Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * async.c
4 : * Asynchronous notification: NOTIFY, LISTEN, UNLISTEN
5 : *
6 : * Portions Copyright (c) 1996-2026, 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 v19:
17 : *
18 : * 1. Multiple backends on same machine. Multiple backends may be listening
19 : * on each of several channels.
20 : *
21 : * 2. There is one central queue in disk-based storage (directory pg_notify/),
22 : * with actively-used pages mapped into shared memory by the slru.c module.
23 : * All notification messages are placed in the queue and later read out
24 : * by listening backends. The single queue allows us to guarantee that
25 : * notifications are received in commit order.
26 : *
27 : * Although there is only one queue, notifications are treated as being
28 : * database-local; this is done by including the sender's database OID
29 : * in each notification message. Listening backends ignore messages
30 : * that don't match their database OID. This is important because it
31 : * ensures senders and receivers have the same database encoding and won't
32 : * misinterpret non-ASCII text in the channel name or payload string.
33 : *
34 : * Since notifications are not expected to survive database crashes,
35 : * we can simply clean out the pg_notify data at any reboot, and there
36 : * is no need for WAL support or fsync'ing.
37 : *
38 : * 3. Every backend that is listening on at least one channel registers by
39 : * entering its PID into the array in AsyncQueueControl. It then scans all
40 : * incoming notifications in the central queue and first compares the
41 : * database OID of the notification with its own database OID and then
42 : * compares the notified channel with the list of channels that it listens
43 : * to. In case there is a match it delivers the notification event to its
44 : * frontend. Non-matching events are simply skipped.
45 : *
46 : * 4. The NOTIFY statement (routine Async_Notify) stores the notification in
47 : * a backend-local list which will not be processed until transaction end.
48 : *
49 : * Duplicate notifications from the same transaction are sent out as one
50 : * notification only. This is done to save work when for example a trigger
51 : * on a 2 million row table fires a notification for each row that has been
52 : * changed. If the application needs to receive every single notification
53 : * that has been sent, it can easily add some unique string into the extra
54 : * payload parameter.
55 : *
56 : * When the transaction is ready to commit, PreCommit_Notify() adds the
57 : * pending notifications to the head of the queue. The head pointer of the
58 : * queue always points to the next free position and a position is just a
59 : * page number and the offset in that page. This is done before marking the
60 : * transaction as committed in clog. If we run into problems writing the
61 : * notifications, we can still call elog(ERROR, ...) and the transaction
62 : * will roll back safely.
63 : *
64 : * Once we have put all of the notifications into the queue, we return to
65 : * CommitTransaction() which will then do the actual transaction commit.
66 : *
67 : * After commit we are called another time (AtCommit_Notify()). Here we
68 : * make any required updates to the effective listen state (see below).
69 : * Then we signal any backends that may be interested in our messages
70 : * (including our own backend, if listening). This is done by
71 : * SignalBackends(), which sends a PROCSIG_NOTIFY_INTERRUPT signal to
72 : * each relevant backend, as described below.
73 : *
74 : * Finally, after we are out of the transaction altogether and about to go
75 : * idle, we scan the queue for messages that need to be sent to our
76 : * frontend (which might be notifies from other backends, or self-notifies
77 : * from our own). This step is not part of the CommitTransaction sequence
78 : * for two important reasons. First, we could get errors while sending
79 : * data to our frontend, and it's really bad for errors to happen in
80 : * post-commit cleanup. Second, in cases where a procedure issues commits
81 : * within a single frontend command, we don't want to send notifies to our
82 : * frontend until the command is done; but notifies to other backends
83 : * should go out immediately after each commit.
84 : *
85 : * 5. Upon receipt of a PROCSIG_NOTIFY_INTERRUPT signal, the signal handler
86 : * sets the process's latch, which triggers the event to be processed
87 : * immediately if this backend is idle (i.e., it is waiting for a frontend
88 : * command and is not within a transaction block. C.f.
89 : * ProcessClientReadInterrupt()). Otherwise the handler may only set a
90 : * flag, which will cause the processing to occur just before we next go
91 : * idle.
92 : *
93 : * Inbound-notify processing consists of reading all of the notifications
94 : * that have arrived since scanning last time. We read every notification
95 : * until we reach either a notification from an uncommitted transaction or
96 : * the head pointer's position.
97 : *
98 : * 6. To limit disk space consumption, the tail pointer needs to be advanced
99 : * so that old pages can be truncated. This is relatively expensive
100 : * (notably, it requires an exclusive lock), so we don't want to do it
101 : * often. We make sending backends do this work if they advanced the queue
102 : * head into a new page, but only once every QUEUE_CLEANUP_DELAY pages.
103 : *
104 : * 7. So far we have not discussed how backends change their listening state,
105 : * nor how notification senders know which backends to awaken. To handle
106 : * the latter, we maintain a global channel table (implemented as a dynamic
107 : * shared hash table, or dshash) that maps channel names to the set of
108 : * backends listening on each channel. This table is created lazily on the
109 : * first LISTEN command and grows dynamically as needed. There is also a
110 : * local channel table (a plain dynahash table) in each listening backend,
111 : * tracking which channels that backend is listening to. The local table
112 : * serves to reduce the number of accesses needed to the shared table.
113 : *
114 : * If the current transaction has executed any LISTEN/UNLISTEN actions,
115 : * PreCommit_Notify() prepares to commit those. For LISTEN, it
116 : * pre-allocates entries in both the per-backend localChannelTable and the
117 : * shared globalChannelTable, marking new shared entries removeOnAbort.
118 : * It also records the final per-channel intent in pendingListenActions,
119 : * so post-commit/abort processing can apply that in a single step.
120 : * Since all these allocations happen before committing to clog, we can
121 : * safely abort the transaction on failure.
122 : *
123 : * After commit, AtCommit_Notify() runs through pendingListenActions and
124 : * applies the final per-channel listen/unlisten state. This happens
125 : * before sending signals.
126 : *
127 : * SignalBackends() consults the shared global channel table to identify
128 : * listeners for the channels that the current transaction sent
129 : * notification(s) to. Each selected backend is marked as having a wakeup
130 : * pending to avoid duplicate signals, and a PROCSIG_NOTIFY_INTERRUPT
131 : * signal is sent to it.
132 : *
133 : * 8. While writing notifications, PreCommit_Notify() records the queue head
134 : * position both before and after the write. Because all writers serialize
135 : * on a cluster-wide heavyweight lock, no other backend can insert entries
136 : * between these two points. SignalBackends() uses this fact to directly
137 : * advance the queue pointer for any backend that is still positioned at
138 : * the old head, or within the range written, but is not interested in any
139 : * of our notifications. This avoids unnecessary wakeups for idle
140 : * listeners that have nothing to read. Backends that are not interested
141 : * in our notifications, but cannot be directly advanced, are signaled only
142 : * if they are far behind the current queue head; that is to ensure that
143 : * we can advance the queue tail without undue delay.
144 : *
145 : * An application that listens on the same channel it notifies will get
146 : * NOTIFY messages for its own NOTIFYs. These can be ignored, if not useful,
147 : * by comparing be_pid in the NOTIFY message to the application's own backend's
148 : * PID. (As of FE/BE protocol 2.0, the backend's PID is provided to the
149 : * frontend during startup.) The above design guarantees that notifies from
150 : * other backends will never be missed by ignoring self-notifies.
151 : *
152 : * The amount of shared memory used for notify management (notify_buffers)
153 : * can be varied without affecting anything but performance. The maximum
154 : * amount of notification data that can be queued at one time is determined
155 : * by the max_notify_queue_pages GUC.
156 : *-------------------------------------------------------------------------
157 : */
158 :
159 : #include "postgres.h"
160 :
161 : #include <limits.h>
162 : #include <unistd.h>
163 : #include <signal.h>
164 :
165 : #include "access/parallel.h"
166 : #include "access/slru.h"
167 : #include "access/transam.h"
168 : #include "access/xact.h"
169 : #include "catalog/pg_database.h"
170 : #include "commands/async.h"
171 : #include "common/hashfn.h"
172 : #include "funcapi.h"
173 : #include "lib/dshash.h"
174 : #include "libpq/libpq.h"
175 : #include "libpq/pqformat.h"
176 : #include "miscadmin.h"
177 : #include "storage/dsm_registry.h"
178 : #include "storage/ipc.h"
179 : #include "storage/latch.h"
180 : #include "storage/lmgr.h"
181 : #include "storage/procsignal.h"
182 : #include "storage/subsystems.h"
183 : #include "tcop/tcopprot.h"
184 : #include "utils/builtins.h"
185 : #include "utils/dsa.h"
186 : #include "utils/guc_hooks.h"
187 : #include "utils/memutils.h"
188 : #include "utils/ps_status.h"
189 : #include "utils/snapmgr.h"
190 : #include "utils/timestamp.h"
191 :
192 :
193 : /*
194 : * Maximum size of a NOTIFY payload, including terminating NULL. This
195 : * must be kept small enough so that a notification message fits on one
196 : * SLRU page. The magic fudge factor here is noncritical as long as it's
197 : * more than AsyncQueueEntryEmptySize --- we make it significantly bigger
198 : * than that, so changes in that data structure won't affect user-visible
199 : * restrictions.
200 : */
201 : #define NOTIFY_PAYLOAD_MAX_LENGTH (BLCKSZ - NAMEDATALEN - 128)
202 :
203 : /*
204 : * Struct representing an entry in the global notify queue
205 : *
206 : * This struct declaration has the maximal length, but in a real queue entry
207 : * the data area is only big enough for the actual channel and payload strings
208 : * (each null-terminated). AsyncQueueEntryEmptySize is the minimum possible
209 : * entry size, if both channel and payload strings are empty (but note it
210 : * doesn't include alignment padding).
211 : *
212 : * The "length" field should always be rounded up to the next QUEUEALIGN
213 : * multiple so that all fields are properly aligned.
214 : */
215 : typedef struct AsyncQueueEntry
216 : {
217 : int length; /* total allocated length of entry */
218 : Oid dboid; /* sender's database OID */
219 : TransactionId xid; /* sender's XID */
220 : int32 srcPid; /* sender's PID */
221 : char data[NAMEDATALEN + NOTIFY_PAYLOAD_MAX_LENGTH];
222 : } AsyncQueueEntry;
223 :
224 : /* Currently, no field of AsyncQueueEntry requires more than int alignment */
225 : #define QUEUEALIGN(len) INTALIGN(len)
226 :
227 : #define AsyncQueueEntryEmptySize (offsetof(AsyncQueueEntry, data) + 2)
228 :
229 : /*
230 : * Struct describing a queue position, and assorted macros for working with it
231 : */
232 : typedef struct QueuePosition
233 : {
234 : int64 page; /* SLRU page number */
235 : int offset; /* byte offset within page */
236 : } QueuePosition;
237 :
238 : #define QUEUE_POS_PAGE(x) ((x).page)
239 : #define QUEUE_POS_OFFSET(x) ((x).offset)
240 :
241 : #define SET_QUEUE_POS(x,y,z) \
242 : do { \
243 : (x).page = (y); \
244 : (x).offset = (z); \
245 : } while (0)
246 :
247 : #define QUEUE_POS_EQUAL(x,y) \
248 : ((x).page == (y).page && (x).offset == (y).offset)
249 :
250 : #define QUEUE_POS_IS_ZERO(x) \
251 : ((x).page == 0 && (x).offset == 0)
252 :
253 : /* choose logically smaller QueuePosition */
254 : #define QUEUE_POS_MIN(x,y) \
255 : (asyncQueuePagePrecedes((x).page, (y).page) ? (x) : \
256 : (x).page != (y).page ? (y) : \
257 : (x).offset < (y).offset ? (x) : (y))
258 :
259 : /* choose logically larger QueuePosition */
260 : #define QUEUE_POS_MAX(x,y) \
261 : (asyncQueuePagePrecedes((x).page, (y).page) ? (y) : \
262 : (x).page != (y).page ? (x) : \
263 : (x).offset > (y).offset ? (x) : (y))
264 :
265 : /* returns true if x comes before y in queue order */
266 : #define QUEUE_POS_PRECEDES(x,y) \
267 : (asyncQueuePagePrecedes((x).page, (y).page) || \
268 : ((x).page == (y).page && (x).offset < (y).offset))
269 :
270 : /*
271 : * Parameter determining how often we try to advance the tail pointer:
272 : * we do that after every QUEUE_CLEANUP_DELAY pages of NOTIFY data. This is
273 : * also the distance by which a backend that's not interested in our
274 : * notifications needs to be behind before we'll decide we need to wake it
275 : * up so it can advance its pointer.
276 : *
277 : * Resist the temptation to make this really large. While that would save
278 : * work in some places, it would add cost in others. In particular, this
279 : * should likely be less than notify_buffers, to ensure that backends
280 : * catch up before the pages they'll need to read fall out of SLRU cache.
281 : */
282 : #define QUEUE_CLEANUP_DELAY 4
283 :
284 : /*
285 : * Struct describing a listening backend's status
286 : */
287 : typedef struct QueueBackendStatus
288 : {
289 : int32 pid; /* either a PID or InvalidPid */
290 : Oid dboid; /* backend's database OID, or InvalidOid */
291 : ProcNumber nextListener; /* id of next listener, or INVALID_PROC_NUMBER */
292 : QueuePosition pos; /* backend has read queue up to here */
293 : bool wakeupPending; /* signal sent to backend, not yet processed */
294 : bool isAdvancing; /* backend is advancing its position */
295 : } QueueBackendStatus;
296 :
297 : /*
298 : * Shared memory state for LISTEN/NOTIFY (excluding its SLRU stuff)
299 : *
300 : * The AsyncQueueControl structure is protected by the NotifyQueueLock and
301 : * NotifyQueueTailLock.
302 : *
303 : * When holding NotifyQueueLock in SHARED mode, backends may only inspect
304 : * their own entries as well as the head and tail pointers. Consequently we
305 : * can allow a backend to update its own record while holding only SHARED lock
306 : * (since no other backend will inspect it).
307 : *
308 : * When holding NotifyQueueLock in EXCLUSIVE mode, backends can inspect the
309 : * entries of other backends and also change the head pointer. They can
310 : * also advance other backends' queue positions, unless the other backend
311 : * has isAdvancing set (i.e., is in process of doing that itself).
312 : *
313 : * When holding both NotifyQueueLock and NotifyQueueTailLock in EXCLUSIVE
314 : * mode, backends can change the tail pointers.
315 : *
316 : * SLRU buffer pool is divided in banks and bank wise SLRU lock is used as
317 : * the control lock for the pg_notify SLRU buffers.
318 : * In order to avoid deadlocks, whenever we need multiple locks, we first get
319 : * NotifyQueueTailLock, then NotifyQueueLock, then SLRU bank lock, and lastly
320 : * globalChannelTable partition locks.
321 : *
322 : * Each backend uses the backend[] array entry with index equal to its
323 : * ProcNumber. We rely on this to make SendProcSignal fast.
324 : *
325 : * The backend[] array entries for actively-listening backends are threaded
326 : * together using firstListener and the nextListener links, so that we can
327 : * scan them without having to iterate over inactive entries. We keep this
328 : * list in order by ProcNumber so that the scan is cache-friendly when there
329 : * are many active entries.
330 : */
331 : typedef struct AsyncQueueControl
332 : {
333 : QueuePosition head; /* head points to the next free location */
334 : QueuePosition tail; /* tail must be <= the queue position of every
335 : * listening backend */
336 : int64 stopPage; /* oldest unrecycled page; must be <=
337 : * tail.page */
338 : ProcNumber firstListener; /* id of first listener, or
339 : * INVALID_PROC_NUMBER */
340 : TimestampTz lastQueueFillWarn; /* time of last queue-full msg */
341 : dsa_handle globalChannelTableDSA; /* global channel table's DSA handle */
342 : dshash_table_handle globalChannelTableDSH; /* and its dshash handle */
343 : /* Array with room for MaxBackends entries: */
344 : QueueBackendStatus backend[FLEXIBLE_ARRAY_MEMBER];
345 : } AsyncQueueControl;
346 :
347 : static AsyncQueueControl *asyncQueueControl;
348 :
349 : static void AsyncShmemRequest(void *arg);
350 : static void AsyncShmemInit(void *arg);
351 :
352 : const ShmemCallbacks AsyncShmemCallbacks = {
353 : .request_fn = AsyncShmemRequest,
354 : .init_fn = AsyncShmemInit,
355 : };
356 :
357 :
358 : #define QUEUE_HEAD (asyncQueueControl->head)
359 : #define QUEUE_TAIL (asyncQueueControl->tail)
360 : #define QUEUE_STOP_PAGE (asyncQueueControl->stopPage)
361 : #define QUEUE_FIRST_LISTENER (asyncQueueControl->firstListener)
362 : #define QUEUE_BACKEND_PID(i) (asyncQueueControl->backend[i].pid)
363 : #define QUEUE_BACKEND_DBOID(i) (asyncQueueControl->backend[i].dboid)
364 : #define QUEUE_NEXT_LISTENER(i) (asyncQueueControl->backend[i].nextListener)
365 : #define QUEUE_BACKEND_POS(i) (asyncQueueControl->backend[i].pos)
366 : #define QUEUE_BACKEND_WAKEUP_PENDING(i) (asyncQueueControl->backend[i].wakeupPending)
367 : #define QUEUE_BACKEND_IS_ADVANCING(i) (asyncQueueControl->backend[i].isAdvancing)
368 :
369 : /*
370 : * The SLRU buffer area through which we access the notification queue
371 : */
372 : static inline bool asyncQueuePagePrecedes(int64 p, int64 q);
373 : static int asyncQueueErrdetailForIoError(const void *opaque_data);
374 :
375 : static SlruDesc NotifySlruDesc;
376 :
377 :
378 : #define NotifyCtl (&NotifySlruDesc)
379 : #define QUEUE_PAGESIZE BLCKSZ
380 :
381 : #define QUEUE_FULL_WARN_INTERVAL 5000 /* warn at most once every 5s */
382 :
383 : /*
384 : * Global channel table definitions
385 : *
386 : * This hash table maps (database OID, channel name) keys to arrays of
387 : * ProcNumbers representing the backends listening or about to listen
388 : * on each channel. The removeOnAbort flags allow us to create hash table
389 : * entries pre-commit and not have to assume that creating them post-commit
390 : * will succeed.
391 : */
392 : #define INITIAL_LISTENERS_ARRAY_SIZE 4
393 :
394 : typedef struct GlobalChannelKey
395 : {
396 : Oid dboid;
397 : char channel[NAMEDATALEN];
398 : } GlobalChannelKey;
399 :
400 : typedef struct ListenerEntry
401 : {
402 : ProcNumber procNo; /* listener's ProcNumber */
403 : bool removeOnAbort; /* remove entry if current xact aborts */
404 : } ListenerEntry;
405 :
406 : typedef struct GlobalChannelEntry
407 : {
408 : GlobalChannelKey key; /* hash key */
409 : dsa_pointer listenersArray; /* DSA pointer to ListenerEntry array */
410 : int numListeners; /* Number of listeners currently stored */
411 : int allocatedListeners; /* Allocated size of array */
412 : } GlobalChannelEntry;
413 :
414 : static dshash_table *globalChannelTable = NULL;
415 : static dsa_area *globalChannelDSA = NULL;
416 :
417 : /*
418 : * localChannelTable caches the channel names this backend is listening on
419 : * (including those we have staged to be listened on, but not yet committed).
420 : * Used by IsListeningOn() for fast lookups when reading notifications.
421 : */
422 : static HTAB *localChannelTable = NULL;
423 :
424 : /* We test this condition to detect that we're not listening at all */
425 : #define LocalChannelTableIsEmpty() \
426 : (localChannelTable == NULL || hash_get_num_entries(localChannelTable) == 0)
427 :
428 : /*
429 : * State for pending LISTEN/UNLISTEN actions consists of an ordered list of
430 : * all actions requested in the current transaction. As explained above,
431 : * we don't actually change listen state until we reach transaction commit.
432 : *
433 : * The list is kept in CurTransactionContext. In subtransactions, each
434 : * subtransaction has its own list in its own CurTransactionContext, but
435 : * successful subtransactions attach their lists to their parent's list.
436 : * Failed subtransactions simply discard their lists.
437 : */
438 : typedef enum
439 : {
440 : LISTEN_LISTEN,
441 : LISTEN_UNLISTEN,
442 : LISTEN_UNLISTEN_ALL,
443 : } ListenActionKind;
444 :
445 : typedef struct
446 : {
447 : ListenActionKind action;
448 : char channel[FLEXIBLE_ARRAY_MEMBER]; /* nul-terminated string */
449 : } ListenAction;
450 :
451 : typedef struct ActionList
452 : {
453 : int nestingLevel; /* current transaction nesting depth */
454 : List *actions; /* list of ListenAction structs */
455 : struct ActionList *upper; /* details for upper transaction levels */
456 : } ActionList;
457 :
458 : static ActionList *pendingActions = NULL;
459 :
460 : /*
461 : * Hash table recording the final listen/unlisten intent per channel for
462 : * the current transaction. Key is channel name, value is PENDING_LISTEN or
463 : * PENDING_UNLISTEN. This keeps critical commit/abort processing to one step
464 : * per channel instead of replaying every action. This is built from the
465 : * pendingActions list by PreCommit_Notify, then used by AtCommit_Notify or
466 : * AtAbort_Notify.
467 : */
468 : typedef enum
469 : {
470 : PENDING_LISTEN,
471 : PENDING_UNLISTEN,
472 : } PendingListenAction;
473 :
474 : typedef struct PendingListenEntry
475 : {
476 : char channel[NAMEDATALEN]; /* hash key */
477 : PendingListenAction action; /* which action should we perform? */
478 : } PendingListenEntry;
479 :
480 : static HTAB *pendingListenActions = NULL;
481 :
482 : /*
483 : * State for outbound notifies consists of a list of all channels+payloads
484 : * NOTIFYed in the current transaction. We do not actually perform a NOTIFY
485 : * until and unless the transaction commits. pendingNotifies is NULL if no
486 : * NOTIFYs have been done in the current (sub) transaction.
487 : *
488 : * We discard duplicate notify events issued in the same transaction.
489 : * Hence, in addition to the list proper (which we need to track the order
490 : * of the events, since we guarantee to deliver them in order), we build a
491 : * hash table which we can probe to detect duplicates. Since building the
492 : * hash table is somewhat expensive, we do so only once we have at least
493 : * MIN_HASHABLE_NOTIFIES events queued in the current (sub) transaction;
494 : * before that we just scan the events linearly.
495 : *
496 : * The list is kept in CurTransactionContext. In subtransactions, each
497 : * subtransaction has its own list in its own CurTransactionContext, but
498 : * successful subtransactions add their entries to their parent's list.
499 : * Failed subtransactions simply discard their lists. Since these lists
500 : * are independent, there may be notify events in a subtransaction's list
501 : * that duplicate events in some ancestor (sub) transaction; we get rid of
502 : * the dups when merging the subtransaction's list into its parent's.
503 : *
504 : * Note: the action and notify lists do not interact within a transaction.
505 : * In particular, if a transaction does NOTIFY and then LISTEN on the same
506 : * condition name, it will get a self-notify at commit. This is a bit odd
507 : * but is consistent with our historical behavior.
508 : */
509 : typedef struct Notification
510 : {
511 : uint16 channel_len; /* length of channel-name string */
512 : uint16 payload_len; /* length of payload string */
513 : /* null-terminated channel name, then null-terminated payload follow */
514 : char data[FLEXIBLE_ARRAY_MEMBER];
515 : } Notification;
516 :
517 : typedef struct NotificationList
518 : {
519 : int nestingLevel; /* current transaction nesting depth */
520 : List *events; /* list of Notification structs */
521 : HTAB *hashtab; /* hash of NotificationHash structs, or NULL */
522 : List *uniqueChannelNames; /* unique channel names being notified */
523 : HTAB *uniqueChannelHash; /* hash of unique channel names, or NULL */
524 : struct NotificationList *upper; /* details for upper transaction levels */
525 : } NotificationList;
526 :
527 : #define MIN_HASHABLE_NOTIFIES 16 /* threshold to build hashtab */
528 :
529 : struct NotificationHash
530 : {
531 : Notification *event; /* => the actual Notification struct */
532 : };
533 :
534 : static NotificationList *pendingNotifies = NULL;
535 :
536 : /*
537 : * Hash entry in NotificationList.uniqueChannelHash or localChannelTable
538 : * (both just carry the channel name, with no payload).
539 : */
540 : typedef struct ChannelName
541 : {
542 : char channel[NAMEDATALEN]; /* hash key */
543 : } ChannelName;
544 :
545 : /*
546 : * Inbound notifications are initially processed by HandleNotifyInterrupt(),
547 : * called from inside a signal handler. That just sets the
548 : * notifyInterruptPending flag and sets the process
549 : * latch. ProcessNotifyInterrupt() will then be called whenever it's safe to
550 : * actually deal with the interrupt.
551 : */
552 : volatile sig_atomic_t notifyInterruptPending = false;
553 :
554 : /* True if we've registered an on_shmem_exit cleanup */
555 : static bool unlistenExitRegistered = false;
556 :
557 : /* True if we're currently registered as a listener in asyncQueueControl */
558 : static bool amRegisteredListener = false;
559 :
560 : /*
561 : * Queue head positions for direct advancement.
562 : * These are captured during PreCommit_Notify while holding the heavyweight
563 : * lock on database 0, ensuring no other backend can insert notifications
564 : * between them. SignalBackends uses these to advance idle backends.
565 : */
566 : static QueuePosition queueHeadBeforeWrite;
567 : static QueuePosition queueHeadAfterWrite;
568 :
569 : /*
570 : * Workspace arrays for SignalBackends. These are preallocated in
571 : * PreCommit_Notify to avoid needing memory allocation after committing to
572 : * clog.
573 : */
574 : static int32 *signalPids = NULL;
575 : static ProcNumber *signalProcnos = NULL;
576 :
577 : /* have we advanced to a page that's a multiple of QUEUE_CLEANUP_DELAY? */
578 : static bool tryAdvanceTail = false;
579 :
580 : /* GUC parameters */
581 : bool Trace_notify = false;
582 :
583 : /* For 8 KB pages this gives 8 GB of disk space */
584 : int max_notify_queue_pages = 1048576;
585 :
586 : /* local function prototypes */
587 : static inline int64 asyncQueuePageDiff(int64 p, int64 q);
588 : static inline void GlobalChannelKeyInit(GlobalChannelKey *key, Oid dboid,
589 : const char *channel);
590 : static dshash_hash globalChannelTableHash(const void *key, size_t size,
591 : void *arg);
592 : static void initGlobalChannelTable(void);
593 : static void initLocalChannelTable(void);
594 : static void queue_listen(ListenActionKind action, const char *channel);
595 : static void Async_UnlistenOnExit(int code, Datum arg);
596 : static void BecomeRegisteredListener(void);
597 : static void PrepareTableEntriesForListen(const char *channel);
598 : static void PrepareTableEntriesForUnlisten(const char *channel);
599 : static void PrepareTableEntriesForUnlistenAll(void);
600 : static void RemoveListenerFromChannel(GlobalChannelEntry **entry_ptr,
601 : ListenerEntry *listeners,
602 : int idx);
603 : static void ApplyPendingListenActions(bool isCommit);
604 : static void CleanupListenersOnExit(void);
605 : static bool IsListeningOn(const char *channel);
606 : static void asyncQueueUnregister(void);
607 : static bool asyncQueueIsFull(void);
608 : static bool asyncQueueAdvance(volatile QueuePosition *position, int entryLength);
609 : static void asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe);
610 : static ListCell *asyncQueueAddEntries(ListCell *nextNotify);
611 : static double asyncQueueUsage(void);
612 : static void asyncQueueFillWarning(void);
613 : static void SignalBackends(void);
614 : static void asyncQueueReadAllNotifications(void);
615 : static bool asyncQueueProcessPageEntries(QueuePosition *current,
616 : QueuePosition stop,
617 : Snapshot snapshot);
618 : static void asyncQueueAdvanceTail(void);
619 : static void ProcessIncomingNotify(bool flush);
620 : static bool AsyncExistsPendingNotify(Notification *n);
621 : static void AddEventToPendingNotifies(Notification *n);
622 : static uint32 notification_hash(const void *key, Size keysize);
623 : static int notification_match(const void *key1, const void *key2, Size keysize);
624 : static void ClearPendingActionsAndNotifies(void);
625 :
626 : static int
627 0 : asyncQueueErrdetailForIoError(const void *opaque_data)
628 : {
629 0 : const QueuePosition *position = opaque_data;
630 :
631 0 : return errdetail("Could not access async queue at page %" PRId64 ", offset %d.",
632 0 : position->page, position->offset);
633 : }
634 :
635 : /*
636 : * Compute the difference between two queue page numbers.
637 : * Previously this function accounted for a wraparound.
638 : */
639 : static inline int64
640 0 : asyncQueuePageDiff(int64 p, int64 q)
641 : {
642 0 : return p - q;
643 : }
644 :
645 : /*
646 : * Determines whether p precedes q.
647 : * Previously this function accounted for a wraparound.
648 : */
649 : static inline bool
650 120 : asyncQueuePagePrecedes(int64 p, int64 q)
651 : {
652 120 : return p < q;
653 : }
654 :
655 : /*
656 : * GlobalChannelKeyInit
657 : * Prepare a global channel table key for hashing.
658 : */
659 : static inline void
660 203 : GlobalChannelKeyInit(GlobalChannelKey *key, Oid dboid, const char *channel)
661 : {
662 203 : memset(key, 0, sizeof(GlobalChannelKey));
663 203 : key->dboid = dboid;
664 203 : strlcpy(key->channel, channel, NAMEDATALEN);
665 203 : }
666 :
667 : /*
668 : * globalChannelTableHash
669 : * Hash function for global channel table keys.
670 : */
671 : static dshash_hash
672 203 : globalChannelTableHash(const void *key, size_t size, void *arg)
673 : {
674 203 : const GlobalChannelKey *k = (const GlobalChannelKey *) key;
675 : dshash_hash h;
676 :
677 203 : h = DatumGetUInt32(hash_uint32(k->dboid));
678 203 : h ^= DatumGetUInt32(hash_any((const unsigned char *) k->channel,
679 203 : strnlen(k->channel, NAMEDATALEN)));
680 :
681 203 : return h;
682 : }
683 :
684 : /* parameters for the global channel table */
685 : static const dshash_parameters globalChannelTableDSHParams = {
686 : sizeof(GlobalChannelKey),
687 : sizeof(GlobalChannelEntry),
688 : dshash_memcmp,
689 : globalChannelTableHash,
690 : dshash_memcpy,
691 : LWTRANCHE_NOTIFY_CHANNEL_HASH
692 : };
693 :
694 : /*
695 : * initGlobalChannelTable
696 : * Lazy initialization of the global channel table.
697 : */
698 : static void
699 151 : initGlobalChannelTable(void)
700 : {
701 : MemoryContext oldcontext;
702 :
703 : /* Quick exit if we already did this */
704 151 : if (asyncQueueControl->globalChannelTableDSH != DSHASH_HANDLE_INVALID &&
705 143 : globalChannelTable != NULL)
706 121 : return;
707 :
708 : /* Otherwise, use a lock to ensure only one process creates the table */
709 30 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
710 :
711 : /* Be sure any local memory allocated by DSA routines is persistent */
712 30 : oldcontext = MemoryContextSwitchTo(TopMemoryContext);
713 :
714 30 : if (asyncQueueControl->globalChannelTableDSH == DSHASH_HANDLE_INVALID)
715 : {
716 : /* Initialize dynamic shared hash table for global channels */
717 8 : globalChannelDSA = dsa_create(LWTRANCHE_NOTIFY_CHANNEL_HASH);
718 8 : dsa_pin(globalChannelDSA);
719 8 : dsa_pin_mapping(globalChannelDSA);
720 8 : globalChannelTable = dshash_create(globalChannelDSA,
721 : &globalChannelTableDSHParams,
722 : NULL);
723 :
724 : /* Store handles in shared memory for other backends to use */
725 8 : asyncQueueControl->globalChannelTableDSA = dsa_get_handle(globalChannelDSA);
726 8 : asyncQueueControl->globalChannelTableDSH =
727 8 : dshash_get_hash_table_handle(globalChannelTable);
728 : }
729 22 : else if (!globalChannelTable)
730 : {
731 : /* Attach to existing dynamic shared hash table */
732 22 : globalChannelDSA = dsa_attach(asyncQueueControl->globalChannelTableDSA);
733 22 : dsa_pin_mapping(globalChannelDSA);
734 22 : globalChannelTable = dshash_attach(globalChannelDSA,
735 : &globalChannelTableDSHParams,
736 22 : asyncQueueControl->globalChannelTableDSH,
737 : NULL);
738 : }
739 :
740 30 : MemoryContextSwitchTo(oldcontext);
741 30 : LWLockRelease(NotifyQueueLock);
742 : }
743 :
744 : /*
745 : * initLocalChannelTable
746 : * Lazy initialization of the local channel table.
747 : * Once created, this table lasts for the life of the session.
748 : */
749 : static void
750 82 : initLocalChannelTable(void)
751 : {
752 : HASHCTL hash_ctl;
753 :
754 : /* Quick exit if we already did this */
755 82 : if (localChannelTable != NULL)
756 64 : return;
757 :
758 : /* Initialize local hash table for this backend's listened channels */
759 18 : hash_ctl.keysize = NAMEDATALEN;
760 18 : hash_ctl.entrysize = sizeof(ChannelName);
761 :
762 18 : localChannelTable =
763 18 : hash_create("Local Listen Channels",
764 : 64,
765 : &hash_ctl,
766 : HASH_ELEM | HASH_STRINGS);
767 : }
768 :
769 : /*
770 : * initPendingListenActions
771 : * Lazy initialization of the pending listen actions hash table.
772 : * This is allocated in CurTransactionContext during PreCommit_Notify,
773 : * and destroyed at transaction end.
774 : */
775 : static void
776 82 : initPendingListenActions(void)
777 : {
778 : HASHCTL hash_ctl;
779 :
780 82 : if (pendingListenActions != NULL)
781 0 : return;
782 :
783 82 : hash_ctl.keysize = NAMEDATALEN;
784 82 : hash_ctl.entrysize = sizeof(PendingListenEntry);
785 82 : hash_ctl.hcxt = CurTransactionContext;
786 :
787 82 : pendingListenActions =
788 82 : hash_create("Pending Listen Actions",
789 82 : list_length(pendingActions->actions),
790 : &hash_ctl,
791 : HASH_ELEM | HASH_STRINGS | HASH_CONTEXT);
792 : }
793 :
794 : /*
795 : * Register our shared memory needs
796 : */
797 : static void
798 1255 : AsyncShmemRequest(void *arg)
799 : {
800 : Size size;
801 :
802 1255 : size = mul_size(MaxBackends, sizeof(QueueBackendStatus));
803 1255 : size = add_size(size, offsetof(AsyncQueueControl, backend));
804 :
805 1255 : ShmemRequestStruct(.name = "Async Queue Control",
806 : .size = size,
807 : .ptr = (void **) &asyncQueueControl,
808 : );
809 :
810 1255 : SimpleLruRequest(.desc = &NotifySlruDesc,
811 : .name = "notify",
812 : .Dir = "pg_notify",
813 :
814 : /* long segment names are used in order to avoid wraparound */
815 : .long_segment_names = true,
816 :
817 : .nslots = notify_buffers,
818 :
819 : .sync_handler = SYNC_HANDLER_NONE,
820 : .PagePrecedes = asyncQueuePagePrecedes,
821 : .errdetail_for_io_error = asyncQueueErrdetailForIoError,
822 :
823 : .buffer_tranche_id = LWTRANCHE_NOTIFY_BUFFER,
824 : .bank_tranche_id = LWTRANCHE_NOTIFY_SLRU,
825 : );
826 1255 : }
827 :
828 : static void
829 1252 : AsyncShmemInit(void *arg)
830 : {
831 1252 : SET_QUEUE_POS(QUEUE_HEAD, 0, 0);
832 1252 : SET_QUEUE_POS(QUEUE_TAIL, 0, 0);
833 1252 : QUEUE_STOP_PAGE = 0;
834 1252 : QUEUE_FIRST_LISTENER = INVALID_PROC_NUMBER;
835 1252 : asyncQueueControl->lastQueueFillWarn = 0;
836 1252 : asyncQueueControl->globalChannelTableDSA = DSA_HANDLE_INVALID;
837 1252 : asyncQueueControl->globalChannelTableDSH = DSHASH_HANDLE_INVALID;
838 117894 : for (int i = 0; i < MaxBackends; i++)
839 : {
840 116642 : QUEUE_BACKEND_PID(i) = InvalidPid;
841 116642 : QUEUE_BACKEND_DBOID(i) = InvalidOid;
842 116642 : QUEUE_NEXT_LISTENER(i) = INVALID_PROC_NUMBER;
843 116642 : SET_QUEUE_POS(QUEUE_BACKEND_POS(i), 0, 0);
844 116642 : QUEUE_BACKEND_WAKEUP_PENDING(i) = false;
845 116642 : QUEUE_BACKEND_IS_ADVANCING(i) = false;
846 : }
847 :
848 : /*
849 : * During start or reboot, clean out the pg_notify directory.
850 : */
851 1252 : (void) SlruScanDirectory(NotifyCtl, SlruScanDirCbDeleteAll, NULL);
852 1252 : }
853 :
854 :
855 : /*
856 : * pg_notify -
857 : * SQL function to send a notification event
858 : */
859 : Datum
860 1078 : pg_notify(PG_FUNCTION_ARGS)
861 : {
862 : const char *channel;
863 : const char *payload;
864 :
865 1078 : if (PG_ARGISNULL(0))
866 4 : channel = "";
867 : else
868 1074 : channel = text_to_cstring(PG_GETARG_TEXT_PP(0));
869 :
870 1078 : if (PG_ARGISNULL(1))
871 7 : payload = "";
872 : else
873 1071 : payload = text_to_cstring(PG_GETARG_TEXT_PP(1));
874 :
875 : /* For NOTIFY as a statement, this is checked in ProcessUtility */
876 1078 : PreventCommandDuringRecovery("NOTIFY");
877 :
878 1078 : Async_Notify(channel, payload);
879 :
880 1066 : PG_RETURN_VOID();
881 : }
882 :
883 :
884 : /*
885 : * Async_Notify
886 : *
887 : * This is executed by the SQL notify command.
888 : *
889 : * Adds the message to the list of pending notifies.
890 : * Actual notification happens during transaction commit.
891 : * ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
892 : */
893 : void
894 1145 : Async_Notify(const char *channel, const char *payload)
895 : {
896 1145 : int my_level = GetCurrentTransactionNestLevel();
897 : size_t channel_len;
898 : size_t payload_len;
899 : Notification *n;
900 : MemoryContext oldcontext;
901 :
902 1145 : if (IsParallelWorker())
903 0 : elog(ERROR, "cannot send notifications from a parallel worker");
904 :
905 1145 : if (Trace_notify)
906 0 : elog(DEBUG1, "Async_Notify(%s)", channel);
907 :
908 1145 : channel_len = channel ? strlen(channel) : 0;
909 1145 : payload_len = payload ? strlen(payload) : 0;
910 :
911 : /* a channel name must be specified */
912 1145 : if (channel_len == 0)
913 8 : ereport(ERROR,
914 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
915 : errmsg("channel name cannot be empty")));
916 :
917 : /* enforce length limits */
918 1137 : if (channel_len >= NAMEDATALEN)
919 4 : ereport(ERROR,
920 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
921 : errmsg("channel name too long")));
922 :
923 1133 : if (payload_len >= NOTIFY_PAYLOAD_MAX_LENGTH)
924 0 : ereport(ERROR,
925 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
926 : errmsg("payload string too long")));
927 :
928 : /*
929 : * We must construct the Notification entry, even if we end up not using
930 : * it, in order to compare it cheaply to existing list entries.
931 : *
932 : * The notification list needs to live until end of transaction, so store
933 : * it in the transaction context.
934 : */
935 1133 : oldcontext = MemoryContextSwitchTo(CurTransactionContext);
936 :
937 1133 : n = (Notification *) palloc(offsetof(Notification, data) +
938 1133 : channel_len + payload_len + 2);
939 1133 : n->channel_len = channel_len;
940 1133 : n->payload_len = payload_len;
941 1133 : strcpy(n->data, channel);
942 1133 : if (payload)
943 1118 : strcpy(n->data + channel_len + 1, payload);
944 : else
945 15 : n->data[channel_len + 1] = '\0';
946 :
947 1133 : if (pendingNotifies == NULL || my_level > pendingNotifies->nestingLevel)
948 71 : {
949 : NotificationList *notifies;
950 :
951 : /*
952 : * First notify event in current (sub)xact. Note that we allocate the
953 : * NotificationList in TopTransactionContext; the nestingLevel might
954 : * get changed later by AtSubCommit_Notify.
955 : */
956 : notifies = (NotificationList *)
957 71 : MemoryContextAlloc(TopTransactionContext,
958 : sizeof(NotificationList));
959 71 : notifies->nestingLevel = my_level;
960 71 : notifies->events = list_make1(n);
961 : /* We certainly don't need a hashtable yet */
962 71 : notifies->hashtab = NULL;
963 : /* We won't build uniqueChannelNames/Hash till later, either */
964 71 : notifies->uniqueChannelNames = NIL;
965 71 : notifies->uniqueChannelHash = NULL;
966 71 : notifies->upper = pendingNotifies;
967 71 : pendingNotifies = notifies;
968 : }
969 : else
970 : {
971 : /* Now check for duplicates */
972 1062 : if (AsyncExistsPendingNotify(n))
973 : {
974 : /* It's a dup, so forget it */
975 13 : pfree(n);
976 13 : MemoryContextSwitchTo(oldcontext);
977 13 : return;
978 : }
979 :
980 : /* Append more events to existing list */
981 1049 : AddEventToPendingNotifies(n);
982 : }
983 :
984 1120 : MemoryContextSwitchTo(oldcontext);
985 : }
986 :
987 : /*
988 : * queue_listen
989 : * Common code for listen, unlisten, unlisten all commands.
990 : *
991 : * Adds the request to the list of pending actions.
992 : * Actual update of localChannelTable and globalChannelTable happens during
993 : * PreCommit_Notify, with staged changes committed in AtCommit_Notify.
994 : */
995 : static void
996 98 : queue_listen(ListenActionKind action, const char *channel)
997 : {
998 : MemoryContext oldcontext;
999 : ListenAction *actrec;
1000 98 : int my_level = GetCurrentTransactionNestLevel();
1001 :
1002 : /*
1003 : * Unlike Async_Notify, we don't try to collapse out duplicates here. We
1004 : * keep the ordered list to preserve interactions like UNLISTEN ALL; the
1005 : * final per-channel intent is computed during PreCommit_Notify.
1006 : */
1007 98 : oldcontext = MemoryContextSwitchTo(CurTransactionContext);
1008 :
1009 : /* space for terminating null is included in sizeof(ListenAction) */
1010 98 : actrec = (ListenAction *) palloc(offsetof(ListenAction, channel) +
1011 98 : strlen(channel) + 1);
1012 98 : actrec->action = action;
1013 98 : strcpy(actrec->channel, channel);
1014 :
1015 98 : if (pendingActions == NULL || my_level > pendingActions->nestingLevel)
1016 85 : {
1017 : ActionList *actions;
1018 :
1019 : /*
1020 : * First action in current sub(xact). Note that we allocate the
1021 : * ActionList in TopTransactionContext; the nestingLevel might get
1022 : * changed later by AtSubCommit_Notify.
1023 : */
1024 : actions = (ActionList *)
1025 85 : MemoryContextAlloc(TopTransactionContext, sizeof(ActionList));
1026 85 : actions->nestingLevel = my_level;
1027 85 : actions->actions = list_make1(actrec);
1028 85 : actions->upper = pendingActions;
1029 85 : pendingActions = actions;
1030 : }
1031 : else
1032 13 : pendingActions->actions = lappend(pendingActions->actions, actrec);
1033 :
1034 98 : MemoryContextSwitchTo(oldcontext);
1035 98 : }
1036 :
1037 : /*
1038 : * Async_Listen
1039 : *
1040 : * This is executed by the SQL listen command.
1041 : */
1042 : void
1043 52 : Async_Listen(const char *channel)
1044 : {
1045 52 : if (Trace_notify)
1046 0 : elog(DEBUG1, "Async_Listen(%s,%d)", channel, MyProcPid);
1047 :
1048 52 : queue_listen(LISTEN_LISTEN, channel);
1049 52 : }
1050 :
1051 : /*
1052 : * Async_Unlisten
1053 : *
1054 : * This is executed by the SQL unlisten command.
1055 : */
1056 : void
1057 4 : Async_Unlisten(const char *channel)
1058 : {
1059 4 : if (Trace_notify)
1060 0 : elog(DEBUG1, "Async_Unlisten(%s,%d)", channel, MyProcPid);
1061 :
1062 : /* If we couldn't possibly be listening, no need to queue anything */
1063 4 : if (pendingActions == NULL && !unlistenExitRegistered)
1064 0 : return;
1065 :
1066 4 : queue_listen(LISTEN_UNLISTEN, channel);
1067 : }
1068 :
1069 : /*
1070 : * Async_UnlistenAll
1071 : *
1072 : * This is invoked by UNLISTEN * command, and also at backend exit.
1073 : */
1074 : void
1075 62 : Async_UnlistenAll(void)
1076 : {
1077 62 : if (Trace_notify)
1078 0 : elog(DEBUG1, "Async_UnlistenAll(%d)", MyProcPid);
1079 :
1080 : /* If we couldn't possibly be listening, no need to queue anything */
1081 62 : if (pendingActions == NULL && !unlistenExitRegistered)
1082 20 : return;
1083 :
1084 42 : queue_listen(LISTEN_UNLISTEN_ALL, "");
1085 : }
1086 :
1087 : /*
1088 : * SQL function: return a set of the channel names this backend is actively
1089 : * listening to.
1090 : *
1091 : * Note: this coding relies on the fact that the localChannelTable cannot
1092 : * change within a transaction.
1093 : */
1094 : Datum
1095 12 : pg_listening_channels(PG_FUNCTION_ARGS)
1096 : {
1097 : FuncCallContext *funcctx;
1098 : HASH_SEQ_STATUS *status;
1099 :
1100 : /* stuff done only on the first call of the function */
1101 12 : if (SRF_IS_FIRSTCALL())
1102 : {
1103 : /* create a function context for cross-call persistence */
1104 8 : funcctx = SRF_FIRSTCALL_INIT();
1105 :
1106 : /* Initialize hash table iteration if we have any channels */
1107 8 : if (localChannelTable != NULL)
1108 : {
1109 : MemoryContext oldcontext;
1110 :
1111 8 : oldcontext = MemoryContextSwitchTo(funcctx->multi_call_memory_ctx);
1112 8 : status = (HASH_SEQ_STATUS *) palloc(sizeof(HASH_SEQ_STATUS));
1113 8 : hash_seq_init(status, localChannelTable);
1114 8 : funcctx->user_fctx = status;
1115 8 : MemoryContextSwitchTo(oldcontext);
1116 : }
1117 : else
1118 : {
1119 0 : funcctx->user_fctx = NULL;
1120 : }
1121 : }
1122 :
1123 : /* stuff done on every call of the function */
1124 12 : funcctx = SRF_PERCALL_SETUP();
1125 12 : status = (HASH_SEQ_STATUS *) funcctx->user_fctx;
1126 :
1127 12 : if (status != NULL)
1128 : {
1129 : ChannelName *entry;
1130 :
1131 12 : entry = (ChannelName *) hash_seq_search(status);
1132 12 : if (entry != NULL)
1133 4 : SRF_RETURN_NEXT(funcctx, CStringGetTextDatum(entry->channel));
1134 : }
1135 :
1136 8 : SRF_RETURN_DONE(funcctx);
1137 : }
1138 :
1139 : /*
1140 : * Async_UnlistenOnExit
1141 : *
1142 : * This is executed at backend exit if we have done any LISTENs in this
1143 : * backend. It might not be necessary anymore, if the user UNLISTENed
1144 : * everything, but we don't try to detect that case.
1145 : */
1146 : static void
1147 18 : Async_UnlistenOnExit(int code, Datum arg)
1148 : {
1149 18 : CleanupListenersOnExit();
1150 18 : asyncQueueUnregister();
1151 18 : }
1152 :
1153 : /*
1154 : * AtPrepare_Notify
1155 : *
1156 : * This is called at the prepare phase of a two-phase
1157 : * transaction. Save the state for possible commit later.
1158 : */
1159 : void
1160 312 : AtPrepare_Notify(void)
1161 : {
1162 : /* It's not allowed to have any pending LISTEN/UNLISTEN/NOTIFY actions */
1163 312 : if (pendingActions || pendingNotifies)
1164 0 : ereport(ERROR,
1165 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1166 : errmsg("cannot PREPARE a transaction that has executed LISTEN, UNLISTEN, or NOTIFY")));
1167 312 : }
1168 :
1169 : /*
1170 : * PreCommit_Notify
1171 : *
1172 : * This is called at transaction commit, before actually committing to
1173 : * clog.
1174 : *
1175 : * If there are pending LISTEN actions, make sure we are listed in the
1176 : * shared-memory listener array. This must happen before commit to
1177 : * ensure we don't miss any notifies from transactions that commit
1178 : * just after ours.
1179 : *
1180 : * If there are outbound notify requests in the pendingNotifies list,
1181 : * add them to the global queue. We do that before commit so that
1182 : * we can still throw error if we run out of queue space.
1183 : */
1184 : void
1185 612594 : PreCommit_Notify(void)
1186 : {
1187 : ListCell *p;
1188 :
1189 612594 : if (!pendingActions && !pendingNotifies)
1190 612443 : return; /* no relevant statements in this xact */
1191 :
1192 151 : if (Trace_notify)
1193 0 : elog(DEBUG1, "PreCommit_Notify");
1194 :
1195 : /* Preflight for any pending listen/unlisten actions */
1196 151 : initGlobalChannelTable();
1197 :
1198 151 : if (pendingActions != NULL)
1199 : {
1200 : /* Ensure we have a local channel table */
1201 82 : initLocalChannelTable();
1202 : /* Create pendingListenActions hash table for this transaction */
1203 82 : initPendingListenActions();
1204 :
1205 : /* Stage all the actions this transaction wants to perform */
1206 177 : foreach(p, pendingActions->actions)
1207 : {
1208 95 : ListenAction *actrec = (ListenAction *) lfirst(p);
1209 :
1210 95 : switch (actrec->action)
1211 : {
1212 50 : case LISTEN_LISTEN:
1213 50 : BecomeRegisteredListener();
1214 50 : PrepareTableEntriesForListen(actrec->channel);
1215 50 : break;
1216 4 : case LISTEN_UNLISTEN:
1217 4 : PrepareTableEntriesForUnlisten(actrec->channel);
1218 4 : break;
1219 41 : case LISTEN_UNLISTEN_ALL:
1220 41 : PrepareTableEntriesForUnlistenAll();
1221 41 : break;
1222 : }
1223 : }
1224 : }
1225 :
1226 : /* Queue any pending notifies (must happen after the above) */
1227 151 : if (pendingNotifies)
1228 : {
1229 : ListCell *nextNotify;
1230 69 : bool firstIteration = true;
1231 :
1232 : /*
1233 : * Build list of unique channel names being notified for use by
1234 : * SignalBackends().
1235 : *
1236 : * If uniqueChannelHash is available, use it to efficiently get the
1237 : * unique channels. Otherwise, fall back to the O(N^2) approach.
1238 : */
1239 69 : pendingNotifies->uniqueChannelNames = NIL;
1240 69 : if (pendingNotifies->uniqueChannelHash != NULL)
1241 : {
1242 : HASH_SEQ_STATUS status;
1243 : ChannelName *channelEntry;
1244 :
1245 2 : hash_seq_init(&status, pendingNotifies->uniqueChannelHash);
1246 4 : while ((channelEntry = (ChannelName *) hash_seq_search(&status)) != NULL)
1247 2 : pendingNotifies->uniqueChannelNames =
1248 2 : lappend(pendingNotifies->uniqueChannelNames,
1249 2 : channelEntry->channel);
1250 : }
1251 : else
1252 : {
1253 : /* O(N^2) approach is better for small number of notifications */
1254 231 : foreach_ptr(Notification, n, pendingNotifies->events)
1255 : {
1256 97 : char *channel = n->data;
1257 97 : bool found = false;
1258 :
1259 : /* Name present in list? */
1260 198 : foreach_ptr(char, oldchan, pendingNotifies->uniqueChannelNames)
1261 : {
1262 31 : if (strcmp(oldchan, channel) == 0)
1263 : {
1264 27 : found = true;
1265 27 : break;
1266 : }
1267 : }
1268 : /* Add if not already in list */
1269 97 : if (!found)
1270 70 : pendingNotifies->uniqueChannelNames =
1271 70 : lappend(pendingNotifies->uniqueChannelNames,
1272 : channel);
1273 : }
1274 : }
1275 :
1276 : /* Preallocate workspace that will be needed by SignalBackends() */
1277 69 : if (signalPids == NULL)
1278 23 : signalPids = MemoryContextAlloc(TopMemoryContext,
1279 : MaxBackends * sizeof(int32));
1280 :
1281 69 : if (signalProcnos == NULL)
1282 23 : signalProcnos = MemoryContextAlloc(TopMemoryContext,
1283 : MaxBackends * sizeof(ProcNumber));
1284 :
1285 : /*
1286 : * Make sure that we have an XID assigned to the current transaction.
1287 : * GetCurrentTransactionId is cheap if we already have an XID, but not
1288 : * so cheap if we don't, and we'd prefer not to do that work while
1289 : * holding NotifyQueueLock.
1290 : */
1291 69 : (void) GetCurrentTransactionId();
1292 :
1293 : /*
1294 : * Serialize writers by acquiring a special lock that we hold till
1295 : * after commit. This ensures that queue entries appear in commit
1296 : * order, and in particular that there are never uncommitted queue
1297 : * entries ahead of committed ones, so an uncommitted transaction
1298 : * can't block delivery of deliverable notifications.
1299 : *
1300 : * We use a heavyweight lock so that it'll automatically be released
1301 : * after either commit or abort. This also allows deadlocks to be
1302 : * detected, though really a deadlock shouldn't be possible here.
1303 : *
1304 : * The lock is on "database 0", which is pretty ugly but it doesn't
1305 : * seem worth inventing a special locktag category just for this.
1306 : * (Historical note: before PG 9.0, a similar lock on "database 0" was
1307 : * used by the flatfiles mechanism.)
1308 : */
1309 69 : LockSharedObject(DatabaseRelationId, InvalidOid, 0,
1310 : AccessExclusiveLock);
1311 :
1312 : /*
1313 : * For the direct advancement optimization in SignalBackends(), we
1314 : * need to ensure that no other backend can insert queue entries
1315 : * between queueHeadBeforeWrite and queueHeadAfterWrite. The
1316 : * heavyweight lock above provides this guarantee, since it serializes
1317 : * all writers.
1318 : *
1319 : * Note: if the heavyweight lock were ever removed for scalability
1320 : * reasons, we could achieve the same guarantee by holding
1321 : * NotifyQueueLock in EXCLUSIVE mode across all our insertions, rather
1322 : * than releasing and reacquiring it for each page as we do below.
1323 : */
1324 :
1325 : /* Initialize values to a safe default in case list is empty */
1326 69 : SET_QUEUE_POS(queueHeadBeforeWrite, 0, 0);
1327 69 : SET_QUEUE_POS(queueHeadAfterWrite, 0, 0);
1328 :
1329 : /* Now push the notifications into the queue */
1330 69 : nextNotify = list_head(pendingNotifies->events);
1331 173 : while (nextNotify != NULL)
1332 : {
1333 : /*
1334 : * Add the pending notifications to the queue. We acquire and
1335 : * release NotifyQueueLock once per page, which might be overkill
1336 : * but it does allow readers to get in while we're doing this.
1337 : *
1338 : * A full queue is very uncommon and should really not happen,
1339 : * given that we have so much space available in the SLRU pages.
1340 : * Nevertheless we need to deal with this possibility. Note that
1341 : * when we get here we are in the process of committing our
1342 : * transaction, but we have not yet committed to clog, so at this
1343 : * point in time we can still roll the transaction back.
1344 : */
1345 104 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
1346 104 : if (firstIteration)
1347 : {
1348 69 : queueHeadBeforeWrite = QUEUE_HEAD;
1349 69 : firstIteration = false;
1350 : }
1351 104 : asyncQueueFillWarning();
1352 104 : if (asyncQueueIsFull())
1353 0 : ereport(ERROR,
1354 : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1355 : errmsg("too many notifications in the NOTIFY queue")));
1356 104 : nextNotify = asyncQueueAddEntries(nextNotify);
1357 104 : queueHeadAfterWrite = QUEUE_HEAD;
1358 104 : LWLockRelease(NotifyQueueLock);
1359 : }
1360 :
1361 : /* Note that we don't clear pendingNotifies; AtCommit_Notify will. */
1362 : }
1363 : }
1364 :
1365 : /*
1366 : * AtCommit_Notify
1367 : *
1368 : * This is called at transaction commit, after committing to clog.
1369 : *
1370 : * Apply pending listen/unlisten changes and clear transaction-local state.
1371 : *
1372 : * If we issued any notifications in the transaction, send signals to
1373 : * listening backends (possibly including ourselves) to process them.
1374 : * Also, if we filled enough queue pages with new notifies, try to
1375 : * advance the queue tail pointer.
1376 : */
1377 : void
1378 612439 : AtCommit_Notify(void)
1379 : {
1380 : /*
1381 : * Allow transactions that have not executed LISTEN/UNLISTEN/NOTIFY to
1382 : * return as soon as possible
1383 : */
1384 612439 : if (!pendingActions && !pendingNotifies)
1385 612288 : return;
1386 :
1387 151 : if (Trace_notify)
1388 0 : elog(DEBUG1, "AtCommit_Notify");
1389 :
1390 : /* Apply staged listen/unlisten changes */
1391 151 : ApplyPendingListenActions(true);
1392 :
1393 : /* If no longer listening to anything, get out of listener array */
1394 151 : if (amRegisteredListener && LocalChannelTableIsEmpty())
1395 21 : asyncQueueUnregister();
1396 :
1397 : /*
1398 : * Send signals to listening backends. We need do this only if there are
1399 : * pending notifies, which were previously added to the shared queue by
1400 : * PreCommit_Notify().
1401 : */
1402 151 : if (pendingNotifies != NULL)
1403 69 : SignalBackends();
1404 :
1405 : /*
1406 : * If it's time to try to advance the global tail pointer, do that.
1407 : *
1408 : * (It might seem odd to do this in the sender, when more than likely the
1409 : * listeners won't yet have read the messages we just sent. However,
1410 : * there's less contention if only the sender does it, and there is little
1411 : * need for urgency in advancing the global tail. So this typically will
1412 : * be clearing out messages that were sent some time ago.)
1413 : */
1414 151 : if (tryAdvanceTail)
1415 : {
1416 8 : tryAdvanceTail = false;
1417 8 : asyncQueueAdvanceTail();
1418 : }
1419 :
1420 : /* And clean up */
1421 151 : ClearPendingActionsAndNotifies();
1422 : }
1423 :
1424 : /*
1425 : * BecomeRegisteredListener --- subroutine for PreCommit_Notify
1426 : *
1427 : * This function must make sure we are ready to catch any incoming messages.
1428 : */
1429 : static void
1430 50 : BecomeRegisteredListener(void)
1431 : {
1432 : QueuePosition head;
1433 : QueuePosition max;
1434 : ProcNumber prevListener;
1435 :
1436 : /*
1437 : * Nothing to do if we are already listening to something, nor if we
1438 : * already ran this routine in this transaction.
1439 : */
1440 50 : if (amRegisteredListener)
1441 23 : return;
1442 :
1443 27 : if (Trace_notify)
1444 0 : elog(DEBUG1, "BecomeRegisteredListener(%d)", MyProcPid);
1445 :
1446 : /*
1447 : * Before registering, make sure we will unlisten before dying. (Note:
1448 : * this action does not get undone if we abort later.)
1449 : */
1450 27 : if (!unlistenExitRegistered)
1451 : {
1452 18 : before_shmem_exit(Async_UnlistenOnExit, 0);
1453 18 : unlistenExitRegistered = true;
1454 : }
1455 :
1456 : /*
1457 : * This is our first LISTEN, so establish our pointer.
1458 : *
1459 : * We set our pointer to the global tail pointer and then move it forward
1460 : * over already-committed notifications. This ensures we cannot miss any
1461 : * not-yet-committed notifications. We might get a few more but that
1462 : * doesn't hurt.
1463 : *
1464 : * In some scenarios there might be a lot of committed notifications that
1465 : * have not yet been pruned away (because some backend is being lazy about
1466 : * reading them). To reduce our startup time, we can look at other
1467 : * backends and adopt the maximum "pos" pointer of any backend that's in
1468 : * our database; any notifications it's already advanced over are surely
1469 : * committed and need not be re-examined by us. (We must consider only
1470 : * backends connected to our DB, because others will not have bothered to
1471 : * check committed-ness of notifications in our DB.)
1472 : *
1473 : * We need exclusive lock here so we can look at other backends' entries
1474 : * and manipulate the list links.
1475 : */
1476 27 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
1477 27 : head = QUEUE_HEAD;
1478 27 : max = QUEUE_TAIL;
1479 27 : prevListener = INVALID_PROC_NUMBER;
1480 29 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
1481 : {
1482 2 : if (QUEUE_BACKEND_DBOID(i) == MyDatabaseId)
1483 2 : max = QUEUE_POS_MAX(max, QUEUE_BACKEND_POS(i));
1484 : /* Also find last listening backend before this one */
1485 2 : if (i < MyProcNumber)
1486 1 : prevListener = i;
1487 : }
1488 27 : QUEUE_BACKEND_POS(MyProcNumber) = max;
1489 27 : QUEUE_BACKEND_PID(MyProcNumber) = MyProcPid;
1490 27 : QUEUE_BACKEND_DBOID(MyProcNumber) = MyDatabaseId;
1491 27 : QUEUE_BACKEND_WAKEUP_PENDING(MyProcNumber) = false;
1492 27 : QUEUE_BACKEND_IS_ADVANCING(MyProcNumber) = false;
1493 : /* Insert backend into list of listeners at correct position */
1494 27 : if (prevListener != INVALID_PROC_NUMBER)
1495 : {
1496 1 : QUEUE_NEXT_LISTENER(MyProcNumber) = QUEUE_NEXT_LISTENER(prevListener);
1497 1 : QUEUE_NEXT_LISTENER(prevListener) = MyProcNumber;
1498 : }
1499 : else
1500 : {
1501 26 : QUEUE_NEXT_LISTENER(MyProcNumber) = QUEUE_FIRST_LISTENER;
1502 26 : QUEUE_FIRST_LISTENER = MyProcNumber;
1503 : }
1504 27 : LWLockRelease(NotifyQueueLock);
1505 :
1506 : /* Now we are listed in the global array, so remember we're listening */
1507 27 : amRegisteredListener = true;
1508 :
1509 : /*
1510 : * Try to move our pointer forward as far as possible. This will skip
1511 : * over already-committed notifications, which we want to do because they
1512 : * might be quite stale. Note that we are not yet listening on anything,
1513 : * so we won't deliver such notifications to our frontend. Also, although
1514 : * our transaction might have executed NOTIFY, those message(s) aren't
1515 : * queued yet so we won't skip them here.
1516 : */
1517 27 : if (!QUEUE_POS_EQUAL(max, head))
1518 18 : asyncQueueReadAllNotifications();
1519 : }
1520 :
1521 : /*
1522 : * PrepareTableEntriesForListen --- subroutine for PreCommit_Notify
1523 : *
1524 : * Prepare a LISTEN by recording it in pendingListenActions, pre-allocating
1525 : * an entry in localChannelTable, and pre-allocating an entry in the shared
1526 : * globalChannelTable with removeOnAbort set. AtCommit_Notify will clear
1527 : * removeOnAbort; abort processing will remove entries still marked so.
1528 : */
1529 : static void
1530 50 : PrepareTableEntriesForListen(const char *channel)
1531 : {
1532 : GlobalChannelKey key;
1533 : GlobalChannelEntry *entry;
1534 : bool found;
1535 : ListenerEntry *listeners;
1536 : PendingListenEntry *pending;
1537 :
1538 : /*
1539 : * Record in local pending hash that we want to LISTEN, overwriting any
1540 : * earlier attempt to UNLISTEN.
1541 : */
1542 : pending = (PendingListenEntry *)
1543 50 : hash_search(pendingListenActions, channel, HASH_ENTER, NULL);
1544 50 : pending->action = PENDING_LISTEN;
1545 :
1546 : /*
1547 : * Ensure that there is an entry for the channel in localChannelTable.
1548 : * (Should this fail, we can just roll back.) If the transaction fails
1549 : * after this point, we will remove the entry if appropriate during
1550 : * ApplyPendingListenActions. Note that this entry allows IsListeningOn()
1551 : * to return TRUE; we assume nothing is going to consult that before
1552 : * AtCommit_Notify/AtAbort_Notify. However, if later actions attempt to
1553 : * UNLISTEN this channel or UNLISTEN *, we need to have the local entry
1554 : * present to ensure they do the right things; see
1555 : * PrepareTableEntriesForUnlisten and PrepareTableEntriesForUnlistenAll.
1556 : */
1557 50 : (void) hash_search(localChannelTable, channel, HASH_ENTER, NULL);
1558 :
1559 : /* Pre-allocate entry in shared globalChannelTable */
1560 50 : GlobalChannelKeyInit(&key, MyDatabaseId, channel);
1561 50 : entry = dshash_find_or_insert(globalChannelTable, &key, &found);
1562 :
1563 50 : if (!found)
1564 : {
1565 : /* New channel entry, so initialize it to a safe state */
1566 37 : entry->listenersArray = InvalidDsaPointer;
1567 37 : entry->numListeners = 0;
1568 37 : entry->allocatedListeners = 0;
1569 : }
1570 :
1571 : /*
1572 : * Create listenersArray if entry doesn't have one. It's tempting to fold
1573 : * this into the !found case, but this coding allows us to cope in case
1574 : * dsa_allocate() failed in an earlier attempt.
1575 : */
1576 50 : if (!DsaPointerIsValid(entry->listenersArray))
1577 : {
1578 37 : entry->listenersArray = dsa_allocate(globalChannelDSA,
1579 : sizeof(ListenerEntry) * INITIAL_LISTENERS_ARRAY_SIZE);
1580 37 : entry->allocatedListeners = INITIAL_LISTENERS_ARRAY_SIZE;
1581 : }
1582 :
1583 : listeners = (ListenerEntry *)
1584 50 : dsa_get_address(globalChannelDSA, entry->listenersArray);
1585 :
1586 : /*
1587 : * Check if we already have a ListenerEntry (possibly from earlier in this
1588 : * transaction)
1589 : */
1590 53 : for (int i = 0; i < entry->numListeners; i++)
1591 : {
1592 13 : if (listeners[i].procNo == MyProcNumber)
1593 : {
1594 : /* Already have an entry; leave removeOnAbort as-is */
1595 10 : dshash_release_lock(globalChannelTable, entry);
1596 10 : return;
1597 : }
1598 : }
1599 :
1600 : /* Need to add a new entry; grow array if necessary */
1601 40 : if (entry->numListeners >= entry->allocatedListeners)
1602 : {
1603 0 : int new_size = entry->allocatedListeners * 2;
1604 0 : dsa_pointer old_array = entry->listenersArray;
1605 0 : dsa_pointer new_array = dsa_allocate(globalChannelDSA,
1606 : sizeof(ListenerEntry) * new_size);
1607 0 : ListenerEntry *new_listeners = (ListenerEntry *) dsa_get_address(globalChannelDSA, new_array);
1608 :
1609 0 : memcpy(new_listeners, listeners, sizeof(ListenerEntry) * entry->numListeners);
1610 0 : entry->listenersArray = new_array;
1611 0 : entry->allocatedListeners = new_size;
1612 0 : dsa_free(globalChannelDSA, old_array);
1613 0 : listeners = new_listeners;
1614 : }
1615 :
1616 40 : listeners[entry->numListeners].procNo = MyProcNumber;
1617 40 : listeners[entry->numListeners].removeOnAbort = true;
1618 40 : entry->numListeners++;
1619 :
1620 40 : dshash_release_lock(globalChannelTable, entry);
1621 : }
1622 :
1623 : /*
1624 : * PrepareTableEntriesForUnlisten --- subroutine for PreCommit_Notify
1625 : *
1626 : * Prepare an UNLISTEN by recording it in pendingListenActions, but only if
1627 : * we're currently listening (committed or staged). We don't touch
1628 : * globalChannelTable yet - the listener keeps receiving signals until
1629 : * commit, when the entry is removed.
1630 : */
1631 : static void
1632 4 : PrepareTableEntriesForUnlisten(const char *channel)
1633 : {
1634 : PendingListenEntry *pending;
1635 :
1636 : /*
1637 : * If the channel name is not in localChannelTable, then we are neither
1638 : * listening on it nor preparing to listen on it, so we don't need to
1639 : * record an UNLISTEN action.
1640 : */
1641 : Assert(localChannelTable != NULL);
1642 4 : if (hash_search(localChannelTable, channel, HASH_FIND, NULL) == NULL)
1643 0 : return;
1644 :
1645 : /*
1646 : * Record in local pending hash that we want to UNLISTEN, overwriting any
1647 : * earlier attempt to LISTEN. Don't touch localChannelTable or
1648 : * globalChannelTable yet - we keep receiving signals until commit.
1649 : */
1650 : pending = (PendingListenEntry *)
1651 4 : hash_search(pendingListenActions, channel, HASH_ENTER, NULL);
1652 4 : pending->action = PENDING_UNLISTEN;
1653 : }
1654 :
1655 : /*
1656 : * PrepareTableEntriesForUnlistenAll --- subroutine for PreCommit_Notify
1657 : *
1658 : * Prepare UNLISTEN * by recording an UNLISTEN for all listened or
1659 : * about-to-be-listened channels in pendingListenActions.
1660 : */
1661 : static void
1662 41 : PrepareTableEntriesForUnlistenAll(void)
1663 : {
1664 : HASH_SEQ_STATUS seq;
1665 : ChannelName *channelEntry;
1666 : PendingListenEntry *pending;
1667 :
1668 : /*
1669 : * Scan localChannelTable, which will have the names of all channels that
1670 : * we are listening on or have prepared to listen on. Record an UNLISTEN
1671 : * action for each one, overwriting any earlier attempt to LISTEN.
1672 : */
1673 41 : hash_seq_init(&seq, localChannelTable);
1674 69 : while ((channelEntry = (ChannelName *) hash_seq_search(&seq)) != NULL)
1675 : {
1676 : pending = (PendingListenEntry *)
1677 28 : hash_search(pendingListenActions, channelEntry->channel, HASH_ENTER, NULL);
1678 28 : pending->action = PENDING_UNLISTEN;
1679 : }
1680 41 : }
1681 :
1682 : /*
1683 : * RemoveListenerFromChannel --- remove idx'th listener from global channel entry
1684 : *
1685 : * Decrements numListeners, compacts the array, and frees the entry if empty.
1686 : * Sets *entry_ptr to NULL if the entry was deleted.
1687 : *
1688 : * We could get the listeners pointer from the entry, but all callers
1689 : * already have it at hand.
1690 : */
1691 : static void
1692 32 : RemoveListenerFromChannel(GlobalChannelEntry **entry_ptr,
1693 : ListenerEntry *listeners,
1694 : int idx)
1695 : {
1696 32 : GlobalChannelEntry *entry = *entry_ptr;
1697 :
1698 32 : entry->numListeners--;
1699 32 : if (idx < entry->numListeners)
1700 3 : memmove(&listeners[idx], &listeners[idx + 1],
1701 3 : sizeof(ListenerEntry) * (entry->numListeners - idx));
1702 :
1703 32 : if (entry->numListeners == 0)
1704 : {
1705 29 : dsa_free(globalChannelDSA, entry->listenersArray);
1706 29 : dshash_delete_entry(globalChannelTable, entry);
1707 : /* tells caller not to release the entry's lock: */
1708 29 : *entry_ptr = NULL;
1709 : }
1710 32 : }
1711 :
1712 : /*
1713 : * ApplyPendingListenActions
1714 : *
1715 : * Apply, or revert, staged listen/unlisten changes to the local and global
1716 : * hash tables.
1717 : */
1718 : static void
1719 35680 : ApplyPendingListenActions(bool isCommit)
1720 : {
1721 : HASH_SEQ_STATUS seq;
1722 : PendingListenEntry *pending;
1723 :
1724 : /* Quick exit if nothing to do */
1725 35680 : if (pendingListenActions == NULL)
1726 35598 : return;
1727 :
1728 : /* We made a globalChannelTable before building pendingListenActions */
1729 82 : if (globalChannelTable == NULL)
1730 0 : elog(PANIC, "global channel table missing post-commit/abort");
1731 :
1732 : /* For each staged action ... */
1733 82 : hash_seq_init(&seq, pendingListenActions);
1734 163 : while ((pending = (PendingListenEntry *) hash_seq_search(&seq)) != NULL)
1735 : {
1736 : GlobalChannelKey key;
1737 : GlobalChannelEntry *entry;
1738 81 : bool removeLocal = true;
1739 81 : bool foundListener = false;
1740 :
1741 : /*
1742 : * Find the global entry for this channel. If isCommit, it had better
1743 : * exist (it was created in PreCommit). In an abort, it might not
1744 : * exist, in which case we are not listening and should discard any
1745 : * local entry that PreCommit may have managed to create.
1746 : */
1747 81 : GlobalChannelKeyInit(&key, MyDatabaseId, pending->channel);
1748 81 : entry = dshash_find(globalChannelTable, &key, true);
1749 81 : if (entry != NULL)
1750 : {
1751 : /* Scan entry to find the ListenerEntry for this backend */
1752 : ListenerEntry *listeners;
1753 :
1754 : listeners = (ListenerEntry *)
1755 81 : dsa_get_address(globalChannelDSA, entry->listenersArray);
1756 :
1757 84 : for (int i = 0; i < entry->numListeners; i++)
1758 : {
1759 84 : if (listeners[i].procNo != MyProcNumber)
1760 3 : continue;
1761 81 : foundListener = true;
1762 81 : if (isCommit)
1763 : {
1764 81 : if (pending->action == PENDING_LISTEN)
1765 : {
1766 : /*
1767 : * LISTEN being committed: entry is now permanent.
1768 : * localChannelTable entry was created during
1769 : * PreCommit and should be kept.
1770 : */
1771 49 : listeners[i].removeOnAbort = false;
1772 49 : removeLocal = false;
1773 : }
1774 : else
1775 : {
1776 : /*
1777 : * UNLISTEN being committed: remove pre-allocated
1778 : * entries from both tables.
1779 : */
1780 32 : RemoveListenerFromChannel(&entry, listeners, i);
1781 : }
1782 : }
1783 : else
1784 : {
1785 : /*
1786 : * Note: this part is reachable only if the transaction
1787 : * aborts after PreCommit_Notify() has made some
1788 : * pendingListenActions entries, so it's pretty hard to
1789 : * test.
1790 : */
1791 0 : if (listeners[i].removeOnAbort)
1792 : {
1793 : /*
1794 : * Staged LISTEN (or LISTEN+UNLISTEN) being aborted,
1795 : * so remove pre-allocated entries from both tables.
1796 : */
1797 0 : RemoveListenerFromChannel(&entry, listeners, i);
1798 : }
1799 : else
1800 : {
1801 : /*
1802 : * Entry predates this transaction, so keep the
1803 : * localChannelTable entry.
1804 : */
1805 0 : removeLocal = false;
1806 : }
1807 : }
1808 81 : break; /* there shouldn't be another match */
1809 : }
1810 :
1811 : /* We might have already released the entry by removing it */
1812 81 : if (entry != NULL)
1813 52 : dshash_release_lock(globalChannelTable, entry);
1814 : }
1815 :
1816 : /*
1817 : * If we're committing a LISTEN action, we should have found a
1818 : * matching ListenerEntry, but otherwise it's okay if we didn't.
1819 : */
1820 81 : if (isCommit && pending->action == PENDING_LISTEN && !foundListener)
1821 0 : elog(PANIC, "could not find listener entry for channel \"%s\" backend %d",
1822 : pending->channel, MyProcNumber);
1823 :
1824 : /*
1825 : * If we did not find a globalChannelTable entry for our backend, or
1826 : * if we are unlistening, remove any localChannelTable entry that may
1827 : * exist. (Note in particular that this cleans up if we created a
1828 : * localChannelTable entry and then failed while trying to create a
1829 : * globalChannelTable entry.)
1830 : */
1831 81 : if (removeLocal && localChannelTable != NULL)
1832 32 : (void) hash_search(localChannelTable, pending->channel,
1833 : HASH_REMOVE, NULL);
1834 : }
1835 : }
1836 :
1837 : /*
1838 : * CleanupListenersOnExit --- called from Async_UnlistenOnExit
1839 : *
1840 : * Remove this backend from all channels in the shared global table.
1841 : */
1842 : static void
1843 18 : CleanupListenersOnExit(void)
1844 : {
1845 : dshash_seq_status status;
1846 : GlobalChannelEntry *entry;
1847 :
1848 18 : if (Trace_notify)
1849 0 : elog(DEBUG1, "CleanupListenersOnExit(%d)", MyProcPid);
1850 :
1851 : /* Clear our local cache (not really necessary, but be consistent) */
1852 18 : if (localChannelTable != NULL)
1853 : {
1854 18 : hash_destroy(localChannelTable);
1855 18 : localChannelTable = NULL;
1856 : }
1857 :
1858 : /* Now remove our entries from the shared globalChannelTable */
1859 18 : if (globalChannelTable == NULL)
1860 0 : return;
1861 :
1862 18 : dshash_seq_init(&status, globalChannelTable, true);
1863 28 : while ((entry = dshash_seq_next(&status)) != NULL)
1864 : {
1865 : ListenerEntry *listeners;
1866 :
1867 10 : if (entry->key.dboid != MyDatabaseId)
1868 0 : continue; /* not relevant */
1869 :
1870 : listeners = (ListenerEntry *)
1871 10 : dsa_get_address(globalChannelDSA, entry->listenersArray);
1872 :
1873 12 : for (int i = 0; i < entry->numListeners; i++)
1874 : {
1875 10 : if (listeners[i].procNo == MyProcNumber)
1876 : {
1877 8 : entry->numListeners--;
1878 8 : if (i < entry->numListeners)
1879 0 : memmove(&listeners[i], &listeners[i + 1],
1880 0 : sizeof(ListenerEntry) * (entry->numListeners - i));
1881 :
1882 8 : if (entry->numListeners == 0)
1883 : {
1884 8 : dsa_free(globalChannelDSA, entry->listenersArray);
1885 8 : dshash_delete_current(&status);
1886 : }
1887 8 : break;
1888 : }
1889 : }
1890 : }
1891 18 : dshash_seq_term(&status);
1892 : }
1893 :
1894 : /*
1895 : * Test whether we are actively listening on the given channel name.
1896 : *
1897 : * Note: this function is executed for every notification found in the queue.
1898 : */
1899 : static bool
1900 53 : IsListeningOn(const char *channel)
1901 : {
1902 53 : if (localChannelTable == NULL)
1903 0 : return false;
1904 :
1905 53 : return (hash_search(localChannelTable, channel, HASH_FIND, NULL) != NULL);
1906 : }
1907 :
1908 : /*
1909 : * Remove our entry from the listeners array when we are no longer listening
1910 : * on any channel. NB: must not fail if we're already not listening.
1911 : */
1912 : static void
1913 39 : asyncQueueUnregister(void)
1914 : {
1915 : Assert(LocalChannelTableIsEmpty()); /* else caller error */
1916 :
1917 39 : if (!amRegisteredListener) /* nothing to do */
1918 12 : return;
1919 :
1920 : /*
1921 : * Need exclusive lock here to manipulate list links.
1922 : */
1923 27 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
1924 : /* Mark our entry as invalid */
1925 27 : QUEUE_BACKEND_PID(MyProcNumber) = InvalidPid;
1926 27 : QUEUE_BACKEND_DBOID(MyProcNumber) = InvalidOid;
1927 27 : QUEUE_BACKEND_WAKEUP_PENDING(MyProcNumber) = false;
1928 27 : QUEUE_BACKEND_IS_ADVANCING(MyProcNumber) = false;
1929 : /* and remove it from the list */
1930 27 : if (QUEUE_FIRST_LISTENER == MyProcNumber)
1931 26 : QUEUE_FIRST_LISTENER = QUEUE_NEXT_LISTENER(MyProcNumber);
1932 : else
1933 : {
1934 1 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
1935 : {
1936 1 : if (QUEUE_NEXT_LISTENER(i) == MyProcNumber)
1937 : {
1938 1 : QUEUE_NEXT_LISTENER(i) = QUEUE_NEXT_LISTENER(MyProcNumber);
1939 1 : break;
1940 : }
1941 : }
1942 : }
1943 27 : QUEUE_NEXT_LISTENER(MyProcNumber) = INVALID_PROC_NUMBER;
1944 27 : LWLockRelease(NotifyQueueLock);
1945 :
1946 : /* mark ourselves as no longer listed in the global array */
1947 27 : amRegisteredListener = false;
1948 : }
1949 :
1950 : /*
1951 : * Test whether there is room to insert more notification messages.
1952 : *
1953 : * Caller must hold at least shared NotifyQueueLock.
1954 : */
1955 : static bool
1956 104 : asyncQueueIsFull(void)
1957 : {
1958 104 : int64 headPage = QUEUE_POS_PAGE(QUEUE_HEAD);
1959 104 : int64 tailPage = QUEUE_POS_PAGE(QUEUE_TAIL);
1960 104 : int64 occupied = headPage - tailPage;
1961 :
1962 104 : return occupied >= max_notify_queue_pages;
1963 : }
1964 :
1965 : /*
1966 : * Advance the QueuePosition to the next entry, assuming that the current
1967 : * entry is of length entryLength. If we jump to a new page the function
1968 : * returns true, else false.
1969 : */
1970 : static bool
1971 2543 : asyncQueueAdvance(volatile QueuePosition *position, int entryLength)
1972 : {
1973 2543 : int64 pageno = QUEUE_POS_PAGE(*position);
1974 2543 : int offset = QUEUE_POS_OFFSET(*position);
1975 2543 : bool pageJump = false;
1976 :
1977 : /*
1978 : * Move to the next writing position: First jump over what we have just
1979 : * written or read.
1980 : */
1981 2543 : offset += entryLength;
1982 : Assert(offset <= QUEUE_PAGESIZE);
1983 :
1984 : /*
1985 : * In a second step check if another entry can possibly be written to the
1986 : * page. If so, stay here, we have reached the next position. If not, then
1987 : * we need to move on to the next page.
1988 : */
1989 2543 : if (offset + QUEUEALIGN(AsyncQueueEntryEmptySize) > QUEUE_PAGESIZE)
1990 : {
1991 38 : pageno++;
1992 38 : offset = 0;
1993 38 : pageJump = true;
1994 : }
1995 :
1996 2543 : SET_QUEUE_POS(*position, pageno, offset);
1997 2543 : return pageJump;
1998 : }
1999 :
2000 : /*
2001 : * Fill the AsyncQueueEntry at *qe with an outbound notification message.
2002 : */
2003 : static void
2004 1146 : asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe)
2005 : {
2006 1146 : size_t channellen = n->channel_len;
2007 1146 : size_t payloadlen = n->payload_len;
2008 : int entryLength;
2009 :
2010 : Assert(channellen < NAMEDATALEN);
2011 : Assert(payloadlen < NOTIFY_PAYLOAD_MAX_LENGTH);
2012 :
2013 : /* The terminators are already included in AsyncQueueEntryEmptySize */
2014 1146 : entryLength = AsyncQueueEntryEmptySize + payloadlen + channellen;
2015 1146 : entryLength = QUEUEALIGN(entryLength);
2016 1146 : qe->length = entryLength;
2017 1146 : qe->dboid = MyDatabaseId;
2018 1146 : qe->xid = GetCurrentTransactionId();
2019 1146 : qe->srcPid = MyProcPid;
2020 1146 : memcpy(qe->data, n->data, channellen + payloadlen + 2);
2021 1146 : }
2022 :
2023 : /*
2024 : * Add pending notifications to the queue.
2025 : *
2026 : * We go page by page here, i.e. we stop once we have to go to a new page but
2027 : * we will be called again and then fill that next page. If an entry does not
2028 : * fit into the current page, we write a dummy entry with an InvalidOid as the
2029 : * database OID in order to fill the page. So every page is always used up to
2030 : * the last byte which simplifies reading the page later.
2031 : *
2032 : * We are passed the list cell (in pendingNotifies->events) containing the next
2033 : * notification to write and return the first still-unwritten cell back.
2034 : * Eventually we will return NULL indicating all is done.
2035 : *
2036 : * We are holding NotifyQueueLock already from the caller and grab
2037 : * page specific SLRU bank lock locally in this function.
2038 : */
2039 : static ListCell *
2040 104 : asyncQueueAddEntries(ListCell *nextNotify)
2041 : {
2042 : AsyncQueueEntry qe;
2043 : QueuePosition queue_head;
2044 : int64 pageno;
2045 : int offset;
2046 : int slotno;
2047 : LWLock *prevlock;
2048 :
2049 : /*
2050 : * We work with a local copy of QUEUE_HEAD, which we write back to shared
2051 : * memory upon exiting. The reason for this is that if we have to advance
2052 : * to a new page, SimpleLruZeroPage might fail (out of disk space, for
2053 : * instance), and we must not advance QUEUE_HEAD if it does. (Otherwise,
2054 : * subsequent insertions would try to put entries into a page that slru.c
2055 : * thinks doesn't exist yet.) So, use a local position variable. Note
2056 : * that if we do fail, any already-inserted queue entries are forgotten;
2057 : * this is okay, since they'd be useless anyway after our transaction
2058 : * rolls back.
2059 : */
2060 104 : queue_head = QUEUE_HEAD;
2061 :
2062 : /*
2063 : * If this is the first write since the postmaster started, we need to
2064 : * initialize the first page of the async SLRU. Otherwise, the current
2065 : * page should be initialized already, so just fetch it.
2066 : */
2067 104 : pageno = QUEUE_POS_PAGE(queue_head);
2068 104 : prevlock = SimpleLruGetBankLock(NotifyCtl, pageno);
2069 :
2070 : /* We hold both NotifyQueueLock and SLRU bank lock during this operation */
2071 104 : LWLockAcquire(prevlock, LW_EXCLUSIVE);
2072 :
2073 104 : if (QUEUE_POS_IS_ZERO(queue_head))
2074 9 : slotno = SimpleLruZeroPage(NotifyCtl, pageno);
2075 : else
2076 95 : slotno = SimpleLruReadPage(NotifyCtl, pageno, true, &queue_head);
2077 :
2078 : /* Note we mark the page dirty before writing in it */
2079 104 : NotifyCtl->shared->page_dirty[slotno] = true;
2080 :
2081 1215 : while (nextNotify != NULL)
2082 : {
2083 1146 : Notification *n = (Notification *) lfirst(nextNotify);
2084 :
2085 : /* Construct a valid queue entry in local variable qe */
2086 1146 : asyncQueueNotificationToEntry(n, &qe);
2087 :
2088 1146 : offset = QUEUE_POS_OFFSET(queue_head);
2089 :
2090 : /* Check whether the entry really fits on the current page */
2091 1146 : if (offset + qe.length <= QUEUE_PAGESIZE)
2092 : {
2093 : /* OK, so advance nextNotify past this item */
2094 1114 : nextNotify = lnext(pendingNotifies->events, nextNotify);
2095 : }
2096 : else
2097 : {
2098 : /*
2099 : * Write a dummy entry to fill up the page. Actually readers will
2100 : * only check dboid and since it won't match any reader's database
2101 : * OID, they will ignore this entry and move on.
2102 : */
2103 32 : qe.length = QUEUE_PAGESIZE - offset;
2104 32 : qe.dboid = InvalidOid;
2105 32 : qe.xid = InvalidTransactionId;
2106 32 : qe.data[0] = '\0'; /* empty channel */
2107 32 : qe.data[1] = '\0'; /* empty payload */
2108 : }
2109 :
2110 : /* Now copy qe into the shared buffer page */
2111 1146 : memcpy(NotifyCtl->shared->page_buffer[slotno] + offset,
2112 : &qe,
2113 1146 : qe.length);
2114 :
2115 : /* Advance queue_head appropriately, and detect if page is full */
2116 1146 : if (asyncQueueAdvance(&(queue_head), qe.length))
2117 : {
2118 : LWLock *lock;
2119 :
2120 35 : pageno = QUEUE_POS_PAGE(queue_head);
2121 35 : lock = SimpleLruGetBankLock(NotifyCtl, pageno);
2122 35 : if (lock != prevlock)
2123 : {
2124 0 : LWLockRelease(prevlock);
2125 0 : LWLockAcquire(lock, LW_EXCLUSIVE);
2126 0 : prevlock = lock;
2127 : }
2128 :
2129 : /*
2130 : * Page is full, so we're done here, but first fill the next page
2131 : * with zeroes. The reason to do this is to ensure that slru.c's
2132 : * idea of the head page is always the same as ours, which avoids
2133 : * boundary problems in SimpleLruTruncate. The test in
2134 : * asyncQueueIsFull() ensured that there is room to create this
2135 : * page without overrunning the queue.
2136 : */
2137 35 : slotno = SimpleLruZeroPage(NotifyCtl, QUEUE_POS_PAGE(queue_head));
2138 :
2139 : /*
2140 : * If the new page address is a multiple of QUEUE_CLEANUP_DELAY,
2141 : * set flag to remember that we should try to advance the tail
2142 : * pointer (we don't want to actually do that right here).
2143 : */
2144 35 : if (QUEUE_POS_PAGE(queue_head) % QUEUE_CLEANUP_DELAY == 0)
2145 8 : tryAdvanceTail = true;
2146 :
2147 : /* And exit the loop */
2148 35 : break;
2149 : }
2150 : }
2151 :
2152 : /* Success, so update the global QUEUE_HEAD */
2153 104 : QUEUE_HEAD = queue_head;
2154 :
2155 104 : LWLockRelease(prevlock);
2156 :
2157 104 : return nextNotify;
2158 : }
2159 :
2160 : /*
2161 : * SQL function to return the fraction of the notification queue currently
2162 : * occupied.
2163 : */
2164 : Datum
2165 6 : pg_notification_queue_usage(PG_FUNCTION_ARGS)
2166 : {
2167 : double usage;
2168 :
2169 : /* Advance the queue tail so we don't report a too-large result */
2170 6 : asyncQueueAdvanceTail();
2171 :
2172 6 : LWLockAcquire(NotifyQueueLock, LW_SHARED);
2173 6 : usage = asyncQueueUsage();
2174 6 : LWLockRelease(NotifyQueueLock);
2175 :
2176 6 : PG_RETURN_FLOAT8(usage);
2177 : }
2178 :
2179 : /*
2180 : * Return the fraction of the queue that is currently occupied.
2181 : *
2182 : * The caller must hold NotifyQueueLock in (at least) shared mode.
2183 : *
2184 : * Note: we measure the distance to the logical tail page, not the physical
2185 : * tail page. In some sense that's wrong, but the relative position of the
2186 : * physical tail is affected by details such as SLRU segment boundaries,
2187 : * so that a result based on that is unpleasantly unstable.
2188 : */
2189 : static double
2190 110 : asyncQueueUsage(void)
2191 : {
2192 110 : int64 headPage = QUEUE_POS_PAGE(QUEUE_HEAD);
2193 110 : int64 tailPage = QUEUE_POS_PAGE(QUEUE_TAIL);
2194 110 : int64 occupied = headPage - tailPage;
2195 :
2196 110 : if (occupied == 0)
2197 71 : return (double) 0; /* fast exit for common case */
2198 :
2199 39 : return (double) occupied / (double) max_notify_queue_pages;
2200 : }
2201 :
2202 : /*
2203 : * Check whether the queue is at least half full, and emit a warning if so.
2204 : *
2205 : * This is unlikely given the size of the queue, but possible.
2206 : * The warnings show up at most once every QUEUE_FULL_WARN_INTERVAL.
2207 : *
2208 : * Caller must hold exclusive NotifyQueueLock.
2209 : */
2210 : static void
2211 104 : asyncQueueFillWarning(void)
2212 : {
2213 : double fillDegree;
2214 : TimestampTz t;
2215 :
2216 104 : fillDegree = asyncQueueUsage();
2217 104 : if (fillDegree < 0.5)
2218 104 : return;
2219 :
2220 0 : t = GetCurrentTimestamp();
2221 :
2222 0 : if (TimestampDifferenceExceeds(asyncQueueControl->lastQueueFillWarn,
2223 : t, QUEUE_FULL_WARN_INTERVAL))
2224 : {
2225 0 : QueuePosition min = QUEUE_HEAD;
2226 0 : int32 minPid = InvalidPid;
2227 :
2228 0 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
2229 : {
2230 : Assert(QUEUE_BACKEND_PID(i) != InvalidPid);
2231 0 : min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
2232 0 : if (QUEUE_POS_EQUAL(min, QUEUE_BACKEND_POS(i)))
2233 0 : minPid = QUEUE_BACKEND_PID(i);
2234 : }
2235 :
2236 0 : ereport(WARNING,
2237 : (errmsg("NOTIFY queue is %.0f%% full", fillDegree * 100),
2238 : (minPid != InvalidPid ?
2239 : errdetail("The server process with PID %d is among those with the oldest transactions.", minPid)
2240 : : 0),
2241 : (minPid != InvalidPid ?
2242 : errhint("The NOTIFY queue cannot be emptied until that process ends its current transaction.")
2243 : : 0)));
2244 :
2245 0 : asyncQueueControl->lastQueueFillWarn = t;
2246 : }
2247 : }
2248 :
2249 : /*
2250 : * Send signals to listening backends.
2251 : *
2252 : * Normally we signal only backends that are interested in the notifies that
2253 : * we just sent. However, that will leave idle listeners falling further and
2254 : * further behind. Waken them anyway if they're far enough behind, so they'll
2255 : * advance their queue position pointers, allowing the global tail to advance.
2256 : *
2257 : * Since we know the ProcNumber and the Pid the signaling is quite cheap.
2258 : *
2259 : * This is called during CommitTransaction(), so it's important for it
2260 : * to have very low probability of failure.
2261 : */
2262 : static void
2263 69 : SignalBackends(void)
2264 : {
2265 : int count;
2266 :
2267 : /* Can't get here without PreCommit_Notify having made the global table */
2268 : Assert(globalChannelTable != NULL);
2269 :
2270 : /* It should have set up these arrays, too */
2271 : Assert(signalPids != NULL && signalProcnos != NULL);
2272 :
2273 : /*
2274 : * Identify backends that we need to signal. We don't want to send
2275 : * signals while holding the NotifyQueueLock, so this part just builds a
2276 : * list of target PIDs in signalPids[] and signalProcnos[].
2277 : */
2278 69 : count = 0;
2279 :
2280 69 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
2281 :
2282 : /* Scan each channel name that we notified in this transaction */
2283 210 : foreach_ptr(char, channel, pendingNotifies->uniqueChannelNames)
2284 : {
2285 : GlobalChannelKey key;
2286 : GlobalChannelEntry *entry;
2287 : ListenerEntry *listeners;
2288 :
2289 72 : GlobalChannelKeyInit(&key, MyDatabaseId, channel);
2290 72 : entry = dshash_find(globalChannelTable, &key, false);
2291 72 : if (entry == NULL)
2292 32 : continue; /* nobody is listening */
2293 :
2294 40 : listeners = (ListenerEntry *) dsa_get_address(globalChannelDSA,
2295 : entry->listenersArray);
2296 :
2297 : /*
2298 : * Identify listeners that now need waking, add them to arrays.
2299 : *
2300 : * Note that we signal listeners regardless of the state of their
2301 : * removeOnAbort flags. Hence a new listener that reached PreCommit,
2302 : * but then failed before AtCommit_Notify, can receive a signal even
2303 : * though it was never really listening. This is okay because it will
2304 : * not do anything in response to that signal. If we did not do it
2305 : * like this then a new listener might miss some messages due to the
2306 : * direct-advance logic below.
2307 : */
2308 83 : for (int j = 0; j < entry->numListeners; j++)
2309 : {
2310 43 : ProcNumber i = listeners[j].procNo;
2311 : int32 pid;
2312 : QueuePosition pos;
2313 :
2314 43 : if (QUEUE_BACKEND_WAKEUP_PENDING(i))
2315 12 : continue; /* already signaled, no need to repeat */
2316 :
2317 31 : pid = QUEUE_BACKEND_PID(i);
2318 31 : pos = QUEUE_BACKEND_POS(i);
2319 :
2320 31 : if (QUEUE_POS_EQUAL(pos, QUEUE_HEAD))
2321 0 : continue; /* it's fully caught up already */
2322 :
2323 : Assert(pid != InvalidPid);
2324 :
2325 31 : QUEUE_BACKEND_WAKEUP_PENDING(i) = true;
2326 31 : signalPids[count] = pid;
2327 31 : signalProcnos[count] = i;
2328 31 : count++;
2329 : }
2330 :
2331 40 : dshash_release_lock(globalChannelTable, entry);
2332 : }
2333 :
2334 : /*
2335 : * Scan all listeners. Any that are not already pending wakeup must not
2336 : * be interested in our notifications (else we'd have set their wakeup
2337 : * flags above). Check to see if we can directly advance their queue
2338 : * pointers to save a wakeup. Otherwise, if they are far behind, wake
2339 : * them anyway so they will catch up.
2340 : */
2341 124 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
2342 : {
2343 : int32 pid;
2344 : QueuePosition pos;
2345 :
2346 55 : if (QUEUE_BACKEND_WAKEUP_PENDING(i))
2347 40 : continue;
2348 :
2349 : /* If it's currently advancing, we should not touch it */
2350 15 : if (QUEUE_BACKEND_IS_ADVANCING(i))
2351 0 : continue;
2352 :
2353 15 : pid = QUEUE_BACKEND_PID(i);
2354 15 : pos = QUEUE_BACKEND_POS(i);
2355 :
2356 : /*
2357 : * We can directly advance the other backend's queue pointer if it's
2358 : * not currently advancing (else there are race conditions), and its
2359 : * current pointer is not behind queueHeadBeforeWrite (else we'd make
2360 : * it miss some older messages), and we'd not be moving the pointer
2361 : * backward.
2362 : */
2363 30 : if (!QUEUE_POS_PRECEDES(pos, queueHeadBeforeWrite) &&
2364 19 : QUEUE_POS_PRECEDES(pos, queueHeadAfterWrite))
2365 : {
2366 : /* We can directly advance its pointer past what we wrote */
2367 15 : QUEUE_BACKEND_POS(i) = queueHeadAfterWrite;
2368 : }
2369 0 : else if (asyncQueuePageDiff(QUEUE_POS_PAGE(QUEUE_HEAD),
2370 : QUEUE_POS_PAGE(pos)) >= QUEUE_CLEANUP_DELAY)
2371 : {
2372 : /* It's idle and far behind, so wake it up */
2373 : Assert(pid != InvalidPid);
2374 :
2375 0 : QUEUE_BACKEND_WAKEUP_PENDING(i) = true;
2376 0 : signalPids[count] = pid;
2377 0 : signalProcnos[count] = i;
2378 0 : count++;
2379 : }
2380 : }
2381 :
2382 69 : LWLockRelease(NotifyQueueLock);
2383 :
2384 : /* Now send signals */
2385 100 : for (int i = 0; i < count; i++)
2386 : {
2387 31 : int32 pid = signalPids[i];
2388 :
2389 : /*
2390 : * If we are signaling our own process, no need to involve the kernel;
2391 : * just set the flag directly.
2392 : */
2393 31 : if (pid == MyProcPid)
2394 : {
2395 21 : notifyInterruptPending = true;
2396 21 : continue;
2397 : }
2398 :
2399 : /*
2400 : * Note: assuming things aren't broken, a signal failure here could
2401 : * only occur if the target backend exited since we released
2402 : * NotifyQueueLock; which is unlikely but certainly possible. So we
2403 : * just log a low-level debug message if it happens.
2404 : */
2405 10 : if (SendProcSignal(pid, PROCSIG_NOTIFY_INTERRUPT, signalProcnos[i]) < 0)
2406 0 : elog(DEBUG3, "could not signal backend with PID %d: %m", pid);
2407 : }
2408 69 : }
2409 :
2410 : /*
2411 : * AtAbort_Notify
2412 : *
2413 : * This is called at transaction abort.
2414 : *
2415 : * Revert any staged listen/unlisten changes and clean up transaction state.
2416 : * This only does anything if we abort after PreCommit_Notify has staged
2417 : * some entries.
2418 : */
2419 : void
2420 35529 : AtAbort_Notify(void)
2421 : {
2422 : /* Revert staged listen/unlisten changes */
2423 35529 : ApplyPendingListenActions(false);
2424 :
2425 : /* If we're no longer listening on anything, unregister */
2426 35529 : if (amRegisteredListener && LocalChannelTableIsEmpty())
2427 0 : asyncQueueUnregister();
2428 :
2429 : /* And clean up */
2430 35529 : ClearPendingActionsAndNotifies();
2431 35529 : }
2432 :
2433 : /*
2434 : * AtSubCommit_Notify() --- Take care of subtransaction commit.
2435 : *
2436 : * Reassign all items in the pending lists to the parent transaction.
2437 : */
2438 : void
2439 5631 : AtSubCommit_Notify(void)
2440 : {
2441 5631 : int my_level = GetCurrentTransactionNestLevel();
2442 :
2443 : /* If there are actions at our nesting level, we must reparent them. */
2444 5631 : if (pendingActions != NULL &&
2445 2 : pendingActions->nestingLevel >= my_level)
2446 : {
2447 2 : if (pendingActions->upper == NULL ||
2448 1 : pendingActions->upper->nestingLevel < my_level - 1)
2449 : {
2450 : /* nothing to merge; give the whole thing to the parent */
2451 1 : --pendingActions->nestingLevel;
2452 : }
2453 : else
2454 : {
2455 1 : ActionList *childPendingActions = pendingActions;
2456 :
2457 1 : pendingActions = pendingActions->upper;
2458 :
2459 : /*
2460 : * Mustn't try to eliminate duplicates here --- see queue_listen()
2461 : */
2462 2 : pendingActions->actions =
2463 1 : list_concat(pendingActions->actions,
2464 1 : childPendingActions->actions);
2465 1 : pfree(childPendingActions);
2466 : }
2467 : }
2468 :
2469 : /* If there are notifies at our nesting level, we must reparent them. */
2470 5631 : if (pendingNotifies != NULL &&
2471 3 : pendingNotifies->nestingLevel >= my_level)
2472 : {
2473 : Assert(pendingNotifies->nestingLevel == my_level);
2474 :
2475 2 : if (pendingNotifies->upper == NULL ||
2476 1 : pendingNotifies->upper->nestingLevel < my_level - 1)
2477 : {
2478 : /* nothing to merge; give the whole thing to the parent */
2479 1 : --pendingNotifies->nestingLevel;
2480 : }
2481 : else
2482 : {
2483 : /*
2484 : * Formerly, we didn't bother to eliminate duplicates here, but
2485 : * now we must, else we fall foul of "Assert(!found)", either here
2486 : * or during a later attempt to build the parent-level hashtable.
2487 : */
2488 1 : NotificationList *childPendingNotifies = pendingNotifies;
2489 : ListCell *l;
2490 :
2491 1 : pendingNotifies = pendingNotifies->upper;
2492 : /* Insert all the subxact's events into parent, except for dups */
2493 5 : foreach(l, childPendingNotifies->events)
2494 : {
2495 4 : Notification *childn = (Notification *) lfirst(l);
2496 :
2497 4 : if (!AsyncExistsPendingNotify(childn))
2498 2 : AddEventToPendingNotifies(childn);
2499 : }
2500 1 : pfree(childPendingNotifies);
2501 : }
2502 : }
2503 5631 : }
2504 :
2505 : /*
2506 : * AtSubAbort_Notify() --- Take care of subtransaction abort.
2507 : */
2508 : void
2509 5421 : AtSubAbort_Notify(void)
2510 : {
2511 5421 : int my_level = GetCurrentTransactionNestLevel();
2512 :
2513 : /*
2514 : * All we have to do is pop the stack --- the actions/notifies made in
2515 : * this subxact are no longer interesting, and the space will be freed
2516 : * when CurTransactionContext is recycled. We still have to free the
2517 : * ActionList and NotificationList objects themselves, though, because
2518 : * those are allocated in TopTransactionContext.
2519 : *
2520 : * Note that there might be no entries at all, or no entries for the
2521 : * current subtransaction level, either because none were ever created, or
2522 : * because we reentered this routine due to trouble during subxact abort.
2523 : */
2524 5422 : while (pendingActions != NULL &&
2525 1 : pendingActions->nestingLevel >= my_level)
2526 : {
2527 1 : ActionList *childPendingActions = pendingActions;
2528 :
2529 1 : pendingActions = pendingActions->upper;
2530 1 : pfree(childPendingActions);
2531 : }
2532 :
2533 5422 : while (pendingNotifies != NULL &&
2534 2 : pendingNotifies->nestingLevel >= my_level)
2535 : {
2536 1 : NotificationList *childPendingNotifies = pendingNotifies;
2537 :
2538 1 : pendingNotifies = pendingNotifies->upper;
2539 1 : pfree(childPendingNotifies);
2540 : }
2541 5421 : }
2542 :
2543 : /*
2544 : * HandleNotifyInterrupt
2545 : *
2546 : * Signal handler portion of interrupt handling. Let the backend know
2547 : * that there's a pending notify interrupt. If we're currently reading
2548 : * from the client, this will interrupt the read and
2549 : * ProcessClientReadInterrupt() will call ProcessNotifyInterrupt().
2550 : */
2551 : void
2552 9 : HandleNotifyInterrupt(void)
2553 : {
2554 : /*
2555 : * Note: this is called by a SIGNAL HANDLER. You must be very wary what
2556 : * you do here.
2557 : */
2558 :
2559 : /* signal that work needs to be done */
2560 9 : notifyInterruptPending = true;
2561 :
2562 : /* latch will be set by procsignal_sigusr1_handler */
2563 9 : }
2564 :
2565 : /*
2566 : * ProcessNotifyInterrupt
2567 : *
2568 : * This is called if we see notifyInterruptPending set, just before
2569 : * transmitting ReadyForQuery at the end of a frontend command, and
2570 : * also if a notify signal occurs while reading from the frontend.
2571 : * HandleNotifyInterrupt() will cause the read to be interrupted
2572 : * via the process's latch, and this routine will get called.
2573 : * If we are truly idle (ie, *not* inside a transaction block),
2574 : * process the incoming notifies.
2575 : *
2576 : * If "flush" is true, force any frontend messages out immediately.
2577 : * This can be false when being called at the end of a frontend command,
2578 : * since we'll flush after sending ReadyForQuery.
2579 : */
2580 : void
2581 35 : ProcessNotifyInterrupt(bool flush)
2582 : {
2583 35 : if (IsTransactionOrTransactionBlock())
2584 6 : return; /* not really idle */
2585 :
2586 : /* Loop in case another signal arrives while sending messages */
2587 58 : while (notifyInterruptPending)
2588 29 : ProcessIncomingNotify(flush);
2589 : }
2590 :
2591 :
2592 : /*
2593 : * Read all pending notifications from the queue, and deliver appropriate
2594 : * ones to my frontend. Stop when we reach queue head or an uncommitted
2595 : * notification.
2596 : */
2597 : static void
2598 47 : asyncQueueReadAllNotifications(void)
2599 : {
2600 : QueuePosition pos;
2601 : QueuePosition head;
2602 : Snapshot snapshot;
2603 :
2604 : /*
2605 : * Fetch current state, indicate to others that we have woken up, and that
2606 : * we are in process of advancing our position.
2607 : */
2608 47 : LWLockAcquire(NotifyQueueLock, LW_SHARED);
2609 : /* Assert checks that we have a valid state entry */
2610 : Assert(MyProcPid == QUEUE_BACKEND_PID(MyProcNumber));
2611 47 : QUEUE_BACKEND_WAKEUP_PENDING(MyProcNumber) = false;
2612 47 : pos = QUEUE_BACKEND_POS(MyProcNumber);
2613 47 : head = QUEUE_HEAD;
2614 :
2615 47 : if (QUEUE_POS_EQUAL(pos, head))
2616 : {
2617 : /* Nothing to do, we have read all notifications already. */
2618 0 : LWLockRelease(NotifyQueueLock);
2619 0 : return;
2620 : }
2621 :
2622 47 : QUEUE_BACKEND_IS_ADVANCING(MyProcNumber) = true;
2623 47 : LWLockRelease(NotifyQueueLock);
2624 :
2625 : /*----------
2626 : * Get snapshot we'll use to decide which xacts are still in progress.
2627 : * This is trickier than it might seem, because of race conditions.
2628 : * Consider the following example:
2629 : *
2630 : * Backend 1: Backend 2:
2631 : *
2632 : * transaction starts
2633 : * UPDATE foo SET ...;
2634 : * NOTIFY foo;
2635 : * commit starts
2636 : * queue the notify message
2637 : * transaction starts
2638 : * LISTEN foo; -- first LISTEN in session
2639 : * SELECT * FROM foo WHERE ...;
2640 : * commit to clog
2641 : * commit starts
2642 : * add backend 2 to array of listeners
2643 : * advance to queue head (this code)
2644 : * commit to clog
2645 : *
2646 : * Transaction 2's SELECT has not seen the UPDATE's effects, since that
2647 : * wasn't committed yet. Ideally we'd ensure that client 2 would
2648 : * eventually get transaction 1's notify message, but there's no way
2649 : * to do that; until we're in the listener array, there's no guarantee
2650 : * that the notify message doesn't get removed from the queue.
2651 : *
2652 : * Therefore the coding technique transaction 2 is using is unsafe:
2653 : * applications must commit a LISTEN before inspecting database state,
2654 : * if they want to ensure they will see notifications about subsequent
2655 : * changes to that state.
2656 : *
2657 : * What we do guarantee is that we'll see all notifications from
2658 : * transactions committing after the snapshot we take here.
2659 : * BecomeRegisteredListener has already added us to the listener array,
2660 : * so no not-yet-committed messages can be removed from the queue
2661 : * before we see them.
2662 : *----------
2663 : */
2664 47 : snapshot = RegisterSnapshot(GetLatestSnapshot());
2665 :
2666 : /*
2667 : * It is possible that we fail while trying to send a message to our
2668 : * frontend (for example, because of encoding conversion failure). If
2669 : * that happens it is critical that we not try to send the same message
2670 : * over and over again. Therefore, we set ExitOnAnyError to upgrade any
2671 : * ERRORs to FATAL, causing the client connection to be closed on error.
2672 : *
2673 : * We used to only skip over the offending message and try to soldier on,
2674 : * but it was somewhat questionable to lose a notification and give the
2675 : * client an ERROR instead. A client application is not be prepared for
2676 : * that and can't tell that a notification was missed. It was also not
2677 : * very useful in practice because notifications are often processed while
2678 : * a connection is idle and reading a message from the client, and in that
2679 : * state, any error is upgraded to FATAL anyway. Closing the connection
2680 : * is a clear signal to the application that it might have missed
2681 : * notifications.
2682 : */
2683 : {
2684 47 : bool save_ExitOnAnyError = ExitOnAnyError;
2685 : bool reachedStop;
2686 :
2687 47 : ExitOnAnyError = true;
2688 :
2689 : do
2690 : {
2691 : /*
2692 : * Process messages up to the stop position, end of page, or an
2693 : * uncommitted message.
2694 : *
2695 : * Our stop position is what we found to be the head's position
2696 : * when we entered this function. It might have changed already.
2697 : * But if it has, we will receive (or have already received and
2698 : * queued) another signal and come here again.
2699 : *
2700 : * We are not holding NotifyQueueLock here! The queue can only
2701 : * extend beyond the head pointer (see above) and we leave our
2702 : * backend's pointer where it is so nobody will truncate or
2703 : * rewrite pages under us. Especially we don't want to hold a lock
2704 : * while sending the notifications to the frontend.
2705 : */
2706 50 : reachedStop = asyncQueueProcessPageEntries(&pos, head, snapshot);
2707 50 : } while (!reachedStop);
2708 :
2709 : /* Update shared state */
2710 47 : LWLockAcquire(NotifyQueueLock, LW_SHARED);
2711 47 : QUEUE_BACKEND_POS(MyProcNumber) = pos;
2712 47 : QUEUE_BACKEND_IS_ADVANCING(MyProcNumber) = false;
2713 47 : LWLockRelease(NotifyQueueLock);
2714 :
2715 47 : ExitOnAnyError = save_ExitOnAnyError;
2716 : }
2717 :
2718 : /* Done with snapshot */
2719 47 : UnregisterSnapshot(snapshot);
2720 : }
2721 :
2722 : /*
2723 : * Fetch notifications from the shared queue, beginning at position current,
2724 : * and deliver relevant ones to my frontend.
2725 : *
2726 : * The function returns true once we have reached the stop position or an
2727 : * uncommitted notification, and false if we have finished with the page.
2728 : * In other words: once it returns true there is no need to look further.
2729 : * The QueuePosition *current is advanced past all processed messages.
2730 : */
2731 : static bool
2732 50 : asyncQueueProcessPageEntries(QueuePosition *current,
2733 : QueuePosition stop,
2734 : Snapshot snapshot)
2735 : {
2736 50 : int64 curpage = QUEUE_POS_PAGE(*current);
2737 : int slotno;
2738 : char *page_buffer;
2739 50 : bool reachedStop = false;
2740 : bool reachedEndOfPage;
2741 :
2742 : /*
2743 : * We copy the entries into a local buffer to avoid holding the SLRU lock
2744 : * while we transmit them to our frontend. The local buffer must be
2745 : * adequately aligned.
2746 : */
2747 : alignas(AsyncQueueEntry) char local_buf[QUEUE_PAGESIZE];
2748 50 : char *local_buf_end = local_buf;
2749 :
2750 50 : slotno = SimpleLruReadPage_ReadOnly(NotifyCtl, curpage, current);
2751 50 : page_buffer = NotifyCtl->shared->page_buffer[slotno];
2752 :
2753 : do
2754 : {
2755 1414 : QueuePosition thisentry = *current;
2756 : AsyncQueueEntry *qe;
2757 :
2758 1414 : if (QUEUE_POS_EQUAL(thisentry, stop))
2759 47 : break;
2760 :
2761 1367 : qe = (AsyncQueueEntry *) (page_buffer + QUEUE_POS_OFFSET(thisentry));
2762 :
2763 : /*
2764 : * Advance *current over this message, possibly to the next page. As
2765 : * noted in the comments for asyncQueueReadAllNotifications, we must
2766 : * do this before possibly failing while processing the message.
2767 : */
2768 1367 : reachedEndOfPage = asyncQueueAdvance(current, qe->length);
2769 :
2770 : /* Ignore messages destined for other databases */
2771 1367 : if (qe->dboid == MyDatabaseId)
2772 : {
2773 1367 : if (XidInMVCCSnapshot(qe->xid, snapshot))
2774 : {
2775 : /*
2776 : * The source transaction is still in progress, so we can't
2777 : * process this message yet. Break out of the loop, but first
2778 : * back up *current so we will reprocess the message next
2779 : * time. (Note: it is unlikely but not impossible for
2780 : * TransactionIdDidCommit to fail, so we can't really avoid
2781 : * this advance-then-back-up behavior when dealing with an
2782 : * uncommitted message.)
2783 : *
2784 : * Note that we must test XidInMVCCSnapshot before we test
2785 : * TransactionIdDidCommit, else we might return a message from
2786 : * a transaction that is not yet visible to snapshots; compare
2787 : * the comments at the head of heapam_visibility.c.
2788 : *
2789 : * Also, while our own xact won't be listed in the snapshot,
2790 : * we need not check for TransactionIdIsCurrentTransactionId
2791 : * because our transaction cannot (yet) have queued any
2792 : * messages.
2793 : */
2794 0 : *current = thisentry;
2795 0 : reachedStop = true;
2796 0 : break;
2797 : }
2798 :
2799 : /*
2800 : * Quick check for the case that we're not listening on any
2801 : * channels, before calling TransactionIdDidCommit(). This makes
2802 : * that case a little faster, but more importantly, it ensures
2803 : * that if there's a bad entry in the queue for which
2804 : * TransactionIdDidCommit() fails for some reason, we can skip
2805 : * over it on the first LISTEN in a session, and not get stuck on
2806 : * it indefinitely. (This is a little trickier than it looks: it
2807 : * works because BecomeRegisteredListener runs this code before we
2808 : * have made the first entry in localChannelTable.)
2809 : */
2810 1367 : if (LocalChannelTableIsEmpty())
2811 1314 : continue;
2812 :
2813 53 : if (TransactionIdDidCommit(qe->xid))
2814 : {
2815 53 : memcpy(local_buf_end, qe, qe->length);
2816 53 : local_buf_end += qe->length;
2817 : }
2818 : else
2819 : {
2820 : /*
2821 : * The source transaction aborted or crashed, so we just
2822 : * ignore its notifications.
2823 : */
2824 : }
2825 : }
2826 :
2827 : /* Loop back if we're not at end of page */
2828 1367 : } while (!reachedEndOfPage);
2829 :
2830 : /* Release lock that we got from SimpleLruReadPage_ReadOnly() */
2831 50 : LWLockRelease(SimpleLruGetBankLock(NotifyCtl, curpage));
2832 :
2833 : /*
2834 : * Now that we have let go of the SLRU bank lock, send the notifications
2835 : * to our backend
2836 : */
2837 : Assert(local_buf_end - local_buf <= BLCKSZ);
2838 103 : for (char *p = local_buf; p < local_buf_end;)
2839 : {
2840 53 : AsyncQueueEntry *qe = (AsyncQueueEntry *) p;
2841 :
2842 : /* qe->data is the null-terminated channel name */
2843 53 : char *channel = qe->data;
2844 :
2845 53 : if (IsListeningOn(channel))
2846 : {
2847 : /* payload follows channel name */
2848 53 : char *payload = qe->data + strlen(channel) + 1;
2849 :
2850 53 : NotifyMyFrontEnd(channel, payload, qe->srcPid);
2851 : }
2852 :
2853 53 : p += qe->length;
2854 : }
2855 :
2856 50 : if (QUEUE_POS_EQUAL(*current, stop))
2857 47 : reachedStop = true;
2858 :
2859 50 : return reachedStop;
2860 : }
2861 :
2862 : /*
2863 : * Advance the shared queue tail variable to the minimum of all the
2864 : * per-backend tail pointers. Truncate pg_notify space if possible.
2865 : *
2866 : * This is (usually) called during CommitTransaction(), so it's important for
2867 : * it to have very low probability of failure.
2868 : */
2869 : static void
2870 14 : asyncQueueAdvanceTail(void)
2871 : {
2872 : QueuePosition min;
2873 : int64 oldtailpage;
2874 : int64 newtailpage;
2875 : int64 boundary;
2876 :
2877 : /* Restrict task to one backend per cluster; see SimpleLruTruncate(). */
2878 14 : LWLockAcquire(NotifyQueueTailLock, LW_EXCLUSIVE);
2879 :
2880 : /*
2881 : * Compute the new tail. Pre-v13, it's essential that QUEUE_TAIL be exact
2882 : * (ie, exactly match at least one backend's queue position), so it must
2883 : * be updated atomically with the actual computation. Since v13, we could
2884 : * get away with not doing it like that, but it seems prudent to keep it
2885 : * so.
2886 : *
2887 : * Also, because incoming backends will scan forward from QUEUE_TAIL, that
2888 : * must be advanced before we can truncate any data. Thus, QUEUE_TAIL is
2889 : * the logical tail, while QUEUE_STOP_PAGE is the physical tail, or oldest
2890 : * un-truncated page. When QUEUE_STOP_PAGE != QUEUE_POS_PAGE(QUEUE_TAIL),
2891 : * there are pages we can truncate but haven't yet finished doing so.
2892 : *
2893 : * For concurrency's sake, we don't want to hold NotifyQueueLock while
2894 : * performing SimpleLruTruncate. This is OK because no backend will try
2895 : * to access the pages we are in the midst of truncating.
2896 : */
2897 14 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
2898 14 : min = QUEUE_HEAD;
2899 24 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
2900 : {
2901 : Assert(QUEUE_BACKEND_PID(i) != InvalidPid);
2902 10 : min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
2903 : }
2904 14 : QUEUE_TAIL = min;
2905 14 : oldtailpage = QUEUE_STOP_PAGE;
2906 14 : LWLockRelease(NotifyQueueLock);
2907 :
2908 : /*
2909 : * We can truncate something if the global tail advanced across an SLRU
2910 : * segment boundary.
2911 : *
2912 : * XXX it might be better to truncate only once every several segments, to
2913 : * reduce the number of directory scans.
2914 : */
2915 14 : newtailpage = QUEUE_POS_PAGE(min);
2916 14 : boundary = newtailpage - (newtailpage % SLRU_PAGES_PER_SEGMENT);
2917 14 : if (asyncQueuePagePrecedes(oldtailpage, boundary))
2918 : {
2919 : /*
2920 : * SimpleLruTruncate() will ask for SLRU bank locks but will also
2921 : * release the lock again.
2922 : */
2923 1 : SimpleLruTruncate(NotifyCtl, newtailpage);
2924 :
2925 1 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
2926 1 : QUEUE_STOP_PAGE = newtailpage;
2927 1 : LWLockRelease(NotifyQueueLock);
2928 : }
2929 :
2930 14 : LWLockRelease(NotifyQueueTailLock);
2931 14 : }
2932 :
2933 : /*
2934 : * AsyncNotifyFreezeXids
2935 : *
2936 : * Prepare the async notification queue for CLOG truncation by freezing
2937 : * transaction IDs that are about to become inaccessible.
2938 : *
2939 : * This function is called by VACUUM before advancing datfrozenxid. It scans
2940 : * the notification queue and replaces XIDs that would become inaccessible
2941 : * after CLOG truncation with special markers:
2942 : * - Committed transactions are set to FrozenTransactionId
2943 : * - Aborted/crashed transactions are set to InvalidTransactionId
2944 : *
2945 : * Only XIDs < newFrozenXid are processed, as those are the ones whose CLOG
2946 : * pages will be truncated. If XID < newFrozenXid, it cannot still be running
2947 : * (or it would have held back newFrozenXid through ProcArray).
2948 : * Therefore, if TransactionIdDidCommit returns false, we know the transaction
2949 : * either aborted explicitly or crashed, and we can safely mark it invalid.
2950 : */
2951 : void
2952 1128 : AsyncNotifyFreezeXids(TransactionId newFrozenXid)
2953 : {
2954 : QueuePosition pos;
2955 : QueuePosition head;
2956 1128 : int64 curpage = -1;
2957 1128 : int slotno = -1;
2958 1128 : char *page_buffer = NULL;
2959 1128 : bool page_dirty = false;
2960 :
2961 : /*
2962 : * Acquire locks in the correct order to avoid deadlocks. As per the
2963 : * locking protocol: NotifyQueueTailLock, then NotifyQueueLock, then SLRU
2964 : * bank locks.
2965 : *
2966 : * We only need SHARED mode since we're just reading the head/tail
2967 : * positions, not modifying them.
2968 : */
2969 1128 : LWLockAcquire(NotifyQueueTailLock, LW_SHARED);
2970 1128 : LWLockAcquire(NotifyQueueLock, LW_SHARED);
2971 :
2972 1128 : pos = QUEUE_TAIL;
2973 1128 : head = QUEUE_HEAD;
2974 :
2975 : /* Release NotifyQueueLock early, we only needed to read the positions */
2976 1128 : LWLockRelease(NotifyQueueLock);
2977 :
2978 : /*
2979 : * Scan the queue from tail to head, freezing XIDs as needed. We hold
2980 : * NotifyQueueTailLock throughout to ensure the tail doesn't move while
2981 : * we're working.
2982 : */
2983 1158 : while (!QUEUE_POS_EQUAL(pos, head))
2984 : {
2985 : AsyncQueueEntry *qe;
2986 : TransactionId xid;
2987 30 : int64 pageno = QUEUE_POS_PAGE(pos);
2988 30 : int offset = QUEUE_POS_OFFSET(pos);
2989 :
2990 : /* If we need a different page, release old lock and get new one */
2991 30 : if (pageno != curpage)
2992 : {
2993 : LWLock *lock;
2994 :
2995 : /* Release previous page if any */
2996 3 : if (slotno >= 0)
2997 : {
2998 0 : if (page_dirty)
2999 : {
3000 0 : NotifyCtl->shared->page_dirty[slotno] = true;
3001 0 : page_dirty = false;
3002 : }
3003 0 : LWLockRelease(SimpleLruGetBankLock(NotifyCtl, curpage));
3004 : }
3005 :
3006 3 : lock = SimpleLruGetBankLock(NotifyCtl, pageno);
3007 3 : LWLockAcquire(lock, LW_EXCLUSIVE);
3008 3 : slotno = SimpleLruReadPage(NotifyCtl, pageno, true, &pos);
3009 3 : page_buffer = NotifyCtl->shared->page_buffer[slotno];
3010 3 : curpage = pageno;
3011 : }
3012 :
3013 30 : qe = (AsyncQueueEntry *) (page_buffer + offset);
3014 30 : xid = qe->xid;
3015 :
3016 60 : if (TransactionIdIsNormal(xid) &&
3017 30 : TransactionIdPrecedes(xid, newFrozenXid))
3018 : {
3019 10 : if (TransactionIdDidCommit(xid))
3020 : {
3021 10 : qe->xid = FrozenTransactionId;
3022 10 : page_dirty = true;
3023 : }
3024 : else
3025 : {
3026 0 : qe->xid = InvalidTransactionId;
3027 0 : page_dirty = true;
3028 : }
3029 : }
3030 :
3031 : /* Advance to next entry */
3032 30 : asyncQueueAdvance(&pos, qe->length);
3033 : }
3034 :
3035 : /* Release final page lock if we acquired one */
3036 1128 : if (slotno >= 0)
3037 : {
3038 3 : if (page_dirty)
3039 1 : NotifyCtl->shared->page_dirty[slotno] = true;
3040 3 : LWLockRelease(SimpleLruGetBankLock(NotifyCtl, curpage));
3041 : }
3042 :
3043 1128 : LWLockRelease(NotifyQueueTailLock);
3044 1128 : }
3045 :
3046 : /*
3047 : * ProcessIncomingNotify
3048 : *
3049 : * Scan the queue for arriving notifications and report them to the front
3050 : * end. The notifications might be from other sessions, or our own;
3051 : * there's no need to distinguish here.
3052 : *
3053 : * If "flush" is true, force any frontend messages out immediately.
3054 : *
3055 : * NOTE: since we are outside any transaction, we must create our own.
3056 : */
3057 : static void
3058 29 : ProcessIncomingNotify(bool flush)
3059 : {
3060 : /* We *must* reset the flag */
3061 29 : notifyInterruptPending = false;
3062 :
3063 : /* Do nothing else if we aren't actively listening */
3064 29 : if (LocalChannelTableIsEmpty())
3065 0 : return;
3066 :
3067 29 : if (Trace_notify)
3068 0 : elog(DEBUG1, "ProcessIncomingNotify");
3069 :
3070 29 : set_ps_display("notify interrupt");
3071 :
3072 : /*
3073 : * We must run asyncQueueReadAllNotifications inside a transaction, else
3074 : * bad things happen if it gets an error.
3075 : */
3076 29 : StartTransactionCommand();
3077 :
3078 29 : asyncQueueReadAllNotifications();
3079 :
3080 29 : CommitTransactionCommand();
3081 :
3082 : /*
3083 : * If this isn't an end-of-command case, we must flush the notify messages
3084 : * to ensure frontend gets them promptly.
3085 : */
3086 29 : if (flush)
3087 7 : pq_flush();
3088 :
3089 29 : set_ps_display("idle");
3090 :
3091 29 : if (Trace_notify)
3092 0 : elog(DEBUG1, "ProcessIncomingNotify: done");
3093 : }
3094 :
3095 : /*
3096 : * Send NOTIFY message to my front end.
3097 : */
3098 : void
3099 53 : NotifyMyFrontEnd(const char *channel, const char *payload, int32 srcPid)
3100 : {
3101 53 : if (whereToSendOutput == DestRemote)
3102 : {
3103 : StringInfoData buf;
3104 :
3105 53 : pq_beginmessage(&buf, PqMsg_NotificationResponse);
3106 53 : pq_sendint32(&buf, srcPid);
3107 53 : pq_sendstring(&buf, channel);
3108 53 : pq_sendstring(&buf, payload);
3109 53 : pq_endmessage(&buf);
3110 :
3111 : /*
3112 : * NOTE: we do not do pq_flush() here. Some level of caller will
3113 : * handle it later, allowing this message to be combined into a packet
3114 : * with other ones.
3115 : */
3116 : }
3117 : else
3118 0 : elog(INFO, "NOTIFY for \"%s\" payload \"%s\"", channel, payload);
3119 53 : }
3120 :
3121 : /* Does pendingNotifies include a match for the given event? */
3122 : static bool
3123 1066 : AsyncExistsPendingNotify(Notification *n)
3124 : {
3125 1066 : if (pendingNotifies == NULL)
3126 0 : return false;
3127 :
3128 1066 : if (pendingNotifies->hashtab != NULL)
3129 : {
3130 : /* Use the hash table to probe for a match */
3131 984 : if (hash_search(pendingNotifies->hashtab,
3132 : &n,
3133 : HASH_FIND,
3134 : NULL))
3135 1 : return true;
3136 : }
3137 : else
3138 : {
3139 : /* Must scan the event list */
3140 : ListCell *l;
3141 :
3142 425 : foreach(l, pendingNotifies->events)
3143 : {
3144 357 : Notification *oldn = (Notification *) lfirst(l);
3145 :
3146 357 : if (n->channel_len == oldn->channel_len &&
3147 357 : n->payload_len == oldn->payload_len &&
3148 190 : memcmp(n->data, oldn->data,
3149 190 : n->channel_len + n->payload_len + 2) == 0)
3150 14 : return true;
3151 : }
3152 : }
3153 :
3154 1051 : return false;
3155 : }
3156 :
3157 : /*
3158 : * Add a notification event to a pre-existing pendingNotifies list.
3159 : *
3160 : * Because pendingNotifies->events is already nonempty, this works
3161 : * correctly no matter what CurrentMemoryContext is.
3162 : */
3163 : static void
3164 1051 : AddEventToPendingNotifies(Notification *n)
3165 : {
3166 : Assert(pendingNotifies->events != NIL);
3167 :
3168 : /* Create the hash tables if it's time to */
3169 1051 : if (list_length(pendingNotifies->events) >= MIN_HASHABLE_NOTIFIES &&
3170 985 : pendingNotifies->hashtab == NULL)
3171 : {
3172 : HASHCTL hash_ctl;
3173 : ListCell *l;
3174 :
3175 : /* Create the hash table */
3176 2 : hash_ctl.keysize = sizeof(Notification *);
3177 2 : hash_ctl.entrysize = sizeof(struct NotificationHash);
3178 2 : hash_ctl.hash = notification_hash;
3179 2 : hash_ctl.match = notification_match;
3180 2 : hash_ctl.hcxt = CurTransactionContext;
3181 4 : pendingNotifies->hashtab =
3182 2 : hash_create("Pending Notifies",
3183 : 256L,
3184 : &hash_ctl,
3185 : HASH_ELEM | HASH_FUNCTION | HASH_COMPARE | HASH_CONTEXT);
3186 :
3187 : /* Create the unique channel name table */
3188 : Assert(pendingNotifies->uniqueChannelHash == NULL);
3189 2 : hash_ctl.keysize = NAMEDATALEN;
3190 2 : hash_ctl.entrysize = sizeof(ChannelName);
3191 2 : hash_ctl.hcxt = CurTransactionContext;
3192 4 : pendingNotifies->uniqueChannelHash =
3193 2 : hash_create("Pending Notify Channel Names",
3194 : 64L,
3195 : &hash_ctl,
3196 : HASH_ELEM | HASH_STRINGS | HASH_CONTEXT);
3197 :
3198 : /* Insert all the already-existing events */
3199 34 : foreach(l, pendingNotifies->events)
3200 : {
3201 32 : Notification *oldn = (Notification *) lfirst(l);
3202 32 : char *channel = oldn->data;
3203 : bool found;
3204 :
3205 32 : (void) hash_search(pendingNotifies->hashtab,
3206 : &oldn,
3207 : HASH_ENTER,
3208 : &found);
3209 : Assert(!found);
3210 :
3211 : /* Add channel name to uniqueChannelHash; might be there already */
3212 32 : (void) hash_search(pendingNotifies->uniqueChannelHash,
3213 : channel,
3214 : HASH_ENTER,
3215 : NULL);
3216 : }
3217 : }
3218 :
3219 : /* Add new event to the list, in order */
3220 1051 : pendingNotifies->events = lappend(pendingNotifies->events, n);
3221 :
3222 : /* Add event to the hash tables if needed */
3223 1051 : if (pendingNotifies->hashtab != NULL)
3224 : {
3225 985 : char *channel = n->data;
3226 : bool found;
3227 :
3228 985 : (void) hash_search(pendingNotifies->hashtab,
3229 : &n,
3230 : HASH_ENTER,
3231 : &found);
3232 : Assert(!found);
3233 :
3234 : /* Add channel name to uniqueChannelHash; might be there already */
3235 985 : (void) hash_search(pendingNotifies->uniqueChannelHash,
3236 : channel,
3237 : HASH_ENTER,
3238 : NULL);
3239 : }
3240 1051 : }
3241 :
3242 : /*
3243 : * notification_hash: hash function for notification hash table
3244 : *
3245 : * The hash "keys" are pointers to Notification structs.
3246 : */
3247 : static uint32
3248 2001 : notification_hash(const void *key, Size keysize)
3249 : {
3250 2001 : const Notification *k = *(const Notification *const *) key;
3251 :
3252 : Assert(keysize == sizeof(Notification *));
3253 : /* We don't bother to include the payload's trailing null in the hash */
3254 2001 : return DatumGetUInt32(hash_any((const unsigned char *) k->data,
3255 2001 : k->channel_len + k->payload_len + 1));
3256 : }
3257 :
3258 : /*
3259 : * notification_match: match function to use with notification_hash
3260 : */
3261 : static int
3262 1 : notification_match(const void *key1, const void *key2, Size keysize)
3263 : {
3264 1 : const Notification *k1 = *(const Notification *const *) key1;
3265 1 : const Notification *k2 = *(const Notification *const *) key2;
3266 :
3267 : Assert(keysize == sizeof(Notification *));
3268 1 : if (k1->channel_len == k2->channel_len &&
3269 1 : k1->payload_len == k2->payload_len &&
3270 1 : memcmp(k1->data, k2->data,
3271 1 : k1->channel_len + k1->payload_len + 2) == 0)
3272 1 : return 0; /* equal */
3273 0 : return 1; /* not equal */
3274 : }
3275 :
3276 : /* Clear the pendingActions and pendingNotifies lists. */
3277 : static void
3278 35680 : ClearPendingActionsAndNotifies(void)
3279 : {
3280 : /*
3281 : * Everything's allocated in either TopTransactionContext or the context
3282 : * for the subtransaction to which it corresponds. So, there's nothing to
3283 : * do here except reset the pointers; the space will be reclaimed when the
3284 : * contexts are deleted.
3285 : */
3286 35680 : pendingActions = NULL;
3287 35680 : pendingNotifies = NULL;
3288 : /* Also clear pendingListenActions, which is derived from pendingActions */
3289 35680 : pendingListenActions = NULL;
3290 35680 : }
3291 :
3292 : /*
3293 : * GUC check_hook for notify_buffers
3294 : */
3295 : bool
3296 1296 : check_notify_buffers(int *newval, void **extra, GucSource source)
3297 : {
3298 1296 : return check_slru_buffers("notify_buffers", newval);
3299 : }
|
