Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * walsender.c
4 : *
5 : * The WAL sender process (walsender) is new as of Postgres 9.0. It takes
6 : * care of sending XLOG from the primary server to a single recipient.
7 : * (Note that there can be more than one walsender process concurrently.)
8 : * It is started by the postmaster when the walreceiver of a standby server
9 : * connects to the primary server and requests XLOG streaming replication.
10 : *
11 : * A walsender is similar to a regular backend, ie. there is a one-to-one
12 : * relationship between a connection and a walsender process, but instead
13 : * of processing SQL queries, it understands a small set of special
14 : * replication-mode commands. The START_REPLICATION command begins streaming
15 : * WAL to the client. While streaming, the walsender keeps reading XLOG
16 : * records from the disk and sends them to the standby server over the
17 : * COPY protocol, until either side ends the replication by exiting COPY
18 : * mode (or until the connection is closed).
19 : *
20 : * Normal termination is by SIGTERM, which instructs the walsender to
21 : * close the connection and exit(0) at the next convenient moment. Emergency
22 : * termination is by SIGQUIT; like any backend, the walsender will simply
23 : * abort and exit on SIGQUIT. A close of the connection and a FATAL error
24 : * are treated as not a crash but approximately normal termination;
25 : * the walsender will exit quickly without sending any more XLOG records.
26 : *
27 : * If the server is shut down, checkpointer sends us
28 : * PROCSIG_WALSND_INIT_STOPPING after all regular backends have exited. If
29 : * the backend is idle or runs an SQL query this causes the backend to
30 : * shutdown, if logical replication is in progress all existing WAL records
31 : * are processed followed by a shutdown. Otherwise this causes the walsender
32 : * to switch to the "stopping" state. In this state, the walsender will reject
33 : * any further replication commands. The checkpointer begins the shutdown
34 : * checkpoint once all walsenders are confirmed as stopping. When the shutdown
35 : * checkpoint finishes, the postmaster sends us SIGUSR2. This instructs
36 : * walsender to send any outstanding WAL, including the shutdown checkpoint
37 : * record, wait for it to be replicated to the standby, and then exit.
38 : *
39 : *
40 : * Portions Copyright (c) 2010-2025, PostgreSQL Global Development Group
41 : *
42 : * IDENTIFICATION
43 : * src/backend/replication/walsender.c
44 : *
45 : *-------------------------------------------------------------------------
46 : */
47 : #include "postgres.h"
48 :
49 : #include <signal.h>
50 : #include <unistd.h>
51 :
52 : #include "access/timeline.h"
53 : #include "access/transam.h"
54 : #include "access/twophase.h"
55 : #include "access/xact.h"
56 : #include "access/xlog_internal.h"
57 : #include "access/xlogreader.h"
58 : #include "access/xlogrecovery.h"
59 : #include "access/xlogutils.h"
60 : #include "backup/basebackup.h"
61 : #include "backup/basebackup_incremental.h"
62 : #include "catalog/pg_authid.h"
63 : #include "catalog/pg_type.h"
64 : #include "commands/defrem.h"
65 : #include "funcapi.h"
66 : #include "libpq/libpq.h"
67 : #include "libpq/pqformat.h"
68 : #include "libpq/protocol.h"
69 : #include "miscadmin.h"
70 : #include "nodes/replnodes.h"
71 : #include "pgstat.h"
72 : #include "postmaster/interrupt.h"
73 : #include "replication/decode.h"
74 : #include "replication/logical.h"
75 : #include "replication/slotsync.h"
76 : #include "replication/slot.h"
77 : #include "replication/snapbuild.h"
78 : #include "replication/syncrep.h"
79 : #include "replication/walreceiver.h"
80 : #include "replication/walsender.h"
81 : #include "replication/walsender_private.h"
82 : #include "storage/condition_variable.h"
83 : #include "storage/aio_subsys.h"
84 : #include "storage/fd.h"
85 : #include "storage/ipc.h"
86 : #include "storage/pmsignal.h"
87 : #include "storage/proc.h"
88 : #include "storage/procarray.h"
89 : #include "tcop/dest.h"
90 : #include "tcop/tcopprot.h"
91 : #include "utils/acl.h"
92 : #include "utils/builtins.h"
93 : #include "utils/guc.h"
94 : #include "utils/lsyscache.h"
95 : #include "utils/memutils.h"
96 : #include "utils/pg_lsn.h"
97 : #include "utils/pgstat_internal.h"
98 : #include "utils/ps_status.h"
99 : #include "utils/timeout.h"
100 : #include "utils/timestamp.h"
101 :
102 : /* Minimum interval used by walsender for stats flushes, in ms */
103 : #define WALSENDER_STATS_FLUSH_INTERVAL 1000
104 :
105 : /*
106 : * Maximum data payload in a WAL data message. Must be >= XLOG_BLCKSZ.
107 : *
108 : * We don't have a good idea of what a good value would be; there's some
109 : * overhead per message in both walsender and walreceiver, but on the other
110 : * hand sending large batches makes walsender less responsive to signals
111 : * because signals are checked only between messages. 128kB (with
112 : * default 8k blocks) seems like a reasonable guess for now.
113 : */
114 : #define MAX_SEND_SIZE (XLOG_BLCKSZ * 16)
115 :
116 : /* Array of WalSnds in shared memory */
117 : WalSndCtlData *WalSndCtl = NULL;
118 :
119 : /* My slot in the shared memory array */
120 : WalSnd *MyWalSnd = NULL;
121 :
122 : /* Global state */
123 : bool am_walsender = false; /* Am I a walsender process? */
124 : bool am_cascading_walsender = false; /* Am I cascading WAL to another
125 : * standby? */
126 : bool am_db_walsender = false; /* Connected to a database? */
127 :
128 : /* GUC variables */
129 : int max_wal_senders = 10; /* the maximum number of concurrent
130 : * walsenders */
131 : int wal_sender_timeout = 60 * 1000; /* maximum time to send one WAL
132 : * data message */
133 : bool log_replication_commands = false;
134 :
135 : /*
136 : * State for WalSndWakeupRequest
137 : */
138 : bool wake_wal_senders = false;
139 :
140 : /*
141 : * xlogreader used for replication. Note that a WAL sender doing physical
142 : * replication does not need xlogreader to read WAL, but it needs one to
143 : * keep a state of its work.
144 : */
145 : static XLogReaderState *xlogreader = NULL;
146 :
147 : /*
148 : * If the UPLOAD_MANIFEST command is used to provide a backup manifest in
149 : * preparation for an incremental backup, uploaded_manifest will be point
150 : * to an object containing information about its contexts, and
151 : * uploaded_manifest_mcxt will point to the memory context that contains
152 : * that object and all of its subordinate data. Otherwise, both values will
153 : * be NULL.
154 : */
155 : static IncrementalBackupInfo *uploaded_manifest = NULL;
156 : static MemoryContext uploaded_manifest_mcxt = NULL;
157 :
158 : /*
159 : * These variables keep track of the state of the timeline we're currently
160 : * sending. sendTimeLine identifies the timeline. If sendTimeLineIsHistoric,
161 : * the timeline is not the latest timeline on this server, and the server's
162 : * history forked off from that timeline at sendTimeLineValidUpto.
163 : */
164 : static TimeLineID sendTimeLine = 0;
165 : static TimeLineID sendTimeLineNextTLI = 0;
166 : static bool sendTimeLineIsHistoric = false;
167 : static XLogRecPtr sendTimeLineValidUpto = InvalidXLogRecPtr;
168 :
169 : /*
170 : * How far have we sent WAL already? This is also advertised in
171 : * MyWalSnd->sentPtr. (Actually, this is the next WAL location to send.)
172 : */
173 : static XLogRecPtr sentPtr = InvalidXLogRecPtr;
174 :
175 : /* Buffers for constructing outgoing messages and processing reply messages. */
176 : static StringInfoData output_message;
177 : static StringInfoData reply_message;
178 : static StringInfoData tmpbuf;
179 :
180 : /* Timestamp of last ProcessRepliesIfAny(). */
181 : static TimestampTz last_processing = 0;
182 :
183 : /*
184 : * Timestamp of last ProcessRepliesIfAny() that saw a reply from the
185 : * standby. Set to 0 if wal_sender_timeout doesn't need to be active.
186 : */
187 : static TimestampTz last_reply_timestamp = 0;
188 :
189 : /* Have we sent a heartbeat message asking for reply, since last reply? */
190 : static bool waiting_for_ping_response = false;
191 :
192 : /*
193 : * While streaming WAL in Copy mode, streamingDoneSending is set to true
194 : * after we have sent CopyDone. We should not send any more CopyData messages
195 : * after that. streamingDoneReceiving is set to true when we receive CopyDone
196 : * from the other end. When both become true, it's time to exit Copy mode.
197 : */
198 : static bool streamingDoneSending;
199 : static bool streamingDoneReceiving;
200 :
201 : /* Are we there yet? */
202 : static bool WalSndCaughtUp = false;
203 :
204 : /* Flags set by signal handlers for later service in main loop */
205 : static volatile sig_atomic_t got_SIGUSR2 = false;
206 : static volatile sig_atomic_t got_STOPPING = false;
207 :
208 : /*
209 : * This is set while we are streaming. When not set
210 : * PROCSIG_WALSND_INIT_STOPPING signal will be handled like SIGTERM. When set,
211 : * the main loop is responsible for checking got_STOPPING and terminating when
212 : * it's set (after streaming any remaining WAL).
213 : */
214 : static volatile sig_atomic_t replication_active = false;
215 :
216 : static LogicalDecodingContext *logical_decoding_ctx = NULL;
217 :
218 : /* A sample associating a WAL location with the time it was written. */
219 : typedef struct
220 : {
221 : XLogRecPtr lsn;
222 : TimestampTz time;
223 : } WalTimeSample;
224 :
225 : /* The size of our buffer of time samples. */
226 : #define LAG_TRACKER_BUFFER_SIZE 8192
227 :
228 : /* A mechanism for tracking replication lag. */
229 : typedef struct
230 : {
231 : XLogRecPtr last_lsn;
232 : WalTimeSample buffer[LAG_TRACKER_BUFFER_SIZE];
233 : int write_head;
234 : int read_heads[NUM_SYNC_REP_WAIT_MODE];
235 : WalTimeSample last_read[NUM_SYNC_REP_WAIT_MODE];
236 :
237 : /*
238 : * Overflow entries for read heads that collide with the write head.
239 : *
240 : * When the cyclic buffer fills (write head is about to collide with a
241 : * read head), we save that read head's current sample here and mark it as
242 : * using overflow (read_heads[i] = -1). This allows the write head to
243 : * continue advancing while the overflowed mode continues lag computation
244 : * using the saved sample.
245 : *
246 : * Once the standby's reported LSN advances past the overflow entry's LSN,
247 : * we transition back to normal buffer-based tracking.
248 : */
249 : WalTimeSample overflowed[NUM_SYNC_REP_WAIT_MODE];
250 : } LagTracker;
251 :
252 : static LagTracker *lag_tracker;
253 :
254 : /* Signal handlers */
255 : static void WalSndLastCycleHandler(SIGNAL_ARGS);
256 :
257 : /* Prototypes for private functions */
258 : typedef void (*WalSndSendDataCallback) (void);
259 : static void WalSndLoop(WalSndSendDataCallback send_data);
260 : static void InitWalSenderSlot(void);
261 : static void WalSndKill(int code, Datum arg);
262 : pg_noreturn static void WalSndShutdown(void);
263 : static void XLogSendPhysical(void);
264 : static void XLogSendLogical(void);
265 : static void WalSndDone(WalSndSendDataCallback send_data);
266 : static void IdentifySystem(void);
267 : static void UploadManifest(void);
268 : static bool HandleUploadManifestPacket(StringInfo buf, off_t *offset,
269 : IncrementalBackupInfo *ib);
270 : static void ReadReplicationSlot(ReadReplicationSlotCmd *cmd);
271 : static void CreateReplicationSlot(CreateReplicationSlotCmd *cmd);
272 : static void DropReplicationSlot(DropReplicationSlotCmd *cmd);
273 : static void StartReplication(StartReplicationCmd *cmd);
274 : static void StartLogicalReplication(StartReplicationCmd *cmd);
275 : static void ProcessStandbyMessage(void);
276 : static void ProcessStandbyReplyMessage(void);
277 : static void ProcessStandbyHSFeedbackMessage(void);
278 : static void ProcessStandbyPSRequestMessage(void);
279 : static void ProcessRepliesIfAny(void);
280 : static void ProcessPendingWrites(void);
281 : static void WalSndKeepalive(bool requestReply, XLogRecPtr writePtr);
282 : static void WalSndKeepaliveIfNecessary(void);
283 : static void WalSndCheckTimeOut(void);
284 : static long WalSndComputeSleeptime(TimestampTz now);
285 : static void WalSndWait(uint32 socket_events, long timeout, uint32 wait_event);
286 : static void WalSndPrepareWrite(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid, bool last_write);
287 : static void WalSndWriteData(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid, bool last_write);
288 : static void WalSndUpdateProgress(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid,
289 : bool skipped_xact);
290 : static XLogRecPtr WalSndWaitForWal(XLogRecPtr loc);
291 : static void LagTrackerWrite(XLogRecPtr lsn, TimestampTz local_flush_time);
292 : static TimeOffset LagTrackerRead(int head, XLogRecPtr lsn, TimestampTz now);
293 : static bool TransactionIdInRecentPast(TransactionId xid, uint32 epoch);
294 :
295 : static void WalSndSegmentOpen(XLogReaderState *state, XLogSegNo nextSegNo,
296 : TimeLineID *tli_p);
297 :
298 :
299 : /* Initialize walsender process before entering the main command loop */
300 : void
301 2318 : InitWalSender(void)
302 : {
303 2318 : am_cascading_walsender = RecoveryInProgress();
304 :
305 : /* Create a per-walsender data structure in shared memory */
306 2318 : InitWalSenderSlot();
307 :
308 : /* need resource owner for e.g. basebackups */
309 2318 : CreateAuxProcessResourceOwner();
310 :
311 : /*
312 : * Let postmaster know that we're a WAL sender. Once we've declared us as
313 : * a WAL sender process, postmaster will let us outlive the bgwriter and
314 : * kill us last in the shutdown sequence, so we get a chance to stream all
315 : * remaining WAL at shutdown, including the shutdown checkpoint. Note that
316 : * there's no going back, and we mustn't write any WAL records after this.
317 : */
318 2318 : MarkPostmasterChildWalSender();
319 2318 : SendPostmasterSignal(PMSIGNAL_ADVANCE_STATE_MACHINE);
320 :
321 : /*
322 : * If the client didn't specify a database to connect to, show in PGPROC
323 : * that our advertised xmin should affect vacuum horizons in all
324 : * databases. This allows physical replication clients to send hot
325 : * standby feedback that will delay vacuum cleanup in all databases.
326 : */
327 2318 : if (MyDatabaseId == InvalidOid)
328 : {
329 : Assert(MyProc->xmin == InvalidTransactionId);
330 898 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
331 898 : MyProc->statusFlags |= PROC_AFFECTS_ALL_HORIZONS;
332 898 : ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
333 898 : LWLockRelease(ProcArrayLock);
334 : }
335 :
336 : /* Initialize empty timestamp buffer for lag tracking. */
337 2318 : lag_tracker = MemoryContextAllocZero(TopMemoryContext, sizeof(LagTracker));
338 2318 : }
339 :
340 : /*
341 : * Clean up after an error.
342 : *
343 : * WAL sender processes don't use transactions like regular backends do.
344 : * This function does any cleanup required after an error in a WAL sender
345 : * process, similar to what transaction abort does in a regular backend.
346 : */
347 : void
348 98 : WalSndErrorCleanup(void)
349 : {
350 98 : LWLockReleaseAll();
351 98 : ConditionVariableCancelSleep();
352 98 : pgstat_report_wait_end();
353 98 : pgaio_error_cleanup();
354 :
355 98 : if (xlogreader != NULL && xlogreader->seg.ws_file >= 0)
356 12 : wal_segment_close(xlogreader);
357 :
358 98 : if (MyReplicationSlot != NULL)
359 30 : ReplicationSlotRelease();
360 :
361 98 : ReplicationSlotCleanup(false);
362 :
363 98 : replication_active = false;
364 :
365 : /*
366 : * If there is a transaction in progress, it will clean up our
367 : * ResourceOwner, but if a replication command set up a resource owner
368 : * without a transaction, we've got to clean that up now.
369 : */
370 98 : if (!IsTransactionOrTransactionBlock())
371 96 : ReleaseAuxProcessResources(false);
372 :
373 98 : if (got_STOPPING || got_SIGUSR2)
374 0 : proc_exit(0);
375 :
376 : /* Revert back to startup state */
377 98 : WalSndSetState(WALSNDSTATE_STARTUP);
378 98 : }
379 :
380 : /*
381 : * Handle a client's connection abort in an orderly manner.
382 : */
383 : static void
384 34 : WalSndShutdown(void)
385 : {
386 : /*
387 : * Reset whereToSendOutput to prevent ereport from attempting to send any
388 : * more messages to the standby.
389 : */
390 34 : if (whereToSendOutput == DestRemote)
391 34 : whereToSendOutput = DestNone;
392 :
393 34 : proc_exit(0);
394 : abort(); /* keep the compiler quiet */
395 : }
396 :
397 : /*
398 : * Handle the IDENTIFY_SYSTEM command.
399 : */
400 : static void
401 1416 : IdentifySystem(void)
402 : {
403 : char sysid[32];
404 : char xloc[MAXFNAMELEN];
405 : XLogRecPtr logptr;
406 1416 : char *dbname = NULL;
407 : DestReceiver *dest;
408 : TupOutputState *tstate;
409 : TupleDesc tupdesc;
410 : Datum values[4];
411 1416 : bool nulls[4] = {0};
412 : TimeLineID currTLI;
413 :
414 : /*
415 : * Reply with a result set with one row, four columns. First col is system
416 : * ID, second is timeline ID, third is current xlog location and the
417 : * fourth contains the database name if we are connected to one.
418 : */
419 :
420 1416 : snprintf(sysid, sizeof(sysid), UINT64_FORMAT,
421 : GetSystemIdentifier());
422 :
423 1416 : am_cascading_walsender = RecoveryInProgress();
424 1416 : if (am_cascading_walsender)
425 112 : logptr = GetStandbyFlushRecPtr(&currTLI);
426 : else
427 1304 : logptr = GetFlushRecPtr(&currTLI);
428 :
429 1416 : snprintf(xloc, sizeof(xloc), "%X/%08X", LSN_FORMAT_ARGS(logptr));
430 :
431 1416 : if (MyDatabaseId != InvalidOid)
432 : {
433 498 : MemoryContext cur = CurrentMemoryContext;
434 :
435 : /* syscache access needs a transaction env. */
436 498 : StartTransactionCommand();
437 498 : dbname = get_database_name(MyDatabaseId);
438 : /* copy dbname out of TX context */
439 498 : dbname = MemoryContextStrdup(cur, dbname);
440 498 : CommitTransactionCommand();
441 : }
442 :
443 1416 : dest = CreateDestReceiver(DestRemoteSimple);
444 :
445 : /* need a tuple descriptor representing four columns */
446 1416 : tupdesc = CreateTemplateTupleDesc(4);
447 1416 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "systemid",
448 : TEXTOID, -1, 0);
449 1416 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "timeline",
450 : INT8OID, -1, 0);
451 1416 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 3, "xlogpos",
452 : TEXTOID, -1, 0);
453 1416 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 4, "dbname",
454 : TEXTOID, -1, 0);
455 :
456 : /* prepare for projection of tuples */
457 1416 : tstate = begin_tup_output_tupdesc(dest, tupdesc, &TTSOpsVirtual);
458 :
459 : /* column 1: system identifier */
460 1416 : values[0] = CStringGetTextDatum(sysid);
461 :
462 : /* column 2: timeline */
463 1416 : values[1] = Int64GetDatum(currTLI);
464 :
465 : /* column 3: wal location */
466 1416 : values[2] = CStringGetTextDatum(xloc);
467 :
468 : /* column 4: database name, or NULL if none */
469 1416 : if (dbname)
470 498 : values[3] = CStringGetTextDatum(dbname);
471 : else
472 918 : nulls[3] = true;
473 :
474 : /* send it to dest */
475 1416 : do_tup_output(tstate, values, nulls);
476 :
477 1416 : end_tup_output(tstate);
478 1416 : }
479 :
480 : /* Handle READ_REPLICATION_SLOT command */
481 : static void
482 12 : ReadReplicationSlot(ReadReplicationSlotCmd *cmd)
483 : {
484 : #define READ_REPLICATION_SLOT_COLS 3
485 : ReplicationSlot *slot;
486 : DestReceiver *dest;
487 : TupOutputState *tstate;
488 : TupleDesc tupdesc;
489 12 : Datum values[READ_REPLICATION_SLOT_COLS] = {0};
490 : bool nulls[READ_REPLICATION_SLOT_COLS];
491 :
492 12 : tupdesc = CreateTemplateTupleDesc(READ_REPLICATION_SLOT_COLS);
493 12 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "slot_type",
494 : TEXTOID, -1, 0);
495 12 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "restart_lsn",
496 : TEXTOID, -1, 0);
497 : /* TimeLineID is unsigned, so int4 is not wide enough. */
498 12 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 3, "restart_tli",
499 : INT8OID, -1, 0);
500 :
501 12 : memset(nulls, true, READ_REPLICATION_SLOT_COLS * sizeof(bool));
502 :
503 12 : LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
504 12 : slot = SearchNamedReplicationSlot(cmd->slotname, false);
505 12 : if (slot == NULL || !slot->in_use)
506 : {
507 4 : LWLockRelease(ReplicationSlotControlLock);
508 : }
509 : else
510 : {
511 : ReplicationSlot slot_contents;
512 8 : int i = 0;
513 :
514 : /* Copy slot contents while holding spinlock */
515 8 : SpinLockAcquire(&slot->mutex);
516 8 : slot_contents = *slot;
517 8 : SpinLockRelease(&slot->mutex);
518 8 : LWLockRelease(ReplicationSlotControlLock);
519 :
520 8 : if (OidIsValid(slot_contents.data.database))
521 2 : ereport(ERROR,
522 : errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
523 : errmsg("cannot use %s with a logical replication slot",
524 : "READ_REPLICATION_SLOT"));
525 :
526 : /* slot type */
527 6 : values[i] = CStringGetTextDatum("physical");
528 6 : nulls[i] = false;
529 6 : i++;
530 :
531 : /* start LSN */
532 6 : if (XLogRecPtrIsValid(slot_contents.data.restart_lsn))
533 : {
534 : char xloc[64];
535 :
536 6 : snprintf(xloc, sizeof(xloc), "%X/%08X",
537 6 : LSN_FORMAT_ARGS(slot_contents.data.restart_lsn));
538 6 : values[i] = CStringGetTextDatum(xloc);
539 6 : nulls[i] = false;
540 : }
541 6 : i++;
542 :
543 : /* timeline this WAL was produced on */
544 6 : if (XLogRecPtrIsValid(slot_contents.data.restart_lsn))
545 : {
546 : TimeLineID slots_position_timeline;
547 : TimeLineID current_timeline;
548 6 : List *timeline_history = NIL;
549 :
550 : /*
551 : * While in recovery, use as timeline the currently-replaying one
552 : * to get the LSN position's history.
553 : */
554 6 : if (RecoveryInProgress())
555 0 : (void) GetXLogReplayRecPtr(¤t_timeline);
556 : else
557 6 : current_timeline = GetWALInsertionTimeLine();
558 :
559 6 : timeline_history = readTimeLineHistory(current_timeline);
560 6 : slots_position_timeline = tliOfPointInHistory(slot_contents.data.restart_lsn,
561 : timeline_history);
562 6 : values[i] = Int64GetDatum((int64) slots_position_timeline);
563 6 : nulls[i] = false;
564 : }
565 6 : i++;
566 :
567 : Assert(i == READ_REPLICATION_SLOT_COLS);
568 : }
569 :
570 10 : dest = CreateDestReceiver(DestRemoteSimple);
571 10 : tstate = begin_tup_output_tupdesc(dest, tupdesc, &TTSOpsVirtual);
572 10 : do_tup_output(tstate, values, nulls);
573 10 : end_tup_output(tstate);
574 10 : }
575 :
576 :
577 : /*
578 : * Handle TIMELINE_HISTORY command.
579 : */
580 : static void
581 22 : SendTimeLineHistory(TimeLineHistoryCmd *cmd)
582 : {
583 : DestReceiver *dest;
584 : TupleDesc tupdesc;
585 : StringInfoData buf;
586 : char histfname[MAXFNAMELEN];
587 : char path[MAXPGPATH];
588 : int fd;
589 : off_t histfilelen;
590 : off_t bytesleft;
591 : Size len;
592 :
593 22 : dest = CreateDestReceiver(DestRemoteSimple);
594 :
595 : /*
596 : * Reply with a result set with one row, and two columns. The first col is
597 : * the name of the history file, 2nd is the contents.
598 : */
599 22 : tupdesc = CreateTemplateTupleDesc(2);
600 22 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "filename", TEXTOID, -1, 0);
601 22 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "content", TEXTOID, -1, 0);
602 :
603 22 : TLHistoryFileName(histfname, cmd->timeline);
604 22 : TLHistoryFilePath(path, cmd->timeline);
605 :
606 : /* Send a RowDescription message */
607 22 : dest->rStartup(dest, CMD_SELECT, tupdesc);
608 :
609 : /* Send a DataRow message */
610 22 : pq_beginmessage(&buf, PqMsg_DataRow);
611 22 : pq_sendint16(&buf, 2); /* # of columns */
612 22 : len = strlen(histfname);
613 22 : pq_sendint32(&buf, len); /* col1 len */
614 22 : pq_sendbytes(&buf, histfname, len);
615 :
616 22 : fd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
617 22 : if (fd < 0)
618 0 : ereport(ERROR,
619 : (errcode_for_file_access(),
620 : errmsg("could not open file \"%s\": %m", path)));
621 :
622 : /* Determine file length and send it to client */
623 22 : histfilelen = lseek(fd, 0, SEEK_END);
624 22 : if (histfilelen < 0)
625 0 : ereport(ERROR,
626 : (errcode_for_file_access(),
627 : errmsg("could not seek to end of file \"%s\": %m", path)));
628 22 : if (lseek(fd, 0, SEEK_SET) != 0)
629 0 : ereport(ERROR,
630 : (errcode_for_file_access(),
631 : errmsg("could not seek to beginning of file \"%s\": %m", path)));
632 :
633 22 : pq_sendint32(&buf, histfilelen); /* col2 len */
634 :
635 22 : bytesleft = histfilelen;
636 44 : while (bytesleft > 0)
637 : {
638 : PGAlignedBlock rbuf;
639 : int nread;
640 :
641 22 : pgstat_report_wait_start(WAIT_EVENT_WALSENDER_TIMELINE_HISTORY_READ);
642 22 : nread = read(fd, rbuf.data, sizeof(rbuf));
643 22 : pgstat_report_wait_end();
644 22 : if (nread < 0)
645 0 : ereport(ERROR,
646 : (errcode_for_file_access(),
647 : errmsg("could not read file \"%s\": %m",
648 : path)));
649 22 : else if (nread == 0)
650 0 : ereport(ERROR,
651 : (errcode(ERRCODE_DATA_CORRUPTED),
652 : errmsg("could not read file \"%s\": read %d of %zu",
653 : path, nread, (Size) bytesleft)));
654 :
655 22 : pq_sendbytes(&buf, rbuf.data, nread);
656 22 : bytesleft -= nread;
657 : }
658 :
659 22 : if (CloseTransientFile(fd) != 0)
660 0 : ereport(ERROR,
661 : (errcode_for_file_access(),
662 : errmsg("could not close file \"%s\": %m", path)));
663 :
664 22 : pq_endmessage(&buf);
665 22 : }
666 :
667 : /*
668 : * Handle UPLOAD_MANIFEST command.
669 : */
670 : static void
671 22 : UploadManifest(void)
672 : {
673 : MemoryContext mcxt;
674 : IncrementalBackupInfo *ib;
675 22 : off_t offset = 0;
676 : StringInfoData buf;
677 :
678 : /*
679 : * parsing the manifest will use the cryptohash stuff, which requires a
680 : * resource owner
681 : */
682 : Assert(AuxProcessResourceOwner != NULL);
683 : Assert(CurrentResourceOwner == AuxProcessResourceOwner ||
684 : CurrentResourceOwner == NULL);
685 22 : CurrentResourceOwner = AuxProcessResourceOwner;
686 :
687 : /* Prepare to read manifest data into a temporary context. */
688 22 : mcxt = AllocSetContextCreate(CurrentMemoryContext,
689 : "incremental backup information",
690 : ALLOCSET_DEFAULT_SIZES);
691 22 : ib = CreateIncrementalBackupInfo(mcxt);
692 :
693 : /* Send a CopyInResponse message */
694 22 : pq_beginmessage(&buf, PqMsg_CopyInResponse);
695 22 : pq_sendbyte(&buf, 0);
696 22 : pq_sendint16(&buf, 0);
697 22 : pq_endmessage_reuse(&buf);
698 22 : pq_flush();
699 :
700 : /* Receive packets from client until done. */
701 86 : while (HandleUploadManifestPacket(&buf, &offset, ib))
702 : ;
703 :
704 : /* Finish up manifest processing. */
705 20 : FinalizeIncrementalManifest(ib);
706 :
707 : /*
708 : * Discard any old manifest information and arrange to preserve the new
709 : * information we just got.
710 : *
711 : * We assume that MemoryContextDelete and MemoryContextSetParent won't
712 : * fail, and thus we shouldn't end up bailing out of here in such a way as
713 : * to leave dangling pointers.
714 : */
715 20 : if (uploaded_manifest_mcxt != NULL)
716 0 : MemoryContextDelete(uploaded_manifest_mcxt);
717 20 : MemoryContextSetParent(mcxt, CacheMemoryContext);
718 20 : uploaded_manifest = ib;
719 20 : uploaded_manifest_mcxt = mcxt;
720 :
721 : /* clean up the resource owner we created */
722 20 : ReleaseAuxProcessResources(true);
723 20 : }
724 :
725 : /*
726 : * Process one packet received during the handling of an UPLOAD_MANIFEST
727 : * operation.
728 : *
729 : * 'buf' is scratch space. This function expects it to be initialized, doesn't
730 : * care what the current contents are, and may override them with completely
731 : * new contents.
732 : *
733 : * The return value is true if the caller should continue processing
734 : * additional packets and false if the UPLOAD_MANIFEST operation is complete.
735 : */
736 : static bool
737 86 : HandleUploadManifestPacket(StringInfo buf, off_t *offset,
738 : IncrementalBackupInfo *ib)
739 : {
740 : int mtype;
741 : int maxmsglen;
742 :
743 86 : HOLD_CANCEL_INTERRUPTS();
744 :
745 86 : pq_startmsgread();
746 86 : mtype = pq_getbyte();
747 86 : if (mtype == EOF)
748 0 : ereport(ERROR,
749 : (errcode(ERRCODE_CONNECTION_FAILURE),
750 : errmsg("unexpected EOF on client connection with an open transaction")));
751 :
752 86 : switch (mtype)
753 : {
754 66 : case PqMsg_CopyData:
755 66 : maxmsglen = PQ_LARGE_MESSAGE_LIMIT;
756 66 : break;
757 20 : case PqMsg_CopyDone:
758 : case PqMsg_CopyFail:
759 : case PqMsg_Flush:
760 : case PqMsg_Sync:
761 20 : maxmsglen = PQ_SMALL_MESSAGE_LIMIT;
762 20 : break;
763 0 : default:
764 0 : ereport(ERROR,
765 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
766 : errmsg("unexpected message type 0x%02X during COPY from stdin",
767 : mtype)));
768 : maxmsglen = 0; /* keep compiler quiet */
769 : break;
770 : }
771 :
772 : /* Now collect the message body */
773 86 : if (pq_getmessage(buf, maxmsglen))
774 0 : ereport(ERROR,
775 : (errcode(ERRCODE_CONNECTION_FAILURE),
776 : errmsg("unexpected EOF on client connection with an open transaction")));
777 86 : RESUME_CANCEL_INTERRUPTS();
778 :
779 : /* Process the message */
780 86 : switch (mtype)
781 : {
782 66 : case PqMsg_CopyData:
783 66 : AppendIncrementalManifestData(ib, buf->data, buf->len);
784 64 : return true;
785 :
786 20 : case PqMsg_CopyDone:
787 20 : return false;
788 :
789 0 : case PqMsg_Sync:
790 : case PqMsg_Flush:
791 : /* Ignore these while in CopyOut mode as we do elsewhere. */
792 0 : return true;
793 :
794 0 : case PqMsg_CopyFail:
795 0 : ereport(ERROR,
796 : (errcode(ERRCODE_QUERY_CANCELED),
797 : errmsg("COPY from stdin failed: %s",
798 : pq_getmsgstring(buf))));
799 : }
800 :
801 : /* Not reached. */
802 : Assert(false);
803 0 : return false;
804 : }
805 :
806 : /*
807 : * Handle START_REPLICATION command.
808 : *
809 : * At the moment, this never returns, but an ereport(ERROR) will take us back
810 : * to the main loop.
811 : */
812 : static void
813 526 : StartReplication(StartReplicationCmd *cmd)
814 : {
815 : StringInfoData buf;
816 : XLogRecPtr FlushPtr;
817 : TimeLineID FlushTLI;
818 :
819 : /* create xlogreader for physical replication */
820 526 : xlogreader =
821 526 : XLogReaderAllocate(wal_segment_size, NULL,
822 526 : XL_ROUTINE(.segment_open = WalSndSegmentOpen,
823 : .segment_close = wal_segment_close),
824 : NULL);
825 :
826 526 : if (!xlogreader)
827 0 : ereport(ERROR,
828 : (errcode(ERRCODE_OUT_OF_MEMORY),
829 : errmsg("out of memory"),
830 : errdetail("Failed while allocating a WAL reading processor.")));
831 :
832 : /*
833 : * We assume here that we're logging enough information in the WAL for
834 : * log-shipping, since this is checked in PostmasterMain().
835 : *
836 : * NOTE: wal_level can only change at shutdown, so in most cases it is
837 : * difficult for there to be WAL data that we can still see that was
838 : * written at wal_level='minimal'.
839 : */
840 :
841 526 : if (cmd->slotname)
842 : {
843 362 : ReplicationSlotAcquire(cmd->slotname, true, true);
844 358 : if (SlotIsLogical(MyReplicationSlot))
845 0 : ereport(ERROR,
846 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
847 : errmsg("cannot use a logical replication slot for physical replication")));
848 :
849 : /*
850 : * We don't need to verify the slot's restart_lsn here; instead we
851 : * rely on the caller requesting the starting point to use. If the
852 : * WAL segment doesn't exist, we'll fail later.
853 : */
854 : }
855 :
856 : /*
857 : * Select the timeline. If it was given explicitly by the client, use
858 : * that. Otherwise use the timeline of the last replayed record.
859 : */
860 522 : am_cascading_walsender = RecoveryInProgress();
861 522 : if (am_cascading_walsender)
862 22 : FlushPtr = GetStandbyFlushRecPtr(&FlushTLI);
863 : else
864 500 : FlushPtr = GetFlushRecPtr(&FlushTLI);
865 :
866 522 : if (cmd->timeline != 0)
867 : {
868 : XLogRecPtr switchpoint;
869 :
870 520 : sendTimeLine = cmd->timeline;
871 520 : if (sendTimeLine == FlushTLI)
872 : {
873 502 : sendTimeLineIsHistoric = false;
874 502 : sendTimeLineValidUpto = InvalidXLogRecPtr;
875 : }
876 : else
877 : {
878 : List *timeLineHistory;
879 :
880 18 : sendTimeLineIsHistoric = true;
881 :
882 : /*
883 : * Check that the timeline the client requested exists, and the
884 : * requested start location is on that timeline.
885 : */
886 18 : timeLineHistory = readTimeLineHistory(FlushTLI);
887 18 : switchpoint = tliSwitchPoint(cmd->timeline, timeLineHistory,
888 : &sendTimeLineNextTLI);
889 18 : list_free_deep(timeLineHistory);
890 :
891 : /*
892 : * Found the requested timeline in the history. Check that
893 : * requested startpoint is on that timeline in our history.
894 : *
895 : * This is quite loose on purpose. We only check that we didn't
896 : * fork off the requested timeline before the switchpoint. We
897 : * don't check that we switched *to* it before the requested
898 : * starting point. This is because the client can legitimately
899 : * request to start replication from the beginning of the WAL
900 : * segment that contains switchpoint, but on the new timeline, so
901 : * that it doesn't end up with a partial segment. If you ask for
902 : * too old a starting point, you'll get an error later when we
903 : * fail to find the requested WAL segment in pg_wal.
904 : *
905 : * XXX: we could be more strict here and only allow a startpoint
906 : * that's older than the switchpoint, if it's still in the same
907 : * WAL segment.
908 : */
909 18 : if (XLogRecPtrIsValid(switchpoint) &&
910 18 : switchpoint < cmd->startpoint)
911 : {
912 0 : ereport(ERROR,
913 : errmsg("requested starting point %X/%08X on timeline %u is not in this server's history",
914 : LSN_FORMAT_ARGS(cmd->startpoint),
915 : cmd->timeline),
916 : errdetail("This server's history forked from timeline %u at %X/%08X.",
917 : cmd->timeline,
918 : LSN_FORMAT_ARGS(switchpoint)));
919 : }
920 18 : sendTimeLineValidUpto = switchpoint;
921 : }
922 : }
923 : else
924 : {
925 2 : sendTimeLine = FlushTLI;
926 2 : sendTimeLineValidUpto = InvalidXLogRecPtr;
927 2 : sendTimeLineIsHistoric = false;
928 : }
929 :
930 522 : streamingDoneSending = streamingDoneReceiving = false;
931 :
932 : /* If there is nothing to stream, don't even enter COPY mode */
933 522 : if (!sendTimeLineIsHistoric || cmd->startpoint < sendTimeLineValidUpto)
934 : {
935 : /*
936 : * When we first start replication the standby will be behind the
937 : * primary. For some applications, for example synchronous
938 : * replication, it is important to have a clear state for this initial
939 : * catchup mode, so we can trigger actions when we change streaming
940 : * state later. We may stay in this state for a long time, which is
941 : * exactly why we want to be able to monitor whether or not we are
942 : * still here.
943 : */
944 522 : WalSndSetState(WALSNDSTATE_CATCHUP);
945 :
946 : /* Send a CopyBothResponse message, and start streaming */
947 522 : pq_beginmessage(&buf, PqMsg_CopyBothResponse);
948 522 : pq_sendbyte(&buf, 0);
949 522 : pq_sendint16(&buf, 0);
950 522 : pq_endmessage(&buf);
951 522 : pq_flush();
952 :
953 : /*
954 : * Don't allow a request to stream from a future point in WAL that
955 : * hasn't been flushed to disk in this server yet.
956 : */
957 522 : if (FlushPtr < cmd->startpoint)
958 : {
959 0 : ereport(ERROR,
960 : errmsg("requested starting point %X/%08X is ahead of the WAL flush position of this server %X/%08X",
961 : LSN_FORMAT_ARGS(cmd->startpoint),
962 : LSN_FORMAT_ARGS(FlushPtr)));
963 : }
964 :
965 : /* Start streaming from the requested point */
966 522 : sentPtr = cmd->startpoint;
967 :
968 : /* Initialize shared memory status, too */
969 522 : SpinLockAcquire(&MyWalSnd->mutex);
970 522 : MyWalSnd->sentPtr = sentPtr;
971 522 : SpinLockRelease(&MyWalSnd->mutex);
972 :
973 522 : SyncRepInitConfig();
974 :
975 : /* Main loop of walsender */
976 522 : replication_active = true;
977 :
978 522 : WalSndLoop(XLogSendPhysical);
979 :
980 294 : replication_active = false;
981 294 : if (got_STOPPING)
982 0 : proc_exit(0);
983 294 : WalSndSetState(WALSNDSTATE_STARTUP);
984 :
985 : Assert(streamingDoneSending && streamingDoneReceiving);
986 : }
987 :
988 294 : if (cmd->slotname)
989 268 : ReplicationSlotRelease();
990 :
991 : /*
992 : * Copy is finished now. Send a single-row result set indicating the next
993 : * timeline.
994 : */
995 294 : if (sendTimeLineIsHistoric)
996 : {
997 : char startpos_str[8 + 1 + 8 + 1];
998 : DestReceiver *dest;
999 : TupOutputState *tstate;
1000 : TupleDesc tupdesc;
1001 : Datum values[2];
1002 18 : bool nulls[2] = {0};
1003 :
1004 18 : snprintf(startpos_str, sizeof(startpos_str), "%X/%08X",
1005 18 : LSN_FORMAT_ARGS(sendTimeLineValidUpto));
1006 :
1007 18 : dest = CreateDestReceiver(DestRemoteSimple);
1008 :
1009 : /*
1010 : * Need a tuple descriptor representing two columns. int8 may seem
1011 : * like a surprising data type for this, but in theory int4 would not
1012 : * be wide enough for this, as TimeLineID is unsigned.
1013 : */
1014 18 : tupdesc = CreateTemplateTupleDesc(2);
1015 18 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "next_tli",
1016 : INT8OID, -1, 0);
1017 18 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "next_tli_startpos",
1018 : TEXTOID, -1, 0);
1019 :
1020 : /* prepare for projection of tuple */
1021 18 : tstate = begin_tup_output_tupdesc(dest, tupdesc, &TTSOpsVirtual);
1022 :
1023 18 : values[0] = Int64GetDatum((int64) sendTimeLineNextTLI);
1024 18 : values[1] = CStringGetTextDatum(startpos_str);
1025 :
1026 : /* send it to dest */
1027 18 : do_tup_output(tstate, values, nulls);
1028 :
1029 18 : end_tup_output(tstate);
1030 : }
1031 :
1032 : /* Send CommandComplete message */
1033 294 : EndReplicationCommand("START_STREAMING");
1034 294 : }
1035 :
1036 : /*
1037 : * XLogReaderRoutine->page_read callback for logical decoding contexts, as a
1038 : * walsender process.
1039 : *
1040 : * Inside the walsender we can do better than read_local_xlog_page,
1041 : * which has to do a plain sleep/busy loop, because the walsender's latch gets
1042 : * set every time WAL is flushed.
1043 : */
1044 : static int
1045 37510 : logical_read_xlog_page(XLogReaderState *state, XLogRecPtr targetPagePtr, int reqLen,
1046 : XLogRecPtr targetRecPtr, char *cur_page)
1047 : {
1048 : XLogRecPtr flushptr;
1049 : int count;
1050 : WALReadError errinfo;
1051 : XLogSegNo segno;
1052 : TimeLineID currTLI;
1053 :
1054 : /*
1055 : * Make sure we have enough WAL available before retrieving the current
1056 : * timeline.
1057 : */
1058 37510 : flushptr = WalSndWaitForWal(targetPagePtr + reqLen);
1059 :
1060 : /* Fail if not enough (implies we are going to shut down) */
1061 37154 : if (flushptr < targetPagePtr + reqLen)
1062 6936 : return -1;
1063 :
1064 : /*
1065 : * Since logical decoding is also permitted on a standby server, we need
1066 : * to check if the server is in recovery to decide how to get the current
1067 : * timeline ID (so that it also covers the promotion or timeline change
1068 : * cases). We must determine am_cascading_walsender after waiting for the
1069 : * required WAL so that it is correct when the walsender wakes up after a
1070 : * promotion.
1071 : */
1072 30218 : am_cascading_walsender = RecoveryInProgress();
1073 :
1074 30218 : if (am_cascading_walsender)
1075 616 : GetXLogReplayRecPtr(&currTLI);
1076 : else
1077 29602 : currTLI = GetWALInsertionTimeLine();
1078 :
1079 30218 : XLogReadDetermineTimeline(state, targetPagePtr, reqLen, currTLI);
1080 30218 : sendTimeLineIsHistoric = (state->currTLI != currTLI);
1081 30218 : sendTimeLine = state->currTLI;
1082 30218 : sendTimeLineValidUpto = state->currTLIValidUntil;
1083 30218 : sendTimeLineNextTLI = state->nextTLI;
1084 :
1085 30218 : if (targetPagePtr + XLOG_BLCKSZ <= flushptr)
1086 26352 : count = XLOG_BLCKSZ; /* more than one block available */
1087 : else
1088 3866 : count = flushptr - targetPagePtr; /* part of the page available */
1089 :
1090 : /* now actually read the data, we know it's there */
1091 30218 : if (!WALRead(state,
1092 : cur_page,
1093 : targetPagePtr,
1094 : count,
1095 : currTLI, /* Pass the current TLI because only
1096 : * WalSndSegmentOpen controls whether new TLI
1097 : * is needed. */
1098 : &errinfo))
1099 0 : WALReadRaiseError(&errinfo);
1100 :
1101 : /*
1102 : * After reading into the buffer, check that what we read was valid. We do
1103 : * this after reading, because even though the segment was present when we
1104 : * opened it, it might get recycled or removed while we read it. The
1105 : * read() succeeds in that case, but the data we tried to read might
1106 : * already have been overwritten with new WAL records.
1107 : */
1108 30218 : XLByteToSeg(targetPagePtr, segno, state->segcxt.ws_segsize);
1109 30218 : CheckXLogRemoved(segno, state->seg.ws_tli);
1110 :
1111 30218 : return count;
1112 : }
1113 :
1114 : /*
1115 : * Process extra options given to CREATE_REPLICATION_SLOT.
1116 : */
1117 : static void
1118 946 : parseCreateReplSlotOptions(CreateReplicationSlotCmd *cmd,
1119 : bool *reserve_wal,
1120 : CRSSnapshotAction *snapshot_action,
1121 : bool *two_phase, bool *failover)
1122 : {
1123 : ListCell *lc;
1124 946 : bool snapshot_action_given = false;
1125 946 : bool reserve_wal_given = false;
1126 946 : bool two_phase_given = false;
1127 946 : bool failover_given = false;
1128 :
1129 : /* Parse options */
1130 1908 : foreach(lc, cmd->options)
1131 : {
1132 962 : DefElem *defel = (DefElem *) lfirst(lc);
1133 :
1134 962 : if (strcmp(defel->defname, "snapshot") == 0)
1135 : {
1136 : char *action;
1137 :
1138 670 : if (snapshot_action_given || cmd->kind != REPLICATION_KIND_LOGICAL)
1139 0 : ereport(ERROR,
1140 : (errcode(ERRCODE_SYNTAX_ERROR),
1141 : errmsg("conflicting or redundant options")));
1142 :
1143 670 : action = defGetString(defel);
1144 670 : snapshot_action_given = true;
1145 :
1146 670 : if (strcmp(action, "export") == 0)
1147 2 : *snapshot_action = CRS_EXPORT_SNAPSHOT;
1148 668 : else if (strcmp(action, "nothing") == 0)
1149 276 : *snapshot_action = CRS_NOEXPORT_SNAPSHOT;
1150 392 : else if (strcmp(action, "use") == 0)
1151 392 : *snapshot_action = CRS_USE_SNAPSHOT;
1152 : else
1153 0 : ereport(ERROR,
1154 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1155 : errmsg("unrecognized value for CREATE_REPLICATION_SLOT option \"%s\": \"%s\"",
1156 : defel->defname, action)));
1157 : }
1158 292 : else if (strcmp(defel->defname, "reserve_wal") == 0)
1159 : {
1160 274 : if (reserve_wal_given || cmd->kind != REPLICATION_KIND_PHYSICAL)
1161 0 : ereport(ERROR,
1162 : (errcode(ERRCODE_SYNTAX_ERROR),
1163 : errmsg("conflicting or redundant options")));
1164 :
1165 274 : reserve_wal_given = true;
1166 274 : *reserve_wal = defGetBoolean(defel);
1167 : }
1168 18 : else if (strcmp(defel->defname, "two_phase") == 0)
1169 : {
1170 4 : if (two_phase_given || cmd->kind != REPLICATION_KIND_LOGICAL)
1171 0 : ereport(ERROR,
1172 : (errcode(ERRCODE_SYNTAX_ERROR),
1173 : errmsg("conflicting or redundant options")));
1174 4 : two_phase_given = true;
1175 4 : *two_phase = defGetBoolean(defel);
1176 : }
1177 14 : else if (strcmp(defel->defname, "failover") == 0)
1178 : {
1179 14 : if (failover_given || cmd->kind != REPLICATION_KIND_LOGICAL)
1180 0 : ereport(ERROR,
1181 : (errcode(ERRCODE_SYNTAX_ERROR),
1182 : errmsg("conflicting or redundant options")));
1183 14 : failover_given = true;
1184 14 : *failover = defGetBoolean(defel);
1185 : }
1186 : else
1187 0 : elog(ERROR, "unrecognized option: %s", defel->defname);
1188 : }
1189 946 : }
1190 :
1191 : /*
1192 : * Create a new replication slot.
1193 : */
1194 : static void
1195 946 : CreateReplicationSlot(CreateReplicationSlotCmd *cmd)
1196 : {
1197 946 : const char *snapshot_name = NULL;
1198 : char xloc[MAXFNAMELEN];
1199 : char *slot_name;
1200 946 : bool reserve_wal = false;
1201 946 : bool two_phase = false;
1202 946 : bool failover = false;
1203 946 : CRSSnapshotAction snapshot_action = CRS_EXPORT_SNAPSHOT;
1204 : DestReceiver *dest;
1205 : TupOutputState *tstate;
1206 : TupleDesc tupdesc;
1207 : Datum values[4];
1208 946 : bool nulls[4] = {0};
1209 :
1210 : Assert(!MyReplicationSlot);
1211 :
1212 946 : parseCreateReplSlotOptions(cmd, &reserve_wal, &snapshot_action, &two_phase,
1213 : &failover);
1214 :
1215 946 : if (cmd->kind == REPLICATION_KIND_PHYSICAL)
1216 : {
1217 276 : ReplicationSlotCreate(cmd->slotname, false,
1218 276 : cmd->temporary ? RS_TEMPORARY : RS_PERSISTENT,
1219 : false, false, false);
1220 :
1221 274 : if (reserve_wal)
1222 : {
1223 272 : ReplicationSlotReserveWal();
1224 :
1225 272 : ReplicationSlotMarkDirty();
1226 :
1227 : /* Write this slot to disk if it's a permanent one. */
1228 272 : if (!cmd->temporary)
1229 6 : ReplicationSlotSave();
1230 : }
1231 : }
1232 : else
1233 : {
1234 : LogicalDecodingContext *ctx;
1235 670 : bool need_full_snapshot = false;
1236 :
1237 : Assert(cmd->kind == REPLICATION_KIND_LOGICAL);
1238 :
1239 670 : CheckLogicalDecodingRequirements();
1240 :
1241 : /*
1242 : * Initially create persistent slot as ephemeral - that allows us to
1243 : * nicely handle errors during initialization because it'll get
1244 : * dropped if this transaction fails. We'll make it persistent at the
1245 : * end. Temporary slots can be created as temporary from beginning as
1246 : * they get dropped on error as well.
1247 : */
1248 670 : ReplicationSlotCreate(cmd->slotname, true,
1249 670 : cmd->temporary ? RS_TEMPORARY : RS_EPHEMERAL,
1250 : two_phase, failover, false);
1251 :
1252 : /*
1253 : * Do options check early so that we can bail before calling the
1254 : * DecodingContextFindStartpoint which can take long time.
1255 : */
1256 670 : if (snapshot_action == CRS_EXPORT_SNAPSHOT)
1257 : {
1258 2 : if (IsTransactionBlock())
1259 0 : ereport(ERROR,
1260 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1261 : (errmsg("%s must not be called inside a transaction",
1262 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'export')")));
1263 :
1264 2 : need_full_snapshot = true;
1265 : }
1266 668 : else if (snapshot_action == CRS_USE_SNAPSHOT)
1267 : {
1268 392 : if (!IsTransactionBlock())
1269 0 : ereport(ERROR,
1270 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1271 : (errmsg("%s must be called inside a transaction",
1272 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1273 :
1274 392 : if (XactIsoLevel != XACT_REPEATABLE_READ)
1275 0 : ereport(ERROR,
1276 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1277 : (errmsg("%s must be called in REPEATABLE READ isolation mode transaction",
1278 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1279 392 : if (!XactReadOnly)
1280 0 : ereport(ERROR,
1281 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1282 : (errmsg("%s must be called in a read-only transaction",
1283 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1284 :
1285 392 : if (FirstSnapshotSet)
1286 0 : ereport(ERROR,
1287 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1288 : (errmsg("%s must be called before any query",
1289 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1290 :
1291 392 : if (IsSubTransaction())
1292 0 : ereport(ERROR,
1293 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1294 : (errmsg("%s must not be called in a subtransaction",
1295 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1296 :
1297 392 : need_full_snapshot = true;
1298 : }
1299 :
1300 670 : ctx = CreateInitDecodingContext(cmd->plugin, NIL, need_full_snapshot,
1301 : InvalidXLogRecPtr,
1302 670 : XL_ROUTINE(.page_read = logical_read_xlog_page,
1303 : .segment_open = WalSndSegmentOpen,
1304 : .segment_close = wal_segment_close),
1305 : WalSndPrepareWrite, WalSndWriteData,
1306 : WalSndUpdateProgress);
1307 :
1308 : /*
1309 : * Signal that we don't need the timeout mechanism. We're just
1310 : * creating the replication slot and don't yet accept feedback
1311 : * messages or send keepalives. As we possibly need to wait for
1312 : * further WAL the walsender would otherwise possibly be killed too
1313 : * soon.
1314 : */
1315 670 : last_reply_timestamp = 0;
1316 :
1317 : /* build initial snapshot, might take a while */
1318 670 : DecodingContextFindStartpoint(ctx);
1319 :
1320 : /*
1321 : * Export or use the snapshot if we've been asked to do so.
1322 : *
1323 : * NB. We will convert the snapbuild.c kind of snapshot to normal
1324 : * snapshot when doing this.
1325 : */
1326 670 : if (snapshot_action == CRS_EXPORT_SNAPSHOT)
1327 : {
1328 2 : snapshot_name = SnapBuildExportSnapshot(ctx->snapshot_builder);
1329 : }
1330 668 : else if (snapshot_action == CRS_USE_SNAPSHOT)
1331 : {
1332 : Snapshot snap;
1333 :
1334 392 : snap = SnapBuildInitialSnapshot(ctx->snapshot_builder);
1335 392 : RestoreTransactionSnapshot(snap, MyProc);
1336 : }
1337 :
1338 : /* don't need the decoding context anymore */
1339 670 : FreeDecodingContext(ctx);
1340 :
1341 670 : if (!cmd->temporary)
1342 670 : ReplicationSlotPersist();
1343 : }
1344 :
1345 944 : snprintf(xloc, sizeof(xloc), "%X/%08X",
1346 944 : LSN_FORMAT_ARGS(MyReplicationSlot->data.confirmed_flush));
1347 :
1348 944 : dest = CreateDestReceiver(DestRemoteSimple);
1349 :
1350 : /*----------
1351 : * Need a tuple descriptor representing four columns:
1352 : * - first field: the slot name
1353 : * - second field: LSN at which we became consistent
1354 : * - third field: exported snapshot's name
1355 : * - fourth field: output plugin
1356 : */
1357 944 : tupdesc = CreateTemplateTupleDesc(4);
1358 944 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "slot_name",
1359 : TEXTOID, -1, 0);
1360 944 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "consistent_point",
1361 : TEXTOID, -1, 0);
1362 944 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 3, "snapshot_name",
1363 : TEXTOID, -1, 0);
1364 944 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 4, "output_plugin",
1365 : TEXTOID, -1, 0);
1366 :
1367 : /* prepare for projection of tuples */
1368 944 : tstate = begin_tup_output_tupdesc(dest, tupdesc, &TTSOpsVirtual);
1369 :
1370 : /* slot_name */
1371 944 : slot_name = NameStr(MyReplicationSlot->data.name);
1372 944 : values[0] = CStringGetTextDatum(slot_name);
1373 :
1374 : /* consistent wal location */
1375 944 : values[1] = CStringGetTextDatum(xloc);
1376 :
1377 : /* snapshot name, or NULL if none */
1378 944 : if (snapshot_name != NULL)
1379 2 : values[2] = CStringGetTextDatum(snapshot_name);
1380 : else
1381 942 : nulls[2] = true;
1382 :
1383 : /* plugin, or NULL if none */
1384 944 : if (cmd->plugin != NULL)
1385 670 : values[3] = CStringGetTextDatum(cmd->plugin);
1386 : else
1387 274 : nulls[3] = true;
1388 :
1389 : /* send it to dest */
1390 944 : do_tup_output(tstate, values, nulls);
1391 944 : end_tup_output(tstate);
1392 :
1393 944 : ReplicationSlotRelease();
1394 944 : }
1395 :
1396 : /*
1397 : * Get rid of a replication slot that is no longer wanted.
1398 : */
1399 : static void
1400 550 : DropReplicationSlot(DropReplicationSlotCmd *cmd)
1401 : {
1402 550 : ReplicationSlotDrop(cmd->slotname, !cmd->wait);
1403 542 : }
1404 :
1405 : /*
1406 : * Change the definition of a replication slot.
1407 : */
1408 : static void
1409 14 : AlterReplicationSlot(AlterReplicationSlotCmd *cmd)
1410 : {
1411 14 : bool failover_given = false;
1412 14 : bool two_phase_given = false;
1413 : bool failover;
1414 : bool two_phase;
1415 :
1416 : /* Parse options */
1417 42 : foreach_ptr(DefElem, defel, cmd->options)
1418 : {
1419 14 : if (strcmp(defel->defname, "failover") == 0)
1420 : {
1421 12 : if (failover_given)
1422 0 : ereport(ERROR,
1423 : (errcode(ERRCODE_SYNTAX_ERROR),
1424 : errmsg("conflicting or redundant options")));
1425 12 : failover_given = true;
1426 12 : failover = defGetBoolean(defel);
1427 : }
1428 2 : else if (strcmp(defel->defname, "two_phase") == 0)
1429 : {
1430 2 : if (two_phase_given)
1431 0 : ereport(ERROR,
1432 : (errcode(ERRCODE_SYNTAX_ERROR),
1433 : errmsg("conflicting or redundant options")));
1434 2 : two_phase_given = true;
1435 2 : two_phase = defGetBoolean(defel);
1436 : }
1437 : else
1438 0 : elog(ERROR, "unrecognized option: %s", defel->defname);
1439 : }
1440 :
1441 14 : ReplicationSlotAlter(cmd->slotname,
1442 : failover_given ? &failover : NULL,
1443 : two_phase_given ? &two_phase : NULL);
1444 10 : }
1445 :
1446 : /*
1447 : * Load previously initiated logical slot and prepare for sending data (via
1448 : * WalSndLoop).
1449 : */
1450 : static void
1451 818 : StartLogicalReplication(StartReplicationCmd *cmd)
1452 : {
1453 : StringInfoData buf;
1454 : QueryCompletion qc;
1455 :
1456 : /* make sure that our requirements are still fulfilled */
1457 818 : CheckLogicalDecodingRequirements();
1458 :
1459 : Assert(!MyReplicationSlot);
1460 :
1461 814 : ReplicationSlotAcquire(cmd->slotname, true, true);
1462 :
1463 : /*
1464 : * Force a disconnect, so that the decoding code doesn't need to care
1465 : * about an eventual switch from running in recovery, to running in a
1466 : * normal environment. Client code is expected to handle reconnects.
1467 : */
1468 804 : if (am_cascading_walsender && !RecoveryInProgress())
1469 : {
1470 0 : ereport(LOG,
1471 : (errmsg("terminating walsender process after promotion")));
1472 0 : got_STOPPING = true;
1473 : }
1474 :
1475 : /*
1476 : * Create our decoding context, making it start at the previously ack'ed
1477 : * position.
1478 : *
1479 : * Do this before sending a CopyBothResponse message, so that any errors
1480 : * are reported early.
1481 : */
1482 802 : logical_decoding_ctx =
1483 804 : CreateDecodingContext(cmd->startpoint, cmd->options, false,
1484 804 : XL_ROUTINE(.page_read = logical_read_xlog_page,
1485 : .segment_open = WalSndSegmentOpen,
1486 : .segment_close = wal_segment_close),
1487 : WalSndPrepareWrite, WalSndWriteData,
1488 : WalSndUpdateProgress);
1489 802 : xlogreader = logical_decoding_ctx->reader;
1490 :
1491 802 : WalSndSetState(WALSNDSTATE_CATCHUP);
1492 :
1493 : /* Send a CopyBothResponse message, and start streaming */
1494 802 : pq_beginmessage(&buf, PqMsg_CopyBothResponse);
1495 802 : pq_sendbyte(&buf, 0);
1496 802 : pq_sendint16(&buf, 0);
1497 802 : pq_endmessage(&buf);
1498 802 : pq_flush();
1499 :
1500 : /* Start reading WAL from the oldest required WAL. */
1501 802 : XLogBeginRead(logical_decoding_ctx->reader,
1502 802 : MyReplicationSlot->data.restart_lsn);
1503 :
1504 : /*
1505 : * Report the location after which we'll send out further commits as the
1506 : * current sentPtr.
1507 : */
1508 802 : sentPtr = MyReplicationSlot->data.confirmed_flush;
1509 :
1510 : /* Also update the sent position status in shared memory */
1511 802 : SpinLockAcquire(&MyWalSnd->mutex);
1512 802 : MyWalSnd->sentPtr = MyReplicationSlot->data.restart_lsn;
1513 802 : SpinLockRelease(&MyWalSnd->mutex);
1514 :
1515 802 : replication_active = true;
1516 :
1517 802 : SyncRepInitConfig();
1518 :
1519 : /* Main loop of walsender */
1520 802 : WalSndLoop(XLogSendLogical);
1521 :
1522 382 : FreeDecodingContext(logical_decoding_ctx);
1523 382 : ReplicationSlotRelease();
1524 :
1525 382 : replication_active = false;
1526 382 : if (got_STOPPING)
1527 0 : proc_exit(0);
1528 382 : WalSndSetState(WALSNDSTATE_STARTUP);
1529 :
1530 : /* Get out of COPY mode (CommandComplete). */
1531 382 : SetQueryCompletion(&qc, CMDTAG_COPY, 0);
1532 382 : EndCommand(&qc, DestRemote, false);
1533 382 : }
1534 :
1535 : /*
1536 : * LogicalDecodingContext 'prepare_write' callback.
1537 : *
1538 : * Prepare a write into a StringInfo.
1539 : *
1540 : * Don't do anything lasting in here, it's quite possible that nothing will be done
1541 : * with the data.
1542 : */
1543 : static void
1544 369990 : WalSndPrepareWrite(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid, bool last_write)
1545 : {
1546 : /* can't have sync rep confused by sending the same LSN several times */
1547 369990 : if (!last_write)
1548 796 : lsn = InvalidXLogRecPtr;
1549 :
1550 369990 : resetStringInfo(ctx->out);
1551 :
1552 369990 : pq_sendbyte(ctx->out, PqReplMsg_WALData);
1553 369990 : pq_sendint64(ctx->out, lsn); /* dataStart */
1554 369990 : pq_sendint64(ctx->out, lsn); /* walEnd */
1555 :
1556 : /*
1557 : * Fill out the sendtime later, just as it's done in XLogSendPhysical, but
1558 : * reserve space here.
1559 : */
1560 369990 : pq_sendint64(ctx->out, 0); /* sendtime */
1561 369990 : }
1562 :
1563 : /*
1564 : * LogicalDecodingContext 'write' callback.
1565 : *
1566 : * Actually write out data previously prepared by WalSndPrepareWrite out to
1567 : * the network. Take as long as needed, but process replies from the other
1568 : * side and check timeouts during that.
1569 : */
1570 : static void
1571 369990 : WalSndWriteData(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid,
1572 : bool last_write)
1573 : {
1574 : TimestampTz now;
1575 :
1576 : /*
1577 : * Fill the send timestamp last, so that it is taken as late as possible.
1578 : * This is somewhat ugly, but the protocol is set as it's already used for
1579 : * several releases by streaming physical replication.
1580 : */
1581 369990 : resetStringInfo(&tmpbuf);
1582 369990 : now = GetCurrentTimestamp();
1583 369990 : pq_sendint64(&tmpbuf, now);
1584 369990 : memcpy(&ctx->out->data[1 + sizeof(int64) + sizeof(int64)],
1585 369990 : tmpbuf.data, sizeof(int64));
1586 :
1587 : /* output previously gathered data in a CopyData packet */
1588 369990 : pq_putmessage_noblock(PqMsg_CopyData, ctx->out->data, ctx->out->len);
1589 :
1590 369990 : CHECK_FOR_INTERRUPTS();
1591 :
1592 : /* Try to flush pending output to the client */
1593 369990 : if (pq_flush_if_writable() != 0)
1594 34 : WalSndShutdown();
1595 :
1596 : /* Try taking fast path unless we get too close to walsender timeout. */
1597 369956 : if (now < TimestampTzPlusMilliseconds(last_reply_timestamp,
1598 369956 : wal_sender_timeout / 2) &&
1599 369956 : !pq_is_send_pending())
1600 : {
1601 369372 : return;
1602 : }
1603 :
1604 : /* If we have pending write here, go to slow path */
1605 584 : ProcessPendingWrites();
1606 : }
1607 :
1608 : /*
1609 : * Wait until there is no pending write. Also process replies from the other
1610 : * side and check timeouts during that.
1611 : */
1612 : static void
1613 584 : ProcessPendingWrites(void)
1614 : {
1615 : for (;;)
1616 720 : {
1617 : long sleeptime;
1618 :
1619 : /* Check for input from the client */
1620 1304 : ProcessRepliesIfAny();
1621 :
1622 : /* die if timeout was reached */
1623 1304 : WalSndCheckTimeOut();
1624 :
1625 : /* Send keepalive if the time has come */
1626 1304 : WalSndKeepaliveIfNecessary();
1627 :
1628 1304 : if (!pq_is_send_pending())
1629 584 : break;
1630 :
1631 720 : sleeptime = WalSndComputeSleeptime(GetCurrentTimestamp());
1632 :
1633 : /* Sleep until something happens or we time out */
1634 720 : WalSndWait(WL_SOCKET_WRITEABLE | WL_SOCKET_READABLE, sleeptime,
1635 : WAIT_EVENT_WAL_SENDER_WRITE_DATA);
1636 :
1637 : /* Clear any already-pending wakeups */
1638 720 : ResetLatch(MyLatch);
1639 :
1640 720 : CHECK_FOR_INTERRUPTS();
1641 :
1642 : /* Process any requests or signals received recently */
1643 720 : if (ConfigReloadPending)
1644 : {
1645 0 : ConfigReloadPending = false;
1646 0 : ProcessConfigFile(PGC_SIGHUP);
1647 0 : SyncRepInitConfig();
1648 : }
1649 :
1650 : /* Try to flush pending output to the client */
1651 720 : if (pq_flush_if_writable() != 0)
1652 0 : WalSndShutdown();
1653 : }
1654 :
1655 : /* reactivate latch so WalSndLoop knows to continue */
1656 584 : SetLatch(MyLatch);
1657 584 : }
1658 :
1659 : /*
1660 : * LogicalDecodingContext 'update_progress' callback.
1661 : *
1662 : * Write the current position to the lag tracker (see XLogSendPhysical).
1663 : *
1664 : * When skipping empty transactions, send a keepalive message if necessary.
1665 : */
1666 : static void
1667 5082 : WalSndUpdateProgress(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid,
1668 : bool skipped_xact)
1669 : {
1670 : static TimestampTz sendTime = 0;
1671 5082 : TimestampTz now = GetCurrentTimestamp();
1672 5082 : bool pending_writes = false;
1673 5082 : bool end_xact = ctx->end_xact;
1674 :
1675 : /*
1676 : * Track lag no more than once per WALSND_LOGICAL_LAG_TRACK_INTERVAL_MS to
1677 : * avoid flooding the lag tracker when we commit frequently.
1678 : *
1679 : * We don't have a mechanism to get the ack for any LSN other than end
1680 : * xact LSN from the downstream. So, we track lag only for end of
1681 : * transaction LSN.
1682 : */
1683 : #define WALSND_LOGICAL_LAG_TRACK_INTERVAL_MS 1000
1684 5082 : if (end_xact && TimestampDifferenceExceeds(sendTime, now,
1685 : WALSND_LOGICAL_LAG_TRACK_INTERVAL_MS))
1686 : {
1687 452 : LagTrackerWrite(lsn, now);
1688 452 : sendTime = now;
1689 : }
1690 :
1691 : /*
1692 : * When skipping empty transactions in synchronous replication, we send a
1693 : * keepalive message to avoid delaying such transactions.
1694 : *
1695 : * It is okay to check sync_standbys_status without lock here as in the
1696 : * worst case we will just send an extra keepalive message when it is
1697 : * really not required.
1698 : */
1699 5082 : if (skipped_xact &&
1700 918 : SyncRepRequested() &&
1701 918 : (((volatile WalSndCtlData *) WalSndCtl)->sync_standbys_status & SYNC_STANDBY_DEFINED))
1702 : {
1703 0 : WalSndKeepalive(false, lsn);
1704 :
1705 : /* Try to flush pending output to the client */
1706 0 : if (pq_flush_if_writable() != 0)
1707 0 : WalSndShutdown();
1708 :
1709 : /* If we have pending write here, make sure it's actually flushed */
1710 0 : if (pq_is_send_pending())
1711 0 : pending_writes = true;
1712 : }
1713 :
1714 : /*
1715 : * Process pending writes if any or try to send a keepalive if required.
1716 : * We don't need to try sending keep alive messages at the transaction end
1717 : * as that will be done at a later point in time. This is required only
1718 : * for large transactions where we don't send any changes to the
1719 : * downstream and the receiver can timeout due to that.
1720 : */
1721 5082 : if (pending_writes || (!end_xact &&
1722 3078 : now >= TimestampTzPlusMilliseconds(last_reply_timestamp,
1723 : wal_sender_timeout / 2)))
1724 0 : ProcessPendingWrites();
1725 5082 : }
1726 :
1727 : /*
1728 : * Wake up the logical walsender processes with logical failover slots if the
1729 : * currently acquired physical slot is specified in synchronized_standby_slots GUC.
1730 : */
1731 : void
1732 44530 : PhysicalWakeupLogicalWalSnd(void)
1733 : {
1734 : Assert(MyReplicationSlot && SlotIsPhysical(MyReplicationSlot));
1735 :
1736 : /*
1737 : * If we are running in a standby, there is no need to wake up walsenders.
1738 : * This is because we do not support syncing slots to cascading standbys,
1739 : * so, there are no walsenders waiting for standbys to catch up.
1740 : */
1741 44530 : if (RecoveryInProgress())
1742 100 : return;
1743 :
1744 44430 : if (SlotExistsInSyncStandbySlots(NameStr(MyReplicationSlot->data.name)))
1745 12 : ConditionVariableBroadcast(&WalSndCtl->wal_confirm_rcv_cv);
1746 : }
1747 :
1748 : /*
1749 : * Returns true if not all standbys have caught up to the flushed position
1750 : * (flushed_lsn) when the current acquired slot is a logical failover
1751 : * slot and we are streaming; otherwise, returns false.
1752 : *
1753 : * If returning true, the function sets the appropriate wait event in
1754 : * wait_event; otherwise, wait_event is set to 0.
1755 : */
1756 : static bool
1757 36978 : NeedToWaitForStandbys(XLogRecPtr flushed_lsn, uint32 *wait_event)
1758 : {
1759 36978 : int elevel = got_STOPPING ? ERROR : WARNING;
1760 : bool failover_slot;
1761 :
1762 36978 : failover_slot = (replication_active && MyReplicationSlot->data.failover);
1763 :
1764 : /*
1765 : * Note that after receiving the shutdown signal, an ERROR is reported if
1766 : * any slots are dropped, invalidated, or inactive. This measure is taken
1767 : * to prevent the walsender from waiting indefinitely.
1768 : */
1769 36978 : if (failover_slot && !StandbySlotsHaveCaughtup(flushed_lsn, elevel))
1770 : {
1771 14 : *wait_event = WAIT_EVENT_WAIT_FOR_STANDBY_CONFIRMATION;
1772 14 : return true;
1773 : }
1774 :
1775 36964 : *wait_event = 0;
1776 36964 : return false;
1777 : }
1778 :
1779 : /*
1780 : * Returns true if we need to wait for WALs to be flushed to disk, or if not
1781 : * all standbys have caught up to the flushed position (flushed_lsn) when the
1782 : * current acquired slot is a logical failover slot and we are
1783 : * streaming; otherwise, returns false.
1784 : *
1785 : * If returning true, the function sets the appropriate wait event in
1786 : * wait_event; otherwise, wait_event is set to 0.
1787 : */
1788 : static bool
1789 54962 : NeedToWaitForWal(XLogRecPtr target_lsn, XLogRecPtr flushed_lsn,
1790 : uint32 *wait_event)
1791 : {
1792 : /* Check if we need to wait for WALs to be flushed to disk */
1793 54962 : if (target_lsn > flushed_lsn)
1794 : {
1795 24732 : *wait_event = WAIT_EVENT_WAL_SENDER_WAIT_FOR_WAL;
1796 24732 : return true;
1797 : }
1798 :
1799 : /* Check if the standby slots have caught up to the flushed position */
1800 30230 : return NeedToWaitForStandbys(flushed_lsn, wait_event);
1801 : }
1802 :
1803 : /*
1804 : * Wait till WAL < loc is flushed to disk so it can be safely sent to client.
1805 : *
1806 : * If the walsender holds a logical failover slot, we also wait for all the
1807 : * specified streaming replication standby servers to confirm receipt of WAL
1808 : * up to RecentFlushPtr. It is beneficial to wait here for the confirmation
1809 : * up to RecentFlushPtr rather than waiting before transmitting each change
1810 : * to logical subscribers, which is already covered by RecentFlushPtr.
1811 : *
1812 : * Returns end LSN of flushed WAL. Normally this will be >= loc, but if we
1813 : * detect a shutdown request (either from postmaster or client) we will return
1814 : * early, so caller must always check.
1815 : */
1816 : static XLogRecPtr
1817 37510 : WalSndWaitForWal(XLogRecPtr loc)
1818 : {
1819 : int wakeEvents;
1820 37510 : uint32 wait_event = 0;
1821 : static XLogRecPtr RecentFlushPtr = InvalidXLogRecPtr;
1822 37510 : TimestampTz last_flush = 0;
1823 :
1824 : /*
1825 : * Fast path to avoid acquiring the spinlock in case we already know we
1826 : * have enough WAL available and all the standby servers have confirmed
1827 : * receipt of WAL up to RecentFlushPtr. This is particularly interesting
1828 : * if we're far behind.
1829 : */
1830 37510 : if (XLogRecPtrIsValid(RecentFlushPtr) &&
1831 36406 : !NeedToWaitForWal(loc, RecentFlushPtr, &wait_event))
1832 26412 : return RecentFlushPtr;
1833 :
1834 : /*
1835 : * Within the loop, we wait for the necessary WALs to be flushed to disk
1836 : * first, followed by waiting for standbys to catch up if there are enough
1837 : * WALs (see NeedToWaitForWal()) or upon receiving the shutdown signal.
1838 : */
1839 : for (;;)
1840 14562 : {
1841 25660 : bool wait_for_standby_at_stop = false;
1842 : long sleeptime;
1843 : TimestampTz now;
1844 :
1845 : /* Clear any already-pending wakeups */
1846 25660 : ResetLatch(MyLatch);
1847 :
1848 25660 : CHECK_FOR_INTERRUPTS();
1849 :
1850 : /* Process any requests or signals received recently */
1851 25648 : if (ConfigReloadPending)
1852 : {
1853 18 : ConfigReloadPending = false;
1854 18 : ProcessConfigFile(PGC_SIGHUP);
1855 18 : SyncRepInitConfig();
1856 : }
1857 :
1858 : /* Check for input from the client */
1859 25648 : ProcessRepliesIfAny();
1860 :
1861 : /*
1862 : * If we're shutting down, trigger pending WAL to be written out,
1863 : * otherwise we'd possibly end up waiting for WAL that never gets
1864 : * written, because walwriter has shut down already.
1865 : */
1866 25304 : if (got_STOPPING)
1867 6748 : XLogBackgroundFlush();
1868 :
1869 : /*
1870 : * To avoid the scenario where standbys need to catch up to a newer
1871 : * WAL location in each iteration, we update our idea of the currently
1872 : * flushed position only if we are not waiting for standbys to catch
1873 : * up.
1874 : */
1875 25304 : if (wait_event != WAIT_EVENT_WAIT_FOR_STANDBY_CONFIRMATION)
1876 : {
1877 25290 : if (!RecoveryInProgress())
1878 24536 : RecentFlushPtr = GetFlushRecPtr(NULL);
1879 : else
1880 754 : RecentFlushPtr = GetXLogReplayRecPtr(NULL);
1881 : }
1882 :
1883 : /*
1884 : * If postmaster asked us to stop and the standby slots have caught up
1885 : * to the flushed position, don't wait anymore.
1886 : *
1887 : * It's important to do this check after the recomputation of
1888 : * RecentFlushPtr, so we can send all remaining data before shutting
1889 : * down.
1890 : */
1891 25304 : if (got_STOPPING)
1892 : {
1893 6748 : if (NeedToWaitForStandbys(RecentFlushPtr, &wait_event))
1894 0 : wait_for_standby_at_stop = true;
1895 : else
1896 6748 : break;
1897 : }
1898 :
1899 : /*
1900 : * We only send regular messages to the client for full decoded
1901 : * transactions, but a synchronous replication and walsender shutdown
1902 : * possibly are waiting for a later location. So, before sleeping, we
1903 : * send a ping containing the flush location. If the receiver is
1904 : * otherwise idle, this keepalive will trigger a reply. Processing the
1905 : * reply will update these MyWalSnd locations.
1906 : */
1907 18556 : if (MyWalSnd->flush < sentPtr &&
1908 6690 : MyWalSnd->write < sentPtr &&
1909 3654 : !waiting_for_ping_response)
1910 3654 : WalSndKeepalive(false, InvalidXLogRecPtr);
1911 :
1912 : /*
1913 : * Exit the loop if already caught up and doesn't need to wait for
1914 : * standby slots.
1915 : */
1916 18556 : if (!wait_for_standby_at_stop &&
1917 18556 : !NeedToWaitForWal(loc, RecentFlushPtr, &wait_event))
1918 3804 : break;
1919 :
1920 : /*
1921 : * Waiting for new WAL or waiting for standbys to catch up. Since we
1922 : * need to wait, we're now caught up.
1923 : */
1924 14752 : WalSndCaughtUp = true;
1925 :
1926 : /*
1927 : * Try to flush any pending output to the client.
1928 : */
1929 14752 : if (pq_flush_if_writable() != 0)
1930 0 : WalSndShutdown();
1931 :
1932 : /*
1933 : * If we have received CopyDone from the client, sent CopyDone
1934 : * ourselves, and the output buffer is empty, it's time to exit
1935 : * streaming, so fail the current WAL fetch request.
1936 : */
1937 14752 : if (streamingDoneReceiving && streamingDoneSending &&
1938 190 : !pq_is_send_pending())
1939 190 : break;
1940 :
1941 : /* die if timeout was reached */
1942 14562 : WalSndCheckTimeOut();
1943 :
1944 : /* Send keepalive if the time has come */
1945 14562 : WalSndKeepaliveIfNecessary();
1946 :
1947 : /*
1948 : * Sleep until something happens or we time out. Also wait for the
1949 : * socket becoming writable, if there's still pending output.
1950 : * Otherwise we might sit on sendable output data while waiting for
1951 : * new WAL to be generated. (But if we have nothing to send, we don't
1952 : * want to wake on socket-writable.)
1953 : */
1954 14562 : now = GetCurrentTimestamp();
1955 14562 : sleeptime = WalSndComputeSleeptime(now);
1956 :
1957 14562 : wakeEvents = WL_SOCKET_READABLE;
1958 :
1959 14562 : if (pq_is_send_pending())
1960 0 : wakeEvents |= WL_SOCKET_WRITEABLE;
1961 :
1962 : Assert(wait_event != 0);
1963 :
1964 : /* Report IO statistics, if needed */
1965 14562 : if (TimestampDifferenceExceeds(last_flush, now,
1966 : WALSENDER_STATS_FLUSH_INTERVAL))
1967 : {
1968 2914 : pgstat_flush_io(false);
1969 2914 : (void) pgstat_flush_backend(false, PGSTAT_BACKEND_FLUSH_IO);
1970 2914 : last_flush = now;
1971 : }
1972 :
1973 14562 : WalSndWait(wakeEvents, sleeptime, wait_event);
1974 : }
1975 :
1976 : /* reactivate latch so WalSndLoop knows to continue */
1977 10742 : SetLatch(MyLatch);
1978 10742 : return RecentFlushPtr;
1979 : }
1980 :
1981 : /*
1982 : * Execute an incoming replication command.
1983 : *
1984 : * Returns true if the cmd_string was recognized as WalSender command, false
1985 : * if not.
1986 : */
1987 : bool
1988 10600 : exec_replication_command(const char *cmd_string)
1989 : {
1990 : yyscan_t scanner;
1991 : int parse_rc;
1992 : Node *cmd_node;
1993 : const char *cmdtag;
1994 10600 : MemoryContext old_context = CurrentMemoryContext;
1995 :
1996 : /* We save and re-use the cmd_context across calls */
1997 : static MemoryContext cmd_context = NULL;
1998 :
1999 : /*
2000 : * If WAL sender has been told that shutdown is getting close, switch its
2001 : * status accordingly to handle the next replication commands correctly.
2002 : */
2003 10600 : if (got_STOPPING)
2004 0 : WalSndSetState(WALSNDSTATE_STOPPING);
2005 :
2006 : /*
2007 : * Throw error if in stopping mode. We need prevent commands that could
2008 : * generate WAL while the shutdown checkpoint is being written. To be
2009 : * safe, we just prohibit all new commands.
2010 : */
2011 10600 : if (MyWalSnd->state == WALSNDSTATE_STOPPING)
2012 0 : ereport(ERROR,
2013 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
2014 : errmsg("cannot execute new commands while WAL sender is in stopping mode")));
2015 :
2016 : /*
2017 : * CREATE_REPLICATION_SLOT ... LOGICAL exports a snapshot until the next
2018 : * command arrives. Clean up the old stuff if there's anything.
2019 : */
2020 10600 : SnapBuildClearExportedSnapshot();
2021 :
2022 10600 : CHECK_FOR_INTERRUPTS();
2023 :
2024 : /*
2025 : * Prepare to parse and execute the command.
2026 : *
2027 : * Because replication command execution can involve beginning or ending
2028 : * transactions, we need a working context that will survive that, so we
2029 : * make it a child of TopMemoryContext. That in turn creates a hazard of
2030 : * long-lived memory leaks if we lose track of the working context. We
2031 : * deal with that by creating it only once per walsender, and resetting it
2032 : * for each new command. (Normally this reset is a no-op, but if the
2033 : * prior exec_replication_command call failed with an error, it won't be.)
2034 : *
2035 : * This is subtler than it looks. The transactions we manage can extend
2036 : * across replication commands, indeed SnapBuildClearExportedSnapshot
2037 : * might have just ended one. Because transaction exit will revert to the
2038 : * memory context that was current at transaction start, we need to be
2039 : * sure that that context is still valid. That motivates re-using the
2040 : * same cmd_context rather than making a new one each time.
2041 : */
2042 10600 : if (cmd_context == NULL)
2043 2314 : cmd_context = AllocSetContextCreate(TopMemoryContext,
2044 : "Replication command context",
2045 : ALLOCSET_DEFAULT_SIZES);
2046 : else
2047 8286 : MemoryContextReset(cmd_context);
2048 :
2049 10600 : MemoryContextSwitchTo(cmd_context);
2050 :
2051 10600 : replication_scanner_init(cmd_string, &scanner);
2052 :
2053 : /*
2054 : * Is it a WalSender command?
2055 : */
2056 10600 : if (!replication_scanner_is_replication_command(scanner))
2057 : {
2058 : /* Nope; clean up and get out. */
2059 4772 : replication_scanner_finish(scanner);
2060 :
2061 4772 : MemoryContextSwitchTo(old_context);
2062 4772 : MemoryContextReset(cmd_context);
2063 :
2064 : /* XXX this is a pretty random place to make this check */
2065 4772 : if (MyDatabaseId == InvalidOid)
2066 0 : ereport(ERROR,
2067 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2068 : errmsg("cannot execute SQL commands in WAL sender for physical replication")));
2069 :
2070 : /* Tell the caller that this wasn't a WalSender command. */
2071 4772 : return false;
2072 : }
2073 :
2074 : /*
2075 : * Looks like a WalSender command, so parse it.
2076 : */
2077 5828 : parse_rc = replication_yyparse(&cmd_node, scanner);
2078 5828 : if (parse_rc != 0)
2079 0 : ereport(ERROR,
2080 : (errcode(ERRCODE_SYNTAX_ERROR),
2081 : errmsg_internal("replication command parser returned %d",
2082 : parse_rc)));
2083 5828 : replication_scanner_finish(scanner);
2084 :
2085 : /*
2086 : * Report query to various monitoring facilities. For this purpose, we
2087 : * report replication commands just like SQL commands.
2088 : */
2089 5828 : debug_query_string = cmd_string;
2090 :
2091 5828 : pgstat_report_activity(STATE_RUNNING, cmd_string);
2092 :
2093 : /*
2094 : * Log replication command if log_replication_commands is enabled. Even
2095 : * when it's disabled, log the command with DEBUG1 level for backward
2096 : * compatibility.
2097 : */
2098 5828 : ereport(log_replication_commands ? LOG : DEBUG1,
2099 : (errmsg("received replication command: %s", cmd_string)));
2100 :
2101 : /*
2102 : * Disallow replication commands in aborted transaction blocks.
2103 : */
2104 5828 : if (IsAbortedTransactionBlockState())
2105 0 : ereport(ERROR,
2106 : (errcode(ERRCODE_IN_FAILED_SQL_TRANSACTION),
2107 : errmsg("current transaction is aborted, "
2108 : "commands ignored until end of transaction block")));
2109 :
2110 5828 : CHECK_FOR_INTERRUPTS();
2111 :
2112 : /*
2113 : * Allocate buffers that will be used for each outgoing and incoming
2114 : * message. We do this just once per command to reduce palloc overhead.
2115 : */
2116 5828 : initStringInfo(&output_message);
2117 5828 : initStringInfo(&reply_message);
2118 5828 : initStringInfo(&tmpbuf);
2119 :
2120 5828 : switch (cmd_node->type)
2121 : {
2122 1416 : case T_IdentifySystemCmd:
2123 1416 : cmdtag = "IDENTIFY_SYSTEM";
2124 1416 : set_ps_display(cmdtag);
2125 1416 : IdentifySystem();
2126 1416 : EndReplicationCommand(cmdtag);
2127 1416 : break;
2128 :
2129 12 : case T_ReadReplicationSlotCmd:
2130 12 : cmdtag = "READ_REPLICATION_SLOT";
2131 12 : set_ps_display(cmdtag);
2132 12 : ReadReplicationSlot((ReadReplicationSlotCmd *) cmd_node);
2133 10 : EndReplicationCommand(cmdtag);
2134 10 : break;
2135 :
2136 362 : case T_BaseBackupCmd:
2137 362 : cmdtag = "BASE_BACKUP";
2138 362 : set_ps_display(cmdtag);
2139 362 : PreventInTransactionBlock(true, cmdtag);
2140 362 : SendBaseBackup((BaseBackupCmd *) cmd_node, uploaded_manifest);
2141 310 : EndReplicationCommand(cmdtag);
2142 310 : break;
2143 :
2144 946 : case T_CreateReplicationSlotCmd:
2145 946 : cmdtag = "CREATE_REPLICATION_SLOT";
2146 946 : set_ps_display(cmdtag);
2147 946 : CreateReplicationSlot((CreateReplicationSlotCmd *) cmd_node);
2148 944 : EndReplicationCommand(cmdtag);
2149 944 : break;
2150 :
2151 550 : case T_DropReplicationSlotCmd:
2152 550 : cmdtag = "DROP_REPLICATION_SLOT";
2153 550 : set_ps_display(cmdtag);
2154 550 : DropReplicationSlot((DropReplicationSlotCmd *) cmd_node);
2155 542 : EndReplicationCommand(cmdtag);
2156 542 : break;
2157 :
2158 14 : case T_AlterReplicationSlotCmd:
2159 14 : cmdtag = "ALTER_REPLICATION_SLOT";
2160 14 : set_ps_display(cmdtag);
2161 14 : AlterReplicationSlot((AlterReplicationSlotCmd *) cmd_node);
2162 10 : EndReplicationCommand(cmdtag);
2163 10 : break;
2164 :
2165 1344 : case T_StartReplicationCmd:
2166 : {
2167 1344 : StartReplicationCmd *cmd = (StartReplicationCmd *) cmd_node;
2168 :
2169 1344 : cmdtag = "START_REPLICATION";
2170 1344 : set_ps_display(cmdtag);
2171 1344 : PreventInTransactionBlock(true, cmdtag);
2172 :
2173 1344 : if (cmd->kind == REPLICATION_KIND_PHYSICAL)
2174 526 : StartReplication(cmd);
2175 : else
2176 818 : StartLogicalReplication(cmd);
2177 :
2178 : /* dupe, but necessary per libpqrcv_endstreaming */
2179 676 : EndReplicationCommand(cmdtag);
2180 :
2181 : Assert(xlogreader != NULL);
2182 676 : break;
2183 : }
2184 :
2185 22 : case T_TimeLineHistoryCmd:
2186 22 : cmdtag = "TIMELINE_HISTORY";
2187 22 : set_ps_display(cmdtag);
2188 22 : PreventInTransactionBlock(true, cmdtag);
2189 22 : SendTimeLineHistory((TimeLineHistoryCmd *) cmd_node);
2190 22 : EndReplicationCommand(cmdtag);
2191 22 : break;
2192 :
2193 1140 : case T_VariableShowStmt:
2194 : {
2195 1140 : DestReceiver *dest = CreateDestReceiver(DestRemoteSimple);
2196 1140 : VariableShowStmt *n = (VariableShowStmt *) cmd_node;
2197 :
2198 1140 : cmdtag = "SHOW";
2199 1140 : set_ps_display(cmdtag);
2200 :
2201 : /* syscache access needs a transaction environment */
2202 1140 : StartTransactionCommand();
2203 1140 : GetPGVariable(n->name, dest);
2204 1140 : CommitTransactionCommand();
2205 1140 : EndReplicationCommand(cmdtag);
2206 : }
2207 1140 : break;
2208 :
2209 22 : case T_UploadManifestCmd:
2210 22 : cmdtag = "UPLOAD_MANIFEST";
2211 22 : set_ps_display(cmdtag);
2212 22 : PreventInTransactionBlock(true, cmdtag);
2213 22 : UploadManifest();
2214 20 : EndReplicationCommand(cmdtag);
2215 20 : break;
2216 :
2217 0 : default:
2218 0 : elog(ERROR, "unrecognized replication command node tag: %u",
2219 : cmd_node->type);
2220 : }
2221 :
2222 : /*
2223 : * Done. Revert to caller's memory context, and clean out the cmd_context
2224 : * to recover memory right away.
2225 : */
2226 5090 : MemoryContextSwitchTo(old_context);
2227 5090 : MemoryContextReset(cmd_context);
2228 :
2229 : /*
2230 : * We need not update ps display or pg_stat_activity, because PostgresMain
2231 : * will reset those to "idle". But we must reset debug_query_string to
2232 : * ensure it doesn't become a dangling pointer.
2233 : */
2234 5090 : debug_query_string = NULL;
2235 :
2236 5090 : return true;
2237 : }
2238 :
2239 : /*
2240 : * Process any incoming messages while streaming. Also checks if the remote
2241 : * end has closed the connection.
2242 : */
2243 : static void
2244 2002724 : ProcessRepliesIfAny(void)
2245 : {
2246 : unsigned char firstchar;
2247 : int maxmsglen;
2248 : int r;
2249 2002724 : bool received = false;
2250 :
2251 2002724 : last_processing = GetCurrentTimestamp();
2252 :
2253 : /*
2254 : * If we already received a CopyDone from the frontend, any subsequent
2255 : * message is the beginning of a new command, and should be processed in
2256 : * the main processing loop.
2257 : */
2258 4204878 : while (!streamingDoneReceiving)
2259 : {
2260 2200812 : pq_startmsgread();
2261 2200812 : r = pq_getbyte_if_available(&firstchar);
2262 2200812 : if (r < 0)
2263 : {
2264 : /* unexpected error or EOF */
2265 30 : ereport(COMMERROR,
2266 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2267 : errmsg("unexpected EOF on standby connection")));
2268 30 : proc_exit(0);
2269 : }
2270 2200782 : if (r == 0)
2271 : {
2272 : /* no data available without blocking */
2273 2000862 : pq_endmsgread();
2274 2000862 : break;
2275 : }
2276 :
2277 : /* Validate message type and set packet size limit */
2278 199920 : switch (firstchar)
2279 : {
2280 198754 : case PqMsg_CopyData:
2281 198754 : maxmsglen = PQ_LARGE_MESSAGE_LIMIT;
2282 198754 : break;
2283 1166 : case PqMsg_CopyDone:
2284 : case PqMsg_Terminate:
2285 1166 : maxmsglen = PQ_SMALL_MESSAGE_LIMIT;
2286 1166 : break;
2287 0 : default:
2288 0 : ereport(FATAL,
2289 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2290 : errmsg("invalid standby message type \"%c\"",
2291 : firstchar)));
2292 : maxmsglen = 0; /* keep compiler quiet */
2293 : break;
2294 : }
2295 :
2296 : /* Read the message contents */
2297 199920 : resetStringInfo(&reply_message);
2298 199920 : if (pq_getmessage(&reply_message, maxmsglen))
2299 : {
2300 0 : ereport(COMMERROR,
2301 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2302 : errmsg("unexpected EOF on standby connection")));
2303 0 : proc_exit(0);
2304 : }
2305 :
2306 : /* ... and process it */
2307 199920 : switch (firstchar)
2308 : {
2309 : /*
2310 : * PqMsg_CopyData means a standby reply wrapped in a CopyData
2311 : * packet.
2312 : */
2313 198754 : case PqMsg_CopyData:
2314 198754 : ProcessStandbyMessage();
2315 198754 : received = true;
2316 198754 : break;
2317 :
2318 : /*
2319 : * PqMsg_CopyDone means the standby requested to finish
2320 : * streaming. Reply with CopyDone, if we had not sent that
2321 : * already.
2322 : */
2323 676 : case PqMsg_CopyDone:
2324 676 : if (!streamingDoneSending)
2325 : {
2326 658 : pq_putmessage_noblock(PqMsg_CopyDone, NULL, 0);
2327 658 : streamingDoneSending = true;
2328 : }
2329 :
2330 676 : streamingDoneReceiving = true;
2331 676 : received = true;
2332 676 : break;
2333 :
2334 : /*
2335 : * PqMsg_Terminate means that the standby is closing down the
2336 : * socket.
2337 : */
2338 490 : case PqMsg_Terminate:
2339 490 : proc_exit(0);
2340 :
2341 199430 : default:
2342 : Assert(false); /* NOT REACHED */
2343 : }
2344 : }
2345 :
2346 : /*
2347 : * Save the last reply timestamp if we've received at least one reply.
2348 : */
2349 2002204 : if (received)
2350 : {
2351 71992 : last_reply_timestamp = last_processing;
2352 71992 : waiting_for_ping_response = false;
2353 : }
2354 2002204 : }
2355 :
2356 : /*
2357 : * Process a status update message received from standby.
2358 : */
2359 : static void
2360 198754 : ProcessStandbyMessage(void)
2361 : {
2362 : char msgtype;
2363 :
2364 : /*
2365 : * Check message type from the first byte.
2366 : */
2367 198754 : msgtype = pq_getmsgbyte(&reply_message);
2368 :
2369 198754 : switch (msgtype)
2370 : {
2371 189234 : case PqReplMsg_StandbyStatusUpdate:
2372 189234 : ProcessStandbyReplyMessage();
2373 189234 : break;
2374 :
2375 264 : case PqReplMsg_HotStandbyFeedback:
2376 264 : ProcessStandbyHSFeedbackMessage();
2377 264 : break;
2378 :
2379 9256 : case PqReplMsg_PrimaryStatusRequest:
2380 9256 : ProcessStandbyPSRequestMessage();
2381 9256 : break;
2382 :
2383 0 : default:
2384 0 : ereport(COMMERROR,
2385 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2386 : errmsg("unexpected message type \"%c\"", msgtype)));
2387 0 : proc_exit(0);
2388 : }
2389 198754 : }
2390 :
2391 : /*
2392 : * Remember that a walreceiver just confirmed receipt of lsn `lsn`.
2393 : */
2394 : static void
2395 93480 : PhysicalConfirmReceivedLocation(XLogRecPtr lsn)
2396 : {
2397 93480 : bool changed = false;
2398 93480 : ReplicationSlot *slot = MyReplicationSlot;
2399 :
2400 : Assert(XLogRecPtrIsValid(lsn));
2401 93480 : SpinLockAcquire(&slot->mutex);
2402 93480 : if (slot->data.restart_lsn != lsn)
2403 : {
2404 44516 : changed = true;
2405 44516 : slot->data.restart_lsn = lsn;
2406 : }
2407 93480 : SpinLockRelease(&slot->mutex);
2408 :
2409 93480 : if (changed)
2410 : {
2411 44516 : ReplicationSlotMarkDirty();
2412 44516 : ReplicationSlotsComputeRequiredLSN();
2413 44516 : PhysicalWakeupLogicalWalSnd();
2414 : }
2415 :
2416 : /*
2417 : * One could argue that the slot should be saved to disk now, but that'd
2418 : * be energy wasted - the worst thing lost information could cause here is
2419 : * to give wrong information in a statistics view - we'll just potentially
2420 : * be more conservative in removing files.
2421 : */
2422 93480 : }
2423 :
2424 : /*
2425 : * Regular reply from standby advising of WAL locations on standby server.
2426 : */
2427 : static void
2428 189234 : ProcessStandbyReplyMessage(void)
2429 : {
2430 : XLogRecPtr writePtr,
2431 : flushPtr,
2432 : applyPtr;
2433 : bool replyRequested;
2434 : TimeOffset writeLag,
2435 : flushLag,
2436 : applyLag;
2437 : bool clearLagTimes;
2438 : TimestampTz now;
2439 : TimestampTz replyTime;
2440 :
2441 : static bool fullyAppliedLastTime = false;
2442 :
2443 : /* the caller already consumed the msgtype byte */
2444 189234 : writePtr = pq_getmsgint64(&reply_message);
2445 189234 : flushPtr = pq_getmsgint64(&reply_message);
2446 189234 : applyPtr = pq_getmsgint64(&reply_message);
2447 189234 : replyTime = pq_getmsgint64(&reply_message);
2448 189234 : replyRequested = pq_getmsgbyte(&reply_message);
2449 :
2450 189234 : if (message_level_is_interesting(DEBUG2))
2451 : {
2452 : char *replyTimeStr;
2453 :
2454 : /* Copy because timestamptz_to_str returns a static buffer */
2455 746 : replyTimeStr = pstrdup(timestamptz_to_str(replyTime));
2456 :
2457 746 : elog(DEBUG2, "write %X/%08X flush %X/%08X apply %X/%08X%s reply_time %s",
2458 : LSN_FORMAT_ARGS(writePtr),
2459 : LSN_FORMAT_ARGS(flushPtr),
2460 : LSN_FORMAT_ARGS(applyPtr),
2461 : replyRequested ? " (reply requested)" : "",
2462 : replyTimeStr);
2463 :
2464 746 : pfree(replyTimeStr);
2465 : }
2466 :
2467 : /* See if we can compute the round-trip lag for these positions. */
2468 189234 : now = GetCurrentTimestamp();
2469 189234 : writeLag = LagTrackerRead(SYNC_REP_WAIT_WRITE, writePtr, now);
2470 189234 : flushLag = LagTrackerRead(SYNC_REP_WAIT_FLUSH, flushPtr, now);
2471 189234 : applyLag = LagTrackerRead(SYNC_REP_WAIT_APPLY, applyPtr, now);
2472 :
2473 : /*
2474 : * If the standby reports that it has fully replayed the WAL in two
2475 : * consecutive reply messages, then the second such message must result
2476 : * from wal_receiver_status_interval expiring on the standby. This is a
2477 : * convenient time to forget the lag times measured when it last
2478 : * wrote/flushed/applied a WAL record, to avoid displaying stale lag data
2479 : * until more WAL traffic arrives.
2480 : */
2481 189234 : clearLagTimes = false;
2482 189234 : if (applyPtr == sentPtr)
2483 : {
2484 7502 : if (fullyAppliedLastTime)
2485 3422 : clearLagTimes = true;
2486 7502 : fullyAppliedLastTime = true;
2487 : }
2488 : else
2489 181732 : fullyAppliedLastTime = false;
2490 :
2491 : /* Send a reply if the standby requested one. */
2492 189234 : if (replyRequested)
2493 0 : WalSndKeepalive(false, InvalidXLogRecPtr);
2494 :
2495 : /*
2496 : * Update shared state for this WalSender process based on reply data from
2497 : * standby.
2498 : */
2499 : {
2500 189234 : WalSnd *walsnd = MyWalSnd;
2501 :
2502 189234 : SpinLockAcquire(&walsnd->mutex);
2503 189234 : walsnd->write = writePtr;
2504 189234 : walsnd->flush = flushPtr;
2505 189234 : walsnd->apply = applyPtr;
2506 189234 : if (writeLag != -1 || clearLagTimes)
2507 90202 : walsnd->writeLag = writeLag;
2508 189234 : if (flushLag != -1 || clearLagTimes)
2509 101108 : walsnd->flushLag = flushLag;
2510 189234 : if (applyLag != -1 || clearLagTimes)
2511 102902 : walsnd->applyLag = applyLag;
2512 189234 : walsnd->replyTime = replyTime;
2513 189234 : SpinLockRelease(&walsnd->mutex);
2514 : }
2515 :
2516 189234 : if (!am_cascading_walsender)
2517 188680 : SyncRepReleaseWaiters();
2518 :
2519 : /*
2520 : * Advance our local xmin horizon when the client confirmed a flush.
2521 : */
2522 189234 : if (MyReplicationSlot && XLogRecPtrIsValid(flushPtr))
2523 : {
2524 182034 : if (SlotIsLogical(MyReplicationSlot))
2525 88554 : LogicalConfirmReceivedLocation(flushPtr);
2526 : else
2527 93480 : PhysicalConfirmReceivedLocation(flushPtr);
2528 : }
2529 189234 : }
2530 :
2531 : /* compute new replication slot xmin horizon if needed */
2532 : static void
2533 126 : PhysicalReplicationSlotNewXmin(TransactionId feedbackXmin, TransactionId feedbackCatalogXmin)
2534 : {
2535 126 : bool changed = false;
2536 126 : ReplicationSlot *slot = MyReplicationSlot;
2537 :
2538 126 : SpinLockAcquire(&slot->mutex);
2539 126 : MyProc->xmin = InvalidTransactionId;
2540 :
2541 : /*
2542 : * For physical replication we don't need the interlock provided by xmin
2543 : * and effective_xmin since the consequences of a missed increase are
2544 : * limited to query cancellations, so set both at once.
2545 : */
2546 126 : if (!TransactionIdIsNormal(slot->data.xmin) ||
2547 66 : !TransactionIdIsNormal(feedbackXmin) ||
2548 66 : TransactionIdPrecedes(slot->data.xmin, feedbackXmin))
2549 : {
2550 78 : changed = true;
2551 78 : slot->data.xmin = feedbackXmin;
2552 78 : slot->effective_xmin = feedbackXmin;
2553 : }
2554 126 : if (!TransactionIdIsNormal(slot->data.catalog_xmin) ||
2555 36 : !TransactionIdIsNormal(feedbackCatalogXmin) ||
2556 36 : TransactionIdPrecedes(slot->data.catalog_xmin, feedbackCatalogXmin))
2557 : {
2558 92 : changed = true;
2559 92 : slot->data.catalog_xmin = feedbackCatalogXmin;
2560 92 : slot->effective_catalog_xmin = feedbackCatalogXmin;
2561 : }
2562 126 : SpinLockRelease(&slot->mutex);
2563 :
2564 126 : if (changed)
2565 : {
2566 98 : ReplicationSlotMarkDirty();
2567 98 : ReplicationSlotsComputeRequiredXmin(false);
2568 : }
2569 126 : }
2570 :
2571 : /*
2572 : * Check that the provided xmin/epoch are sane, that is, not in the future
2573 : * and not so far back as to be already wrapped around.
2574 : *
2575 : * Epoch of nextXid should be same as standby, or if the counter has
2576 : * wrapped, then one greater than standby.
2577 : *
2578 : * This check doesn't care about whether clog exists for these xids
2579 : * at all.
2580 : */
2581 : static bool
2582 140 : TransactionIdInRecentPast(TransactionId xid, uint32 epoch)
2583 : {
2584 : FullTransactionId nextFullXid;
2585 : TransactionId nextXid;
2586 : uint32 nextEpoch;
2587 :
2588 140 : nextFullXid = ReadNextFullTransactionId();
2589 140 : nextXid = XidFromFullTransactionId(nextFullXid);
2590 140 : nextEpoch = EpochFromFullTransactionId(nextFullXid);
2591 :
2592 140 : if (xid <= nextXid)
2593 : {
2594 140 : if (epoch != nextEpoch)
2595 0 : return false;
2596 : }
2597 : else
2598 : {
2599 0 : if (epoch + 1 != nextEpoch)
2600 0 : return false;
2601 : }
2602 :
2603 140 : if (!TransactionIdPrecedesOrEquals(xid, nextXid))
2604 0 : return false; /* epoch OK, but it's wrapped around */
2605 :
2606 140 : return true;
2607 : }
2608 :
2609 : /*
2610 : * Hot Standby feedback
2611 : */
2612 : static void
2613 264 : ProcessStandbyHSFeedbackMessage(void)
2614 : {
2615 : TransactionId feedbackXmin;
2616 : uint32 feedbackEpoch;
2617 : TransactionId feedbackCatalogXmin;
2618 : uint32 feedbackCatalogEpoch;
2619 : TimestampTz replyTime;
2620 :
2621 : /*
2622 : * Decipher the reply message. The caller already consumed the msgtype
2623 : * byte. See XLogWalRcvSendHSFeedback() in walreceiver.c for the creation
2624 : * of this message.
2625 : */
2626 264 : replyTime = pq_getmsgint64(&reply_message);
2627 264 : feedbackXmin = pq_getmsgint(&reply_message, 4);
2628 264 : feedbackEpoch = pq_getmsgint(&reply_message, 4);
2629 264 : feedbackCatalogXmin = pq_getmsgint(&reply_message, 4);
2630 264 : feedbackCatalogEpoch = pq_getmsgint(&reply_message, 4);
2631 :
2632 264 : if (message_level_is_interesting(DEBUG2))
2633 : {
2634 : char *replyTimeStr;
2635 :
2636 : /* Copy because timestamptz_to_str returns a static buffer */
2637 8 : replyTimeStr = pstrdup(timestamptz_to_str(replyTime));
2638 :
2639 8 : elog(DEBUG2, "hot standby feedback xmin %u epoch %u, catalog_xmin %u epoch %u reply_time %s",
2640 : feedbackXmin,
2641 : feedbackEpoch,
2642 : feedbackCatalogXmin,
2643 : feedbackCatalogEpoch,
2644 : replyTimeStr);
2645 :
2646 8 : pfree(replyTimeStr);
2647 : }
2648 :
2649 : /*
2650 : * Update shared state for this WalSender process based on reply data from
2651 : * standby.
2652 : */
2653 : {
2654 264 : WalSnd *walsnd = MyWalSnd;
2655 :
2656 264 : SpinLockAcquire(&walsnd->mutex);
2657 264 : walsnd->replyTime = replyTime;
2658 264 : SpinLockRelease(&walsnd->mutex);
2659 : }
2660 :
2661 : /*
2662 : * Unset WalSender's xmins if the feedback message values are invalid.
2663 : * This happens when the downstream turned hot_standby_feedback off.
2664 : */
2665 264 : if (!TransactionIdIsNormal(feedbackXmin)
2666 172 : && !TransactionIdIsNormal(feedbackCatalogXmin))
2667 : {
2668 172 : MyProc->xmin = InvalidTransactionId;
2669 172 : if (MyReplicationSlot != NULL)
2670 42 : PhysicalReplicationSlotNewXmin(feedbackXmin, feedbackCatalogXmin);
2671 172 : return;
2672 : }
2673 :
2674 : /*
2675 : * Check that the provided xmin/epoch are sane, that is, not in the future
2676 : * and not so far back as to be already wrapped around. Ignore if not.
2677 : */
2678 92 : if (TransactionIdIsNormal(feedbackXmin) &&
2679 92 : !TransactionIdInRecentPast(feedbackXmin, feedbackEpoch))
2680 0 : return;
2681 :
2682 92 : if (TransactionIdIsNormal(feedbackCatalogXmin) &&
2683 48 : !TransactionIdInRecentPast(feedbackCatalogXmin, feedbackCatalogEpoch))
2684 0 : return;
2685 :
2686 : /*
2687 : * Set the WalSender's xmin equal to the standby's requested xmin, so that
2688 : * the xmin will be taken into account by GetSnapshotData() /
2689 : * ComputeXidHorizons(). This will hold back the removal of dead rows and
2690 : * thereby prevent the generation of cleanup conflicts on the standby
2691 : * server.
2692 : *
2693 : * There is a small window for a race condition here: although we just
2694 : * checked that feedbackXmin precedes nextXid, the nextXid could have
2695 : * gotten advanced between our fetching it and applying the xmin below,
2696 : * perhaps far enough to make feedbackXmin wrap around. In that case the
2697 : * xmin we set here would be "in the future" and have no effect. No point
2698 : * in worrying about this since it's too late to save the desired data
2699 : * anyway. Assuming that the standby sends us an increasing sequence of
2700 : * xmins, this could only happen during the first reply cycle, else our
2701 : * own xmin would prevent nextXid from advancing so far.
2702 : *
2703 : * We don't bother taking the ProcArrayLock here. Setting the xmin field
2704 : * is assumed atomic, and there's no real need to prevent concurrent
2705 : * horizon determinations. (If we're moving our xmin forward, this is
2706 : * obviously safe, and if we're moving it backwards, well, the data is at
2707 : * risk already since a VACUUM could already have determined the horizon.)
2708 : *
2709 : * If we're using a replication slot we reserve the xmin via that,
2710 : * otherwise via the walsender's PGPROC entry. We can only track the
2711 : * catalog xmin separately when using a slot, so we store the least of the
2712 : * two provided when not using a slot.
2713 : *
2714 : * XXX: It might make sense to generalize the ephemeral slot concept and
2715 : * always use the slot mechanism to handle the feedback xmin.
2716 : */
2717 92 : if (MyReplicationSlot != NULL) /* XXX: persistency configurable? */
2718 84 : PhysicalReplicationSlotNewXmin(feedbackXmin, feedbackCatalogXmin);
2719 : else
2720 : {
2721 8 : if (TransactionIdIsNormal(feedbackCatalogXmin)
2722 0 : && TransactionIdPrecedes(feedbackCatalogXmin, feedbackXmin))
2723 0 : MyProc->xmin = feedbackCatalogXmin;
2724 : else
2725 8 : MyProc->xmin = feedbackXmin;
2726 : }
2727 : }
2728 :
2729 : /*
2730 : * Process the request for a primary status update message.
2731 : */
2732 : static void
2733 9256 : ProcessStandbyPSRequestMessage(void)
2734 : {
2735 9256 : XLogRecPtr lsn = InvalidXLogRecPtr;
2736 : TransactionId oldestXidInCommit;
2737 : TransactionId oldestGXidInCommit;
2738 : FullTransactionId nextFullXid;
2739 : FullTransactionId fullOldestXidInCommit;
2740 9256 : WalSnd *walsnd = MyWalSnd;
2741 : TimestampTz replyTime;
2742 :
2743 : /*
2744 : * This shouldn't happen because we don't support getting primary status
2745 : * message from standby.
2746 : */
2747 9256 : if (RecoveryInProgress())
2748 0 : elog(ERROR, "the primary status is unavailable during recovery");
2749 :
2750 9256 : replyTime = pq_getmsgint64(&reply_message);
2751 :
2752 : /*
2753 : * Update shared state for this WalSender process based on reply data from
2754 : * standby.
2755 : */
2756 9256 : SpinLockAcquire(&walsnd->mutex);
2757 9256 : walsnd->replyTime = replyTime;
2758 9256 : SpinLockRelease(&walsnd->mutex);
2759 :
2760 : /*
2761 : * Consider transactions in the current database, as only these are the
2762 : * ones replicated.
2763 : */
2764 9256 : oldestXidInCommit = GetOldestActiveTransactionId(true, false);
2765 9256 : oldestGXidInCommit = TwoPhaseGetOldestXidInCommit();
2766 :
2767 : /*
2768 : * Update the oldest xid for standby transmission if an older prepared
2769 : * transaction exists and is currently in commit phase.
2770 : */
2771 15884 : if (TransactionIdIsValid(oldestGXidInCommit) &&
2772 6628 : TransactionIdPrecedes(oldestGXidInCommit, oldestXidInCommit))
2773 6628 : oldestXidInCommit = oldestGXidInCommit;
2774 :
2775 9256 : nextFullXid = ReadNextFullTransactionId();
2776 9256 : fullOldestXidInCommit = FullTransactionIdFromAllowableAt(nextFullXid,
2777 : oldestXidInCommit);
2778 9256 : lsn = GetXLogWriteRecPtr();
2779 :
2780 9256 : elog(DEBUG2, "sending primary status");
2781 :
2782 : /* construct the message... */
2783 9256 : resetStringInfo(&output_message);
2784 9256 : pq_sendbyte(&output_message, PqReplMsg_PrimaryStatusUpdate);
2785 9256 : pq_sendint64(&output_message, lsn);
2786 9256 : pq_sendint64(&output_message, (int64) U64FromFullTransactionId(fullOldestXidInCommit));
2787 9256 : pq_sendint64(&output_message, (int64) U64FromFullTransactionId(nextFullXid));
2788 9256 : pq_sendint64(&output_message, GetCurrentTimestamp());
2789 :
2790 : /* ... and send it wrapped in CopyData */
2791 9256 : pq_putmessage_noblock(PqMsg_CopyData, output_message.data, output_message.len);
2792 9256 : }
2793 :
2794 : /*
2795 : * Compute how long send/receive loops should sleep.
2796 : *
2797 : * If wal_sender_timeout is enabled we want to wake up in time to send
2798 : * keepalives and to abort the connection if wal_sender_timeout has been
2799 : * reached.
2800 : */
2801 : static long
2802 117858 : WalSndComputeSleeptime(TimestampTz now)
2803 : {
2804 117858 : long sleeptime = 10000; /* 10 s */
2805 :
2806 117858 : if (wal_sender_timeout > 0 && last_reply_timestamp > 0)
2807 : {
2808 : TimestampTz wakeup_time;
2809 :
2810 : /*
2811 : * At the latest stop sleeping once wal_sender_timeout has been
2812 : * reached.
2813 : */
2814 117808 : wakeup_time = TimestampTzPlusMilliseconds(last_reply_timestamp,
2815 : wal_sender_timeout);
2816 :
2817 : /*
2818 : * If no ping has been sent yet, wakeup when it's time to do so.
2819 : * WalSndKeepaliveIfNecessary() wants to send a keepalive once half of
2820 : * the timeout passed without a response.
2821 : */
2822 117808 : if (!waiting_for_ping_response)
2823 107600 : wakeup_time = TimestampTzPlusMilliseconds(last_reply_timestamp,
2824 : wal_sender_timeout / 2);
2825 :
2826 : /* Compute relative time until wakeup. */
2827 117808 : sleeptime = TimestampDifferenceMilliseconds(now, wakeup_time);
2828 : }
2829 :
2830 117858 : return sleeptime;
2831 : }
2832 :
2833 : /*
2834 : * Check whether there have been responses by the client within
2835 : * wal_sender_timeout and shutdown if not. Using last_processing as the
2836 : * reference point avoids counting server-side stalls against the client.
2837 : * However, a long server-side stall can make WalSndKeepaliveIfNecessary()
2838 : * postdate last_processing by more than wal_sender_timeout. If that happens,
2839 : * the client must reply almost immediately to avoid a timeout. This rarely
2840 : * affects the default configuration, under which clients spontaneously send a
2841 : * message every standby_message_timeout = wal_sender_timeout/6 = 10s. We
2842 : * could eliminate that problem by recognizing timeout expiration at
2843 : * wal_sender_timeout/2 after the keepalive.
2844 : */
2845 : static void
2846 1990320 : WalSndCheckTimeOut(void)
2847 : {
2848 : TimestampTz timeout;
2849 :
2850 : /* don't bail out if we're doing something that doesn't require timeouts */
2851 1990320 : if (last_reply_timestamp <= 0)
2852 50 : return;
2853 :
2854 1990270 : timeout = TimestampTzPlusMilliseconds(last_reply_timestamp,
2855 : wal_sender_timeout);
2856 :
2857 1990270 : if (wal_sender_timeout > 0 && last_processing >= timeout)
2858 : {
2859 : /*
2860 : * Since typically expiration of replication timeout means
2861 : * communication problem, we don't send the error message to the
2862 : * standby.
2863 : */
2864 0 : ereport(COMMERROR,
2865 : (errmsg("terminating walsender process due to replication timeout")));
2866 :
2867 0 : WalSndShutdown();
2868 : }
2869 : }
2870 :
2871 : /* Main loop of walsender process that streams the WAL over Copy messages. */
2872 : static void
2873 1324 : WalSndLoop(WalSndSendDataCallback send_data)
2874 : {
2875 1324 : TimestampTz last_flush = 0;
2876 :
2877 : /*
2878 : * Initialize the last reply timestamp. That enables timeout processing
2879 : * from hereon.
2880 : */
2881 1324 : last_reply_timestamp = GetCurrentTimestamp();
2882 1324 : waiting_for_ping_response = false;
2883 :
2884 : /*
2885 : * Loop until we reach the end of this timeline or the client requests to
2886 : * stop streaming.
2887 : */
2888 : for (;;)
2889 : {
2890 : /* Clear any already-pending wakeups */
2891 1975778 : ResetLatch(MyLatch);
2892 :
2893 1975778 : CHECK_FOR_INTERRUPTS();
2894 :
2895 : /* Process any requests or signals received recently */
2896 1975772 : if (ConfigReloadPending)
2897 : {
2898 34 : ConfigReloadPending = false;
2899 34 : ProcessConfigFile(PGC_SIGHUP);
2900 34 : SyncRepInitConfig();
2901 : }
2902 :
2903 : /* Check for input from the client */
2904 1975772 : ProcessRepliesIfAny();
2905 :
2906 : /*
2907 : * If we have received CopyDone from the client, sent CopyDone
2908 : * ourselves, and the output buffer is empty, it's time to exit
2909 : * streaming.
2910 : */
2911 1975596 : if (streamingDoneReceiving && streamingDoneSending &&
2912 1152 : !pq_is_send_pending())
2913 676 : break;
2914 :
2915 : /*
2916 : * If we don't have any pending data in the output buffer, try to send
2917 : * some more. If there is some, we don't bother to call send_data
2918 : * again until we've flushed it ... but we'd better assume we are not
2919 : * caught up.
2920 : */
2921 1974920 : if (!pq_is_send_pending())
2922 1895736 : send_data();
2923 : else
2924 79184 : WalSndCaughtUp = false;
2925 :
2926 : /* Try to flush pending output to the client */
2927 1974528 : if (pq_flush_if_writable() != 0)
2928 0 : WalSndShutdown();
2929 :
2930 : /* If nothing remains to be sent right now ... */
2931 1974528 : if (WalSndCaughtUp && !pq_is_send_pending())
2932 : {
2933 : /*
2934 : * If we're in catchup state, move to streaming. This is an
2935 : * important state change for users to know about, since before
2936 : * this point data loss might occur if the primary dies and we
2937 : * need to failover to the standby. The state change is also
2938 : * important for synchronous replication, since commits that
2939 : * started to wait at that point might wait for some time.
2940 : */
2941 37732 : if (MyWalSnd->state == WALSNDSTATE_CATCHUP)
2942 : {
2943 1092 : ereport(DEBUG1,
2944 : (errmsg_internal("\"%s\" has now caught up with upstream server",
2945 : application_name)));
2946 1092 : WalSndSetState(WALSNDSTATE_STREAMING);
2947 : }
2948 :
2949 : /*
2950 : * When SIGUSR2 arrives, we send any outstanding logs up to the
2951 : * shutdown checkpoint record (i.e., the latest record), wait for
2952 : * them to be replicated to the standby, and exit. This may be a
2953 : * normal termination at shutdown, or a promotion, the walsender
2954 : * is not sure which.
2955 : */
2956 37732 : if (got_SIGUSR2)
2957 12456 : WalSndDone(send_data);
2958 : }
2959 :
2960 : /* Check for replication timeout. */
2961 1974454 : WalSndCheckTimeOut();
2962 :
2963 : /* Send keepalive if the time has come */
2964 1974454 : WalSndKeepaliveIfNecessary();
2965 :
2966 : /*
2967 : * Block if we have unsent data. XXX For logical replication, let
2968 : * WalSndWaitForWal() handle any other blocking; idle receivers need
2969 : * its additional actions. For physical replication, also block if
2970 : * caught up; its send_data does not block.
2971 : *
2972 : * The IO statistics are reported in WalSndWaitForWal() for the
2973 : * logical WAL senders.
2974 : */
2975 1974454 : if ((WalSndCaughtUp && send_data != XLogSendLogical &&
2976 1975746 : !streamingDoneSending) ||
2977 1945868 : pq_is_send_pending())
2978 : {
2979 : long sleeptime;
2980 : int wakeEvents;
2981 : TimestampTz now;
2982 :
2983 102576 : if (!streamingDoneReceiving)
2984 102568 : wakeEvents = WL_SOCKET_READABLE;
2985 : else
2986 8 : wakeEvents = 0;
2987 :
2988 : /*
2989 : * Use fresh timestamp, not last_processing, to reduce the chance
2990 : * of reaching wal_sender_timeout before sending a keepalive.
2991 : */
2992 102576 : now = GetCurrentTimestamp();
2993 102576 : sleeptime = WalSndComputeSleeptime(now);
2994 :
2995 102576 : if (pq_is_send_pending())
2996 78880 : wakeEvents |= WL_SOCKET_WRITEABLE;
2997 :
2998 : /* Report IO statistics, if needed */
2999 102576 : if (TimestampDifferenceExceeds(last_flush, now,
3000 : WALSENDER_STATS_FLUSH_INTERVAL))
3001 : {
3002 1036 : pgstat_flush_io(false);
3003 1036 : (void) pgstat_flush_backend(false, PGSTAT_BACKEND_FLUSH_IO);
3004 1036 : last_flush = now;
3005 : }
3006 :
3007 : /* Sleep until something happens or we time out */
3008 102576 : WalSndWait(wakeEvents, sleeptime, WAIT_EVENT_WAL_SENDER_MAIN);
3009 : }
3010 : }
3011 676 : }
3012 :
3013 : /* Initialize a per-walsender data structure for this walsender process */
3014 : static void
3015 2318 : InitWalSenderSlot(void)
3016 : {
3017 : int i;
3018 :
3019 : /*
3020 : * WalSndCtl should be set up already (we inherit this by fork() or
3021 : * EXEC_BACKEND mechanism from the postmaster).
3022 : */
3023 : Assert(WalSndCtl != NULL);
3024 : Assert(MyWalSnd == NULL);
3025 :
3026 : /*
3027 : * Find a free walsender slot and reserve it. This must not fail due to
3028 : * the prior check for free WAL senders in InitProcess().
3029 : */
3030 3432 : for (i = 0; i < max_wal_senders; i++)
3031 : {
3032 3432 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3033 :
3034 3432 : SpinLockAcquire(&walsnd->mutex);
3035 :
3036 3432 : if (walsnd->pid != 0)
3037 : {
3038 1114 : SpinLockRelease(&walsnd->mutex);
3039 1114 : continue;
3040 : }
3041 : else
3042 : {
3043 : /*
3044 : * Found a free slot. Reserve it for us.
3045 : */
3046 2318 : walsnd->pid = MyProcPid;
3047 2318 : walsnd->state = WALSNDSTATE_STARTUP;
3048 2318 : walsnd->sentPtr = InvalidXLogRecPtr;
3049 2318 : walsnd->needreload = false;
3050 2318 : walsnd->write = InvalidXLogRecPtr;
3051 2318 : walsnd->flush = InvalidXLogRecPtr;
3052 2318 : walsnd->apply = InvalidXLogRecPtr;
3053 2318 : walsnd->writeLag = -1;
3054 2318 : walsnd->flushLag = -1;
3055 2318 : walsnd->applyLag = -1;
3056 2318 : walsnd->sync_standby_priority = 0;
3057 2318 : walsnd->replyTime = 0;
3058 :
3059 : /*
3060 : * The kind assignment is done here and not in StartReplication()
3061 : * and StartLogicalReplication(). Indeed, the logical walsender
3062 : * needs to read WAL records (like snapshot of running
3063 : * transactions) during the slot creation. So it needs to be woken
3064 : * up based on its kind.
3065 : *
3066 : * The kind assignment could also be done in StartReplication(),
3067 : * StartLogicalReplication() and CREATE_REPLICATION_SLOT but it
3068 : * seems better to set it on one place.
3069 : */
3070 2318 : if (MyDatabaseId == InvalidOid)
3071 898 : walsnd->kind = REPLICATION_KIND_PHYSICAL;
3072 : else
3073 1420 : walsnd->kind = REPLICATION_KIND_LOGICAL;
3074 :
3075 2318 : SpinLockRelease(&walsnd->mutex);
3076 : /* don't need the lock anymore */
3077 2318 : MyWalSnd = (WalSnd *) walsnd;
3078 :
3079 2318 : break;
3080 : }
3081 : }
3082 :
3083 : Assert(MyWalSnd != NULL);
3084 :
3085 : /* Arrange to clean up at walsender exit */
3086 2318 : on_shmem_exit(WalSndKill, 0);
3087 2318 : }
3088 :
3089 : /* Destroy the per-walsender data structure for this walsender process */
3090 : static void
3091 2318 : WalSndKill(int code, Datum arg)
3092 : {
3093 2318 : WalSnd *walsnd = MyWalSnd;
3094 :
3095 : Assert(walsnd != NULL);
3096 :
3097 2318 : MyWalSnd = NULL;
3098 :
3099 2318 : SpinLockAcquire(&walsnd->mutex);
3100 : /* Mark WalSnd struct as no longer being in use. */
3101 2318 : walsnd->pid = 0;
3102 2318 : SpinLockRelease(&walsnd->mutex);
3103 2318 : }
3104 :
3105 : /* XLogReaderRoutine->segment_open callback */
3106 : static void
3107 9918 : WalSndSegmentOpen(XLogReaderState *state, XLogSegNo nextSegNo,
3108 : TimeLineID *tli_p)
3109 : {
3110 : char path[MAXPGPATH];
3111 :
3112 : /*-------
3113 : * When reading from a historic timeline, and there is a timeline switch
3114 : * within this segment, read from the WAL segment belonging to the new
3115 : * timeline.
3116 : *
3117 : * For example, imagine that this server is currently on timeline 5, and
3118 : * we're streaming timeline 4. The switch from timeline 4 to 5 happened at
3119 : * 0/13002088. In pg_wal, we have these files:
3120 : *
3121 : * ...
3122 : * 000000040000000000000012
3123 : * 000000040000000000000013
3124 : * 000000050000000000000013
3125 : * 000000050000000000000014
3126 : * ...
3127 : *
3128 : * In this situation, when requested to send the WAL from segment 0x13, on
3129 : * timeline 4, we read the WAL from file 000000050000000000000013. Archive
3130 : * recovery prefers files from newer timelines, so if the segment was
3131 : * restored from the archive on this server, the file belonging to the old
3132 : * timeline, 000000040000000000000013, might not exist. Their contents are
3133 : * equal up to the switchpoint, because at a timeline switch, the used
3134 : * portion of the old segment is copied to the new file.
3135 : */
3136 9918 : *tli_p = sendTimeLine;
3137 9918 : if (sendTimeLineIsHistoric)
3138 : {
3139 : XLogSegNo endSegNo;
3140 :
3141 18 : XLByteToSeg(sendTimeLineValidUpto, endSegNo, state->segcxt.ws_segsize);
3142 18 : if (nextSegNo == endSegNo)
3143 16 : *tli_p = sendTimeLineNextTLI;
3144 : }
3145 :
3146 9918 : XLogFilePath(path, *tli_p, nextSegNo, state->segcxt.ws_segsize);
3147 9918 : state->seg.ws_file = BasicOpenFile(path, O_RDONLY | PG_BINARY);
3148 9918 : if (state->seg.ws_file >= 0)
3149 9918 : return;
3150 :
3151 : /*
3152 : * If the file is not found, assume it's because the standby asked for a
3153 : * too old WAL segment that has already been removed or recycled.
3154 : */
3155 0 : if (errno == ENOENT)
3156 : {
3157 : char xlogfname[MAXFNAMELEN];
3158 0 : int save_errno = errno;
3159 :
3160 0 : XLogFileName(xlogfname, *tli_p, nextSegNo, wal_segment_size);
3161 0 : errno = save_errno;
3162 0 : ereport(ERROR,
3163 : (errcode_for_file_access(),
3164 : errmsg("requested WAL segment %s has already been removed",
3165 : xlogfname)));
3166 : }
3167 : else
3168 0 : ereport(ERROR,
3169 : (errcode_for_file_access(),
3170 : errmsg("could not open file \"%s\": %m",
3171 : path)));
3172 : }
3173 :
3174 : /*
3175 : * Send out the WAL in its normal physical/stored form.
3176 : *
3177 : * Read up to MAX_SEND_SIZE bytes of WAL that's been flushed to disk,
3178 : * but not yet sent to the client, and buffer it in the libpq output
3179 : * buffer.
3180 : *
3181 : * If there is no unsent WAL remaining, WalSndCaughtUp is set to true,
3182 : * otherwise WalSndCaughtUp is set to false.
3183 : */
3184 : static void
3185 230014 : XLogSendPhysical(void)
3186 : {
3187 : XLogRecPtr SendRqstPtr;
3188 : XLogRecPtr startptr;
3189 : XLogRecPtr endptr;
3190 : Size nbytes;
3191 : XLogSegNo segno;
3192 : WALReadError errinfo;
3193 : Size rbytes;
3194 :
3195 : /* If requested switch the WAL sender to the stopping state. */
3196 230014 : if (got_STOPPING)
3197 18778 : WalSndSetState(WALSNDSTATE_STOPPING);
3198 :
3199 230014 : if (streamingDoneSending)
3200 : {
3201 1274 : WalSndCaughtUp = true;
3202 31634 : return;
3203 : }
3204 :
3205 : /* Figure out how far we can safely send the WAL. */
3206 228740 : if (sendTimeLineIsHistoric)
3207 : {
3208 : /*
3209 : * Streaming an old timeline that's in this server's history, but is
3210 : * not the one we're currently inserting or replaying. It can be
3211 : * streamed up to the point where we switched off that timeline.
3212 : */
3213 62 : SendRqstPtr = sendTimeLineValidUpto;
3214 : }
3215 228678 : else if (am_cascading_walsender)
3216 : {
3217 : TimeLineID SendRqstTLI;
3218 :
3219 : /*
3220 : * Streaming the latest timeline on a standby.
3221 : *
3222 : * Attempt to send all WAL that has already been replayed, so that we
3223 : * know it's valid. If we're receiving WAL through streaming
3224 : * replication, it's also OK to send any WAL that has been received
3225 : * but not replayed.
3226 : *
3227 : * The timeline we're recovering from can change, or we can be
3228 : * promoted. In either case, the current timeline becomes historic. We
3229 : * need to detect that so that we don't try to stream past the point
3230 : * where we switched to another timeline. We check for promotion or
3231 : * timeline switch after calculating FlushPtr, to avoid a race
3232 : * condition: if the timeline becomes historic just after we checked
3233 : * that it was still current, it's still be OK to stream it up to the
3234 : * FlushPtr that was calculated before it became historic.
3235 : */
3236 1376 : bool becameHistoric = false;
3237 :
3238 1376 : SendRqstPtr = GetStandbyFlushRecPtr(&SendRqstTLI);
3239 :
3240 1376 : if (!RecoveryInProgress())
3241 : {
3242 : /* We have been promoted. */
3243 0 : SendRqstTLI = GetWALInsertionTimeLine();
3244 0 : am_cascading_walsender = false;
3245 0 : becameHistoric = true;
3246 : }
3247 : else
3248 : {
3249 : /*
3250 : * Still a cascading standby. But is the timeline we're sending
3251 : * still the one recovery is recovering from?
3252 : */
3253 1376 : if (sendTimeLine != SendRqstTLI)
3254 0 : becameHistoric = true;
3255 : }
3256 :
3257 1376 : if (becameHistoric)
3258 : {
3259 : /*
3260 : * The timeline we were sending has become historic. Read the
3261 : * timeline history file of the new timeline to see where exactly
3262 : * we forked off from the timeline we were sending.
3263 : */
3264 : List *history;
3265 :
3266 0 : history = readTimeLineHistory(SendRqstTLI);
3267 0 : sendTimeLineValidUpto = tliSwitchPoint(sendTimeLine, history, &sendTimeLineNextTLI);
3268 :
3269 : Assert(sendTimeLine < sendTimeLineNextTLI);
3270 0 : list_free_deep(history);
3271 :
3272 0 : sendTimeLineIsHistoric = true;
3273 :
3274 0 : SendRqstPtr = sendTimeLineValidUpto;
3275 : }
3276 : }
3277 : else
3278 : {
3279 : /*
3280 : * Streaming the current timeline on a primary.
3281 : *
3282 : * Attempt to send all data that's already been written out and
3283 : * fsync'd to disk. We cannot go further than what's been written out
3284 : * given the current implementation of WALRead(). And in any case
3285 : * it's unsafe to send WAL that is not securely down to disk on the
3286 : * primary: if the primary subsequently crashes and restarts, standbys
3287 : * must not have applied any WAL that got lost on the primary.
3288 : */
3289 227302 : SendRqstPtr = GetFlushRecPtr(NULL);
3290 : }
3291 :
3292 : /*
3293 : * Record the current system time as an approximation of the time at which
3294 : * this WAL location was written for the purposes of lag tracking.
3295 : *
3296 : * In theory we could make XLogFlush() record a time in shmem whenever WAL
3297 : * is flushed and we could get that time as well as the LSN when we call
3298 : * GetFlushRecPtr() above (and likewise for the cascading standby
3299 : * equivalent), but rather than putting any new code into the hot WAL path
3300 : * it seems good enough to capture the time here. We should reach this
3301 : * after XLogFlush() runs WalSndWakeupProcessRequests(), and although that
3302 : * may take some time, we read the WAL flush pointer and take the time
3303 : * very close to together here so that we'll get a later position if it is
3304 : * still moving.
3305 : *
3306 : * Because LagTrackerWrite ignores samples when the LSN hasn't advanced,
3307 : * this gives us a cheap approximation for the WAL flush time for this
3308 : * LSN.
3309 : *
3310 : * Note that the LSN is not necessarily the LSN for the data contained in
3311 : * the present message; it's the end of the WAL, which might be further
3312 : * ahead. All the lag tracking machinery cares about is finding out when
3313 : * that arbitrary LSN is eventually reported as written, flushed and
3314 : * applied, so that it can measure the elapsed time.
3315 : */
3316 228740 : LagTrackerWrite(SendRqstPtr, GetCurrentTimestamp());
3317 :
3318 : /*
3319 : * If this is a historic timeline and we've reached the point where we
3320 : * forked to the next timeline, stop streaming.
3321 : *
3322 : * Note: We might already have sent WAL > sendTimeLineValidUpto. The
3323 : * startup process will normally replay all WAL that has been received
3324 : * from the primary, before promoting, but if the WAL streaming is
3325 : * terminated at a WAL page boundary, the valid portion of the timeline
3326 : * might end in the middle of a WAL record. We might've already sent the
3327 : * first half of that partial WAL record to the cascading standby, so that
3328 : * sentPtr > sendTimeLineValidUpto. That's OK; the cascading standby can't
3329 : * replay the partial WAL record either, so it can still follow our
3330 : * timeline switch.
3331 : */
3332 228740 : if (sendTimeLineIsHistoric && sendTimeLineValidUpto <= sentPtr)
3333 : {
3334 : /* close the current file. */
3335 18 : if (xlogreader->seg.ws_file >= 0)
3336 18 : wal_segment_close(xlogreader);
3337 :
3338 : /* Send CopyDone */
3339 18 : pq_putmessage_noblock(PqMsg_CopyDone, NULL, 0);
3340 18 : streamingDoneSending = true;
3341 :
3342 18 : WalSndCaughtUp = true;
3343 :
3344 18 : elog(DEBUG1, "walsender reached end of timeline at %X/%08X (sent up to %X/%08X)",
3345 : LSN_FORMAT_ARGS(sendTimeLineValidUpto),
3346 : LSN_FORMAT_ARGS(sentPtr));
3347 18 : return;
3348 : }
3349 :
3350 : /* Do we have any work to do? */
3351 : Assert(sentPtr <= SendRqstPtr);
3352 228722 : if (SendRqstPtr <= sentPtr)
3353 : {
3354 30342 : WalSndCaughtUp = true;
3355 30342 : return;
3356 : }
3357 :
3358 : /*
3359 : * Figure out how much to send in one message. If there's no more than
3360 : * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
3361 : * MAX_SEND_SIZE bytes, but round back to logfile or page boundary.
3362 : *
3363 : * The rounding is not only for performance reasons. Walreceiver relies on
3364 : * the fact that we never split a WAL record across two messages. Since a
3365 : * long WAL record is split at page boundary into continuation records,
3366 : * page boundary is always a safe cut-off point. We also assume that
3367 : * SendRqstPtr never points to the middle of a WAL record.
3368 : */
3369 198380 : startptr = sentPtr;
3370 198380 : endptr = startptr;
3371 198380 : endptr += MAX_SEND_SIZE;
3372 :
3373 : /* if we went beyond SendRqstPtr, back off */
3374 198380 : if (SendRqstPtr <= endptr)
3375 : {
3376 7394 : endptr = SendRqstPtr;
3377 7394 : if (sendTimeLineIsHistoric)
3378 18 : WalSndCaughtUp = false;
3379 : else
3380 7376 : WalSndCaughtUp = true;
3381 : }
3382 : else
3383 : {
3384 : /* round down to page boundary. */
3385 190986 : endptr -= (endptr % XLOG_BLCKSZ);
3386 190986 : WalSndCaughtUp = false;
3387 : }
3388 :
3389 198380 : nbytes = endptr - startptr;
3390 : Assert(nbytes <= MAX_SEND_SIZE);
3391 :
3392 : /*
3393 : * OK to read and send the slice.
3394 : */
3395 198380 : resetStringInfo(&output_message);
3396 198380 : pq_sendbyte(&output_message, PqReplMsg_WALData);
3397 :
3398 198380 : pq_sendint64(&output_message, startptr); /* dataStart */
3399 198380 : pq_sendint64(&output_message, SendRqstPtr); /* walEnd */
3400 198380 : pq_sendint64(&output_message, 0); /* sendtime, filled in last */
3401 :
3402 : /*
3403 : * Read the log directly into the output buffer to avoid extra memcpy
3404 : * calls.
3405 : */
3406 198380 : enlargeStringInfo(&output_message, nbytes);
3407 :
3408 198380 : retry:
3409 : /* attempt to read WAL from WAL buffers first */
3410 198380 : rbytes = WALReadFromBuffers(&output_message.data[output_message.len],
3411 198380 : startptr, nbytes, xlogreader->seg.ws_tli);
3412 198380 : output_message.len += rbytes;
3413 198380 : startptr += rbytes;
3414 198380 : nbytes -= rbytes;
3415 :
3416 : /* now read the remaining WAL from WAL file */
3417 198380 : if (nbytes > 0 &&
3418 194684 : !WALRead(xlogreader,
3419 194684 : &output_message.data[output_message.len],
3420 : startptr,
3421 : nbytes,
3422 194684 : xlogreader->seg.ws_tli, /* Pass the current TLI because
3423 : * only WalSndSegmentOpen controls
3424 : * whether new TLI is needed. */
3425 : &errinfo))
3426 0 : WALReadRaiseError(&errinfo);
3427 :
3428 : /* See logical_read_xlog_page(). */
3429 198380 : XLByteToSeg(startptr, segno, xlogreader->segcxt.ws_segsize);
3430 198380 : CheckXLogRemoved(segno, xlogreader->seg.ws_tli);
3431 :
3432 : /*
3433 : * During recovery, the currently-open WAL file might be replaced with the
3434 : * file of the same name retrieved from archive. So we always need to
3435 : * check what we read was valid after reading into the buffer. If it's
3436 : * invalid, we try to open and read the file again.
3437 : */
3438 198380 : if (am_cascading_walsender)
3439 : {
3440 1018 : WalSnd *walsnd = MyWalSnd;
3441 : bool reload;
3442 :
3443 1018 : SpinLockAcquire(&walsnd->mutex);
3444 1018 : reload = walsnd->needreload;
3445 1018 : walsnd->needreload = false;
3446 1018 : SpinLockRelease(&walsnd->mutex);
3447 :
3448 1018 : if (reload && xlogreader->seg.ws_file >= 0)
3449 : {
3450 0 : wal_segment_close(xlogreader);
3451 :
3452 0 : goto retry;
3453 : }
3454 : }
3455 :
3456 198380 : output_message.len += nbytes;
3457 198380 : output_message.data[output_message.len] = '\0';
3458 :
3459 : /*
3460 : * Fill the send timestamp last, so that it is taken as late as possible.
3461 : */
3462 198380 : resetStringInfo(&tmpbuf);
3463 198380 : pq_sendint64(&tmpbuf, GetCurrentTimestamp());
3464 198380 : memcpy(&output_message.data[1 + sizeof(int64) + sizeof(int64)],
3465 198380 : tmpbuf.data, sizeof(int64));
3466 :
3467 198380 : pq_putmessage_noblock(PqMsg_CopyData, output_message.data, output_message.len);
3468 :
3469 198380 : sentPtr = endptr;
3470 :
3471 : /* Update shared memory status */
3472 : {
3473 198380 : WalSnd *walsnd = MyWalSnd;
3474 :
3475 198380 : SpinLockAcquire(&walsnd->mutex);
3476 198380 : walsnd->sentPtr = sentPtr;
3477 198380 : SpinLockRelease(&walsnd->mutex);
3478 : }
3479 :
3480 : /* Report progress of XLOG streaming in PS display */
3481 198380 : if (update_process_title)
3482 : {
3483 : char activitymsg[50];
3484 :
3485 198380 : snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%08X",
3486 198380 : LSN_FORMAT_ARGS(sentPtr));
3487 198380 : set_ps_display(activitymsg);
3488 : }
3489 : }
3490 :
3491 : /*
3492 : * Stream out logically decoded data.
3493 : */
3494 : static void
3495 1678178 : XLogSendLogical(void)
3496 : {
3497 : XLogRecord *record;
3498 : char *errm;
3499 :
3500 : /*
3501 : * We'll use the current flush point to determine whether we've caught up.
3502 : * This variable is static in order to cache it across calls. Caching is
3503 : * helpful because GetFlushRecPtr() needs to acquire a heavily-contended
3504 : * spinlock.
3505 : */
3506 : static XLogRecPtr flushPtr = InvalidXLogRecPtr;
3507 :
3508 : /*
3509 : * Don't know whether we've caught up yet. We'll set WalSndCaughtUp to
3510 : * true in WalSndWaitForWal, if we're actually waiting. We also set to
3511 : * true if XLogReadRecord() had to stop reading but WalSndWaitForWal
3512 : * didn't wait - i.e. when we're shutting down.
3513 : */
3514 1678178 : WalSndCaughtUp = false;
3515 :
3516 1678178 : record = XLogReadRecord(logical_decoding_ctx->reader, &errm);
3517 :
3518 : /* xlog record was invalid */
3519 1677822 : if (errm != NULL)
3520 0 : elog(ERROR, "could not find record while sending logically-decoded data: %s",
3521 : errm);
3522 :
3523 1677822 : if (record != NULL)
3524 : {
3525 : /*
3526 : * Note the lack of any call to LagTrackerWrite() which is handled by
3527 : * WalSndUpdateProgress which is called by output plugin through
3528 : * logical decoding write api.
3529 : */
3530 1670886 : LogicalDecodingProcessRecord(logical_decoding_ctx, logical_decoding_ctx->reader);
3531 :
3532 1670850 : sentPtr = logical_decoding_ctx->reader->EndRecPtr;
3533 : }
3534 :
3535 : /*
3536 : * If first time through in this session, initialize flushPtr. Otherwise,
3537 : * we only need to update flushPtr if EndRecPtr is past it.
3538 : */
3539 1677786 : if (!XLogRecPtrIsValid(flushPtr) ||
3540 1677166 : logical_decoding_ctx->reader->EndRecPtr >= flushPtr)
3541 : {
3542 : /*
3543 : * For cascading logical WAL senders, we use the replay LSN instead of
3544 : * the flush LSN, since logical decoding on a standby only processes
3545 : * WAL that has been replayed. This distinction becomes particularly
3546 : * important during shutdown, as new WAL is no longer replayed and the
3547 : * last replayed LSN marks the furthest point up to which decoding can
3548 : * proceed.
3549 : */
3550 12076 : if (am_cascading_walsender)
3551 542 : flushPtr = GetXLogReplayRecPtr(NULL);
3552 : else
3553 11534 : flushPtr = GetFlushRecPtr(NULL);
3554 : }
3555 :
3556 : /* If EndRecPtr is still past our flushPtr, it means we caught up. */
3557 1677786 : if (logical_decoding_ctx->reader->EndRecPtr >= flushPtr)
3558 9802 : WalSndCaughtUp = true;
3559 :
3560 : /*
3561 : * If we're caught up and have been requested to stop, have WalSndLoop()
3562 : * terminate the connection in an orderly manner, after writing out all
3563 : * the pending data.
3564 : */
3565 1677786 : if (WalSndCaughtUp && got_STOPPING)
3566 6750 : got_SIGUSR2 = true;
3567 :
3568 : /* Update shared memory status */
3569 : {
3570 1677786 : WalSnd *walsnd = MyWalSnd;
3571 :
3572 1677786 : SpinLockAcquire(&walsnd->mutex);
3573 1677786 : walsnd->sentPtr = sentPtr;
3574 1677786 : SpinLockRelease(&walsnd->mutex);
3575 : }
3576 1677786 : }
3577 :
3578 : /*
3579 : * Shutdown if the sender is caught up.
3580 : *
3581 : * NB: This should only be called when the shutdown signal has been received
3582 : * from postmaster.
3583 : *
3584 : * Note that if we determine that there's still more data to send, this
3585 : * function will return control to the caller.
3586 : */
3587 : static void
3588 12456 : WalSndDone(WalSndSendDataCallback send_data)
3589 : {
3590 : XLogRecPtr replicatedPtr;
3591 :
3592 : /* ... let's just be real sure we're caught up ... */
3593 12456 : send_data();
3594 :
3595 : /*
3596 : * To figure out whether all WAL has successfully been replicated, check
3597 : * flush location if valid, write otherwise. Tools like pg_receivewal will
3598 : * usually (unless in synchronous mode) return an invalid flush location.
3599 : */
3600 24912 : replicatedPtr = XLogRecPtrIsValid(MyWalSnd->flush) ?
3601 12456 : MyWalSnd->flush : MyWalSnd->write;
3602 :
3603 12456 : if (WalSndCaughtUp && sentPtr == replicatedPtr &&
3604 74 : !pq_is_send_pending())
3605 : {
3606 : QueryCompletion qc;
3607 :
3608 : /* Inform the standby that XLOG streaming is done */
3609 74 : SetQueryCompletion(&qc, CMDTAG_COPY, 0);
3610 74 : EndCommand(&qc, DestRemote, false);
3611 74 : pq_flush();
3612 :
3613 74 : proc_exit(0);
3614 : }
3615 12382 : if (!waiting_for_ping_response)
3616 5696 : WalSndKeepalive(true, InvalidXLogRecPtr);
3617 12382 : }
3618 :
3619 : /*
3620 : * Returns the latest point in WAL that has been safely flushed to disk.
3621 : * This should only be called when in recovery.
3622 : *
3623 : * This is called either by cascading walsender to find WAL position to be sent
3624 : * to a cascaded standby or by slot synchronization operation to validate remote
3625 : * slot's lsn before syncing it locally.
3626 : *
3627 : * As a side-effect, *tli is updated to the TLI of the last
3628 : * replayed WAL record.
3629 : */
3630 : XLogRecPtr
3631 1584 : GetStandbyFlushRecPtr(TimeLineID *tli)
3632 : {
3633 : XLogRecPtr replayPtr;
3634 : TimeLineID replayTLI;
3635 : XLogRecPtr receivePtr;
3636 : TimeLineID receiveTLI;
3637 : XLogRecPtr result;
3638 :
3639 : Assert(am_cascading_walsender || IsSyncingReplicationSlots());
3640 :
3641 : /*
3642 : * We can safely send what's already been replayed. Also, if walreceiver
3643 : * is streaming WAL from the same timeline, we can send anything that it
3644 : * has streamed, but hasn't been replayed yet.
3645 : */
3646 :
3647 1584 : receivePtr = GetWalRcvFlushRecPtr(NULL, &receiveTLI);
3648 1584 : replayPtr = GetXLogReplayRecPtr(&replayTLI);
3649 :
3650 1584 : if (tli)
3651 1510 : *tli = replayTLI;
3652 :
3653 1584 : result = replayPtr;
3654 1584 : if (receiveTLI == replayTLI && receivePtr > replayPtr)
3655 50 : result = receivePtr;
3656 :
3657 1584 : return result;
3658 : }
3659 :
3660 : /*
3661 : * Request walsenders to reload the currently-open WAL file
3662 : */
3663 : void
3664 56 : WalSndRqstFileReload(void)
3665 : {
3666 : int i;
3667 :
3668 568 : for (i = 0; i < max_wal_senders; i++)
3669 : {
3670 512 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3671 :
3672 512 : SpinLockAcquire(&walsnd->mutex);
3673 512 : if (walsnd->pid == 0)
3674 : {
3675 512 : SpinLockRelease(&walsnd->mutex);
3676 512 : continue;
3677 : }
3678 0 : walsnd->needreload = true;
3679 0 : SpinLockRelease(&walsnd->mutex);
3680 : }
3681 56 : }
3682 :
3683 : /*
3684 : * Handle PROCSIG_WALSND_INIT_STOPPING signal.
3685 : */
3686 : void
3687 74 : HandleWalSndInitStopping(void)
3688 : {
3689 : Assert(am_walsender);
3690 :
3691 : /*
3692 : * If replication has not yet started, die like with SIGTERM. If
3693 : * replication is active, only set a flag and wake up the main loop. It
3694 : * will send any outstanding WAL, wait for it to be replicated to the
3695 : * standby, and then exit gracefully.
3696 : */
3697 74 : if (!replication_active)
3698 0 : kill(MyProcPid, SIGTERM);
3699 : else
3700 74 : got_STOPPING = true;
3701 74 : }
3702 :
3703 : /*
3704 : * SIGUSR2: set flag to do a last cycle and shut down afterwards. The WAL
3705 : * sender should already have been switched to WALSNDSTATE_STOPPING at
3706 : * this point.
3707 : */
3708 : static void
3709 64 : WalSndLastCycleHandler(SIGNAL_ARGS)
3710 : {
3711 64 : got_SIGUSR2 = true;
3712 64 : SetLatch(MyLatch);
3713 64 : }
3714 :
3715 : /* Set up signal handlers */
3716 : void
3717 2318 : WalSndSignals(void)
3718 : {
3719 : /* Set up signal handlers */
3720 2318 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
3721 2318 : pqsignal(SIGINT, StatementCancelHandler); /* query cancel */
3722 2318 : pqsignal(SIGTERM, die); /* request shutdown */
3723 : /* SIGQUIT handler was already set up by InitPostmasterChild */
3724 2318 : InitializeTimeouts(); /* establishes SIGALRM handler */
3725 2318 : pqsignal(SIGPIPE, SIG_IGN);
3726 2318 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
3727 2318 : pqsignal(SIGUSR2, WalSndLastCycleHandler); /* request a last cycle and
3728 : * shutdown */
3729 :
3730 : /* Reset some signals that are accepted by postmaster but not here */
3731 2318 : pqsignal(SIGCHLD, SIG_DFL);
3732 2318 : }
3733 :
3734 : /* Report shared-memory space needed by WalSndShmemInit */
3735 : Size
3736 8500 : WalSndShmemSize(void)
3737 : {
3738 8500 : Size size = 0;
3739 :
3740 8500 : size = offsetof(WalSndCtlData, walsnds);
3741 8500 : size = add_size(size, mul_size(max_wal_senders, sizeof(WalSnd)));
3742 :
3743 8500 : return size;
3744 : }
3745 :
3746 : /* Allocate and initialize walsender-related shared memory */
3747 : void
3748 2200 : WalSndShmemInit(void)
3749 : {
3750 : bool found;
3751 : int i;
3752 :
3753 2200 : WalSndCtl = (WalSndCtlData *)
3754 2200 : ShmemInitStruct("Wal Sender Ctl", WalSndShmemSize(), &found);
3755 :
3756 2200 : if (!found)
3757 : {
3758 : /* First time through, so initialize */
3759 15740 : MemSet(WalSndCtl, 0, WalSndShmemSize());
3760 :
3761 8800 : for (i = 0; i < NUM_SYNC_REP_WAIT_MODE; i++)
3762 6600 : dlist_init(&(WalSndCtl->SyncRepQueue[i]));
3763 :
3764 16596 : for (i = 0; i < max_wal_senders; i++)
3765 : {
3766 14396 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3767 :
3768 14396 : SpinLockInit(&walsnd->mutex);
3769 : }
3770 :
3771 2200 : ConditionVariableInit(&WalSndCtl->wal_flush_cv);
3772 2200 : ConditionVariableInit(&WalSndCtl->wal_replay_cv);
3773 2200 : ConditionVariableInit(&WalSndCtl->wal_confirm_rcv_cv);
3774 : }
3775 2200 : }
3776 :
3777 : /*
3778 : * Wake up physical, logical or both kinds of walsenders
3779 : *
3780 : * The distinction between physical and logical walsenders is done, because:
3781 : * - physical walsenders can't send data until it's been flushed
3782 : * - logical walsenders on standby can't decode and send data until it's been
3783 : * applied
3784 : *
3785 : * For cascading replication we need to wake up physical walsenders separately
3786 : * from logical walsenders (see the comment before calling WalSndWakeup() in
3787 : * ApplyWalRecord() for more details).
3788 : *
3789 : * This will be called inside critical sections, so throwing an error is not
3790 : * advisable.
3791 : */
3792 : void
3793 5377618 : WalSndWakeup(bool physical, bool logical)
3794 : {
3795 : /*
3796 : * Wake up all the walsenders waiting on WAL being flushed or replayed
3797 : * respectively. Note that waiting walsender would have prepared to sleep
3798 : * on the CV (i.e., added itself to the CV's waitlist) in WalSndWait()
3799 : * before actually waiting.
3800 : */
3801 5377618 : if (physical)
3802 254230 : ConditionVariableBroadcast(&WalSndCtl->wal_flush_cv);
3803 :
3804 5377618 : if (logical)
3805 5329296 : ConditionVariableBroadcast(&WalSndCtl->wal_replay_cv);
3806 5377618 : }
3807 :
3808 : /*
3809 : * Wait for readiness on the FeBe socket, or a timeout. The mask should be
3810 : * composed of optional WL_SOCKET_WRITEABLE and WL_SOCKET_READABLE flags. Exit
3811 : * on postmaster death.
3812 : */
3813 : static void
3814 117858 : WalSndWait(uint32 socket_events, long timeout, uint32 wait_event)
3815 : {
3816 : WaitEvent event;
3817 :
3818 117858 : ModifyWaitEvent(FeBeWaitSet, FeBeWaitSetSocketPos, socket_events, NULL);
3819 :
3820 : /*
3821 : * We use a condition variable to efficiently wake up walsenders in
3822 : * WalSndWakeup().
3823 : *
3824 : * Every walsender prepares to sleep on a shared memory CV. Note that it
3825 : * just prepares to sleep on the CV (i.e., adds itself to the CV's
3826 : * waitlist), but does not actually wait on the CV (IOW, it never calls
3827 : * ConditionVariableSleep()). It still uses WaitEventSetWait() for
3828 : * waiting, because we also need to wait for socket events. The processes
3829 : * (startup process, walreceiver etc.) wanting to wake up walsenders use
3830 : * ConditionVariableBroadcast(), which in turn calls SetLatch(), helping
3831 : * walsenders come out of WaitEventSetWait().
3832 : *
3833 : * This approach is simple and efficient because, one doesn't have to loop
3834 : * through all the walsenders slots, with a spinlock acquisition and
3835 : * release for every iteration, just to wake up only the waiting
3836 : * walsenders. It makes WalSndWakeup() callers' life easy.
3837 : *
3838 : * XXX: A desirable future improvement would be to add support for CVs
3839 : * into WaitEventSetWait().
3840 : *
3841 : * And, we use separate shared memory CVs for physical and logical
3842 : * walsenders for selective wake ups, see WalSndWakeup() for more details.
3843 : *
3844 : * If the wait event is WAIT_FOR_STANDBY_CONFIRMATION, wait on another CV
3845 : * until awakened by physical walsenders after the walreceiver confirms
3846 : * the receipt of the LSN.
3847 : */
3848 117858 : if (wait_event == WAIT_EVENT_WAIT_FOR_STANDBY_CONFIRMATION)
3849 14 : ConditionVariablePrepareToSleep(&WalSndCtl->wal_confirm_rcv_cv);
3850 117844 : else if (MyWalSnd->kind == REPLICATION_KIND_PHYSICAL)
3851 101396 : ConditionVariablePrepareToSleep(&WalSndCtl->wal_flush_cv);
3852 16448 : else if (MyWalSnd->kind == REPLICATION_KIND_LOGICAL)
3853 16448 : ConditionVariablePrepareToSleep(&WalSndCtl->wal_replay_cv);
3854 :
3855 117858 : if (WaitEventSetWait(FeBeWaitSet, timeout, &event, 1, wait_event) == 1 &&
3856 117858 : (event.events & WL_POSTMASTER_DEATH))
3857 : {
3858 0 : ConditionVariableCancelSleep();
3859 0 : proc_exit(1);
3860 : }
3861 :
3862 117858 : ConditionVariableCancelSleep();
3863 117858 : }
3864 :
3865 : /*
3866 : * Signal all walsenders to move to stopping state.
3867 : *
3868 : * This will trigger walsenders to move to a state where no further WAL can be
3869 : * generated. See this file's header for details.
3870 : */
3871 : void
3872 1314 : WalSndInitStopping(void)
3873 : {
3874 : int i;
3875 :
3876 10078 : for (i = 0; i < max_wal_senders; i++)
3877 : {
3878 8764 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3879 : pid_t pid;
3880 :
3881 8764 : SpinLockAcquire(&walsnd->mutex);
3882 8764 : pid = walsnd->pid;
3883 8764 : SpinLockRelease(&walsnd->mutex);
3884 :
3885 8764 : if (pid == 0)
3886 8690 : continue;
3887 :
3888 74 : SendProcSignal(pid, PROCSIG_WALSND_INIT_STOPPING, INVALID_PROC_NUMBER);
3889 : }
3890 1314 : }
3891 :
3892 : /*
3893 : * Wait that all the WAL senders have quit or reached the stopping state. This
3894 : * is used by the checkpointer to control when the shutdown checkpoint can
3895 : * safely be performed.
3896 : */
3897 : void
3898 1314 : WalSndWaitStopping(void)
3899 : {
3900 : for (;;)
3901 68 : {
3902 : int i;
3903 1382 : bool all_stopped = true;
3904 :
3905 10150 : for (i = 0; i < max_wal_senders; i++)
3906 : {
3907 8836 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3908 :
3909 8836 : SpinLockAcquire(&walsnd->mutex);
3910 :
3911 8836 : if (walsnd->pid == 0)
3912 : {
3913 8716 : SpinLockRelease(&walsnd->mutex);
3914 8716 : continue;
3915 : }
3916 :
3917 120 : if (walsnd->state != WALSNDSTATE_STOPPING)
3918 : {
3919 68 : all_stopped = false;
3920 68 : SpinLockRelease(&walsnd->mutex);
3921 68 : break;
3922 : }
3923 52 : SpinLockRelease(&walsnd->mutex);
3924 : }
3925 :
3926 : /* safe to leave if confirmation is done for all WAL senders */
3927 1382 : if (all_stopped)
3928 1314 : return;
3929 :
3930 68 : pg_usleep(10000L); /* wait for 10 msec */
3931 : }
3932 : }
3933 :
3934 : /* Set state for current walsender (only called in walsender) */
3935 : void
3936 22300 : WalSndSetState(WalSndState state)
3937 : {
3938 22300 : WalSnd *walsnd = MyWalSnd;
3939 :
3940 : Assert(am_walsender);
3941 :
3942 22300 : if (walsnd->state == state)
3943 18798 : return;
3944 :
3945 3502 : SpinLockAcquire(&walsnd->mutex);
3946 3502 : walsnd->state = state;
3947 3502 : SpinLockRelease(&walsnd->mutex);
3948 : }
3949 :
3950 : /*
3951 : * Return a string constant representing the state. This is used
3952 : * in system views, and should *not* be translated.
3953 : */
3954 : static const char *
3955 1810 : WalSndGetStateString(WalSndState state)
3956 : {
3957 1810 : switch (state)
3958 : {
3959 12 : case WALSNDSTATE_STARTUP:
3960 12 : return "startup";
3961 0 : case WALSNDSTATE_BACKUP:
3962 0 : return "backup";
3963 30 : case WALSNDSTATE_CATCHUP:
3964 30 : return "catchup";
3965 1768 : case WALSNDSTATE_STREAMING:
3966 1768 : return "streaming";
3967 0 : case WALSNDSTATE_STOPPING:
3968 0 : return "stopping";
3969 : }
3970 0 : return "UNKNOWN";
3971 : }
3972 :
3973 : static Interval *
3974 2698 : offset_to_interval(TimeOffset offset)
3975 : {
3976 2698 : Interval *result = palloc(sizeof(Interval));
3977 :
3978 2698 : result->month = 0;
3979 2698 : result->day = 0;
3980 2698 : result->time = offset;
3981 :
3982 2698 : return result;
3983 : }
3984 :
3985 : /*
3986 : * Returns activity of walsenders, including pids and xlog locations sent to
3987 : * standby servers.
3988 : */
3989 : Datum
3990 1534 : pg_stat_get_wal_senders(PG_FUNCTION_ARGS)
3991 : {
3992 : #define PG_STAT_GET_WAL_SENDERS_COLS 12
3993 1534 : ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
3994 : SyncRepStandbyData *sync_standbys;
3995 : int num_standbys;
3996 : int i;
3997 :
3998 1534 : InitMaterializedSRF(fcinfo, 0);
3999 :
4000 : /*
4001 : * Get the currently active synchronous standbys. This could be out of
4002 : * date before we're done, but we'll use the data anyway.
4003 : */
4004 1534 : num_standbys = SyncRepGetCandidateStandbys(&sync_standbys);
4005 :
4006 16342 : for (i = 0; i < max_wal_senders; i++)
4007 : {
4008 14808 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
4009 : XLogRecPtr sent_ptr;
4010 : XLogRecPtr write;
4011 : XLogRecPtr flush;
4012 : XLogRecPtr apply;
4013 : TimeOffset writeLag;
4014 : TimeOffset flushLag;
4015 : TimeOffset applyLag;
4016 : int priority;
4017 : int pid;
4018 : WalSndState state;
4019 : TimestampTz replyTime;
4020 : bool is_sync_standby;
4021 : Datum values[PG_STAT_GET_WAL_SENDERS_COLS];
4022 14808 : bool nulls[PG_STAT_GET_WAL_SENDERS_COLS] = {0};
4023 : int j;
4024 :
4025 : /* Collect data from shared memory */
4026 14808 : SpinLockAcquire(&walsnd->mutex);
4027 14808 : if (walsnd->pid == 0)
4028 : {
4029 12998 : SpinLockRelease(&walsnd->mutex);
4030 12998 : continue;
4031 : }
4032 1810 : pid = walsnd->pid;
4033 1810 : sent_ptr = walsnd->sentPtr;
4034 1810 : state = walsnd->state;
4035 1810 : write = walsnd->write;
4036 1810 : flush = walsnd->flush;
4037 1810 : apply = walsnd->apply;
4038 1810 : writeLag = walsnd->writeLag;
4039 1810 : flushLag = walsnd->flushLag;
4040 1810 : applyLag = walsnd->applyLag;
4041 1810 : priority = walsnd->sync_standby_priority;
4042 1810 : replyTime = walsnd->replyTime;
4043 1810 : SpinLockRelease(&walsnd->mutex);
4044 :
4045 : /*
4046 : * Detect whether walsender is/was considered synchronous. We can
4047 : * provide some protection against stale data by checking the PID
4048 : * along with walsnd_index.
4049 : */
4050 1810 : is_sync_standby = false;
4051 1892 : for (j = 0; j < num_standbys; j++)
4052 : {
4053 136 : if (sync_standbys[j].walsnd_index == i &&
4054 54 : sync_standbys[j].pid == pid)
4055 : {
4056 54 : is_sync_standby = true;
4057 54 : break;
4058 : }
4059 : }
4060 :
4061 1810 : values[0] = Int32GetDatum(pid);
4062 :
4063 1810 : if (!has_privs_of_role(GetUserId(), ROLE_PG_READ_ALL_STATS))
4064 : {
4065 : /*
4066 : * Only superusers and roles with privileges of pg_read_all_stats
4067 : * can see details. Other users only get the pid value to know
4068 : * it's a walsender, but no details.
4069 : */
4070 0 : MemSet(&nulls[1], true, PG_STAT_GET_WAL_SENDERS_COLS - 1);
4071 : }
4072 : else
4073 : {
4074 1810 : values[1] = CStringGetTextDatum(WalSndGetStateString(state));
4075 :
4076 1810 : if (!XLogRecPtrIsValid(sent_ptr))
4077 10 : nulls[2] = true;
4078 1810 : values[2] = LSNGetDatum(sent_ptr);
4079 :
4080 1810 : if (!XLogRecPtrIsValid(write))
4081 24 : nulls[3] = true;
4082 1810 : values[3] = LSNGetDatum(write);
4083 :
4084 1810 : if (!XLogRecPtrIsValid(flush))
4085 24 : nulls[4] = true;
4086 1810 : values[4] = LSNGetDatum(flush);
4087 :
4088 1810 : if (!XLogRecPtrIsValid(apply))
4089 24 : nulls[5] = true;
4090 1810 : values[5] = LSNGetDatum(apply);
4091 :
4092 : /*
4093 : * Treat a standby such as a pg_basebackup background process
4094 : * which always returns an invalid flush location, as an
4095 : * asynchronous standby.
4096 : */
4097 1810 : priority = XLogRecPtrIsValid(flush) ? priority : 0;
4098 :
4099 1810 : if (writeLag < 0)
4100 958 : nulls[6] = true;
4101 : else
4102 852 : values[6] = IntervalPGetDatum(offset_to_interval(writeLag));
4103 :
4104 1810 : if (flushLag < 0)
4105 812 : nulls[7] = true;
4106 : else
4107 998 : values[7] = IntervalPGetDatum(offset_to_interval(flushLag));
4108 :
4109 1810 : if (applyLag < 0)
4110 962 : nulls[8] = true;
4111 : else
4112 848 : values[8] = IntervalPGetDatum(offset_to_interval(applyLag));
4113 :
4114 1810 : values[9] = Int32GetDatum(priority);
4115 :
4116 : /*
4117 : * More easily understood version of standby state. This is purely
4118 : * informational.
4119 : *
4120 : * In quorum-based sync replication, the role of each standby
4121 : * listed in synchronous_standby_names can be changing very
4122 : * frequently. Any standbys considered as "sync" at one moment can
4123 : * be switched to "potential" ones at the next moment. So, it's
4124 : * basically useless to report "sync" or "potential" as their sync
4125 : * states. We report just "quorum" for them.
4126 : */
4127 1810 : if (priority == 0)
4128 1734 : values[10] = CStringGetTextDatum("async");
4129 76 : else if (is_sync_standby)
4130 54 : values[10] = SyncRepConfig->syncrep_method == SYNC_REP_PRIORITY ?
4131 54 : CStringGetTextDatum("sync") : CStringGetTextDatum("quorum");
4132 : else
4133 22 : values[10] = CStringGetTextDatum("potential");
4134 :
4135 1810 : if (replyTime == 0)
4136 10 : nulls[11] = true;
4137 : else
4138 1800 : values[11] = TimestampTzGetDatum(replyTime);
4139 : }
4140 :
4141 1810 : tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc,
4142 : values, nulls);
4143 : }
4144 :
4145 1534 : return (Datum) 0;
4146 : }
4147 :
4148 : /*
4149 : * Send a keepalive message to standby.
4150 : *
4151 : * If requestReply is set, the message requests the other party to send
4152 : * a message back to us, for heartbeat purposes. We also set a flag to
4153 : * let nearby code know that we're waiting for that response, to avoid
4154 : * repeated requests.
4155 : *
4156 : * writePtr is the location up to which the WAL is sent. It is essentially
4157 : * the same as sentPtr but in some cases, we need to send keep alive before
4158 : * sentPtr is updated like when skipping empty transactions.
4159 : */
4160 : static void
4161 9350 : WalSndKeepalive(bool requestReply, XLogRecPtr writePtr)
4162 : {
4163 9350 : elog(DEBUG2, "sending replication keepalive");
4164 :
4165 : /* construct the message... */
4166 9350 : resetStringInfo(&output_message);
4167 9350 : pq_sendbyte(&output_message, PqReplMsg_Keepalive);
4168 9350 : pq_sendint64(&output_message, XLogRecPtrIsValid(writePtr) ? writePtr : sentPtr);
4169 9350 : pq_sendint64(&output_message, GetCurrentTimestamp());
4170 9350 : pq_sendbyte(&output_message, requestReply ? 1 : 0);
4171 :
4172 : /* ... and send it wrapped in CopyData */
4173 9350 : pq_putmessage_noblock(PqMsg_CopyData, output_message.data, output_message.len);
4174 :
4175 : /* Set local flag */
4176 9350 : if (requestReply)
4177 5696 : waiting_for_ping_response = true;
4178 9350 : }
4179 :
4180 : /*
4181 : * Send keepalive message if too much time has elapsed.
4182 : */
4183 : static void
4184 1990320 : WalSndKeepaliveIfNecessary(void)
4185 : {
4186 : TimestampTz ping_time;
4187 :
4188 : /*
4189 : * Don't send keepalive messages if timeouts are globally disabled or
4190 : * we're doing something not partaking in timeouts.
4191 : */
4192 1990320 : if (wal_sender_timeout <= 0 || last_reply_timestamp <= 0)
4193 50 : return;
4194 :
4195 1990270 : if (waiting_for_ping_response)
4196 18076 : return;
4197 :
4198 : /*
4199 : * If half of wal_sender_timeout has lapsed without receiving any reply
4200 : * from the standby, send a keep-alive message to the standby requesting
4201 : * an immediate reply.
4202 : */
4203 1972194 : ping_time = TimestampTzPlusMilliseconds(last_reply_timestamp,
4204 : wal_sender_timeout / 2);
4205 1972194 : if (last_processing >= ping_time)
4206 : {
4207 0 : WalSndKeepalive(true, InvalidXLogRecPtr);
4208 :
4209 : /* Try to flush pending output to the client */
4210 0 : if (pq_flush_if_writable() != 0)
4211 0 : WalSndShutdown();
4212 : }
4213 : }
4214 :
4215 : /*
4216 : * Record the end of the WAL and the time it was flushed locally, so that
4217 : * LagTrackerRead can compute the elapsed time (lag) when this WAL location is
4218 : * eventually reported to have been written, flushed and applied by the
4219 : * standby in a reply message.
4220 : */
4221 : static void
4222 229192 : LagTrackerWrite(XLogRecPtr lsn, TimestampTz local_flush_time)
4223 : {
4224 : int new_write_head;
4225 : int i;
4226 :
4227 229192 : if (!am_walsender)
4228 0 : return;
4229 :
4230 : /*
4231 : * If the lsn hasn't advanced since last time, then do nothing. This way
4232 : * we only record a new sample when new WAL has been written.
4233 : */
4234 229192 : if (lag_tracker->last_lsn == lsn)
4235 217568 : return;
4236 11624 : lag_tracker->last_lsn = lsn;
4237 :
4238 : /*
4239 : * If advancing the write head of the circular buffer would crash into any
4240 : * of the read heads, then the buffer is full. In other words, the
4241 : * slowest reader (presumably apply) is the one that controls the release
4242 : * of space.
4243 : */
4244 11624 : new_write_head = (lag_tracker->write_head + 1) % LAG_TRACKER_BUFFER_SIZE;
4245 46496 : for (i = 0; i < NUM_SYNC_REP_WAIT_MODE; ++i)
4246 : {
4247 : /*
4248 : * If the buffer is full, move the slowest reader to a separate
4249 : * overflow entry and free its space in the buffer so the write head
4250 : * can advance.
4251 : */
4252 34872 : if (new_write_head == lag_tracker->read_heads[i])
4253 : {
4254 0 : lag_tracker->overflowed[i] =
4255 0 : lag_tracker->buffer[lag_tracker->read_heads[i]];
4256 0 : lag_tracker->read_heads[i] = -1;
4257 : }
4258 : }
4259 :
4260 : /* Store a sample at the current write head position. */
4261 11624 : lag_tracker->buffer[lag_tracker->write_head].lsn = lsn;
4262 11624 : lag_tracker->buffer[lag_tracker->write_head].time = local_flush_time;
4263 11624 : lag_tracker->write_head = new_write_head;
4264 : }
4265 :
4266 : /*
4267 : * Find out how much time has elapsed between the moment WAL location 'lsn'
4268 : * (or the highest known earlier LSN) was flushed locally and the time 'now'.
4269 : * We have a separate read head for each of the reported LSN locations we
4270 : * receive in replies from standby; 'head' controls which read head is
4271 : * used. Whenever a read head crosses an LSN which was written into the
4272 : * lag buffer with LagTrackerWrite, we can use the associated timestamp to
4273 : * find out the time this LSN (or an earlier one) was flushed locally, and
4274 : * therefore compute the lag.
4275 : *
4276 : * Return -1 if no new sample data is available, and otherwise the elapsed
4277 : * time in microseconds.
4278 : */
4279 : static TimeOffset
4280 567702 : LagTrackerRead(int head, XLogRecPtr lsn, TimestampTz now)
4281 : {
4282 567702 : TimestampTz time = 0;
4283 :
4284 : /*
4285 : * If 'lsn' has not passed the WAL position stored in the overflow entry,
4286 : * return the elapsed time (in microseconds) since the saved local flush
4287 : * time. If the flush time is in the future (due to clock drift), return
4288 : * -1 to treat as no valid sample.
4289 : *
4290 : * Otherwise, switch back to using the buffer to control the read head and
4291 : * compute the elapsed time. The read head is then reset to point to the
4292 : * oldest entry in the buffer.
4293 : */
4294 567702 : if (lag_tracker->read_heads[head] == -1)
4295 : {
4296 0 : if (lag_tracker->overflowed[head].lsn > lsn)
4297 0 : return (now >= lag_tracker->overflowed[head].time) ?
4298 0 : now - lag_tracker->overflowed[head].time : -1;
4299 :
4300 0 : time = lag_tracker->overflowed[head].time;
4301 0 : lag_tracker->last_read[head] = lag_tracker->overflowed[head];
4302 0 : lag_tracker->read_heads[head] =
4303 0 : (lag_tracker->write_head + 1) % LAG_TRACKER_BUFFER_SIZE;
4304 : }
4305 :
4306 : /* Read all unread samples up to this LSN or end of buffer. */
4307 600314 : while (lag_tracker->read_heads[head] != lag_tracker->write_head &&
4308 301498 : lag_tracker->buffer[lag_tracker->read_heads[head]].lsn <= lsn)
4309 : {
4310 32612 : time = lag_tracker->buffer[lag_tracker->read_heads[head]].time;
4311 32612 : lag_tracker->last_read[head] =
4312 32612 : lag_tracker->buffer[lag_tracker->read_heads[head]];
4313 32612 : lag_tracker->read_heads[head] =
4314 32612 : (lag_tracker->read_heads[head] + 1) % LAG_TRACKER_BUFFER_SIZE;
4315 : }
4316 :
4317 : /*
4318 : * If the lag tracker is empty, that means the standby has processed
4319 : * everything we've ever sent so we should now clear 'last_read'. If we
4320 : * didn't do that, we'd risk using a stale and irrelevant sample for
4321 : * interpolation at the beginning of the next burst of WAL after a period
4322 : * of idleness.
4323 : */
4324 567702 : if (lag_tracker->read_heads[head] == lag_tracker->write_head)
4325 298816 : lag_tracker->last_read[head].time = 0;
4326 :
4327 567702 : if (time > now)
4328 : {
4329 : /* If the clock somehow went backwards, treat as not found. */
4330 0 : return -1;
4331 : }
4332 567702 : else if (time == 0)
4333 : {
4334 : /*
4335 : * We didn't cross a time. If there is a future sample that we
4336 : * haven't reached yet, and we've already reached at least one sample,
4337 : * let's interpolate the local flushed time. This is mainly useful
4338 : * for reporting a completely stuck apply position as having
4339 : * increasing lag, since otherwise we'd have to wait for it to
4340 : * eventually start moving again and cross one of our samples before
4341 : * we can show the lag increasing.
4342 : */
4343 544824 : if (lag_tracker->read_heads[head] == lag_tracker->write_head)
4344 : {
4345 : /* There are no future samples, so we can't interpolate. */
4346 281378 : return -1;
4347 : }
4348 263446 : else if (lag_tracker->last_read[head].time != 0)
4349 : {
4350 : /* We can interpolate between last_read and the next sample. */
4351 : double fraction;
4352 223866 : WalTimeSample prev = lag_tracker->last_read[head];
4353 223866 : WalTimeSample next = lag_tracker->buffer[lag_tracker->read_heads[head]];
4354 :
4355 223866 : if (lsn < prev.lsn)
4356 : {
4357 : /*
4358 : * Reported LSNs shouldn't normally go backwards, but it's
4359 : * possible when there is a timeline change. Treat as not
4360 : * found.
4361 : */
4362 0 : return -1;
4363 : }
4364 :
4365 : Assert(prev.lsn < next.lsn);
4366 :
4367 223866 : if (prev.time > next.time)
4368 : {
4369 : /* If the clock somehow went backwards, treat as not found. */
4370 0 : return -1;
4371 : }
4372 :
4373 : /* See how far we are between the previous and next samples. */
4374 223866 : fraction =
4375 223866 : (double) (lsn - prev.lsn) / (double) (next.lsn - prev.lsn);
4376 :
4377 : /* Scale the local flush time proportionally. */
4378 223866 : time = (TimestampTz)
4379 223866 : ((double) prev.time + (next.time - prev.time) * fraction);
4380 : }
4381 : else
4382 : {
4383 : /*
4384 : * We have only a future sample, implying that we were entirely
4385 : * caught up but and now there is a new burst of WAL and the
4386 : * standby hasn't processed the first sample yet. Until the
4387 : * standby reaches the future sample the best we can do is report
4388 : * the hypothetical lag if that sample were to be replayed now.
4389 : */
4390 39580 : time = lag_tracker->buffer[lag_tracker->read_heads[head]].time;
4391 : }
4392 : }
4393 :
4394 : /* Return the elapsed time since local flush time in microseconds. */
4395 : Assert(time != 0);
4396 286324 : return now - time;
4397 : }
|
