pre-existing pendingNotifies list.
2338 : *
2339 : * Because pendingNotifies->events is already nonempty, this works
2340 : * correctly no matter what CurrentMemoryContext is.
2341 : */
2342 : static void
2343 1034 : AddEventToPendingNotifies(Notification *n)
2344 : {
2345 1034 : Assert(pendingNotifies->events != NIL);
2346 :
2347 : /* Create the hash table if it's time to */
2348 1034 : if (list_length(pendingNotifies->events) >= MIN_HASHABLE_NOTIFIES &&
2349 984 : pendingNotifies->hashtab == NULL)
2350 : {
2351 : HASHCTL hash_ctl;
2352 : ListCell *l;
2353 :
2354 : /* Create the hash table */
2355 1 : hash_ctl.keysize = sizeof(Notification *);
2356 1 : hash_ctl.entrysize = sizeof(NotificationHash);
2357 1 : hash_ctl.hash = notification_hash;
2358 1 : hash_ctl.match = notification_match;
2359 1 : hash_ctl.hcxt = CurTransactionContext;
2360 2 : pendingNotifies->hashtab =
2361 1 : hash_create("Pending Notifies",
2362 : 256L,
2363 : &hash_ctl,
2364 : HASH_ELEM | HASH_FUNCTION | HASH_COMPARE | HASH_CONTEXT);
2365 :
2366 : /* Insert all the already-existing events */
2367 17 : foreach(l, pendingNotifies->events)
2368 : {
2369 16 : Notification *oldn = (Notification *) lfirst(l);
2370 : NotificationHash *hentry;
2371 : bool found;
2372 :
2373 16 : hentry = (NotificationHash *) hash_search(pendingNotifies->hashtab,
2374 : &oldn,
2375 : HASH_ENTER,
2376 : &found);
2377 16 : Assert(!found);
2378 16 : hentry->event = oldn;
2379 : }
2380 : }
2381 :
2382 : /* Add new event to the list, in order */
2383 1034 : pendingNotifies->events = lappend(pendingNotifies->events, n);
2384 :
2385 : /* Add event to the hash table if needed */
2386 1034 : if (pendingNotifies->hashtab != NULL)
2387 : {
2388 : NotificationHash *hentry;
2389 : bool found;
2390 :
2391 984 : hentry = (NotificationHash *) hash_search(pendingNotifies->hashtab,
2392 : &n,
2393 : HASH_ENTER,
2394 : &found);
2395 984 : Assert(!found);
2396 984 : hentry->event = n;
2397 : }
2398 1034 : }
2399 :
2400 : /*
2401 : * notification_hash: hash function for notification hash table
2402 : *
2403 : * The hash "keys" are pointers to Notification structs.
2404 : */
2405 : static uint32
2406 1983 : notification_hash(const void *key, Size keysize)
2407 : {
2408 1983 : const Notification *k = *(const Notification *const *) key;
2409 :
2410 1983 : Assert(keysize == sizeof(Notification *));
2411 : /* We don't bother to include the payload's trailing null in the hash */
2412 1983 : return DatumGetUInt32(hash_any((const unsigned char *) k->data,
2413 1983 : k->channel_len + k->payload_len + 1));
2414 : }
2415 :
2416 : /*
2417 : * notification_match: match function to use with notification_hash
2418 : */
2419 : static int
1333 tgl 2420 UBC 0 : notification_match(const void *key1, const void *key2, Size keysize)
2421 : {
2422 0 : const Notification *k1 = *(const Notification *const *) key1;
2423 0 : const Notification *k2 = *(const Notification *const *) key2;
2424 :
2425 0 : Assert(keysize == sizeof(Notification *));
2426 0 : if (k1->channel_len == k2->channel_len &&
2427 0 : k1->payload_len == k2->payload_len &&
2428 0 : memcmp(k1->data, k2->data,
2429 0 : k1->channel_len + k1->payload_len + 2) == 0)
2430 0 : return 0; /* equal */
2431 0 : return 1; /* not equal */
2432 : }
2433 :
2434 : /* Clear the pendingActions and pendingNotifies lists. */
2435 : static void
5506 tgl 2436 CBC 20196 : ClearPendingActionsAndNotifies(void)
2437 : {
2438 : /*
2439 : * Everything's allocated in either TopTransactionContext or the context
2440 : * for the subtransaction to which it corresponds. So, there's nothing to
2441 : * do here except reset the pointers; the space will be reclaimed when the
2442 : * contexts are deleted.
2443 : */
1283 rhaas 2444 20196 : pendingActions = NULL;
1333 tgl 2445 20196 : pendingNotifies = NULL;
9345 bruce 2446 20196 : }
|
