Age Owner Branch data TLA Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : *
3 : : * async.c
4 : : * Asynchronous notification: NOTIFY, LISTEN, UNLISTEN
5 : : *
6 : : * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
7 : : * Portions Copyright (c) 1994, Regents of the University of California
8 : : *
9 : : * IDENTIFICATION
10 : : * src/backend/commands/async.c
11 : : *
12 : : *-------------------------------------------------------------------------
13 : : */
14 : :
15 : : /*-------------------------------------------------------------------------
16 : : * Async Notification Model as of 9.0:
17 : : *
18 : : * 1. Multiple backends on same machine. Multiple backends listening on
19 : : * several channels. (Channels are also called "conditions" in other
20 : : * parts of the code.)
21 : : *
22 : : * 2. There is one central queue in disk-based storage (directory pg_notify/),
23 : : * with actively-used pages mapped into shared memory by the slru.c module.
24 : : * All notification messages are placed in the queue and later read out
25 : : * by listening backends.
26 : : *
27 : : * There is no central knowledge of which backend listens on which channel;
28 : : * every backend has its own list of interesting channels.
29 : : *
30 : : * Although there is only one queue, notifications are treated as being
31 : : * database-local; this is done by including the sender's database OID
32 : : * in each notification message. Listening backends ignore messages
33 : : * that don't match their database OID. This is important because it
34 : : * ensures senders and receivers have the same database encoding and won't
35 : : * misinterpret non-ASCII text in the channel name or payload string.
36 : : *
37 : : * Since notifications are not expected to survive database crashes,
38 : : * we can simply clean out the pg_notify data at any reboot, and there
39 : : * is no need for WAL support or fsync'ing.
40 : : *
41 : : * 3. Every backend that is listening on at least one channel registers by
42 : : * entering its PID into the array in AsyncQueueControl. It then scans all
43 : : * incoming notifications in the central queue and first compares the
44 : : * database OID of the notification with its own database OID and then
45 : : * compares the notified channel with the list of channels that it listens
46 : : * to. In case there is a match it delivers the notification event to its
47 : : * frontend. Non-matching events are simply skipped.
48 : : *
49 : : * 4. The NOTIFY statement (routine Async_Notify) stores the notification in
50 : : * a backend-local list which will not be processed until transaction end.
51 : : *
52 : : * Duplicate notifications from the same transaction are sent out as one
53 : : * notification only. This is done to save work when for example a trigger
54 : : * on a 2 million row table fires a notification for each row that has been
55 : : * changed. If the application needs to receive every single notification
56 : : * that has been sent, it can easily add some unique string into the extra
57 : : * payload parameter.
58 : : *
59 : : * When the transaction is ready to commit, PreCommit_Notify() adds the
60 : : * pending notifications to the head of the queue. The head pointer of the
61 : : * queue always points to the next free position and a position is just a
62 : : * page number and the offset in that page. This is done before marking the
63 : : * transaction as committed in clog. If we run into problems writing the
64 : : * notifications, we can still call elog(ERROR, ...) and the transaction
65 : : * will roll back.
66 : : *
67 : : * Once we have put all of the notifications into the queue, we return to
68 : : * CommitTransaction() which will then do the actual transaction commit.
69 : : *
70 : : * After commit we are called another time (AtCommit_Notify()). Here we
71 : : * make any actual updates to the effective listen state (listenChannels).
72 : : * Then we signal any backends that may be interested in our messages
73 : : * (including our own backend, if listening). This is done by
74 : : * SignalBackends(), which scans the list of listening backends and sends a
75 : : * PROCSIG_NOTIFY_INTERRUPT signal to every listening backend (we don't
76 : : * know which backend is listening on which channel so we must signal them
77 : : * all). We can exclude backends that are already up to date, though, and
78 : : * we can also exclude backends that are in other databases (unless they
79 : : * are way behind and should be kicked to make them advance their
80 : : * pointers).
81 : : *
82 : : * Finally, after we are out of the transaction altogether and about to go
83 : : * idle, we scan the queue for messages that need to be sent to our
84 : : * frontend (which might be notifies from other backends, or self-notifies
85 : : * from our own). This step is not part of the CommitTransaction sequence
86 : : * for two important reasons. First, we could get errors while sending
87 : : * data to our frontend, and it's really bad for errors to happen in
88 : : * post-commit cleanup. Second, in cases where a procedure issues commits
89 : : * within a single frontend command, we don't want to send notifies to our
90 : : * frontend until the command is done; but notifies to other backends
91 : : * should go out immediately after each commit.
92 : : *
93 : : * 5. Upon receipt of a PROCSIG_NOTIFY_INTERRUPT signal, the signal handler
94 : : * sets the process's latch, which triggers the event to be processed
95 : : * immediately if this backend is idle (i.e., it is waiting for a frontend
96 : : * command and is not within a transaction block. C.f.
97 : : * ProcessClientReadInterrupt()). Otherwise the handler may only set a
98 : : * flag, which will cause the processing to occur just before we next go
99 : : * idle.
100 : : *
101 : : * Inbound-notify processing consists of reading all of the notifications
102 : : * that have arrived since scanning last time. We read every notification
103 : : * until we reach either a notification from an uncommitted transaction or
104 : : * the head pointer's position.
105 : : *
106 : : * 6. To limit disk space consumption, the tail pointer needs to be advanced
107 : : * so that old pages can be truncated. This is relatively expensive
108 : : * (notably, it requires an exclusive lock), so we don't want to do it
109 : : * often. We make sending backends do this work if they advanced the queue
110 : : * head into a new page, but only once every QUEUE_CLEANUP_DELAY pages.
111 : : *
112 : : * An application that listens on the same channel it notifies will get
113 : : * NOTIFY messages for its own NOTIFYs. These can be ignored, if not useful,
114 : : * by comparing be_pid in the NOTIFY message to the application's own backend's
115 : : * PID. (As of FE/BE protocol 2.0, the backend's PID is provided to the
116 : : * frontend during startup.) The above design guarantees that notifies from
117 : : * other backends will never be missed by ignoring self-notifies.
118 : : *
119 : : * The amount of shared memory used for notify management (notify_buffers)
120 : : * can be varied without affecting anything but performance. The maximum
121 : : * amount of notification data that can be queued at one time is determined
122 : : * by max_notify_queue_pages GUC.
123 : : *-------------------------------------------------------------------------
124 : : */
125 : :
126 : : #include "postgres.h"
127 : :
128 : : #include <limits.h>
129 : : #include <unistd.h>
130 : : #include <signal.h>
131 : :
132 : : #include "access/parallel.h"
133 : : #include "access/slru.h"
134 : : #include "access/transam.h"
135 : : #include "access/xact.h"
136 : : #include "catalog/pg_database.h"
137 : : #include "commands/async.h"
138 : : #include "common/hashfn.h"
139 : : #include "funcapi.h"
140 : : #include "libpq/libpq.h"
141 : : #include "libpq/pqformat.h"
142 : : #include "miscadmin.h"
143 : : #include "storage/ipc.h"
144 : : #include "storage/lmgr.h"
145 : : #include "storage/procsignal.h"
146 : : #include "tcop/tcopprot.h"
147 : : #include "utils/builtins.h"
148 : : #include "utils/guc_hooks.h"
149 : : #include "utils/memutils.h"
150 : : #include "utils/ps_status.h"
151 : : #include "utils/snapmgr.h"
152 : : #include "utils/timestamp.h"
153 : :
154 : :
155 : : /*
156 : : * Maximum size of a NOTIFY payload, including terminating NULL. This
157 : : * must be kept small enough so that a notification message fits on one
158 : : * SLRU page. The magic fudge factor here is noncritical as long as it's
159 : : * more than AsyncQueueEntryEmptySize --- we make it significantly bigger
160 : : * than that, so changes in that data structure won't affect user-visible
161 : : * restrictions.
162 : : */
163 : : #define NOTIFY_PAYLOAD_MAX_LENGTH (BLCKSZ - NAMEDATALEN - 128)
164 : :
165 : : /*
166 : : * Struct representing an entry in the global notify queue
167 : : *
168 : : * This struct declaration has the maximal length, but in a real queue entry
169 : : * the data area is only big enough for the actual channel and payload strings
170 : : * (each null-terminated). AsyncQueueEntryEmptySize is the minimum possible
171 : : * entry size, if both channel and payload strings are empty (but note it
172 : : * doesn't include alignment padding).
173 : : *
174 : : * The "length" field should always be rounded up to the next QUEUEALIGN
175 : : * multiple so that all fields are properly aligned.
176 : : */
177 : : typedef struct AsyncQueueEntry
178 : : {
179 : : int length; /* total allocated length of entry */
180 : : Oid dboid; /* sender's database OID */
181 : : TransactionId xid; /* sender's XID */
182 : : int32 srcPid; /* sender's PID */
183 : : char data[NAMEDATALEN + NOTIFY_PAYLOAD_MAX_LENGTH];
184 : : } AsyncQueueEntry;
185 : :
186 : : /* Currently, no field of AsyncQueueEntry requires more than int alignment */
187 : : #define QUEUEALIGN(len) INTALIGN(len)
188 : :
189 : : #define AsyncQueueEntryEmptySize (offsetof(AsyncQueueEntry, data) + 2)
190 : :
191 : : /*
192 : : * Struct describing a queue position, and assorted macros for working with it
193 : : */
194 : : typedef struct QueuePosition
195 : : {
196 : : int64 page; /* SLRU page number */
197 : : int offset; /* byte offset within page */
198 : : } QueuePosition;
199 : :
200 : : #define QUEUE_POS_PAGE(x) ((x).page)
201 : : #define QUEUE_POS_OFFSET(x) ((x).offset)
202 : :
203 : : #define SET_QUEUE_POS(x,y,z) \
204 : : do { \
205 : : (x).page = (y); \
206 : : (x).offset = (z); \
207 : : } while (0)
208 : :
209 : : #define QUEUE_POS_EQUAL(x,y) \
210 : : ((x).page == (y).page && (x).offset == (y).offset)
211 : :
212 : : #define QUEUE_POS_IS_ZERO(x) \
213 : : ((x).page == 0 && (x).offset == 0)
214 : :
215 : : /* choose logically smaller QueuePosition */
216 : : #define QUEUE_POS_MIN(x,y) \
217 : : (asyncQueuePagePrecedes((x).page, (y).page) ? (x) : \
218 : : (x).page != (y).page ? (y) : \
219 : : (x).offset < (y).offset ? (x) : (y))
220 : :
221 : : /* choose logically larger QueuePosition */
222 : : #define QUEUE_POS_MAX(x,y) \
223 : : (asyncQueuePagePrecedes((x).page, (y).page) ? (y) : \
224 : : (x).page != (y).page ? (x) : \
225 : : (x).offset > (y).offset ? (x) : (y))
226 : :
227 : : /*
228 : : * Parameter determining how often we try to advance the tail pointer:
229 : : * we do that after every QUEUE_CLEANUP_DELAY pages of NOTIFY data. This is
230 : : * also the distance by which a backend in another database needs to be
231 : : * behind before we'll decide we need to wake it up to advance its pointer.
232 : : *
233 : : * Resist the temptation to make this really large. While that would save
234 : : * work in some places, it would add cost in others. In particular, this
235 : : * should likely be less than notify_buffers, to ensure that backends
236 : : * catch up before the pages they'll need to read fall out of SLRU cache.
237 : : */
238 : : #define QUEUE_CLEANUP_DELAY 4
239 : :
240 : : /*
241 : : * Struct describing a listening backend's status
242 : : */
243 : : typedef struct QueueBackendStatus
244 : : {
245 : : int32 pid; /* either a PID or InvalidPid */
246 : : Oid dboid; /* backend's database OID, or InvalidOid */
247 : : ProcNumber nextListener; /* id of next listener, or INVALID_PROC_NUMBER */
248 : : QueuePosition pos; /* backend has read queue up to here */
249 : : } QueueBackendStatus;
250 : :
251 : : /*
252 : : * Shared memory state for LISTEN/NOTIFY (excluding its SLRU stuff)
253 : : *
254 : : * The AsyncQueueControl structure is protected by the NotifyQueueLock and
255 : : * NotifyQueueTailLock.
256 : : *
257 : : * When holding NotifyQueueLock in SHARED mode, backends may only inspect
258 : : * their own entries as well as the head and tail pointers. Consequently we
259 : : * can allow a backend to update its own record while holding only SHARED lock
260 : : * (since no other backend will inspect it).
261 : : *
262 : : * When holding NotifyQueueLock in EXCLUSIVE mode, backends can inspect the
263 : : * entries of other backends and also change the head pointer. When holding
264 : : * both NotifyQueueLock and NotifyQueueTailLock in EXCLUSIVE mode, backends
265 : : * can change the tail pointers.
266 : : *
267 : : * SLRU buffer pool is divided in banks and bank wise SLRU lock is used as
268 : : * the control lock for the pg_notify SLRU buffers.
269 : : * In order to avoid deadlocks, whenever we need multiple locks, we first get
270 : : * NotifyQueueTailLock, then NotifyQueueLock, and lastly SLRU bank lock.
271 : : *
272 : : * Each backend uses the backend[] array entry with index equal to its
273 : : * ProcNumber. We rely on this to make SendProcSignal fast.
274 : : *
275 : : * The backend[] array entries for actively-listening backends are threaded
276 : : * together using firstListener and the nextListener links, so that we can
277 : : * scan them without having to iterate over inactive entries. We keep this
278 : : * list in order by ProcNumber so that the scan is cache-friendly when there
279 : : * are many active entries.
280 : : */
281 : : typedef struct AsyncQueueControl
282 : : {
283 : : QueuePosition head; /* head points to the next free location */
284 : : QueuePosition tail; /* tail must be <= the queue position of every
285 : : * listening backend */
286 : : int stopPage; /* oldest unrecycled page; must be <=
287 : : * tail.page */
288 : : ProcNumber firstListener; /* id of first listener, or
289 : : * INVALID_PROC_NUMBER */
290 : : TimestampTz lastQueueFillWarn; /* time of last queue-full msg */
291 : : QueueBackendStatus backend[FLEXIBLE_ARRAY_MEMBER];
292 : : } AsyncQueueControl;
293 : :
294 : : static AsyncQueueControl *asyncQueueControl;
295 : :
296 : : #define QUEUE_HEAD (asyncQueueControl->head)
297 : : #define QUEUE_TAIL (asyncQueueControl->tail)
298 : : #define QUEUE_STOP_PAGE (asyncQueueControl->stopPage)
299 : : #define QUEUE_FIRST_LISTENER (asyncQueueControl->firstListener)
300 : : #define QUEUE_BACKEND_PID(i) (asyncQueueControl->backend[i].pid)
301 : : #define QUEUE_BACKEND_DBOID(i) (asyncQueueControl->backend[i].dboid)
302 : : #define QUEUE_NEXT_LISTENER(i) (asyncQueueControl->backend[i].nextListener)
303 : : #define QUEUE_BACKEND_POS(i) (asyncQueueControl->backend[i].pos)
304 : :
305 : : /*
306 : : * The SLRU buffer area through which we access the notification queue
307 : : */
308 : : static SlruCtlData NotifyCtlData;
309 : :
310 : : #define NotifyCtl (&NotifyCtlData)
311 : : #define QUEUE_PAGESIZE BLCKSZ
312 : :
313 : : #define QUEUE_FULL_WARN_INTERVAL 5000 /* warn at most once every 5s */
314 : :
315 : : /*
316 : : * listenChannels identifies the channels we are actually listening to
317 : : * (ie, have committed a LISTEN on). It is a simple list of channel names,
318 : : * allocated in TopMemoryContext.
319 : : */
320 : : static List *listenChannels = NIL; /* list of C strings */
321 : :
322 : : /*
323 : : * State for pending LISTEN/UNLISTEN actions consists of an ordered list of
324 : : * all actions requested in the current transaction. As explained above,
325 : : * we don't actually change listenChannels until we reach transaction commit.
326 : : *
327 : : * The list is kept in CurTransactionContext. In subtransactions, each
328 : : * subtransaction has its own list in its own CurTransactionContext, but
329 : : * successful subtransactions attach their lists to their parent's list.
330 : : * Failed subtransactions simply discard their lists.
331 : : */
332 : : typedef enum
333 : : {
334 : : LISTEN_LISTEN,
335 : : LISTEN_UNLISTEN,
336 : : LISTEN_UNLISTEN_ALL,
337 : : } ListenActionKind;
338 : :
339 : : typedef struct
340 : : {
341 : : ListenActionKind action;
342 : : char channel[FLEXIBLE_ARRAY_MEMBER]; /* nul-terminated string */
343 : : } ListenAction;
344 : :
345 : : typedef struct ActionList
346 : : {
347 : : int nestingLevel; /* current transaction nesting depth */
348 : : List *actions; /* list of ListenAction structs */
349 : : struct ActionList *upper; /* details for upper transaction levels */
350 : : } ActionList;
351 : :
352 : : static ActionList *pendingActions = NULL;
353 : :
354 : : /*
355 : : * State for outbound notifies consists of a list of all channels+payloads
356 : : * NOTIFYed in the current transaction. We do not actually perform a NOTIFY
357 : : * until and unless the transaction commits. pendingNotifies is NULL if no
358 : : * NOTIFYs have been done in the current (sub) transaction.
359 : : *
360 : : * We discard duplicate notify events issued in the same transaction.
361 : : * Hence, in addition to the list proper (which we need to track the order
362 : : * of the events, since we guarantee to deliver them in order), we build a
363 : : * hash table which we can probe to detect duplicates. Since building the
364 : : * hash table is somewhat expensive, we do so only once we have at least
365 : : * MIN_HASHABLE_NOTIFIES events queued in the current (sub) transaction;
366 : : * before that we just scan the events linearly.
367 : : *
368 : : * The list is kept in CurTransactionContext. In subtransactions, each
369 : : * subtransaction has its own list in its own CurTransactionContext, but
370 : : * successful subtransactions add their entries to their parent's list.
371 : : * Failed subtransactions simply discard their lists. Since these lists
372 : : * are independent, there may be notify events in a subtransaction's list
373 : : * that duplicate events in some ancestor (sub) transaction; we get rid of
374 : : * the dups when merging the subtransaction's list into its parent's.
375 : : *
376 : : * Note: the action and notify lists do not interact within a transaction.
377 : : * In particular, if a transaction does NOTIFY and then LISTEN on the same
378 : : * condition name, it will get a self-notify at commit. This is a bit odd
379 : : * but is consistent with our historical behavior.
380 : : */
381 : : typedef struct Notification
382 : : {
383 : : uint16 channel_len; /* length of channel-name string */
384 : : uint16 payload_len; /* length of payload string */
385 : : /* null-terminated channel name, then null-terminated payload follow */
386 : : char data[FLEXIBLE_ARRAY_MEMBER];
387 : : } Notification;
388 : :
389 : : typedef struct NotificationList
390 : : {
391 : : int nestingLevel; /* current transaction nesting depth */
392 : : List *events; /* list of Notification structs */
393 : : HTAB *hashtab; /* hash of NotificationHash structs, or NULL */
394 : : struct NotificationList *upper; /* details for upper transaction levels */
395 : : } NotificationList;
396 : :
397 : : #define MIN_HASHABLE_NOTIFIES 16 /* threshold to build hashtab */
398 : :
399 : : typedef struct NotificationHash
400 : : {
401 : : Notification *event; /* => the actual Notification struct */
402 : : } NotificationHash;
403 : :
404 : : static NotificationList *pendingNotifies = NULL;
405 : :
406 : : /*
407 : : * Inbound notifications are initially processed by HandleNotifyInterrupt(),
408 : : * called from inside a signal handler. That just sets the
409 : : * notifyInterruptPending flag and sets the process
410 : : * latch. ProcessNotifyInterrupt() will then be called whenever it's safe to
411 : : * actually deal with the interrupt.
412 : : */
413 : : volatile sig_atomic_t notifyInterruptPending = false;
414 : :
415 : : /* True if we've registered an on_shmem_exit cleanup */
416 : : static bool unlistenExitRegistered = false;
417 : :
418 : : /* True if we're currently registered as a listener in asyncQueueControl */
419 : : static bool amRegisteredListener = false;
420 : :
421 : : /* have we advanced to a page that's a multiple of QUEUE_CLEANUP_DELAY? */
422 : : static bool tryAdvanceTail = false;
423 : :
424 : : /* GUC parameters */
425 : : bool Trace_notify = false;
426 : :
427 : : /* For 8 KB pages this gives 8 GB of disk space */
428 : : int max_notify_queue_pages = 1048576;
429 : :
430 : : /* local function prototypes */
431 : : static inline int64 asyncQueuePageDiff(int64 p, int64 q);
432 : : static inline bool asyncQueuePagePrecedes(int64 p, int64 q);
433 : : static void queue_listen(ListenActionKind action, const char *channel);
434 : : static void Async_UnlistenOnExit(int code, Datum arg);
435 : : static void Exec_ListenPreCommit(void);
436 : : static void Exec_ListenCommit(const char *channel);
437 : : static void Exec_UnlistenCommit(const char *channel);
438 : : static void Exec_UnlistenAllCommit(void);
439 : : static bool IsListeningOn(const char *channel);
440 : : static void asyncQueueUnregister(void);
441 : : static bool asyncQueueIsFull(void);
442 : : static bool asyncQueueAdvance(volatile QueuePosition *position, int entryLength);
443 : : static void asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe);
444 : : static ListCell *asyncQueueAddEntries(ListCell *nextNotify);
445 : : static double asyncQueueUsage(void);
446 : : static void asyncQueueFillWarning(void);
447 : : static void SignalBackends(void);
448 : : static void asyncQueueReadAllNotifications(void);
449 : : static bool asyncQueueProcessPageEntries(volatile QueuePosition *current,
450 : : QueuePosition stop,
451 : : char *page_buffer,
452 : : Snapshot snapshot);
453 : : static void asyncQueueAdvanceTail(void);
454 : : static void ProcessIncomingNotify(bool flush);
455 : : static bool AsyncExistsPendingNotify(Notification *n);
456 : : static void AddEventToPendingNotifies(Notification *n);
457 : : static uint32 notification_hash(const void *key, Size keysize);
458 : : static int notification_match(const void *key1, const void *key2, Size keysize);
459 : : static void ClearPendingActionsAndNotifies(void);
460 : :
461 : : /*
462 : : * Compute the difference between two queue page numbers.
463 : : * Previously this function accounted for a wraparound.
464 : : */
465 : : static inline int64
137 akorotkov@postgresql 466 :UNC 0 : asyncQueuePageDiff(int64 p, int64 q)
467 : : {
468 : 0 : return p - q;
469 : : }
470 : :
471 : : /*
472 : : * Determines whether p precedes q.
473 : : * Previously this function accounted for a wraparound.
474 : : */
475 : : static inline bool
137 akorotkov@postgresql 476 :GNC 25 : asyncQueuePagePrecedes(int64 p, int64 q)
477 : : {
478 : 25 : return p < q;
479 : : }
480 : :
481 : : /*
482 : : * Report space needed for our shared memory area
483 : : */
484 : : Size
5171 tgl@sss.pgh.pa.us 485 :CBC 1679 : AsyncShmemSize(void)
486 : : {
487 : : Size size;
488 : :
489 : : /* This had better match AsyncShmemInit */
42 heikki.linnakangas@i 490 :GNC 1679 : size = mul_size(MaxBackends, sizeof(QueueBackendStatus));
3340 tgl@sss.pgh.pa.us 491 :CBC 1679 : size = add_size(size, offsetof(AsyncQueueControl, backend));
492 : :
46 alvherre@alvh.no-ip. 493 :GNC 1679 : size = add_size(size, SimpleLruShmemSize(notify_buffers, 0));
494 : :
5171 tgl@sss.pgh.pa.us 495 :CBC 1679 : return size;
496 : : }
497 : :
498 : : /*
499 : : * Initialize our shared memory area
500 : : */
501 : : void
502 : 898 : AsyncShmemInit(void)
503 : : {
504 : : bool found;
505 : : Size size;
506 : :
507 : : /*
508 : : * Create or attach to the AsyncQueueControl structure.
509 : : */
42 heikki.linnakangas@i 510 :GNC 898 : size = mul_size(MaxBackends, sizeof(QueueBackendStatus));
3340 tgl@sss.pgh.pa.us 511 :CBC 898 : size = add_size(size, offsetof(AsyncQueueControl, backend));
512 : :
5171 513 : 898 : asyncQueueControl = (AsyncQueueControl *)
514 : 898 : ShmemInitStruct("Async Queue Control", size, &found);
515 : :
516 [ + - ]: 898 : if (!found)
517 : : {
518 : : /* First time through, so initialize it */
519 : 898 : SET_QUEUE_POS(QUEUE_HEAD, 0, 0);
520 : 898 : SET_QUEUE_POS(QUEUE_TAIL, 0, 0);
1233 521 : 898 : QUEUE_STOP_PAGE = 0;
42 heikki.linnakangas@i 522 :GNC 898 : QUEUE_FIRST_LISTENER = INVALID_PROC_NUMBER;
5171 tgl@sss.pgh.pa.us 523 :CBC 898 : asyncQueueControl->lastQueueFillWarn = 0;
42 heikki.linnakangas@i 524 [ + + ]:GNC 75332 : for (int i = 0; i < MaxBackends; i++)
525 : : {
5171 tgl@sss.pgh.pa.us 526 :CBC 74434 : QUEUE_BACKEND_PID(i) = InvalidPid;
3119 527 : 74434 : QUEUE_BACKEND_DBOID(i) = InvalidOid;
42 heikki.linnakangas@i 528 :GNC 74434 : QUEUE_NEXT_LISTENER(i) = INVALID_PROC_NUMBER;
5171 tgl@sss.pgh.pa.us 529 :CBC 74434 : SET_QUEUE_POS(QUEUE_BACKEND_POS(i), 0, 0);
530 : : }
531 : : }
532 : :
533 : : /*
534 : : * Set up SLRU management of the pg_notify data. Note that long segment
535 : : * names are used in order to avoid wraparound.
536 : : */
1430 537 : 898 : NotifyCtl->PagePrecedes = asyncQueuePagePrecedes;
46 alvherre@alvh.no-ip. 538 :GNC 898 : SimpleLruInit(NotifyCtl, "notify", notify_buffers, 0,
539 : : "pg_notify", LWTRANCHE_NOTIFY_BUFFER, LWTRANCHE_NOTIFY_SLRU,
540 : : SYNC_HANDLER_NONE, true);
541 : :
5171 tgl@sss.pgh.pa.us 542 [ + - ]:CBC 898 : if (!found)
543 : : {
544 : : /*
545 : : * During start or reboot, clean out the pg_notify directory.
546 : : */
1430 547 : 898 : (void) SlruScanDirectory(NotifyCtl, SlruScanDirCbDeleteAll, NULL);
548 : : }
5171 549 : 898 : }
550 : :
551 : :
552 : : /*
553 : : * pg_notify -
554 : : * SQL function to send a notification event
555 : : */
556 : : Datum
557 : 1054 : pg_notify(PG_FUNCTION_ARGS)
558 : : {
559 : : const char *channel;
560 : : const char *payload;
561 : :
562 [ + + ]: 1054 : if (PG_ARGISNULL(0))
563 : 3 : channel = "";
564 : : else
565 : 1051 : channel = text_to_cstring(PG_GETARG_TEXT_PP(0));
566 : :
567 [ + + ]: 1054 : if (PG_ARGISNULL(1))
568 : 6 : payload = "";
569 : : else
570 : 1048 : payload = text_to_cstring(PG_GETARG_TEXT_PP(1));
571 : :
572 : : /* For NOTIFY as a statement, this is checked in ProcessUtility */
5167 573 : 1054 : PreventCommandDuringRecovery("NOTIFY");
574 : :
5171 575 : 1054 : Async_Notify(channel, payload);
576 : :
577 : 1045 : PG_RETURN_VOID();
578 : : }
579 : :
580 : :
581 : : /*
582 : : * Async_Notify
583 : : *
584 : : * This is executed by the SQL notify command.
585 : : *
586 : : * Adds the message to the list of pending notifies.
587 : : * Actual notification happens during transaction commit.
588 : : * ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
589 : : */
590 : : void
591 : 1103 : Async_Notify(const char *channel, const char *payload)
592 : : {
1654 rhaas@postgresql.org 593 : 1103 : int my_level = GetCurrentTransactionNestLevel();
594 : : size_t channel_len;
595 : : size_t payload_len;
596 : : Notification *n;
597 : : MemoryContext oldcontext;
598 : :
3097 599 [ - + ]: 1103 : if (IsParallelWorker())
3097 rhaas@postgresql.org 600 [ # # ]:UBC 0 : elog(ERROR, "cannot send notifications from a parallel worker");
601 : :
8719 peter_e@gmx.net 602 [ - + ]:CBC 1103 : if (Trace_notify)
5171 tgl@sss.pgh.pa.us 603 [ # # ]:UBC 0 : elog(DEBUG1, "Async_Notify(%s)", channel);
604 : :
1704 tgl@sss.pgh.pa.us 605 [ + - ]:CBC 1103 : channel_len = channel ? strlen(channel) : 0;
606 [ + + ]: 1103 : payload_len = payload ? strlen(payload) : 0;
607 : :
608 : : /* a channel name must be specified */
609 [ + + ]: 1103 : if (channel_len == 0)
5171 610 [ + - ]: 6 : ereport(ERROR,
611 : : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
612 : : errmsg("channel name cannot be empty")));
613 : :
614 : : /* enforce length limits */
1704 615 [ + + ]: 1097 : if (channel_len >= NAMEDATALEN)
5171 616 [ + - ]: 3 : ereport(ERROR,
617 : : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
618 : : errmsg("channel name too long")));
619 : :
1704 620 [ - + ]: 1094 : if (payload_len >= NOTIFY_PAYLOAD_MAX_LENGTH)
1704 tgl@sss.pgh.pa.us 621 [ # # ]:UBC 0 : ereport(ERROR,
622 : : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
623 : : errmsg("payload string too long")));
624 : :
625 : : /*
626 : : * We must construct the Notification entry, even if we end up not using
627 : : * it, in order to compare it cheaply to existing list entries.
628 : : *
629 : : * The notification list needs to live until end of transaction, so store
630 : : * it in the transaction context.
631 : : */
5171 tgl@sss.pgh.pa.us 632 :CBC 1094 : oldcontext = MemoryContextSwitchTo(CurTransactionContext);
633 : :
1704 634 : 1094 : n = (Notification *) palloc(offsetof(Notification, data) +
635 : 1094 : channel_len + payload_len + 2);
636 : 1094 : n->channel_len = channel_len;
637 : 1094 : n->payload_len = payload_len;
638 : 1094 : strcpy(n->data, channel);
5171 639 [ + + ]: 1094 : if (payload)
1704 640 : 1081 : strcpy(n->data + channel_len + 1, payload);
641 : : else
642 : 13 : n->data[channel_len + 1] = '\0';
643 : :
1654 rhaas@postgresql.org 644 [ + + + + ]: 1094 : if (pendingNotifies == NULL || my_level > pendingNotifies->nestingLevel)
1704 tgl@sss.pgh.pa.us 645 : 50 : {
646 : : NotificationList *notifies;
647 : :
648 : : /*
649 : : * First notify event in current (sub)xact. Note that we allocate the
650 : : * NotificationList in TopTransactionContext; the nestingLevel might
651 : : * get changed later by AtSubCommit_Notify.
652 : : */
653 : : notifies = (NotificationList *)
1654 rhaas@postgresql.org 654 : 50 : MemoryContextAlloc(TopTransactionContext,
655 : : sizeof(NotificationList));
656 : 50 : notifies->nestingLevel = my_level;
657 : 50 : notifies->events = list_make1(n);
658 : : /* We certainly don't need a hashtable yet */
659 : 50 : notifies->hashtab = NULL;
660 : 50 : notifies->upper = pendingNotifies;
661 : 50 : pendingNotifies = notifies;
662 : : }
663 : : else
664 : : {
665 : : /* Now check for duplicates */
666 [ + + ]: 1044 : if (AsyncExistsPendingNotify(n))
667 : : {
668 : : /* It's a dup, so forget it */
669 : 12 : pfree(n);
670 : 12 : MemoryContextSwitchTo(oldcontext);
671 : 12 : return;
672 : : }
673 : :
674 : : /* Append more events to existing list */
1704 tgl@sss.pgh.pa.us 675 : 1032 : AddEventToPendingNotifies(n);
676 : : }
677 : :
5171 678 : 1082 : MemoryContextSwitchTo(oldcontext);
679 : : }
680 : :
681 : : /*
682 : : * queue_listen
683 : : * Common code for listen, unlisten, unlisten all commands.
684 : : *
685 : : * Adds the request to the list of pending actions.
686 : : * Actual update of the listenChannels list happens during transaction
687 : : * commit.
688 : : */
689 : : static void
690 : 57 : queue_listen(ListenActionKind action, const char *channel)
691 : : {
692 : : MemoryContext oldcontext;
693 : : ListenAction *actrec;
1654 rhaas@postgresql.org 694 : 57 : int my_level = GetCurrentTransactionNestLevel();
695 : :
696 : : /*
697 : : * Unlike Async_Notify, we don't try to collapse out duplicates. It would
698 : : * be too complicated to ensure we get the right interactions of
699 : : * conflicting LISTEN/UNLISTEN/UNLISTEN_ALL, and it's unlikely that there
700 : : * would be any performance benefit anyway in sane applications.
701 : : */
5877 tgl@sss.pgh.pa.us 702 : 57 : oldcontext = MemoryContextSwitchTo(CurTransactionContext);
703 : :
704 : : /* space for terminating null is included in sizeof(ListenAction) */
3340 705 : 57 : actrec = (ListenAction *) palloc(offsetof(ListenAction, channel) +
706 : 57 : strlen(channel) + 1);
5877 707 : 57 : actrec->action = action;
5171 708 : 57 : strcpy(actrec->channel, channel);
709 : :
1654 rhaas@postgresql.org 710 [ + + - + ]: 57 : if (pendingActions == NULL || my_level > pendingActions->nestingLevel)
711 : 50 : {
712 : : ActionList *actions;
713 : :
714 : : /*
715 : : * First action in current sub(xact). Note that we allocate the
716 : : * ActionList in TopTransactionContext; the nestingLevel might get
717 : : * changed later by AtSubCommit_Notify.
718 : : */
719 : : actions = (ActionList *)
720 : 50 : MemoryContextAlloc(TopTransactionContext, sizeof(ActionList));
721 : 50 : actions->nestingLevel = my_level;
722 : 50 : actions->actions = list_make1(actrec);
723 : 50 : actions->upper = pendingActions;
724 : 50 : pendingActions = actions;
725 : : }
726 : : else
727 : 7 : pendingActions->actions = lappend(pendingActions->actions, actrec);
728 : :
5877 tgl@sss.pgh.pa.us 729 : 57 : MemoryContextSwitchTo(oldcontext);
730 : 57 : }
731 : :
732 : : /*
733 : : * Async_Listen
734 : : *
735 : : * This is executed by the SQL listen command.
736 : : */
737 : : void
5171 738 : 37 : Async_Listen(const char *channel)
739 : : {
5877 740 [ - + ]: 37 : if (Trace_notify)
5171 tgl@sss.pgh.pa.us 741 [ # # ]:UBC 0 : elog(DEBUG1, "Async_Listen(%s,%d)", channel, MyProcPid);
742 : :
5171 tgl@sss.pgh.pa.us 743 :CBC 37 : queue_listen(LISTEN_LISTEN, channel);
5877 744 : 37 : }
745 : :
746 : : /*
747 : : * Async_Unlisten
748 : : *
749 : : * This is executed by the SQL unlisten command.
750 : : */
751 : : void
5171 752 : 3 : Async_Unlisten(const char *channel)
753 : : {
5706 754 [ - + ]: 3 : if (Trace_notify)
5171 tgl@sss.pgh.pa.us 755 [ # # ]:UBC 0 : elog(DEBUG1, "Async_Unlisten(%s,%d)", channel, MyProcPid);
756 : :
757 : : /* If we couldn't possibly be listening, no need to queue anything */
1654 rhaas@postgresql.org 758 [ + - - + ]:CBC 3 : if (pendingActions == NULL && !unlistenExitRegistered)
5539 tgl@sss.pgh.pa.us 759 :UBC 0 : return;
760 : :
5171 tgl@sss.pgh.pa.us 761 :CBC 3 : queue_listen(LISTEN_UNLISTEN, channel);
762 : : }
763 : :
764 : : /*
765 : : * Async_UnlistenAll
766 : : *
767 : : * This is invoked by UNLISTEN * command, and also at backend exit.
768 : : */
769 : : void
5877 770 : 19 : Async_UnlistenAll(void)
771 : : {
772 [ - + ]: 19 : if (Trace_notify)
5877 tgl@sss.pgh.pa.us 773 [ # # ]:UBC 0 : elog(DEBUG1, "Async_UnlistenAll(%d)", MyProcPid);
774 : :
775 : : /* If we couldn't possibly be listening, no need to queue anything */
1654 rhaas@postgresql.org 776 [ + - + + ]:CBC 19 : if (pendingActions == NULL && !unlistenExitRegistered)
5539 tgl@sss.pgh.pa.us 777 : 2 : return;
778 : :
5877 779 : 17 : queue_listen(LISTEN_UNLISTEN_ALL, "");
780 : : }
781 : :
782 : : /*
783 : : * SQL function: return a set of the channel names this backend is actively
784 : : * listening to.
785 : : *
786 : : * Note: this coding relies on the fact that the listenChannels list cannot
787 : : * change within a transaction.
788 : : */
789 : : Datum
5171 790 : 9 : pg_listening_channels(PG_FUNCTION_ARGS)
791 : : {
792 : : FuncCallContext *funcctx;
793 : :
794 : : /* stuff done only on the first call of the function */
795 [ + + ]: 9 : if (SRF_IS_FIRSTCALL())
796 : : {
797 : : /* create a function context for cross-call persistence */
798 : 6 : funcctx = SRF_FIRSTCALL_INIT();
799 : : }
800 : :
801 : : /* stuff done on every call of the function */
802 : 9 : funcctx = SRF_PERCALL_SETUP();
803 : :
1735 804 [ + + ]: 9 : if (funcctx->call_cntr < list_length(listenChannels))
805 : : {
806 : 3 : char *channel = (char *) list_nth(listenChannels,
807 : 3 : funcctx->call_cntr);
808 : :
5171 809 : 3 : SRF_RETURN_NEXT(funcctx, CStringGetTextDatum(channel));
810 : : }
811 : :
812 : 6 : SRF_RETURN_DONE(funcctx);
813 : : }
814 : :
815 : : /*
816 : : * Async_UnlistenOnExit
817 : : *
818 : : * This is executed at backend exit if we have done any LISTENs in this
819 : : * backend. It might not be necessary anymore, if the user UNLISTENed
820 : : * everything, but we don't try to detect that case.
821 : : */
822 : : static void
5877 823 : 14 : Async_UnlistenOnExit(int code, Datum arg)
824 : : {
5171 825 : 14 : Exec_UnlistenAllCommit();
4078 826 : 14 : asyncQueueUnregister();
5877 827 : 14 : }
828 : :
829 : : /*
830 : : * AtPrepare_Notify
831 : : *
832 : : * This is called at the prepare phase of a two-phase
833 : : * transaction. Save the state for possible commit later.
834 : : */
835 : : void
836 : 393 : AtPrepare_Notify(void)
837 : : {
838 : : /* It's not allowed to have any pending LISTEN/UNLISTEN/NOTIFY actions */
5171 839 [ + - - + ]: 393 : if (pendingActions || pendingNotifies)
5877 tgl@sss.pgh.pa.us 840 [ # # ]:UBC 0 : ereport(ERROR,
841 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
842 : : errmsg("cannot PREPARE a transaction that has executed LISTEN, UNLISTEN, or NOTIFY")));
5171 tgl@sss.pgh.pa.us 843 :CBC 393 : }
844 : :
845 : : /*
846 : : * PreCommit_Notify
847 : : *
848 : : * This is called at transaction commit, before actually committing to
849 : : * clog.
850 : : *
851 : : * If there are pending LISTEN actions, make sure we are listed in the
852 : : * shared-memory listener array. This must happen before commit to
853 : : * ensure we don't miss any notifies from transactions that commit
854 : : * just after ours.
855 : : *
856 : : * If there are outbound notify requests in the pendingNotifies list,
857 : : * add them to the global queue. We do that before commit so that
858 : : * we can still throw error if we run out of queue space.
859 : : */
860 : : void
861 : 408453 : PreCommit_Notify(void)
862 : : {
863 : : ListCell *p;
864 : :
1704 865 [ + + + + ]: 408453 : if (!pendingActions && !pendingNotifies)
5171 866 : 408356 : return; /* no relevant statements in this xact */
867 : :
868 [ - + ]: 97 : if (Trace_notify)
5171 tgl@sss.pgh.pa.us 869 [ # # ]:UBC 0 : elog(DEBUG1, "PreCommit_Notify");
870 : :
871 : : /* Preflight for any pending listen/unlisten actions */
1654 rhaas@postgresql.org 872 [ + + ]:CBC 97 : if (pendingActions != NULL)
873 : : {
874 [ + - + + : 105 : foreach(p, pendingActions->actions)
+ + ]
875 : : {
876 : 56 : ListenAction *actrec = (ListenAction *) lfirst(p);
877 : :
878 [ + + + - ]: 56 : switch (actrec->action)
879 : : {
880 : 37 : case LISTEN_LISTEN:
881 : 37 : Exec_ListenPreCommit();
882 : 37 : break;
883 : 3 : case LISTEN_UNLISTEN:
884 : : /* there is no Exec_UnlistenPreCommit() */
885 : 3 : break;
886 : 16 : case LISTEN_UNLISTEN_ALL:
887 : : /* there is no Exec_UnlistenAllPreCommit() */
888 : 16 : break;
889 : : }
890 : : }
891 : : }
892 : :
893 : : /* Queue any pending notifies (must happen after the above) */
5171 tgl@sss.pgh.pa.us 894 [ + + ]: 97 : if (pendingNotifies)
895 : : {
896 : : ListCell *nextNotify;
897 : :
898 : : /*
899 : : * Make sure that we have an XID assigned to the current transaction.
900 : : * GetCurrentTransactionId is cheap if we already have an XID, but not
901 : : * so cheap if we don't, and we'd prefer not to do that work while
902 : : * holding NotifyQueueLock.
903 : : */
904 : 48 : (void) GetCurrentTransactionId();
905 : :
906 : : /*
907 : : * Serialize writers by acquiring a special lock that we hold till
908 : : * after commit. This ensures that queue entries appear in commit
909 : : * order, and in particular that there are never uncommitted queue
910 : : * entries ahead of committed ones, so an uncommitted transaction
911 : : * can't block delivery of deliverable notifications.
912 : : *
913 : : * We use a heavyweight lock so that it'll automatically be released
914 : : * after either commit or abort. This also allows deadlocks to be
915 : : * detected, though really a deadlock shouldn't be possible here.
916 : : *
917 : : * The lock is on "database 0", which is pretty ugly but it doesn't
918 : : * seem worth inventing a special locktag category just for this.
919 : : * (Historical note: before PG 9.0, a similar lock on "database 0" was
920 : : * used by the flatfiles mechanism.)
921 : : */
922 : 48 : LockSharedObject(DatabaseRelationId, InvalidOid, 0,
923 : : AccessExclusiveLock);
924 : :
925 : : /* Now push the notifications into the queue */
1704 926 : 48 : nextNotify = list_head(pendingNotifies->events);
5171 927 [ + + ]: 131 : while (nextNotify != NULL)
928 : : {
929 : : /*
930 : : * Add the pending notifications to the queue. We acquire and
931 : : * release NotifyQueueLock once per page, which might be overkill
932 : : * but it does allow readers to get in while we're doing this.
933 : : *
934 : : * A full queue is very uncommon and should really not happen,
935 : : * given that we have so much space available in the SLRU pages.
936 : : * Nevertheless we need to deal with this possibility. Note that
937 : : * when we get here we are in the process of committing our
938 : : * transaction, but we have not yet committed to clog, so at this
939 : : * point in time we can still roll the transaction back.
940 : : */
1430 941 : 83 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
5171 942 : 83 : asyncQueueFillWarning();
943 [ - + ]: 83 : if (asyncQueueIsFull())
5171 tgl@sss.pgh.pa.us 944 [ # # ]:UBC 0 : ereport(ERROR,
945 : : (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
946 : : errmsg("too many notifications in the NOTIFY queue")));
5171 tgl@sss.pgh.pa.us 947 :CBC 83 : nextNotify = asyncQueueAddEntries(nextNotify);
1430 948 : 83 : LWLockRelease(NotifyQueueLock);
949 : : }
950 : :
951 : : /* Note that we don't clear pendingNotifies; AtCommit_Notify will. */
952 : : }
953 : : }
954 : :
955 : : /*
956 : : * AtCommit_Notify
957 : : *
958 : : * This is called at transaction commit, after committing to clog.
959 : : *
960 : : * Update listenChannels and clear transaction-local state.
961 : : *
962 : : * If we issued any notifications in the transaction, send signals to
963 : : * listening backends (possibly including ourselves) to process them.
964 : : * Also, if we filled enough queue pages with new notifies, try to
965 : : * advance the queue tail pointer.
966 : : */
967 : : void
5877 968 : 408298 : AtCommit_Notify(void)
969 : : {
970 : : ListCell *p;
971 : :
972 : : /*
973 : : * Allow transactions that have not executed LISTEN/UNLISTEN/NOTIFY to
974 : : * return as soon as possible
975 : : */
5171 976 [ + + + + ]: 408298 : if (!pendingActions && !pendingNotifies)
5877 977 : 408201 : return;
978 : :
979 [ - + ]: 97 : if (Trace_notify)
5877 tgl@sss.pgh.pa.us 980 [ # # ]:UBC 0 : elog(DEBUG1, "AtCommit_Notify");
981 : :
982 : : /* Perform any pending listen/unlisten actions */
1654 rhaas@postgresql.org 983 [ + + ]:CBC 97 : if (pendingActions != NULL)
984 : : {
985 [ + - + + : 105 : foreach(p, pendingActions->actions)
+ + ]
986 : : {
987 : 56 : ListenAction *actrec = (ListenAction *) lfirst(p);
988 : :
989 [ + + + - ]: 56 : switch (actrec->action)
990 : : {
991 : 37 : case LISTEN_LISTEN:
992 : 37 : Exec_ListenCommit(actrec->channel);
993 : 37 : break;
994 : 3 : case LISTEN_UNLISTEN:
995 : 3 : Exec_UnlistenCommit(actrec->channel);
996 : 3 : break;
997 : 16 : case LISTEN_UNLISTEN_ALL:
998 : 16 : Exec_UnlistenAllCommit();
999 : 16 : break;
1000 : : }
1001 : : }
1002 : : }
1003 : :
1004 : : /* If no longer listening to anything, get out of listener array */
4078 tgl@sss.pgh.pa.us 1005 [ + + + + ]: 97 : if (amRegisteredListener && listenChannels == NIL)
1006 : 13 : asyncQueueUnregister();
1007 : :
1008 : : /*
1009 : : * Send signals to listening backends. We need do this only if there are
1010 : : * pending notifies, which were previously added to the shared queue by
1011 : : * PreCommit_Notify().
1012 : : */
943 1013 [ + + ]: 97 : if (pendingNotifies != NULL)
1014 : 48 : SignalBackends();
1015 : :
1016 : : /*
1017 : : * If it's time to try to advance the global tail pointer, do that.
1018 : : *
1019 : : * (It might seem odd to do this in the sender, when more than likely the
1020 : : * listeners won't yet have read the messages we just sent. However,
1021 : : * there's less contention if only the sender does it, and there is little
1022 : : * need for urgency in advancing the global tail. So this typically will
1023 : : * be clearing out messages that were sent some time ago.)
1024 : : */
1025 [ + + ]: 97 : if (tryAdvanceTail)
1026 : : {
1027 : 8 : tryAdvanceTail = false;
1028 : 8 : asyncQueueAdvanceTail();
1029 : : }
1030 : :
1031 : : /* And clean up */
5171 1032 : 97 : ClearPendingActionsAndNotifies();
1033 : : }
1034 : :
1035 : : /*
1036 : : * Exec_ListenPreCommit --- subroutine for PreCommit_Notify
1037 : : *
1038 : : * This function must make sure we are ready to catch any incoming messages.
1039 : : */
1040 : : static void
1041 : 37 : Exec_ListenPreCommit(void)
1042 : : {
1043 : : QueuePosition head;
1044 : : QueuePosition max;
1045 : : ProcNumber prevListener;
1046 : :
1047 : : /*
1048 : : * Nothing to do if we are already listening to something, nor if we
1049 : : * already ran this routine in this transaction.
1050 : : */
4078 1051 [ + + ]: 37 : if (amRegisteredListener)
5171 1052 : 18 : return;
1053 : :
1054 [ - + ]: 19 : if (Trace_notify)
5171 tgl@sss.pgh.pa.us 1055 [ # # ]:UBC 0 : elog(DEBUG1, "Exec_ListenPreCommit(%d)", MyProcPid);
1056 : :
1057 : : /*
1058 : : * Before registering, make sure we will unlisten before dying. (Note:
1059 : : * this action does not get undone if we abort later.)
1060 : : */
5171 tgl@sss.pgh.pa.us 1061 [ + + ]:CBC 19 : if (!unlistenExitRegistered)
1062 : : {
3770 rhaas@postgresql.org 1063 : 14 : before_shmem_exit(Async_UnlistenOnExit, 0);
5171 tgl@sss.pgh.pa.us 1064 : 14 : unlistenExitRegistered = true;
1065 : : }
1066 : :
1067 : : /*
1068 : : * This is our first LISTEN, so establish our pointer.
1069 : : *
1070 : : * We set our pointer to the global tail pointer and then move it forward
1071 : : * over already-committed notifications. This ensures we cannot miss any
1072 : : * not-yet-committed notifications. We might get a few more but that
1073 : : * doesn't hurt.
1074 : : *
1075 : : * In some scenarios there might be a lot of committed notifications that
1076 : : * have not yet been pruned away (because some backend is being lazy about
1077 : : * reading them). To reduce our startup time, we can look at other
1078 : : * backends and adopt the maximum "pos" pointer of any backend that's in
1079 : : * our database; any notifications it's already advanced over are surely
1080 : : * committed and need not be re-examined by us. (We must consider only
1081 : : * backends connected to our DB, because others will not have bothered to
1082 : : * check committed-ness of notifications in our DB.)
1083 : : *
1084 : : * We need exclusive lock here so we can look at other backends' entries
1085 : : * and manipulate the list links.
1086 : : */
1430 1087 : 19 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
3119 1088 : 19 : head = QUEUE_HEAD;
1089 : 19 : max = QUEUE_TAIL;
42 heikki.linnakangas@i 1090 :GNC 19 : prevListener = INVALID_PROC_NUMBER;
1091 [ + + ]: 21 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
1092 : : {
1678 tgl@sss.pgh.pa.us 1093 [ + - ]:CBC 2 : if (QUEUE_BACKEND_DBOID(i) == MyDatabaseId)
1094 [ - + + - : 2 : max = QUEUE_POS_MAX(max, QUEUE_BACKEND_POS(i));
+ - ]
1095 : : /* Also find last listening backend before this one */
42 heikki.linnakangas@i 1096 [ - + ]:GNC 2 : if (i < MyProcNumber)
1678 tgl@sss.pgh.pa.us 1097 :LBC (1) : prevListener = i;
1098 : : }
42 heikki.linnakangas@i 1099 :GNC 19 : QUEUE_BACKEND_POS(MyProcNumber) = max;
1100 : 19 : QUEUE_BACKEND_PID(MyProcNumber) = MyProcPid;
1101 : 19 : QUEUE_BACKEND_DBOID(MyProcNumber) = MyDatabaseId;
1102 : : /* Insert backend into list of listeners at correct position */
1103 [ - + ]: 19 : if (prevListener != INVALID_PROC_NUMBER)
1104 : : {
42 heikki.linnakangas@i 1105 :UNC 0 : QUEUE_NEXT_LISTENER(MyProcNumber) = QUEUE_NEXT_LISTENER(prevListener);
1106 : 0 : QUEUE_NEXT_LISTENER(prevListener) = MyProcNumber;
1107 : : }
1108 : : else
1109 : : {
42 heikki.linnakangas@i 1110 :GNC 19 : QUEUE_NEXT_LISTENER(MyProcNumber) = QUEUE_FIRST_LISTENER;
1111 : 19 : QUEUE_FIRST_LISTENER = MyProcNumber;
1112 : : }
1430 tgl@sss.pgh.pa.us 1113 :CBC 19 : LWLockRelease(NotifyQueueLock);
1114 : :
1115 : : /* Now we are listed in the global array, so remember we're listening */
4078 1116 : 19 : amRegisteredListener = true;
1117 : :
1118 : : /*
1119 : : * Try to move our pointer forward as far as possible. This will skip
1120 : : * over already-committed notifications, which we want to do because they
1121 : : * might be quite stale. Note that we are not yet listening on anything,
1122 : : * so we won't deliver such notifications to our frontend. Also, although
1123 : : * our transaction might have executed NOTIFY, those message(s) aren't
1124 : : * queued yet so we won't skip them here.
1125 : : */
3119 1126 [ + + + + ]: 19 : if (!QUEUE_POS_EQUAL(max, head))
1127 : 11 : asyncQueueReadAllNotifications();
1128 : : }
1129 : :
1130 : : /*
1131 : : * Exec_ListenCommit --- subroutine for AtCommit_Notify
1132 : : *
1133 : : * Add the channel to the list of channels we are listening on.
1134 : : */
1135 : : static void
5171 1136 : 37 : Exec_ListenCommit(const char *channel)
1137 : : {
1138 : : MemoryContext oldcontext;
1139 : :
1140 : : /* Do nothing if we are already listening on this channel */
1141 [ + + ]: 37 : if (IsListeningOn(channel))
1142 : 10 : return;
1143 : :
1144 : : /*
1145 : : * Add the new channel name to listenChannels.
1146 : : *
1147 : : * XXX It is theoretically possible to get an out-of-memory failure here,
1148 : : * which would be bad because we already committed. For the moment it
1149 : : * doesn't seem worth trying to guard against that, but maybe improve this
1150 : : * later.
1151 : : */
1152 : 27 : oldcontext = MemoryContextSwitchTo(TopMemoryContext);
1153 : 27 : listenChannels = lappend(listenChannels, pstrdup(channel));
1154 : 27 : MemoryContextSwitchTo(oldcontext);
1155 : : }
1156 : :
1157 : : /*
1158 : : * Exec_UnlistenCommit --- subroutine for AtCommit_Notify
1159 : : *
1160 : : * Remove the specified channel name from listenChannels.
1161 : : */
1162 : : static void
1163 : 3 : Exec_UnlistenCommit(const char *channel)
1164 : : {
1165 : : ListCell *q;
1166 : :
1167 [ - + ]: 3 : if (Trace_notify)
5171 tgl@sss.pgh.pa.us 1168 [ # # ]:UBC 0 : elog(DEBUG1, "Exec_UnlistenCommit(%s,%d)", channel, MyProcPid);
1169 : :
5171 tgl@sss.pgh.pa.us 1170 [ + - + - :CBC 3 : foreach(q, listenChannels)
+ - ]
1171 : : {
5161 bruce@momjian.us 1172 : 3 : char *lchan = (char *) lfirst(q);
1173 : :
5171 tgl@sss.pgh.pa.us 1174 [ + - ]: 3 : if (strcmp(lchan, channel) == 0)
1175 : : {
1735 1176 : 3 : listenChannels = foreach_delete_current(listenChannels, q);
5171 1177 : 3 : pfree(lchan);
1178 : 3 : break;
1179 : : }
1180 : : }
1181 : :
1182 : : /*
1183 : : * We do not complain about unlistening something not being listened;
1184 : : * should we?
1185 : : */
1186 : 3 : }
1187 : :
1188 : : /*
1189 : : * Exec_UnlistenAllCommit --- subroutine for AtCommit_Notify
1190 : : *
1191 : : * Unlisten on all channels for this backend.
1192 : : */
1193 : : static void
1194 : 30 : Exec_UnlistenAllCommit(void)
1195 : : {
1196 [ - + ]: 30 : if (Trace_notify)
5171 tgl@sss.pgh.pa.us 1197 [ # # ]:UBC 0 : elog(DEBUG1, "Exec_UnlistenAllCommit(%d)", MyProcPid);
1198 : :
5171 tgl@sss.pgh.pa.us 1199 :CBC 30 : list_free_deep(listenChannels);
1200 : 30 : listenChannels = NIL;
1201 : 30 : }
1202 : :
1203 : : /*
1204 : : * Test whether we are actively listening on the given channel name.
1205 : : *
1206 : : * Note: this function is executed for every notification found in the queue.
1207 : : * Perhaps it is worth further optimization, eg convert the list to a sorted
1208 : : * array so we can binary-search it. In practice the list is likely to be
1209 : : * fairly short, though.
1210 : : */
1211 : : static bool
1212 : 1196 : IsListeningOn(const char *channel)
1213 : : {
1214 : : ListCell *p;
1215 : :
1216 [ + + + + : 1257 : foreach(p, listenChannels)
+ + ]
1217 : : {
1218 : 102 : char *lchan = (char *) lfirst(p);
1219 : :
1220 [ + + ]: 102 : if (strcmp(lchan, channel) == 0)
1221 : 41 : return true;
1222 : : }
1223 : 1155 : return false;
1224 : : }
1225 : :
1226 : : /*
1227 : : * Remove our entry from the listeners array when we are no longer listening
1228 : : * on any channel. NB: must not fail if we're already not listening.
1229 : : */
1230 : : static void
1231 : 27 : asyncQueueUnregister(void)
1232 : : {
2489 1233 [ - + ]: 27 : Assert(listenChannels == NIL); /* else caller error */
1234 : :
3973 bruce@momjian.us 1235 [ + + ]: 27 : if (!amRegisteredListener) /* nothing to do */
4078 tgl@sss.pgh.pa.us 1236 : 8 : return;
1237 : :
1238 : : /*
1239 : : * Need exclusive lock here to manipulate list links.
1240 : : */
1430 1241 : 19 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
1242 : : /* Mark our entry as invalid */
42 heikki.linnakangas@i 1243 :GNC 19 : QUEUE_BACKEND_PID(MyProcNumber) = InvalidPid;
1244 : 19 : QUEUE_BACKEND_DBOID(MyProcNumber) = InvalidOid;
1245 : : /* and remove it from the list */
1246 [ + + ]: 19 : if (QUEUE_FIRST_LISTENER == MyProcNumber)
1247 : 17 : QUEUE_FIRST_LISTENER = QUEUE_NEXT_LISTENER(MyProcNumber);
1248 : : else
1249 : : {
1250 [ + - ]: 2 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
1251 : : {
1252 [ + - ]: 2 : if (QUEUE_NEXT_LISTENER(i) == MyProcNumber)
1253 : : {
1254 : 2 : QUEUE_NEXT_LISTENER(i) = QUEUE_NEXT_LISTENER(MyProcNumber);
1678 tgl@sss.pgh.pa.us 1255 :CBC 2 : break;
1256 : : }
1257 : : }
1258 : : }
42 heikki.linnakangas@i 1259 :GNC 19 : QUEUE_NEXT_LISTENER(MyProcNumber) = INVALID_PROC_NUMBER;
1430 tgl@sss.pgh.pa.us 1260 :CBC 19 : LWLockRelease(NotifyQueueLock);
1261 : :
1262 : : /* mark ourselves as no longer listed in the global array */
4078 1263 : 19 : amRegisteredListener = false;
1264 : : }
1265 : :
1266 : : /*
1267 : : * Test whether there is room to insert more notification messages.
1268 : : *
1269 : : * Caller must hold at least shared NotifyQueueLock.
1270 : : */
1271 : : static bool
5171 1272 : 83 : asyncQueueIsFull(void)
1273 : : {
137 akorotkov@postgresql 1274 :GNC 83 : int headPage = QUEUE_POS_PAGE(QUEUE_HEAD);
1275 : 83 : int tailPage = QUEUE_POS_PAGE(QUEUE_TAIL);
1276 : 83 : int occupied = headPage - tailPage;
1277 : :
1278 : 83 : return occupied >= max_notify_queue_pages;
1279 : : }
1280 : :
1281 : : /*
1282 : : * Advance the QueuePosition to the next entry, assuming that the current
1283 : : * entry is of length entryLength. If we jump to a new page the function
1284 : : * returns true, else false.
1285 : : */
1286 : : static bool
3366 tgl@sss.pgh.pa.us 1287 :CBC 2301 : asyncQueueAdvance(volatile QueuePosition *position, int entryLength)
1288 : : {
137 akorotkov@postgresql 1289 :GNC 2301 : int64 pageno = QUEUE_POS_PAGE(*position);
5161 bruce@momjian.us 1290 :CBC 2301 : int offset = QUEUE_POS_OFFSET(*position);
1291 : 2301 : bool pageJump = false;
1292 : :
1293 : : /*
1294 : : * Move to the next writing position: First jump over what we have just
1295 : : * written or read.
1296 : : */
5171 tgl@sss.pgh.pa.us 1297 : 2301 : offset += entryLength;
1298 [ - + ]: 2301 : Assert(offset <= QUEUE_PAGESIZE);
1299 : :
1300 : : /*
1301 : : * In a second step check if another entry can possibly be written to the
1302 : : * page. If so, stay here, we have reached the next position. If not, then
1303 : : * we need to move on to the next page.
1304 : : */
1305 [ + + ]: 2301 : if (offset + QUEUEALIGN(AsyncQueueEntryEmptySize) > QUEUE_PAGESIZE)
1306 : : {
1307 : 70 : pageno++;
1308 : 70 : offset = 0;
1309 : 70 : pageJump = true;
1310 : : }
1311 : :
1312 : 2301 : SET_QUEUE_POS(*position, pageno, offset);
1313 : 2301 : return pageJump;
1314 : : }
1315 : :
1316 : : /*
1317 : : * Fill the AsyncQueueEntry at *qe with an outbound notification message.
1318 : : */
1319 : : static void
1320 : 1109 : asyncQueueNotificationToEntry(Notification *n, AsyncQueueEntry *qe)
1321 : : {
1704 1322 : 1109 : size_t channellen = n->channel_len;
1323 : 1109 : size_t payloadlen = n->payload_len;
1324 : : int entryLength;
1325 : :
5171 1326 [ - + ]: 1109 : Assert(channellen < NAMEDATALEN);
1327 [ - + ]: 1109 : Assert(payloadlen < NOTIFY_PAYLOAD_MAX_LENGTH);
1328 : :
1329 : : /* The terminators are already included in AsyncQueueEntryEmptySize */
1330 : 1109 : entryLength = AsyncQueueEntryEmptySize + payloadlen + channellen;
1331 : 1109 : entryLength = QUEUEALIGN(entryLength);
1332 : 1109 : qe->length = entryLength;
1333 : 1109 : qe->dboid = MyDatabaseId;
1334 : 1109 : qe->xid = GetCurrentTransactionId();
1335 : 1109 : qe->srcPid = MyProcPid;
1704 1336 : 1109 : memcpy(qe->data, n->data, channellen + payloadlen + 2);
5171 1337 : 1109 : }
1338 : :
1339 : : /*
1340 : : * Add pending notifications to the queue.
1341 : : *
1342 : : * We go page by page here, i.e. we stop once we have to go to a new page but
1343 : : * we will be called again and then fill that next page. If an entry does not
1344 : : * fit into the current page, we write a dummy entry with an InvalidOid as the
1345 : : * database OID in order to fill the page. So every page is always used up to
1346 : : * the last byte which simplifies reading the page later.
1347 : : *
1348 : : * We are passed the list cell (in pendingNotifies->events) containing the next
1349 : : * notification to write and return the first still-unwritten cell back.
1350 : : * Eventually we will return NULL indicating all is done.
1351 : : *
1352 : : * We are holding NotifyQueueLock already from the caller and grab
1353 : : * page specific SLRU bank lock locally in this function.
1354 : : */
1355 : : static ListCell *
1356 : 83 : asyncQueueAddEntries(ListCell *nextNotify)
1357 : : {
1358 : : AsyncQueueEntry qe;
1359 : : QueuePosition queue_head;
1360 : : int64 pageno;
1361 : : int offset;
1362 : : int slotno;
1363 : : LWLock *prevlock;
1364 : :
1365 : : /*
1366 : : * We work with a local copy of QUEUE_HEAD, which we write back to shared
1367 : : * memory upon exiting. The reason for this is that if we have to advance
1368 : : * to a new page, SimpleLruZeroPage might fail (out of disk space, for
1369 : : * instance), and we must not advance QUEUE_HEAD if it does. (Otherwise,
1370 : : * subsequent insertions would try to put entries into a page that slru.c
1371 : : * thinks doesn't exist yet.) So, use a local position variable. Note
1372 : : * that if we do fail, any already-inserted queue entries are forgotten;
1373 : : * this is okay, since they'd be useless anyway after our transaction
1374 : : * rolls back.
1375 : : */
4307 1376 : 83 : queue_head = QUEUE_HEAD;
1377 : :
1378 : : /*
1379 : : * If this is the first write since the postmaster started, we need to
1380 : : * initialize the first page of the async SLRU. Otherwise, the current
1381 : : * page should be initialized already, so just fetch it.
1382 : : */
1383 : 83 : pageno = QUEUE_POS_PAGE(queue_head);
46 alvherre@alvh.no-ip. 1384 :GNC 83 : prevlock = SimpleLruGetBankLock(NotifyCtl, pageno);
1385 : :
1386 : : /* We hold both NotifyQueueLock and SLRU bank lock during this operation */
1387 : 83 : LWLockAcquire(prevlock, LW_EXCLUSIVE);
1388 : :
1432 tgl@sss.pgh.pa.us 1389 [ + + + + ]:CBC 83 : if (QUEUE_POS_IS_ZERO(queue_head))
1430 1390 : 7 : slotno = SimpleLruZeroPage(NotifyCtl, pageno);
1391 : : else
1392 : 76 : slotno = SimpleLruReadPage(NotifyCtl, pageno, true,
1393 : : InvalidTransactionId);
1394 : :
1395 : : /* Note we mark the page dirty before writing in it */
1396 : 83 : NotifyCtl->shared->page_dirty[slotno] = true;
1397 : :
5171 1398 [ + + ]: 1157 : while (nextNotify != NULL)
1399 : : {
5161 bruce@momjian.us 1400 : 1109 : Notification *n = (Notification *) lfirst(nextNotify);
1401 : :
1402 : : /* Construct a valid queue entry in local variable qe */
5171 tgl@sss.pgh.pa.us 1403 : 1109 : asyncQueueNotificationToEntry(n, &qe);
1404 : :
4307 1405 : 1109 : offset = QUEUE_POS_OFFSET(queue_head);
1406 : :
1407 : : /* Check whether the entry really fits on the current page */
5171 1408 [ + + ]: 1109 : if (offset + qe.length <= QUEUE_PAGESIZE)
1409 : : {
1410 : : /* OK, so advance nextNotify past this item */
1704 1411 : 1076 : nextNotify = lnext(pendingNotifies->events, nextNotify);
1412 : : }
1413 : : else
1414 : : {
1415 : : /*
1416 : : * Write a dummy entry to fill up the page. Actually readers will
1417 : : * only check dboid and since it won't match any reader's database
1418 : : * OID, they will ignore this entry and move on.
1419 : : */
5171 1420 : 33 : qe.length = QUEUE_PAGESIZE - offset;
1421 : 33 : qe.dboid = InvalidOid;
5161 bruce@momjian.us 1422 : 33 : qe.data[0] = '\0'; /* empty channel */
1423 : 33 : qe.data[1] = '\0'; /* empty payload */
1424 : : }
1425 : :
1426 : : /* Now copy qe into the shared buffer page */
1430 tgl@sss.pgh.pa.us 1427 : 1109 : memcpy(NotifyCtl->shared->page_buffer[slotno] + offset,
1428 : : &qe,
5171 1429 : 1109 : qe.length);
1430 : :
1431 : : /* Advance queue_head appropriately, and detect if page is full */
4307 1432 [ + + ]: 1109 : if (asyncQueueAdvance(&(queue_head), qe.length))
1433 : : {
1434 : : LWLock *lock;
1435 : :
46 alvherre@alvh.no-ip. 1436 :GNC 35 : pageno = QUEUE_POS_PAGE(queue_head);
1437 : 35 : lock = SimpleLruGetBankLock(NotifyCtl, pageno);
1438 [ - + ]: 35 : if (lock != prevlock)
1439 : : {
46 alvherre@alvh.no-ip. 1440 :UNC 0 : LWLockRelease(prevlock);
1441 : 0 : LWLockAcquire(lock, LW_EXCLUSIVE);
1442 : 0 : prevlock = lock;
1443 : : }
1444 : :
1445 : : /*
1446 : : * Page is full, so we're done here, but first fill the next page
1447 : : * with zeroes. The reason to do this is to ensure that slru.c's
1448 : : * idea of the head page is always the same as ours, which avoids
1449 : : * boundary problems in SimpleLruTruncate. The test in
1450 : : * asyncQueueIsFull() ensured that there is room to create this
1451 : : * page without overrunning the queue.
1452 : : */
1430 tgl@sss.pgh.pa.us 1453 :CBC 35 : slotno = SimpleLruZeroPage(NotifyCtl, QUEUE_POS_PAGE(queue_head));
1454 : :
1455 : : /*
1456 : : * If the new page address is a multiple of QUEUE_CLEANUP_DELAY,
1457 : : * set flag to remember that we should try to advance the tail
1458 : : * pointer (we don't want to actually do that right here).
1459 : : */
1666 1460 [ + + ]: 35 : if (QUEUE_POS_PAGE(queue_head) % QUEUE_CLEANUP_DELAY == 0)
943 1461 : 8 : tryAdvanceTail = true;
1462 : :
1463 : : /* And exit the loop */
8342 1464 : 35 : break;
1465 : : }
1466 : : }
1467 : :
1468 : : /* Success, so update the global QUEUE_HEAD */
4307 1469 : 83 : QUEUE_HEAD = queue_head;
1470 : :
46 alvherre@alvh.no-ip. 1471 :GNC 83 : LWLockRelease(prevlock);
1472 : :
5171 tgl@sss.pgh.pa.us 1473 :CBC 83 : return nextNotify;
1474 : : }
1475 : :
1476 : : /*
1477 : : * SQL function to return the fraction of the notification queue currently
1478 : : * occupied.
1479 : : */
1480 : : Datum
3194 rhaas@postgresql.org 1481 : 5 : pg_notification_queue_usage(PG_FUNCTION_ARGS)
1482 : : {
1483 : : double usage;
1484 : :
1485 : : /* Advance the queue tail so we don't report a too-large result */
1603 tgl@sss.pgh.pa.us 1486 : 5 : asyncQueueAdvanceTail();
1487 : :
1430 1488 : 5 : LWLockAcquire(NotifyQueueLock, LW_SHARED);
3194 rhaas@postgresql.org 1489 : 5 : usage = asyncQueueUsage();
1430 tgl@sss.pgh.pa.us 1490 : 5 : LWLockRelease(NotifyQueueLock);
1491 : :
3194 rhaas@postgresql.org 1492 : 5 : PG_RETURN_FLOAT8(usage);
1493 : : }
1494 : :
1495 : : /*
1496 : : * Return the fraction of the queue that is currently occupied.
1497 : : *
1498 : : * The caller must hold NotifyQueueLock in (at least) shared mode.
1499 : : *
1500 : : * Note: we measure the distance to the logical tail page, not the physical
1501 : : * tail page. In some sense that's wrong, but the relative position of the
1502 : : * physical tail is affected by details such as SLRU segment boundaries,
1503 : : * so that a result based on that is unpleasantly unstable.
1504 : : */
1505 : : static double
1506 : 88 : asyncQueueUsage(void)
1507 : : {
5161 bruce@momjian.us 1508 : 88 : int headPage = QUEUE_POS_PAGE(QUEUE_HEAD);
1509 : 88 : int tailPage = QUEUE_POS_PAGE(QUEUE_TAIL);
137 akorotkov@postgresql 1510 :GNC 88 : int occupied = headPage - tailPage;
1511 : :
5171 tgl@sss.pgh.pa.us 1512 [ + + ]:CBC 88 : if (occupied == 0)
3194 rhaas@postgresql.org 1513 : 41 : return (double) 0; /* fast exit for common case */
1514 : :
137 akorotkov@postgresql 1515 :GNC 47 : return (double) occupied / (double) max_notify_queue_pages;
1516 : : }
1517 : :
1518 : : /*
1519 : : * Check whether the queue is at least half full, and emit a warning if so.
1520 : : *
1521 : : * This is unlikely given the size of the queue, but possible.
1522 : : * The warnings show up at most once every QUEUE_FULL_WARN_INTERVAL.
1523 : : *
1524 : : * Caller must hold exclusive NotifyQueueLock.
1525 : : */
1526 : : static void
3194 rhaas@postgresql.org 1527 :CBC 83 : asyncQueueFillWarning(void)
1528 : : {
1529 : : double fillDegree;
1530 : : TimestampTz t;
1531 : :
1532 : 83 : fillDegree = asyncQueueUsage();
5171 tgl@sss.pgh.pa.us 1533 [ + - ]: 83 : if (fillDegree < 0.5)
1534 : 83 : return;
1535 : :
5171 tgl@sss.pgh.pa.us 1536 :UBC 0 : t = GetCurrentTimestamp();
1537 : :
1538 [ # # ]: 0 : if (TimestampDifferenceExceeds(asyncQueueControl->lastQueueFillWarn,
1539 : : t, QUEUE_FULL_WARN_INTERVAL))
1540 : : {
5161 bruce@momjian.us 1541 : 0 : QueuePosition min = QUEUE_HEAD;
1542 : 0 : int32 minPid = InvalidPid;
1543 : :
42 heikki.linnakangas@i 1544 [ # # ]:UNC 0 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
1545 : : {
1678 tgl@sss.pgh.pa.us 1546 [ # # ]:UBC 0 : Assert(QUEUE_BACKEND_PID(i) != InvalidPid);
1547 [ # # # # : 0 : min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
# # ]
1548 [ # # # # ]: 0 : if (QUEUE_POS_EQUAL(min, QUEUE_BACKEND_POS(i)))
1549 : 0 : minPid = QUEUE_BACKEND_PID(i);
1550 : : }
1551 : :
5171 1552 [ # # # # : 0 : ereport(WARNING,
# # ]
1553 : : (errmsg("NOTIFY queue is %.0f%% full", fillDegree * 100),
1554 : : (minPid != InvalidPid ?
1555 : : errdetail("The server process with PID %d is among those with the oldest transactions.", minPid)
1556 : : : 0),
1557 : : (minPid != InvalidPid ?
1558 : : errhint("The NOTIFY queue cannot be emptied until that process ends its current transaction.")
1559 : : : 0)));
1560 : :
1561 : 0 : asyncQueueControl->lastQueueFillWarn = t;
1562 : : }
1563 : : }
1564 : :
1565 : : /*
1566 : : * Send signals to listening backends.
1567 : : *
1568 : : * Normally we signal only backends in our own database, since only those
1569 : : * backends could be interested in notifies we send. However, if there's
1570 : : * notify traffic in our database but no traffic in another database that
1571 : : * does have listener(s), those listeners will fall further and further
1572 : : * behind. Waken them anyway if they're far enough behind, so that they'll
1573 : : * advance their queue position pointers, allowing the global tail to advance.
1574 : : *
1575 : : * Since we know the ProcNumber and the Pid the signaling is quite cheap.
1576 : : *
1577 : : * This is called during CommitTransaction(), so it's important for it
1578 : : * to have very low probability of failure.
1579 : : */
1580 : : static void
5171 tgl@sss.pgh.pa.us 1581 :CBC 48 : SignalBackends(void)
1582 : : {
1583 : : int32 *pids;
1584 : : ProcNumber *procnos;
1585 : : int count;
1586 : :
1587 : : /*
1588 : : * Identify backends that we need to signal. We don't want to send
1589 : : * signals while holding the NotifyQueueLock, so this loop just builds a
1590 : : * list of target PIDs.
1591 : : *
1592 : : * XXX in principle these pallocs could fail, which would be bad. Maybe
1593 : : * preallocate the arrays? They're not that large, though.
1594 : : */
733 rhaas@postgresql.org 1595 : 48 : pids = (int32 *) palloc(MaxBackends * sizeof(int32));
42 heikki.linnakangas@i 1596 :GNC 48 : procnos = (ProcNumber *) palloc(MaxBackends * sizeof(ProcNumber));
5171 tgl@sss.pgh.pa.us 1597 :CBC 48 : count = 0;
1598 : :
1430 1599 : 48 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
42 heikki.linnakangas@i 1600 [ + + ]:GNC 88 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
1601 : : {
1666 tgl@sss.pgh.pa.us 1602 :CBC 40 : int32 pid = QUEUE_BACKEND_PID(i);
1603 : : QueuePosition pos;
1604 : :
1678 1605 [ - + ]: 40 : Assert(pid != InvalidPid);
1666 1606 : 40 : pos = QUEUE_BACKEND_POS(i);
1607 [ + - ]: 40 : if (QUEUE_BACKEND_DBOID(i) == MyDatabaseId)
1608 : : {
1609 : : /*
1610 : : * Always signal listeners in our own database, unless they're
1611 : : * already caught up (unlikely, but possible).
1612 : : */
1613 [ + + - + ]: 40 : if (QUEUE_POS_EQUAL(pos, QUEUE_HEAD))
1666 tgl@sss.pgh.pa.us 1614 :UBC 0 : continue;
1615 : : }
1616 : : else
1617 : : {
1618 : : /*
1619 : : * Listeners in other databases should be signaled only if they
1620 : : * are far behind.
1621 : : */
1622 [ # # ]: 0 : if (asyncQueuePageDiff(QUEUE_POS_PAGE(QUEUE_HEAD),
1623 : : QUEUE_POS_PAGE(pos)) < QUEUE_CLEANUP_DELAY)
1624 : 0 : continue;
1625 : : }
1626 : : /* OK, need to signal this one */
1666 tgl@sss.pgh.pa.us 1627 :CBC 40 : pids[count] = pid;
42 heikki.linnakangas@i 1628 :GNC 40 : procnos[count] = i;
1666 tgl@sss.pgh.pa.us 1629 :CBC 40 : count++;
1630 : : }
1430 1631 : 48 : LWLockRelease(NotifyQueueLock);
1632 : :
1633 : : /* Now send signals */
1678 1634 [ + + ]: 88 : for (int i = 0; i < count; i++)
1635 : : {
1666 1636 : 40 : int32 pid = pids[i];
1637 : :
1638 : : /*
1639 : : * If we are signaling our own process, no need to involve the kernel;
1640 : : * just set the flag directly.
1641 : : */
943 1642 [ + + ]: 40 : if (pid == MyProcPid)
1643 : : {
1644 : 20 : notifyInterruptPending = true;
1645 : 20 : continue;
1646 : : }
1647 : :
1648 : : /*
1649 : : * Note: assuming things aren't broken, a signal failure here could
1650 : : * only occur if the target backend exited since we released
1651 : : * NotifyQueueLock; which is unlikely but certainly possible. So we
1652 : : * just log a low-level debug message if it happens.
1653 : : */
42 heikki.linnakangas@i 1654 [ - + ]:GNC 20 : if (SendProcSignal(pid, PROCSIG_NOTIFY_INTERRUPT, procnos[i]) < 0)
5171 tgl@sss.pgh.pa.us 1655 [ # # ]:UBC 0 : elog(DEBUG3, "could not signal backend with PID %d: %m", pid);
1656 : : }
1657 : :
5171 tgl@sss.pgh.pa.us 1658 :CBC 48 : pfree(pids);
42 heikki.linnakangas@i 1659 :GNC 48 : pfree(procnos);
9322 tgl@sss.pgh.pa.us 1660 :CBC 48 : }
1661 : :
1662 : : /*
1663 : : * AtAbort_Notify
1664 : : *
1665 : : * This is called at transaction abort.
1666 : : *
1667 : : * Gets rid of pending actions and outbound notifies that we would have
1668 : : * executed if the transaction got committed.
1669 : : */
1670 : : void
8337 1671 : 22765 : AtAbort_Notify(void)
1672 : : {
1673 : : /*
1674 : : * If we LISTEN but then roll back the transaction after PreCommit_Notify,
1675 : : * we have registered as a listener but have not made any entry in
1676 : : * listenChannels. In that case, deregister again.
1677 : : */
4078 1678 [ - + - - ]: 22765 : if (amRegisteredListener && listenChannels == NIL)
4078 tgl@sss.pgh.pa.us 1679 :UBC 0 : asyncQueueUnregister();
1680 : :
1681 : : /* And clean up */
5877 tgl@sss.pgh.pa.us 1682 :CBC 22765 : ClearPendingActionsAndNotifies();
9322 1683 : 22765 : }
1684 : :
1685 : : /*
1686 : : * AtSubCommit_Notify() --- Take care of subtransaction commit.
1687 : : *
1688 : : * Reassign all items in the pending lists to the parent transaction.
1689 : : */
1690 : : void
7227 1691 : 5361 : AtSubCommit_Notify(void)
1692 : : {
1654 rhaas@postgresql.org 1693 : 5361 : int my_level = GetCurrentTransactionNestLevel();
1694 : :
1695 : : /* If there are actions at our nesting level, we must reparent them. */
1696 [ - + ]: 5361 : if (pendingActions != NULL &&
1654 rhaas@postgresql.org 1697 [ # # ]:UBC 0 : pendingActions->nestingLevel >= my_level)
1698 : : {
1699 [ # # ]: 0 : if (pendingActions->upper == NULL ||
1700 [ # # ]: 0 : pendingActions->upper->nestingLevel < my_level - 1)
1701 : : {
1702 : : /* nothing to merge; give the whole thing to the parent */
1703 : 0 : --pendingActions->nestingLevel;
1704 : : }
1705 : : else
1706 : : {
1707 : 0 : ActionList *childPendingActions = pendingActions;
1708 : :
1709 : 0 : pendingActions = pendingActions->upper;
1710 : :
1711 : : /*
1712 : : * Mustn't try to eliminate duplicates here --- see queue_listen()
1713 : : */
1714 : 0 : pendingActions->actions =
1715 : 0 : list_concat(pendingActions->actions,
1716 : 0 : childPendingActions->actions);
1717 : 0 : pfree(childPendingActions);
1718 : : }
1719 : : }
1720 : :
1721 : : /* If there are notifies at our nesting level, we must reparent them. */
1654 rhaas@postgresql.org 1722 [ + + ]:CBC 5361 : if (pendingNotifies != NULL &&
1723 [ + + ]: 2 : pendingNotifies->nestingLevel >= my_level)
1724 : : {
1725 [ - + ]: 1 : Assert(pendingNotifies->nestingLevel == my_level);
1726 : :
1727 [ + - ]: 1 : if (pendingNotifies->upper == NULL ||
1728 [ - + ]: 1 : pendingNotifies->upper->nestingLevel < my_level - 1)
1729 : : {
1730 : : /* nothing to merge; give the whole thing to the parent */
1654 rhaas@postgresql.org 1731 :UBC 0 : --pendingNotifies->nestingLevel;
1732 : : }
1733 : : else
1734 : : {
1735 : : /*
1736 : : * Formerly, we didn't bother to eliminate duplicates here, but
1737 : : * now we must, else we fall foul of "Assert(!found)", either here
1738 : : * or during a later attempt to build the parent-level hashtable.
1739 : : */
1654 rhaas@postgresql.org 1740 :CBC 1 : NotificationList *childPendingNotifies = pendingNotifies;
1741 : : ListCell *l;
1742 : :
1743 : 1 : pendingNotifies = pendingNotifies->upper;
1744 : : /* Insert all the subxact's events into parent, except for dups */
1745 [ + - + + : 5 : foreach(l, childPendingNotifies->events)
+ + ]
1746 : : {
1747 : 4 : Notification *childn = (Notification *) lfirst(l);
1748 : :
1749 [ + + ]: 4 : if (!AsyncExistsPendingNotify(childn))
1750 : 2 : AddEventToPendingNotifies(childn);
1751 : : }
1752 : 1 : pfree(childPendingNotifies);
1753 : : }
1754 : : }
7227 tgl@sss.pgh.pa.us 1755 : 5361 : }
1756 : :
1757 : : /*
1758 : : * AtSubAbort_Notify() --- Take care of subtransaction abort.
1759 : : */
1760 : : void
1761 : 4566 : AtSubAbort_Notify(void)
1762 : : {
7160 1763 : 4566 : int my_level = GetCurrentTransactionNestLevel();
1764 : :
1765 : : /*
1766 : : * All we have to do is pop the stack --- the actions/notifies made in
1767 : : * this subxact are no longer interesting, and the space will be freed
1768 : : * when CurTransactionContext is recycled. We still have to free the
1769 : : * ActionList and NotificationList objects themselves, though, because
1770 : : * those are allocated in TopTransactionContext.
1771 : : *
1772 : : * Note that there might be no entries at all, or no entries for the
1773 : : * current subtransaction level, either because none were ever created, or
1774 : : * because we reentered this routine due to trouble during subxact abort.
1775 : : */
1654 rhaas@postgresql.org 1776 [ - + ]: 4566 : while (pendingActions != NULL &&
1654 rhaas@postgresql.org 1777 [ # # ]:UBC 0 : pendingActions->nestingLevel >= my_level)
1778 : : {
1779 : 0 : ActionList *childPendingActions = pendingActions;
1780 : :
1781 : 0 : pendingActions = pendingActions->upper;
1782 : 0 : pfree(childPendingActions);
1783 : : }
1784 : :
1654 rhaas@postgresql.org 1785 [ + + ]:CBC 4567 : while (pendingNotifies != NULL &&
1786 [ + + ]: 2 : pendingNotifies->nestingLevel >= my_level)
1787 : : {
1788 : 1 : NotificationList *childPendingNotifies = pendingNotifies;
1789 : :
1790 : 1 : pendingNotifies = pendingNotifies->upper;
1791 : 1 : pfree(childPendingNotifies);
1792 : : }
7227 tgl@sss.pgh.pa.us 1793 : 4566 : }
1794 : :
1795 : : /*
1796 : : * HandleNotifyInterrupt
1797 : : *
1798 : : * Signal handler portion of interrupt handling. Let the backend know
1799 : : * that there's a pending notify interrupt. If we're currently reading
1800 : : * from the client, this will interrupt the read and
1801 : : * ProcessClientReadInterrupt() will call ProcessNotifyInterrupt().
1802 : : */
1803 : : void
5371 1804 : 20 : HandleNotifyInterrupt(void)
1805 : : {
1806 : : /*
1807 : : * Note: this is called by a SIGNAL HANDLER. You must be very wary what
1808 : : * you do here.
1809 : : */
1810 : :
1811 : : /* signal that work needs to be done */
3358 andres@anarazel.de 1812 : 20 : notifyInterruptPending = true;
1813 : :
1814 : : /* make sure the event is processed in due course */
1815 : 20 : SetLatch(MyLatch);
9322 tgl@sss.pgh.pa.us 1816 : 20 : }
1817 : :
1818 : : /*
1819 : : * ProcessNotifyInterrupt
1820 : : *
1821 : : * This is called if we see notifyInterruptPending set, just before
1822 : : * transmitting ReadyForQuery at the end of a frontend command, and
1823 : : * also if a notify signal occurs while reading from the frontend.
1824 : : * HandleNotifyInterrupt() will cause the read to be interrupted
1825 : : * via the process's latch, and this routine will get called.
1826 : : * If we are truly idle (ie, *not* inside a transaction block),
1827 : : * process the incoming notifies.
1828 : : *
1829 : : * If "flush" is true, force any frontend messages out immediately.
1830 : : * This can be false when being called at the end of a frontend command,
1831 : : * since we'll flush after sending ReadyForQuery.
1832 : : */
1833 : : void
943 1834 : 95 : ProcessNotifyInterrupt(bool flush)
1835 : : {
7486 1836 [ + + ]: 95 : if (IsTransactionOrTransactionBlock())
9322 1837 : 56 : return; /* not really idle */
1838 : :
1839 : : /* Loop in case another signal arrives while sending messages */
3358 andres@anarazel.de 1840 [ + + ]: 78 : while (notifyInterruptPending)
943 tgl@sss.pgh.pa.us 1841 : 39 : ProcessIncomingNotify(flush);
1842 : : }
1843 : :
1844 : :
1845 : : /*
1846 : : * Read all pending notifications from the queue, and deliver appropriate
1847 : : * ones to my frontend. Stop when we reach queue head or an uncommitted
1848 : : * notification.
1849 : : */
1850 : : static void
5171 1851 : 50 : asyncQueueReadAllNotifications(void)
1852 : : {
1853 : : volatile QueuePosition pos;
1854 : : QueuePosition head;
1855 : : Snapshot snapshot;
1856 : :
1857 : : /* page_buffer must be adequately aligned, so use a union */
1858 : : union
1859 : : {
1860 : : char buf[QUEUE_PAGESIZE];
1861 : : AsyncQueueEntry align;
1862 : : } page_buffer;
1863 : :
1864 : : /* Fetch current state */
1430 1865 : 50 : LWLockAcquire(NotifyQueueLock, LW_SHARED);
1866 : : /* Assert checks that we have a valid state entry */
42 heikki.linnakangas@i 1867 [ - + ]:GNC 50 : Assert(MyProcPid == QUEUE_BACKEND_PID(MyProcNumber));
1868 : 50 : pos = QUEUE_BACKEND_POS(MyProcNumber);
5171 tgl@sss.pgh.pa.us 1869 :CBC 50 : head = QUEUE_HEAD;
1430 1870 : 50 : LWLockRelease(NotifyQueueLock);
1871 : :
5171 1872 [ + + - + ]: 50 : if (QUEUE_POS_EQUAL(pos, head))
1873 : : {
1874 : : /* Nothing to do, we have read all notifications already. */
5171 tgl@sss.pgh.pa.us 1875 :UBC 0 : return;
1876 : : }
1877 : :
1878 : : /*----------
1879 : : * Get snapshot we'll use to decide which xacts are still in progress.
1880 : : * This is trickier than it might seem, because of race conditions.
1881 : : * Consider the following example:
1882 : : *
1883 : : * Backend 1: Backend 2:
1884 : : *
1885 : : * transaction starts
1886 : : * UPDATE foo SET ...;
1887 : : * NOTIFY foo;
1888 : : * commit starts
1889 : : * queue the notify message
1890 : : * transaction starts
1891 : : * LISTEN foo; -- first LISTEN in session
1892 : : * SELECT * FROM foo WHERE ...;
1893 : : * commit to clog
1894 : : * commit starts
1895 : : * add backend 2 to array of listeners
1896 : : * advance to queue head (this code)
1897 : : * commit to clog
1898 : : *
1899 : : * Transaction 2's SELECT has not seen the UPDATE's effects, since that
1900 : : * wasn't committed yet. Ideally we'd ensure that client 2 would
1901 : : * eventually get transaction 1's notify message, but there's no way
1902 : : * to do that; until we're in the listener array, there's no guarantee
1903 : : * that the notify message doesn't get removed from the queue.
1904 : : *
1905 : : * Therefore the coding technique transaction 2 is using is unsafe:
1906 : : * applications must commit a LISTEN before inspecting database state,
1907 : : * if they want to ensure they will see notifications about subsequent
1908 : : * changes to that state.
1909 : : *
1910 : : * What we do guarantee is that we'll see all notifications from
1911 : : * transactions committing after the snapshot we take here.
1912 : : * Exec_ListenPreCommit has already added us to the listener array,
1913 : : * so no not-yet-committed messages can be removed from the queue
1914 : : * before we see them.
1915 : : *----------
1916 : : */
1603 tgl@sss.pgh.pa.us 1917 :CBC 50 : snapshot = RegisterSnapshot(GetLatestSnapshot());
1918 : :
1919 : : /*
1920 : : * It is possible that we fail while trying to send a message to our
1921 : : * frontend (for example, because of encoding conversion failure). If
1922 : : * that happens it is critical that we not try to send the same message
1923 : : * over and over again. Therefore, we place a PG_TRY block here that will
1924 : : * forcibly advance our queue position before we lose control to an error.
1925 : : * (We could alternatively retake NotifyQueueLock and move the position
1926 : : * before handling each individual message, but that seems like too much
1927 : : * lock traffic.)
1928 : : */
5171 1929 [ + - ]: 50 : PG_TRY();
1930 : : {
1931 : : bool reachedStop;
1932 : :
1933 : : do
1934 : : {
1935 : 85 : int curpage = QUEUE_POS_PAGE(pos);
1936 : 85 : int curoffset = QUEUE_POS_OFFSET(pos);
1937 : : int slotno;
1938 : : int copysize;
1939 : :
1940 : : /*
1941 : : * We copy the data from SLRU into a local buffer, so as to avoid
1942 : : * holding the SLRU lock while we are examining the entries and
1943 : : * possibly transmitting them to our frontend. Copy only the part
1944 : : * of the page we will actually inspect.
1945 : : */
1430 1946 : 85 : slotno = SimpleLruReadPage_ReadOnly(NotifyCtl, curpage,
1947 : : InvalidTransactionId);
5171 1948 [ + + ]: 85 : if (curpage == QUEUE_POS_PAGE(head))
1949 : : {
1950 : : /* we only want to read as far as head */
1951 : 50 : copysize = QUEUE_POS_OFFSET(head) - curoffset;
1952 [ - + ]: 50 : if (copysize < 0)
2489 tgl@sss.pgh.pa.us 1953 :UBC 0 : copysize = 0; /* just for safety */
1954 : : }
1955 : : else
1956 : : {
1957 : : /* fetch all the rest of the page */
5171 tgl@sss.pgh.pa.us 1958 :CBC 35 : copysize = QUEUE_PAGESIZE - curoffset;
1959 : : }
1960 : 85 : memcpy(page_buffer.buf + curoffset,
1430 1961 : 85 : NotifyCtl->shared->page_buffer[slotno] + curoffset,
1962 : : copysize);
1963 : : /* Release lock that we got from SimpleLruReadPage_ReadOnly() */
46 alvherre@alvh.no-ip. 1964 :GNC 85 : LWLockRelease(SimpleLruGetBankLock(NotifyCtl, curpage));
1965 : :
1966 : : /*
1967 : : * Process messages up to the stop position, end of page, or an
1968 : : * uncommitted message.
1969 : : *
1970 : : * Our stop position is what we found to be the head's position
1971 : : * when we entered this function. It might have changed already.
1972 : : * But if it has, we will receive (or have already received and
1973 : : * queued) another signal and come here again.
1974 : : *
1975 : : * We are not holding NotifyQueueLock here! The queue can only
1976 : : * extend beyond the head pointer (see above) and we leave our
1977 : : * backend's pointer where it is so nobody will truncate or
1978 : : * rewrite pages under us. Especially we don't want to hold a lock
1979 : : * while sending the notifications to the frontend.
1980 : : */
5171 tgl@sss.pgh.pa.us 1981 :CBC 85 : reachedStop = asyncQueueProcessPageEntries(&pos, head,
1982 : : page_buffer.buf,
1983 : : snapshot);
1984 [ + + ]: 85 : } while (!reachedStop);
1985 : : }
1626 peter@eisentraut.org 1986 :UBC 0 : PG_FINALLY();
1987 : : {
1988 : : /* Update shared state */
1430 tgl@sss.pgh.pa.us 1989 :CBC 50 : LWLockAcquire(NotifyQueueLock, LW_SHARED);
42 heikki.linnakangas@i 1990 :GNC 50 : QUEUE_BACKEND_POS(MyProcNumber) = pos;
1430 tgl@sss.pgh.pa.us 1991 :CBC 50 : LWLockRelease(NotifyQueueLock);
1992 : : }
5171 1993 [ - + ]: 50 : PG_END_TRY();
1994 : :
1995 : : /* Done with snapshot */
2377 1996 : 50 : UnregisterSnapshot(snapshot);
1997 : : }
1998 : :
1999 : : /*
2000 : : * Fetch notifications from the shared queue, beginning at position current,
2001 : : * and deliver relevant ones to my frontend.
2002 : : *
2003 : : * The current page must have been fetched into page_buffer from shared
2004 : : * memory. (We could access the page right in shared memory, but that
2005 : : * would imply holding the SLRU bank lock throughout this routine.)
2006 : : *
2007 : : * We stop if we reach the "stop" position, or reach a notification from an
2008 : : * uncommitted transaction, or reach the end of the page.
2009 : : *
2010 : : * The function returns true once we have reached the stop position or an
2011 : : * uncommitted notification, and false if we have finished with the page.
2012 : : * In other words: once it returns true there is no need to look further.
2013 : : * The QueuePosition *current is advanced past all processed messages.
2014 : : */
2015 : : static bool
3366 2016 : 85 : asyncQueueProcessPageEntries(volatile QueuePosition *current,
2017 : : QueuePosition stop,
2018 : : char *page_buffer,
2019 : : Snapshot snapshot)
2020 : : {
5171 2021 : 85 : bool reachedStop = false;
2022 : : bool reachedEndOfPage;
2023 : : AsyncQueueEntry *qe;
2024 : :
2025 : : do
2026 : : {
5161 bruce@momjian.us 2027 : 1242 : QueuePosition thisentry = *current;
2028 : :
5170 tgl@sss.pgh.pa.us 2029 [ + + + + ]: 1242 : if (QUEUE_POS_EQUAL(thisentry, stop))
5171 2030 : 50 : break;
2031 : :
5170 2032 : 1192 : qe = (AsyncQueueEntry *) (page_buffer + QUEUE_POS_OFFSET(thisentry));
2033 : :
2034 : : /*
2035 : : * Advance *current over this message, possibly to the next page. As
2036 : : * noted in the comments for asyncQueueReadAllNotifications, we must
2037 : : * do this before possibly failing while processing the message.
2038 : : */
5171 2039 : 1192 : reachedEndOfPage = asyncQueueAdvance(current, qe->length);
2040 : :
2041 : : /* Ignore messages destined for other databases */
2042 [ + + ]: 1192 : if (qe->dboid == MyDatabaseId)
2043 : : {
2377 2044 [ - + ]: 1159 : if (XidInMVCCSnapshot(qe->xid, snapshot))
2045 : : {
2046 : : /*
2047 : : * The source transaction is still in progress, so we can't
2048 : : * process this message yet. Break out of the loop, but first
2049 : : * back up *current so we will reprocess the message next
2050 : : * time. (Note: it is unlikely but not impossible for
2051 : : * TransactionIdDidCommit to fail, so we can't really avoid
2052 : : * this advance-then-back-up behavior when dealing with an
2053 : : * uncommitted message.)
2054 : : *
2055 : : * Note that we must test XidInMVCCSnapshot before we test
2056 : : * TransactionIdDidCommit, else we might return a message from
2057 : : * a transaction that is not yet visible to snapshots; compare
2058 : : * the comments at the head of heapam_visibility.c.
2059 : : *
2060 : : * Also, while our own xact won't be listed in the snapshot,
2061 : : * we need not check for TransactionIdIsCurrentTransactionId
2062 : : * because our transaction cannot (yet) have queued any
2063 : : * messages.
2064 : : */
3685 tgl@sss.pgh.pa.us 2065 :UBC 0 : *current = thisentry;
2066 : 0 : reachedStop = true;
2067 : 0 : break;
2068 : : }
3685 tgl@sss.pgh.pa.us 2069 [ + - ]:CBC 1159 : else if (TransactionIdDidCommit(qe->xid))
2070 : : {
2071 : : /* qe->data is the null-terminated channel name */
5161 bruce@momjian.us 2072 : 1159 : char *channel = qe->data;
2073 : :
5171 tgl@sss.pgh.pa.us 2074 [ + + ]: 1159 : if (IsListeningOn(channel))
2075 : : {
2076 : : /* payload follows channel name */
5161 bruce@momjian.us 2077 : 31 : char *payload = qe->data + strlen(channel) + 1;
2078 : :
5171 tgl@sss.pgh.pa.us 2079 : 31 : NotifyMyFrontEnd(channel, payload, qe->srcPid);
2080 : : }
2081 : : }
2082 : : else
2083 : : {
2084 : : /*
2085 : : * The source transaction aborted or crashed, so we just
2086 : : * ignore its notifications.
2087 : : */
2088 : : }
2089 : : }
2090 : :
2091 : : /* Loop back if we're not at end of page */
2092 [ + + ]: 1192 : } while (!reachedEndOfPage);
2093 : :
2094 [ + + + + ]: 85 : if (QUEUE_POS_EQUAL(*current, stop))
2095 : 50 : reachedStop = true;
2096 : :
2097 : 85 : return reachedStop;
2098 : : }
2099 : :
2100 : : /*
2101 : : * Advance the shared queue tail variable to the minimum of all the
2102 : : * per-backend tail pointers. Truncate pg_notify space if possible.
2103 : : *
2104 : : * This is (usually) called during CommitTransaction(), so it's important for
2105 : : * it to have very low probability of failure.
2106 : : */
2107 : : static void
2108 : 13 : asyncQueueAdvanceTail(void)
2109 : : {
2110 : : QueuePosition min;
2111 : : int oldtailpage;
2112 : : int newtailpage;
2113 : : int boundary;
2114 : :
2115 : : /* Restrict task to one backend per cluster; see SimpleLruTruncate(). */
1338 noah@leadboat.com 2116 : 13 : LWLockAcquire(NotifyQueueTailLock, LW_EXCLUSIVE);
2117 : :
2118 : : /*
2119 : : * Compute the new tail. Pre-v13, it's essential that QUEUE_TAIL be exact
2120 : : * (ie, exactly match at least one backend's queue position), so it must
2121 : : * be updated atomically with the actual computation. Since v13, we could
2122 : : * get away with not doing it like that, but it seems prudent to keep it
2123 : : * so.
2124 : : *
2125 : : * Also, because incoming backends will scan forward from QUEUE_TAIL, that
2126 : : * must be advanced before we can truncate any data. Thus, QUEUE_TAIL is
2127 : : * the logical tail, while QUEUE_STOP_PAGE is the physical tail, or oldest
2128 : : * un-truncated page. When QUEUE_STOP_PAGE != QUEUE_POS_PAGE(QUEUE_TAIL),
2129 : : * there are pages we can truncate but haven't yet finished doing so.
2130 : : *
2131 : : * For concurrency's sake, we don't want to hold NotifyQueueLock while
2132 : : * performing SimpleLruTruncate. This is OK because no backend will try
2133 : : * to access the pages we are in the midst of truncating.
2134 : : */
1430 tgl@sss.pgh.pa.us 2135 : 13 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
5171 2136 : 13 : min = QUEUE_HEAD;
42 heikki.linnakangas@i 2137 [ + + ]:GNC 23 : for (ProcNumber i = QUEUE_FIRST_LISTENER; i != INVALID_PROC_NUMBER; i = QUEUE_NEXT_LISTENER(i))
2138 : : {
1678 tgl@sss.pgh.pa.us 2139 [ - + ]:CBC 10 : Assert(QUEUE_BACKEND_PID(i) != InvalidPid);
2140 [ + - + + : 10 : min = QUEUE_POS_MIN(min, QUEUE_BACKEND_POS(i));
+ - ]
2141 : : }
1233 2142 : 13 : QUEUE_TAIL = min;
2143 : 13 : oldtailpage = QUEUE_STOP_PAGE;
1430 2144 : 13 : LWLockRelease(NotifyQueueLock);
2145 : :
2146 : : /*
2147 : : * We can truncate something if the global tail advanced across an SLRU
2148 : : * segment boundary.
2149 : : *
2150 : : * XXX it might be better to truncate only once every several segments, to
2151 : : * reduce the number of directory scans.
2152 : : */
5171 2153 : 13 : newtailpage = QUEUE_POS_PAGE(min);
2154 : 13 : boundary = newtailpage - (newtailpage % SLRU_PAGES_PER_SEGMENT);
4582 alvherre@alvh.no-ip. 2155 [ - + ]: 13 : if (asyncQueuePagePrecedes(oldtailpage, boundary))
2156 : : {
2157 : : /*
2158 : : * SimpleLruTruncate() will ask for SLRU bank locks but will also
2159 : : * release the lock again.
2160 : : */
1430 tgl@sss.pgh.pa.us 2161 :UBC 0 : SimpleLruTruncate(NotifyCtl, newtailpage);
2162 : :
1233 2163 : 0 : LWLockAcquire(NotifyQueueLock, LW_EXCLUSIVE);
2164 : 0 : QUEUE_STOP_PAGE = newtailpage;
2165 : 0 : LWLockRelease(NotifyQueueLock);
2166 : : }
2167 : :
1338 noah@leadboat.com 2168 :CBC 13 : LWLockRelease(NotifyQueueTailLock);
5171 tgl@sss.pgh.pa.us 2169 : 13 : }
2170 : :
2171 : : /*
2172 : : * ProcessIncomingNotify
2173 : : *
2174 : : * Scan the queue for arriving notifications and report them to the front
2175 : : * end. The notifications might be from other sessions, or our own;
2176 : : * there's no need to distinguish here.
2177 : : *
2178 : : * If "flush" is true, force any frontend messages out immediately.
2179 : : *
2180 : : * NOTE: since we are outside any transaction, we must create our own.
2181 : : */
2182 : : static void
943 2183 : 39 : ProcessIncomingNotify(bool flush)
2184 : : {
2185 : : /* We *must* reset the flag */
3358 andres@anarazel.de 2186 : 39 : notifyInterruptPending = false;
2187 : :
2188 : : /* Do nothing else if we aren't actively listening */
5171 tgl@sss.pgh.pa.us 2189 [ - + ]: 39 : if (listenChannels == NIL)
5171 tgl@sss.pgh.pa.us 2190 :UBC 0 : return;
2191 : :
8719 peter_e@gmx.net 2192 [ - + ]:CBC 39 : if (Trace_notify)
7628 bruce@momjian.us 2193 [ # # ]:UBC 0 : elog(DEBUG1, "ProcessIncomingNotify");
2194 : :
1495 peter@eisentraut.org 2195 :CBC 39 : set_ps_display("notify interrupt");
2196 : :
2197 : : /*
2198 : : * We must run asyncQueueReadAllNotifications inside a transaction, else
2199 : : * bad things happen if it gets an error.
2200 : : */
5171 tgl@sss.pgh.pa.us 2201 : 39 : StartTransactionCommand();
2202 : :
2203 : 39 : asyncQueueReadAllNotifications();
2204 : :
7641 2205 : 39 : CommitTransactionCommand();
2206 : :
2207 : : /*
2208 : : * If this isn't an end-of-command case, we must flush the notify messages
2209 : : * to ensure frontend gets them promptly.
2210 : : */
943 2211 [ + + ]: 39 : if (flush)
2212 : 10 : pq_flush();
2213 : :
1495 peter@eisentraut.org 2214 : 39 : set_ps_display("idle");
2215 : :
8719 peter_e@gmx.net 2216 [ - + ]: 39 : if (Trace_notify)
7628 bruce@momjian.us 2217 [ # # ]:UBC 0 : elog(DEBUG1, "ProcessIncomingNotify: done");
2218 : : }
2219 : :
2220 : : /*
2221 : : * Send NOTIFY message to my front end.
2222 : : */
2223 : : void
5171 tgl@sss.pgh.pa.us 2224 :CBC 31 : NotifyMyFrontEnd(const char *channel, const char *payload, int32 srcPid)
2225 : : {
6737 alvherre@alvh.no-ip. 2226 [ + - ]: 31 : if (whereToSendOutput == DestRemote)
2227 : : {
2228 : : StringInfoData buf;
2229 : :
236 nathan@postgresql.or 2230 :GNC 31 : pq_beginmessage(&buf, PqMsg_NotificationResponse);
2377 andres@anarazel.de 2231 :CBC 31 : pq_sendint32(&buf, srcPid);
5171 tgl@sss.pgh.pa.us 2232 : 31 : pq_sendstring(&buf, channel);
1137 heikki.linnakangas@i 2233 : 31 : pq_sendstring(&buf, payload);
9121 tgl@sss.pgh.pa.us 2234 : 31 : pq_endmessage(&buf);
2235 : :
2236 : : /*
2237 : : * NOTE: we do not do pq_flush() here. Some level of caller will
2238 : : * handle it later, allowing this message to be combined into a packet
2239 : : * with other ones.
2240 : : */
2241 : : }
2242 : : else
5171 tgl@sss.pgh.pa.us 2243 [ # # ]:UBC 0 : elog(INFO, "NOTIFY for \"%s\" payload \"%s\"", channel, payload);
9322 tgl@sss.pgh.pa.us 2244 :CBC 31 : }
2245 : :
2246 : : /* Does pendingNotifies include a match for the given event? */
2247 : : static bool
1704 2248 : 1048 : AsyncExistsPendingNotify(Notification *n)
2249 : : {
2250 [ - + ]: 1048 : if (pendingNotifies == NULL)
5171 tgl@sss.pgh.pa.us 2251 :UBC 0 : return false;
2252 : :
1704 tgl@sss.pgh.pa.us 2253 [ + + ]:CBC 1048 : if (pendingNotifies->hashtab != NULL)
2254 : : {
2255 : : /* Use the hash table to probe for a match */
2256 [ - + ]: 983 : if (hash_search(pendingNotifies->hashtab,
2257 : : &n,
2258 : : HASH_FIND,
2259 : : NULL))
1704 tgl@sss.pgh.pa.us 2260 :UBC 0 : return true;
2261 : : }
2262 : : else
2263 : : {
2264 : : /* Must scan the event list */
2265 : : ListCell *l;
2266 : :
1704 tgl@sss.pgh.pa.us 2267 [ + - + + :CBC 271 : foreach(l, pendingNotifies->events)
+ + ]
2268 : : {
2269 : 220 : Notification *oldn = (Notification *) lfirst(l);
2270 : :
2271 [ + - ]: 220 : if (n->channel_len == oldn->channel_len &&
2272 [ + + ]: 220 : n->payload_len == oldn->payload_len &&
2273 : 125 : memcmp(n->data, oldn->data,
2274 [ + + ]: 125 : n->channel_len + n->payload_len + 2) == 0)
2275 : 14 : return true;
2276 : : }
2277 : : }
2278 : :
2279 : 1034 : return false;
2280 : : }
2281 : :
2282 : : /*
2283 : : * Add a notification event to a pre-existing pendingNotifies list.
2284 : : *
2285 : : * Because pendingNotifies->events is already nonempty, this works
2286 : : * correctly no matter what CurrentMemoryContext is.
2287 : : */
2288 : : static void
2289 : 1034 : AddEventToPendingNotifies(Notification *n)
2290 : : {
2291 [ - + ]: 1034 : Assert(pendingNotifies->events != NIL);
2292 : :
2293 : : /* Create the hash table if it's time to */
2294 [ + + ]: 1034 : if (list_length(pendingNotifies->events) >= MIN_HASHABLE_NOTIFIES &&
2295 [ + + ]: 984 : pendingNotifies->hashtab == NULL)
2296 : : {
2297 : : HASHCTL hash_ctl;
2298 : : ListCell *l;
2299 : :
2300 : : /* Create the hash table */
2301 : 1 : hash_ctl.keysize = sizeof(Notification *);
2302 : 1 : hash_ctl.entrysize = sizeof(NotificationHash);
2303 : 1 : hash_ctl.hash = notification_hash;
2304 : 1 : hash_ctl.match = notification_match;
2305 : 1 : hash_ctl.hcxt = CurTransactionContext;
2306 : 2 : pendingNotifies->hashtab =
2307 : 1 : hash_create("Pending Notifies",
2308 : : 256L,
2309 : : &hash_ctl,
2310 : : HASH_ELEM | HASH_FUNCTION | HASH_COMPARE | HASH_CONTEXT);
2311 : :
2312 : : /* Insert all the already-existing events */
2313 [ + - + + : 17 : foreach(l, pendingNotifies->events)
+ + ]
2314 : : {
2315 : 16 : Notification *oldn = (Notification *) lfirst(l);
2316 : : bool found;
2317 : :
136 john.naylor@postgres 2318 :GNC 16 : (void) hash_search(pendingNotifies->hashtab,
2319 : : &oldn,
2320 : : HASH_ENTER,
2321 : : &found);
1704 tgl@sss.pgh.pa.us 2322 [ - + ]:CBC 16 : Assert(!found);
2323 : : }
2324 : : }
2325 : :
2326 : : /* Add new event to the list, in order */
2327 : 1034 : pendingNotifies->events = lappend(pendingNotifies->events, n);
2328 : :
2329 : : /* Add event to the hash table if needed */
2330 [ + + ]: 1034 : if (pendingNotifies->hashtab != NULL)
2331 : : {
2332 : : bool found;
2333 : :
136 john.naylor@postgres 2334 :GNC 984 : (void) hash_search(pendingNotifies->hashtab,
2335 : : &n,
2336 : : HASH_ENTER,
2337 : : &found);
1704 tgl@sss.pgh.pa.us 2338 [ - + ]:CBC 984 : Assert(!found);
2339 : : }
2340 : 1034 : }
2341 : :
2342 : : /*
2343 : : * notification_hash: hash function for notification hash table
2344 : : *
2345 : : * The hash "keys" are pointers to Notification structs.
2346 : : */
2347 : : static uint32
2348 : 1983 : notification_hash(const void *key, Size keysize)
2349 : : {
2350 : 1983 : const Notification *k = *(const Notification *const *) key;
2351 : :
2352 [ - + ]: 1983 : Assert(keysize == sizeof(Notification *));
2353 : : /* We don't bother to include the payload's trailing null in the hash */
2354 : 1983 : return DatumGetUInt32(hash_any((const unsigned char *) k->data,
2355 : 1983 : k->channel_len + k->payload_len + 1));
2356 : : }
2357 : :
2358 : : /*
2359 : : * notification_match: match function to use with notification_hash
2360 : : */
2361 : : static int
1704 tgl@sss.pgh.pa.us 2362 :UBC 0 : notification_match(const void *key1, const void *key2, Size keysize)
2363 : : {
2364 : 0 : const Notification *k1 = *(const Notification *const *) key1;
2365 : 0 : const Notification *k2 = *(const Notification *const *) key2;
2366 : :
2367 [ # # ]: 0 : Assert(keysize == sizeof(Notification *));
2368 [ # # ]: 0 : if (k1->channel_len == k2->channel_len &&
2369 [ # # ]: 0 : k1->payload_len == k2->payload_len &&
2370 : 0 : memcmp(k1->data, k2->data,
2371 [ # # ]: 0 : k1->channel_len + k1->payload_len + 2) == 0)
2372 : 0 : return 0; /* equal */
2373 : 0 : return 1; /* not equal */
2374 : : }
2375 : :
2376 : : /* Clear the pendingActions and pendingNotifies lists. */
2377 : : static void
5877 tgl@sss.pgh.pa.us 2378 :CBC 22862 : ClearPendingActionsAndNotifies(void)
2379 : : {
2380 : : /*
2381 : : * Everything's allocated in either TopTransactionContext or the context
2382 : : * for the subtransaction to which it corresponds. So, there's nothing to
2383 : : * do here except reset the pointers; the space will be reclaimed when the
2384 : : * contexts are deleted.
2385 : : */
1654 rhaas@postgresql.org 2386 : 22862 : pendingActions = NULL;
1704 tgl@sss.pgh.pa.us 2387 : 22862 : pendingNotifies = NULL;
9716 bruce@momjian.us 2388 : 22862 : }
2389 : :
2390 : : /*
2391 : : * GUC check_hook for notify_buffers
2392 : : */
2393 : : bool
46 alvherre@alvh.no-ip. 2394 :GNC 928 : check_notify_buffers(int *newval, void **extra, GucSource source)
2395 : : {
2396 : 928 : return check_slru_buffers("notify_buffers", newval);
2397 : : }
|
