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/xact.h"
55 : #include "access/xlog_internal.h"
56 : #include "access/xlogreader.h"
57 : #include "access/xlogrecovery.h"
58 : #include "access/xlogutils.h"
59 : #include "backup/basebackup.h"
60 : #include "backup/basebackup_incremental.h"
61 : #include "catalog/pg_authid.h"
62 : #include "catalog/pg_type.h"
63 : #include "commands/defrem.h"
64 : #include "funcapi.h"
65 : #include "libpq/libpq.h"
66 : #include "libpq/pqformat.h"
67 : #include "libpq/protocol.h"
68 : #include "miscadmin.h"
69 : #include "nodes/replnodes.h"
70 : #include "pgstat.h"
71 : #include "postmaster/interrupt.h"
72 : #include "replication/decode.h"
73 : #include "replication/logical.h"
74 : #include "replication/slotsync.h"
75 : #include "replication/slot.h"
76 : #include "replication/snapbuild.h"
77 : #include "replication/syncrep.h"
78 : #include "replication/walreceiver.h"
79 : #include "replication/walsender.h"
80 : #include "replication/walsender_private.h"
81 : #include "storage/condition_variable.h"
82 : #include "storage/aio_subsys.h"
83 : #include "storage/fd.h"
84 : #include "storage/ipc.h"
85 : #include "storage/pmsignal.h"
86 : #include "storage/proc.h"
87 : #include "storage/procarray.h"
88 : #include "tcop/dest.h"
89 : #include "tcop/tcopprot.h"
90 : #include "utils/acl.h"
91 : #include "utils/builtins.h"
92 : #include "utils/guc.h"
93 : #include "utils/lsyscache.h"
94 : #include "utils/memutils.h"
95 : #include "utils/pg_lsn.h"
96 : #include "utils/pgstat_internal.h"
97 : #include "utils/ps_status.h"
98 : #include "utils/timeout.h"
99 : #include "utils/timestamp.h"
100 :
101 : /* Minimum interval used by walsender for stats flushes, in ms */
102 : #define WALSENDER_STATS_FLUSH_INTERVAL 1000
103 :
104 : /*
105 : * Maximum data payload in a WAL data message. Must be >= XLOG_BLCKSZ.
106 : *
107 : * We don't have a good idea of what a good value would be; there's some
108 : * overhead per message in both walsender and walreceiver, but on the other
109 : * hand sending large batches makes walsender less responsive to signals
110 : * because signals are checked only between messages. 128kB (with
111 : * default 8k blocks) seems like a reasonable guess for now.
112 : */
113 : #define MAX_SEND_SIZE (XLOG_BLCKSZ * 16)
114 :
115 : /* Array of WalSnds in shared memory */
116 : WalSndCtlData *WalSndCtl = NULL;
117 :
118 : /* My slot in the shared memory array */
119 : WalSnd *MyWalSnd = NULL;
120 :
121 : /* Global state */
122 : bool am_walsender = false; /* Am I a walsender process? */
123 : bool am_cascading_walsender = false; /* Am I cascading WAL to another
124 : * standby? */
125 : bool am_db_walsender = false; /* Connected to a database? */
126 :
127 : /* GUC variables */
128 : int max_wal_senders = 10; /* the maximum number of concurrent
129 : * walsenders */
130 : int wal_sender_timeout = 60 * 1000; /* maximum time to send one WAL
131 : * data message */
132 : bool log_replication_commands = false;
133 :
134 : /*
135 : * State for WalSndWakeupRequest
136 : */
137 : bool wake_wal_senders = false;
138 :
139 : /*
140 : * xlogreader used for replication. Note that a WAL sender doing physical
141 : * replication does not need xlogreader to read WAL, but it needs one to
142 : * keep a state of its work.
143 : */
144 : static XLogReaderState *xlogreader = NULL;
145 :
146 : /*
147 : * If the UPLOAD_MANIFEST command is used to provide a backup manifest in
148 : * preparation for an incremental backup, uploaded_manifest will be point
149 : * to an object containing information about its contexts, and
150 : * uploaded_manifest_mcxt will point to the memory context that contains
151 : * that object and all of its subordinate data. Otherwise, both values will
152 : * be NULL.
153 : */
154 : static IncrementalBackupInfo *uploaded_manifest = NULL;
155 : static MemoryContext uploaded_manifest_mcxt = NULL;
156 :
157 : /*
158 : * These variables keep track of the state of the timeline we're currently
159 : * sending. sendTimeLine identifies the timeline. If sendTimeLineIsHistoric,
160 : * the timeline is not the latest timeline on this server, and the server's
161 : * history forked off from that timeline at sendTimeLineValidUpto.
162 : */
163 : static TimeLineID sendTimeLine = 0;
164 : static TimeLineID sendTimeLineNextTLI = 0;
165 : static bool sendTimeLineIsHistoric = false;
166 : static XLogRecPtr sendTimeLineValidUpto = InvalidXLogRecPtr;
167 :
168 : /*
169 : * How far have we sent WAL already? This is also advertised in
170 : * MyWalSnd->sentPtr. (Actually, this is the next WAL location to send.)
171 : */
172 : static XLogRecPtr sentPtr = InvalidXLogRecPtr;
173 :
174 : /* Buffers for constructing outgoing messages and processing reply messages. */
175 : static StringInfoData output_message;
176 : static StringInfoData reply_message;
177 : static StringInfoData tmpbuf;
178 :
179 : /* Timestamp of last ProcessRepliesIfAny(). */
180 : static TimestampTz last_processing = 0;
181 :
182 : /*
183 : * Timestamp of last ProcessRepliesIfAny() that saw a reply from the
184 : * standby. Set to 0 if wal_sender_timeout doesn't need to be active.
185 : */
186 : static TimestampTz last_reply_timestamp = 0;
187 :
188 : /* Have we sent a heartbeat message asking for reply, since last reply? */
189 : static bool waiting_for_ping_response = false;
190 :
191 : /*
192 : * While streaming WAL in Copy mode, streamingDoneSending is set to true
193 : * after we have sent CopyDone. We should not send any more CopyData messages
194 : * after that. streamingDoneReceiving is set to true when we receive CopyDone
195 : * from the other end. When both become true, it's time to exit Copy mode.
196 : */
197 : static bool streamingDoneSending;
198 : static bool streamingDoneReceiving;
199 :
200 : /* Are we there yet? */
201 : static bool WalSndCaughtUp = false;
202 :
203 : /* Flags set by signal handlers for later service in main loop */
204 : static volatile sig_atomic_t got_SIGUSR2 = false;
205 : static volatile sig_atomic_t got_STOPPING = false;
206 :
207 : /*
208 : * This is set while we are streaming. When not set
209 : * PROCSIG_WALSND_INIT_STOPPING signal will be handled like SIGTERM. When set,
210 : * the main loop is responsible for checking got_STOPPING and terminating when
211 : * it's set (after streaming any remaining WAL).
212 : */
213 : static volatile sig_atomic_t replication_active = false;
214 :
215 : static LogicalDecodingContext *logical_decoding_ctx = NULL;
216 :
217 : /* A sample associating a WAL location with the time it was written. */
218 : typedef struct
219 : {
220 : XLogRecPtr lsn;
221 : TimestampTz time;
222 : } WalTimeSample;
223 :
224 : /* The size of our buffer of time samples. */
225 : #define LAG_TRACKER_BUFFER_SIZE 8192
226 :
227 : /* A mechanism for tracking replication lag. */
228 : typedef struct
229 : {
230 : XLogRecPtr last_lsn;
231 : WalTimeSample buffer[LAG_TRACKER_BUFFER_SIZE];
232 : int write_head;
233 : int read_heads[NUM_SYNC_REP_WAIT_MODE];
234 : WalTimeSample last_read[NUM_SYNC_REP_WAIT_MODE];
235 : } LagTracker;
236 :
237 : static LagTracker *lag_tracker;
238 :
239 : /* Signal handlers */
240 : static void WalSndLastCycleHandler(SIGNAL_ARGS);
241 :
242 : /* Prototypes for private functions */
243 : typedef void (*WalSndSendDataCallback) (void);
244 : static void WalSndLoop(WalSndSendDataCallback send_data);
245 : static void InitWalSenderSlot(void);
246 : static void WalSndKill(int code, Datum arg);
247 : pg_noreturn static void WalSndShutdown(void);
248 : static void XLogSendPhysical(void);
249 : static void XLogSendLogical(void);
250 : static void WalSndDone(WalSndSendDataCallback send_data);
251 : static void IdentifySystem(void);
252 : static void UploadManifest(void);
253 : static bool HandleUploadManifestPacket(StringInfo buf, off_t *offset,
254 : IncrementalBackupInfo *ib);
255 : static void ReadReplicationSlot(ReadReplicationSlotCmd *cmd);
256 : static void CreateReplicationSlot(CreateReplicationSlotCmd *cmd);
257 : static void DropReplicationSlot(DropReplicationSlotCmd *cmd);
258 : static void StartReplication(StartReplicationCmd *cmd);
259 : static void StartLogicalReplication(StartReplicationCmd *cmd);
260 : static void ProcessStandbyMessage(void);
261 : static void ProcessStandbyReplyMessage(void);
262 : static void ProcessStandbyHSFeedbackMessage(void);
263 : static void ProcessStandbyPSRequestMessage(void);
264 : static void ProcessRepliesIfAny(void);
265 : static void ProcessPendingWrites(void);
266 : static void WalSndKeepalive(bool requestReply, XLogRecPtr writePtr);
267 : static void WalSndKeepaliveIfNecessary(void);
268 : static void WalSndCheckTimeOut(void);
269 : static long WalSndComputeSleeptime(TimestampTz now);
270 : static void WalSndWait(uint32 socket_events, long timeout, uint32 wait_event);
271 : static void WalSndPrepareWrite(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid, bool last_write);
272 : static void WalSndWriteData(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid, bool last_write);
273 : static void WalSndUpdateProgress(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid,
274 : bool skipped_xact);
275 : static XLogRecPtr WalSndWaitForWal(XLogRecPtr loc);
276 : static void LagTrackerWrite(XLogRecPtr lsn, TimestampTz local_flush_time);
277 : static TimeOffset LagTrackerRead(int head, XLogRecPtr lsn, TimestampTz now);
278 : static bool TransactionIdInRecentPast(TransactionId xid, uint32 epoch);
279 :
280 : static void WalSndSegmentOpen(XLogReaderState *state, XLogSegNo nextSegNo,
281 : TimeLineID *tli_p);
282 :
283 :
284 : /* Initialize walsender process before entering the main command loop */
285 : void
286 2282 : InitWalSender(void)
287 : {
288 2282 : am_cascading_walsender = RecoveryInProgress();
289 :
290 : /* Create a per-walsender data structure in shared memory */
291 2282 : InitWalSenderSlot();
292 :
293 : /* need resource owner for e.g. basebackups */
294 2282 : CreateAuxProcessResourceOwner();
295 :
296 : /*
297 : * Let postmaster know that we're a WAL sender. Once we've declared us as
298 : * a WAL sender process, postmaster will let us outlive the bgwriter and
299 : * kill us last in the shutdown sequence, so we get a chance to stream all
300 : * remaining WAL at shutdown, including the shutdown checkpoint. Note that
301 : * there's no going back, and we mustn't write any WAL records after this.
302 : */
303 2282 : MarkPostmasterChildWalSender();
304 2282 : SendPostmasterSignal(PMSIGNAL_ADVANCE_STATE_MACHINE);
305 :
306 : /*
307 : * If the client didn't specify a database to connect to, show in PGPROC
308 : * that our advertised xmin should affect vacuum horizons in all
309 : * databases. This allows physical replication clients to send hot
310 : * standby feedback that will delay vacuum cleanup in all databases.
311 : */
312 2282 : if (MyDatabaseId == InvalidOid)
313 : {
314 : Assert(MyProc->xmin == InvalidTransactionId);
315 906 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
316 906 : MyProc->statusFlags |= PROC_AFFECTS_ALL_HORIZONS;
317 906 : ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
318 906 : LWLockRelease(ProcArrayLock);
319 : }
320 :
321 : /* Initialize empty timestamp buffer for lag tracking. */
322 2282 : lag_tracker = MemoryContextAllocZero(TopMemoryContext, sizeof(LagTracker));
323 2282 : }
324 :
325 : /*
326 : * Clean up after an error.
327 : *
328 : * WAL sender processes don't use transactions like regular backends do.
329 : * This function does any cleanup required after an error in a WAL sender
330 : * process, similar to what transaction abort does in a regular backend.
331 : */
332 : void
333 96 : WalSndErrorCleanup(void)
334 : {
335 96 : LWLockReleaseAll();
336 96 : ConditionVariableCancelSleep();
337 96 : pgstat_report_wait_end();
338 96 : pgaio_error_cleanup();
339 :
340 96 : if (xlogreader != NULL && xlogreader->seg.ws_file >= 0)
341 12 : wal_segment_close(xlogreader);
342 :
343 96 : if (MyReplicationSlot != NULL)
344 30 : ReplicationSlotRelease();
345 :
346 96 : ReplicationSlotCleanup(false);
347 :
348 96 : replication_active = false;
349 :
350 : /*
351 : * If there is a transaction in progress, it will clean up our
352 : * ResourceOwner, but if a replication command set up a resource owner
353 : * without a transaction, we've got to clean that up now.
354 : */
355 96 : if (!IsTransactionOrTransactionBlock())
356 94 : ReleaseAuxProcessResources(false);
357 :
358 96 : if (got_STOPPING || got_SIGUSR2)
359 0 : proc_exit(0);
360 :
361 : /* Revert back to startup state */
362 96 : WalSndSetState(WALSNDSTATE_STARTUP);
363 96 : }
364 :
365 : /*
366 : * Handle a client's connection abort in an orderly manner.
367 : */
368 : static void
369 58 : WalSndShutdown(void)
370 : {
371 : /*
372 : * Reset whereToSendOutput to prevent ereport from attempting to send any
373 : * more messages to the standby.
374 : */
375 58 : if (whereToSendOutput == DestRemote)
376 58 : whereToSendOutput = DestNone;
377 :
378 58 : proc_exit(0);
379 : abort(); /* keep the compiler quiet */
380 : }
381 :
382 : /*
383 : * Handle the IDENTIFY_SYSTEM command.
384 : */
385 : static void
386 1412 : IdentifySystem(void)
387 : {
388 : char sysid[32];
389 : char xloc[MAXFNAMELEN];
390 : XLogRecPtr logptr;
391 1412 : char *dbname = NULL;
392 : DestReceiver *dest;
393 : TupOutputState *tstate;
394 : TupleDesc tupdesc;
395 : Datum values[4];
396 1412 : bool nulls[4] = {0};
397 : TimeLineID currTLI;
398 :
399 : /*
400 : * Reply with a result set with one row, four columns. First col is system
401 : * ID, second is timeline ID, third is current xlog location and the
402 : * fourth contains the database name if we are connected to one.
403 : */
404 :
405 1412 : snprintf(sysid, sizeof(sysid), UINT64_FORMAT,
406 : GetSystemIdentifier());
407 :
408 1412 : am_cascading_walsender = RecoveryInProgress();
409 1412 : if (am_cascading_walsender)
410 116 : logptr = GetStandbyFlushRecPtr(&currTLI);
411 : else
412 1296 : logptr = GetFlushRecPtr(&currTLI);
413 :
414 1412 : snprintf(xloc, sizeof(xloc), "%X/%08X", LSN_FORMAT_ARGS(logptr));
415 :
416 1412 : if (MyDatabaseId != InvalidOid)
417 : {
418 506 : MemoryContext cur = CurrentMemoryContext;
419 :
420 : /* syscache access needs a transaction env. */
421 506 : StartTransactionCommand();
422 506 : dbname = get_database_name(MyDatabaseId);
423 : /* copy dbname out of TX context */
424 506 : dbname = MemoryContextStrdup(cur, dbname);
425 506 : CommitTransactionCommand();
426 : }
427 :
428 1412 : dest = CreateDestReceiver(DestRemoteSimple);
429 :
430 : /* need a tuple descriptor representing four columns */
431 1412 : tupdesc = CreateTemplateTupleDesc(4);
432 1412 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "systemid",
433 : TEXTOID, -1, 0);
434 1412 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "timeline",
435 : INT8OID, -1, 0);
436 1412 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 3, "xlogpos",
437 : TEXTOID, -1, 0);
438 1412 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 4, "dbname",
439 : TEXTOID, -1, 0);
440 :
441 : /* prepare for projection of tuples */
442 1412 : tstate = begin_tup_output_tupdesc(dest, tupdesc, &TTSOpsVirtual);
443 :
444 : /* column 1: system identifier */
445 1412 : values[0] = CStringGetTextDatum(sysid);
446 :
447 : /* column 2: timeline */
448 1412 : values[1] = Int64GetDatum(currTLI);
449 :
450 : /* column 3: wal location */
451 1412 : values[2] = CStringGetTextDatum(xloc);
452 :
453 : /* column 4: database name, or NULL if none */
454 1412 : if (dbname)
455 506 : values[3] = CStringGetTextDatum(dbname);
456 : else
457 906 : nulls[3] = true;
458 :
459 : /* send it to dest */
460 1412 : do_tup_output(tstate, values, nulls);
461 :
462 1412 : end_tup_output(tstate);
463 1412 : }
464 :
465 : /* Handle READ_REPLICATION_SLOT command */
466 : static void
467 12 : ReadReplicationSlot(ReadReplicationSlotCmd *cmd)
468 : {
469 : #define READ_REPLICATION_SLOT_COLS 3
470 : ReplicationSlot *slot;
471 : DestReceiver *dest;
472 : TupOutputState *tstate;
473 : TupleDesc tupdesc;
474 12 : Datum values[READ_REPLICATION_SLOT_COLS] = {0};
475 : bool nulls[READ_REPLICATION_SLOT_COLS];
476 :
477 12 : tupdesc = CreateTemplateTupleDesc(READ_REPLICATION_SLOT_COLS);
478 12 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "slot_type",
479 : TEXTOID, -1, 0);
480 12 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "restart_lsn",
481 : TEXTOID, -1, 0);
482 : /* TimeLineID is unsigned, so int4 is not wide enough. */
483 12 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 3, "restart_tli",
484 : INT8OID, -1, 0);
485 :
486 12 : memset(nulls, true, READ_REPLICATION_SLOT_COLS * sizeof(bool));
487 :
488 12 : LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
489 12 : slot = SearchNamedReplicationSlot(cmd->slotname, false);
490 12 : if (slot == NULL || !slot->in_use)
491 : {
492 4 : LWLockRelease(ReplicationSlotControlLock);
493 : }
494 : else
495 : {
496 : ReplicationSlot slot_contents;
497 8 : int i = 0;
498 :
499 : /* Copy slot contents while holding spinlock */
500 8 : SpinLockAcquire(&slot->mutex);
501 8 : slot_contents = *slot;
502 8 : SpinLockRelease(&slot->mutex);
503 8 : LWLockRelease(ReplicationSlotControlLock);
504 :
505 8 : if (OidIsValid(slot_contents.data.database))
506 2 : ereport(ERROR,
507 : errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
508 : errmsg("cannot use %s with a logical replication slot",
509 : "READ_REPLICATION_SLOT"));
510 :
511 : /* slot type */
512 6 : values[i] = CStringGetTextDatum("physical");
513 6 : nulls[i] = false;
514 6 : i++;
515 :
516 : /* start LSN */
517 6 : if (!XLogRecPtrIsInvalid(slot_contents.data.restart_lsn))
518 : {
519 : char xloc[64];
520 :
521 6 : snprintf(xloc, sizeof(xloc), "%X/%08X",
522 6 : LSN_FORMAT_ARGS(slot_contents.data.restart_lsn));
523 6 : values[i] = CStringGetTextDatum(xloc);
524 6 : nulls[i] = false;
525 : }
526 6 : i++;
527 :
528 : /* timeline this WAL was produced on */
529 6 : if (!XLogRecPtrIsInvalid(slot_contents.data.restart_lsn))
530 : {
531 : TimeLineID slots_position_timeline;
532 : TimeLineID current_timeline;
533 6 : List *timeline_history = NIL;
534 :
535 : /*
536 : * While in recovery, use as timeline the currently-replaying one
537 : * to get the LSN position's history.
538 : */
539 6 : if (RecoveryInProgress())
540 0 : (void) GetXLogReplayRecPtr(¤t_timeline);
541 : else
542 6 : current_timeline = GetWALInsertionTimeLine();
543 :
544 6 : timeline_history = readTimeLineHistory(current_timeline);
545 6 : slots_position_timeline = tliOfPointInHistory(slot_contents.data.restart_lsn,
546 : timeline_history);
547 6 : values[i] = Int64GetDatum((int64) slots_position_timeline);
548 6 : nulls[i] = false;
549 : }
550 6 : i++;
551 :
552 : Assert(i == READ_REPLICATION_SLOT_COLS);
553 : }
554 :
555 10 : dest = CreateDestReceiver(DestRemoteSimple);
556 10 : tstate = begin_tup_output_tupdesc(dest, tupdesc, &TTSOpsVirtual);
557 10 : do_tup_output(tstate, values, nulls);
558 10 : end_tup_output(tstate);
559 10 : }
560 :
561 :
562 : /*
563 : * Handle TIMELINE_HISTORY command.
564 : */
565 : static void
566 28 : SendTimeLineHistory(TimeLineHistoryCmd *cmd)
567 : {
568 : DestReceiver *dest;
569 : TupleDesc tupdesc;
570 : StringInfoData buf;
571 : char histfname[MAXFNAMELEN];
572 : char path[MAXPGPATH];
573 : int fd;
574 : off_t histfilelen;
575 : off_t bytesleft;
576 : Size len;
577 :
578 28 : dest = CreateDestReceiver(DestRemoteSimple);
579 :
580 : /*
581 : * Reply with a result set with one row, and two columns. The first col is
582 : * the name of the history file, 2nd is the contents.
583 : */
584 28 : tupdesc = CreateTemplateTupleDesc(2);
585 28 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "filename", TEXTOID, -1, 0);
586 28 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "content", TEXTOID, -1, 0);
587 :
588 28 : TLHistoryFileName(histfname, cmd->timeline);
589 28 : TLHistoryFilePath(path, cmd->timeline);
590 :
591 : /* Send a RowDescription message */
592 28 : dest->rStartup(dest, CMD_SELECT, tupdesc);
593 :
594 : /* Send a DataRow message */
595 28 : pq_beginmessage(&buf, PqMsg_DataRow);
596 28 : pq_sendint16(&buf, 2); /* # of columns */
597 28 : len = strlen(histfname);
598 28 : pq_sendint32(&buf, len); /* col1 len */
599 28 : pq_sendbytes(&buf, histfname, len);
600 :
601 28 : fd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
602 28 : if (fd < 0)
603 0 : ereport(ERROR,
604 : (errcode_for_file_access(),
605 : errmsg("could not open file \"%s\": %m", path)));
606 :
607 : /* Determine file length and send it to client */
608 28 : histfilelen = lseek(fd, 0, SEEK_END);
609 28 : if (histfilelen < 0)
610 0 : ereport(ERROR,
611 : (errcode_for_file_access(),
612 : errmsg("could not seek to end of file \"%s\": %m", path)));
613 28 : if (lseek(fd, 0, SEEK_SET) != 0)
614 0 : ereport(ERROR,
615 : (errcode_for_file_access(),
616 : errmsg("could not seek to beginning of file \"%s\": %m", path)));
617 :
618 28 : pq_sendint32(&buf, histfilelen); /* col2 len */
619 :
620 28 : bytesleft = histfilelen;
621 56 : while (bytesleft > 0)
622 : {
623 : PGAlignedBlock rbuf;
624 : int nread;
625 :
626 28 : pgstat_report_wait_start(WAIT_EVENT_WALSENDER_TIMELINE_HISTORY_READ);
627 28 : nread = read(fd, rbuf.data, sizeof(rbuf));
628 28 : pgstat_report_wait_end();
629 28 : if (nread < 0)
630 0 : ereport(ERROR,
631 : (errcode_for_file_access(),
632 : errmsg("could not read file \"%s\": %m",
633 : path)));
634 28 : else if (nread == 0)
635 0 : ereport(ERROR,
636 : (errcode(ERRCODE_DATA_CORRUPTED),
637 : errmsg("could not read file \"%s\": read %d of %zu",
638 : path, nread, (Size) bytesleft)));
639 :
640 28 : pq_sendbytes(&buf, rbuf.data, nread);
641 28 : bytesleft -= nread;
642 : }
643 :
644 28 : if (CloseTransientFile(fd) != 0)
645 0 : ereport(ERROR,
646 : (errcode_for_file_access(),
647 : errmsg("could not close file \"%s\": %m", path)));
648 :
649 28 : pq_endmessage(&buf);
650 28 : }
651 :
652 : /*
653 : * Handle UPLOAD_MANIFEST command.
654 : */
655 : static void
656 22 : UploadManifest(void)
657 : {
658 : MemoryContext mcxt;
659 : IncrementalBackupInfo *ib;
660 22 : off_t offset = 0;
661 : StringInfoData buf;
662 :
663 : /*
664 : * parsing the manifest will use the cryptohash stuff, which requires a
665 : * resource owner
666 : */
667 : Assert(AuxProcessResourceOwner != NULL);
668 : Assert(CurrentResourceOwner == AuxProcessResourceOwner ||
669 : CurrentResourceOwner == NULL);
670 22 : CurrentResourceOwner = AuxProcessResourceOwner;
671 :
672 : /* Prepare to read manifest data into a temporary context. */
673 22 : mcxt = AllocSetContextCreate(CurrentMemoryContext,
674 : "incremental backup information",
675 : ALLOCSET_DEFAULT_SIZES);
676 22 : ib = CreateIncrementalBackupInfo(mcxt);
677 :
678 : /* Send a CopyInResponse message */
679 22 : pq_beginmessage(&buf, PqMsg_CopyInResponse);
680 22 : pq_sendbyte(&buf, 0);
681 22 : pq_sendint16(&buf, 0);
682 22 : pq_endmessage_reuse(&buf);
683 22 : pq_flush();
684 :
685 : /* Receive packets from client until done. */
686 86 : while (HandleUploadManifestPacket(&buf, &offset, ib))
687 : ;
688 :
689 : /* Finish up manifest processing. */
690 20 : FinalizeIncrementalManifest(ib);
691 :
692 : /*
693 : * Discard any old manifest information and arrange to preserve the new
694 : * information we just got.
695 : *
696 : * We assume that MemoryContextDelete and MemoryContextSetParent won't
697 : * fail, and thus we shouldn't end up bailing out of here in such a way as
698 : * to leave dangling pointers.
699 : */
700 20 : if (uploaded_manifest_mcxt != NULL)
701 0 : MemoryContextDelete(uploaded_manifest_mcxt);
702 20 : MemoryContextSetParent(mcxt, CacheMemoryContext);
703 20 : uploaded_manifest = ib;
704 20 : uploaded_manifest_mcxt = mcxt;
705 :
706 : /* clean up the resource owner we created */
707 20 : ReleaseAuxProcessResources(true);
708 20 : }
709 :
710 : /*
711 : * Process one packet received during the handling of an UPLOAD_MANIFEST
712 : * operation.
713 : *
714 : * 'buf' is scratch space. This function expects it to be initialized, doesn't
715 : * care what the current contents are, and may override them with completely
716 : * new contents.
717 : *
718 : * The return value is true if the caller should continue processing
719 : * additional packets and false if the UPLOAD_MANIFEST operation is complete.
720 : */
721 : static bool
722 86 : HandleUploadManifestPacket(StringInfo buf, off_t *offset,
723 : IncrementalBackupInfo *ib)
724 : {
725 : int mtype;
726 : int maxmsglen;
727 :
728 86 : HOLD_CANCEL_INTERRUPTS();
729 :
730 86 : pq_startmsgread();
731 86 : mtype = pq_getbyte();
732 86 : if (mtype == EOF)
733 0 : ereport(ERROR,
734 : (errcode(ERRCODE_CONNECTION_FAILURE),
735 : errmsg("unexpected EOF on client connection with an open transaction")));
736 :
737 86 : switch (mtype)
738 : {
739 66 : case PqMsg_CopyData:
740 66 : maxmsglen = PQ_LARGE_MESSAGE_LIMIT;
741 66 : break;
742 20 : case PqMsg_CopyDone:
743 : case PqMsg_CopyFail:
744 : case PqMsg_Flush:
745 : case PqMsg_Sync:
746 20 : maxmsglen = PQ_SMALL_MESSAGE_LIMIT;
747 20 : break;
748 0 : default:
749 0 : ereport(ERROR,
750 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
751 : errmsg("unexpected message type 0x%02X during COPY from stdin",
752 : mtype)));
753 : maxmsglen = 0; /* keep compiler quiet */
754 : break;
755 : }
756 :
757 : /* Now collect the message body */
758 86 : if (pq_getmessage(buf, maxmsglen))
759 0 : ereport(ERROR,
760 : (errcode(ERRCODE_CONNECTION_FAILURE),
761 : errmsg("unexpected EOF on client connection with an open transaction")));
762 86 : RESUME_CANCEL_INTERRUPTS();
763 :
764 : /* Process the message */
765 86 : switch (mtype)
766 : {
767 66 : case PqMsg_CopyData:
768 66 : AppendIncrementalManifestData(ib, buf->data, buf->len);
769 64 : return true;
770 :
771 20 : case PqMsg_CopyDone:
772 20 : return false;
773 :
774 0 : case PqMsg_Sync:
775 : case PqMsg_Flush:
776 : /* Ignore these while in CopyOut mode as we do elsewhere. */
777 0 : return true;
778 :
779 0 : case PqMsg_CopyFail:
780 0 : ereport(ERROR,
781 : (errcode(ERRCODE_QUERY_CANCELED),
782 : errmsg("COPY from stdin failed: %s",
783 : pq_getmsgstring(buf))));
784 : }
785 :
786 : /* Not reached. */
787 : Assert(false);
788 0 : return false;
789 : }
790 :
791 : /*
792 : * Handle START_REPLICATION command.
793 : *
794 : * At the moment, this never returns, but an ereport(ERROR) will take us back
795 : * to the main loop.
796 : */
797 : static void
798 520 : StartReplication(StartReplicationCmd *cmd)
799 : {
800 : StringInfoData buf;
801 : XLogRecPtr FlushPtr;
802 : TimeLineID FlushTLI;
803 :
804 : /* create xlogreader for physical replication */
805 520 : xlogreader =
806 520 : XLogReaderAllocate(wal_segment_size, NULL,
807 520 : XL_ROUTINE(.segment_open = WalSndSegmentOpen,
808 : .segment_close = wal_segment_close),
809 : NULL);
810 :
811 520 : if (!xlogreader)
812 0 : ereport(ERROR,
813 : (errcode(ERRCODE_OUT_OF_MEMORY),
814 : errmsg("out of memory"),
815 : errdetail("Failed while allocating a WAL reading processor.")));
816 :
817 : /*
818 : * We assume here that we're logging enough information in the WAL for
819 : * log-shipping, since this is checked in PostmasterMain().
820 : *
821 : * NOTE: wal_level can only change at shutdown, so in most cases it is
822 : * difficult for there to be WAL data that we can still see that was
823 : * written at wal_level='minimal'.
824 : */
825 :
826 520 : if (cmd->slotname)
827 : {
828 358 : ReplicationSlotAcquire(cmd->slotname, true, true);
829 354 : if (SlotIsLogical(MyReplicationSlot))
830 0 : ereport(ERROR,
831 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
832 : errmsg("cannot use a logical replication slot for physical replication")));
833 :
834 : /*
835 : * We don't need to verify the slot's restart_lsn here; instead we
836 : * rely on the caller requesting the starting point to use. If the
837 : * WAL segment doesn't exist, we'll fail later.
838 : */
839 : }
840 :
841 : /*
842 : * Select the timeline. If it was given explicitly by the client, use
843 : * that. Otherwise use the timeline of the last replayed record.
844 : */
845 516 : am_cascading_walsender = RecoveryInProgress();
846 516 : if (am_cascading_walsender)
847 26 : FlushPtr = GetStandbyFlushRecPtr(&FlushTLI);
848 : else
849 490 : FlushPtr = GetFlushRecPtr(&FlushTLI);
850 :
851 516 : if (cmd->timeline != 0)
852 : {
853 : XLogRecPtr switchpoint;
854 :
855 514 : sendTimeLine = cmd->timeline;
856 514 : if (sendTimeLine == FlushTLI)
857 : {
858 490 : sendTimeLineIsHistoric = false;
859 490 : sendTimeLineValidUpto = InvalidXLogRecPtr;
860 : }
861 : else
862 : {
863 : List *timeLineHistory;
864 :
865 24 : sendTimeLineIsHistoric = true;
866 :
867 : /*
868 : * Check that the timeline the client requested exists, and the
869 : * requested start location is on that timeline.
870 : */
871 24 : timeLineHistory = readTimeLineHistory(FlushTLI);
872 24 : switchpoint = tliSwitchPoint(cmd->timeline, timeLineHistory,
873 : &sendTimeLineNextTLI);
874 24 : list_free_deep(timeLineHistory);
875 :
876 : /*
877 : * Found the requested timeline in the history. Check that
878 : * requested startpoint is on that timeline in our history.
879 : *
880 : * This is quite loose on purpose. We only check that we didn't
881 : * fork off the requested timeline before the switchpoint. We
882 : * don't check that we switched *to* it before the requested
883 : * starting point. This is because the client can legitimately
884 : * request to start replication from the beginning of the WAL
885 : * segment that contains switchpoint, but on the new timeline, so
886 : * that it doesn't end up with a partial segment. If you ask for
887 : * too old a starting point, you'll get an error later when we
888 : * fail to find the requested WAL segment in pg_wal.
889 : *
890 : * XXX: we could be more strict here and only allow a startpoint
891 : * that's older than the switchpoint, if it's still in the same
892 : * WAL segment.
893 : */
894 24 : if (!XLogRecPtrIsInvalid(switchpoint) &&
895 24 : switchpoint < cmd->startpoint)
896 : {
897 0 : ereport(ERROR,
898 : errmsg("requested starting point %X/%08X on timeline %u is not in this server's history",
899 : LSN_FORMAT_ARGS(cmd->startpoint),
900 : cmd->timeline),
901 : errdetail("This server's history forked from timeline %u at %X/%08X.",
902 : cmd->timeline,
903 : LSN_FORMAT_ARGS(switchpoint)));
904 : }
905 24 : sendTimeLineValidUpto = switchpoint;
906 : }
907 : }
908 : else
909 : {
910 2 : sendTimeLine = FlushTLI;
911 2 : sendTimeLineValidUpto = InvalidXLogRecPtr;
912 2 : sendTimeLineIsHistoric = false;
913 : }
914 :
915 516 : streamingDoneSending = streamingDoneReceiving = false;
916 :
917 : /* If there is nothing to stream, don't even enter COPY mode */
918 516 : if (!sendTimeLineIsHistoric || cmd->startpoint < sendTimeLineValidUpto)
919 : {
920 : /*
921 : * When we first start replication the standby will be behind the
922 : * primary. For some applications, for example synchronous
923 : * replication, it is important to have a clear state for this initial
924 : * catchup mode, so we can trigger actions when we change streaming
925 : * state later. We may stay in this state for a long time, which is
926 : * exactly why we want to be able to monitor whether or not we are
927 : * still here.
928 : */
929 516 : WalSndSetState(WALSNDSTATE_CATCHUP);
930 :
931 : /* Send a CopyBothResponse message, and start streaming */
932 516 : pq_beginmessage(&buf, PqMsg_CopyBothResponse);
933 516 : pq_sendbyte(&buf, 0);
934 516 : pq_sendint16(&buf, 0);
935 516 : pq_endmessage(&buf);
936 516 : pq_flush();
937 :
938 : /*
939 : * Don't allow a request to stream from a future point in WAL that
940 : * hasn't been flushed to disk in this server yet.
941 : */
942 516 : if (FlushPtr < cmd->startpoint)
943 : {
944 0 : ereport(ERROR,
945 : errmsg("requested starting point %X/%08X is ahead of the WAL flush position of this server %X/%08X",
946 : LSN_FORMAT_ARGS(cmd->startpoint),
947 : LSN_FORMAT_ARGS(FlushPtr)));
948 : }
949 :
950 : /* Start streaming from the requested point */
951 516 : sentPtr = cmd->startpoint;
952 :
953 : /* Initialize shared memory status, too */
954 516 : SpinLockAcquire(&MyWalSnd->mutex);
955 516 : MyWalSnd->sentPtr = sentPtr;
956 516 : SpinLockRelease(&MyWalSnd->mutex);
957 :
958 516 : SyncRepInitConfig();
959 :
960 : /* Main loop of walsender */
961 516 : replication_active = true;
962 :
963 516 : WalSndLoop(XLogSendPhysical);
964 :
965 298 : replication_active = false;
966 298 : if (got_STOPPING)
967 0 : proc_exit(0);
968 298 : WalSndSetState(WALSNDSTATE_STARTUP);
969 :
970 : Assert(streamingDoneSending && streamingDoneReceiving);
971 : }
972 :
973 298 : if (cmd->slotname)
974 266 : ReplicationSlotRelease();
975 :
976 : /*
977 : * Copy is finished now. Send a single-row result set indicating the next
978 : * timeline.
979 : */
980 298 : if (sendTimeLineIsHistoric)
981 : {
982 : char startpos_str[8 + 1 + 8 + 1];
983 : DestReceiver *dest;
984 : TupOutputState *tstate;
985 : TupleDesc tupdesc;
986 : Datum values[2];
987 26 : bool nulls[2] = {0};
988 :
989 26 : snprintf(startpos_str, sizeof(startpos_str), "%X/%08X",
990 26 : LSN_FORMAT_ARGS(sendTimeLineValidUpto));
991 :
992 26 : dest = CreateDestReceiver(DestRemoteSimple);
993 :
994 : /*
995 : * Need a tuple descriptor representing two columns. int8 may seem
996 : * like a surprising data type for this, but in theory int4 would not
997 : * be wide enough for this, as TimeLineID is unsigned.
998 : */
999 26 : tupdesc = CreateTemplateTupleDesc(2);
1000 26 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "next_tli",
1001 : INT8OID, -1, 0);
1002 26 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "next_tli_startpos",
1003 : TEXTOID, -1, 0);
1004 :
1005 : /* prepare for projection of tuple */
1006 26 : tstate = begin_tup_output_tupdesc(dest, tupdesc, &TTSOpsVirtual);
1007 :
1008 26 : values[0] = Int64GetDatum((int64) sendTimeLineNextTLI);
1009 26 : values[1] = CStringGetTextDatum(startpos_str);
1010 :
1011 : /* send it to dest */
1012 26 : do_tup_output(tstate, values, nulls);
1013 :
1014 26 : end_tup_output(tstate);
1015 : }
1016 :
1017 : /* Send CommandComplete message */
1018 298 : EndReplicationCommand("START_STREAMING");
1019 298 : }
1020 :
1021 : /*
1022 : * XLogReaderRoutine->page_read callback for logical decoding contexts, as a
1023 : * walsender process.
1024 : *
1025 : * Inside the walsender we can do better than read_local_xlog_page,
1026 : * which has to do a plain sleep/busy loop, because the walsender's latch gets
1027 : * set every time WAL is flushed.
1028 : */
1029 : static int
1030 32392 : logical_read_xlog_page(XLogReaderState *state, XLogRecPtr targetPagePtr, int reqLen,
1031 : XLogRecPtr targetRecPtr, char *cur_page)
1032 : {
1033 : XLogRecPtr flushptr;
1034 : int count;
1035 : WALReadError errinfo;
1036 : XLogSegNo segno;
1037 : TimeLineID currTLI;
1038 :
1039 : /*
1040 : * Make sure we have enough WAL available before retrieving the current
1041 : * timeline.
1042 : */
1043 32392 : flushptr = WalSndWaitForWal(targetPagePtr + reqLen);
1044 :
1045 : /* Fail if not enough (implies we are going to shut down) */
1046 32050 : if (flushptr < targetPagePtr + reqLen)
1047 7154 : return -1;
1048 :
1049 : /*
1050 : * Since logical decoding is also permitted on a standby server, we need
1051 : * to check if the server is in recovery to decide how to get the current
1052 : * timeline ID (so that it also covers the promotion or timeline change
1053 : * cases). We must determine am_cascading_walsender after waiting for the
1054 : * required WAL so that it is correct when the walsender wakes up after a
1055 : * promotion.
1056 : */
1057 24896 : am_cascading_walsender = RecoveryInProgress();
1058 :
1059 24896 : if (am_cascading_walsender)
1060 1612 : GetXLogReplayRecPtr(&currTLI);
1061 : else
1062 23284 : currTLI = GetWALInsertionTimeLine();
1063 :
1064 24896 : XLogReadDetermineTimeline(state, targetPagePtr, reqLen, currTLI);
1065 24896 : sendTimeLineIsHistoric = (state->currTLI != currTLI);
1066 24896 : sendTimeLine = state->currTLI;
1067 24896 : sendTimeLineValidUpto = state->currTLIValidUntil;
1068 24896 : sendTimeLineNextTLI = state->nextTLI;
1069 :
1070 24896 : if (targetPagePtr + XLOG_BLCKSZ <= flushptr)
1071 21212 : count = XLOG_BLCKSZ; /* more than one block available */
1072 : else
1073 3684 : count = flushptr - targetPagePtr; /* part of the page available */
1074 :
1075 : /* now actually read the data, we know it's there */
1076 24896 : if (!WALRead(state,
1077 : cur_page,
1078 : targetPagePtr,
1079 : count,
1080 : currTLI, /* Pass the current TLI because only
1081 : * WalSndSegmentOpen controls whether new TLI
1082 : * is needed. */
1083 : &errinfo))
1084 0 : WALReadRaiseError(&errinfo);
1085 :
1086 : /*
1087 : * After reading into the buffer, check that what we read was valid. We do
1088 : * this after reading, because even though the segment was present when we
1089 : * opened it, it might get recycled or removed while we read it. The
1090 : * read() succeeds in that case, but the data we tried to read might
1091 : * already have been overwritten with new WAL records.
1092 : */
1093 24896 : XLByteToSeg(targetPagePtr, segno, state->segcxt.ws_segsize);
1094 24896 : CheckXLogRemoved(segno, state->seg.ws_tli);
1095 :
1096 24896 : return count;
1097 : }
1098 :
1099 : /*
1100 : * Process extra options given to CREATE_REPLICATION_SLOT.
1101 : */
1102 : static void
1103 934 : parseCreateReplSlotOptions(CreateReplicationSlotCmd *cmd,
1104 : bool *reserve_wal,
1105 : CRSSnapshotAction *snapshot_action,
1106 : bool *two_phase, bool *failover)
1107 : {
1108 : ListCell *lc;
1109 934 : bool snapshot_action_given = false;
1110 934 : bool reserve_wal_given = false;
1111 934 : bool two_phase_given = false;
1112 934 : bool failover_given = false;
1113 :
1114 : /* Parse options */
1115 1884 : foreach(lc, cmd->options)
1116 : {
1117 950 : DefElem *defel = (DefElem *) lfirst(lc);
1118 :
1119 950 : if (strcmp(defel->defname, "snapshot") == 0)
1120 : {
1121 : char *action;
1122 :
1123 662 : if (snapshot_action_given || cmd->kind != REPLICATION_KIND_LOGICAL)
1124 0 : ereport(ERROR,
1125 : (errcode(ERRCODE_SYNTAX_ERROR),
1126 : errmsg("conflicting or redundant options")));
1127 :
1128 662 : action = defGetString(defel);
1129 662 : snapshot_action_given = true;
1130 :
1131 662 : if (strcmp(action, "export") == 0)
1132 2 : *snapshot_action = CRS_EXPORT_SNAPSHOT;
1133 660 : else if (strcmp(action, "nothing") == 0)
1134 274 : *snapshot_action = CRS_NOEXPORT_SNAPSHOT;
1135 386 : else if (strcmp(action, "use") == 0)
1136 386 : *snapshot_action = CRS_USE_SNAPSHOT;
1137 : else
1138 0 : ereport(ERROR,
1139 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1140 : errmsg("unrecognized value for CREATE_REPLICATION_SLOT option \"%s\": \"%s\"",
1141 : defel->defname, action)));
1142 : }
1143 288 : else if (strcmp(defel->defname, "reserve_wal") == 0)
1144 : {
1145 270 : if (reserve_wal_given || cmd->kind != REPLICATION_KIND_PHYSICAL)
1146 0 : ereport(ERROR,
1147 : (errcode(ERRCODE_SYNTAX_ERROR),
1148 : errmsg("conflicting or redundant options")));
1149 :
1150 270 : reserve_wal_given = true;
1151 270 : *reserve_wal = defGetBoolean(defel);
1152 : }
1153 18 : else if (strcmp(defel->defname, "two_phase") == 0)
1154 : {
1155 4 : if (two_phase_given || cmd->kind != REPLICATION_KIND_LOGICAL)
1156 0 : ereport(ERROR,
1157 : (errcode(ERRCODE_SYNTAX_ERROR),
1158 : errmsg("conflicting or redundant options")));
1159 4 : two_phase_given = true;
1160 4 : *two_phase = defGetBoolean(defel);
1161 : }
1162 14 : else if (strcmp(defel->defname, "failover") == 0)
1163 : {
1164 14 : if (failover_given || cmd->kind != REPLICATION_KIND_LOGICAL)
1165 0 : ereport(ERROR,
1166 : (errcode(ERRCODE_SYNTAX_ERROR),
1167 : errmsg("conflicting or redundant options")));
1168 14 : failover_given = true;
1169 14 : *failover = defGetBoolean(defel);
1170 : }
1171 : else
1172 0 : elog(ERROR, "unrecognized option: %s", defel->defname);
1173 : }
1174 934 : }
1175 :
1176 : /*
1177 : * Create a new replication slot.
1178 : */
1179 : static void
1180 934 : CreateReplicationSlot(CreateReplicationSlotCmd *cmd)
1181 : {
1182 934 : const char *snapshot_name = NULL;
1183 : char xloc[MAXFNAMELEN];
1184 : char *slot_name;
1185 934 : bool reserve_wal = false;
1186 934 : bool two_phase = false;
1187 934 : bool failover = false;
1188 934 : CRSSnapshotAction snapshot_action = CRS_EXPORT_SNAPSHOT;
1189 : DestReceiver *dest;
1190 : TupOutputState *tstate;
1191 : TupleDesc tupdesc;
1192 : Datum values[4];
1193 934 : bool nulls[4] = {0};
1194 :
1195 : Assert(!MyReplicationSlot);
1196 :
1197 934 : parseCreateReplSlotOptions(cmd, &reserve_wal, &snapshot_action, &two_phase,
1198 : &failover);
1199 :
1200 934 : if (cmd->kind == REPLICATION_KIND_PHYSICAL)
1201 : {
1202 272 : ReplicationSlotCreate(cmd->slotname, false,
1203 272 : cmd->temporary ? RS_TEMPORARY : RS_PERSISTENT,
1204 : false, false, false);
1205 :
1206 270 : if (reserve_wal)
1207 : {
1208 268 : ReplicationSlotReserveWal();
1209 :
1210 268 : ReplicationSlotMarkDirty();
1211 :
1212 : /* Write this slot to disk if it's a permanent one. */
1213 268 : if (!cmd->temporary)
1214 6 : ReplicationSlotSave();
1215 : }
1216 : }
1217 : else
1218 : {
1219 : LogicalDecodingContext *ctx;
1220 662 : bool need_full_snapshot = false;
1221 :
1222 : Assert(cmd->kind == REPLICATION_KIND_LOGICAL);
1223 :
1224 662 : CheckLogicalDecodingRequirements();
1225 :
1226 : /*
1227 : * Initially create persistent slot as ephemeral - that allows us to
1228 : * nicely handle errors during initialization because it'll get
1229 : * dropped if this transaction fails. We'll make it persistent at the
1230 : * end. Temporary slots can be created as temporary from beginning as
1231 : * they get dropped on error as well.
1232 : */
1233 662 : ReplicationSlotCreate(cmd->slotname, true,
1234 662 : cmd->temporary ? RS_TEMPORARY : RS_EPHEMERAL,
1235 : two_phase, failover, false);
1236 :
1237 : /*
1238 : * Do options check early so that we can bail before calling the
1239 : * DecodingContextFindStartpoint which can take long time.
1240 : */
1241 662 : if (snapshot_action == CRS_EXPORT_SNAPSHOT)
1242 : {
1243 2 : if (IsTransactionBlock())
1244 0 : ereport(ERROR,
1245 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1246 : (errmsg("%s must not be called inside a transaction",
1247 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'export')")));
1248 :
1249 2 : need_full_snapshot = true;
1250 : }
1251 660 : else if (snapshot_action == CRS_USE_SNAPSHOT)
1252 : {
1253 386 : if (!IsTransactionBlock())
1254 0 : ereport(ERROR,
1255 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1256 : (errmsg("%s must be called inside a transaction",
1257 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1258 :
1259 386 : if (XactIsoLevel != XACT_REPEATABLE_READ)
1260 0 : ereport(ERROR,
1261 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1262 : (errmsg("%s must be called in REPEATABLE READ isolation mode transaction",
1263 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1264 386 : if (!XactReadOnly)
1265 0 : ereport(ERROR,
1266 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1267 : (errmsg("%s must be called in a read-only transaction",
1268 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1269 :
1270 386 : if (FirstSnapshotSet)
1271 0 : ereport(ERROR,
1272 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1273 : (errmsg("%s must be called before any query",
1274 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1275 :
1276 386 : if (IsSubTransaction())
1277 0 : ereport(ERROR,
1278 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1279 : (errmsg("%s must not be called in a subtransaction",
1280 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1281 :
1282 386 : need_full_snapshot = true;
1283 : }
1284 :
1285 662 : ctx = CreateInitDecodingContext(cmd->plugin, NIL, need_full_snapshot,
1286 : InvalidXLogRecPtr,
1287 662 : XL_ROUTINE(.page_read = logical_read_xlog_page,
1288 : .segment_open = WalSndSegmentOpen,
1289 : .segment_close = wal_segment_close),
1290 : WalSndPrepareWrite, WalSndWriteData,
1291 : WalSndUpdateProgress);
1292 :
1293 : /*
1294 : * Signal that we don't need the timeout mechanism. We're just
1295 : * creating the replication slot and don't yet accept feedback
1296 : * messages or send keepalives. As we possibly need to wait for
1297 : * further WAL the walsender would otherwise possibly be killed too
1298 : * soon.
1299 : */
1300 662 : last_reply_timestamp = 0;
1301 :
1302 : /* build initial snapshot, might take a while */
1303 662 : DecodingContextFindStartpoint(ctx);
1304 :
1305 : /*
1306 : * Export or use the snapshot if we've been asked to do so.
1307 : *
1308 : * NB. We will convert the snapbuild.c kind of snapshot to normal
1309 : * snapshot when doing this.
1310 : */
1311 662 : if (snapshot_action == CRS_EXPORT_SNAPSHOT)
1312 : {
1313 2 : snapshot_name = SnapBuildExportSnapshot(ctx->snapshot_builder);
1314 : }
1315 660 : else if (snapshot_action == CRS_USE_SNAPSHOT)
1316 : {
1317 : Snapshot snap;
1318 :
1319 386 : snap = SnapBuildInitialSnapshot(ctx->snapshot_builder);
1320 386 : RestoreTransactionSnapshot(snap, MyProc);
1321 : }
1322 :
1323 : /* don't need the decoding context anymore */
1324 662 : FreeDecodingContext(ctx);
1325 :
1326 662 : if (!cmd->temporary)
1327 662 : ReplicationSlotPersist();
1328 : }
1329 :
1330 932 : snprintf(xloc, sizeof(xloc), "%X/%08X",
1331 932 : LSN_FORMAT_ARGS(MyReplicationSlot->data.confirmed_flush));
1332 :
1333 932 : dest = CreateDestReceiver(DestRemoteSimple);
1334 :
1335 : /*----------
1336 : * Need a tuple descriptor representing four columns:
1337 : * - first field: the slot name
1338 : * - second field: LSN at which we became consistent
1339 : * - third field: exported snapshot's name
1340 : * - fourth field: output plugin
1341 : */
1342 932 : tupdesc = CreateTemplateTupleDesc(4);
1343 932 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "slot_name",
1344 : TEXTOID, -1, 0);
1345 932 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "consistent_point",
1346 : TEXTOID, -1, 0);
1347 932 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 3, "snapshot_name",
1348 : TEXTOID, -1, 0);
1349 932 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 4, "output_plugin",
1350 : TEXTOID, -1, 0);
1351 :
1352 : /* prepare for projection of tuples */
1353 932 : tstate = begin_tup_output_tupdesc(dest, tupdesc, &TTSOpsVirtual);
1354 :
1355 : /* slot_name */
1356 932 : slot_name = NameStr(MyReplicationSlot->data.name);
1357 932 : values[0] = CStringGetTextDatum(slot_name);
1358 :
1359 : /* consistent wal location */
1360 932 : values[1] = CStringGetTextDatum(xloc);
1361 :
1362 : /* snapshot name, or NULL if none */
1363 932 : if (snapshot_name != NULL)
1364 2 : values[2] = CStringGetTextDatum(snapshot_name);
1365 : else
1366 930 : nulls[2] = true;
1367 :
1368 : /* plugin, or NULL if none */
1369 932 : if (cmd->plugin != NULL)
1370 662 : values[3] = CStringGetTextDatum(cmd->plugin);
1371 : else
1372 270 : nulls[3] = true;
1373 :
1374 : /* send it to dest */
1375 932 : do_tup_output(tstate, values, nulls);
1376 932 : end_tup_output(tstate);
1377 :
1378 932 : ReplicationSlotRelease();
1379 932 : }
1380 :
1381 : /*
1382 : * Get rid of a replication slot that is no longer wanted.
1383 : */
1384 : static void
1385 534 : DropReplicationSlot(DropReplicationSlotCmd *cmd)
1386 : {
1387 534 : ReplicationSlotDrop(cmd->slotname, !cmd->wait);
1388 528 : }
1389 :
1390 : /*
1391 : * Change the definition of a replication slot.
1392 : */
1393 : static void
1394 12 : AlterReplicationSlot(AlterReplicationSlotCmd *cmd)
1395 : {
1396 12 : bool failover_given = false;
1397 12 : bool two_phase_given = false;
1398 : bool failover;
1399 : bool two_phase;
1400 :
1401 : /* Parse options */
1402 36 : foreach_ptr(DefElem, defel, cmd->options)
1403 : {
1404 12 : if (strcmp(defel->defname, "failover") == 0)
1405 : {
1406 10 : if (failover_given)
1407 0 : ereport(ERROR,
1408 : (errcode(ERRCODE_SYNTAX_ERROR),
1409 : errmsg("conflicting or redundant options")));
1410 10 : failover_given = true;
1411 10 : failover = defGetBoolean(defel);
1412 : }
1413 2 : else if (strcmp(defel->defname, "two_phase") == 0)
1414 : {
1415 2 : if (two_phase_given)
1416 0 : ereport(ERROR,
1417 : (errcode(ERRCODE_SYNTAX_ERROR),
1418 : errmsg("conflicting or redundant options")));
1419 2 : two_phase_given = true;
1420 2 : two_phase = defGetBoolean(defel);
1421 : }
1422 : else
1423 0 : elog(ERROR, "unrecognized option: %s", defel->defname);
1424 : }
1425 :
1426 12 : ReplicationSlotAlter(cmd->slotname,
1427 : failover_given ? &failover : NULL,
1428 : two_phase_given ? &two_phase : NULL);
1429 8 : }
1430 :
1431 : /*
1432 : * Load previously initiated logical slot and prepare for sending data (via
1433 : * WalSndLoop).
1434 : */
1435 : static void
1436 822 : StartLogicalReplication(StartReplicationCmd *cmd)
1437 : {
1438 : StringInfoData buf;
1439 : QueryCompletion qc;
1440 :
1441 : /* make sure that our requirements are still fulfilled */
1442 822 : CheckLogicalDecodingRequirements();
1443 :
1444 : Assert(!MyReplicationSlot);
1445 :
1446 818 : ReplicationSlotAcquire(cmd->slotname, true, true);
1447 :
1448 : /*
1449 : * Force a disconnect, so that the decoding code doesn't need to care
1450 : * about an eventual switch from running in recovery, to running in a
1451 : * normal environment. Client code is expected to handle reconnects.
1452 : */
1453 808 : if (am_cascading_walsender && !RecoveryInProgress())
1454 : {
1455 0 : ereport(LOG,
1456 : (errmsg("terminating walsender process after promotion")));
1457 0 : got_STOPPING = true;
1458 : }
1459 :
1460 : /*
1461 : * Create our decoding context, making it start at the previously ack'ed
1462 : * position.
1463 : *
1464 : * Do this before sending a CopyBothResponse message, so that any errors
1465 : * are reported early.
1466 : */
1467 806 : logical_decoding_ctx =
1468 808 : CreateDecodingContext(cmd->startpoint, cmd->options, false,
1469 808 : XL_ROUTINE(.page_read = logical_read_xlog_page,
1470 : .segment_open = WalSndSegmentOpen,
1471 : .segment_close = wal_segment_close),
1472 : WalSndPrepareWrite, WalSndWriteData,
1473 : WalSndUpdateProgress);
1474 806 : xlogreader = logical_decoding_ctx->reader;
1475 :
1476 806 : WalSndSetState(WALSNDSTATE_CATCHUP);
1477 :
1478 : /* Send a CopyBothResponse message, and start streaming */
1479 806 : pq_beginmessage(&buf, PqMsg_CopyBothResponse);
1480 806 : pq_sendbyte(&buf, 0);
1481 806 : pq_sendint16(&buf, 0);
1482 806 : pq_endmessage(&buf);
1483 806 : pq_flush();
1484 :
1485 : /* Start reading WAL from the oldest required WAL. */
1486 806 : XLogBeginRead(logical_decoding_ctx->reader,
1487 806 : MyReplicationSlot->data.restart_lsn);
1488 :
1489 : /*
1490 : * Report the location after which we'll send out further commits as the
1491 : * current sentPtr.
1492 : */
1493 806 : sentPtr = MyReplicationSlot->data.confirmed_flush;
1494 :
1495 : /* Also update the sent position status in shared memory */
1496 806 : SpinLockAcquire(&MyWalSnd->mutex);
1497 806 : MyWalSnd->sentPtr = MyReplicationSlot->data.restart_lsn;
1498 806 : SpinLockRelease(&MyWalSnd->mutex);
1499 :
1500 806 : replication_active = true;
1501 :
1502 806 : SyncRepInitConfig();
1503 :
1504 : /* Main loop of walsender */
1505 806 : WalSndLoop(XLogSendLogical);
1506 :
1507 378 : FreeDecodingContext(logical_decoding_ctx);
1508 378 : ReplicationSlotRelease();
1509 :
1510 378 : replication_active = false;
1511 378 : if (got_STOPPING)
1512 0 : proc_exit(0);
1513 378 : WalSndSetState(WALSNDSTATE_STARTUP);
1514 :
1515 : /* Get out of COPY mode (CommandComplete). */
1516 378 : SetQueryCompletion(&qc, CMDTAG_COPY, 0);
1517 378 : EndCommand(&qc, DestRemote, false);
1518 378 : }
1519 :
1520 : /*
1521 : * LogicalDecodingContext 'prepare_write' callback.
1522 : *
1523 : * Prepare a write into a StringInfo.
1524 : *
1525 : * Don't do anything lasting in here, it's quite possible that nothing will be done
1526 : * with the data.
1527 : */
1528 : static void
1529 369906 : WalSndPrepareWrite(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid, bool last_write)
1530 : {
1531 : /* can't have sync rep confused by sending the same LSN several times */
1532 369906 : if (!last_write)
1533 806 : lsn = InvalidXLogRecPtr;
1534 :
1535 369906 : resetStringInfo(ctx->out);
1536 :
1537 369906 : pq_sendbyte(ctx->out, PqReplMsg_WALData);
1538 369906 : pq_sendint64(ctx->out, lsn); /* dataStart */
1539 369906 : pq_sendint64(ctx->out, lsn); /* walEnd */
1540 :
1541 : /*
1542 : * Fill out the sendtime later, just as it's done in XLogSendPhysical, but
1543 : * reserve space here.
1544 : */
1545 369906 : pq_sendint64(ctx->out, 0); /* sendtime */
1546 369906 : }
1547 :
1548 : /*
1549 : * LogicalDecodingContext 'write' callback.
1550 : *
1551 : * Actually write out data previously prepared by WalSndPrepareWrite out to
1552 : * the network. Take as long as needed, but process replies from the other
1553 : * side and check timeouts during that.
1554 : */
1555 : static void
1556 369906 : WalSndWriteData(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid,
1557 : bool last_write)
1558 : {
1559 : TimestampTz now;
1560 :
1561 : /*
1562 : * Fill the send timestamp last, so that it is taken as late as possible.
1563 : * This is somewhat ugly, but the protocol is set as it's already used for
1564 : * several releases by streaming physical replication.
1565 : */
1566 369906 : resetStringInfo(&tmpbuf);
1567 369906 : now = GetCurrentTimestamp();
1568 369906 : pq_sendint64(&tmpbuf, now);
1569 369906 : memcpy(&ctx->out->data[1 + sizeof(int64) + sizeof(int64)],
1570 369906 : tmpbuf.data, sizeof(int64));
1571 :
1572 : /* output previously gathered data in a CopyData packet */
1573 369906 : pq_putmessage_noblock(PqMsg_CopyData, ctx->out->data, ctx->out->len);
1574 :
1575 369906 : CHECK_FOR_INTERRUPTS();
1576 :
1577 : /* Try to flush pending output to the client */
1578 369906 : if (pq_flush_if_writable() != 0)
1579 58 : WalSndShutdown();
1580 :
1581 : /* Try taking fast path unless we get too close to walsender timeout. */
1582 369848 : if (now < TimestampTzPlusMilliseconds(last_reply_timestamp,
1583 369848 : wal_sender_timeout / 2) &&
1584 369848 : !pq_is_send_pending())
1585 : {
1586 369172 : return;
1587 : }
1588 :
1589 : /* If we have pending write here, go to slow path */
1590 676 : ProcessPendingWrites();
1591 : }
1592 :
1593 : /*
1594 : * Wait until there is no pending write. Also process replies from the other
1595 : * side and check timeouts during that.
1596 : */
1597 : static void
1598 676 : ProcessPendingWrites(void)
1599 : {
1600 : for (;;)
1601 814 : {
1602 : long sleeptime;
1603 :
1604 : /* Check for input from the client */
1605 1490 : ProcessRepliesIfAny();
1606 :
1607 : /* die if timeout was reached */
1608 1490 : WalSndCheckTimeOut();
1609 :
1610 : /* Send keepalive if the time has come */
1611 1490 : WalSndKeepaliveIfNecessary();
1612 :
1613 1490 : if (!pq_is_send_pending())
1614 676 : break;
1615 :
1616 814 : sleeptime = WalSndComputeSleeptime(GetCurrentTimestamp());
1617 :
1618 : /* Sleep until something happens or we time out */
1619 814 : WalSndWait(WL_SOCKET_WRITEABLE | WL_SOCKET_READABLE, sleeptime,
1620 : WAIT_EVENT_WAL_SENDER_WRITE_DATA);
1621 :
1622 : /* Clear any already-pending wakeups */
1623 814 : ResetLatch(MyLatch);
1624 :
1625 814 : CHECK_FOR_INTERRUPTS();
1626 :
1627 : /* Process any requests or signals received recently */
1628 814 : if (ConfigReloadPending)
1629 : {
1630 0 : ConfigReloadPending = false;
1631 0 : ProcessConfigFile(PGC_SIGHUP);
1632 0 : SyncRepInitConfig();
1633 : }
1634 :
1635 : /* Try to flush pending output to the client */
1636 814 : if (pq_flush_if_writable() != 0)
1637 0 : WalSndShutdown();
1638 : }
1639 :
1640 : /* reactivate latch so WalSndLoop knows to continue */
1641 676 : SetLatch(MyLatch);
1642 676 : }
1643 :
1644 : /*
1645 : * LogicalDecodingContext 'update_progress' callback.
1646 : *
1647 : * Write the current position to the lag tracker (see XLogSendPhysical).
1648 : *
1649 : * When skipping empty transactions, send a keepalive message if necessary.
1650 : */
1651 : static void
1652 4866 : WalSndUpdateProgress(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid,
1653 : bool skipped_xact)
1654 : {
1655 : static TimestampTz sendTime = 0;
1656 4866 : TimestampTz now = GetCurrentTimestamp();
1657 4866 : bool pending_writes = false;
1658 4866 : bool end_xact = ctx->end_xact;
1659 :
1660 : /*
1661 : * Track lag no more than once per WALSND_LOGICAL_LAG_TRACK_INTERVAL_MS to
1662 : * avoid flooding the lag tracker when we commit frequently.
1663 : *
1664 : * We don't have a mechanism to get the ack for any LSN other than end
1665 : * xact LSN from the downstream. So, we track lag only for end of
1666 : * transaction LSN.
1667 : */
1668 : #define WALSND_LOGICAL_LAG_TRACK_INTERVAL_MS 1000
1669 4866 : if (end_xact && TimestampDifferenceExceeds(sendTime, now,
1670 : WALSND_LOGICAL_LAG_TRACK_INTERVAL_MS))
1671 : {
1672 544 : LagTrackerWrite(lsn, now);
1673 544 : sendTime = now;
1674 : }
1675 :
1676 : /*
1677 : * When skipping empty transactions in synchronous replication, we send a
1678 : * keepalive message to avoid delaying such transactions.
1679 : *
1680 : * It is okay to check sync_standbys_status without lock here as in the
1681 : * worst case we will just send an extra keepalive message when it is
1682 : * really not required.
1683 : */
1684 4866 : if (skipped_xact &&
1685 702 : SyncRepRequested() &&
1686 702 : (((volatile WalSndCtlData *) WalSndCtl)->sync_standbys_status & SYNC_STANDBY_DEFINED))
1687 : {
1688 0 : WalSndKeepalive(false, lsn);
1689 :
1690 : /* Try to flush pending output to the client */
1691 0 : if (pq_flush_if_writable() != 0)
1692 0 : WalSndShutdown();
1693 :
1694 : /* If we have pending write here, make sure it's actually flushed */
1695 0 : if (pq_is_send_pending())
1696 0 : pending_writes = true;
1697 : }
1698 :
1699 : /*
1700 : * Process pending writes if any or try to send a keepalive if required.
1701 : * We don't need to try sending keep alive messages at the transaction end
1702 : * as that will be done at a later point in time. This is required only
1703 : * for large transactions where we don't send any changes to the
1704 : * downstream and the receiver can timeout due to that.
1705 : */
1706 4866 : if (pending_writes || (!end_xact &&
1707 3080 : now >= TimestampTzPlusMilliseconds(last_reply_timestamp,
1708 : wal_sender_timeout / 2)))
1709 0 : ProcessPendingWrites();
1710 4866 : }
1711 :
1712 : /*
1713 : * Wake up the logical walsender processes with logical failover slots if the
1714 : * currently acquired physical slot is specified in synchronized_standby_slots GUC.
1715 : */
1716 : void
1717 59480 : PhysicalWakeupLogicalWalSnd(void)
1718 : {
1719 : Assert(MyReplicationSlot && SlotIsPhysical(MyReplicationSlot));
1720 :
1721 : /*
1722 : * If we are running in a standby, there is no need to wake up walsenders.
1723 : * This is because we do not support syncing slots to cascading standbys,
1724 : * so, there are no walsenders waiting for standbys to catch up.
1725 : */
1726 59480 : if (RecoveryInProgress())
1727 106 : return;
1728 :
1729 59374 : if (SlotExistsInSyncStandbySlots(NameStr(MyReplicationSlot->data.name)))
1730 14 : ConditionVariableBroadcast(&WalSndCtl->wal_confirm_rcv_cv);
1731 : }
1732 :
1733 : /*
1734 : * Returns true if not all standbys have caught up to the flushed position
1735 : * (flushed_lsn) when the current acquired slot is a logical failover
1736 : * slot and we are streaming; otherwise, returns false.
1737 : *
1738 : * If returning true, the function sets the appropriate wait event in
1739 : * wait_event; otherwise, wait_event is set to 0.
1740 : */
1741 : static bool
1742 31914 : NeedToWaitForStandbys(XLogRecPtr flushed_lsn, uint32 *wait_event)
1743 : {
1744 31914 : int elevel = got_STOPPING ? ERROR : WARNING;
1745 : bool failover_slot;
1746 :
1747 31914 : failover_slot = (replication_active && MyReplicationSlot->data.failover);
1748 :
1749 : /*
1750 : * Note that after receiving the shutdown signal, an ERROR is reported if
1751 : * any slots are dropped, invalidated, or inactive. This measure is taken
1752 : * to prevent the walsender from waiting indefinitely.
1753 : */
1754 31914 : if (failover_slot && !StandbySlotsHaveCaughtup(flushed_lsn, elevel))
1755 : {
1756 10 : *wait_event = WAIT_EVENT_WAIT_FOR_STANDBY_CONFIRMATION;
1757 10 : return true;
1758 : }
1759 :
1760 31904 : *wait_event = 0;
1761 31904 : return false;
1762 : }
1763 :
1764 : /*
1765 : * Returns true if we need to wait for WALs to be flushed to disk, or if not
1766 : * all standbys have caught up to the flushed position (flushed_lsn) when the
1767 : * current acquired slot is a logical failover slot and we are
1768 : * streaming; otherwise, returns false.
1769 : *
1770 : * If returning true, the function sets the appropriate wait event in
1771 : * wait_event; otherwise, wait_event is set to 0.
1772 : */
1773 : static bool
1774 43290 : NeedToWaitForWal(XLogRecPtr target_lsn, XLogRecPtr flushed_lsn,
1775 : uint32 *wait_event)
1776 : {
1777 : /* Check if we need to wait for WALs to be flushed to disk */
1778 43290 : if (target_lsn > flushed_lsn)
1779 : {
1780 18384 : *wait_event = WAIT_EVENT_WAL_SENDER_WAIT_FOR_WAL;
1781 18384 : return true;
1782 : }
1783 :
1784 : /* Check if the standby slots have caught up to the flushed position */
1785 24906 : return NeedToWaitForStandbys(flushed_lsn, wait_event);
1786 : }
1787 :
1788 : /*
1789 : * Wait till WAL < loc is flushed to disk so it can be safely sent to client.
1790 : *
1791 : * If the walsender holds a logical failover slot, we also wait for all the
1792 : * specified streaming replication standby servers to confirm receipt of WAL
1793 : * up to RecentFlushPtr. It is beneficial to wait here for the confirmation
1794 : * up to RecentFlushPtr rather than waiting before transmitting each change
1795 : * to logical subscribers, which is already covered by RecentFlushPtr.
1796 : *
1797 : * Returns end LSN of flushed WAL. Normally this will be >= loc, but if we
1798 : * detect a shutdown request (either from postmaster or client) we will return
1799 : * early, so caller must always check.
1800 : */
1801 : static XLogRecPtr
1802 32392 : WalSndWaitForWal(XLogRecPtr loc)
1803 : {
1804 : int wakeEvents;
1805 32392 : uint32 wait_event = 0;
1806 : static XLogRecPtr RecentFlushPtr = InvalidXLogRecPtr;
1807 32392 : TimestampTz last_flush = 0;
1808 :
1809 : /*
1810 : * Fast path to avoid acquiring the spinlock in case we already know we
1811 : * have enough WAL available and all the standby servers have confirmed
1812 : * receipt of WAL up to RecentFlushPtr. This is particularly interesting
1813 : * if we're far behind.
1814 : */
1815 32392 : if (!XLogRecPtrIsInvalid(RecentFlushPtr) &&
1816 31290 : !NeedToWaitForWal(loc, RecentFlushPtr, &wait_event))
1817 21188 : return RecentFlushPtr;
1818 :
1819 : /*
1820 : * Within the loop, we wait for the necessary WALs to be flushed to disk
1821 : * first, followed by waiting for standbys to catch up if there are enough
1822 : * WALs (see NeedToWaitForWal()) or upon receiving the shutdown signal.
1823 : */
1824 : for (;;)
1825 8146 : {
1826 19350 : bool wait_for_standby_at_stop = false;
1827 : long sleeptime;
1828 : TimestampTz now;
1829 :
1830 : /* Clear any already-pending wakeups */
1831 19350 : ResetLatch(MyLatch);
1832 :
1833 19350 : CHECK_FOR_INTERRUPTS();
1834 :
1835 : /* Process any requests or signals received recently */
1836 19338 : if (ConfigReloadPending)
1837 : {
1838 14 : ConfigReloadPending = false;
1839 14 : ProcessConfigFile(PGC_SIGHUP);
1840 14 : SyncRepInitConfig();
1841 : }
1842 :
1843 : /* Check for input from the client */
1844 19338 : ProcessRepliesIfAny();
1845 :
1846 : /*
1847 : * If we're shutting down, trigger pending WAL to be written out,
1848 : * otherwise we'd possibly end up waiting for WAL that never gets
1849 : * written, because walwriter has shut down already.
1850 : */
1851 19008 : if (got_STOPPING)
1852 7008 : XLogBackgroundFlush();
1853 :
1854 : /*
1855 : * To avoid the scenario where standbys need to catch up to a newer
1856 : * WAL location in each iteration, we update our idea of the currently
1857 : * flushed position only if we are not waiting for standbys to catch
1858 : * up.
1859 : */
1860 19008 : if (wait_event != WAIT_EVENT_WAIT_FOR_STANDBY_CONFIRMATION)
1861 : {
1862 18998 : if (!RecoveryInProgress())
1863 17244 : RecentFlushPtr = GetFlushRecPtr(NULL);
1864 : else
1865 1754 : RecentFlushPtr = GetXLogReplayRecPtr(NULL);
1866 : }
1867 :
1868 : /*
1869 : * If postmaster asked us to stop and the standby slots have caught up
1870 : * to the flushed position, don't wait anymore.
1871 : *
1872 : * It's important to do this check after the recomputation of
1873 : * RecentFlushPtr, so we can send all remaining data before shutting
1874 : * down.
1875 : */
1876 19008 : if (got_STOPPING)
1877 : {
1878 7008 : if (NeedToWaitForStandbys(RecentFlushPtr, &wait_event))
1879 0 : wait_for_standby_at_stop = true;
1880 : else
1881 7008 : break;
1882 : }
1883 :
1884 : /*
1885 : * We only send regular messages to the client for full decoded
1886 : * transactions, but a synchronous replication and walsender shutdown
1887 : * possibly are waiting for a later location. So, before sleeping, we
1888 : * send a ping containing the flush location. If the receiver is
1889 : * otherwise idle, this keepalive will trigger a reply. Processing the
1890 : * reply will update these MyWalSnd locations.
1891 : */
1892 12000 : if (MyWalSnd->flush < sentPtr &&
1893 5972 : MyWalSnd->write < sentPtr &&
1894 3480 : !waiting_for_ping_response)
1895 3480 : WalSndKeepalive(false, InvalidXLogRecPtr);
1896 :
1897 : /*
1898 : * Exit the loop if already caught up and doesn't need to wait for
1899 : * standby slots.
1900 : */
1901 12000 : if (!wait_for_standby_at_stop &&
1902 12000 : !NeedToWaitForWal(loc, RecentFlushPtr, &wait_event))
1903 3708 : break;
1904 :
1905 : /*
1906 : * Waiting for new WAL or waiting for standbys to catch up. Since we
1907 : * need to wait, we're now caught up.
1908 : */
1909 8292 : WalSndCaughtUp = true;
1910 :
1911 : /*
1912 : * Try to flush any pending output to the client.
1913 : */
1914 8292 : if (pq_flush_if_writable() != 0)
1915 0 : WalSndShutdown();
1916 :
1917 : /*
1918 : * If we have received CopyDone from the client, sent CopyDone
1919 : * ourselves, and the output buffer is empty, it's time to exit
1920 : * streaming, so fail the current WAL fetch request.
1921 : */
1922 8292 : if (streamingDoneReceiving && streamingDoneSending &&
1923 146 : !pq_is_send_pending())
1924 146 : break;
1925 :
1926 : /* die if timeout was reached */
1927 8146 : WalSndCheckTimeOut();
1928 :
1929 : /* Send keepalive if the time has come */
1930 8146 : WalSndKeepaliveIfNecessary();
1931 :
1932 : /*
1933 : * Sleep until something happens or we time out. Also wait for the
1934 : * socket becoming writable, if there's still pending output.
1935 : * Otherwise we might sit on sendable output data while waiting for
1936 : * new WAL to be generated. (But if we have nothing to send, we don't
1937 : * want to wake on socket-writable.)
1938 : */
1939 8146 : now = GetCurrentTimestamp();
1940 8146 : sleeptime = WalSndComputeSleeptime(now);
1941 :
1942 8146 : wakeEvents = WL_SOCKET_READABLE;
1943 :
1944 8146 : if (pq_is_send_pending())
1945 0 : wakeEvents |= WL_SOCKET_WRITEABLE;
1946 :
1947 : Assert(wait_event != 0);
1948 :
1949 : /* Report IO statistics, if needed */
1950 8146 : if (TimestampDifferenceExceeds(last_flush, now,
1951 : WALSENDER_STATS_FLUSH_INTERVAL))
1952 : {
1953 2896 : pgstat_flush_io(false);
1954 2896 : (void) pgstat_flush_backend(false, PGSTAT_BACKEND_FLUSH_IO);
1955 2896 : last_flush = now;
1956 : }
1957 :
1958 8146 : WalSndWait(wakeEvents, sleeptime, wait_event);
1959 : }
1960 :
1961 : /* reactivate latch so WalSndLoop knows to continue */
1962 10862 : SetLatch(MyLatch);
1963 10862 : return RecentFlushPtr;
1964 : }
1965 :
1966 : /*
1967 : * Execute an incoming replication command.
1968 : *
1969 : * Returns true if the cmd_string was recognized as WalSender command, false
1970 : * if not.
1971 : */
1972 : bool
1973 10402 : exec_replication_command(const char *cmd_string)
1974 : {
1975 : yyscan_t scanner;
1976 : int parse_rc;
1977 : Node *cmd_node;
1978 : const char *cmdtag;
1979 10402 : MemoryContext old_context = CurrentMemoryContext;
1980 :
1981 : /* We save and re-use the cmd_context across calls */
1982 : static MemoryContext cmd_context = NULL;
1983 :
1984 : /*
1985 : * If WAL sender has been told that shutdown is getting close, switch its
1986 : * status accordingly to handle the next replication commands correctly.
1987 : */
1988 10402 : if (got_STOPPING)
1989 0 : WalSndSetState(WALSNDSTATE_STOPPING);
1990 :
1991 : /*
1992 : * Throw error if in stopping mode. We need prevent commands that could
1993 : * generate WAL while the shutdown checkpoint is being written. To be
1994 : * safe, we just prohibit all new commands.
1995 : */
1996 10402 : if (MyWalSnd->state == WALSNDSTATE_STOPPING)
1997 0 : ereport(ERROR,
1998 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1999 : errmsg("cannot execute new commands while WAL sender is in stopping mode")));
2000 :
2001 : /*
2002 : * CREATE_REPLICATION_SLOT ... LOGICAL exports a snapshot until the next
2003 : * command arrives. Clean up the old stuff if there's anything.
2004 : */
2005 10402 : SnapBuildClearExportedSnapshot();
2006 :
2007 10402 : CHECK_FOR_INTERRUPTS();
2008 :
2009 : /*
2010 : * Prepare to parse and execute the command.
2011 : *
2012 : * Because replication command execution can involve beginning or ending
2013 : * transactions, we need a working context that will survive that, so we
2014 : * make it a child of TopMemoryContext. That in turn creates a hazard of
2015 : * long-lived memory leaks if we lose track of the working context. We
2016 : * deal with that by creating it only once per walsender, and resetting it
2017 : * for each new command. (Normally this reset is a no-op, but if the
2018 : * prior exec_replication_command call failed with an error, it won't be.)
2019 : *
2020 : * This is subtler than it looks. The transactions we manage can extend
2021 : * across replication commands, indeed SnapBuildClearExportedSnapshot
2022 : * might have just ended one. Because transaction exit will revert to the
2023 : * memory context that was current at transaction start, we need to be
2024 : * sure that that context is still valid. That motivates re-using the
2025 : * same cmd_context rather than making a new one each time.
2026 : */
2027 10402 : if (cmd_context == NULL)
2028 2278 : cmd_context = AllocSetContextCreate(TopMemoryContext,
2029 : "Replication command context",
2030 : ALLOCSET_DEFAULT_SIZES);
2031 : else
2032 8124 : MemoryContextReset(cmd_context);
2033 :
2034 10402 : MemoryContextSwitchTo(cmd_context);
2035 :
2036 10402 : replication_scanner_init(cmd_string, &scanner);
2037 :
2038 : /*
2039 : * Is it a WalSender command?
2040 : */
2041 10402 : if (!replication_scanner_is_replication_command(scanner))
2042 : {
2043 : /* Nope; clean up and get out. */
2044 4626 : replication_scanner_finish(scanner);
2045 :
2046 4626 : MemoryContextSwitchTo(old_context);
2047 4626 : MemoryContextReset(cmd_context);
2048 :
2049 : /* XXX this is a pretty random place to make this check */
2050 4626 : if (MyDatabaseId == InvalidOid)
2051 0 : ereport(ERROR,
2052 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2053 : errmsg("cannot execute SQL commands in WAL sender for physical replication")));
2054 :
2055 : /* Tell the caller that this wasn't a WalSender command. */
2056 4626 : return false;
2057 : }
2058 :
2059 : /*
2060 : * Looks like a WalSender command, so parse it.
2061 : */
2062 5776 : parse_rc = replication_yyparse(&cmd_node, scanner);
2063 5776 : if (parse_rc != 0)
2064 0 : ereport(ERROR,
2065 : (errcode(ERRCODE_SYNTAX_ERROR),
2066 : errmsg_internal("replication command parser returned %d",
2067 : parse_rc)));
2068 5776 : replication_scanner_finish(scanner);
2069 :
2070 : /*
2071 : * Report query to various monitoring facilities. For this purpose, we
2072 : * report replication commands just like SQL commands.
2073 : */
2074 5776 : debug_query_string = cmd_string;
2075 :
2076 5776 : pgstat_report_activity(STATE_RUNNING, cmd_string);
2077 :
2078 : /*
2079 : * Log replication command if log_replication_commands is enabled. Even
2080 : * when it's disabled, log the command with DEBUG1 level for backward
2081 : * compatibility.
2082 : */
2083 5776 : ereport(log_replication_commands ? LOG : DEBUG1,
2084 : (errmsg("received replication command: %s", cmd_string)));
2085 :
2086 : /*
2087 : * Disallow replication commands in aborted transaction blocks.
2088 : */
2089 5776 : if (IsAbortedTransactionBlockState())
2090 0 : ereport(ERROR,
2091 : (errcode(ERRCODE_IN_FAILED_SQL_TRANSACTION),
2092 : errmsg("current transaction is aborted, "
2093 : "commands ignored until end of transaction block")));
2094 :
2095 5776 : CHECK_FOR_INTERRUPTS();
2096 :
2097 : /*
2098 : * Allocate buffers that will be used for each outgoing and incoming
2099 : * message. We do this just once per command to reduce palloc overhead.
2100 : */
2101 5776 : initStringInfo(&output_message);
2102 5776 : initStringInfo(&reply_message);
2103 5776 : initStringInfo(&tmpbuf);
2104 :
2105 5776 : switch (cmd_node->type)
2106 : {
2107 1412 : case T_IdentifySystemCmd:
2108 1412 : cmdtag = "IDENTIFY_SYSTEM";
2109 1412 : set_ps_display(cmdtag);
2110 1412 : IdentifySystem();
2111 1412 : EndReplicationCommand(cmdtag);
2112 1412 : break;
2113 :
2114 12 : case T_ReadReplicationSlotCmd:
2115 12 : cmdtag = "READ_REPLICATION_SLOT";
2116 12 : set_ps_display(cmdtag);
2117 12 : ReadReplicationSlot((ReadReplicationSlotCmd *) cmd_node);
2118 10 : EndReplicationCommand(cmdtag);
2119 10 : break;
2120 :
2121 356 : case T_BaseBackupCmd:
2122 356 : cmdtag = "BASE_BACKUP";
2123 356 : set_ps_display(cmdtag);
2124 356 : PreventInTransactionBlock(true, cmdtag);
2125 356 : SendBaseBackup((BaseBackupCmd *) cmd_node, uploaded_manifest);
2126 304 : EndReplicationCommand(cmdtag);
2127 304 : break;
2128 :
2129 934 : case T_CreateReplicationSlotCmd:
2130 934 : cmdtag = "CREATE_REPLICATION_SLOT";
2131 934 : set_ps_display(cmdtag);
2132 934 : CreateReplicationSlot((CreateReplicationSlotCmd *) cmd_node);
2133 932 : EndReplicationCommand(cmdtag);
2134 932 : break;
2135 :
2136 534 : case T_DropReplicationSlotCmd:
2137 534 : cmdtag = "DROP_REPLICATION_SLOT";
2138 534 : set_ps_display(cmdtag);
2139 534 : DropReplicationSlot((DropReplicationSlotCmd *) cmd_node);
2140 528 : EndReplicationCommand(cmdtag);
2141 528 : break;
2142 :
2143 12 : case T_AlterReplicationSlotCmd:
2144 12 : cmdtag = "ALTER_REPLICATION_SLOT";
2145 12 : set_ps_display(cmdtag);
2146 12 : AlterReplicationSlot((AlterReplicationSlotCmd *) cmd_node);
2147 8 : EndReplicationCommand(cmdtag);
2148 8 : break;
2149 :
2150 1342 : case T_StartReplicationCmd:
2151 : {
2152 1342 : StartReplicationCmd *cmd = (StartReplicationCmd *) cmd_node;
2153 :
2154 1342 : cmdtag = "START_REPLICATION";
2155 1342 : set_ps_display(cmdtag);
2156 1342 : PreventInTransactionBlock(true, cmdtag);
2157 :
2158 1342 : if (cmd->kind == REPLICATION_KIND_PHYSICAL)
2159 520 : StartReplication(cmd);
2160 : else
2161 822 : StartLogicalReplication(cmd);
2162 :
2163 : /* dupe, but necessary per libpqrcv_endstreaming */
2164 676 : EndReplicationCommand(cmdtag);
2165 :
2166 : Assert(xlogreader != NULL);
2167 676 : break;
2168 : }
2169 :
2170 28 : case T_TimeLineHistoryCmd:
2171 28 : cmdtag = "TIMELINE_HISTORY";
2172 28 : set_ps_display(cmdtag);
2173 28 : PreventInTransactionBlock(true, cmdtag);
2174 28 : SendTimeLineHistory((TimeLineHistoryCmd *) cmd_node);
2175 28 : EndReplicationCommand(cmdtag);
2176 28 : break;
2177 :
2178 1124 : case T_VariableShowStmt:
2179 : {
2180 1124 : DestReceiver *dest = CreateDestReceiver(DestRemoteSimple);
2181 1124 : VariableShowStmt *n = (VariableShowStmt *) cmd_node;
2182 :
2183 1124 : cmdtag = "SHOW";
2184 1124 : set_ps_display(cmdtag);
2185 :
2186 : /* syscache access needs a transaction environment */
2187 1124 : StartTransactionCommand();
2188 1124 : GetPGVariable(n->name, dest);
2189 1124 : CommitTransactionCommand();
2190 1124 : EndReplicationCommand(cmdtag);
2191 : }
2192 1124 : break;
2193 :
2194 22 : case T_UploadManifestCmd:
2195 22 : cmdtag = "UPLOAD_MANIFEST";
2196 22 : set_ps_display(cmdtag);
2197 22 : PreventInTransactionBlock(true, cmdtag);
2198 22 : UploadManifest();
2199 20 : EndReplicationCommand(cmdtag);
2200 20 : break;
2201 :
2202 0 : default:
2203 0 : elog(ERROR, "unrecognized replication command node tag: %u",
2204 : cmd_node->type);
2205 : }
2206 :
2207 : /*
2208 : * Done. Revert to caller's memory context, and clean out the cmd_context
2209 : * to recover memory right away.
2210 : */
2211 5042 : MemoryContextSwitchTo(old_context);
2212 5042 : MemoryContextReset(cmd_context);
2213 :
2214 : /*
2215 : * We need not update ps display or pg_stat_activity, because PostgresMain
2216 : * will reset those to "idle". But we must reset debug_query_string to
2217 : * ensure it doesn't become a dangling pointer.
2218 : */
2219 5042 : debug_query_string = NULL;
2220 :
2221 5042 : return true;
2222 : }
2223 :
2224 : /*
2225 : * Process any incoming messages while streaming. Also checks if the remote
2226 : * end has closed the connection.
2227 : */
2228 : static void
2229 1530162 : ProcessRepliesIfAny(void)
2230 : {
2231 : unsigned char firstchar;
2232 : int maxmsglen;
2233 : int r;
2234 1530162 : bool received = false;
2235 :
2236 1530162 : last_processing = GetCurrentTimestamp();
2237 :
2238 : /*
2239 : * If we already received a CopyDone from the frontend, any subsequent
2240 : * message is the beginning of a new command, and should be processed in
2241 : * the main processing loop.
2242 : */
2243 3311302 : while (!streamingDoneReceiving)
2244 : {
2245 1779796 : pq_startmsgread();
2246 1779796 : r = pq_getbyte_if_available(&firstchar);
2247 1779796 : if (r < 0)
2248 : {
2249 : /* unexpected error or EOF */
2250 26 : ereport(COMMERROR,
2251 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2252 : errmsg("unexpected EOF on standby connection")));
2253 26 : proc_exit(0);
2254 : }
2255 1779770 : if (r == 0)
2256 : {
2257 : /* no data available without blocking */
2258 1528324 : pq_endmsgread();
2259 1528324 : break;
2260 : }
2261 :
2262 : /* Validate message type and set packet size limit */
2263 251446 : switch (firstchar)
2264 : {
2265 250302 : case PqMsg_CopyData:
2266 250302 : maxmsglen = PQ_LARGE_MESSAGE_LIMIT;
2267 250302 : break;
2268 1144 : case PqMsg_CopyDone:
2269 : case PqMsg_Terminate:
2270 1144 : maxmsglen = PQ_SMALL_MESSAGE_LIMIT;
2271 1144 : break;
2272 0 : default:
2273 0 : ereport(FATAL,
2274 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2275 : errmsg("invalid standby message type \"%c\"",
2276 : firstchar)));
2277 : maxmsglen = 0; /* keep compiler quiet */
2278 : break;
2279 : }
2280 :
2281 : /* Read the message contents */
2282 251446 : resetStringInfo(&reply_message);
2283 251446 : if (pq_getmessage(&reply_message, maxmsglen))
2284 : {
2285 0 : ereport(COMMERROR,
2286 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2287 : errmsg("unexpected EOF on standby connection")));
2288 0 : proc_exit(0);
2289 : }
2290 :
2291 : /* ... and process it */
2292 251446 : switch (firstchar)
2293 : {
2294 : /*
2295 : * PqMsg_CopyData means a standby reply wrapped in a CopyData
2296 : * packet.
2297 : */
2298 250302 : case PqMsg_CopyData:
2299 250302 : ProcessStandbyMessage();
2300 250302 : received = true;
2301 250302 : break;
2302 :
2303 : /*
2304 : * PqMsg_CopyDone means the standby requested to finish
2305 : * streaming. Reply with CopyDone, if we had not sent that
2306 : * already.
2307 : */
2308 676 : case PqMsg_CopyDone:
2309 676 : if (!streamingDoneSending)
2310 : {
2311 650 : pq_putmessage_noblock(PqMsg_CopyDone, NULL, 0);
2312 650 : streamingDoneSending = true;
2313 : }
2314 :
2315 676 : streamingDoneReceiving = true;
2316 676 : received = true;
2317 676 : break;
2318 :
2319 : /*
2320 : * PqMsg_Terminate means that the standby is closing down the
2321 : * socket.
2322 : */
2323 468 : case PqMsg_Terminate:
2324 468 : proc_exit(0);
2325 :
2326 250978 : default:
2327 : Assert(false); /* NOT REACHED */
2328 : }
2329 : }
2330 :
2331 : /*
2332 : * Save the last reply timestamp if we've received at least one reply.
2333 : */
2334 1529668 : if (received)
2335 : {
2336 84404 : last_reply_timestamp = last_processing;
2337 84404 : waiting_for_ping_response = false;
2338 : }
2339 1529668 : }
2340 :
2341 : /*
2342 : * Process a status update message received from standby.
2343 : */
2344 : static void
2345 250302 : ProcessStandbyMessage(void)
2346 : {
2347 : char msgtype;
2348 :
2349 : /*
2350 : * Check message type from the first byte.
2351 : */
2352 250302 : msgtype = pq_getmsgbyte(&reply_message);
2353 :
2354 250302 : switch (msgtype)
2355 : {
2356 247366 : case PqReplMsg_StandbyStatusUpdate:
2357 247366 : ProcessStandbyReplyMessage();
2358 247366 : break;
2359 :
2360 272 : case PqReplMsg_HotStandbyFeedback:
2361 272 : ProcessStandbyHSFeedbackMessage();
2362 272 : break;
2363 :
2364 2664 : case PqReplMsg_PrimaryStatusRequest:
2365 2664 : ProcessStandbyPSRequestMessage();
2366 2664 : break;
2367 :
2368 0 : default:
2369 0 : ereport(COMMERROR,
2370 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2371 : errmsg("unexpected message type \"%c\"", msgtype)));
2372 0 : proc_exit(0);
2373 : }
2374 250302 : }
2375 :
2376 : /*
2377 : * Remember that a walreceiver just confirmed receipt of lsn `lsn`.
2378 : */
2379 : static void
2380 133326 : PhysicalConfirmReceivedLocation(XLogRecPtr lsn)
2381 : {
2382 133326 : bool changed = false;
2383 133326 : ReplicationSlot *slot = MyReplicationSlot;
2384 :
2385 : Assert(lsn != InvalidXLogRecPtr);
2386 133326 : SpinLockAcquire(&slot->mutex);
2387 133326 : if (slot->data.restart_lsn != lsn)
2388 : {
2389 59466 : changed = true;
2390 59466 : slot->data.restart_lsn = lsn;
2391 : }
2392 133326 : SpinLockRelease(&slot->mutex);
2393 :
2394 133326 : if (changed)
2395 : {
2396 59466 : ReplicationSlotMarkDirty();
2397 59466 : ReplicationSlotsComputeRequiredLSN();
2398 59466 : PhysicalWakeupLogicalWalSnd();
2399 : }
2400 :
2401 : /*
2402 : * One could argue that the slot should be saved to disk now, but that'd
2403 : * be energy wasted - the worst thing lost information could cause here is
2404 : * to give wrong information in a statistics view - we'll just potentially
2405 : * be more conservative in removing files.
2406 : */
2407 133326 : }
2408 :
2409 : /*
2410 : * Regular reply from standby advising of WAL locations on standby server.
2411 : */
2412 : static void
2413 247366 : ProcessStandbyReplyMessage(void)
2414 : {
2415 : XLogRecPtr writePtr,
2416 : flushPtr,
2417 : applyPtr;
2418 : bool replyRequested;
2419 : TimeOffset writeLag,
2420 : flushLag,
2421 : applyLag;
2422 : bool clearLagTimes;
2423 : TimestampTz now;
2424 : TimestampTz replyTime;
2425 :
2426 : static bool fullyAppliedLastTime = false;
2427 :
2428 : /* the caller already consumed the msgtype byte */
2429 247366 : writePtr = pq_getmsgint64(&reply_message);
2430 247366 : flushPtr = pq_getmsgint64(&reply_message);
2431 247366 : applyPtr = pq_getmsgint64(&reply_message);
2432 247366 : replyTime = pq_getmsgint64(&reply_message);
2433 247366 : replyRequested = pq_getmsgbyte(&reply_message);
2434 :
2435 247366 : if (message_level_is_interesting(DEBUG2))
2436 : {
2437 : char *replyTimeStr;
2438 :
2439 : /* Copy because timestamptz_to_str returns a static buffer */
2440 1328 : replyTimeStr = pstrdup(timestamptz_to_str(replyTime));
2441 :
2442 1328 : elog(DEBUG2, "write %X/%08X flush %X/%08X apply %X/%08X%s reply_time %s",
2443 : LSN_FORMAT_ARGS(writePtr),
2444 : LSN_FORMAT_ARGS(flushPtr),
2445 : LSN_FORMAT_ARGS(applyPtr),
2446 : replyRequested ? " (reply requested)" : "",
2447 : replyTimeStr);
2448 :
2449 1328 : pfree(replyTimeStr);
2450 : }
2451 :
2452 : /* See if we can compute the round-trip lag for these positions. */
2453 247366 : now = GetCurrentTimestamp();
2454 247366 : writeLag = LagTrackerRead(SYNC_REP_WAIT_WRITE, writePtr, now);
2455 247366 : flushLag = LagTrackerRead(SYNC_REP_WAIT_FLUSH, flushPtr, now);
2456 247366 : applyLag = LagTrackerRead(SYNC_REP_WAIT_APPLY, applyPtr, now);
2457 :
2458 : /*
2459 : * If the standby reports that it has fully replayed the WAL in two
2460 : * consecutive reply messages, then the second such message must result
2461 : * from wal_receiver_status_interval expiring on the standby. This is a
2462 : * convenient time to forget the lag times measured when it last
2463 : * wrote/flushed/applied a WAL record, to avoid displaying stale lag data
2464 : * until more WAL traffic arrives.
2465 : */
2466 247366 : clearLagTimes = false;
2467 247366 : if (applyPtr == sentPtr)
2468 : {
2469 9404 : if (fullyAppliedLastTime)
2470 2512 : clearLagTimes = true;
2471 9404 : fullyAppliedLastTime = true;
2472 : }
2473 : else
2474 237962 : fullyAppliedLastTime = false;
2475 :
2476 : /* Send a reply if the standby requested one. */
2477 247366 : if (replyRequested)
2478 0 : WalSndKeepalive(false, InvalidXLogRecPtr);
2479 :
2480 : /*
2481 : * Update shared state for this WalSender process based on reply data from
2482 : * standby.
2483 : */
2484 : {
2485 247366 : WalSnd *walsnd = MyWalSnd;
2486 :
2487 247366 : SpinLockAcquire(&walsnd->mutex);
2488 247366 : walsnd->write = writePtr;
2489 247366 : walsnd->flush = flushPtr;
2490 247366 : walsnd->apply = applyPtr;
2491 247366 : if (writeLag != -1 || clearLagTimes)
2492 124556 : walsnd->writeLag = writeLag;
2493 247366 : if (flushLag != -1 || clearLagTimes)
2494 141768 : walsnd->flushLag = flushLag;
2495 247366 : if (applyLag != -1 || clearLagTimes)
2496 144468 : walsnd->applyLag = applyLag;
2497 247366 : walsnd->replyTime = replyTime;
2498 247366 : SpinLockRelease(&walsnd->mutex);
2499 : }
2500 :
2501 247366 : if (!am_cascading_walsender)
2502 246678 : SyncRepReleaseWaiters();
2503 :
2504 : /*
2505 : * Advance our local xmin horizon when the client confirmed a flush.
2506 : */
2507 247366 : if (MyReplicationSlot && flushPtr != InvalidXLogRecPtr)
2508 : {
2509 237586 : if (SlotIsLogical(MyReplicationSlot))
2510 104260 : LogicalConfirmReceivedLocation(flushPtr);
2511 : else
2512 133326 : PhysicalConfirmReceivedLocation(flushPtr);
2513 : }
2514 247366 : }
2515 :
2516 : /* compute new replication slot xmin horizon if needed */
2517 : static void
2518 122 : PhysicalReplicationSlotNewXmin(TransactionId feedbackXmin, TransactionId feedbackCatalogXmin)
2519 : {
2520 122 : bool changed = false;
2521 122 : ReplicationSlot *slot = MyReplicationSlot;
2522 :
2523 122 : SpinLockAcquire(&slot->mutex);
2524 122 : MyProc->xmin = InvalidTransactionId;
2525 :
2526 : /*
2527 : * For physical replication we don't need the interlock provided by xmin
2528 : * and effective_xmin since the consequences of a missed increase are
2529 : * limited to query cancellations, so set both at once.
2530 : */
2531 122 : if (!TransactionIdIsNormal(slot->data.xmin) ||
2532 60 : !TransactionIdIsNormal(feedbackXmin) ||
2533 60 : TransactionIdPrecedes(slot->data.xmin, feedbackXmin))
2534 : {
2535 80 : changed = true;
2536 80 : slot->data.xmin = feedbackXmin;
2537 80 : slot->effective_xmin = feedbackXmin;
2538 : }
2539 122 : if (!TransactionIdIsNormal(slot->data.catalog_xmin) ||
2540 30 : !TransactionIdIsNormal(feedbackCatalogXmin) ||
2541 30 : TransactionIdPrecedes(slot->data.catalog_xmin, feedbackCatalogXmin))
2542 : {
2543 94 : changed = true;
2544 94 : slot->data.catalog_xmin = feedbackCatalogXmin;
2545 94 : slot->effective_catalog_xmin = feedbackCatalogXmin;
2546 : }
2547 122 : SpinLockRelease(&slot->mutex);
2548 :
2549 122 : if (changed)
2550 : {
2551 100 : ReplicationSlotMarkDirty();
2552 100 : ReplicationSlotsComputeRequiredXmin(false);
2553 : }
2554 122 : }
2555 :
2556 : /*
2557 : * Check that the provided xmin/epoch are sane, that is, not in the future
2558 : * and not so far back as to be already wrapped around.
2559 : *
2560 : * Epoch of nextXid should be same as standby, or if the counter has
2561 : * wrapped, then one greater than standby.
2562 : *
2563 : * This check doesn't care about whether clog exists for these xids
2564 : * at all.
2565 : */
2566 : static bool
2567 130 : TransactionIdInRecentPast(TransactionId xid, uint32 epoch)
2568 : {
2569 : FullTransactionId nextFullXid;
2570 : TransactionId nextXid;
2571 : uint32 nextEpoch;
2572 :
2573 130 : nextFullXid = ReadNextFullTransactionId();
2574 130 : nextXid = XidFromFullTransactionId(nextFullXid);
2575 130 : nextEpoch = EpochFromFullTransactionId(nextFullXid);
2576 :
2577 130 : if (xid <= nextXid)
2578 : {
2579 130 : if (epoch != nextEpoch)
2580 0 : return false;
2581 : }
2582 : else
2583 : {
2584 0 : if (epoch + 1 != nextEpoch)
2585 0 : return false;
2586 : }
2587 :
2588 130 : if (!TransactionIdPrecedesOrEquals(xid, nextXid))
2589 0 : return false; /* epoch OK, but it's wrapped around */
2590 :
2591 130 : return true;
2592 : }
2593 :
2594 : /*
2595 : * Hot Standby feedback
2596 : */
2597 : static void
2598 272 : ProcessStandbyHSFeedbackMessage(void)
2599 : {
2600 : TransactionId feedbackXmin;
2601 : uint32 feedbackEpoch;
2602 : TransactionId feedbackCatalogXmin;
2603 : uint32 feedbackCatalogEpoch;
2604 : TimestampTz replyTime;
2605 :
2606 : /*
2607 : * Decipher the reply message. The caller already consumed the msgtype
2608 : * byte. See XLogWalRcvSendHSFeedback() in walreceiver.c for the creation
2609 : * of this message.
2610 : */
2611 272 : replyTime = pq_getmsgint64(&reply_message);
2612 272 : feedbackXmin = pq_getmsgint(&reply_message, 4);
2613 272 : feedbackEpoch = pq_getmsgint(&reply_message, 4);
2614 272 : feedbackCatalogXmin = pq_getmsgint(&reply_message, 4);
2615 272 : feedbackCatalogEpoch = pq_getmsgint(&reply_message, 4);
2616 :
2617 272 : if (message_level_is_interesting(DEBUG2))
2618 : {
2619 : char *replyTimeStr;
2620 :
2621 : /* Copy because timestamptz_to_str returns a static buffer */
2622 8 : replyTimeStr = pstrdup(timestamptz_to_str(replyTime));
2623 :
2624 8 : elog(DEBUG2, "hot standby feedback xmin %u epoch %u, catalog_xmin %u epoch %u reply_time %s",
2625 : feedbackXmin,
2626 : feedbackEpoch,
2627 : feedbackCatalogXmin,
2628 : feedbackCatalogEpoch,
2629 : replyTimeStr);
2630 :
2631 8 : pfree(replyTimeStr);
2632 : }
2633 :
2634 : /*
2635 : * Update shared state for this WalSender process based on reply data from
2636 : * standby.
2637 : */
2638 : {
2639 272 : WalSnd *walsnd = MyWalSnd;
2640 :
2641 272 : SpinLockAcquire(&walsnd->mutex);
2642 272 : walsnd->replyTime = replyTime;
2643 272 : SpinLockRelease(&walsnd->mutex);
2644 : }
2645 :
2646 : /*
2647 : * Unset WalSender's xmins if the feedback message values are invalid.
2648 : * This happens when the downstream turned hot_standby_feedback off.
2649 : */
2650 272 : if (!TransactionIdIsNormal(feedbackXmin)
2651 184 : && !TransactionIdIsNormal(feedbackCatalogXmin))
2652 : {
2653 184 : MyProc->xmin = InvalidTransactionId;
2654 184 : if (MyReplicationSlot != NULL)
2655 42 : PhysicalReplicationSlotNewXmin(feedbackXmin, feedbackCatalogXmin);
2656 184 : return;
2657 : }
2658 :
2659 : /*
2660 : * Check that the provided xmin/epoch are sane, that is, not in the future
2661 : * and not so far back as to be already wrapped around. Ignore if not.
2662 : */
2663 88 : if (TransactionIdIsNormal(feedbackXmin) &&
2664 88 : !TransactionIdInRecentPast(feedbackXmin, feedbackEpoch))
2665 0 : return;
2666 :
2667 88 : if (TransactionIdIsNormal(feedbackCatalogXmin) &&
2668 42 : !TransactionIdInRecentPast(feedbackCatalogXmin, feedbackCatalogEpoch))
2669 0 : return;
2670 :
2671 : /*
2672 : * Set the WalSender's xmin equal to the standby's requested xmin, so that
2673 : * the xmin will be taken into account by GetSnapshotData() /
2674 : * ComputeXidHorizons(). This will hold back the removal of dead rows and
2675 : * thereby prevent the generation of cleanup conflicts on the standby
2676 : * server.
2677 : *
2678 : * There is a small window for a race condition here: although we just
2679 : * checked that feedbackXmin precedes nextXid, the nextXid could have
2680 : * gotten advanced between our fetching it and applying the xmin below,
2681 : * perhaps far enough to make feedbackXmin wrap around. In that case the
2682 : * xmin we set here would be "in the future" and have no effect. No point
2683 : * in worrying about this since it's too late to save the desired data
2684 : * anyway. Assuming that the standby sends us an increasing sequence of
2685 : * xmins, this could only happen during the first reply cycle, else our
2686 : * own xmin would prevent nextXid from advancing so far.
2687 : *
2688 : * We don't bother taking the ProcArrayLock here. Setting the xmin field
2689 : * is assumed atomic, and there's no real need to prevent concurrent
2690 : * horizon determinations. (If we're moving our xmin forward, this is
2691 : * obviously safe, and if we're moving it backwards, well, the data is at
2692 : * risk already since a VACUUM could already have determined the horizon.)
2693 : *
2694 : * If we're using a replication slot we reserve the xmin via that,
2695 : * otherwise via the walsender's PGPROC entry. We can only track the
2696 : * catalog xmin separately when using a slot, so we store the least of the
2697 : * two provided when not using a slot.
2698 : *
2699 : * XXX: It might make sense to generalize the ephemeral slot concept and
2700 : * always use the slot mechanism to handle the feedback xmin.
2701 : */
2702 88 : if (MyReplicationSlot != NULL) /* XXX: persistency configurable? */
2703 80 : PhysicalReplicationSlotNewXmin(feedbackXmin, feedbackCatalogXmin);
2704 : else
2705 : {
2706 8 : if (TransactionIdIsNormal(feedbackCatalogXmin)
2707 0 : && TransactionIdPrecedes(feedbackCatalogXmin, feedbackXmin))
2708 0 : MyProc->xmin = feedbackCatalogXmin;
2709 : else
2710 8 : MyProc->xmin = feedbackXmin;
2711 : }
2712 : }
2713 :
2714 : /*
2715 : * Process the request for a primary status update message.
2716 : */
2717 : static void
2718 2664 : ProcessStandbyPSRequestMessage(void)
2719 : {
2720 2664 : XLogRecPtr lsn = InvalidXLogRecPtr;
2721 : TransactionId oldestXidInCommit;
2722 : FullTransactionId nextFullXid;
2723 : FullTransactionId fullOldestXidInCommit;
2724 2664 : WalSnd *walsnd = MyWalSnd;
2725 : TimestampTz replyTime;
2726 :
2727 : /*
2728 : * This shouldn't happen because we don't support getting primary status
2729 : * message from standby.
2730 : */
2731 2664 : if (RecoveryInProgress())
2732 0 : elog(ERROR, "the primary status is unavailable during recovery");
2733 :
2734 2664 : replyTime = pq_getmsgint64(&reply_message);
2735 :
2736 : /*
2737 : * Update shared state for this WalSender process based on reply data from
2738 : * standby.
2739 : */
2740 2664 : SpinLockAcquire(&walsnd->mutex);
2741 2664 : walsnd->replyTime = replyTime;
2742 2664 : SpinLockRelease(&walsnd->mutex);
2743 :
2744 : /*
2745 : * Consider transactions in the current database, as only these are the
2746 : * ones replicated.
2747 : */
2748 2664 : oldestXidInCommit = GetOldestActiveTransactionId(true, false);
2749 2664 : nextFullXid = ReadNextFullTransactionId();
2750 2664 : fullOldestXidInCommit = FullTransactionIdFromAllowableAt(nextFullXid,
2751 : oldestXidInCommit);
2752 2664 : lsn = GetXLogWriteRecPtr();
2753 :
2754 2664 : elog(DEBUG2, "sending primary status");
2755 :
2756 : /* construct the message... */
2757 2664 : resetStringInfo(&output_message);
2758 2664 : pq_sendbyte(&output_message, PqReplMsg_PrimaryStatusUpdate);
2759 2664 : pq_sendint64(&output_message, lsn);
2760 2664 : pq_sendint64(&output_message, (int64) U64FromFullTransactionId(fullOldestXidInCommit));
2761 2664 : pq_sendint64(&output_message, (int64) U64FromFullTransactionId(nextFullXid));
2762 2664 : pq_sendint64(&output_message, GetCurrentTimestamp());
2763 :
2764 : /* ... and send it wrapped in CopyData */
2765 2664 : pq_putmessage_noblock(PqMsg_CopyData, output_message.data, output_message.len);
2766 2664 : }
2767 :
2768 : /*
2769 : * Compute how long send/receive loops should sleep.
2770 : *
2771 : * If wal_sender_timeout is enabled we want to wake up in time to send
2772 : * keepalives and to abort the connection if wal_sender_timeout has been
2773 : * reached.
2774 : */
2775 : static long
2776 126342 : WalSndComputeSleeptime(TimestampTz now)
2777 : {
2778 126342 : long sleeptime = 10000; /* 10 s */
2779 :
2780 126342 : if (wal_sender_timeout > 0 && last_reply_timestamp > 0)
2781 : {
2782 : TimestampTz wakeup_time;
2783 :
2784 : /*
2785 : * At the latest stop sleeping once wal_sender_timeout has been
2786 : * reached.
2787 : */
2788 126292 : wakeup_time = TimestampTzPlusMilliseconds(last_reply_timestamp,
2789 : wal_sender_timeout);
2790 :
2791 : /*
2792 : * If no ping has been sent yet, wakeup when it's time to do so.
2793 : * WalSndKeepaliveIfNecessary() wants to send a keepalive once half of
2794 : * the timeout passed without a response.
2795 : */
2796 126292 : if (!waiting_for_ping_response)
2797 115426 : wakeup_time = TimestampTzPlusMilliseconds(last_reply_timestamp,
2798 : wal_sender_timeout / 2);
2799 :
2800 : /* Compute relative time until wakeup. */
2801 126292 : sleeptime = TimestampDifferenceMilliseconds(now, wakeup_time);
2802 : }
2803 :
2804 126342 : return sleeptime;
2805 : }
2806 :
2807 : /*
2808 : * Check whether there have been responses by the client within
2809 : * wal_sender_timeout and shutdown if not. Using last_processing as the
2810 : * reference point avoids counting server-side stalls against the client.
2811 : * However, a long server-side stall can make WalSndKeepaliveIfNecessary()
2812 : * postdate last_processing by more than wal_sender_timeout. If that happens,
2813 : * the client must reply almost immediately to avoid a timeout. This rarely
2814 : * affects the default configuration, under which clients spontaneously send a
2815 : * message every standby_message_timeout = wal_sender_timeout/6 = 10s. We
2816 : * could eliminate that problem by recognizing timeout expiration at
2817 : * wal_sender_timeout/2 after the keepalive.
2818 : */
2819 : static void
2820 1517656 : WalSndCheckTimeOut(void)
2821 : {
2822 : TimestampTz timeout;
2823 :
2824 : /* don't bail out if we're doing something that doesn't require timeouts */
2825 1517656 : if (last_reply_timestamp <= 0)
2826 50 : return;
2827 :
2828 1517606 : timeout = TimestampTzPlusMilliseconds(last_reply_timestamp,
2829 : wal_sender_timeout);
2830 :
2831 1517606 : if (wal_sender_timeout > 0 && last_processing >= timeout)
2832 : {
2833 : /*
2834 : * Since typically expiration of replication timeout means
2835 : * communication problem, we don't send the error message to the
2836 : * standby.
2837 : */
2838 0 : ereport(COMMERROR,
2839 : (errmsg("terminating walsender process due to replication timeout")));
2840 :
2841 0 : WalSndShutdown();
2842 : }
2843 : }
2844 :
2845 : /* Main loop of walsender process that streams the WAL over Copy messages. */
2846 : static void
2847 1322 : WalSndLoop(WalSndSendDataCallback send_data)
2848 : {
2849 1322 : TimestampTz last_flush = 0;
2850 :
2851 : /*
2852 : * Initialize the last reply timestamp. That enables timeout processing
2853 : * from hereon.
2854 : */
2855 1322 : last_reply_timestamp = GetCurrentTimestamp();
2856 1322 : waiting_for_ping_response = false;
2857 :
2858 : /*
2859 : * Loop until we reach the end of this timeline or the client requests to
2860 : * stop streaming.
2861 : */
2862 : for (;;)
2863 : {
2864 : /* Clear any already-pending wakeups */
2865 1509342 : ResetLatch(MyLatch);
2866 :
2867 1509342 : CHECK_FOR_INTERRUPTS();
2868 :
2869 : /* Process any requests or signals received recently */
2870 1509334 : if (ConfigReloadPending)
2871 : {
2872 34 : ConfigReloadPending = false;
2873 34 : ProcessConfigFile(PGC_SIGHUP);
2874 34 : SyncRepInitConfig();
2875 : }
2876 :
2877 : /* Check for input from the client */
2878 1509334 : ProcessRepliesIfAny();
2879 :
2880 : /*
2881 : * If we have received CopyDone from the client, sent CopyDone
2882 : * ourselves, and the output buffer is empty, it's time to exit
2883 : * streaming.
2884 : */
2885 1509170 : if (streamingDoneReceiving && streamingDoneSending &&
2886 1198 : !pq_is_send_pending())
2887 676 : break;
2888 :
2889 : /*
2890 : * If we don't have any pending data in the output buffer, try to send
2891 : * some more. If there is some, we don't bother to call send_data
2892 : * again until we've flushed it ... but we'd better assume we are not
2893 : * caught up.
2894 : */
2895 1508494 : if (!pq_is_send_pending())
2896 1426112 : send_data();
2897 : else
2898 82382 : WalSndCaughtUp = false;
2899 :
2900 : /* Try to flush pending output to the client */
2901 1508092 : if (pq_flush_if_writable() != 0)
2902 0 : WalSndShutdown();
2903 :
2904 : /* If nothing remains to be sent right now ... */
2905 1508092 : if (WalSndCaughtUp && !pq_is_send_pending())
2906 : {
2907 : /*
2908 : * If we're in catchup state, move to streaming. This is an
2909 : * important state change for users to know about, since before
2910 : * this point data loss might occur if the primary dies and we
2911 : * need to failover to the standby. The state change is also
2912 : * important for synchronous replication, since commits that
2913 : * started to wait at that point might wait for some time.
2914 : */
2915 96880 : if (MyWalSnd->state == WALSNDSTATE_CATCHUP)
2916 : {
2917 1024 : ereport(DEBUG1,
2918 : (errmsg_internal("\"%s\" has now caught up with upstream server",
2919 : application_name)));
2920 1024 : WalSndSetState(WALSNDSTATE_STREAMING);
2921 : }
2922 :
2923 : /*
2924 : * When SIGUSR2 arrives, we send any outstanding logs up to the
2925 : * shutdown checkpoint record (i.e., the latest record), wait for
2926 : * them to be replicated to the standby, and exit. This may be a
2927 : * normal termination at shutdown, or a promotion, the walsender
2928 : * is not sure which.
2929 : */
2930 96880 : if (got_SIGUSR2)
2931 12356 : WalSndDone(send_data);
2932 : }
2933 :
2934 : /* Check for replication timeout. */
2935 1508020 : WalSndCheckTimeOut();
2936 :
2937 : /* Send keepalive if the time has come */
2938 1508020 : WalSndKeepaliveIfNecessary();
2939 :
2940 : /*
2941 : * Block if we have unsent data. XXX For logical replication, let
2942 : * WalSndWaitForWal() handle any other blocking; idle receivers need
2943 : * its additional actions. For physical replication, also block if
2944 : * caught up; its send_data does not block.
2945 : *
2946 : * The IO statistics are reported in WalSndWaitForWal() for the
2947 : * logical WAL senders.
2948 : */
2949 1508020 : if ((WalSndCaughtUp && send_data != XLogSendLogical &&
2950 1556914 : !streamingDoneSending) ||
2951 1467182 : pq_is_send_pending())
2952 : {
2953 : long sleeptime;
2954 : int wakeEvents;
2955 : TimestampTz now;
2956 :
2957 117382 : if (!streamingDoneReceiving)
2958 117364 : wakeEvents = WL_SOCKET_READABLE;
2959 : else
2960 18 : wakeEvents = 0;
2961 :
2962 : /*
2963 : * Use fresh timestamp, not last_processing, to reduce the chance
2964 : * of reaching wal_sender_timeout before sending a keepalive.
2965 : */
2966 117382 : now = GetCurrentTimestamp();
2967 117382 : sleeptime = WalSndComputeSleeptime(now);
2968 :
2969 117382 : if (pq_is_send_pending())
2970 82052 : wakeEvents |= WL_SOCKET_WRITEABLE;
2971 :
2972 : /* Report IO statistics, if needed */
2973 117382 : if (TimestampDifferenceExceeds(last_flush, now,
2974 : WALSENDER_STATS_FLUSH_INTERVAL))
2975 : {
2976 1246 : pgstat_flush_io(false);
2977 1246 : (void) pgstat_flush_backend(false, PGSTAT_BACKEND_FLUSH_IO);
2978 1246 : last_flush = now;
2979 : }
2980 :
2981 : /* Sleep until something happens or we time out */
2982 117382 : WalSndWait(wakeEvents, sleeptime, WAIT_EVENT_WAL_SENDER_MAIN);
2983 : }
2984 : }
2985 676 : }
2986 :
2987 : /* Initialize a per-walsender data structure for this walsender process */
2988 : static void
2989 2282 : InitWalSenderSlot(void)
2990 : {
2991 : int i;
2992 :
2993 : /*
2994 : * WalSndCtl should be set up already (we inherit this by fork() or
2995 : * EXEC_BACKEND mechanism from the postmaster).
2996 : */
2997 : Assert(WalSndCtl != NULL);
2998 : Assert(MyWalSnd == NULL);
2999 :
3000 : /*
3001 : * Find a free walsender slot and reserve it. This must not fail due to
3002 : * the prior check for free WAL senders in InitProcess().
3003 : */
3004 3338 : for (i = 0; i < max_wal_senders; i++)
3005 : {
3006 3338 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3007 :
3008 3338 : SpinLockAcquire(&walsnd->mutex);
3009 :
3010 3338 : if (walsnd->pid != 0)
3011 : {
3012 1056 : SpinLockRelease(&walsnd->mutex);
3013 1056 : continue;
3014 : }
3015 : else
3016 : {
3017 : /*
3018 : * Found a free slot. Reserve it for us.
3019 : */
3020 2282 : walsnd->pid = MyProcPid;
3021 2282 : walsnd->state = WALSNDSTATE_STARTUP;
3022 2282 : walsnd->sentPtr = InvalidXLogRecPtr;
3023 2282 : walsnd->needreload = false;
3024 2282 : walsnd->write = InvalidXLogRecPtr;
3025 2282 : walsnd->flush = InvalidXLogRecPtr;
3026 2282 : walsnd->apply = InvalidXLogRecPtr;
3027 2282 : walsnd->writeLag = -1;
3028 2282 : walsnd->flushLag = -1;
3029 2282 : walsnd->applyLag = -1;
3030 2282 : walsnd->sync_standby_priority = 0;
3031 2282 : walsnd->replyTime = 0;
3032 :
3033 : /*
3034 : * The kind assignment is done here and not in StartReplication()
3035 : * and StartLogicalReplication(). Indeed, the logical walsender
3036 : * needs to read WAL records (like snapshot of running
3037 : * transactions) during the slot creation. So it needs to be woken
3038 : * up based on its kind.
3039 : *
3040 : * The kind assignment could also be done in StartReplication(),
3041 : * StartLogicalReplication() and CREATE_REPLICATION_SLOT but it
3042 : * seems better to set it on one place.
3043 : */
3044 2282 : if (MyDatabaseId == InvalidOid)
3045 906 : walsnd->kind = REPLICATION_KIND_PHYSICAL;
3046 : else
3047 1376 : walsnd->kind = REPLICATION_KIND_LOGICAL;
3048 :
3049 2282 : SpinLockRelease(&walsnd->mutex);
3050 : /* don't need the lock anymore */
3051 2282 : MyWalSnd = (WalSnd *) walsnd;
3052 :
3053 2282 : break;
3054 : }
3055 : }
3056 :
3057 : Assert(MyWalSnd != NULL);
3058 :
3059 : /* Arrange to clean up at walsender exit */
3060 2282 : on_shmem_exit(WalSndKill, 0);
3061 2282 : }
3062 :
3063 : /* Destroy the per-walsender data structure for this walsender process */
3064 : static void
3065 2282 : WalSndKill(int code, Datum arg)
3066 : {
3067 2282 : WalSnd *walsnd = MyWalSnd;
3068 :
3069 : Assert(walsnd != NULL);
3070 :
3071 2282 : MyWalSnd = NULL;
3072 :
3073 2282 : SpinLockAcquire(&walsnd->mutex);
3074 : /* Mark WalSnd struct as no longer being in use. */
3075 2282 : walsnd->pid = 0;
3076 2282 : SpinLockRelease(&walsnd->mutex);
3077 2282 : }
3078 :
3079 : /* XLogReaderRoutine->segment_open callback */
3080 : static void
3081 10122 : WalSndSegmentOpen(XLogReaderState *state, XLogSegNo nextSegNo,
3082 : TimeLineID *tli_p)
3083 : {
3084 : char path[MAXPGPATH];
3085 :
3086 : /*-------
3087 : * When reading from a historic timeline, and there is a timeline switch
3088 : * within this segment, read from the WAL segment belonging to the new
3089 : * timeline.
3090 : *
3091 : * For example, imagine that this server is currently on timeline 5, and
3092 : * we're streaming timeline 4. The switch from timeline 4 to 5 happened at
3093 : * 0/13002088. In pg_wal, we have these files:
3094 : *
3095 : * ...
3096 : * 000000040000000000000012
3097 : * 000000040000000000000013
3098 : * 000000050000000000000013
3099 : * 000000050000000000000014
3100 : * ...
3101 : *
3102 : * In this situation, when requested to send the WAL from segment 0x13, on
3103 : * timeline 4, we read the WAL from file 000000050000000000000013. Archive
3104 : * recovery prefers files from newer timelines, so if the segment was
3105 : * restored from the archive on this server, the file belonging to the old
3106 : * timeline, 000000040000000000000013, might not exist. Their contents are
3107 : * equal up to the switchpoint, because at a timeline switch, the used
3108 : * portion of the old segment is copied to the new file.
3109 : */
3110 10122 : *tli_p = sendTimeLine;
3111 10122 : if (sendTimeLineIsHistoric)
3112 : {
3113 : XLogSegNo endSegNo;
3114 :
3115 24 : XLByteToSeg(sendTimeLineValidUpto, endSegNo, state->segcxt.ws_segsize);
3116 24 : if (nextSegNo == endSegNo)
3117 20 : *tli_p = sendTimeLineNextTLI;
3118 : }
3119 :
3120 10122 : XLogFilePath(path, *tli_p, nextSegNo, state->segcxt.ws_segsize);
3121 10122 : state->seg.ws_file = BasicOpenFile(path, O_RDONLY | PG_BINARY);
3122 10122 : if (state->seg.ws_file >= 0)
3123 10122 : return;
3124 :
3125 : /*
3126 : * If the file is not found, assume it's because the standby asked for a
3127 : * too old WAL segment that has already been removed or recycled.
3128 : */
3129 0 : if (errno == ENOENT)
3130 : {
3131 : char xlogfname[MAXFNAMELEN];
3132 0 : int save_errno = errno;
3133 :
3134 0 : XLogFileName(xlogfname, *tli_p, nextSegNo, wal_segment_size);
3135 0 : errno = save_errno;
3136 0 : ereport(ERROR,
3137 : (errcode_for_file_access(),
3138 : errmsg("requested WAL segment %s has already been removed",
3139 : xlogfname)));
3140 : }
3141 : else
3142 0 : ereport(ERROR,
3143 : (errcode_for_file_access(),
3144 : errmsg("could not open file \"%s\": %m",
3145 : path)));
3146 : }
3147 :
3148 : /*
3149 : * Send out the WAL in its normal physical/stored form.
3150 : *
3151 : * Read up to MAX_SEND_SIZE bytes of WAL that's been flushed to disk,
3152 : * but not yet sent to the client, and buffer it in the libpq output
3153 : * buffer.
3154 : *
3155 : * If there is no unsent WAL remaining, WalSndCaughtUp is set to true,
3156 : * otherwise WalSndCaughtUp is set to false.
3157 : */
3158 : static void
3159 286480 : XLogSendPhysical(void)
3160 : {
3161 : XLogRecPtr SendRqstPtr;
3162 : XLogRecPtr startptr;
3163 : XLogRecPtr endptr;
3164 : Size nbytes;
3165 : XLogSegNo segno;
3166 : WALReadError errinfo;
3167 : Size rbytes;
3168 :
3169 : /* If requested switch the WAL sender to the stopping state. */
3170 286480 : if (got_STOPPING)
3171 18332 : WalSndSetState(WALSNDSTATE_STOPPING);
3172 :
3173 286480 : if (streamingDoneSending)
3174 : {
3175 48868 : WalSndCaughtUp = true;
3176 86708 : return;
3177 : }
3178 :
3179 : /* Figure out how far we can safely send the WAL. */
3180 237612 : if (sendTimeLineIsHistoric)
3181 : {
3182 : /*
3183 : * Streaming an old timeline that's in this server's history, but is
3184 : * not the one we're currently inserting or replaying. It can be
3185 : * streamed up to the point where we switched off that timeline.
3186 : */
3187 330 : SendRqstPtr = sendTimeLineValidUpto;
3188 : }
3189 237282 : else if (am_cascading_walsender)
3190 : {
3191 : TimeLineID SendRqstTLI;
3192 :
3193 : /*
3194 : * Streaming the latest timeline on a standby.
3195 : *
3196 : * Attempt to send all WAL that has already been replayed, so that we
3197 : * know it's valid. If we're receiving WAL through streaming
3198 : * replication, it's also OK to send any WAL that has been received
3199 : * but not replayed.
3200 : *
3201 : * The timeline we're recovering from can change, or we can be
3202 : * promoted. In either case, the current timeline becomes historic. We
3203 : * need to detect that so that we don't try to stream past the point
3204 : * where we switched to another timeline. We check for promotion or
3205 : * timeline switch after calculating FlushPtr, to avoid a race
3206 : * condition: if the timeline becomes historic just after we checked
3207 : * that it was still current, it's still be OK to stream it up to the
3208 : * FlushPtr that was calculated before it became historic.
3209 : */
3210 1518 : bool becameHistoric = false;
3211 :
3212 1518 : SendRqstPtr = GetStandbyFlushRecPtr(&SendRqstTLI);
3213 :
3214 1518 : if (!RecoveryInProgress())
3215 : {
3216 : /* We have been promoted. */
3217 2 : SendRqstTLI = GetWALInsertionTimeLine();
3218 2 : am_cascading_walsender = false;
3219 2 : becameHistoric = true;
3220 : }
3221 : else
3222 : {
3223 : /*
3224 : * Still a cascading standby. But is the timeline we're sending
3225 : * still the one recovery is recovering from?
3226 : */
3227 1516 : if (sendTimeLine != SendRqstTLI)
3228 0 : becameHistoric = true;
3229 : }
3230 :
3231 1518 : if (becameHistoric)
3232 : {
3233 : /*
3234 : * The timeline we were sending has become historic. Read the
3235 : * timeline history file of the new timeline to see where exactly
3236 : * we forked off from the timeline we were sending.
3237 : */
3238 : List *history;
3239 :
3240 2 : history = readTimeLineHistory(SendRqstTLI);
3241 2 : sendTimeLineValidUpto = tliSwitchPoint(sendTimeLine, history, &sendTimeLineNextTLI);
3242 :
3243 : Assert(sendTimeLine < sendTimeLineNextTLI);
3244 2 : list_free_deep(history);
3245 :
3246 2 : sendTimeLineIsHistoric = true;
3247 :
3248 2 : SendRqstPtr = sendTimeLineValidUpto;
3249 : }
3250 : }
3251 : else
3252 : {
3253 : /*
3254 : * Streaming the current timeline on a primary.
3255 : *
3256 : * Attempt to send all data that's already been written out and
3257 : * fsync'd to disk. We cannot go further than what's been written out
3258 : * given the current implementation of WALRead(). And in any case
3259 : * it's unsafe to send WAL that is not securely down to disk on the
3260 : * primary: if the primary subsequently crashes and restarts, standbys
3261 : * must not have applied any WAL that got lost on the primary.
3262 : */
3263 235764 : SendRqstPtr = GetFlushRecPtr(NULL);
3264 : }
3265 :
3266 : /*
3267 : * Record the current system time as an approximation of the time at which
3268 : * this WAL location was written for the purposes of lag tracking.
3269 : *
3270 : * In theory we could make XLogFlush() record a time in shmem whenever WAL
3271 : * is flushed and we could get that time as well as the LSN when we call
3272 : * GetFlushRecPtr() above (and likewise for the cascading standby
3273 : * equivalent), but rather than putting any new code into the hot WAL path
3274 : * it seems good enough to capture the time here. We should reach this
3275 : * after XLogFlush() runs WalSndWakeupProcessRequests(), and although that
3276 : * may take some time, we read the WAL flush pointer and take the time
3277 : * very close to together here so that we'll get a later position if it is
3278 : * still moving.
3279 : *
3280 : * Because LagTrackerWrite ignores samples when the LSN hasn't advanced,
3281 : * this gives us a cheap approximation for the WAL flush time for this
3282 : * LSN.
3283 : *
3284 : * Note that the LSN is not necessarily the LSN for the data contained in
3285 : * the present message; it's the end of the WAL, which might be further
3286 : * ahead. All the lag tracking machinery cares about is finding out when
3287 : * that arbitrary LSN is eventually reported as written, flushed and
3288 : * applied, so that it can measure the elapsed time.
3289 : */
3290 237612 : LagTrackerWrite(SendRqstPtr, GetCurrentTimestamp());
3291 :
3292 : /*
3293 : * If this is a historic timeline and we've reached the point where we
3294 : * forked to the next timeline, stop streaming.
3295 : *
3296 : * Note: We might already have sent WAL > sendTimeLineValidUpto. The
3297 : * startup process will normally replay all WAL that has been received
3298 : * from the primary, before promoting, but if the WAL streaming is
3299 : * terminated at a WAL page boundary, the valid portion of the timeline
3300 : * might end in the middle of a WAL record. We might've already sent the
3301 : * first half of that partial WAL record to the cascading standby, so that
3302 : * sentPtr > sendTimeLineValidUpto. That's OK; the cascading standby can't
3303 : * replay the partial WAL record either, so it can still follow our
3304 : * timeline switch.
3305 : */
3306 237612 : if (sendTimeLineIsHistoric && sendTimeLineValidUpto <= sentPtr)
3307 : {
3308 : /* close the current file. */
3309 26 : if (xlogreader->seg.ws_file >= 0)
3310 26 : wal_segment_close(xlogreader);
3311 :
3312 : /* Send CopyDone */
3313 26 : pq_putmessage_noblock(PqMsg_CopyDone, NULL, 0);
3314 26 : streamingDoneSending = true;
3315 :
3316 26 : WalSndCaughtUp = true;
3317 :
3318 26 : elog(DEBUG1, "walsender reached end of timeline at %X/%08X (sent up to %X/%08X)",
3319 : LSN_FORMAT_ARGS(sendTimeLineValidUpto),
3320 : LSN_FORMAT_ARGS(sentPtr));
3321 26 : return;
3322 : }
3323 :
3324 : /* Do we have any work to do? */
3325 : Assert(sentPtr <= SendRqstPtr);
3326 237586 : if (SendRqstPtr <= sentPtr)
3327 : {
3328 37814 : WalSndCaughtUp = true;
3329 37814 : return;
3330 : }
3331 :
3332 : /*
3333 : * Figure out how much to send in one message. If there's no more than
3334 : * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
3335 : * MAX_SEND_SIZE bytes, but round back to logfile or page boundary.
3336 : *
3337 : * The rounding is not only for performance reasons. Walreceiver relies on
3338 : * the fact that we never split a WAL record across two messages. Since a
3339 : * long WAL record is split at page boundary into continuation records,
3340 : * page boundary is always a safe cut-off point. We also assume that
3341 : * SendRqstPtr never points to the middle of a WAL record.
3342 : */
3343 199772 : startptr = sentPtr;
3344 199772 : endptr = startptr;
3345 199772 : endptr += MAX_SEND_SIZE;
3346 :
3347 : /* if we went beyond SendRqstPtr, back off */
3348 199772 : if (SendRqstPtr <= endptr)
3349 : {
3350 11952 : endptr = SendRqstPtr;
3351 11952 : if (sendTimeLineIsHistoric)
3352 24 : WalSndCaughtUp = false;
3353 : else
3354 11928 : WalSndCaughtUp = true;
3355 : }
3356 : else
3357 : {
3358 : /* round down to page boundary. */
3359 187820 : endptr -= (endptr % XLOG_BLCKSZ);
3360 187820 : WalSndCaughtUp = false;
3361 : }
3362 :
3363 199772 : nbytes = endptr - startptr;
3364 : Assert(nbytes <= MAX_SEND_SIZE);
3365 :
3366 : /*
3367 : * OK to read and send the slice.
3368 : */
3369 199772 : resetStringInfo(&output_message);
3370 199772 : pq_sendbyte(&output_message, PqReplMsg_WALData);
3371 :
3372 199772 : pq_sendint64(&output_message, startptr); /* dataStart */
3373 199772 : pq_sendint64(&output_message, SendRqstPtr); /* walEnd */
3374 199772 : pq_sendint64(&output_message, 0); /* sendtime, filled in last */
3375 :
3376 : /*
3377 : * Read the log directly into the output buffer to avoid extra memcpy
3378 : * calls.
3379 : */
3380 199772 : enlargeStringInfo(&output_message, nbytes);
3381 :
3382 199772 : retry:
3383 : /* attempt to read WAL from WAL buffers first */
3384 199772 : rbytes = WALReadFromBuffers(&output_message.data[output_message.len],
3385 199772 : startptr, nbytes, xlogreader->seg.ws_tli);
3386 199772 : output_message.len += rbytes;
3387 199772 : startptr += rbytes;
3388 199772 : nbytes -= rbytes;
3389 :
3390 : /* now read the remaining WAL from WAL file */
3391 199772 : if (nbytes > 0 &&
3392 193468 : !WALRead(xlogreader,
3393 193468 : &output_message.data[output_message.len],
3394 : startptr,
3395 : nbytes,
3396 193468 : xlogreader->seg.ws_tli, /* Pass the current TLI because
3397 : * only WalSndSegmentOpen controls
3398 : * whether new TLI is needed. */
3399 : &errinfo))
3400 0 : WALReadRaiseError(&errinfo);
3401 :
3402 : /* See logical_read_xlog_page(). */
3403 199772 : XLByteToSeg(startptr, segno, xlogreader->segcxt.ws_segsize);
3404 199772 : CheckXLogRemoved(segno, xlogreader->seg.ws_tli);
3405 :
3406 : /*
3407 : * During recovery, the currently-open WAL file might be replaced with the
3408 : * file of the same name retrieved from archive. So we always need to
3409 : * check what we read was valid after reading into the buffer. If it's
3410 : * invalid, we try to open and read the file again.
3411 : */
3412 199772 : if (am_cascading_walsender)
3413 : {
3414 1112 : WalSnd *walsnd = MyWalSnd;
3415 : bool reload;
3416 :
3417 1112 : SpinLockAcquire(&walsnd->mutex);
3418 1112 : reload = walsnd->needreload;
3419 1112 : walsnd->needreload = false;
3420 1112 : SpinLockRelease(&walsnd->mutex);
3421 :
3422 1112 : if (reload && xlogreader->seg.ws_file >= 0)
3423 : {
3424 0 : wal_segment_close(xlogreader);
3425 :
3426 0 : goto retry;
3427 : }
3428 : }
3429 :
3430 199772 : output_message.len += nbytes;
3431 199772 : output_message.data[output_message.len] = '\0';
3432 :
3433 : /*
3434 : * Fill the send timestamp last, so that it is taken as late as possible.
3435 : */
3436 199772 : resetStringInfo(&tmpbuf);
3437 199772 : pq_sendint64(&tmpbuf, GetCurrentTimestamp());
3438 199772 : memcpy(&output_message.data[1 + sizeof(int64) + sizeof(int64)],
3439 199772 : tmpbuf.data, sizeof(int64));
3440 :
3441 199772 : pq_putmessage_noblock(PqMsg_CopyData, output_message.data, output_message.len);
3442 :
3443 199772 : sentPtr = endptr;
3444 :
3445 : /* Update shared memory status */
3446 : {
3447 199772 : WalSnd *walsnd = MyWalSnd;
3448 :
3449 199772 : SpinLockAcquire(&walsnd->mutex);
3450 199772 : walsnd->sentPtr = sentPtr;
3451 199772 : SpinLockRelease(&walsnd->mutex);
3452 : }
3453 :
3454 : /* Report progress of XLOG streaming in PS display */
3455 199772 : if (update_process_title)
3456 : {
3457 : char activitymsg[50];
3458 :
3459 199772 : snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%08X",
3460 199772 : LSN_FORMAT_ARGS(sentPtr));
3461 199772 : set_ps_display(activitymsg);
3462 : }
3463 : }
3464 :
3465 : /*
3466 : * Stream out logically decoded data.
3467 : */
3468 : static void
3469 1151988 : XLogSendLogical(void)
3470 : {
3471 : XLogRecord *record;
3472 : char *errm;
3473 :
3474 : /*
3475 : * We'll use the current flush point to determine whether we've caught up.
3476 : * This variable is static in order to cache it across calls. Caching is
3477 : * helpful because GetFlushRecPtr() needs to acquire a heavily-contended
3478 : * spinlock.
3479 : */
3480 : static XLogRecPtr flushPtr = InvalidXLogRecPtr;
3481 :
3482 : /*
3483 : * Don't know whether we've caught up yet. We'll set WalSndCaughtUp to
3484 : * true in WalSndWaitForWal, if we're actually waiting. We also set to
3485 : * true if XLogReadRecord() had to stop reading but WalSndWaitForWal
3486 : * didn't wait - i.e. when we're shutting down.
3487 : */
3488 1151988 : WalSndCaughtUp = false;
3489 :
3490 1151988 : record = XLogReadRecord(logical_decoding_ctx->reader, &errm);
3491 :
3492 : /* xlog record was invalid */
3493 1151646 : if (errm != NULL)
3494 0 : elog(ERROR, "could not find record while sending logically-decoded data: %s",
3495 : errm);
3496 :
3497 1151646 : if (record != NULL)
3498 : {
3499 : /*
3500 : * Note the lack of any call to LagTrackerWrite() which is handled by
3501 : * WalSndUpdateProgress which is called by output plugin through
3502 : * logical decoding write api.
3503 : */
3504 1144492 : LogicalDecodingProcessRecord(logical_decoding_ctx, logical_decoding_ctx->reader);
3505 :
3506 1144432 : sentPtr = logical_decoding_ctx->reader->EndRecPtr;
3507 : }
3508 :
3509 : /*
3510 : * If first time through in this session, initialize flushPtr. Otherwise,
3511 : * we only need to update flushPtr if EndRecPtr is past it.
3512 : */
3513 1151586 : if (flushPtr == InvalidXLogRecPtr ||
3514 1151006 : logical_decoding_ctx->reader->EndRecPtr >= flushPtr)
3515 : {
3516 : /*
3517 : * For cascading logical WAL senders, we use the replay LSN instead of
3518 : * the flush LSN, since logical decoding on a standby only processes
3519 : * WAL that has been replayed. This distinction becomes particularly
3520 : * important during shutdown, as new WAL is no longer replayed and the
3521 : * last replayed LSN marks the furthest point up to which decoding can
3522 : * proceed.
3523 : */
3524 12128 : if (am_cascading_walsender)
3525 1534 : flushPtr = GetXLogReplayRecPtr(NULL);
3526 : else
3527 10594 : flushPtr = GetFlushRecPtr(NULL);
3528 : }
3529 :
3530 : /* If EndRecPtr is still past our flushPtr, it means we caught up. */
3531 1151586 : if (logical_decoding_ctx->reader->EndRecPtr >= flushPtr)
3532 9924 : WalSndCaughtUp = true;
3533 :
3534 : /*
3535 : * If we're caught up and have been requested to stop, have WalSndLoop()
3536 : * terminate the connection in an orderly manner, after writing out all
3537 : * the pending data.
3538 : */
3539 1151586 : if (WalSndCaughtUp && got_STOPPING)
3540 7008 : got_SIGUSR2 = true;
3541 :
3542 : /* Update shared memory status */
3543 : {
3544 1151586 : WalSnd *walsnd = MyWalSnd;
3545 :
3546 1151586 : SpinLockAcquire(&walsnd->mutex);
3547 1151586 : walsnd->sentPtr = sentPtr;
3548 1151586 : SpinLockRelease(&walsnd->mutex);
3549 : }
3550 1151586 : }
3551 :
3552 : /*
3553 : * Shutdown if the sender is caught up.
3554 : *
3555 : * NB: This should only be called when the shutdown signal has been received
3556 : * from postmaster.
3557 : *
3558 : * Note that if we determine that there's still more data to send, this
3559 : * function will return control to the caller.
3560 : */
3561 : static void
3562 12356 : WalSndDone(WalSndSendDataCallback send_data)
3563 : {
3564 : XLogRecPtr replicatedPtr;
3565 :
3566 : /* ... let's just be real sure we're caught up ... */
3567 12356 : send_data();
3568 :
3569 : /*
3570 : * To figure out whether all WAL has successfully been replicated, check
3571 : * flush location if valid, write otherwise. Tools like pg_receivewal will
3572 : * usually (unless in synchronous mode) return an invalid flush location.
3573 : */
3574 24712 : replicatedPtr = XLogRecPtrIsInvalid(MyWalSnd->flush) ?
3575 12356 : MyWalSnd->write : MyWalSnd->flush;
3576 :
3577 12356 : if (WalSndCaughtUp && sentPtr == replicatedPtr &&
3578 72 : !pq_is_send_pending())
3579 : {
3580 : QueryCompletion qc;
3581 :
3582 : /* Inform the standby that XLOG streaming is done */
3583 72 : SetQueryCompletion(&qc, CMDTAG_COPY, 0);
3584 72 : EndCommand(&qc, DestRemote, false);
3585 72 : pq_flush();
3586 :
3587 72 : proc_exit(0);
3588 : }
3589 12284 : if (!waiting_for_ping_response)
3590 6468 : WalSndKeepalive(true, InvalidXLogRecPtr);
3591 12284 : }
3592 :
3593 : /*
3594 : * Returns the latest point in WAL that has been safely flushed to disk.
3595 : * This should only be called when in recovery.
3596 : *
3597 : * This is called either by cascading walsender to find WAL position to be sent
3598 : * to a cascaded standby or by slot synchronization operation to validate remote
3599 : * slot's lsn before syncing it locally.
3600 : *
3601 : * As a side-effect, *tli is updated to the TLI of the last
3602 : * replayed WAL record.
3603 : */
3604 : XLogRecPtr
3605 1734 : GetStandbyFlushRecPtr(TimeLineID *tli)
3606 : {
3607 : XLogRecPtr replayPtr;
3608 : TimeLineID replayTLI;
3609 : XLogRecPtr receivePtr;
3610 : TimeLineID receiveTLI;
3611 : XLogRecPtr result;
3612 :
3613 : Assert(am_cascading_walsender || IsSyncingReplicationSlots());
3614 :
3615 : /*
3616 : * We can safely send what's already been replayed. Also, if walreceiver
3617 : * is streaming WAL from the same timeline, we can send anything that it
3618 : * has streamed, but hasn't been replayed yet.
3619 : */
3620 :
3621 1734 : receivePtr = GetWalRcvFlushRecPtr(NULL, &receiveTLI);
3622 1734 : replayPtr = GetXLogReplayRecPtr(&replayTLI);
3623 :
3624 1734 : if (tli)
3625 1660 : *tli = replayTLI;
3626 :
3627 1734 : result = replayPtr;
3628 1734 : if (receiveTLI == replayTLI && receivePtr > replayPtr)
3629 62 : result = receivePtr;
3630 :
3631 1734 : return result;
3632 : }
3633 :
3634 : /*
3635 : * Request walsenders to reload the currently-open WAL file
3636 : */
3637 : void
3638 58 : WalSndRqstFileReload(void)
3639 : {
3640 : int i;
3641 :
3642 590 : for (i = 0; i < max_wal_senders; i++)
3643 : {
3644 532 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3645 :
3646 532 : SpinLockAcquire(&walsnd->mutex);
3647 532 : if (walsnd->pid == 0)
3648 : {
3649 532 : SpinLockRelease(&walsnd->mutex);
3650 532 : continue;
3651 : }
3652 0 : walsnd->needreload = true;
3653 0 : SpinLockRelease(&walsnd->mutex);
3654 : }
3655 58 : }
3656 :
3657 : /*
3658 : * Handle PROCSIG_WALSND_INIT_STOPPING signal.
3659 : */
3660 : void
3661 72 : HandleWalSndInitStopping(void)
3662 : {
3663 : Assert(am_walsender);
3664 :
3665 : /*
3666 : * If replication has not yet started, die like with SIGTERM. If
3667 : * replication is active, only set a flag and wake up the main loop. It
3668 : * will send any outstanding WAL, wait for it to be replicated to the
3669 : * standby, and then exit gracefully.
3670 : */
3671 72 : if (!replication_active)
3672 0 : kill(MyProcPid, SIGTERM);
3673 : else
3674 72 : got_STOPPING = true;
3675 72 : }
3676 :
3677 : /*
3678 : * SIGUSR2: set flag to do a last cycle and shut down afterwards. The WAL
3679 : * sender should already have been switched to WALSNDSTATE_STOPPING at
3680 : * this point.
3681 : */
3682 : static void
3683 66 : WalSndLastCycleHandler(SIGNAL_ARGS)
3684 : {
3685 66 : got_SIGUSR2 = true;
3686 66 : SetLatch(MyLatch);
3687 66 : }
3688 :
3689 : /* Set up signal handlers */
3690 : void
3691 2282 : WalSndSignals(void)
3692 : {
3693 : /* Set up signal handlers */
3694 2282 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
3695 2282 : pqsignal(SIGINT, StatementCancelHandler); /* query cancel */
3696 2282 : pqsignal(SIGTERM, die); /* request shutdown */
3697 : /* SIGQUIT handler was already set up by InitPostmasterChild */
3698 2282 : InitializeTimeouts(); /* establishes SIGALRM handler */
3699 2282 : pqsignal(SIGPIPE, SIG_IGN);
3700 2282 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
3701 2282 : pqsignal(SIGUSR2, WalSndLastCycleHandler); /* request a last cycle and
3702 : * shutdown */
3703 :
3704 : /* Reset some signals that are accepted by postmaster but not here */
3705 2282 : pqsignal(SIGCHLD, SIG_DFL);
3706 2282 : }
3707 :
3708 : /* Report shared-memory space needed by WalSndShmemInit */
3709 : Size
3710 8308 : WalSndShmemSize(void)
3711 : {
3712 8308 : Size size = 0;
3713 :
3714 8308 : size = offsetof(WalSndCtlData, walsnds);
3715 8308 : size = add_size(size, mul_size(max_wal_senders, sizeof(WalSnd)));
3716 :
3717 8308 : return size;
3718 : }
3719 :
3720 : /* Allocate and initialize walsender-related shared memory */
3721 : void
3722 2152 : WalSndShmemInit(void)
3723 : {
3724 : bool found;
3725 : int i;
3726 :
3727 2152 : WalSndCtl = (WalSndCtlData *)
3728 2152 : ShmemInitStruct("Wal Sender Ctl", WalSndShmemSize(), &found);
3729 :
3730 2152 : if (!found)
3731 : {
3732 : /* First time through, so initialize */
3733 15272 : MemSet(WalSndCtl, 0, WalSndShmemSize());
3734 :
3735 8608 : for (i = 0; i < NUM_SYNC_REP_WAIT_MODE; i++)
3736 6456 : dlist_init(&(WalSndCtl->SyncRepQueue[i]));
3737 :
3738 16368 : for (i = 0; i < max_wal_senders; i++)
3739 : {
3740 14216 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3741 :
3742 14216 : SpinLockInit(&walsnd->mutex);
3743 : }
3744 :
3745 2152 : ConditionVariableInit(&WalSndCtl->wal_flush_cv);
3746 2152 : ConditionVariableInit(&WalSndCtl->wal_replay_cv);
3747 2152 : ConditionVariableInit(&WalSndCtl->wal_confirm_rcv_cv);
3748 : }
3749 2152 : }
3750 :
3751 : /*
3752 : * Wake up physical, logical or both kinds of walsenders
3753 : *
3754 : * The distinction between physical and logical walsenders is done, because:
3755 : * - physical walsenders can't send data until it's been flushed
3756 : * - logical walsenders on standby can't decode and send data until it's been
3757 : * applied
3758 : *
3759 : * For cascading replication we need to wake up physical walsenders separately
3760 : * from logical walsenders (see the comment before calling WalSndWakeup() in
3761 : * ApplyWalRecord() for more details).
3762 : *
3763 : * This will be called inside critical sections, so throwing an error is not
3764 : * advisable.
3765 : */
3766 : void
3767 5381358 : WalSndWakeup(bool physical, bool logical)
3768 : {
3769 : /*
3770 : * Wake up all the walsenders waiting on WAL being flushed or replayed
3771 : * respectively. Note that waiting walsender would have prepared to sleep
3772 : * on the CV (i.e., added itself to the CV's waitlist) in WalSndWait()
3773 : * before actually waiting.
3774 : */
3775 5381358 : if (physical)
3776 268738 : ConditionVariableBroadcast(&WalSndCtl->wal_flush_cv);
3777 :
3778 5381358 : if (logical)
3779 5318136 : ConditionVariableBroadcast(&WalSndCtl->wal_replay_cv);
3780 5381358 : }
3781 :
3782 : /*
3783 : * Wait for readiness on the FeBe socket, or a timeout. The mask should be
3784 : * composed of optional WL_SOCKET_WRITEABLE and WL_SOCKET_READABLE flags. Exit
3785 : * on postmaster death.
3786 : */
3787 : static void
3788 126342 : WalSndWait(uint32 socket_events, long timeout, uint32 wait_event)
3789 : {
3790 : WaitEvent event;
3791 :
3792 126342 : ModifyWaitEvent(FeBeWaitSet, FeBeWaitSetSocketPos, socket_events, NULL);
3793 :
3794 : /*
3795 : * We use a condition variable to efficiently wake up walsenders in
3796 : * WalSndWakeup().
3797 : *
3798 : * Every walsender prepares to sleep on a shared memory CV. Note that it
3799 : * just prepares to sleep on the CV (i.e., adds itself to the CV's
3800 : * waitlist), but does not actually wait on the CV (IOW, it never calls
3801 : * ConditionVariableSleep()). It still uses WaitEventSetWait() for
3802 : * waiting, because we also need to wait for socket events. The processes
3803 : * (startup process, walreceiver etc.) wanting to wake up walsenders use
3804 : * ConditionVariableBroadcast(), which in turn calls SetLatch(), helping
3805 : * walsenders come out of WaitEventSetWait().
3806 : *
3807 : * This approach is simple and efficient because, one doesn't have to loop
3808 : * through all the walsenders slots, with a spinlock acquisition and
3809 : * release for every iteration, just to wake up only the waiting
3810 : * walsenders. It makes WalSndWakeup() callers' life easy.
3811 : *
3812 : * XXX: A desirable future improvement would be to add support for CVs
3813 : * into WaitEventSetWait().
3814 : *
3815 : * And, we use separate shared memory CVs for physical and logical
3816 : * walsenders for selective wake ups, see WalSndWakeup() for more details.
3817 : *
3818 : * If the wait event is WAIT_FOR_STANDBY_CONFIRMATION, wait on another CV
3819 : * until awakened by physical walsenders after the walreceiver confirms
3820 : * the receipt of the LSN.
3821 : */
3822 126342 : if (wait_event == WAIT_EVENT_WAIT_FOR_STANDBY_CONFIRMATION)
3823 10 : ConditionVariablePrepareToSleep(&WalSndCtl->wal_confirm_rcv_cv);
3824 126332 : else if (MyWalSnd->kind == REPLICATION_KIND_PHYSICAL)
3825 115316 : ConditionVariablePrepareToSleep(&WalSndCtl->wal_flush_cv);
3826 11016 : else if (MyWalSnd->kind == REPLICATION_KIND_LOGICAL)
3827 11016 : ConditionVariablePrepareToSleep(&WalSndCtl->wal_replay_cv);
3828 :
3829 126342 : if (WaitEventSetWait(FeBeWaitSet, timeout, &event, 1, wait_event) == 1 &&
3830 126342 : (event.events & WL_POSTMASTER_DEATH))
3831 : {
3832 0 : ConditionVariableCancelSleep();
3833 0 : proc_exit(1);
3834 : }
3835 :
3836 126342 : ConditionVariableCancelSleep();
3837 126342 : }
3838 :
3839 : /*
3840 : * Signal all walsenders to move to stopping state.
3841 : *
3842 : * This will trigger walsenders to move to a state where no further WAL can be
3843 : * generated. See this file's header for details.
3844 : */
3845 : void
3846 1272 : WalSndInitStopping(void)
3847 : {
3848 : int i;
3849 :
3850 9876 : for (i = 0; i < max_wal_senders; i++)
3851 : {
3852 8604 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3853 : pid_t pid;
3854 :
3855 8604 : SpinLockAcquire(&walsnd->mutex);
3856 8604 : pid = walsnd->pid;
3857 8604 : SpinLockRelease(&walsnd->mutex);
3858 :
3859 8604 : if (pid == 0)
3860 8532 : continue;
3861 :
3862 72 : SendProcSignal(pid, PROCSIG_WALSND_INIT_STOPPING, INVALID_PROC_NUMBER);
3863 : }
3864 1272 : }
3865 :
3866 : /*
3867 : * Wait that all the WAL senders have quit or reached the stopping state. This
3868 : * is used by the checkpointer to control when the shutdown checkpoint can
3869 : * safely be performed.
3870 : */
3871 : void
3872 1272 : WalSndWaitStopping(void)
3873 : {
3874 : for (;;)
3875 70 : {
3876 : int i;
3877 1342 : bool all_stopped = true;
3878 :
3879 9952 : for (i = 0; i < max_wal_senders; i++)
3880 : {
3881 8680 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3882 :
3883 8680 : SpinLockAcquire(&walsnd->mutex);
3884 :
3885 8680 : if (walsnd->pid == 0)
3886 : {
3887 8558 : SpinLockRelease(&walsnd->mutex);
3888 8558 : continue;
3889 : }
3890 :
3891 122 : if (walsnd->state != WALSNDSTATE_STOPPING)
3892 : {
3893 70 : all_stopped = false;
3894 70 : SpinLockRelease(&walsnd->mutex);
3895 70 : break;
3896 : }
3897 52 : SpinLockRelease(&walsnd->mutex);
3898 : }
3899 :
3900 : /* safe to leave if confirmation is done for all WAL senders */
3901 1342 : if (all_stopped)
3902 1272 : return;
3903 :
3904 70 : pg_usleep(10000L); /* wait for 10 msec */
3905 : }
3906 : }
3907 :
3908 : /* Set state for current walsender (only called in walsender) */
3909 : void
3910 21776 : WalSndSetState(WalSndState state)
3911 : {
3912 21776 : WalSnd *walsnd = MyWalSnd;
3913 :
3914 : Assert(am_walsender);
3915 :
3916 21776 : if (walsnd->state == state)
3917 18350 : return;
3918 :
3919 3426 : SpinLockAcquire(&walsnd->mutex);
3920 3426 : walsnd->state = state;
3921 3426 : SpinLockRelease(&walsnd->mutex);
3922 : }
3923 :
3924 : /*
3925 : * Return a string constant representing the state. This is used
3926 : * in system views, and should *not* be translated.
3927 : */
3928 : static const char *
3929 1982 : WalSndGetStateString(WalSndState state)
3930 : {
3931 1982 : switch (state)
3932 : {
3933 8 : case WALSNDSTATE_STARTUP:
3934 8 : return "startup";
3935 0 : case WALSNDSTATE_BACKUP:
3936 0 : return "backup";
3937 22 : case WALSNDSTATE_CATCHUP:
3938 22 : return "catchup";
3939 1952 : case WALSNDSTATE_STREAMING:
3940 1952 : return "streaming";
3941 0 : case WALSNDSTATE_STOPPING:
3942 0 : return "stopping";
3943 : }
3944 0 : return "UNKNOWN";
3945 : }
3946 :
3947 : static Interval *
3948 3374 : offset_to_interval(TimeOffset offset)
3949 : {
3950 3374 : Interval *result = palloc(sizeof(Interval));
3951 :
3952 3374 : result->month = 0;
3953 3374 : result->day = 0;
3954 3374 : result->time = offset;
3955 :
3956 3374 : return result;
3957 : }
3958 :
3959 : /*
3960 : * Returns activity of walsenders, including pids and xlog locations sent to
3961 : * standby servers.
3962 : */
3963 : Datum
3964 1696 : pg_stat_get_wal_senders(PG_FUNCTION_ARGS)
3965 : {
3966 : #define PG_STAT_GET_WAL_SENDERS_COLS 12
3967 1696 : ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
3968 : SyncRepStandbyData *sync_standbys;
3969 : int num_standbys;
3970 : int i;
3971 :
3972 1696 : InitMaterializedSRF(fcinfo, 0);
3973 :
3974 : /*
3975 : * Get the currently active synchronous standbys. This could be out of
3976 : * date before we're done, but we'll use the data anyway.
3977 : */
3978 1696 : num_standbys = SyncRepGetCandidateStandbys(&sync_standbys);
3979 :
3980 18148 : for (i = 0; i < max_wal_senders; i++)
3981 : {
3982 16452 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3983 : XLogRecPtr sent_ptr;
3984 : XLogRecPtr write;
3985 : XLogRecPtr flush;
3986 : XLogRecPtr apply;
3987 : TimeOffset writeLag;
3988 : TimeOffset flushLag;
3989 : TimeOffset applyLag;
3990 : int priority;
3991 : int pid;
3992 : WalSndState state;
3993 : TimestampTz replyTime;
3994 : bool is_sync_standby;
3995 : Datum values[PG_STAT_GET_WAL_SENDERS_COLS];
3996 16452 : bool nulls[PG_STAT_GET_WAL_SENDERS_COLS] = {0};
3997 : int j;
3998 :
3999 : /* Collect data from shared memory */
4000 16452 : SpinLockAcquire(&walsnd->mutex);
4001 16452 : if (walsnd->pid == 0)
4002 : {
4003 14470 : SpinLockRelease(&walsnd->mutex);
4004 14470 : continue;
4005 : }
4006 1982 : pid = walsnd->pid;
4007 1982 : sent_ptr = walsnd->sentPtr;
4008 1982 : state = walsnd->state;
4009 1982 : write = walsnd->write;
4010 1982 : flush = walsnd->flush;
4011 1982 : apply = walsnd->apply;
4012 1982 : writeLag = walsnd->writeLag;
4013 1982 : flushLag = walsnd->flushLag;
4014 1982 : applyLag = walsnd->applyLag;
4015 1982 : priority = walsnd->sync_standby_priority;
4016 1982 : replyTime = walsnd->replyTime;
4017 1982 : SpinLockRelease(&walsnd->mutex);
4018 :
4019 : /*
4020 : * Detect whether walsender is/was considered synchronous. We can
4021 : * provide some protection against stale data by checking the PID
4022 : * along with walsnd_index.
4023 : */
4024 1982 : is_sync_standby = false;
4025 2064 : for (j = 0; j < num_standbys; j++)
4026 : {
4027 136 : if (sync_standbys[j].walsnd_index == i &&
4028 54 : sync_standbys[j].pid == pid)
4029 : {
4030 54 : is_sync_standby = true;
4031 54 : break;
4032 : }
4033 : }
4034 :
4035 1982 : values[0] = Int32GetDatum(pid);
4036 :
4037 1982 : if (!has_privs_of_role(GetUserId(), ROLE_PG_READ_ALL_STATS))
4038 : {
4039 : /*
4040 : * Only superusers and roles with privileges of pg_read_all_stats
4041 : * can see details. Other users only get the pid value to know
4042 : * it's a walsender, but no details.
4043 : */
4044 0 : MemSet(&nulls[1], true, PG_STAT_GET_WAL_SENDERS_COLS - 1);
4045 : }
4046 : else
4047 : {
4048 1982 : values[1] = CStringGetTextDatum(WalSndGetStateString(state));
4049 :
4050 1982 : if (XLogRecPtrIsInvalid(sent_ptr))
4051 6 : nulls[2] = true;
4052 1982 : values[2] = LSNGetDatum(sent_ptr);
4053 :
4054 1982 : if (XLogRecPtrIsInvalid(write))
4055 18 : nulls[3] = true;
4056 1982 : values[3] = LSNGetDatum(write);
4057 :
4058 1982 : if (XLogRecPtrIsInvalid(flush))
4059 18 : nulls[4] = true;
4060 1982 : values[4] = LSNGetDatum(flush);
4061 :
4062 1982 : if (XLogRecPtrIsInvalid(apply))
4063 18 : nulls[5] = true;
4064 1982 : values[5] = LSNGetDatum(apply);
4065 :
4066 : /*
4067 : * Treat a standby such as a pg_basebackup background process
4068 : * which always returns an invalid flush location, as an
4069 : * asynchronous standby.
4070 : */
4071 1982 : priority = XLogRecPtrIsInvalid(flush) ? 0 : priority;
4072 :
4073 1982 : if (writeLag < 0)
4074 922 : nulls[6] = true;
4075 : else
4076 1060 : values[6] = IntervalPGetDatum(offset_to_interval(writeLag));
4077 :
4078 1982 : if (flushLag < 0)
4079 726 : nulls[7] = true;
4080 : else
4081 1256 : values[7] = IntervalPGetDatum(offset_to_interval(flushLag));
4082 :
4083 1982 : if (applyLag < 0)
4084 924 : nulls[8] = true;
4085 : else
4086 1058 : values[8] = IntervalPGetDatum(offset_to_interval(applyLag));
4087 :
4088 1982 : values[9] = Int32GetDatum(priority);
4089 :
4090 : /*
4091 : * More easily understood version of standby state. This is purely
4092 : * informational.
4093 : *
4094 : * In quorum-based sync replication, the role of each standby
4095 : * listed in synchronous_standby_names can be changing very
4096 : * frequently. Any standbys considered as "sync" at one moment can
4097 : * be switched to "potential" ones at the next moment. So, it's
4098 : * basically useless to report "sync" or "potential" as their sync
4099 : * states. We report just "quorum" for them.
4100 : */
4101 1982 : if (priority == 0)
4102 1906 : values[10] = CStringGetTextDatum("async");
4103 76 : else if (is_sync_standby)
4104 54 : values[10] = SyncRepConfig->syncrep_method == SYNC_REP_PRIORITY ?
4105 54 : CStringGetTextDatum("sync") : CStringGetTextDatum("quorum");
4106 : else
4107 22 : values[10] = CStringGetTextDatum("potential");
4108 :
4109 1982 : if (replyTime == 0)
4110 6 : nulls[11] = true;
4111 : else
4112 1976 : values[11] = TimestampTzGetDatum(replyTime);
4113 : }
4114 :
4115 1982 : tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc,
4116 : values, nulls);
4117 : }
4118 :
4119 1696 : return (Datum) 0;
4120 : }
4121 :
4122 : /*
4123 : * Send a keepalive message to standby.
4124 : *
4125 : * If requestReply is set, the message requests the other party to send
4126 : * a message back to us, for heartbeat purposes. We also set a flag to
4127 : * let nearby code know that we're waiting for that response, to avoid
4128 : * repeated requests.
4129 : *
4130 : * writePtr is the location up to which the WAL is sent. It is essentially
4131 : * the same as sentPtr but in some cases, we need to send keep alive before
4132 : * sentPtr is updated like when skipping empty transactions.
4133 : */
4134 : static void
4135 9948 : WalSndKeepalive(bool requestReply, XLogRecPtr writePtr)
4136 : {
4137 9948 : elog(DEBUG2, "sending replication keepalive");
4138 :
4139 : /* construct the message... */
4140 9948 : resetStringInfo(&output_message);
4141 9948 : pq_sendbyte(&output_message, PqReplMsg_Keepalive);
4142 9948 : pq_sendint64(&output_message, XLogRecPtrIsInvalid(writePtr) ? sentPtr : writePtr);
4143 9948 : pq_sendint64(&output_message, GetCurrentTimestamp());
4144 9948 : pq_sendbyte(&output_message, requestReply ? 1 : 0);
4145 :
4146 : /* ... and send it wrapped in CopyData */
4147 9948 : pq_putmessage_noblock(PqMsg_CopyData, output_message.data, output_message.len);
4148 :
4149 : /* Set local flag */
4150 9948 : if (requestReply)
4151 6468 : waiting_for_ping_response = true;
4152 9948 : }
4153 :
4154 : /*
4155 : * Send keepalive message if too much time has elapsed.
4156 : */
4157 : static void
4158 1517656 : WalSndKeepaliveIfNecessary(void)
4159 : {
4160 : TimestampTz ping_time;
4161 :
4162 : /*
4163 : * Don't send keepalive messages if timeouts are globally disabled or
4164 : * we're doing something not partaking in timeouts.
4165 : */
4166 1517656 : if (wal_sender_timeout <= 0 || last_reply_timestamp <= 0)
4167 50 : return;
4168 :
4169 1517606 : if (waiting_for_ping_response)
4170 18748 : return;
4171 :
4172 : /*
4173 : * If half of wal_sender_timeout has lapsed without receiving any reply
4174 : * from the standby, send a keep-alive message to the standby requesting
4175 : * an immediate reply.
4176 : */
4177 1498858 : ping_time = TimestampTzPlusMilliseconds(last_reply_timestamp,
4178 : wal_sender_timeout / 2);
4179 1498858 : if (last_processing >= ping_time)
4180 : {
4181 0 : WalSndKeepalive(true, InvalidXLogRecPtr);
4182 :
4183 : /* Try to flush pending output to the client */
4184 0 : if (pq_flush_if_writable() != 0)
4185 0 : WalSndShutdown();
4186 : }
4187 : }
4188 :
4189 : /*
4190 : * Record the end of the WAL and the time it was flushed locally, so that
4191 : * LagTrackerRead can compute the elapsed time (lag) when this WAL location is
4192 : * eventually reported to have been written, flushed and applied by the
4193 : * standby in a reply message.
4194 : */
4195 : static void
4196 238156 : LagTrackerWrite(XLogRecPtr lsn, TimestampTz local_flush_time)
4197 : {
4198 : bool buffer_full;
4199 : int new_write_head;
4200 : int i;
4201 :
4202 238156 : if (!am_walsender)
4203 0 : return;
4204 :
4205 : /*
4206 : * If the lsn hasn't advanced since last time, then do nothing. This way
4207 : * we only record a new sample when new WAL has been written.
4208 : */
4209 238156 : if (lag_tracker->last_lsn == lsn)
4210 222220 : return;
4211 15936 : lag_tracker->last_lsn = lsn;
4212 :
4213 : /*
4214 : * If advancing the write head of the circular buffer would crash into any
4215 : * of the read heads, then the buffer is full. In other words, the
4216 : * slowest reader (presumably apply) is the one that controls the release
4217 : * of space.
4218 : */
4219 15936 : new_write_head = (lag_tracker->write_head + 1) % LAG_TRACKER_BUFFER_SIZE;
4220 15936 : buffer_full = false;
4221 63744 : for (i = 0; i < NUM_SYNC_REP_WAIT_MODE; ++i)
4222 : {
4223 47808 : if (new_write_head == lag_tracker->read_heads[i])
4224 0 : buffer_full = true;
4225 : }
4226 :
4227 : /*
4228 : * If the buffer is full, for now we just rewind by one slot and overwrite
4229 : * the last sample, as a simple (if somewhat uneven) way to lower the
4230 : * sampling rate. There may be better adaptive compaction algorithms.
4231 : */
4232 15936 : if (buffer_full)
4233 : {
4234 0 : new_write_head = lag_tracker->write_head;
4235 0 : if (lag_tracker->write_head > 0)
4236 0 : lag_tracker->write_head--;
4237 : else
4238 0 : lag_tracker->write_head = LAG_TRACKER_BUFFER_SIZE - 1;
4239 : }
4240 :
4241 : /* Store a sample at the current write head position. */
4242 15936 : lag_tracker->buffer[lag_tracker->write_head].lsn = lsn;
4243 15936 : lag_tracker->buffer[lag_tracker->write_head].time = local_flush_time;
4244 15936 : lag_tracker->write_head = new_write_head;
4245 : }
4246 :
4247 : /*
4248 : * Find out how much time has elapsed between the moment WAL location 'lsn'
4249 : * (or the highest known earlier LSN) was flushed locally and the time 'now'.
4250 : * We have a separate read head for each of the reported LSN locations we
4251 : * receive in replies from standby; 'head' controls which read head is
4252 : * used. Whenever a read head crosses an LSN which was written into the
4253 : * lag buffer with LagTrackerWrite, we can use the associated timestamp to
4254 : * find out the time this LSN (or an earlier one) was flushed locally, and
4255 : * therefore compute the lag.
4256 : *
4257 : * Return -1 if no new sample data is available, and otherwise the elapsed
4258 : * time in microseconds.
4259 : */
4260 : static TimeOffset
4261 742098 : LagTrackerRead(int head, XLogRecPtr lsn, TimestampTz now)
4262 : {
4263 742098 : TimestampTz time = 0;
4264 :
4265 : /* Read all unread samples up to this LSN or end of buffer. */
4266 787804 : while (lag_tracker->read_heads[head] != lag_tracker->write_head &&
4267 420844 : lag_tracker->buffer[lag_tracker->read_heads[head]].lsn <= lsn)
4268 : {
4269 45706 : time = lag_tracker->buffer[lag_tracker->read_heads[head]].time;
4270 45706 : lag_tracker->last_read[head] =
4271 45706 : lag_tracker->buffer[lag_tracker->read_heads[head]];
4272 45706 : lag_tracker->read_heads[head] =
4273 45706 : (lag_tracker->read_heads[head] + 1) % LAG_TRACKER_BUFFER_SIZE;
4274 : }
4275 :
4276 : /*
4277 : * If the lag tracker is empty, that means the standby has processed
4278 : * everything we've ever sent so we should now clear 'last_read'. If we
4279 : * didn't do that, we'd risk using a stale and irrelevant sample for
4280 : * interpolation at the beginning of the next burst of WAL after a period
4281 : * of idleness.
4282 : */
4283 742098 : if (lag_tracker->read_heads[head] == lag_tracker->write_head)
4284 366960 : lag_tracker->last_read[head].time = 0;
4285 :
4286 742098 : if (time > now)
4287 : {
4288 : /* If the clock somehow went backwards, treat as not found. */
4289 0 : return -1;
4290 : }
4291 742098 : else if (time == 0)
4292 : {
4293 : /*
4294 : * We didn't cross a time. If there is a future sample that we
4295 : * haven't reached yet, and we've already reached at least one sample,
4296 : * let's interpolate the local flushed time. This is mainly useful
4297 : * for reporting a completely stuck apply position as having
4298 : * increasing lag, since otherwise we'd have to wait for it to
4299 : * eventually start moving again and cross one of our samples before
4300 : * we can show the lag increasing.
4301 : */
4302 706832 : if (lag_tracker->read_heads[head] == lag_tracker->write_head)
4303 : {
4304 : /* There are no future samples, so we can't interpolate. */
4305 337334 : return -1;
4306 : }
4307 369498 : else if (lag_tracker->last_read[head].time != 0)
4308 : {
4309 : /* We can interpolate between last_read and the next sample. */
4310 : double fraction;
4311 295178 : WalTimeSample prev = lag_tracker->last_read[head];
4312 295178 : WalTimeSample next = lag_tracker->buffer[lag_tracker->read_heads[head]];
4313 :
4314 295178 : if (lsn < prev.lsn)
4315 : {
4316 : /*
4317 : * Reported LSNs shouldn't normally go backwards, but it's
4318 : * possible when there is a timeline change. Treat as not
4319 : * found.
4320 : */
4321 0 : return -1;
4322 : }
4323 :
4324 : Assert(prev.lsn < next.lsn);
4325 :
4326 295178 : if (prev.time > next.time)
4327 : {
4328 : /* If the clock somehow went backwards, treat as not found. */
4329 0 : return -1;
4330 : }
4331 :
4332 : /* See how far we are between the previous and next samples. */
4333 295178 : fraction =
4334 295178 : (double) (lsn - prev.lsn) / (double) (next.lsn - prev.lsn);
4335 :
4336 : /* Scale the local flush time proportionally. */
4337 295178 : time = (TimestampTz)
4338 295178 : ((double) prev.time + (next.time - prev.time) * fraction);
4339 : }
4340 : else
4341 : {
4342 : /*
4343 : * We have only a future sample, implying that we were entirely
4344 : * caught up but and now there is a new burst of WAL and the
4345 : * standby hasn't processed the first sample yet. Until the
4346 : * standby reaches the future sample the best we can do is report
4347 : * the hypothetical lag if that sample were to be replayed now.
4348 : */
4349 74320 : time = lag_tracker->buffer[lag_tracker->read_heads[head]].time;
4350 : }
4351 : }
4352 :
4353 : /* Return the elapsed time since local flush time in microseconds. */
4354 : Assert(time != 0);
4355 404764 : return now - time;
4356 : }
|
