Branch data 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 : : #include "utils/tuplestore.h"
192 : :
193 : :
194 : : /*
195 : : * Maximum size of a NOTIFY payload, including terminating NULL. This
196 : : * must be kept small enough so that a notification message fits on one
197 : : * SLRU page. The magic fudge factor here is noncritical as long as it's
198 : : * more than AsyncQueueEntryEmptySize --- we make it significantly bigger
199 : : * than that, so changes in that data structure won't affect user-visible
200 : : * restrictions.
201 : : */
202 : : #define NOTIFY_PAYLOAD_MAX_LENGTH (BLCKSZ - NAMEDATALEN - 128)
203 : :
204 : : /*
205 : : * Struct representing an entry in the global notify queue
206 : : *
207 : : * This struct declaration has the maximal length, but in a real queue entry
208 : : * the data area is only big enough for the actual channel and payload strings
209 : : * (each null-terminated). AsyncQueueEntryEmptySize is the minimum possible
210 : : * entry size, if both channel and payload strings are empty (but note it
211 : : * doesn't include alignment padding).
212 : : *
213 : : * The "length" field should always be rounded up to the next QUEUEALIGN
214 : : * multiple so that all fields are properly aligned.
215 : : */
216 : : typedef struct AsyncQueueEntry
217 : : {
218 : : int length; /* total allocated length of entry */
219 : : Oid dboid; /* sender's database OID */
220 : : TransactionId xid; /* sender's XID */
221 : : int32 srcPid; /* sender's PID */
222 : : char data[NAMEDATALEN + NOTIFY_PAYLOAD_MAX_LENGTH];
223 : : } AsyncQueueEntry;
224 : :
225 : : /* Currently, no field of AsyncQueueEntry requires more than int alignment */
226 : : #define QUEUEALIGN(len) INTALIGN(len)
227 : :
228 : : #define AsyncQueueEntryEmptySize (offsetof(AsyncQueueEntry, data) + 2)
229 : :
230 : : /*
231 : : * Struct describing a queue position, and assorted macros for working with it
232 : : */
233 : : typedef struct QueuePosition
234 : : {
235 : : int64 page; /* SLRU page number */
236 : : int offset; /* byte offset within page */
237 : : } QueuePosition;
238 : :
239 : : #define QUEUE_POS_PAGE(x) ((x).page)
240 : : #define QUEUE_POS_OFFSET(x) ((x).offset)
241 : :
242 : : #define SET_QUEUE_POS(x,y,z) \
243 : : do { \
244 : : (x).page = (y); \
245 : : (x).offset = (z); \
246 : : } while (0)
247 : :
248 : : #define QUEUE_POS_EQUAL(x,y) \
249 : : ((x).page == (y).page && (x).offset == (y).offset)
250 : :
251 : : #define QUEUE_POS_IS_ZERO(x) \
252 : : ((x).page == 0 && (x).offset == 0)
253 : :
254 : : /* choose logically smaller QueuePosition */
255 : : #define QUEUE_POS_MIN(x,y) \
256 : : (asyncQueuePagePrecedes((x).page, (y).page) ? (x) : \
257 : : (x).page != (y).page ? (y) : \
258 : : (x).offset < (y).offset ? (x) : (y))
259 : :
260 : : /* choose logically larger QueuePosition */
261 : : #define QUEUE_POS_MAX(x,y) \
262 : : (asyncQueuePagePrecedes((x).page, (y).page) ? (y) : \
263 : : (x).page != (y).page ? (x) : \
264 : : (x).offset > (y).offset ? (x) : (y))
265 : :
266 : : /* returns true if x comes before y in queue order */
267 : : #define QUEUE_POS_PRECEDES(x,y) \
268 : : (asyncQueuePagePrecedes((x).page, (y).page) || \
269 : : ((x).page == (y).page && (x).offset < (y).offset))
270 : :
271 : : /*
272 : : * Parameter determining how often we try to advance the tail pointer:
273 : : * we do that after every QUEUE_CLEANUP_DELAY pages of NOTIFY data. This is
274 : : * also the distance by which a backend that's not interested in our
275 : : * notifications needs to be behind before we'll decide we need to wake it
276 : : * up so it can advance its pointer.
277 : : *
278 : : * Resist the temptation to make this really large. While that would save
279 : : * work in some places, it would add cost in others. In particular, this
280 : : * should likely be less than notify_buffers, to ensure that backends
281 : : * catch up before the pages they'll need to read fall out of SLRU cache.
282 : : */
283 : : #define QUEUE_CLEANUP_DELAY 4
284 : :
285 : : /*
286 : : * Struct describing a listening backend's status
287 : : */
288 : : typedef struct QueueBackendStatus
289 : : {
290 : : int32 pid; /* either a PID or InvalidPid */
291 : : Oid dboid; /* backend's database OID, or InvalidOid */
292 : : ProcNumber nextListener; /* id of next listener, or INVALID_PROC_NUMBER */
293 : : QueuePosition pos; /* backend has read queue up to here */
294 : : bool wakeupPending; /* signal sent to backend, not yet processed */
295 : : bool isAdvancing; /* backend is advancing its position */
296 : : } QueueBackendStatus;
297 : :
298 : : /*
299 : : * Shared memory state for LISTEN/NOTIFY (excluding its SLRU stuff)
300 : : *
301 : : * The AsyncQueueControl structure is protected by the NotifyQueueLock and
302 : : * NotifyQueueTailLock.
303 : : *
304 : : * When holding NotifyQueueLock in SHARED mode, backends may only inspect
305 : : * their own entries as well as the head and tail pointers. Consequently we
306 : : * can allow a backend to update its own record while holding only SHARED lock
307 : : * (since no other backend will inspect it).
308 : : *
309 : : * When holding NotifyQueueLock in EXCLUSIVE mode, backends can inspect the
310 : : * entries of other backends and also change the head pointer. They can
311 : : * also advance other backends' queue positions, unless the other backend
312 : : * has isAdvancing set (i.e., is in process of doing that itself).
313 : : *
314 : : * When holding both NotifyQueueLock and NotifyQueueTailLock in EXCLUSIVE
315 : : * mode, backends can change the tail pointers.
316 : : *
317 : : * SLRU buffer pool is divided in banks and bank wise SLRU lock is used as
318 : : * the control lock for the pg_notify SLRU buffers.
319 : : * In order to avoid deadlocks, whenever we need multiple locks, we first get
320 : : * NotifyQueueTailLock, then NotifyQueueLock, then SLRU bank lock, and lastly
321 : : * globalChannelTable partition locks.
322 : : *
323 : : * Each backend uses the backend[] array entry with index equal to its
324 : : * ProcNumber. We rely on this to make SendProcSignal fast.
325 : : *
326 : : * The backend[] array entries for actively-listening backends are threaded
327 : : * together using firstListener and the nextListener links, so that we can
328 : : * scan them without having to iterate over inactive entries. We keep this
329 : : * list in order by ProcNumber so that the scan is cache-friendly when there
330 : : * are many active entries.
331 : : */
332 : : typedef struct AsyncQueueControl
333 : : {
334 : : QueuePosition head; /* head points to the next free location */
335 : : QueuePosition tail; /* tail must be <= the queue position of every
336 : : * listening backend */
337 : : int64 stopPage; /* oldest unrecycled page; must be <=
338 : : * tail.page */
339 : : ProcNumber firstListener; /* id of first listener, or
340 : : * INVALID_PROC_NUMBER */
341 : : TimestampTz lastQueueFillWarn; /* time of last queue-full msg */
342 : : dsa_handle globalChannelTableDSA; /* global channel table's DSA handle */
343 : : dshash_table_handle globalChannelTableDSH; /* and its dshash handle */
344 : : /* Array with room for MaxBackends entries: */
345 : : QueueBackendStatus backend[FLEXIBLE_ARRAY_MEMBER];
346 : : } AsyncQueueControl;
347 : :
348 : : static AsyncQueueControl *asyncQueueControl;
349 : :
350 : : static void AsyncShmemRequest(void *arg);
351 : : static void AsyncShmemInit(void *arg);
352 : :
353 : : const ShmemCallbacks AsyncShmemCallbacks = {
354 : : .request_fn = AsyncShmemRequest,
355 : : .init_fn = AsyncShmemInit,
356 : : };
357 : :
358 : :
359 : : #define QUEUE_HEAD (asyncQueueControl->head)
360 : : #define QUEUE_TAIL (asyncQueueControl->tail)
361 : : #define QUEUE_STOP_PAGE (asyncQueueControl->stopPage)
362 : : #define QUEUE_FIRST_LISTENER (asyncQueueControl->firstListener)
363 : : #define QUEUE_BACKEND_PID(i) (asyncQueueControl->backend[i].pid)
364 : : #define QUEUE_BACKEND_DBOID(i) (asyncQueueControl->backend[i].dboid)
365 : : #define QUEUE_NEXT_LISTENER(i) (asyncQueueControl->backend[i].nextListener)
366 : : #define QUEUE_BACKEND_POS(i) (asyncQueueControl->backend[i].pos)
367 : : #define QUEUE_BACKEND_WAKEUP_PENDING(i) (asyncQueueControl->backend[i].wakeupPending)
368 : : #define QUEUE_BACKEND_IS_ADVANCING(i) (asyncQueueControl->backend[i].isAdvancing)
369 : :
370 : : /*
371 : : * The SLRU buffer area through which we access the notification queue
372 : : */
373 : : static inline bool asyncQueuePagePrecedes(int64 p, int64 q);
374 : : static int asyncQueueErrdetailForIoError(const void *opaque_data);
375 : :
376 : : static SlruDesc NotifySlruDesc;
377 : :
378 : :
379 : : #define NotifyCtl (&NotifySlruDesc)
380 : : #define QUEUE_PAGESIZE BLCKSZ
381 : :
382 : : #define QUEUE_FULL_WARN_INTERVAL 5000 /* warn at most once every 5s */
383 : :
384 : : /*
385 : : * Global channel table definitions
386 : : *
387 : : * This hash table maps (database OID, channel name) keys to arrays of
388 : : * ProcNumbers representing the backends listening or about to listen
389 : : * on each channel. The removeOnAbort flags allow us to create hash table
390 : : * entries pre-commit and not have to assume that creating them post-commit
391 : : * will succeed.
392 : : */
393 : : #define INITIAL_LISTENERS_ARRAY_SIZE 4
394 : :
395 : : typedef struct GlobalChannelKey
396 : : {
397 : : Oid dboid;
398 : : char channel[NAMEDATALEN];
399 : : } GlobalChannelKey;
400 : :
401 : : typedef struct ListenerEntry
402 : : {
403 : : ProcNumber procNo; /* listener's ProcNumber */
404 : : bool removeOnAbort; /* remove entry if current xact aborts */
405 : : } ListenerEntry;
406 : :
407 : : typedef struct GlobalChannelEntry
408 : : {
409 : : GlobalChannelKey key; /* hash key */
410 : : dsa_pointer listenersArray; /* DSA pointer to ListenerEntry array */
411 : : int numListeners; /* Number of listeners currently stored */
412 : : int allocatedListeners; /* Allocated size of array */
413 : : } GlobalChannelEntry;
414 : :
415 : : static dshash_table *globalChannelTable = NULL;
416 : : static dsa_area *globalChannelDSA = NULL;
417 : :
418 : : /*
419 : : * localChannelTable caches the channel names this backend is listening on
420 : : * (including those we have staged to be listened on, but not yet committed).
421 : : * Used by IsListeningOn() for fast lookups when reading notifications.
422 : : */
423 : : static HTAB *localChannelTable = NULL;
424 : :
425 : : /* We test this condition to detect that we're not listening at all */
426 : : #define LocalChannelTableIsEmpty() \
427 : : (localChannelTable == NULL || hash_get_num_entries(localChannelTable) == 0)
428 : :
429 : : /*
430 : : * State for pending LISTEN/UNLISTEN actions consists of an ordered list of
431 : : * all actions requested in the current transaction. As explained above,
432 : : * we don't actually change listen state until we reach transaction commit.
433 : : *
434 : : * The list is kept in CurTransactionContext. In subtransactions, each
435 : : * subtransaction has its own list in its own CurTransactionContext, but
436 : : * successful subtransactions attach their lists to their parent's list.
437 : : * Failed subtransactions simply discard their lists.
438 : : */
439 : : typedef enum
440 : : {
441 : : LISTEN_LISTEN,
442 : : LISTEN_UNLISTEN,
443 : : LISTEN_UNLISTEN_ALL,
444 : : } ListenActionKind;
445 : :
446 : : typedef struct
447 : : {
448 : : ListenActionKind action;
449 : : char channel[FLEXIBLE_ARRAY_MEMBER]; /* nul-terminated string */
450 : : } ListenAction;
451 : :
452 : : typedef struct ActionList
453 : : {
454 : : int nestingLevel; /* current transaction nesting depth */
455 : : List *actions; /* list of ListenAction structs */
456 : : struct ActionList *upper; /* details for upper transaction levels */
457 : : } ActionList;
458 : :
459 : : static ActionList *pendingActions = NULL;
460 : :
461 : : /*
462 : : * Hash table recording the final listen/unlisten intent per channel for
463 : : * the current transaction. Key is channel name, value is PENDING_LISTEN or
464 : : * PENDING_UNLISTEN. This keeps critical commit/abort processing to one step
465 : : * per channel instead of replaying every action. This is built from the
466 : : * pendingActions list by PreCommit_Notify, then used by AtCommit_Notify or
467 : : * AtAbort_Notify.
468 : : */
469 : : typedef enum
470 : : {
471 : : PENDING_LISTEN,
472 : : PENDING_UNLISTEN,
473 : : } PendingListenAction;
474 : :
475 : : typedef struct PendingListenEntry
476 : : {
477 : : char channel[NAMEDATALEN]; /* hash key */
478 : : PendingListenAction action; /* which action should we perform? */
479 : : } PendingListenEntry;
480 : :
481 : : static HTAB *pendingListenActions = NULL;
482 : :
483 : : /*
484 : : * State for outbound notifies consists of a list of all channels+payloads
485 : : * NOTIFYed in the current transaction. We do not actually perform a NOTIFY
486 : : * until and unless the transaction commits. pendingNotifies is NULL if no
487 : : * NOTIFYs have been done in the current (sub) transaction.
488 : : *
489 : : * We discard duplicate notify events issued in the same transaction.
490 : : * Hence, in addition to the list proper (which we need to track the order
491 : : * of the events, since we guarantee to deliver them in order), we build a
492 : : * hash table which we can probe to detect duplicates. Since building the
493 : : * hash table is somewhat expensive, we do so only once we have at least
494 : : * MIN_HASHABLE_NOTIFIES events queued in the current (sub) transaction;
495 : : * before that we just scan the events linearly.
496 : : *
497 : : * The list is kept in CurTransactionContext. In subtransactions, each
498 : : * subtransaction has its own list in its own CurTransactionContext, but
499 : : * successful subtransactions add their entries to their parent's list.
500 : : * Failed subtransactions simply discard their lists. Since these lists
501 : : * are independent, there may be notify events in a subtransaction's list
502 : : * that duplicate events in some ancestor (sub) transaction; we get rid of
503 : : * the dups when merging the subtransaction's list into its parent's.
504 : : *
505 : : * Note: the action and notify lists do not interact within a transaction.
506 : : * In particular, if a transaction does NOTIFY and then LISTEN on the same
507 : : * condition name, it will get a self-notify at commit. This is a bit odd
508 : : * but is consistent with our historical behavior.
509 : : */
510 : : typedef struct Notification
511 : : {
512 : : uint16 channel_len; /* length of channel-name string */
513 : : uint16 payload_len; /* length of payload string */
514 : : /* null-terminated channel name, then null-terminated payload follow */
515 : : char data[FLEXIBLE_ARRAY_MEMBER];
516 : : } Notification;
517 : :
518 : : typedef struct NotificationList
519 : : {
520 : : int nestingLevel; /* current transaction nesting depth */
521 : : List *events; /* list of Notification structs */
522 : : HTAB *hashtab; /* hash of NotificationHash structs, or NULL */
523 : : List *uniqueChannelNames; /* unique channel names being notified */
524 : : HTAB *uniqueChannelHash; /* hash of unique channel names, or NULL */
525 : : struct NotificationList *upper; /* details for upper transaction levels */
526 : : } NotificationList;
527 : :
528 : : #define MIN_HASHABLE_NOTIFIES 16 /* threshold to build hashtab */
529 : :
530 : : struct NotificationHash
531 : : {
532 : : Notification *event; /* => the actual Notification struct */
533 : : };
534 : :
535 : : static NotificationList *pendingNotifies = NULL;
536 : :
537 : : /*
538 : : * Hash entry in NotificationList.uniqueChannelHash or localChannelTable
539 : : * (both just carry the channel name, with no payload).
540 : : */
541 : : typedef struct ChannelName
542 : : {
543 : : char channel[NAMEDATALEN]; /* hash key */
544 : : } ChannelName;
545 : :
546 : : /*
547 : : * Inbound notifications are initially processed by HandleNotifyInterrupt(),
548 : : * called from inside a signal handler. That just sets the
549 : : * notifyInterruptPending flag and sets the process
550 : : * latch. ProcessNotifyInterrupt() will then be called whenever it's safe to
551 : : * actually deal with the interrupt.
552 : : */
553 : : volatile sig_atomic_t notifyInterruptPending = false;
554 : :
555 : : /* True if we've registered an on_shmem_exit cleanup */
556 : : static bool unlistenExitRegistered = false;
557 : :
558 : : /* True if we're currently registered as a listener in asyncQueueControl */
559 : : static bool amRegisteredListener = false;
560 : :
561 : : /*
562 : : * Queue head positions for direct advancement.
563 : : * These are captured during PreCommit_Notify while holding the heavyweight
564 : : * lock on database 0, ensuring no other backend can insert notifications
565 : : * between them. SignalBackends uses these to advance idle backends.
566 : : */
567 : : static QueuePosition queueHeadBeforeWrite;
568 : : static QueuePosition queueHeadAfterWrite;
569 : :
570 : : /*
571 : : * Workspace arrays for SignalBackends. These are preallocated in
572 : : * PreCommit_Notify to avoid needing memory allocation after committing to
573 : : * clog.
574 : : */
575 : : static int32 *signalPids = NULL;
576 : : static ProcNumber *signalProcnos = NULL;
577 : :
578 : : /* have we advanced to a page that's a multiple of QUEUE_CLEANUP_DELAY? */
579 : : static bool tryAdvanceTail = false;
580 : :
581 : : /* GUC parameters */
582 : : bool Trace_notify = false;
583 : :
584 : : /* For 8 KB pages this gives 8 GB of disk space */
585 : : int max_notify_queue_pages = 1048576;
586 : :
587 : : /* local function prototypes */
588 : : static inline int64 asyncQueuePageDiff(int64 p, int64 q);
589 : : static inline void GlobalChannelKeyInit(GlobalChannelKey *key, Oid dboid,
590 : : const char *channel);
591 : : static dshash_hash globalChannelTableHash(const void *key, size_t size,
592 : : void *arg);
593 : : static void initGlobalChannelTable(void);
594 : : static void initLocalChannelTable(void);
595 : : static void queue_listen(ListenActionKind action, const char *channel);
596 : : static void Async_UnlistenOnExit(int code, Datum arg);
597 : : static void BecomeRegisteredListener(void);
598 : : static void PrepareTableEntriesForListen(const char *channel);
599 : : static void PrepareTableEntriesForUnlisten(const char *channel);
600 : : static void PrepareTableEntriesForUnlistenAll(void);
601 : : static void RemoveListenerFromChannel(GlobalChannelEntry **entry_ptr,
602 : : ListenerEntry *listeners,
603 : : int idx);
604 : : static void ApplyPendingListenActions(bool isCommit);
605 : : static void CleanupListenersOnExit(void);
606 : : static bool IsListeningOn(const char *channel);
607 : : static void asyncQueueUnregister(void);
608 : : static bool asyncQueueIsFull(void);
609 : : static bool asyncQueueAdvance(QueuePosition *position, int entryLength);
610 : : static void asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe);
611 : : static ListCell *asyncQueueAddEntries(ListCell *nextNotify);
612 : : static double asyncQueueUsage(void);
613 : : static void asyncQueueFillWarning(void);
614 : : static void SignalBackends(void);
615 : : static void asyncQueueReadAllNotifications(void);
616 : : static bool asyncQueueProcessPageEntries(QueuePosition *current,
617 : : QueuePosition stop,
618 : : Snapshot snapshot);
619 : : static void asyncQueueAdvanceTail(void);
620 : : static void ProcessIncomingNotify(bool flush);
621 : : static bool AsyncExistsPendingNotify(Notification *n);
622 : : static void AddEventToPendingNotifies(Notification *n);
623 : : static uint32 notification_hash(const void *key, Size keysize);
624 : : static int notification_match(const void *key1, const void *key2, Size keysize);
625 : : static void ClearPendingActionsAndNotifies(void);
626 : :
627 : : static int
628 : 0 : asyncQueueErrdetailForIoError(const void *opaque_data)
629 : : {
630 : 0 : const QueuePosition *position = opaque_data;
631 : :
632 : 0 : return errdetail("Could not access async queue at page %" PRId64 ", offset %d.",
633 : 0 : position->page, position->offset);
634 : : }
635 : :
636 : : /*
637 : : * Compute the difference between two queue page numbers.
638 : : * Previously this function accounted for a wraparound.
639 : : */
640 : : static inline int64
641 : 0 : asyncQueuePageDiff(int64 p, int64 q)
642 : : {
643 : 0 : return p - q;
644 : : }
645 : :
646 : : /*
647 : : * Determines whether p precedes q.
648 : : * Previously this function accounted for a wraparound.
649 : : */
650 : : static inline bool
651 : 118 : asyncQueuePagePrecedes(int64 p, int64 q)
652 : : {
653 : 118 : return p < q;
654 : : }
655 : :
656 : : /*
657 : : * GlobalChannelKeyInit
658 : : * Prepare a global channel table key for hashing.
659 : : */
660 : : static inline void
661 : 203 : GlobalChannelKeyInit(GlobalChannelKey *key, Oid dboid, const char *channel)
662 : : {
663 : 203 : memset(key, 0, sizeof(GlobalChannelKey));
664 : 203 : key->dboid = dboid;
665 : 203 : strlcpy(key->channel, channel, NAMEDATALEN);
666 : 203 : }
667 : :
668 : : /*
669 : : * globalChannelTableHash
670 : : * Hash function for global channel table keys.
671 : : */
672 : : static dshash_hash
673 : 203 : globalChannelTableHash(const void *key, size_t size, void *arg)
674 : : {
675 : 203 : const GlobalChannelKey *k = (const GlobalChannelKey *) key;
676 : : dshash_hash h;
677 : :
678 : 203 : h = DatumGetUInt32(hash_uint32(k->dboid));
679 : 406 : h ^= hash_bytes((const unsigned char *) k->channel,
680 : 203 : strnlen(k->channel, NAMEDATALEN));
681 : :
682 : 203 : return h;
683 : : }
684 : :
685 : : /* parameters for the global channel table */
686 : : static const dshash_parameters globalChannelTableDSHParams = {
687 : : sizeof(GlobalChannelKey),
688 : : sizeof(GlobalChannelEntry),
689 : : dshash_memcmp,
690 : : globalChannelTableHash,
691 : : dshash_memcpy,
692 : : LWTRANCHE_NOTIFY_CHANNEL_HASH
693 : : };
694 : :
695 : : /*
696 : : * initGlobalChannelTable
697 : : * Lazy initialization of the global channel table.
698 : : */
699 : : static void
700 : 151 : initGlobalChannelTable(void)
701 : : {
702 : : MemoryContext oldcontext;
703 : :
704 : : /* Quick exit if we already did this */
705 [ + + ]: 151 : if (asyncQueueControl->globalChannelTableDSH != DSHASH_HANDLE_INVALID &&
706 [ + + ]: 143 : globalChannelTable != NULL)
707 : 121 : return;
708 : :
709 : : /* Otherwise, use a lock to ensure only one process creates the table */
710 : 30 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
711 : :
712 : : /* Be sure any local memory allocated by DSA routines is persistent */
713 : 30 : oldcontext = MemoryContextSwitchTo(TopMemoryContext);
714 : :
715 [ + + ]: 30 : if (asyncQueueControl->globalChannelTableDSH == DSHASH_HANDLE_INVALID)
716 : : {
717 : : /* Initialize dynamic shared hash table for global channels */
718 : 8 : globalChannelDSA = dsa_create(LWTRANCHE_NOTIFY_CHANNEL_HASH);
719 : 8 : dsa_pin(globalChannelDSA);
720 : 8 : dsa_pin_mapping(globalChannelDSA);
721 : 8 : globalChannelTable = dshash_create(globalChannelDSA,
722 : : &globalChannelTableDSHParams,
723 : : NULL);
724 : :
725 : : /* Store handles in shared memory for other backends to use */
726 : 8 : asyncQueueControl->globalChannelTableDSA = dsa_get_handle(globalChannelDSA);
727 : 8 : asyncQueueControl->globalChannelTableDSH =
728 : 8 : dshash_get_hash_table_handle(globalChannelTable);
729 : : }
730 [ + - ]: 22 : else if (!globalChannelTable)
731 : : {
732 : : /* Attach to existing dynamic shared hash table */
733 : 22 : globalChannelDSA = dsa_attach(asyncQueueControl->globalChannelTableDSA);
734 : 22 : dsa_pin_mapping(globalChannelDSA);
735 : 22 : globalChannelTable = dshash_attach(globalChannelDSA,
736 : : &globalChannelTableDSHParams,
737 : 22 : asyncQueueControl->globalChannelTableDSH,
738 : : NULL);
739 : : }
740 : :
741 : 30 : MemoryContextSwitchTo(oldcontext);
742 : 30 : LWLockRelease(NotifyQueueLock);
743 : : }
744 : :
745 : : /*
746 : : * initLocalChannelTable
747 : : * Lazy initialization of the local channel table.
748 : : * Once created, this table lasts for the life of the session.
749 : : */
750 : : static void
751 : 82 : initLocalChannelTable(void)
752 : : {
753 : : HASHCTL hash_ctl;
754 : :
755 : : /* Quick exit if we already did this */
756 [ + + ]: 82 : if (localChannelTable != NULL)
757 : 64 : return;
758 : :
759 : : /* Initialize local hash table for this backend's listened channels */
760 : 18 : hash_ctl.keysize = NAMEDATALEN;
761 : 18 : hash_ctl.entrysize = sizeof(ChannelName);
762 : :
763 : 18 : localChannelTable =
764 : 18 : hash_create("Local Listen Channels",
765 : : 64,
766 : : &hash_ctl,
767 : : HASH_ELEM | HASH_STRINGS);
768 : : }
769 : :
770 : : /*
771 : : * initPendingListenActions
772 : : * Lazy initialization of the pending listen actions hash table.
773 : : * This is allocated in CurTransactionContext during PreCommit_Notify,
774 : : * and destroyed at transaction end.
775 : : */
776 : : static void
777 : 82 : initPendingListenActions(void)
778 : : {
779 : : HASHCTL hash_ctl;
780 : :
781 [ - + ]: 82 : if (pendingListenActions != NULL)
782 : 0 : return;
783 : :
784 : 82 : hash_ctl.keysize = NAMEDATALEN;
785 : 82 : hash_ctl.entrysize = sizeof(PendingListenEntry);
786 : 82 : hash_ctl.hcxt = CurTransactionContext;
787 : :
788 : 82 : pendingListenActions =
789 : 82 : hash_create("Pending Listen Actions",
790 : 82 : list_length(pendingActions->actions),
791 : : &hash_ctl,
792 : : HASH_ELEM | HASH_STRINGS | HASH_CONTEXT);
793 : : }
794 : :
795 : : /*
796 : : * Register our shared memory needs
797 : : */
798 : : static void
799 : 1335 : AsyncShmemRequest(void *arg)
800 : : {
801 : : Size size;
802 : :
803 : 1335 : size = mul_size(MaxBackends, sizeof(QueueBackendStatus));
804 : 1335 : size = add_size(size, offsetof(AsyncQueueControl, backend));
805 : :
806 : 1335 : ShmemRequestStruct(.name = "Async Queue Control",
807 : : .size = size,
808 : : .ptr = (void **) &asyncQueueControl,
809 : : );
810 : :
811 : 1335 : SimpleLruRequest(.desc = &NotifySlruDesc,
812 : : .name = "notify",
813 : : .Dir = "pg_notify",
814 : :
815 : : /* long segment names are used in order to avoid wraparound */
816 : : .long_segment_names = true,
817 : :
818 : : .nslots = notify_buffers,
819 : :
820 : : .sync_handler = SYNC_HANDLER_NONE,
821 : : .PagePrecedes = asyncQueuePagePrecedes,
822 : : .errdetail_for_io_error = asyncQueueErrdetailForIoError,
823 : :
824 : : .buffer_tranche_id = LWTRANCHE_NOTIFY_BUFFER,
825 : : .bank_tranche_id = LWTRANCHE_NOTIFY_SLRU,
826 : : );
827 : 1335 : }
828 : :
829 : : static void
830 : 1331 : AsyncShmemInit(void *arg)
831 : : {
832 : 1331 : SET_QUEUE_POS(QUEUE_HEAD, 0, 0);
833 : 1331 : SET_QUEUE_POS(QUEUE_TAIL, 0, 0);
834 : 1331 : QUEUE_STOP_PAGE = 0;
835 : 1331 : QUEUE_FIRST_LISTENER = INVALID_PROC_NUMBER;
836 : 1331 : asyncQueueControl->lastQueueFillWarn = 0;
837 : 1331 : asyncQueueControl->globalChannelTableDSA = DSA_HANDLE_INVALID;
838 : 1331 : asyncQueueControl->globalChannelTableDSH = DSHASH_HANDLE_INVALID;
839 [ + + ]: 123408 : for (int i = 0; i < MaxBackends; i++)
840 : : {
841 : 122077 : QUEUE_BACKEND_PID(i) = InvalidPid;
842 : 122077 : QUEUE_BACKEND_DBOID(i) = InvalidOid;
843 : 122077 : QUEUE_NEXT_LISTENER(i) = INVALID_PROC_NUMBER;
844 : 122077 : SET_QUEUE_POS(QUEUE_BACKEND_POS(i), 0, 0);
845 : 122077 : QUEUE_BACKEND_WAKEUP_PENDING(i) = false;
846 : 122077 : QUEUE_BACKEND_IS_ADVANCING(i) = false;
847 : : }
848 : :
849 : : /*
850 : : * During start or reboot, clean out the pg_notify directory.
851 : : */
852 : 1331 : (void) SlruScanDirectory(NotifyCtl, SlruScanDirCbDeleteAll, NULL);
853 : 1331 : }
854 : :
855 : :
856 : : /*
857 : : * pg_notify -
858 : : * SQL function to send a notification event
859 : : */
860 : : Datum
861 : 1078 : pg_notify(PG_FUNCTION_ARGS)
862 : : {
863 : : const char *channel;
864 : : const char *payload;
865 : :
866 [ + + ]: 1078 : if (PG_ARGISNULL(0))
867 : 4 : channel = "";
868 : : else
869 : 1074 : channel = text_to_cstring(PG_GETARG_TEXT_PP(0));
870 : :
871 [ + + ]: 1078 : if (PG_ARGISNULL(1))
872 : 7 : payload = "";
873 : : else
874 : 1071 : payload = text_to_cstring(PG_GETARG_TEXT_PP(1));
875 : :
876 : : /* For NOTIFY as a statement, this is checked in ProcessUtility */
877 : 1078 : PreventCommandDuringRecovery("NOTIFY");
878 : :
879 : 1078 : Async_Notify(channel, payload);
880 : :
881 : 1066 : PG_RETURN_VOID();
882 : : }
883 : :
884 : :
885 : : /*
886 : : * Async_Notify
887 : : *
888 : : * This is executed by the SQL notify command.
889 : : *
890 : : * Adds the message to the list of pending notifies.
891 : : * Actual notification happens during transaction commit.
892 : : * ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
893 : : */
894 : : void
895 : 1145 : Async_Notify(const char *channel, const char *payload)
896 : : {
897 : 1145 : int my_level = GetCurrentTransactionNestLevel();
898 : : size_t channel_len;
899 : : size_t payload_len;
900 : : Notification *n;
901 : : MemoryContext oldcontext;
902 : :
903 [ - + ]: 1145 : if (IsParallelWorker())
904 [ # # ]: 0 : elog(ERROR, "cannot send notifications from a parallel worker");
905 : :
906 [ - + ]: 1145 : if (Trace_notify)
907 [ # # ]: 0 : elog(DEBUG1, "Async_Notify(%s)", channel);
908 : :
909 [ + - ]: 1145 : channel_len = channel ? strlen(channel) : 0;
910 [ + + ]: 1145 : payload_len = payload ? strlen(payload) : 0;
911 : :
912 : : /* a channel name must be specified */
913 [ + + ]: 1145 : if (channel_len == 0)
914 [ + - ]: 8 : ereport(ERROR,
915 : : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
916 : : errmsg("channel name cannot be empty")));
917 : :
918 : : /* enforce length limits */
919 [ + + ]: 1137 : if (channel_len >= NAMEDATALEN)
920 [ + - ]: 4 : ereport(ERROR,
921 : : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
922 : : errmsg("channel name too long")));
923 : :
924 [ - + ]: 1133 : if (payload_len >= NOTIFY_PAYLOAD_MAX_LENGTH)
925 [ # # ]: 0 : ereport(ERROR,
926 : : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
927 : : errmsg("payload string too long")));
928 : :
929 : : /*
930 : : * We must construct the Notification entry, even if we end up not using
931 : : * it, in order to compare it cheaply to existing list entries.
932 : : *
933 : : * The notification list needs to live until end of transaction, so store
934 : : * it in the transaction context.
935 : : */
936 : 1133 : oldcontext = MemoryContextSwitchTo(CurTransactionContext);
937 : :
938 : 1133 : n = (Notification *) palloc(offsetof(Notification, data) +
939 : 1133 : channel_len + payload_len + 2);
940 : 1133 : n->channel_len = channel_len;
941 : 1133 : n->payload_len = payload_len;
942 : 1133 : strcpy(n->data, channel);
943 [ + + ]: 1133 : if (payload)
944 : 1118 : strcpy(n->data + channel_len + 1, payload);
945 : : else
946 : 15 : n->data[channel_len + 1] = '\0';
947 : :
948 [ + + + + ]: 1133 : if (pendingNotifies == NULL || my_level > pendingNotifies->nestingLevel)
949 : 71 : {
950 : : NotificationList *notifies;
951 : :
952 : : /*
953 : : * First notify event in current (sub)xact. Note that we allocate the
954 : : * NotificationList in TopTransactionContext; the nestingLevel might
955 : : * get changed later by AtSubCommit_Notify.
956 : : */
957 : : notifies = (NotificationList *)
958 : 71 : MemoryContextAlloc(TopTransactionContext,
959 : : sizeof(NotificationList));
960 : 71 : notifies->nestingLevel = my_level;
961 : 71 : notifies->events = list_make1(n);
962 : : /* We certainly don't need a hashtable yet */
963 : 71 : notifies->hashtab = NULL;
964 : : /* We won't build uniqueChannelNames/Hash till later, either */
965 : 71 : notifies->uniqueChannelNames = NIL;
966 : 71 : notifies->uniqueChannelHash = NULL;
967 : 71 : notifies->upper = pendingNotifies;
968 : 71 : pendingNotifies = notifies;
969 : : }
970 : : else
971 : : {
972 : : /* Now check for duplicates */
973 [ + + ]: 1062 : if (AsyncExistsPendingNotify(n))
974 : : {
975 : : /* It's a dup, so forget it */
976 : 13 : pfree(n);
977 : 13 : MemoryContextSwitchTo(oldcontext);
978 : 13 : return;
979 : : }
980 : :
981 : : /* Append more events to existing list */
982 : 1049 : AddEventToPendingNotifies(n);
983 : : }
984 : :
985 : 1120 : MemoryContextSwitchTo(oldcontext);
986 : : }
987 : :
988 : : /*
989 : : * queue_listen
990 : : * Common code for listen, unlisten, unlisten all commands.
991 : : *
992 : : * Adds the request to the list of pending actions.
993 : : * Actual update of localChannelTable and globalChannelTable happens during
994 : : * PreCommit_Notify, with staged changes committed in AtCommit_Notify.
995 : : */
996 : : static void
997 : 98 : queue_listen(ListenActionKind action, const char *channel)
998 : : {
999 : : MemoryContext oldcontext;
1000 : : ListenAction *actrec;
1001 : 98 : int my_level = GetCurrentTransactionNestLevel();
1002 : :
1003 : : /*
1004 : : * Unlike Async_Notify, we don't try to collapse out duplicates here. We
1005 : : * keep the ordered list to preserve interactions like UNLISTEN ALL; the
1006 : : * final per-channel intent is computed during PreCommit_Notify.
1007 : : */
1008 : 98 : oldcontext = MemoryContextSwitchTo(CurTransactionContext);
1009 : :
1010 : : /* space for terminating null is included in sizeof(ListenAction) */
1011 : 98 : actrec = (ListenAction *) palloc(offsetof(ListenAction, channel) +
1012 : 98 : strlen(channel) + 1);
1013 : 98 : actrec->action = action;
1014 : 98 : strcpy(actrec->channel, channel);
1015 : :
1016 [ + + + + ]: 98 : if (pendingActions == NULL || my_level > pendingActions->nestingLevel)
1017 : 85 : {
1018 : : ActionList *actions;
1019 : :
1020 : : /*
1021 : : * First action in current sub(xact). Note that we allocate the
1022 : : * ActionList in TopTransactionContext; the nestingLevel might get
1023 : : * changed later by AtSubCommit_Notify.
1024 : : */
1025 : : actions = (ActionList *)
1026 : 85 : MemoryContextAlloc(TopTransactionContext, sizeof(ActionList));
1027 : 85 : actions->nestingLevel = my_level;
1028 : 85 : actions->actions = list_make1(actrec);
1029 : 85 : actions->upper = pendingActions;
1030 : 85 : pendingActions = actions;
1031 : : }
1032 : : else
1033 : 13 : pendingActions->actions = lappend(pendingActions->actions, actrec);
1034 : :
1035 : 98 : MemoryContextSwitchTo(oldcontext);
1036 : 98 : }
1037 : :
1038 : : /*
1039 : : * Async_Listen
1040 : : *
1041 : : * This is executed by the SQL listen command.
1042 : : */
1043 : : void
1044 : 52 : Async_Listen(const char *channel)
1045 : : {
1046 [ - + ]: 52 : if (Trace_notify)
1047 [ # # ]: 0 : elog(DEBUG1, "Async_Listen(%s,%d)", channel, MyProcPid);
1048 : :
1049 : 52 : queue_listen(LISTEN_LISTEN, channel);
1050 : 52 : }
1051 : :
1052 : : /*
1053 : : * Async_Unlisten
1054 : : *
1055 : : * This is executed by the SQL unlisten command.
1056 : : */
1057 : : void
1058 : 4 : Async_Unlisten(const char *channel)
1059 : : {
1060 [ - + ]: 4 : if (Trace_notify)
1061 [ # # ]: 0 : elog(DEBUG1, "Async_Unlisten(%s,%d)", channel, MyProcPid);
1062 : :
1063 : : /* If we couldn't possibly be listening, no need to queue anything */
1064 [ + - - + ]: 4 : if (pendingActions == NULL && !unlistenExitRegistered)
1065 : 0 : return;
1066 : :
1067 : 4 : queue_listen(LISTEN_UNLISTEN, channel);
1068 : : }
1069 : :
1070 : : /*
1071 : : * Async_UnlistenAll
1072 : : *
1073 : : * This is invoked by UNLISTEN * command, and also at backend exit.
1074 : : */
1075 : : void
1076 : 62 : Async_UnlistenAll(void)
1077 : : {
1078 [ - + ]: 62 : if (Trace_notify)
1079 [ # # ]: 0 : elog(DEBUG1, "Async_UnlistenAll(%d)", MyProcPid);
1080 : :
1081 : : /* If we couldn't possibly be listening, no need to queue anything */
1082 [ + + + + ]: 62 : if (pendingActions == NULL && !unlistenExitRegistered)
1083 : 20 : return;
1084 : :
1085 : 42 : queue_listen(LISTEN_UNLISTEN_ALL, "");
1086 : : }
1087 : :
1088 : : /*
1089 : : * SQL function: return a set of the channel names this backend is actively
1090 : : * listening to.
1091 : : */
1092 : : Datum
1093 : 12 : pg_listening_channels(PG_FUNCTION_ARGS)
1094 : : {
1095 : 12 : ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
1096 : :
1097 : 12 : InitMaterializedSRF(fcinfo, MAT_SRF_USE_EXPECTED_DESC);
1098 : :
1099 : : /*
1100 : : * Complete the hash scan in one call, since the caller might not fetch
1101 : : * all the rows and would otherwise leave the scan registered until
1102 : : * transaction end.
1103 : : */
1104 [ + - ]: 12 : if (localChannelTable != NULL)
1105 : : {
1106 : : HASH_SEQ_STATUS status;
1107 : : ChannelName *channelEntry;
1108 : :
1109 : 12 : hash_seq_init(&status, localChannelTable);
1110 [ + + ]: 20 : while ((channelEntry = (ChannelName *) hash_seq_search(&status)) != NULL)
1111 : : {
1112 : : Datum values[1];
1113 : : bool nulls[1];
1114 : :
1115 : 8 : values[0] = CStringGetTextDatum(channelEntry->channel);
1116 : 8 : nulls[0] = false;
1117 : 8 : tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc,
1118 : : values, nulls);
1119 : : }
1120 : : }
1121 : :
1122 : 12 : return (Datum) 0;
1123 : : }
1124 : :
1125 : : /*
1126 : : * Async_UnlistenOnExit
1127 : : *
1128 : : * This is executed at backend exit if we have done any LISTENs in this
1129 : : * backend. It might not be necessary anymore, if the user UNLISTENed
1130 : : * everything, but we don't try to detect that case.
1131 : : */
1132 : : static void
1133 : 18 : Async_UnlistenOnExit(int code, Datum arg)
1134 : : {
1135 : 18 : CleanupListenersOnExit();
1136 : 18 : asyncQueueUnregister();
1137 : 18 : }
1138 : :
1139 : : /*
1140 : : * AtPrepare_Notify
1141 : : *
1142 : : * This is called at the prepare phase of a two-phase
1143 : : * transaction. Save the state for possible commit later.
1144 : : */
1145 : : void
1146 : 321 : AtPrepare_Notify(void)
1147 : : {
1148 : : /* It's not allowed to have any pending LISTEN/UNLISTEN/NOTIFY actions */
1149 [ + - - + ]: 321 : if (pendingActions || pendingNotifies)
1150 [ # # ]: 0 : ereport(ERROR,
1151 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1152 : : errmsg("cannot PREPARE a transaction that has executed LISTEN, UNLISTEN, or NOTIFY")));
1153 : 321 : }
1154 : :
1155 : : /*
1156 : : * PreCommit_Notify
1157 : : *
1158 : : * This is called at transaction commit, before actually committing to
1159 : : * clog.
1160 : : *
1161 : : * If there are pending LISTEN actions, make sure we are listed in the
1162 : : * shared-memory listener array. This must happen before commit to
1163 : : * ensure we don't miss any notifies from transactions that commit
1164 : : * just after ours.
1165 : : *
1166 : : * If there are outbound notify requests in the pendingNotifies list,
1167 : : * add them to the global queue. We do that before commit so that
1168 : : * we can still throw error if we run out of queue space.
1169 : : */
1170 : : void
1171 : 612805 : PreCommit_Notify(void)
1172 : : {
1173 : : ListCell *p;
1174 : :
1175 [ + + + + ]: 612805 : if (!pendingActions && !pendingNotifies)
1176 : 612654 : return; /* no relevant statements in this xact */
1177 : :
1178 [ - + ]: 151 : if (Trace_notify)
1179 [ # # ]: 0 : elog(DEBUG1, "PreCommit_Notify");
1180 : :
1181 : : /* Preflight for any pending listen/unlisten actions */
1182 : 151 : initGlobalChannelTable();
1183 : :
1184 [ + + ]: 151 : if (pendingActions != NULL)
1185 : : {
1186 : : /* Ensure we have a local channel table */
1187 : 82 : initLocalChannelTable();
1188 : : /* Create pendingListenActions hash table for this transaction */
1189 : 82 : initPendingListenActions();
1190 : :
1191 : : /* Stage all the actions this transaction wants to perform */
1192 [ + - + + : 177 : foreach(p, pendingActions->actions)
+ + ]
1193 : : {
1194 : 95 : ListenAction *actrec = (ListenAction *) lfirst(p);
1195 : :
1196 [ + + + - ]: 95 : switch (actrec->action)
1197 : : {
1198 : 50 : case LISTEN_LISTEN:
1199 : 50 : BecomeRegisteredListener();
1200 : 50 : PrepareTableEntriesForListen(actrec->channel);
1201 : 50 : break;
1202 : 4 : case LISTEN_UNLISTEN:
1203 : 4 : PrepareTableEntriesForUnlisten(actrec->channel);
1204 : 4 : break;
1205 : 41 : case LISTEN_UNLISTEN_ALL:
1206 : 41 : PrepareTableEntriesForUnlistenAll();
1207 : 41 : break;
1208 : : }
1209 : : }
1210 : : }
1211 : :
1212 : : /* Queue any pending notifies (must happen after the above) */
1213 [ + + ]: 151 : if (pendingNotifies)
1214 : : {
1215 : : ListCell *nextNotify;
1216 : 69 : bool firstIteration = true;
1217 : :
1218 : : /*
1219 : : * Build list of unique channel names being notified for use by
1220 : : * SignalBackends().
1221 : : *
1222 : : * If uniqueChannelHash is available, use it to efficiently get the
1223 : : * unique channels. Otherwise, fall back to the O(N^2) approach.
1224 : : */
1225 : 69 : pendingNotifies->uniqueChannelNames = NIL;
1226 [ + + ]: 69 : if (pendingNotifies->uniqueChannelHash != NULL)
1227 : : {
1228 : : HASH_SEQ_STATUS status;
1229 : : ChannelName *channelEntry;
1230 : :
1231 : 2 : hash_seq_init(&status, pendingNotifies->uniqueChannelHash);
1232 [ + + ]: 4 : while ((channelEntry = (ChannelName *) hash_seq_search(&status)) != NULL)
1233 : 2 : pendingNotifies->uniqueChannelNames =
1234 : 2 : lappend(pendingNotifies->uniqueChannelNames,
1235 : 2 : channelEntry->channel);
1236 : : }
1237 : : else
1238 : : {
1239 : : /* O(N^2) approach is better for small number of notifications */
1240 [ + - + + : 231 : foreach_ptr(Notification, n, pendingNotifies->events)
+ + ]
1241 : : {
1242 : 97 : char *channel = n->data;
1243 : 97 : bool found = false;
1244 : :
1245 : : /* Name present in list? */
1246 [ + + + + : 198 : foreach_ptr(char, oldchan, pendingNotifies->uniqueChannelNames)
+ + ]
1247 : : {
1248 [ + + ]: 31 : if (strcmp(oldchan, channel) == 0)
1249 : : {
1250 : 27 : found = true;
1251 : 27 : break;
1252 : : }
1253 : : }
1254 : : /* Add if not already in list */
1255 [ + + ]: 97 : if (!found)
1256 : 70 : pendingNotifies->uniqueChannelNames =
1257 : 70 : lappend(pendingNotifies->uniqueChannelNames,
1258 : : channel);
1259 : : }
1260 : : }
1261 : :
1262 : : /* Preallocate workspace that will be needed by SignalBackends() */
1263 [ + + ]: 69 : if (signalPids == NULL)
1264 : 23 : signalPids = MemoryContextAlloc(TopMemoryContext,
1265 : : MaxBackends * sizeof(int32));
1266 : :
1267 [ + + ]: 69 : if (signalProcnos == NULL)
1268 : 23 : signalProcnos = MemoryContextAlloc(TopMemoryContext,
1269 : : MaxBackends * sizeof(ProcNumber));
1270 : :
1271 : : /*
1272 : : * Make sure that we have an XID assigned to the current transaction.
1273 : : * GetCurrentTransactionId is cheap if we already have an XID, but not
1274 : : * so cheap if we don't, and we'd prefer not to do that work while
1275 : : * holding NotifyQueueLock.
1276 : : */
1277 : 69 : (void) GetCurrentTransactionId();
1278 : :
1279 : : /*
1280 : : * Serialize writers by acquiring a special lock that we hold till
1281 : : * after commit. This ensures that queue entries appear in commit
1282 : : * order, and in particular that there are never uncommitted queue
1283 : : * entries ahead of committed ones, so an uncommitted transaction
1284 : : * can't block delivery of deliverable notifications.
1285 : : *
1286 : : * We use a heavyweight lock so that it'll automatically be released
1287 : : * after either commit or abort. This also allows deadlocks to be
1288 : : * detected, though really a deadlock shouldn't be possible here.
1289 : : *
1290 : : * The lock is on "database 0", which is pretty ugly but it doesn't
1291 : : * seem worth inventing a special locktag category just for this.
1292 : : * (Historical note: before PG 9.0, a similar lock on "database 0" was
1293 : : * used by the flatfiles mechanism.)
1294 : : */
1295 : 69 : LockSharedObject(DatabaseRelationId, InvalidOid, 0,
1296 : : AccessExclusiveLock);
1297 : :
1298 : : /*
1299 : : * For the direct advancement optimization in SignalBackends(), we
1300 : : * need to ensure that no other backend can insert queue entries
1301 : : * between queueHeadBeforeWrite and queueHeadAfterWrite. The
1302 : : * heavyweight lock above provides this guarantee, since it serializes
1303 : : * all writers.
1304 : : *
1305 : : * Note: if the heavyweight lock were ever removed for scalability
1306 : : * reasons, we could achieve the same guarantee by holding
1307 : : * NotifyQueueLock in EXCLUSIVE mode across all our insertions, rather
1308 : : * than releasing and reacquiring it for each page as we do below.
1309 : : */
1310 : :
1311 : : /* Initialize values to a safe default in case list is empty */
1312 : 69 : SET_QUEUE_POS(queueHeadBeforeWrite, 0, 0);
1313 : 69 : SET_QUEUE_POS(queueHeadAfterWrite, 0, 0);
1314 : :
1315 : : /* Now push the notifications into the queue */
1316 : 69 : nextNotify = list_head(pendingNotifies->events);
1317 [ + + ]: 173 : while (nextNotify != NULL)
1318 : : {
1319 : : /*
1320 : : * Add the pending notifications to the queue. We acquire and
1321 : : * release NotifyQueueLock once per page, which might be overkill
1322 : : * but it does allow readers to get in while we're doing this.
1323 : : *
1324 : : * A full queue is very uncommon and should really not happen,
1325 : : * given that we have so much space available in the SLRU pages.
1326 : : * Nevertheless we need to deal with this possibility. Note that
1327 : : * when we get here we are in the process of committing our
1328 : : * transaction, but we have not yet committed to clog, so at this
1329 : : * point in time we can still roll the transaction back.
1330 : : */
1331 : 104 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
1332 [ + + ]: 104 : if (firstIteration)
1333 : : {
1334 : 69 : queueHeadBeforeWrite = QUEUE_HEAD;
1335 : 69 : firstIteration = false;
1336 : : }
1337 : 104 : asyncQueueFillWarning();
1338 [ - + ]: 104 : if (asyncQueueIsFull())
1339 [ # # ]: 0 : ereport(ERROR,
1340 : : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
1341 : : errmsg("too many notifications in the NOTIFY queue")));
1342 : 104 : nextNotify = asyncQueueAddEntries(nextNotify);
1343 : 104 : queueHeadAfterWrite = QUEUE_HEAD;
1344 : 104 : LWLockRelease(NotifyQueueLock);
1345 : : }
1346 : :
1347 : : /* Note that we don't clear pendingNotifies; AtCommit_Notify will. */
1348 : : }
1349 : : }
1350 : :
1351 : : /*
1352 : : * AtCommit_Notify
1353 : : *
1354 : : * This is called at transaction commit, after committing to clog.
1355 : : *
1356 : : * Apply pending listen/unlisten changes and clear transaction-local state.
1357 : : *
1358 : : * If we issued any notifications in the transaction, send signals to
1359 : : * listening backends (possibly including ourselves) to process them.
1360 : : * Also, if we filled enough queue pages with new notifies, try to
1361 : : * advance the queue tail pointer.
1362 : : */
1363 : : void
1364 : 612645 : AtCommit_Notify(void)
1365 : : {
1366 : : /*
1367 : : * Allow transactions that have not executed LISTEN/UNLISTEN/NOTIFY to
1368 : : * return as soon as possible
1369 : : */
1370 [ + + + + ]: 612645 : if (!pendingActions && !pendingNotifies)
1371 : 612494 : return;
1372 : :
1373 [ - + ]: 151 : if (Trace_notify)
1374 [ # # ]: 0 : elog(DEBUG1, "AtCommit_Notify");
1375 : :
1376 : : /* Apply staged listen/unlisten changes */
1377 : 151 : ApplyPendingListenActions(true);
1378 : :
1379 : : /* If no longer listening to anything, get out of listener array */
1380 [ + + + - : 151 : if (amRegisteredListener && LocalChannelTableIsEmpty())
+ + ]
1381 : 21 : asyncQueueUnregister();
1382 : :
1383 : : /*
1384 : : * Send signals to listening backends. We need do this only if there are
1385 : : * pending notifies, which were previously added to the shared queue by
1386 : : * PreCommit_Notify().
1387 : : */
1388 [ + + ]: 151 : if (pendingNotifies != NULL)
1389 : 69 : SignalBackends();
1390 : :
1391 : : /*
1392 : : * If it's time to try to advance the global tail pointer, do that.
1393 : : *
1394 : : * (It might seem odd to do this in the sender, when more than likely the
1395 : : * listeners won't yet have read the messages we just sent. However,
1396 : : * there's less contention if only the sender does it, and there is little
1397 : : * need for urgency in advancing the global tail. So this typically will
1398 : : * be clearing out messages that were sent some time ago.)
1399 : : */
1400 [ + + ]: 151 : if (tryAdvanceTail)
1401 : : {
1402 : 8 : tryAdvanceTail = false;
1403 : 8 : asyncQueueAdvanceTail();
1404 : : }
1405 : :
1406 : : /* And clean up */
1407 : 151 : ClearPendingActionsAndNotifies();
1408 : : }
1409 : :
1410 : : /*
1411 : : * BecomeRegisteredListener --- subroutine for PreCommit_Notify
1412 : : *
1413 : : * This function must make sure we are ready to catch any incoming messages.
1414 : : */
1415 : : static void
1416 : 50 : BecomeRegisteredListener(void)
1417 : : {
1418 : : QueuePosition head;
1419 : : QueuePosition max;
1420 : : ProcNumber prevListener;
1421 : :
1422 : : /*
1423 : : * Nothing to do if we are already listening to something, nor if we
1424 : : * already ran this routine in this transaction.
1425 : : */
1426 [ + + ]: 50 : if (amRegisteredListener)
1427 : 23 : return;
1428 : :
1429 [ - + ]: 27 : if (Trace_notify)
1430 [ # # ]: 0 : elog(DEBUG1, "BecomeRegisteredListener(%d)", MyProcPid);
1431 : :
1432 : : /*
1433 : : * Before registering, make sure we will unlisten before dying. (Note:
1434 : : * this action does not get undone if we abort later.)
1435 : : */
1436 [ + + ]: 27 : if (!unlistenExitRegistered)
1437 : : {
1438 : 18 : before_shmem_exit(Async_UnlistenOnExit, 0);
1439 : 18 : unlistenExitRegistered = true;
1440 : : }
1441 : :
1442 : : /*
1443 : : * This is our first LISTEN, so establish our pointer.
1444 : : *
1445 : : * We set our pointer to the global tail pointer and then move it forward
1446 : : * over already-committed notifications. This ensures we cannot miss any
1447 : : * not-yet-committed notifications. We might get a few more but that
1448 : : * doesn't hurt.
1449 : : *
1450 : : * In some scenarios there might be a lot of committed notifications that
1451 : : * have not yet been pruned away (because some backend is being lazy about
1452 : : * reading them). To reduce our startup time, we can look at other
1453 : : * backends and adopt the maximum "pos" pointer of any backend that's in
1454 : : * our database; any notifications it's already advanced over are surely
1455 : : * committed and need not be re-examined by us. (We must consider only
1456 : : * backends connected to our DB, because others will not have bothered to
1457 : : * check committed-ness of notifications in our DB.)
1458 : : *
1459 : : * We need exclusive lock here so we can look at other backends' entries
1460 : : * and manipulate the list links.
1461 : : */
1462 : 27 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
1463 : 27 : head = QUEUE_HEAD;
1464 : 27 : max = QUEUE_TAIL;
1465 : 27 : prevListener = INVALID_PROC_NUMBER;
1466 [ + + ]: 29 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
1467 : : {
1468 [ + - ]: 2 : if (QUEUE_BACKEND_DBOID(i) == MyDatabaseId)
1469 [ - + + - : 2 : max = QUEUE_POS_MAX(max, QUEUE_BACKEND_POS(i));
+ - ]
1470 : : /* Also find last listening backend before this one */
1471 [ + + ]: 2 : if (i < MyProcNumber)
1472 : 1 : prevListener = i;
1473 : : }
1474 : 27 : QUEUE_BACKEND_POS(MyProcNumber) = max;
1475 : 27 : QUEUE_BACKEND_PID(MyProcNumber) = MyProcPid;
1476 : 27 : QUEUE_BACKEND_DBOID(MyProcNumber) = MyDatabaseId;
1477 : 27 : QUEUE_BACKEND_WAKEUP_PENDING(MyProcNumber) = false;
1478 : 27 : QUEUE_BACKEND_IS_ADVANCING(MyProcNumber) = false;
1479 : : /* Insert backend into list of listeners at correct position */
1480 [ + + ]: 27 : if (prevListener != INVALID_PROC_NUMBER)
1481 : : {
1482 : 1 : QUEUE_NEXT_LISTENER(MyProcNumber) = QUEUE_NEXT_LISTENER(prevListener);
1483 : 1 : QUEUE_NEXT_LISTENER(prevListener) = MyProcNumber;
1484 : : }
1485 : : else
1486 : : {
1487 : 26 : QUEUE_NEXT_LISTENER(MyProcNumber) = QUEUE_FIRST_LISTENER;
1488 : 26 : QUEUE_FIRST_LISTENER = MyProcNumber;
1489 : : }
1490 : 27 : LWLockRelease(NotifyQueueLock);
1491 : :
1492 : : /* Now we are listed in the global array, so remember we're listening */
1493 : 27 : amRegisteredListener = true;
1494 : :
1495 : : /*
1496 : : * Try to move our pointer forward as far as possible. This will skip
1497 : : * over already-committed notifications, which we want to do because they
1498 : : * might be quite stale. Note that we are not yet listening on anything,
1499 : : * so we won't deliver such notifications to our frontend. Also, although
1500 : : * our transaction might have executed NOTIFY, those message(s) aren't
1501 : : * queued yet so we won't skip them here.
1502 : : */
1503 [ + + + + ]: 27 : if (!QUEUE_POS_EQUAL(max, head))
1504 : 18 : asyncQueueReadAllNotifications();
1505 : : }
1506 : :
1507 : : /*
1508 : : * PrepareTableEntriesForListen --- subroutine for PreCommit_Notify
1509 : : *
1510 : : * Prepare a LISTEN by recording it in pendingListenActions, pre-allocating
1511 : : * an entry in localChannelTable, and pre-allocating an entry in the shared
1512 : : * globalChannelTable with removeOnAbort set. AtCommit_Notify will clear
1513 : : * removeOnAbort; abort processing will remove entries still marked so.
1514 : : */
1515 : : static void
1516 : 50 : PrepareTableEntriesForListen(const char *channel)
1517 : : {
1518 : : GlobalChannelKey key;
1519 : : GlobalChannelEntry *entry;
1520 : : bool found;
1521 : : ListenerEntry *listeners;
1522 : : PendingListenEntry *pending;
1523 : :
1524 : : /*
1525 : : * Record in local pending hash that we want to LISTEN, overwriting any
1526 : : * earlier attempt to UNLISTEN.
1527 : : */
1528 : : pending = (PendingListenEntry *)
1529 : 50 : hash_search(pendingListenActions, channel, HASH_ENTER, NULL);
1530 : 50 : pending->action = PENDING_LISTEN;
1531 : :
1532 : : /*
1533 : : * Ensure that there is an entry for the channel in localChannelTable.
1534 : : * (Should this fail, we can just roll back.) If the transaction fails
1535 : : * after this point, we will remove the entry if appropriate during
1536 : : * ApplyPendingListenActions. Note that this entry allows IsListeningOn()
1537 : : * to return TRUE; we assume nothing is going to consult that before
1538 : : * AtCommit_Notify/AtAbort_Notify. However, if later actions attempt to
1539 : : * UNLISTEN this channel or UNLISTEN *, we need to have the local entry
1540 : : * present to ensure they do the right things; see
1541 : : * PrepareTableEntriesForUnlisten and PrepareTableEntriesForUnlistenAll.
1542 : : */
1543 : 50 : (void) hash_search(localChannelTable, channel, HASH_ENTER, NULL);
1544 : :
1545 : : /* Pre-allocate entry in shared globalChannelTable */
1546 : 50 : GlobalChannelKeyInit(&key, MyDatabaseId, channel);
1547 : 50 : entry = dshash_find_or_insert(globalChannelTable, &key, &found);
1548 : :
1549 [ + + ]: 50 : if (!found)
1550 : : {
1551 : : /* New channel entry, so initialize it to a safe state */
1552 : 37 : entry->listenersArray = InvalidDsaPointer;
1553 : 37 : entry->numListeners = 0;
1554 : 37 : entry->allocatedListeners = 0;
1555 : : }
1556 : :
1557 : : /*
1558 : : * Create listenersArray if entry doesn't have one. It's tempting to fold
1559 : : * this into the !found case, but this coding allows us to cope in case
1560 : : * dsa_allocate() failed in an earlier attempt.
1561 : : */
1562 [ + + ]: 50 : if (!DsaPointerIsValid(entry->listenersArray))
1563 : : {
1564 : 37 : entry->listenersArray = dsa_allocate(globalChannelDSA,
1565 : : sizeof(ListenerEntry) * INITIAL_LISTENERS_ARRAY_SIZE);
1566 : 37 : entry->allocatedListeners = INITIAL_LISTENERS_ARRAY_SIZE;
1567 : : }
1568 : :
1569 : : listeners = (ListenerEntry *)
1570 : 50 : dsa_get_address(globalChannelDSA, entry->listenersArray);
1571 : :
1572 : : /*
1573 : : * Check if we already have a ListenerEntry (possibly from earlier in this
1574 : : * transaction)
1575 : : */
1576 [ + + ]: 53 : for (int i = 0; i < entry->numListeners; i++)
1577 : : {
1578 [ + + ]: 13 : if (listeners[i].procNo == MyProcNumber)
1579 : : {
1580 : : /* Already have an entry; leave removeOnAbort as-is */
1581 : 10 : dshash_release_lock(globalChannelTable, entry);
1582 : 10 : return;
1583 : : }
1584 : : }
1585 : :
1586 : : /* Need to add a new entry; grow array if necessary */
1587 [ - + ]: 40 : if (entry->numListeners >= entry->allocatedListeners)
1588 : : {
1589 : 0 : int new_size = entry->allocatedListeners * 2;
1590 : 0 : dsa_pointer old_array = entry->listenersArray;
1591 : 0 : dsa_pointer new_array = dsa_allocate(globalChannelDSA,
1592 : : sizeof(ListenerEntry) * new_size);
1593 : 0 : ListenerEntry *new_listeners = (ListenerEntry *) dsa_get_address(globalChannelDSA, new_array);
1594 : :
1595 : 0 : memcpy(new_listeners, listeners, sizeof(ListenerEntry) * entry->numListeners);
1596 : 0 : entry->listenersArray = new_array;
1597 : 0 : entry->allocatedListeners = new_size;
1598 : 0 : dsa_free(globalChannelDSA, old_array);
1599 : 0 : listeners = new_listeners;
1600 : : }
1601 : :
1602 : 40 : listeners[entry->numListeners].procNo = MyProcNumber;
1603 : 40 : listeners[entry->numListeners].removeOnAbort = true;
1604 : 40 : entry->numListeners++;
1605 : :
1606 : 40 : dshash_release_lock(globalChannelTable, entry);
1607 : : }
1608 : :
1609 : : /*
1610 : : * PrepareTableEntriesForUnlisten --- subroutine for PreCommit_Notify
1611 : : *
1612 : : * Prepare an UNLISTEN by recording it in pendingListenActions, but only if
1613 : : * we're currently listening (committed or staged). We don't touch
1614 : : * globalChannelTable yet - the listener keeps receiving signals until
1615 : : * commit, when the entry is removed.
1616 : : */
1617 : : static void
1618 : 4 : PrepareTableEntriesForUnlisten(const char *channel)
1619 : : {
1620 : : PendingListenEntry *pending;
1621 : :
1622 : : /*
1623 : : * If the channel name is not in localChannelTable, then we are neither
1624 : : * listening on it nor preparing to listen on it, so we don't need to
1625 : : * record an UNLISTEN action.
1626 : : */
1627 : : Assert(localChannelTable != NULL);
1628 [ - + ]: 4 : if (hash_search(localChannelTable, channel, HASH_FIND, NULL) == NULL)
1629 : 0 : return;
1630 : :
1631 : : /*
1632 : : * Record in local pending hash that we want to UNLISTEN, overwriting any
1633 : : * earlier attempt to LISTEN. Don't touch localChannelTable or
1634 : : * globalChannelTable yet - we keep receiving signals until commit.
1635 : : */
1636 : : pending = (PendingListenEntry *)
1637 : 4 : hash_search(pendingListenActions, channel, HASH_ENTER, NULL);
1638 : 4 : pending->action = PENDING_UNLISTEN;
1639 : : }
1640 : :
1641 : : /*
1642 : : * PrepareTableEntriesForUnlistenAll --- subroutine for PreCommit_Notify
1643 : : *
1644 : : * Prepare UNLISTEN * by recording an UNLISTEN for all listened or
1645 : : * about-to-be-listened channels in pendingListenActions.
1646 : : */
1647 : : static void
1648 : 41 : PrepareTableEntriesForUnlistenAll(void)
1649 : : {
1650 : : HASH_SEQ_STATUS seq;
1651 : : ChannelName *channelEntry;
1652 : : PendingListenEntry *pending;
1653 : :
1654 : : /*
1655 : : * Scan localChannelTable, which will have the names of all channels that
1656 : : * we are listening on or have prepared to listen on. Record an UNLISTEN
1657 : : * action for each one, overwriting any earlier attempt to LISTEN.
1658 : : */
1659 : 41 : hash_seq_init(&seq, localChannelTable);
1660 [ + + ]: 69 : while ((channelEntry = (ChannelName *) hash_seq_search(&seq)) != NULL)
1661 : : {
1662 : : pending = (PendingListenEntry *)
1663 : 28 : hash_search(pendingListenActions, channelEntry->channel, HASH_ENTER, NULL);
1664 : 28 : pending->action = PENDING_UNLISTEN;
1665 : : }
1666 : 41 : }
1667 : :
1668 : : /*
1669 : : * RemoveListenerFromChannel --- remove idx'th listener from global channel entry
1670 : : *
1671 : : * Decrements numListeners, compacts the array, and frees the entry if empty.
1672 : : * Sets *entry_ptr to NULL if the entry was deleted.
1673 : : *
1674 : : * We could get the listeners pointer from the entry, but all callers
1675 : : * already have it at hand.
1676 : : */
1677 : : static void
1678 : 32 : RemoveListenerFromChannel(GlobalChannelEntry **entry_ptr,
1679 : : ListenerEntry *listeners,
1680 : : int idx)
1681 : : {
1682 : 32 : GlobalChannelEntry *entry = *entry_ptr;
1683 : :
1684 : 32 : entry->numListeners--;
1685 [ + + ]: 32 : if (idx < entry->numListeners)
1686 : 3 : memmove(&listeners[idx], &listeners[idx + 1],
1687 : 3 : sizeof(ListenerEntry) * (entry->numListeners - idx));
1688 : :
1689 [ + + ]: 32 : if (entry->numListeners == 0)
1690 : : {
1691 : 29 : dsa_free(globalChannelDSA, entry->listenersArray);
1692 : 29 : dshash_delete_entry(globalChannelTable, entry);
1693 : : /* tells caller not to release the entry's lock: */
1694 : 29 : *entry_ptr = NULL;
1695 : : }
1696 : 32 : }
1697 : :
1698 : : /*
1699 : : * ApplyPendingListenActions
1700 : : *
1701 : : * Apply, or revert, staged listen/unlisten changes to the local and global
1702 : : * hash tables.
1703 : : */
1704 : : static void
1705 : 36228 : ApplyPendingListenActions(bool isCommit)
1706 : : {
1707 : : HASH_SEQ_STATUS seq;
1708 : : PendingListenEntry *pending;
1709 : :
1710 : : /* Quick exit if nothing to do */
1711 [ + + ]: 36228 : if (pendingListenActions == NULL)
1712 : 36146 : return;
1713 : :
1714 : : /* We made a globalChannelTable before building pendingListenActions */
1715 [ - + ]: 82 : if (globalChannelTable == NULL)
1716 [ # # ]: 0 : elog(PANIC, "global channel table missing post-commit/abort");
1717 : :
1718 : : /* For each staged action ... */
1719 : 82 : hash_seq_init(&seq, pendingListenActions);
1720 [ + + ]: 163 : while ((pending = (PendingListenEntry *) hash_seq_search(&seq)) != NULL)
1721 : : {
1722 : : GlobalChannelKey key;
1723 : : GlobalChannelEntry *entry;
1724 : 81 : bool removeLocal = true;
1725 : 81 : bool foundListener = false;
1726 : :
1727 : : /*
1728 : : * Find the global entry for this channel. If isCommit, it had better
1729 : : * exist (it was created in PreCommit). In an abort, it might not
1730 : : * exist, in which case we are not listening and should discard any
1731 : : * local entry that PreCommit may have managed to create.
1732 : : */
1733 : 81 : GlobalChannelKeyInit(&key, MyDatabaseId, pending->channel);
1734 : 81 : entry = dshash_find(globalChannelTable, &key, true);
1735 [ + - ]: 81 : if (entry != NULL)
1736 : : {
1737 : : /* Scan entry to find the ListenerEntry for this backend */
1738 : : ListenerEntry *listeners;
1739 : :
1740 : : listeners = (ListenerEntry *)
1741 : 81 : dsa_get_address(globalChannelDSA, entry->listenersArray);
1742 : :
1743 [ + - ]: 84 : for (int i = 0; i < entry->numListeners; i++)
1744 : : {
1745 [ + + ]: 84 : if (listeners[i].procNo != MyProcNumber)
1746 : 3 : continue;
1747 : 81 : foundListener = true;
1748 [ + - ]: 81 : if (isCommit)
1749 : : {
1750 [ + + ]: 81 : if (pending->action == PENDING_LISTEN)
1751 : : {
1752 : : /*
1753 : : * LISTEN being committed: entry is now permanent.
1754 : : * localChannelTable entry was created during
1755 : : * PreCommit and should be kept.
1756 : : */
1757 : 49 : listeners[i].removeOnAbort = false;
1758 : 49 : removeLocal = false;
1759 : : }
1760 : : else
1761 : : {
1762 : : /*
1763 : : * UNLISTEN being committed: remove pre-allocated
1764 : : * entries from both tables.
1765 : : */
1766 : 32 : RemoveListenerFromChannel(&entry, listeners, i);
1767 : : }
1768 : : }
1769 : : else
1770 : : {
1771 : : /*
1772 : : * Note: this part is reachable only if the transaction
1773 : : * aborts after PreCommit_Notify() has made some
1774 : : * pendingListenActions entries, so it's pretty hard to
1775 : : * test.
1776 : : */
1777 [ # # ]: 0 : if (listeners[i].removeOnAbort)
1778 : : {
1779 : : /*
1780 : : * Staged LISTEN (or LISTEN+UNLISTEN) being aborted,
1781 : : * so remove pre-allocated entries from both tables.
1782 : : */
1783 : 0 : RemoveListenerFromChannel(&entry, listeners, i);
1784 : : }
1785 : : else
1786 : : {
1787 : : /*
1788 : : * Entry predates this transaction, so keep the
1789 : : * localChannelTable entry.
1790 : : */
1791 : 0 : removeLocal = false;
1792 : : }
1793 : : }
1794 : 81 : break; /* there shouldn't be another match */
1795 : : }
1796 : :
1797 : : /* We might have already released the entry by removing it */
1798 [ + + ]: 81 : if (entry != NULL)
1799 : 52 : dshash_release_lock(globalChannelTable, entry);
1800 : : }
1801 : :
1802 : : /*
1803 : : * If we're committing a LISTEN action, we should have found a
1804 : : * matching ListenerEntry, but otherwise it's okay if we didn't.
1805 : : */
1806 [ + - + + : 81 : if (isCommit && pending->action == PENDING_LISTEN && !foundListener)
- + ]
1807 [ # # ]: 0 : elog(PANIC, "could not find listener entry for channel \"%s\" backend %d",
1808 : : pending->channel, MyProcNumber);
1809 : :
1810 : : /*
1811 : : * If we did not find a globalChannelTable entry for our backend, or
1812 : : * if we are unlistening, remove any localChannelTable entry that may
1813 : : * exist. (Note in particular that this cleans up if we created a
1814 : : * localChannelTable entry and then failed while trying to create a
1815 : : * globalChannelTable entry.)
1816 : : */
1817 [ + + + - ]: 81 : if (removeLocal && localChannelTable != NULL)
1818 : 32 : (void) hash_search(localChannelTable, pending->channel,
1819 : : HASH_REMOVE, NULL);
1820 : : }
1821 : : }
1822 : :
1823 : : /*
1824 : : * CleanupListenersOnExit --- called from Async_UnlistenOnExit
1825 : : *
1826 : : * Remove this backend from all channels in the shared global table.
1827 : : */
1828 : : static void
1829 : 18 : CleanupListenersOnExit(void)
1830 : : {
1831 : : dshash_seq_status status;
1832 : : GlobalChannelEntry *entry;
1833 : :
1834 [ - + ]: 18 : if (Trace_notify)
1835 [ # # ]: 0 : elog(DEBUG1, "CleanupListenersOnExit(%d)", MyProcPid);
1836 : :
1837 : : /* Clear our local cache (not really necessary, but be consistent) */
1838 [ + - ]: 18 : if (localChannelTable != NULL)
1839 : : {
1840 : 18 : hash_destroy(localChannelTable);
1841 : 18 : localChannelTable = NULL;
1842 : : }
1843 : :
1844 : : /* Now remove our entries from the shared globalChannelTable */
1845 [ - + ]: 18 : if (globalChannelTable == NULL)
1846 : 0 : return;
1847 : :
1848 : 18 : dshash_seq_init(&status, globalChannelTable, true);
1849 [ + + ]: 26 : while ((entry = dshash_seq_next(&status)) != NULL)
1850 : : {
1851 : : ListenerEntry *listeners;
1852 : :
1853 [ - + ]: 8 : if (entry->key.dboid != MyDatabaseId)
1854 : 0 : continue; /* not relevant */
1855 : :
1856 : : listeners = (ListenerEntry *)
1857 : 8 : dsa_get_address(globalChannelDSA, entry->listenersArray);
1858 : :
1859 [ + - ]: 8 : for (int i = 0; i < entry->numListeners; i++)
1860 : : {
1861 [ + - ]: 8 : if (listeners[i].procNo == MyProcNumber)
1862 : : {
1863 : 8 : entry->numListeners--;
1864 [ - + ]: 8 : if (i < entry->numListeners)
1865 : 0 : memmove(&listeners[i], &listeners[i + 1],
1866 : 0 : sizeof(ListenerEntry) * (entry->numListeners - i));
1867 : :
1868 [ + - ]: 8 : if (entry->numListeners == 0)
1869 : : {
1870 : 8 : dsa_free(globalChannelDSA, entry->listenersArray);
1871 : 8 : dshash_delete_current(&status);
1872 : : }
1873 : 8 : break;
1874 : : }
1875 : : }
1876 : : }
1877 : 18 : dshash_seq_term(&status);
1878 : : }
1879 : :
1880 : : /*
1881 : : * Test whether we are actively listening on the given channel name.
1882 : : *
1883 : : * Note: this function is executed for every notification found in the queue.
1884 : : */
1885 : : static bool
1886 : 54 : IsListeningOn(const char *channel)
1887 : : {
1888 [ - + ]: 54 : if (localChannelTable == NULL)
1889 : 0 : return false;
1890 : :
1891 : 54 : return (hash_search(localChannelTable, channel, HASH_FIND, NULL) != NULL);
1892 : : }
1893 : :
1894 : : /*
1895 : : * Remove our entry from the listeners array when we are no longer listening
1896 : : * on any channel. NB: must not fail if we're already not listening.
1897 : : */
1898 : : static void
1899 : 39 : asyncQueueUnregister(void)
1900 : : {
1901 : : Assert(LocalChannelTableIsEmpty()); /* else caller error */
1902 : :
1903 [ + + ]: 39 : if (!amRegisteredListener) /* nothing to do */
1904 : 12 : return;
1905 : :
1906 : : /*
1907 : : * Need exclusive lock here to manipulate list links.
1908 : : */
1909 : 27 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
1910 : : /* Mark our entry as invalid */
1911 : 27 : QUEUE_BACKEND_PID(MyProcNumber) = InvalidPid;
1912 : 27 : QUEUE_BACKEND_DBOID(MyProcNumber) = InvalidOid;
1913 : 27 : QUEUE_BACKEND_WAKEUP_PENDING(MyProcNumber) = false;
1914 : 27 : QUEUE_BACKEND_IS_ADVANCING(MyProcNumber) = false;
1915 : : /* and remove it from the list */
1916 [ + + ]: 27 : if (QUEUE_FIRST_LISTENER == MyProcNumber)
1917 : 26 : QUEUE_FIRST_LISTENER = QUEUE_NEXT_LISTENER(MyProcNumber);
1918 : : else
1919 : : {
1920 [ + - ]: 1 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
1921 : : {
1922 [ + - ]: 1 : if (QUEUE_NEXT_LISTENER(i) == MyProcNumber)
1923 : : {
1924 : 1 : QUEUE_NEXT_LISTENER(i) = QUEUE_NEXT_LISTENER(MyProcNumber);
1925 : 1 : break;
1926 : : }
1927 : : }
1928 : : }
1929 : 27 : QUEUE_NEXT_LISTENER(MyProcNumber) = INVALID_PROC_NUMBER;
1930 : 27 : LWLockRelease(NotifyQueueLock);
1931 : :
1932 : : /* mark ourselves as no longer listed in the global array */
1933 : 27 : amRegisteredListener = false;
1934 : : }
1935 : :
1936 : : /*
1937 : : * Test whether there is room to insert more notification messages.
1938 : : *
1939 : : * Caller must hold at least shared NotifyQueueLock.
1940 : : */
1941 : : static bool
1942 : 104 : asyncQueueIsFull(void)
1943 : : {
1944 : 104 : int64 headPage = QUEUE_POS_PAGE(QUEUE_HEAD);
1945 : 104 : int64 tailPage = QUEUE_POS_PAGE(QUEUE_TAIL);
1946 : 104 : int64 occupied = headPage - tailPage;
1947 : :
1948 : 104 : return occupied >= max_notify_queue_pages;
1949 : : }
1950 : :
1951 : : /*
1952 : : * Advance the QueuePosition to the next entry, assuming that the current
1953 : : * entry is of length entryLength. If we jump to a new page the function
1954 : : * returns true, else false.
1955 : : */
1956 : : static bool
1957 : 2544 : asyncQueueAdvance(QueuePosition *position, int entryLength)
1958 : : {
1959 : 2544 : int64 pageno = QUEUE_POS_PAGE(*position);
1960 : 2544 : int offset = QUEUE_POS_OFFSET(*position);
1961 : 2544 : bool pageJump = false;
1962 : :
1963 : : /*
1964 : : * Move to the next writing position: First jump over what we have just
1965 : : * written or read.
1966 : : */
1967 : 2544 : offset += entryLength;
1968 : : Assert(offset <= QUEUE_PAGESIZE);
1969 : :
1970 : : /*
1971 : : * In a second step check if another entry can possibly be written to the
1972 : : * page. If so, stay here, we have reached the next position. If not, then
1973 : : * we need to move on to the next page.
1974 : : */
1975 [ + + ]: 2544 : if (offset + QUEUEALIGN(AsyncQueueEntryEmptySize) > QUEUE_PAGESIZE)
1976 : : {
1977 : 38 : pageno++;
1978 : 38 : offset = 0;
1979 : 38 : pageJump = true;
1980 : : }
1981 : :
1982 : 2544 : SET_QUEUE_POS(*position, pageno, offset);
1983 : 2544 : return pageJump;
1984 : : }
1985 : :
1986 : : /*
1987 : : * Fill the AsyncQueueEntry at *qe with an outbound notification message.
1988 : : */
1989 : : static void
1990 : 1146 : asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe)
1991 : : {
1992 : 1146 : size_t channellen = n->channel_len;
1993 : 1146 : size_t payloadlen = n->payload_len;
1994 : : int entryLength;
1995 : :
1996 : : Assert(channellen < NAMEDATALEN);
1997 : : Assert(payloadlen < NOTIFY_PAYLOAD_MAX_LENGTH);
1998 : :
1999 : : /* The terminators are already included in AsyncQueueEntryEmptySize */
2000 : 1146 : entryLength = AsyncQueueEntryEmptySize + payloadlen + channellen;
2001 : 1146 : entryLength = QUEUEALIGN(entryLength);
2002 : 1146 : qe->length = entryLength;
2003 : 1146 : qe->dboid = MyDatabaseId;
2004 : 1146 : qe->xid = GetCurrentTransactionId();
2005 : 1146 : qe->srcPid = MyProcPid;
2006 : 1146 : memcpy(qe->data, n->data, channellen + payloadlen + 2);
2007 : 1146 : }
2008 : :
2009 : : /*
2010 : : * Add pending notifications to the queue.
2011 : : *
2012 : : * We go page by page here, i.e. we stop once we have to go to a new page but
2013 : : * we will be called again and then fill that next page. If an entry does not
2014 : : * fit into the current page, we write a dummy entry with an InvalidOid as the
2015 : : * database OID in order to fill the page. So every page is always used up to
2016 : : * the last byte which simplifies reading the page later.
2017 : : *
2018 : : * We are passed the list cell (in pendingNotifies->events) containing the next
2019 : : * notification to write and return the first still-unwritten cell back.
2020 : : * Eventually we will return NULL indicating all is done.
2021 : : *
2022 : : * We are holding NotifyQueueLock already from the caller and grab
2023 : : * page specific SLRU bank lock locally in this function.
2024 : : */
2025 : : static ListCell *
2026 : 104 : asyncQueueAddEntries(ListCell *nextNotify)
2027 : : {
2028 : : AsyncQueueEntry qe;
2029 : : QueuePosition queue_head;
2030 : : int64 pageno;
2031 : : int offset;
2032 : : int slotno;
2033 : : LWLock *prevlock;
2034 : :
2035 : : /*
2036 : : * We work with a local copy of QUEUE_HEAD, which we write back to shared
2037 : : * memory upon exiting. The reason for this is that if we have to advance
2038 : : * to a new page, SimpleLruZeroPage might fail (out of disk space, for
2039 : : * instance), and we must not advance QUEUE_HEAD if it does. (Otherwise,
2040 : : * subsequent insertions would try to put entries into a page that slru.c
2041 : : * thinks doesn't exist yet.) So, use a local position variable. Note
2042 : : * that if we do fail, any already-inserted queue entries are forgotten;
2043 : : * this is okay, since they'd be useless anyway after our transaction
2044 : : * rolls back.
2045 : : */
2046 : 104 : queue_head = QUEUE_HEAD;
2047 : :
2048 : : /*
2049 : : * If this is the first write since the postmaster started, we need to
2050 : : * initialize the first page of the async SLRU. Otherwise, the current
2051 : : * page should be initialized already, so just fetch it.
2052 : : */
2053 : 104 : pageno = QUEUE_POS_PAGE(queue_head);
2054 : 104 : prevlock = SimpleLruGetBankLock(NotifyCtl, pageno);
2055 : :
2056 : : /* We hold both NotifyQueueLock and SLRU bank lock during this operation */
2057 : 104 : LWLockAcquire(prevlock, LW_EXCLUSIVE);
2058 : :
2059 [ + + + + ]: 104 : if (QUEUE_POS_IS_ZERO(queue_head))
2060 : 9 : slotno = SimpleLruZeroPage(NotifyCtl, pageno);
2061 : : else
2062 : 95 : slotno = SimpleLruReadPage(NotifyCtl, pageno, true, &queue_head);
2063 : :
2064 : : /* Note we mark the page dirty before writing in it */
2065 : 104 : NotifyCtl->shared->page_dirty[slotno] = true;
2066 : :
2067 [ + + ]: 1215 : while (nextNotify != NULL)
2068 : : {
2069 : 1146 : Notification *n = (Notification *) lfirst(nextNotify);
2070 : :
2071 : : /* Construct a valid queue entry in local variable qe */
2072 : 1146 : asyncQueueNotificationToEntry(n, &qe);
2073 : :
2074 : 1146 : offset = QUEUE_POS_OFFSET(queue_head);
2075 : :
2076 : : /* Check whether the entry really fits on the current page */
2077 [ + + ]: 1146 : if (offset + qe.length <= QUEUE_PAGESIZE)
2078 : : {
2079 : : /* OK, so advance nextNotify past this item */
2080 : 1114 : nextNotify = lnext(pendingNotifies->events, nextNotify);
2081 : : }
2082 : : else
2083 : : {
2084 : : /*
2085 : : * Write a dummy entry to fill up the page. Actually readers will
2086 : : * only check dboid and since it won't match any reader's database
2087 : : * OID, they will ignore this entry and move on.
2088 : : */
2089 : 32 : qe.length = QUEUE_PAGESIZE - offset;
2090 : 32 : qe.dboid = InvalidOid;
2091 : 32 : qe.xid = InvalidTransactionId;
2092 : 32 : qe.data[0] = '\0'; /* empty channel */
2093 : 32 : qe.data[1] = '\0'; /* empty payload */
2094 : : }
2095 : :
2096 : : /* Now copy qe into the shared buffer page */
2097 : 1146 : memcpy(NotifyCtl->shared->page_buffer[slotno] + offset,
2098 : : &qe,
2099 : 1146 : qe.length);
2100 : :
2101 : : /* Advance queue_head appropriately, and detect if page is full */
2102 [ + + ]: 1146 : if (asyncQueueAdvance(&(queue_head), qe.length))
2103 : : {
2104 : : LWLock *lock;
2105 : :
2106 : 35 : pageno = QUEUE_POS_PAGE(queue_head);
2107 : 35 : lock = SimpleLruGetBankLock(NotifyCtl, pageno);
2108 [ - + ]: 35 : if (lock != prevlock)
2109 : : {
2110 : 0 : LWLockRelease(prevlock);
2111 : 0 : LWLockAcquire(lock, LW_EXCLUSIVE);
2112 : 0 : prevlock = lock;
2113 : : }
2114 : :
2115 : : /*
2116 : : * Page is full, so we're done here, but first fill the next page
2117 : : * with zeroes. The reason to do this is to ensure that slru.c's
2118 : : * idea of the head page is always the same as ours, which avoids
2119 : : * boundary problems in SimpleLruTruncate. The test in
2120 : : * asyncQueueIsFull() ensured that there is room to create this
2121 : : * page without overrunning the queue.
2122 : : */
2123 : 35 : slotno = SimpleLruZeroPage(NotifyCtl, QUEUE_POS_PAGE(queue_head));
2124 : :
2125 : : /*
2126 : : * If the new page address is a multiple of QUEUE_CLEANUP_DELAY,
2127 : : * set flag to remember that we should try to advance the tail
2128 : : * pointer (we don't want to actually do that right here).
2129 : : */
2130 [ + + ]: 35 : if (QUEUE_POS_PAGE(queue_head) % QUEUE_CLEANUP_DELAY == 0)
2131 : 8 : tryAdvanceTail = true;
2132 : :
2133 : : /* And exit the loop */
2134 : 35 : break;
2135 : : }
2136 : : }
2137 : :
2138 : : /* Success, so update the global QUEUE_HEAD */
2139 : 104 : QUEUE_HEAD = queue_head;
2140 : :
2141 : 104 : LWLockRelease(prevlock);
2142 : :
2143 : 104 : return nextNotify;
2144 : : }
2145 : :
2146 : : /*
2147 : : * SQL function to return the fraction of the notification queue currently
2148 : : * occupied.
2149 : : */
2150 : : Datum
2151 : 6 : pg_notification_queue_usage(PG_FUNCTION_ARGS)
2152 : : {
2153 : : double usage;
2154 : :
2155 : : /* Advance the queue tail so we don't report a too-large result */
2156 : 6 : asyncQueueAdvanceTail();
2157 : :
2158 : 6 : LWLockAcquire(NotifyQueueLock, LW_SHARED);
2159 : 6 : usage = asyncQueueUsage();
2160 : 6 : LWLockRelease(NotifyQueueLock);
2161 : :
2162 : 6 : PG_RETURN_FLOAT8(usage);
2163 : : }
2164 : :
2165 : : /*
2166 : : * Return the fraction of the queue that is currently occupied.
2167 : : *
2168 : : * The caller must hold NotifyQueueLock in (at least) shared mode.
2169 : : *
2170 : : * Note: we measure the distance to the logical tail page, not the physical
2171 : : * tail page. In some sense that's wrong, but the relative position of the
2172 : : * physical tail is affected by details such as SLRU segment boundaries,
2173 : : * so that a result based on that is unpleasantly unstable.
2174 : : */
2175 : : static double
2176 : 110 : asyncQueueUsage(void)
2177 : : {
2178 : 110 : int64 headPage = QUEUE_POS_PAGE(QUEUE_HEAD);
2179 : 110 : int64 tailPage = QUEUE_POS_PAGE(QUEUE_TAIL);
2180 : 110 : int64 occupied = headPage - tailPage;
2181 : :
2182 [ + + ]: 110 : if (occupied == 0)
2183 : 71 : return (double) 0; /* fast exit for common case */
2184 : :
2185 : 39 : return (double) occupied / (double) max_notify_queue_pages;
2186 : : }
2187 : :
2188 : : /*
2189 : : * Check whether the queue is at least half full, and emit a warning if so.
2190 : : *
2191 : : * This is unlikely given the size of the queue, but possible.
2192 : : * The warnings show up at most once every QUEUE_FULL_WARN_INTERVAL.
2193 : : *
2194 : : * Caller must hold exclusive NotifyQueueLock.
2195 : : */
2196 : : static void
2197 : 104 : asyncQueueFillWarning(void)
2198 : : {
2199 : : double fillDegree;
2200 : : TimestampTz t;
2201 : :
2202 : 104 : fillDegree = asyncQueueUsage();
2203 [ + - ]: 104 : if (fillDegree < 0.5)
2204 : 104 : return;
2205 : :
2206 : 0 : t = GetCurrentTimestamp();
2207 : :
2208 [ # # ]: 0 : if (TimestampDifferenceExceeds(asyncQueueControl->lastQueueFillWarn,
2209 : : t, QUEUE_FULL_WARN_INTERVAL))
2210 : : {
2211 : 0 : QueuePosition min = QUEUE_HEAD;
2212 : 0 : int32 minPid = InvalidPid;
2213 : :
2214 [ # # ]: 0 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
2215 : : {
2216 : : Assert(QUEUE_BACKEND_PID(i) != InvalidPid);
2217 [ # # # # : 0 : min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
# # ]
2218 [ # # # # ]: 0 : if (QUEUE_POS_EQUAL(min, QUEUE_BACKEND_POS(i)))
2219 : 0 : minPid = QUEUE_BACKEND_PID(i);
2220 : : }
2221 : :
2222 [ # # # # : 0 : ereport(WARNING,
# # ]
2223 : : (errmsg("NOTIFY queue is %.0f%% full", fillDegree * 100),
2224 : : (minPid != InvalidPid ?
2225 : : errdetail("The server process with PID %d is among those with the oldest transactions.", minPid)
2226 : : : 0),
2227 : : (minPid != InvalidPid ?
2228 : : errhint("The NOTIFY queue cannot be emptied until that process ends its current transaction.")
2229 : : : 0)));
2230 : :
2231 : 0 : asyncQueueControl->lastQueueFillWarn = t;
2232 : : }
2233 : : }
2234 : :
2235 : : /*
2236 : : * Send signals to listening backends.
2237 : : *
2238 : : * Normally we signal only backends that are interested in the notifies that
2239 : : * we just sent. However, that will leave idle listeners falling further and
2240 : : * further behind. Waken them anyway if they're far enough behind, so they'll
2241 : : * advance their queue position pointers, allowing the global tail to advance.
2242 : : *
2243 : : * Since we know the ProcNumber and the Pid the signaling is quite cheap.
2244 : : *
2245 : : * This is called during CommitTransaction(), so it's important for it
2246 : : * to have very low probability of failure.
2247 : : */
2248 : : static void
2249 : 69 : SignalBackends(void)
2250 : : {
2251 : : int count;
2252 : :
2253 : : /* Can't get here without PreCommit_Notify having made the global table */
2254 : : Assert(globalChannelTable != NULL);
2255 : :
2256 : : /* It should have set up these arrays, too */
2257 : : Assert(signalPids != NULL && signalProcnos != NULL);
2258 : :
2259 : : /*
2260 : : * Identify backends that we need to signal. We don't want to send
2261 : : * signals while holding the NotifyQueueLock, so this part just builds a
2262 : : * list of target PIDs in signalPids[] and signalProcnos[].
2263 : : */
2264 : 69 : count = 0;
2265 : :
2266 : 69 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
2267 : :
2268 : : /* Scan each channel name that we notified in this transaction */
2269 [ + - + + : 210 : foreach_ptr(char, channel, pendingNotifies->uniqueChannelNames)
+ + ]
2270 : : {
2271 : : GlobalChannelKey key;
2272 : : GlobalChannelEntry *entry;
2273 : : ListenerEntry *listeners;
2274 : :
2275 : 72 : GlobalChannelKeyInit(&key, MyDatabaseId, channel);
2276 : 72 : entry = dshash_find(globalChannelTable, &key, false);
2277 [ + + ]: 72 : if (entry == NULL)
2278 : 32 : continue; /* nobody is listening */
2279 : :
2280 : 40 : listeners = (ListenerEntry *) dsa_get_address(globalChannelDSA,
2281 : : entry->listenersArray);
2282 : :
2283 : : /*
2284 : : * Identify listeners that now need waking, add them to arrays.
2285 : : *
2286 : : * Note that we signal listeners regardless of the state of their
2287 : : * removeOnAbort flags. Hence a new listener that reached PreCommit,
2288 : : * but then failed before AtCommit_Notify, can receive a signal even
2289 : : * though it was never really listening. This is okay because it will
2290 : : * not do anything in response to that signal. If we did not do it
2291 : : * like this then a new listener might miss some messages due to the
2292 : : * direct-advance logic below.
2293 : : */
2294 [ + + ]: 83 : for (int j = 0; j < entry->numListeners; j++)
2295 : : {
2296 : 43 : ProcNumber i = listeners[j].procNo;
2297 : : int32 pid;
2298 : : QueuePosition pos;
2299 : :
2300 [ + + ]: 43 : if (QUEUE_BACKEND_WAKEUP_PENDING(i))
2301 : 14 : continue; /* already signaled, no need to repeat */
2302 : :
2303 : 29 : pid = QUEUE_BACKEND_PID(i);
2304 : 29 : pos = QUEUE_BACKEND_POS(i);
2305 : :
2306 [ + + - + ]: 29 : if (QUEUE_POS_EQUAL(pos, QUEUE_HEAD))
2307 : 0 : continue; /* it's fully caught up already */
2308 : :
2309 : : Assert(pid != InvalidPid);
2310 : :
2311 : 29 : QUEUE_BACKEND_WAKEUP_PENDING(i) = true;
2312 : 29 : signalPids[count] = pid;
2313 : 29 : signalProcnos[count] = i;
2314 : 29 : count++;
2315 : : }
2316 : :
2317 : 40 : dshash_release_lock(globalChannelTable, entry);
2318 : : }
2319 : :
2320 : : /*
2321 : : * Scan all listeners. Any that are not already pending wakeup must not
2322 : : * be interested in our notifications (else we'd have set their wakeup
2323 : : * flags above). Check to see if we can directly advance their queue
2324 : : * pointers to save a wakeup. Otherwise, if they are far behind, wake
2325 : : * them anyway so they will catch up.
2326 : : */
2327 [ + + ]: 124 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
2328 : : {
2329 : : int32 pid;
2330 : : QueuePosition pos;
2331 : :
2332 [ + + ]: 55 : if (QUEUE_BACKEND_WAKEUP_PENDING(i))
2333 : 41 : continue;
2334 : :
2335 : : /* If it's currently advancing, we should not touch it */
2336 [ - + ]: 14 : if (QUEUE_BACKEND_IS_ADVANCING(i))
2337 : 0 : continue;
2338 : :
2339 : 14 : pid = QUEUE_BACKEND_PID(i);
2340 : 14 : pos = QUEUE_BACKEND_POS(i);
2341 : :
2342 : : /*
2343 : : * We can directly advance the other backend's queue pointer if it's
2344 : : * not currently advancing (else there are race conditions), and its
2345 : : * current pointer is not behind queueHeadBeforeWrite (else we'd make
2346 : : * it miss some older messages), and we'd not be moving the pointer
2347 : : * backward.
2348 : : */
2349 [ + - + - : 28 : if (!QUEUE_POS_PRECEDES(pos, queueHeadBeforeWrite) &&
+ - + + ]
2350 [ + - + - ]: 17 : QUEUE_POS_PRECEDES(pos, queueHeadAfterWrite))
2351 : : {
2352 : : /* We can directly advance its pointer past what we wrote */
2353 : 14 : QUEUE_BACKEND_POS(i) = queueHeadAfterWrite;
2354 : : }
2355 [ # # ]: 0 : else if (asyncQueuePageDiff(QUEUE_POS_PAGE(QUEUE_HEAD),
2356 : : QUEUE_POS_PAGE(pos)) >= QUEUE_CLEANUP_DELAY)
2357 : : {
2358 : : /* It's idle and far behind, so wake it up */
2359 : : Assert(pid != InvalidPid);
2360 : :
2361 : 0 : QUEUE_BACKEND_WAKEUP_PENDING(i) = true;
2362 : 0 : signalPids[count] = pid;
2363 : 0 : signalProcnos[count] = i;
2364 : 0 : count++;
2365 : : }
2366 : : }
2367 : :
2368 : 69 : LWLockRelease(NotifyQueueLock);
2369 : :
2370 : : /* Now send signals */
2371 [ + + ]: 98 : for (int i = 0; i < count; i++)
2372 : : {
2373 : 29 : int32 pid = signalPids[i];
2374 : :
2375 : : /*
2376 : : * If we are signaling our own process, no need to involve the kernel;
2377 : : * just set the flag directly.
2378 : : */
2379 [ + + ]: 29 : if (pid == MyProcPid)
2380 : : {
2381 : 21 : notifyInterruptPending = true;
2382 : 21 : continue;
2383 : : }
2384 : :
2385 : : /*
2386 : : * Note: assuming things aren't broken, a signal failure here could
2387 : : * only occur if the target backend exited since we released
2388 : : * NotifyQueueLock; which is unlikely but certainly possible. So we
2389 : : * just log a low-level debug message if it happens.
2390 : : */
2391 [ - + ]: 8 : if (SendProcSignal(pid, PROCSIG_NOTIFY_INTERRUPT, signalProcnos[i]) < 0)
2392 [ # # ]: 0 : elog(DEBUG3, "could not signal backend with PID %d: %m", pid);
2393 : : }
2394 : 69 : }
2395 : :
2396 : : /*
2397 : : * AtAbort_Notify
2398 : : *
2399 : : * This is called at transaction abort.
2400 : : *
2401 : : * Revert any staged listen/unlisten changes and clean up transaction state.
2402 : : */
2403 : : void
2404 : 36077 : AtAbort_Notify(void)
2405 : : {
2406 : : /*
2407 : : * If we failed during asyncQueueReadAllNotifications(), clear our
2408 : : * isAdvancing flag, else SignalBackends() would never wake us again.
2409 : : * Since only our own process changes this flag, it is safe to check it
2410 : : * without the lock.
2411 : : */
2412 [ - + - - ]: 36077 : if (amRegisteredListener && QUEUE_BACKEND_IS_ADVANCING(MyProcNumber))
2413 : : {
2414 : 0 : LWLockAcquire(NotifyQueueLock, LW_SHARED);
2415 : 0 : QUEUE_BACKEND_IS_ADVANCING(MyProcNumber) = false;
2416 : 0 : LWLockRelease(NotifyQueueLock);
2417 : : }
2418 : :
2419 : : /* Revert any staged listen/unlisten changes */
2420 : 36077 : ApplyPendingListenActions(false);
2421 : :
2422 : : /* If we're no longer listening on anything, unregister */
2423 [ - + - - : 36077 : if (amRegisteredListener && LocalChannelTableIsEmpty())
- - ]
2424 : 0 : asyncQueueUnregister();
2425 : :
2426 : : /* And clean up */
2427 : 36077 : ClearPendingActionsAndNotifies();
2428 : 36077 : }
2429 : :
2430 : : /*
2431 : : * AtSubCommit_Notify() --- Take care of subtransaction commit.
2432 : : *
2433 : : * Reassign all items in the pending lists to the parent transaction.
2434 : : */
2435 : : void
2436 : 17390 : AtSubCommit_Notify(void)
2437 : : {
2438 : 17390 : int my_level = GetCurrentTransactionNestLevel();
2439 : :
2440 : : /* If there are actions at our nesting level, we must reparent them. */
2441 [ + + ]: 17390 : if (pendingActions != NULL &&
2442 [ + - ]: 2 : pendingActions->nestingLevel >= my_level)
2443 : : {
2444 [ + + ]: 2 : if (pendingActions->upper == NULL ||
2445 [ - + ]: 1 : pendingActions->upper->nestingLevel < my_level - 1)
2446 : : {
2447 : : /* nothing to merge; give the whole thing to the parent */
2448 : 1 : --pendingActions->nestingLevel;
2449 : : }
2450 : : else
2451 : : {
2452 : 1 : ActionList *childPendingActions = pendingActions;
2453 : :
2454 : 1 : pendingActions = pendingActions->upper;
2455 : :
2456 : : /*
2457 : : * Mustn't try to eliminate duplicates here --- see queue_listen()
2458 : : */
2459 : 2 : pendingActions->actions =
2460 : 1 : list_concat(pendingActions->actions,
2461 : 1 : childPendingActions->actions);
2462 : 1 : pfree(childPendingActions);
2463 : : }
2464 : : }
2465 : :
2466 : : /* If there are notifies at our nesting level, we must reparent them. */
2467 [ + + ]: 17390 : if (pendingNotifies != NULL &&
2468 [ + + ]: 3 : pendingNotifies->nestingLevel >= my_level)
2469 : : {
2470 : : Assert(pendingNotifies->nestingLevel == my_level);
2471 : :
2472 [ + + ]: 2 : if (pendingNotifies->upper == NULL ||
2473 [ - + ]: 1 : pendingNotifies->upper->nestingLevel < my_level - 1)
2474 : : {
2475 : : /* nothing to merge; give the whole thing to the parent */
2476 : 1 : --pendingNotifies->nestingLevel;
2477 : : }
2478 : : else
2479 : : {
2480 : : /*
2481 : : * Formerly, we didn't bother to eliminate duplicates here, but
2482 : : * now we must, else we fall foul of "Assert(!found)", either here
2483 : : * or during a later attempt to build the parent-level hashtable.
2484 : : */
2485 : 1 : NotificationList *childPendingNotifies = pendingNotifies;
2486 : : ListCell *l;
2487 : :
2488 : 1 : pendingNotifies = pendingNotifies->upper;
2489 : : /* Insert all the subxact's events into parent, except for dups */
2490 [ + - + + : 5 : foreach(l, childPendingNotifies->events)
+ + ]
2491 : : {
2492 : 4 : Notification *childn = (Notification *) lfirst(l);
2493 : :
2494 [ + + ]: 4 : if (!AsyncExistsPendingNotify(childn))
2495 : 2 : AddEventToPendingNotifies(childn);
2496 : : }
2497 : 1 : pfree(childPendingNotifies);
2498 : : }
2499 : : }
2500 : 17390 : }
2501 : :
2502 : : /*
2503 : : * AtSubAbort_Notify() --- Take care of subtransaction abort.
2504 : : */
2505 : : void
2506 : 5614 : AtSubAbort_Notify(void)
2507 : : {
2508 : 5614 : int my_level = GetCurrentTransactionNestLevel();
2509 : :
2510 : : /*
2511 : : * All we have to do is pop the stack --- the actions/notifies made in
2512 : : * this subxact are no longer interesting, and the space will be freed
2513 : : * when CurTransactionContext is recycled. We still have to free the
2514 : : * ActionList and NotificationList objects themselves, though, because
2515 : : * those are allocated in TopTransactionContext.
2516 : : *
2517 : : * Note that there might be no entries at all, or no entries for the
2518 : : * current subtransaction level, either because none were ever created, or
2519 : : * because we reentered this routine due to trouble during subxact abort.
2520 : : */
2521 [ + + ]: 5615 : while (pendingActions != NULL &&
2522 [ + - ]: 1 : pendingActions->nestingLevel >= my_level)
2523 : : {
2524 : 1 : ActionList *childPendingActions = pendingActions;
2525 : :
2526 : 1 : pendingActions = pendingActions->upper;
2527 : 1 : pfree(childPendingActions);
2528 : : }
2529 : :
2530 [ + + ]: 5615 : while (pendingNotifies != NULL &&
2531 [ + + ]: 2 : pendingNotifies->nestingLevel >= my_level)
2532 : : {
2533 : 1 : NotificationList *childPendingNotifies = pendingNotifies;
2534 : :
2535 : 1 : pendingNotifies = pendingNotifies->upper;
2536 : 1 : pfree(childPendingNotifies);
2537 : : }
2538 : 5614 : }
2539 : :
2540 : : /*
2541 : : * HandleNotifyInterrupt
2542 : : *
2543 : : * Signal handler portion of interrupt handling. Let the backend know
2544 : : * that there's a pending notify interrupt. If we're currently reading
2545 : : * from the client, this will interrupt the read and
2546 : : * ProcessClientReadInterrupt() will call ProcessNotifyInterrupt().
2547 : : */
2548 : : void
2549 : 7 : HandleNotifyInterrupt(void)
2550 : : {
2551 : : /*
2552 : : * Note: this is called by a SIGNAL HANDLER. You must be very wary what
2553 : : * you do here.
2554 : : */
2555 : :
2556 : : /* signal that work needs to be done */
2557 : 7 : notifyInterruptPending = true;
2558 : :
2559 : : /* latch will be set by procsignal_sigusr1_handler */
2560 : 7 : }
2561 : :
2562 : : /*
2563 : : * ProcessNotifyInterrupt
2564 : : *
2565 : : * This is called if we see notifyInterruptPending set, just before
2566 : : * transmitting ReadyForQuery at the end of a frontend command, and
2567 : : * also if a notify signal occurs while reading from the frontend.
2568 : : * HandleNotifyInterrupt() will cause the read to be interrupted
2569 : : * via the process's latch, and this routine will get called.
2570 : : * If we are truly idle (ie, *not* inside a transaction block),
2571 : : * process the incoming notifies.
2572 : : *
2573 : : * If "flush" is true, force any frontend messages out immediately.
2574 : : * This can be false when being called at the end of a frontend command,
2575 : : * since we'll flush after sending ReadyForQuery.
2576 : : */
2577 : : void
2578 : 33 : ProcessNotifyInterrupt(bool flush)
2579 : : {
2580 [ + + ]: 33 : if (IsTransactionOrTransactionBlock())
2581 : 6 : return; /* not really idle */
2582 : :
2583 : : /* Loop in case another signal arrives while sending messages */
2584 [ + + ]: 54 : while (notifyInterruptPending)
2585 : 27 : ProcessIncomingNotify(flush);
2586 : : }
2587 : :
2588 : :
2589 : : /*
2590 : : * Read all pending notifications from the queue, and deliver appropriate
2591 : : * ones to my frontend. Stop when we reach queue head or an uncommitted
2592 : : * notification.
2593 : : */
2594 : : static void
2595 : 45 : asyncQueueReadAllNotifications(void)
2596 : : {
2597 : : QueuePosition pos;
2598 : : QueuePosition head;
2599 : : Snapshot snapshot;
2600 : :
2601 : : /*
2602 : : * Fetch current state, indicate to others that we have woken up, and that
2603 : : * we are in process of advancing our position.
2604 : : */
2605 : 45 : LWLockAcquire(NotifyQueueLock, LW_SHARED);
2606 : : /* Assert checks that we have a valid state entry */
2607 : : Assert(MyProcPid == QUEUE_BACKEND_PID(MyProcNumber));
2608 : 45 : QUEUE_BACKEND_WAKEUP_PENDING(MyProcNumber) = false;
2609 : 45 : pos = QUEUE_BACKEND_POS(MyProcNumber);
2610 : 45 : head = QUEUE_HEAD;
2611 : :
2612 [ + + - + ]: 45 : if (QUEUE_POS_EQUAL(pos, head))
2613 : : {
2614 : : /* Nothing to do, we have read all notifications already. */
2615 : 0 : LWLockRelease(NotifyQueueLock);
2616 : 0 : return;
2617 : : }
2618 : :
2619 : 45 : QUEUE_BACKEND_IS_ADVANCING(MyProcNumber) = true;
2620 : 45 : LWLockRelease(NotifyQueueLock);
2621 : :
2622 : : /*----------
2623 : : * Get snapshot we'll use to decide which xacts are still in progress.
2624 : : * This is trickier than it might seem, because of race conditions.
2625 : : * Consider the following example:
2626 : : *
2627 : : * Backend 1: Backend 2:
2628 : : *
2629 : : * transaction starts
2630 : : * UPDATE foo SET ...;
2631 : : * NOTIFY foo;
2632 : : * commit starts
2633 : : * queue the notify message
2634 : : * transaction starts
2635 : : * LISTEN foo; -- first LISTEN in session
2636 : : * SELECT * FROM foo WHERE ...;
2637 : : * commit to clog
2638 : : * commit starts
2639 : : * add backend 2 to array of listeners
2640 : : * advance to queue head (this code)
2641 : : * commit to clog
2642 : : *
2643 : : * Transaction 2's SELECT has not seen the UPDATE's effects, since that
2644 : : * wasn't committed yet. Ideally we'd ensure that client 2 would
2645 : : * eventually get transaction 1's notify message, but there's no way
2646 : : * to do that; until we're in the listener array, there's no guarantee
2647 : : * that the notify message doesn't get removed from the queue.
2648 : : *
2649 : : * Therefore the coding technique transaction 2 is using is unsafe:
2650 : : * applications must commit a LISTEN before inspecting database state,
2651 : : * if they want to ensure they will see notifications about subsequent
2652 : : * changes to that state.
2653 : : *
2654 : : * What we do guarantee is that we'll see all notifications from
2655 : : * transactions committing after the snapshot we take here.
2656 : : * BecomeRegisteredListener has already added us to the listener array,
2657 : : * so no not-yet-committed messages can be removed from the queue
2658 : : * before we see them.
2659 : : *
2660 : : * Since we'll only use this snapshot to check whether XIDs are still
2661 : : * running, and we will release it before returning, it's okay to use
2662 : : * GetSecondarySnapshot(); that avoids possibly blocking on other
2663 : : * transactions in SERIALIZABLE READ ONLY DEFERRABLE mode.
2664 : : *----------
2665 : : */
2666 : 45 : snapshot = RegisterSnapshot(GetSecondarySnapshot());
2667 : :
2668 : : /*
2669 : : * It is possible that we fail while trying to send a message to our
2670 : : * frontend (for example, because of encoding conversion failure). If
2671 : : * that happens it is critical that we not try to send the same message
2672 : : * over and over again. Therefore, we set ExitOnAnyError to upgrade any
2673 : : * ERRORs to FATAL, causing the client connection to be closed on error.
2674 : : *
2675 : : * We used to only skip over the offending message and try to soldier on,
2676 : : * but it was somewhat questionable to lose a notification and give the
2677 : : * client an ERROR instead. A client application is not be prepared for
2678 : : * that and can't tell that a notification was missed. It was also not
2679 : : * very useful in practice because notifications are often processed while
2680 : : * a connection is idle and reading a message from the client, and in that
2681 : : * state, any error is upgraded to FATAL anyway. Closing the connection
2682 : : * is a clear signal to the application that it might have missed
2683 : : * notifications.
2684 : : */
2685 : : {
2686 : 45 : bool save_ExitOnAnyError = ExitOnAnyError;
2687 : : bool reachedStop;
2688 : :
2689 : 45 : ExitOnAnyError = true;
2690 : :
2691 : : do
2692 : : {
2693 : : /*
2694 : : * Process messages up to the stop position, end of page, or an
2695 : : * uncommitted message.
2696 : : *
2697 : : * Our stop position is what we found to be the head's position
2698 : : * when we entered this function. It might have changed already.
2699 : : * But if it has, we will receive (or have already received and
2700 : : * queued) another signal and come here again.
2701 : : *
2702 : : * We are not holding NotifyQueueLock here! The queue can only
2703 : : * extend beyond the head pointer (see above) and we leave our
2704 : : * backend's pointer where it is so nobody will truncate or
2705 : : * rewrite pages under us. Especially we don't want to hold a lock
2706 : : * while sending the notifications to the frontend.
2707 : : */
2708 : 48 : reachedStop = asyncQueueProcessPageEntries(&pos, head, snapshot);
2709 [ + + ]: 48 : } while (!reachedStop);
2710 : :
2711 : : /* Update shared state */
2712 : 45 : LWLockAcquire(NotifyQueueLock, LW_SHARED);
2713 : 45 : QUEUE_BACKEND_POS(MyProcNumber) = pos;
2714 : 45 : QUEUE_BACKEND_IS_ADVANCING(MyProcNumber) = false;
2715 : 45 : LWLockRelease(NotifyQueueLock);
2716 : :
2717 : 45 : ExitOnAnyError = save_ExitOnAnyError;
2718 : : }
2719 : :
2720 : : /* Done with snapshot */
2721 : 45 : UnregisterSnapshot(snapshot);
2722 : : }
2723 : :
2724 : : /*
2725 : : * Fetch notifications from the shared queue, beginning at position current,
2726 : : * and deliver relevant ones to my frontend.
2727 : : *
2728 : : * The function returns true once we have reached the stop position or an
2729 : : * uncommitted notification, and false if we have finished with the page.
2730 : : * In other words: once it returns true there is no need to look further.
2731 : : * The QueuePosition *current is advanced past all processed messages.
2732 : : */
2733 : : static bool
2734 : 48 : asyncQueueProcessPageEntries(QueuePosition *current,
2735 : : QueuePosition stop,
2736 : : Snapshot snapshot)
2737 : : {
2738 : 48 : int64 curpage = QUEUE_POS_PAGE(*current);
2739 : : int slotno;
2740 : : char *page_buffer;
2741 : 48 : bool reachedStop = false;
2742 : : bool reachedEndOfPage;
2743 : :
2744 : : /*
2745 : : * We copy the entries into a local buffer to avoid holding the SLRU lock
2746 : : * while we transmit them to our frontend. The local buffer must be
2747 : : * adequately aligned.
2748 : : */
2749 : : alignas(AsyncQueueEntry) char local_buf[QUEUE_PAGESIZE];
2750 : 48 : char *local_buf_end = local_buf;
2751 : :
2752 : 48 : slotno = SimpleLruReadPage_ReadOnly(NotifyCtl, curpage, current);
2753 : 48 : page_buffer = NotifyCtl->shared->page_buffer[slotno];
2754 : :
2755 : : do
2756 : : {
2757 : 1413 : QueuePosition thisentry = *current;
2758 : : AsyncQueueEntry *qe;
2759 : :
2760 [ + + + + ]: 1413 : if (QUEUE_POS_EQUAL(thisentry, stop))
2761 : 45 : break;
2762 : :
2763 : 1368 : qe = (AsyncQueueEntry *) (page_buffer + QUEUE_POS_OFFSET(thisentry));
2764 : :
2765 : : /*
2766 : : * Advance *current over this message, possibly to the next page. As
2767 : : * noted in the comments for asyncQueueReadAllNotifications, we must
2768 : : * do this before possibly failing while processing the message.
2769 : : */
2770 : 1368 : reachedEndOfPage = asyncQueueAdvance(current, qe->length);
2771 : :
2772 : : /* Ignore messages destined for other databases */
2773 [ + - ]: 1368 : if (qe->dboid == MyDatabaseId)
2774 : : {
2775 [ - + ]: 1368 : if (XidInMVCCSnapshot(qe->xid, snapshot))
2776 : : {
2777 : : /*
2778 : : * The source transaction is still in progress, so we can't
2779 : : * process this message yet. Break out of the loop, but first
2780 : : * back up *current so we will reprocess the message next
2781 : : * time. (Note: it is unlikely but not impossible for
2782 : : * TransactionIdDidCommit to fail, so we can't really avoid
2783 : : * this advance-then-back-up behavior when dealing with an
2784 : : * uncommitted message.)
2785 : : *
2786 : : * Note that we must test XidInMVCCSnapshot before we test
2787 : : * TransactionIdDidCommit, else we might return a message from
2788 : : * a transaction that is not yet visible to snapshots; compare
2789 : : * the comments at the head of heapam_visibility.c.
2790 : : *
2791 : : * Also, while our own xact won't be listed in the snapshot,
2792 : : * we need not check for TransactionIdIsCurrentTransactionId
2793 : : * because our transaction cannot (yet) have queued any
2794 : : * messages.
2795 : : */
2796 : 0 : *current = thisentry;
2797 : 0 : reachedStop = true;
2798 : 0 : break;
2799 : : }
2800 : :
2801 : : /*
2802 : : * Quick check for the case that we're not listening on any
2803 : : * channels, before calling TransactionIdDidCommit(). This makes
2804 : : * that case a little faster, but more importantly, it ensures
2805 : : * that if there's a bad entry in the queue for which
2806 : : * TransactionIdDidCommit() fails for some reason, we can skip
2807 : : * over it on the first LISTEN in a session, and not get stuck on
2808 : : * it indefinitely. (This is a little trickier than it looks: it
2809 : : * works because BecomeRegisteredListener runs this code before we
2810 : : * have made the first entry in localChannelTable.)
2811 : : */
2812 [ + - + + ]: 1368 : if (LocalChannelTableIsEmpty())
2813 : 1314 : continue;
2814 : :
2815 [ + - ]: 54 : if (TransactionIdDidCommit(qe->xid))
2816 : : {
2817 : 54 : memcpy(local_buf_end, qe, qe->length);
2818 : 54 : local_buf_end += qe->length;
2819 : : }
2820 : : else
2821 : : {
2822 : : /*
2823 : : * The source transaction aborted or crashed, so we just
2824 : : * ignore its notifications.
2825 : : */
2826 : : }
2827 : : }
2828 : :
2829 : : /* Loop back if we're not at end of page */
2830 [ + + ]: 1368 : } while (!reachedEndOfPage);
2831 : :
2832 : : /* Release lock that we got from SimpleLruReadPage_ReadOnly() */
2833 : 48 : LWLockRelease(SimpleLruGetBankLock(NotifyCtl, curpage));
2834 : :
2835 : : /*
2836 : : * Now that we have let go of the SLRU bank lock, send the notifications
2837 : : * to our backend
2838 : : */
2839 : : Assert(local_buf_end - local_buf <= BLCKSZ);
2840 [ + + ]: 102 : for (char *p = local_buf; p < local_buf_end;)
2841 : : {
2842 : 54 : AsyncQueueEntry *qe = (AsyncQueueEntry *) p;
2843 : :
2844 : : /* qe->data is the null-terminated channel name */
2845 : 54 : char *channel = qe->data;
2846 : :
2847 [ + + ]: 54 : if (IsListeningOn(channel))
2848 : : {
2849 : : /* payload follows channel name */
2850 : 53 : char *payload = qe->data + strlen(channel) + 1;
2851 : :
2852 : 53 : NotifyMyFrontEnd(channel, payload, qe->srcPid);
2853 : : }
2854 : :
2855 : 54 : p += qe->length;
2856 : : }
2857 : :
2858 [ + + + + ]: 48 : if (QUEUE_POS_EQUAL(*current, stop))
2859 : 45 : reachedStop = true;
2860 : :
2861 : 48 : return reachedStop;
2862 : : }
2863 : :
2864 : : /*
2865 : : * Advance the shared queue tail variable to the minimum of all the
2866 : : * per-backend tail pointers. Truncate pg_notify space if possible.
2867 : : *
2868 : : * This is (usually) called during CommitTransaction(), so it's important for
2869 : : * it to have very low probability of failure.
2870 : : */
2871 : : static void
2872 : 14 : asyncQueueAdvanceTail(void)
2873 : : {
2874 : : QueuePosition min;
2875 : : int64 oldtailpage;
2876 : : int64 newtailpage;
2877 : : int64 boundary;
2878 : :
2879 : : /* Restrict task to one backend per cluster; see SimpleLruTruncate(). */
2880 : 14 : LWLockAcquire(NotifyQueueTailLock, LW_EXCLUSIVE);
2881 : :
2882 : : /*
2883 : : * Compute the new tail. Pre-v13, it's essential that QUEUE_TAIL be exact
2884 : : * (ie, exactly match at least one backend's queue position), so it must
2885 : : * be updated atomically with the actual computation. Since v13, we could
2886 : : * get away with not doing it like that, but it seems prudent to keep it
2887 : : * so.
2888 : : *
2889 : : * Also, because incoming backends will scan forward from QUEUE_TAIL, that
2890 : : * must be advanced before we can truncate any data. Thus, QUEUE_TAIL is
2891 : : * the logical tail, while QUEUE_STOP_PAGE is the physical tail, or oldest
2892 : : * un-truncated page. When QUEUE_STOP_PAGE != QUEUE_POS_PAGE(QUEUE_TAIL),
2893 : : * there are pages we can truncate but haven't yet finished doing so.
2894 : : *
2895 : : * For concurrency's sake, we don't want to hold NotifyQueueLock while
2896 : : * performing SimpleLruTruncate. This is OK because no backend will try
2897 : : * to access the pages we are in the midst of truncating.
2898 : : */
2899 : 14 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
2900 : 14 : min = QUEUE_HEAD;
2901 [ + + ]: 24 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
2902 : : {
2903 : : Assert(QUEUE_BACKEND_PID(i) != InvalidPid);
2904 [ + - + + : 10 : min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
+ - ]
2905 : : }
2906 : 14 : QUEUE_TAIL = min;
2907 : 14 : oldtailpage = QUEUE_STOP_PAGE;
2908 : 14 : LWLockRelease(NotifyQueueLock);
2909 : :
2910 : : /*
2911 : : * We can truncate something if the global tail advanced across an SLRU
2912 : : * segment boundary.
2913 : : *
2914 : : * XXX it might be better to truncate only once every several segments, to
2915 : : * reduce the number of directory scans.
2916 : : */
2917 : 14 : newtailpage = QUEUE_POS_PAGE(min);
2918 : 14 : boundary = newtailpage - (newtailpage % SLRU_PAGES_PER_SEGMENT);
2919 [ + + ]: 14 : if (asyncQueuePagePrecedes(oldtailpage, boundary))
2920 : : {
2921 : : /*
2922 : : * SimpleLruTruncate() will ask for SLRU bank locks but will also
2923 : : * release the lock again.
2924 : : */
2925 : 1 : SimpleLruTruncate(NotifyCtl, newtailpage);
2926 : :
2927 : 1 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
2928 : 1 : QUEUE_STOP_PAGE = newtailpage;
2929 : 1 : LWLockRelease(NotifyQueueLock);
2930 : : }
2931 : :
2932 : 14 : LWLockRelease(NotifyQueueTailLock);
2933 : 14 : }
2934 : :
2935 : : /*
2936 : : * AsyncNotifyFreezeXids
2937 : : *
2938 : : * Prepare the async notification queue for CLOG truncation by freezing
2939 : : * transaction IDs that are about to become inaccessible.
2940 : : *
2941 : : * This function is called by VACUUM before advancing datfrozenxid. It scans
2942 : : * the notification queue and replaces XIDs that would become inaccessible
2943 : : * after CLOG truncation with special markers:
2944 : : * - Committed transactions are set to FrozenTransactionId
2945 : : * - Aborted/crashed transactions are set to InvalidTransactionId
2946 : : *
2947 : : * Only XIDs < newFrozenXid are processed, as those are the ones whose CLOG
2948 : : * pages will be truncated. If XID < newFrozenXid, it cannot still be running
2949 : : * (or it would have held back newFrozenXid through ProcArray).
2950 : : * Therefore, if TransactionIdDidCommit returns false, we know the transaction
2951 : : * either aborted explicitly or crashed, and we can safely mark it invalid.
2952 : : */
2953 : : void
2954 : 1198 : AsyncNotifyFreezeXids(TransactionId newFrozenXid)
2955 : : {
2956 : : QueuePosition pos;
2957 : : QueuePosition head;
2958 : 1198 : int64 curpage = -1;
2959 : 1198 : int slotno = -1;
2960 : 1198 : char *page_buffer = NULL;
2961 : 1198 : bool page_dirty = false;
2962 : :
2963 : : /*
2964 : : * Acquire locks in the correct order to avoid deadlocks. As per the
2965 : : * locking protocol: NotifyQueueTailLock, then NotifyQueueLock, then SLRU
2966 : : * bank locks.
2967 : : *
2968 : : * We only need SHARED mode since we're just reading the head/tail
2969 : : * positions, not modifying them.
2970 : : */
2971 : 1198 : LWLockAcquire(NotifyQueueTailLock, LW_SHARED);
2972 : 1198 : LWLockAcquire(NotifyQueueLock, LW_SHARED);
2973 : :
2974 : 1198 : pos = QUEUE_TAIL;
2975 : 1198 : head = QUEUE_HEAD;
2976 : :
2977 : : /* Release NotifyQueueLock early, we only needed to read the positions */
2978 : 1198 : LWLockRelease(NotifyQueueLock);
2979 : :
2980 : : /*
2981 : : * Scan the queue from tail to head, freezing XIDs as needed. We hold
2982 : : * NotifyQueueTailLock throughout to ensure the tail doesn't move while
2983 : : * we're working.
2984 : : */
2985 [ - + + + ]: 1228 : while (!QUEUE_POS_EQUAL(pos, head))
2986 : : {
2987 : : AsyncQueueEntry *qe;
2988 : : TransactionId xid;
2989 : 30 : int64 pageno = QUEUE_POS_PAGE(pos);
2990 : 30 : int offset = QUEUE_POS_OFFSET(pos);
2991 : :
2992 : : /* If we need a different page, release old lock and get new one */
2993 [ + + ]: 30 : if (pageno != curpage)
2994 : : {
2995 : : LWLock *lock;
2996 : :
2997 : : /* Release previous page if any */
2998 [ - + ]: 3 : if (slotno >= 0)
2999 : : {
3000 [ # # ]: 0 : if (page_dirty)
3001 : : {
3002 : 0 : NotifyCtl->shared->page_dirty[slotno] = true;
3003 : 0 : page_dirty = false;
3004 : : }
3005 : 0 : LWLockRelease(SimpleLruGetBankLock(NotifyCtl, curpage));
3006 : : }
3007 : :
3008 : 3 : lock = SimpleLruGetBankLock(NotifyCtl, pageno);
3009 : 3 : LWLockAcquire(lock, LW_EXCLUSIVE);
3010 : 3 : slotno = SimpleLruReadPage(NotifyCtl, pageno, true, &pos);
3011 : 3 : page_buffer = NotifyCtl->shared->page_buffer[slotno];
3012 : 3 : curpage = pageno;
3013 : : }
3014 : :
3015 : 30 : qe = (AsyncQueueEntry *) (page_buffer + offset);
3016 : 30 : xid = qe->xid;
3017 : :
3018 [ + - + + ]: 60 : if (TransactionIdIsNormal(xid) &&
3019 : 30 : TransactionIdPrecedes(xid, newFrozenXid))
3020 : : {
3021 [ + - ]: 10 : if (TransactionIdDidCommit(xid))
3022 : : {
3023 : 10 : qe->xid = FrozenTransactionId;
3024 : 10 : page_dirty = true;
3025 : : }
3026 : : else
3027 : : {
3028 : 0 : qe->xid = InvalidTransactionId;
3029 : 0 : page_dirty = true;
3030 : : }
3031 : : }
3032 : :
3033 : : /* Advance to next entry */
3034 : 30 : asyncQueueAdvance(&pos, qe->length);
3035 : : }
3036 : :
3037 : : /* Release final page lock if we acquired one */
3038 [ + + ]: 1198 : if (slotno >= 0)
3039 : : {
3040 [ + + ]: 3 : if (page_dirty)
3041 : 1 : NotifyCtl->shared->page_dirty[slotno] = true;
3042 : 3 : LWLockRelease(SimpleLruGetBankLock(NotifyCtl, curpage));
3043 : : }
3044 : :
3045 : 1198 : LWLockRelease(NotifyQueueTailLock);
3046 : 1198 : }
3047 : :
3048 : : /*
3049 : : * ProcessIncomingNotify
3050 : : *
3051 : : * Scan the queue for arriving notifications and report them to the front
3052 : : * end. The notifications might be from other sessions, or our own;
3053 : : * there's no need to distinguish here.
3054 : : *
3055 : : * If "flush" is true, force any frontend messages out immediately.
3056 : : *
3057 : : * NOTE: since we are outside any transaction, we must create our own.
3058 : : */
3059 : : static void
3060 : 27 : ProcessIncomingNotify(bool flush)
3061 : : {
3062 : : /* We *must* reset the flag */
3063 : 27 : notifyInterruptPending = false;
3064 : :
3065 : : /* Do nothing else if we aren't actively listening */
3066 [ + - - + ]: 27 : if (LocalChannelTableIsEmpty())
3067 : 0 : return;
3068 : :
3069 [ - + ]: 27 : if (Trace_notify)
3070 [ # # ]: 0 : elog(DEBUG1, "ProcessIncomingNotify");
3071 : :
3072 : 27 : set_ps_display("notify interrupt");
3073 : :
3074 : : /*
3075 : : * We must run asyncQueueReadAllNotifications inside a transaction, else
3076 : : * bad things happen if it gets an error.
3077 : : */
3078 : 27 : StartTransactionCommand();
3079 : :
3080 : 27 : asyncQueueReadAllNotifications();
3081 : :
3082 : 27 : CommitTransactionCommand();
3083 : :
3084 : : /*
3085 : : * If this isn't an end-of-command case, we must flush the notify messages
3086 : : * to ensure frontend gets them promptly.
3087 : : */
3088 [ + + ]: 27 : if (flush)
3089 : 5 : pq_flush();
3090 : :
3091 : 27 : set_ps_display("idle");
3092 : :
3093 [ - + ]: 27 : if (Trace_notify)
3094 [ # # ]: 0 : elog(DEBUG1, "ProcessIncomingNotify: done");
3095 : : }
3096 : :
3097 : : /*
3098 : : * Send NOTIFY message to my front end.
3099 : : */
3100 : : void
3101 : 53 : NotifyMyFrontEnd(const char *channel, const char *payload, int32 srcPid)
3102 : : {
3103 [ + - ]: 53 : if (whereToSendOutput == DestRemote)
3104 : : {
3105 : : StringInfoData buf;
3106 : :
3107 : 53 : pq_beginmessage(&buf, PqMsg_NotificationResponse);
3108 : 53 : pq_sendint32(&buf, srcPid);
3109 : 53 : pq_sendstring(&buf, channel);
3110 : 53 : pq_sendstring(&buf, payload);
3111 : 53 : pq_endmessage(&buf);
3112 : :
3113 : : /*
3114 : : * NOTE: we do not do pq_flush() here. Some level of caller will
3115 : : * handle it later, allowing this message to be combined into a packet
3116 : : * with other ones.
3117 : : */
3118 : : }
3119 : : else
3120 [ # # ]: 0 : elog(INFO, "NOTIFY for \"%s\" payload \"%s\"", channel, payload);
3121 : 53 : }
3122 : :
3123 : : /* Does pendingNotifies include a match for the given event? */
3124 : : static bool
3125 : 1066 : AsyncExistsPendingNotify(Notification *n)
3126 : : {
3127 [ - + ]: 1066 : if (pendingNotifies == NULL)
3128 : 0 : return false;
3129 : :
3130 [ + + ]: 1066 : if (pendingNotifies->hashtab != NULL)
3131 : : {
3132 : : /* Use the hash table to probe for a match */
3133 [ + + ]: 984 : if (hash_search(pendingNotifies->hashtab,
3134 : : &n,
3135 : : HASH_FIND,
3136 : : NULL))
3137 : 1 : return true;
3138 : : }
3139 : : else
3140 : : {
3141 : : /* Must scan the event list */
3142 : : ListCell *l;
3143 : :
3144 [ + - + + : 425 : foreach(l, pendingNotifies->events)
+ + ]
3145 : : {
3146 : 357 : Notification *oldn = (Notification *) lfirst(l);
3147 : :
3148 [ + - ]: 357 : if (n->channel_len == oldn->channel_len &&
3149 [ + + ]: 357 : n->payload_len == oldn->payload_len &&
3150 : 190 : memcmp(n->data, oldn->data,
3151 [ + + ]: 190 : n->channel_len + n->payload_len + 2) == 0)
3152 : 14 : return true;
3153 : : }
3154 : : }
3155 : :
3156 : 1051 : return false;
3157 : : }
3158 : :
3159 : : /*
3160 : : * Add a notification event to a pre-existing pendingNotifies list.
3161 : : *
3162 : : * Because pendingNotifies->events is already nonempty, this works
3163 : : * correctly no matter what CurrentMemoryContext is.
3164 : : */
3165 : : static void
3166 : 1051 : AddEventToPendingNotifies(Notification *n)
3167 : : {
3168 : : Assert(pendingNotifies->events != NIL);
3169 : :
3170 : : /* Create the hash tables if it's time to */
3171 [ + + ]: 1051 : if (list_length(pendingNotifies->events) >= MIN_HASHABLE_NOTIFIES &&
3172 [ + + ]: 985 : pendingNotifies->hashtab == NULL)
3173 : : {
3174 : : HASHCTL hash_ctl;
3175 : : ListCell *l;
3176 : :
3177 : : /* Create the hash table */
3178 : 2 : hash_ctl.keysize = sizeof(Notification *);
3179 : 2 : hash_ctl.entrysize = sizeof(struct NotificationHash);
3180 : 2 : hash_ctl.hash = notification_hash;
3181 : 2 : hash_ctl.match = notification_match;
3182 : 2 : hash_ctl.hcxt = CurTransactionContext;
3183 : 4 : pendingNotifies->hashtab =
3184 : 2 : hash_create("Pending Notifies",
3185 : : 256L,
3186 : : &hash_ctl,
3187 : : HASH_ELEM | HASH_FUNCTION | HASH_COMPARE | HASH_CONTEXT);
3188 : :
3189 : : /* Create the unique channel name table */
3190 : : Assert(pendingNotifies->uniqueChannelHash == NULL);
3191 : 2 : hash_ctl.keysize = NAMEDATALEN;
3192 : 2 : hash_ctl.entrysize = sizeof(ChannelName);
3193 : 2 : hash_ctl.hcxt = CurTransactionContext;
3194 : 4 : pendingNotifies->uniqueChannelHash =
3195 : 2 : hash_create("Pending Notify Channel Names",
3196 : : 64L,
3197 : : &hash_ctl,
3198 : : HASH_ELEM | HASH_STRINGS | HASH_CONTEXT);
3199 : :
3200 : : /* Insert all the already-existing events */
3201 [ + - + + : 34 : foreach(l, pendingNotifies->events)
+ + ]
3202 : : {
3203 : 32 : Notification *oldn = (Notification *) lfirst(l);
3204 : 32 : char *channel = oldn->data;
3205 : : bool found;
3206 : :
3207 : 32 : (void) hash_search(pendingNotifies->hashtab,
3208 : : &oldn,
3209 : : HASH_ENTER,
3210 : : &found);
3211 : : Assert(!found);
3212 : :
3213 : : /* Add channel name to uniqueChannelHash; might be there already */
3214 : 32 : (void) hash_search(pendingNotifies->uniqueChannelHash,
3215 : : channel,
3216 : : HASH_ENTER,
3217 : : NULL);
3218 : : }
3219 : : }
3220 : :
3221 : : /* Add new event to the list, in order */
3222 : 1051 : pendingNotifies->events = lappend(pendingNotifies->events, n);
3223 : :
3224 : : /* Add event to the hash tables if needed */
3225 [ + + ]: 1051 : if (pendingNotifies->hashtab != NULL)
3226 : : {
3227 : 985 : char *channel = n->data;
3228 : : bool found;
3229 : :
3230 : 985 : (void) hash_search(pendingNotifies->hashtab,
3231 : : &n,
3232 : : HASH_ENTER,
3233 : : &found);
3234 : : Assert(!found);
3235 : :
3236 : : /* Add channel name to uniqueChannelHash; might be there already */
3237 : 985 : (void) hash_search(pendingNotifies->uniqueChannelHash,
3238 : : channel,
3239 : : HASH_ENTER,
3240 : : NULL);
3241 : : }
3242 : 1051 : }
3243 : :
3244 : : /*
3245 : : * notification_hash: hash function for notification hash table
3246 : : *
3247 : : * The hash "keys" are pointers to Notification structs.
3248 : : */
3249 : : static uint32
3250 : 2001 : notification_hash(const void *key, Size keysize)
3251 : : {
3252 : 2001 : const Notification *k = *(const Notification *const *) key;
3253 : :
3254 : : Assert(keysize == sizeof(Notification *));
3255 : : /* We don't bother to include the payload's trailing null in the hash */
3256 : 4002 : return hash_bytes((const unsigned char *) k->data,
3257 : 2001 : k->channel_len + k->payload_len + 1);
3258 : : }
3259 : :
3260 : : /*
3261 : : * notification_match: match function to use with notification_hash
3262 : : */
3263 : : static int
3264 : 1 : notification_match(const void *key1, const void *key2, Size keysize)
3265 : : {
3266 : 1 : const Notification *k1 = *(const Notification *const *) key1;
3267 : 1 : const Notification *k2 = *(const Notification *const *) key2;
3268 : :
3269 : : Assert(keysize == sizeof(Notification *));
3270 [ + - ]: 1 : if (k1->channel_len == k2->channel_len &&
3271 [ + - ]: 1 : k1->payload_len == k2->payload_len &&
3272 : 1 : memcmp(k1->data, k2->data,
3273 [ + - ]: 1 : k1->channel_len + k1->payload_len + 2) == 0)
3274 : 1 : return 0; /* equal */
3275 : 0 : return 1; /* not equal */
3276 : : }
3277 : :
3278 : : /* Clear the pendingActions and pendingNotifies lists. */
3279 : : static void
3280 : 36228 : ClearPendingActionsAndNotifies(void)
3281 : : {
3282 : : /*
3283 : : * Everything's allocated in either TopTransactionContext or the context
3284 : : * for the subtransaction to which it corresponds. So, there's nothing to
3285 : : * do here except reset the pointers; the space will be reclaimed when the
3286 : : * contexts are deleted.
3287 : : */
3288 : 36228 : pendingActions = NULL;
3289 : 36228 : pendingNotifies = NULL;
3290 : : /* Also clear pendingListenActions, which is derived from pendingActions */
3291 : 36228 : pendingListenActions = NULL;
3292 : 36228 : }
3293 : :
3294 : : /*
3295 : : * GUC check_hook for notify_buffers
3296 : : */
3297 : : bool
3298 : 1376 : check_notify_buffers(int *newval, void **extra, GucSource source)
3299 : : {
3300 : 1376 : return check_slru_buffers("notify_buffers", newval);
3301 : : }
|
