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/dbcommands.h"
64 : #include "commands/defrem.h"
65 : #include "funcapi.h"
66 : #include "libpq/libpq.h"
67 : #include "libpq/pqformat.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/fd.h"
83 : #include "storage/ipc.h"
84 : #include "storage/pmsignal.h"
85 : #include "storage/proc.h"
86 : #include "tcop/dest.h"
87 : #include "tcop/tcopprot.h"
88 : #include "utils/acl.h"
89 : #include "utils/builtins.h"
90 : #include "utils/guc.h"
91 : #include "utils/memutils.h"
92 : #include "utils/pg_lsn.h"
93 : #include "utils/ps_status.h"
94 : #include "utils/timeout.h"
95 : #include "utils/timestamp.h"
96 :
97 : /*
98 : * Maximum data payload in a WAL data message. Must be >= XLOG_BLCKSZ.
99 : *
100 : * We don't have a good idea of what a good value would be; there's some
101 : * overhead per message in both walsender and walreceiver, but on the other
102 : * hand sending large batches makes walsender less responsive to signals
103 : * because signals are checked only between messages. 128kB (with
104 : * default 8k blocks) seems like a reasonable guess for now.
105 : */
106 : #define MAX_SEND_SIZE (XLOG_BLCKSZ * 16)
107 :
108 : /* Array of WalSnds in shared memory */
109 : WalSndCtlData *WalSndCtl = NULL;
110 :
111 : /* My slot in the shared memory array */
112 : WalSnd *MyWalSnd = NULL;
113 :
114 : /* Global state */
115 : bool am_walsender = false; /* Am I a walsender process? */
116 : bool am_cascading_walsender = false; /* Am I cascading WAL to another
117 : * standby? */
118 : bool am_db_walsender = false; /* Connected to a database? */
119 :
120 : /* GUC variables */
121 : int max_wal_senders = 10; /* the maximum number of concurrent
122 : * walsenders */
123 : int wal_sender_timeout = 60 * 1000; /* maximum time to send one WAL
124 : * data message */
125 : bool log_replication_commands = false;
126 :
127 : /*
128 : * State for WalSndWakeupRequest
129 : */
130 : bool wake_wal_senders = false;
131 :
132 : /*
133 : * xlogreader used for replication. Note that a WAL sender doing physical
134 : * replication does not need xlogreader to read WAL, but it needs one to
135 : * keep a state of its work.
136 : */
137 : static XLogReaderState *xlogreader = NULL;
138 :
139 : /*
140 : * If the UPLOAD_MANIFEST command is used to provide a backup manifest in
141 : * preparation for an incremental backup, uploaded_manifest will be point
142 : * to an object containing information about its contexts, and
143 : * uploaded_manifest_mcxt will point to the memory context that contains
144 : * that object and all of its subordinate data. Otherwise, both values will
145 : * be NULL.
146 : */
147 : static IncrementalBackupInfo *uploaded_manifest = NULL;
148 : static MemoryContext uploaded_manifest_mcxt = NULL;
149 :
150 : /*
151 : * These variables keep track of the state of the timeline we're currently
152 : * sending. sendTimeLine identifies the timeline. If sendTimeLineIsHistoric,
153 : * the timeline is not the latest timeline on this server, and the server's
154 : * history forked off from that timeline at sendTimeLineValidUpto.
155 : */
156 : static TimeLineID sendTimeLine = 0;
157 : static TimeLineID sendTimeLineNextTLI = 0;
158 : static bool sendTimeLineIsHistoric = false;
159 : static XLogRecPtr sendTimeLineValidUpto = InvalidXLogRecPtr;
160 :
161 : /*
162 : * How far have we sent WAL already? This is also advertised in
163 : * MyWalSnd->sentPtr. (Actually, this is the next WAL location to send.)
164 : */
165 : static XLogRecPtr sentPtr = InvalidXLogRecPtr;
166 :
167 : /* Buffers for constructing outgoing messages and processing reply messages. */
168 : static StringInfoData output_message;
169 : static StringInfoData reply_message;
170 : static StringInfoData tmpbuf;
171 :
172 : /* Timestamp of last ProcessRepliesIfAny(). */
173 : static TimestampTz last_processing = 0;
174 :
175 : /*
176 : * Timestamp of last ProcessRepliesIfAny() that saw a reply from the
177 : * standby. Set to 0 if wal_sender_timeout doesn't need to be active.
178 : */
179 : static TimestampTz last_reply_timestamp = 0;
180 :
181 : /* Have we sent a heartbeat message asking for reply, since last reply? */
182 : static bool waiting_for_ping_response = false;
183 :
184 : /*
185 : * While streaming WAL in Copy mode, streamingDoneSending is set to true
186 : * after we have sent CopyDone. We should not send any more CopyData messages
187 : * after that. streamingDoneReceiving is set to true when we receive CopyDone
188 : * from the other end. When both become true, it's time to exit Copy mode.
189 : */
190 : static bool streamingDoneSending;
191 : static bool streamingDoneReceiving;
192 :
193 : /* Are we there yet? */
194 : static bool WalSndCaughtUp = false;
195 :
196 : /* Flags set by signal handlers for later service in main loop */
197 : static volatile sig_atomic_t got_SIGUSR2 = false;
198 : static volatile sig_atomic_t got_STOPPING = false;
199 :
200 : /*
201 : * This is set while we are streaming. When not set
202 : * PROCSIG_WALSND_INIT_STOPPING signal will be handled like SIGTERM. When set,
203 : * the main loop is responsible for checking got_STOPPING and terminating when
204 : * it's set (after streaming any remaining WAL).
205 : */
206 : static volatile sig_atomic_t replication_active = false;
207 :
208 : static LogicalDecodingContext *logical_decoding_ctx = NULL;
209 :
210 : /* A sample associating a WAL location with the time it was written. */
211 : typedef struct
212 : {
213 : XLogRecPtr lsn;
214 : TimestampTz time;
215 : } WalTimeSample;
216 :
217 : /* The size of our buffer of time samples. */
218 : #define LAG_TRACKER_BUFFER_SIZE 8192
219 :
220 : /* A mechanism for tracking replication lag. */
221 : typedef struct
222 : {
223 : XLogRecPtr last_lsn;
224 : WalTimeSample buffer[LAG_TRACKER_BUFFER_SIZE];
225 : int write_head;
226 : int read_heads[NUM_SYNC_REP_WAIT_MODE];
227 : WalTimeSample last_read[NUM_SYNC_REP_WAIT_MODE];
228 : } LagTracker;
229 :
230 : static LagTracker *lag_tracker;
231 :
232 : /* Signal handlers */
233 : static void WalSndLastCycleHandler(SIGNAL_ARGS);
234 :
235 : /* Prototypes for private functions */
236 : typedef void (*WalSndSendDataCallback) (void);
237 : static void WalSndLoop(WalSndSendDataCallback send_data);
238 : static void InitWalSenderSlot(void);
239 : static void WalSndKill(int code, Datum arg);
240 : static void WalSndShutdown(void) pg_attribute_noreturn();
241 : static void XLogSendPhysical(void);
242 : static void XLogSendLogical(void);
243 : static void WalSndDone(WalSndSendDataCallback send_data);
244 : static void IdentifySystem(void);
245 : static void UploadManifest(void);
246 : static bool HandleUploadManifestPacket(StringInfo buf, off_t *offset,
247 : IncrementalBackupInfo *ib);
248 : static void ReadReplicationSlot(ReadReplicationSlotCmd *cmd);
249 : static void CreateReplicationSlot(CreateReplicationSlotCmd *cmd);
250 : static void DropReplicationSlot(DropReplicationSlotCmd *cmd);
251 : static void StartReplication(StartReplicationCmd *cmd);
252 : static void StartLogicalReplication(StartReplicationCmd *cmd);
253 : static void ProcessStandbyMessage(void);
254 : static void ProcessStandbyReplyMessage(void);
255 : static void ProcessStandbyHSFeedbackMessage(void);
256 : static void ProcessRepliesIfAny(void);
257 : static void ProcessPendingWrites(void);
258 : static void WalSndKeepalive(bool requestReply, XLogRecPtr writePtr);
259 : static void WalSndKeepaliveIfNecessary(void);
260 : static void WalSndCheckTimeOut(void);
261 : static long WalSndComputeSleeptime(TimestampTz now);
262 : static void WalSndWait(uint32 socket_events, long timeout, uint32 wait_event);
263 : static void WalSndPrepareWrite(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid, bool last_write);
264 : static void WalSndWriteData(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid, bool last_write);
265 : static void WalSndUpdateProgress(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid,
266 : bool skipped_xact);
267 : static XLogRecPtr WalSndWaitForWal(XLogRecPtr loc);
268 : static void LagTrackerWrite(XLogRecPtr lsn, TimestampTz local_flush_time);
269 : static TimeOffset LagTrackerRead(int head, XLogRecPtr lsn, TimestampTz now);
270 : static bool TransactionIdInRecentPast(TransactionId xid, uint32 epoch);
271 :
272 : static void WalSndSegmentOpen(XLogReaderState *state, XLogSegNo nextSegNo,
273 : TimeLineID *tli_p);
274 :
275 :
276 : /* Initialize walsender process before entering the main command loop */
277 : void
278 2124 : InitWalSender(void)
279 : {
280 2124 : am_cascading_walsender = RecoveryInProgress();
281 :
282 : /* Create a per-walsender data structure in shared memory */
283 2124 : InitWalSenderSlot();
284 :
285 : /* need resource owner for e.g. basebackups */
286 2124 : CreateAuxProcessResourceOwner();
287 :
288 : /*
289 : * Let postmaster know that we're a WAL sender. Once we've declared us as
290 : * a WAL sender process, postmaster will let us outlive the bgwriter and
291 : * kill us last in the shutdown sequence, so we get a chance to stream all
292 : * remaining WAL at shutdown, including the shutdown checkpoint. Note that
293 : * there's no going back, and we mustn't write any WAL records after this.
294 : */
295 2124 : MarkPostmasterChildWalSender();
296 2124 : SendPostmasterSignal(PMSIGNAL_ADVANCE_STATE_MACHINE);
297 :
298 : /*
299 : * If the client didn't specify a database to connect to, show in PGPROC
300 : * that our advertised xmin should affect vacuum horizons in all
301 : * databases. This allows physical replication clients to send hot
302 : * standby feedback that will delay vacuum cleanup in all databases.
303 : */
304 2124 : if (MyDatabaseId == InvalidOid)
305 : {
306 : Assert(MyProc->xmin == InvalidTransactionId);
307 892 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
308 892 : MyProc->statusFlags |= PROC_AFFECTS_ALL_HORIZONS;
309 892 : ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
310 892 : LWLockRelease(ProcArrayLock);
311 : }
312 :
313 : /* Initialize empty timestamp buffer for lag tracking. */
314 2124 : lag_tracker = MemoryContextAllocZero(TopMemoryContext, sizeof(LagTracker));
315 2124 : }
316 :
317 : /*
318 : * Clean up after an error.
319 : *
320 : * WAL sender processes don't use transactions like regular backends do.
321 : * This function does any cleanup required after an error in a WAL sender
322 : * process, similar to what transaction abort does in a regular backend.
323 : */
324 : void
325 102 : WalSndErrorCleanup(void)
326 : {
327 102 : LWLockReleaseAll();
328 102 : ConditionVariableCancelSleep();
329 102 : pgstat_report_wait_end();
330 :
331 102 : if (xlogreader != NULL && xlogreader->seg.ws_file >= 0)
332 16 : wal_segment_close(xlogreader);
333 :
334 102 : if (MyReplicationSlot != NULL)
335 36 : ReplicationSlotRelease();
336 :
337 102 : ReplicationSlotCleanup(false);
338 :
339 102 : replication_active = false;
340 :
341 : /*
342 : * If there is a transaction in progress, it will clean up our
343 : * ResourceOwner, but if a replication command set up a resource owner
344 : * without a transaction, we've got to clean that up now.
345 : */
346 102 : if (!IsTransactionOrTransactionBlock())
347 100 : ReleaseAuxProcessResources(false);
348 :
349 102 : if (got_STOPPING || got_SIGUSR2)
350 0 : proc_exit(0);
351 :
352 : /* Revert back to startup state */
353 102 : WalSndSetState(WALSNDSTATE_STARTUP);
354 102 : }
355 :
356 : /*
357 : * Handle a client's connection abort in an orderly manner.
358 : */
359 : static void
360 2 : WalSndShutdown(void)
361 : {
362 : /*
363 : * Reset whereToSendOutput to prevent ereport from attempting to send any
364 : * more messages to the standby.
365 : */
366 2 : if (whereToSendOutput == DestRemote)
367 2 : whereToSendOutput = DestNone;
368 :
369 2 : proc_exit(0);
370 : abort(); /* keep the compiler quiet */
371 : }
372 :
373 : /*
374 : * Handle the IDENTIFY_SYSTEM command.
375 : */
376 : static void
377 1318 : IdentifySystem(void)
378 : {
379 : char sysid[32];
380 : char xloc[MAXFNAMELEN];
381 : XLogRecPtr logptr;
382 1318 : char *dbname = NULL;
383 : DestReceiver *dest;
384 : TupOutputState *tstate;
385 : TupleDesc tupdesc;
386 : Datum values[4];
387 1318 : bool nulls[4] = {0};
388 : TimeLineID currTLI;
389 :
390 : /*
391 : * Reply with a result set with one row, four columns. First col is system
392 : * ID, second is timeline ID, third is current xlog location and the
393 : * fourth contains the database name if we are connected to one.
394 : */
395 :
396 1318 : snprintf(sysid, sizeof(sysid), UINT64_FORMAT,
397 : GetSystemIdentifier());
398 :
399 1318 : am_cascading_walsender = RecoveryInProgress();
400 1318 : if (am_cascading_walsender)
401 114 : logptr = GetStandbyFlushRecPtr(&currTLI);
402 : else
403 1204 : logptr = GetFlushRecPtr(&currTLI);
404 :
405 1318 : snprintf(xloc, sizeof(xloc), "%X/%X", LSN_FORMAT_ARGS(logptr));
406 :
407 1318 : if (MyDatabaseId != InvalidOid)
408 : {
409 426 : MemoryContext cur = CurrentMemoryContext;
410 :
411 : /* syscache access needs a transaction env. */
412 426 : StartTransactionCommand();
413 426 : dbname = get_database_name(MyDatabaseId);
414 : /* copy dbname out of TX context */
415 426 : dbname = MemoryContextStrdup(cur, dbname);
416 426 : CommitTransactionCommand();
417 : }
418 :
419 1318 : dest = CreateDestReceiver(DestRemoteSimple);
420 :
421 : /* need a tuple descriptor representing four columns */
422 1318 : tupdesc = CreateTemplateTupleDesc(4);
423 1318 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "systemid",
424 : TEXTOID, -1, 0);
425 1318 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "timeline",
426 : INT8OID, -1, 0);
427 1318 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 3, "xlogpos",
428 : TEXTOID, -1, 0);
429 1318 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 4, "dbname",
430 : TEXTOID, -1, 0);
431 :
432 : /* prepare for projection of tuples */
433 1318 : tstate = begin_tup_output_tupdesc(dest, tupdesc, &TTSOpsVirtual);
434 :
435 : /* column 1: system identifier */
436 1318 : values[0] = CStringGetTextDatum(sysid);
437 :
438 : /* column 2: timeline */
439 1318 : values[1] = Int64GetDatum(currTLI);
440 :
441 : /* column 3: wal location */
442 1318 : values[2] = CStringGetTextDatum(xloc);
443 :
444 : /* column 4: database name, or NULL if none */
445 1318 : if (dbname)
446 426 : values[3] = CStringGetTextDatum(dbname);
447 : else
448 892 : nulls[3] = true;
449 :
450 : /* send it to dest */
451 1318 : do_tup_output(tstate, values, nulls);
452 :
453 1318 : end_tup_output(tstate);
454 1318 : }
455 :
456 : /* Handle READ_REPLICATION_SLOT command */
457 : static void
458 12 : ReadReplicationSlot(ReadReplicationSlotCmd *cmd)
459 : {
460 : #define READ_REPLICATION_SLOT_COLS 3
461 : ReplicationSlot *slot;
462 : DestReceiver *dest;
463 : TupOutputState *tstate;
464 : TupleDesc tupdesc;
465 12 : Datum values[READ_REPLICATION_SLOT_COLS] = {0};
466 : bool nulls[READ_REPLICATION_SLOT_COLS];
467 :
468 12 : tupdesc = CreateTemplateTupleDesc(READ_REPLICATION_SLOT_COLS);
469 12 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "slot_type",
470 : TEXTOID, -1, 0);
471 12 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "restart_lsn",
472 : TEXTOID, -1, 0);
473 : /* TimeLineID is unsigned, so int4 is not wide enough. */
474 12 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 3, "restart_tli",
475 : INT8OID, -1, 0);
476 :
477 12 : memset(nulls, true, READ_REPLICATION_SLOT_COLS * sizeof(bool));
478 :
479 12 : LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
480 12 : slot = SearchNamedReplicationSlot(cmd->slotname, false);
481 12 : if (slot == NULL || !slot->in_use)
482 : {
483 4 : LWLockRelease(ReplicationSlotControlLock);
484 : }
485 : else
486 : {
487 : ReplicationSlot slot_contents;
488 8 : int i = 0;
489 :
490 : /* Copy slot contents while holding spinlock */
491 8 : SpinLockAcquire(&slot->mutex);
492 8 : slot_contents = *slot;
493 8 : SpinLockRelease(&slot->mutex);
494 8 : LWLockRelease(ReplicationSlotControlLock);
495 :
496 8 : if (OidIsValid(slot_contents.data.database))
497 2 : ereport(ERROR,
498 : errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
499 : errmsg("cannot use %s with a logical replication slot",
500 : "READ_REPLICATION_SLOT"));
501 :
502 : /* slot type */
503 6 : values[i] = CStringGetTextDatum("physical");
504 6 : nulls[i] = false;
505 6 : i++;
506 :
507 : /* start LSN */
508 6 : if (!XLogRecPtrIsInvalid(slot_contents.data.restart_lsn))
509 : {
510 : char xloc[64];
511 :
512 6 : snprintf(xloc, sizeof(xloc), "%X/%X",
513 6 : LSN_FORMAT_ARGS(slot_contents.data.restart_lsn));
514 6 : values[i] = CStringGetTextDatum(xloc);
515 6 : nulls[i] = false;
516 : }
517 6 : i++;
518 :
519 : /* timeline this WAL was produced on */
520 6 : if (!XLogRecPtrIsInvalid(slot_contents.data.restart_lsn))
521 : {
522 : TimeLineID slots_position_timeline;
523 : TimeLineID current_timeline;
524 6 : List *timeline_history = NIL;
525 :
526 : /*
527 : * While in recovery, use as timeline the currently-replaying one
528 : * to get the LSN position's history.
529 : */
530 6 : if (RecoveryInProgress())
531 0 : (void) GetXLogReplayRecPtr(¤t_timeline);
532 : else
533 6 : current_timeline = GetWALInsertionTimeLine();
534 :
535 6 : timeline_history = readTimeLineHistory(current_timeline);
536 6 : slots_position_timeline = tliOfPointInHistory(slot_contents.data.restart_lsn,
537 : timeline_history);
538 6 : values[i] = Int64GetDatum((int64) slots_position_timeline);
539 6 : nulls[i] = false;
540 : }
541 6 : i++;
542 :
543 : Assert(i == READ_REPLICATION_SLOT_COLS);
544 : }
545 :
546 10 : dest = CreateDestReceiver(DestRemoteSimple);
547 10 : tstate = begin_tup_output_tupdesc(dest, tupdesc, &TTSOpsVirtual);
548 10 : do_tup_output(tstate, values, nulls);
549 10 : end_tup_output(tstate);
550 10 : }
551 :
552 :
553 : /*
554 : * Handle TIMELINE_HISTORY command.
555 : */
556 : static void
557 28 : SendTimeLineHistory(TimeLineHistoryCmd *cmd)
558 : {
559 : DestReceiver *dest;
560 : TupleDesc tupdesc;
561 : StringInfoData buf;
562 : char histfname[MAXFNAMELEN];
563 : char path[MAXPGPATH];
564 : int fd;
565 : off_t histfilelen;
566 : off_t bytesleft;
567 : Size len;
568 :
569 28 : dest = CreateDestReceiver(DestRemoteSimple);
570 :
571 : /*
572 : * Reply with a result set with one row, and two columns. The first col is
573 : * the name of the history file, 2nd is the contents.
574 : */
575 28 : tupdesc = CreateTemplateTupleDesc(2);
576 28 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "filename", TEXTOID, -1, 0);
577 28 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "content", TEXTOID, -1, 0);
578 :
579 28 : TLHistoryFileName(histfname, cmd->timeline);
580 28 : TLHistoryFilePath(path, cmd->timeline);
581 :
582 : /* Send a RowDescription message */
583 28 : dest->rStartup(dest, CMD_SELECT, tupdesc);
584 :
585 : /* Send a DataRow message */
586 28 : pq_beginmessage(&buf, PqMsg_DataRow);
587 28 : pq_sendint16(&buf, 2); /* # of columns */
588 28 : len = strlen(histfname);
589 28 : pq_sendint32(&buf, len); /* col1 len */
590 28 : pq_sendbytes(&buf, histfname, len);
591 :
592 28 : fd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
593 28 : if (fd < 0)
594 0 : ereport(ERROR,
595 : (errcode_for_file_access(),
596 : errmsg("could not open file \"%s\": %m", path)));
597 :
598 : /* Determine file length and send it to client */
599 28 : histfilelen = lseek(fd, 0, SEEK_END);
600 28 : if (histfilelen < 0)
601 0 : ereport(ERROR,
602 : (errcode_for_file_access(),
603 : errmsg("could not seek to end of file \"%s\": %m", path)));
604 28 : if (lseek(fd, 0, SEEK_SET) != 0)
605 0 : ereport(ERROR,
606 : (errcode_for_file_access(),
607 : errmsg("could not seek to beginning of file \"%s\": %m", path)));
608 :
609 28 : pq_sendint32(&buf, histfilelen); /* col2 len */
610 :
611 28 : bytesleft = histfilelen;
612 56 : while (bytesleft > 0)
613 : {
614 : PGAlignedBlock rbuf;
615 : int nread;
616 :
617 28 : pgstat_report_wait_start(WAIT_EVENT_WALSENDER_TIMELINE_HISTORY_READ);
618 28 : nread = read(fd, rbuf.data, sizeof(rbuf));
619 28 : pgstat_report_wait_end();
620 28 : if (nread < 0)
621 0 : ereport(ERROR,
622 : (errcode_for_file_access(),
623 : errmsg("could not read file \"%s\": %m",
624 : path)));
625 28 : else if (nread == 0)
626 0 : ereport(ERROR,
627 : (errcode(ERRCODE_DATA_CORRUPTED),
628 : errmsg("could not read file \"%s\": read %d of %zu",
629 : path, nread, (Size) bytesleft)));
630 :
631 28 : pq_sendbytes(&buf, rbuf.data, nread);
632 28 : bytesleft -= nread;
633 : }
634 :
635 28 : if (CloseTransientFile(fd) != 0)
636 0 : ereport(ERROR,
637 : (errcode_for_file_access(),
638 : errmsg("could not close file \"%s\": %m", path)));
639 :
640 28 : pq_endmessage(&buf);
641 28 : }
642 :
643 : /*
644 : * Handle UPLOAD_MANIFEST command.
645 : */
646 : static void
647 22 : UploadManifest(void)
648 : {
649 : MemoryContext mcxt;
650 : IncrementalBackupInfo *ib;
651 22 : off_t offset = 0;
652 : StringInfoData buf;
653 :
654 : /*
655 : * parsing the manifest will use the cryptohash stuff, which requires a
656 : * resource owner
657 : */
658 : Assert(AuxProcessResourceOwner != NULL);
659 : Assert(CurrentResourceOwner == AuxProcessResourceOwner ||
660 : CurrentResourceOwner == NULL);
661 22 : CurrentResourceOwner = AuxProcessResourceOwner;
662 :
663 : /* Prepare to read manifest data into a temporary context. */
664 22 : mcxt = AllocSetContextCreate(CurrentMemoryContext,
665 : "incremental backup information",
666 : ALLOCSET_DEFAULT_SIZES);
667 22 : ib = CreateIncrementalBackupInfo(mcxt);
668 :
669 : /* Send a CopyInResponse message */
670 22 : pq_beginmessage(&buf, PqMsg_CopyInResponse);
671 22 : pq_sendbyte(&buf, 0);
672 22 : pq_sendint16(&buf, 0);
673 22 : pq_endmessage_reuse(&buf);
674 22 : pq_flush();
675 :
676 : /* Receive packets from client until done. */
677 86 : while (HandleUploadManifestPacket(&buf, &offset, ib))
678 : ;
679 :
680 : /* Finish up manifest processing. */
681 20 : FinalizeIncrementalManifest(ib);
682 :
683 : /*
684 : * Discard any old manifest information and arrange to preserve the new
685 : * information we just got.
686 : *
687 : * We assume that MemoryContextDelete and MemoryContextSetParent won't
688 : * fail, and thus we shouldn't end up bailing out of here in such a way as
689 : * to leave dangling pointers.
690 : */
691 20 : if (uploaded_manifest_mcxt != NULL)
692 0 : MemoryContextDelete(uploaded_manifest_mcxt);
693 20 : MemoryContextSetParent(mcxt, CacheMemoryContext);
694 20 : uploaded_manifest = ib;
695 20 : uploaded_manifest_mcxt = mcxt;
696 :
697 : /* clean up the resource owner we created */
698 20 : ReleaseAuxProcessResources(true);
699 20 : }
700 :
701 : /*
702 : * Process one packet received during the handling of an UPLOAD_MANIFEST
703 : * operation.
704 : *
705 : * 'buf' is scratch space. This function expects it to be initialized, doesn't
706 : * care what the current contents are, and may override them with completely
707 : * new contents.
708 : *
709 : * The return value is true if the caller should continue processing
710 : * additional packets and false if the UPLOAD_MANIFEST operation is complete.
711 : */
712 : static bool
713 86 : HandleUploadManifestPacket(StringInfo buf, off_t *offset,
714 : IncrementalBackupInfo *ib)
715 : {
716 : int mtype;
717 : int maxmsglen;
718 :
719 86 : HOLD_CANCEL_INTERRUPTS();
720 :
721 86 : pq_startmsgread();
722 86 : mtype = pq_getbyte();
723 86 : if (mtype == EOF)
724 0 : ereport(ERROR,
725 : (errcode(ERRCODE_CONNECTION_FAILURE),
726 : errmsg("unexpected EOF on client connection with an open transaction")));
727 :
728 86 : switch (mtype)
729 : {
730 66 : case 'd': /* CopyData */
731 66 : maxmsglen = PQ_LARGE_MESSAGE_LIMIT;
732 66 : break;
733 20 : case 'c': /* CopyDone */
734 : case 'f': /* CopyFail */
735 : case 'H': /* Flush */
736 : case 'S': /* Sync */
737 20 : maxmsglen = PQ_SMALL_MESSAGE_LIMIT;
738 20 : break;
739 0 : default:
740 0 : ereport(ERROR,
741 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
742 : errmsg("unexpected message type 0x%02X during COPY from stdin",
743 : mtype)));
744 : maxmsglen = 0; /* keep compiler quiet */
745 : break;
746 : }
747 :
748 : /* Now collect the message body */
749 86 : if (pq_getmessage(buf, maxmsglen))
750 0 : ereport(ERROR,
751 : (errcode(ERRCODE_CONNECTION_FAILURE),
752 : errmsg("unexpected EOF on client connection with an open transaction")));
753 86 : RESUME_CANCEL_INTERRUPTS();
754 :
755 : /* Process the message */
756 86 : switch (mtype)
757 : {
758 66 : case 'd': /* CopyData */
759 66 : AppendIncrementalManifestData(ib, buf->data, buf->len);
760 64 : return true;
761 :
762 20 : case 'c': /* CopyDone */
763 20 : return false;
764 :
765 0 : case 'H': /* Sync */
766 : case 'S': /* Flush */
767 : /* Ignore these while in CopyOut mode as we do elsewhere. */
768 0 : return true;
769 :
770 0 : case 'f':
771 0 : ereport(ERROR,
772 : (errcode(ERRCODE_QUERY_CANCELED),
773 : errmsg("COPY from stdin failed: %s",
774 : pq_getmsgstring(buf))));
775 : }
776 :
777 : /* Not reached. */
778 : Assert(false);
779 0 : return false;
780 : }
781 :
782 : /*
783 : * Handle START_REPLICATION command.
784 : *
785 : * At the moment, this never returns, but an ereport(ERROR) will take us back
786 : * to the main loop.
787 : */
788 : static void
789 512 : StartReplication(StartReplicationCmd *cmd)
790 : {
791 : StringInfoData buf;
792 : XLogRecPtr FlushPtr;
793 : TimeLineID FlushTLI;
794 :
795 : /* create xlogreader for physical replication */
796 512 : xlogreader =
797 512 : XLogReaderAllocate(wal_segment_size, NULL,
798 512 : XL_ROUTINE(.segment_open = WalSndSegmentOpen,
799 : .segment_close = wal_segment_close),
800 : NULL);
801 :
802 512 : if (!xlogreader)
803 0 : ereport(ERROR,
804 : (errcode(ERRCODE_OUT_OF_MEMORY),
805 : errmsg("out of memory"),
806 : errdetail("Failed while allocating a WAL reading processor.")));
807 :
808 : /*
809 : * We assume here that we're logging enough information in the WAL for
810 : * log-shipping, since this is checked in PostmasterMain().
811 : *
812 : * NOTE: wal_level can only change at shutdown, so in most cases it is
813 : * difficult for there to be WAL data that we can still see that was
814 : * written at wal_level='minimal'.
815 : */
816 :
817 512 : if (cmd->slotname)
818 : {
819 354 : ReplicationSlotAcquire(cmd->slotname, true);
820 350 : if (SlotIsLogical(MyReplicationSlot))
821 0 : ereport(ERROR,
822 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
823 : errmsg("cannot use a logical replication slot for physical replication")));
824 :
825 : /*
826 : * We don't need to verify the slot's restart_lsn here; instead we
827 : * rely on the caller requesting the starting point to use. If the
828 : * WAL segment doesn't exist, we'll fail later.
829 : */
830 : }
831 :
832 : /*
833 : * Select the timeline. If it was given explicitly by the client, use
834 : * that. Otherwise use the timeline of the last replayed record.
835 : */
836 508 : am_cascading_walsender = RecoveryInProgress();
837 508 : if (am_cascading_walsender)
838 26 : FlushPtr = GetStandbyFlushRecPtr(&FlushTLI);
839 : else
840 482 : FlushPtr = GetFlushRecPtr(&FlushTLI);
841 :
842 508 : if (cmd->timeline != 0)
843 : {
844 : XLogRecPtr switchpoint;
845 :
846 506 : sendTimeLine = cmd->timeline;
847 506 : if (sendTimeLine == FlushTLI)
848 : {
849 482 : sendTimeLineIsHistoric = false;
850 482 : sendTimeLineValidUpto = InvalidXLogRecPtr;
851 : }
852 : else
853 : {
854 : List *timeLineHistory;
855 :
856 24 : sendTimeLineIsHistoric = true;
857 :
858 : /*
859 : * Check that the timeline the client requested exists, and the
860 : * requested start location is on that timeline.
861 : */
862 24 : timeLineHistory = readTimeLineHistory(FlushTLI);
863 24 : switchpoint = tliSwitchPoint(cmd->timeline, timeLineHistory,
864 : &sendTimeLineNextTLI);
865 24 : list_free_deep(timeLineHistory);
866 :
867 : /*
868 : * Found the requested timeline in the history. Check that
869 : * requested startpoint is on that timeline in our history.
870 : *
871 : * This is quite loose on purpose. We only check that we didn't
872 : * fork off the requested timeline before the switchpoint. We
873 : * don't check that we switched *to* it before the requested
874 : * starting point. This is because the client can legitimately
875 : * request to start replication from the beginning of the WAL
876 : * segment that contains switchpoint, but on the new timeline, so
877 : * that it doesn't end up with a partial segment. If you ask for
878 : * too old a starting point, you'll get an error later when we
879 : * fail to find the requested WAL segment in pg_wal.
880 : *
881 : * XXX: we could be more strict here and only allow a startpoint
882 : * that's older than the switchpoint, if it's still in the same
883 : * WAL segment.
884 : */
885 24 : if (!XLogRecPtrIsInvalid(switchpoint) &&
886 24 : switchpoint < cmd->startpoint)
887 : {
888 0 : ereport(ERROR,
889 : (errmsg("requested starting point %X/%X on timeline %u is not in this server's history",
890 : LSN_FORMAT_ARGS(cmd->startpoint),
891 : cmd->timeline),
892 : errdetail("This server's history forked from timeline %u at %X/%X.",
893 : cmd->timeline,
894 : LSN_FORMAT_ARGS(switchpoint))));
895 : }
896 24 : sendTimeLineValidUpto = switchpoint;
897 : }
898 : }
899 : else
900 : {
901 2 : sendTimeLine = FlushTLI;
902 2 : sendTimeLineValidUpto = InvalidXLogRecPtr;
903 2 : sendTimeLineIsHistoric = false;
904 : }
905 :
906 508 : streamingDoneSending = streamingDoneReceiving = false;
907 :
908 : /* If there is nothing to stream, don't even enter COPY mode */
909 508 : if (!sendTimeLineIsHistoric || cmd->startpoint < sendTimeLineValidUpto)
910 : {
911 : /*
912 : * When we first start replication the standby will be behind the
913 : * primary. For some applications, for example synchronous
914 : * replication, it is important to have a clear state for this initial
915 : * catchup mode, so we can trigger actions when we change streaming
916 : * state later. We may stay in this state for a long time, which is
917 : * exactly why we want to be able to monitor whether or not we are
918 : * still here.
919 : */
920 508 : WalSndSetState(WALSNDSTATE_CATCHUP);
921 :
922 : /* Send a CopyBothResponse message, and start streaming */
923 508 : pq_beginmessage(&buf, PqMsg_CopyBothResponse);
924 508 : pq_sendbyte(&buf, 0);
925 508 : pq_sendint16(&buf, 0);
926 508 : pq_endmessage(&buf);
927 508 : pq_flush();
928 :
929 : /*
930 : * Don't allow a request to stream from a future point in WAL that
931 : * hasn't been flushed to disk in this server yet.
932 : */
933 508 : if (FlushPtr < cmd->startpoint)
934 : {
935 0 : ereport(ERROR,
936 : (errmsg("requested starting point %X/%X is ahead of the WAL flush position of this server %X/%X",
937 : LSN_FORMAT_ARGS(cmd->startpoint),
938 : LSN_FORMAT_ARGS(FlushPtr))));
939 : }
940 :
941 : /* Start streaming from the requested point */
942 508 : sentPtr = cmd->startpoint;
943 :
944 : /* Initialize shared memory status, too */
945 508 : SpinLockAcquire(&MyWalSnd->mutex);
946 508 : MyWalSnd->sentPtr = sentPtr;
947 508 : SpinLockRelease(&MyWalSnd->mutex);
948 :
949 508 : SyncRepInitConfig();
950 :
951 : /* Main loop of walsender */
952 508 : replication_active = true;
953 :
954 508 : WalSndLoop(XLogSendPhysical);
955 :
956 294 : replication_active = false;
957 294 : if (got_STOPPING)
958 0 : proc_exit(0);
959 294 : WalSndSetState(WALSNDSTATE_STARTUP);
960 :
961 : Assert(streamingDoneSending && streamingDoneReceiving);
962 : }
963 :
964 294 : if (cmd->slotname)
965 262 : ReplicationSlotRelease();
966 :
967 : /*
968 : * Copy is finished now. Send a single-row result set indicating the next
969 : * timeline.
970 : */
971 294 : if (sendTimeLineIsHistoric)
972 : {
973 : char startpos_str[8 + 1 + 8 + 1];
974 : DestReceiver *dest;
975 : TupOutputState *tstate;
976 : TupleDesc tupdesc;
977 : Datum values[2];
978 26 : bool nulls[2] = {0};
979 :
980 26 : snprintf(startpos_str, sizeof(startpos_str), "%X/%X",
981 26 : LSN_FORMAT_ARGS(sendTimeLineValidUpto));
982 :
983 26 : dest = CreateDestReceiver(DestRemoteSimple);
984 :
985 : /*
986 : * Need a tuple descriptor representing two columns. int8 may seem
987 : * like a surprising data type for this, but in theory int4 would not
988 : * be wide enough for this, as TimeLineID is unsigned.
989 : */
990 26 : tupdesc = CreateTemplateTupleDesc(2);
991 26 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "next_tli",
992 : INT8OID, -1, 0);
993 26 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "next_tli_startpos",
994 : TEXTOID, -1, 0);
995 :
996 : /* prepare for projection of tuple */
997 26 : tstate = begin_tup_output_tupdesc(dest, tupdesc, &TTSOpsVirtual);
998 :
999 26 : values[0] = Int64GetDatum((int64) sendTimeLineNextTLI);
1000 26 : values[1] = CStringGetTextDatum(startpos_str);
1001 :
1002 : /* send it to dest */
1003 26 : do_tup_output(tstate, values, nulls);
1004 :
1005 26 : end_tup_output(tstate);
1006 : }
1007 :
1008 : /* Send CommandComplete message */
1009 294 : EndReplicationCommand("START_STREAMING");
1010 294 : }
1011 :
1012 : /*
1013 : * XLogReaderRoutine->page_read callback for logical decoding contexts, as a
1014 : * walsender process.
1015 : *
1016 : * Inside the walsender we can do better than read_local_xlog_page,
1017 : * which has to do a plain sleep/busy loop, because the walsender's latch gets
1018 : * set every time WAL is flushed.
1019 : */
1020 : static int
1021 37000 : logical_read_xlog_page(XLogReaderState *state, XLogRecPtr targetPagePtr, int reqLen,
1022 : XLogRecPtr targetRecPtr, char *cur_page)
1023 : {
1024 : XLogRecPtr flushptr;
1025 : int count;
1026 : WALReadError errinfo;
1027 : XLogSegNo segno;
1028 : TimeLineID currTLI;
1029 :
1030 : /*
1031 : * Make sure we have enough WAL available before retrieving the current
1032 : * timeline.
1033 : */
1034 37000 : flushptr = WalSndWaitForWal(targetPagePtr + reqLen);
1035 :
1036 : /* Fail if not enough (implies we are going to shut down) */
1037 36674 : if (flushptr < targetPagePtr + reqLen)
1038 7064 : return -1;
1039 :
1040 : /*
1041 : * Since logical decoding is also permitted on a standby server, we need
1042 : * to check if the server is in recovery to decide how to get the current
1043 : * timeline ID (so that it also covers the promotion or timeline change
1044 : * cases). We must determine am_cascading_walsender after waiting for the
1045 : * required WAL so that it is correct when the walsender wakes up after a
1046 : * promotion.
1047 : */
1048 29610 : am_cascading_walsender = RecoveryInProgress();
1049 :
1050 29610 : if (am_cascading_walsender)
1051 628 : GetXLogReplayRecPtr(&currTLI);
1052 : else
1053 28982 : currTLI = GetWALInsertionTimeLine();
1054 :
1055 29610 : XLogReadDetermineTimeline(state, targetPagePtr, reqLen, currTLI);
1056 29610 : sendTimeLineIsHistoric = (state->currTLI != currTLI);
1057 29610 : sendTimeLine = state->currTLI;
1058 29610 : sendTimeLineValidUpto = state->currTLIValidUntil;
1059 29610 : sendTimeLineNextTLI = state->nextTLI;
1060 :
1061 29610 : if (targetPagePtr + XLOG_BLCKSZ <= flushptr)
1062 25894 : count = XLOG_BLCKSZ; /* more than one block available */
1063 : else
1064 3716 : count = flushptr - targetPagePtr; /* part of the page available */
1065 :
1066 : /* now actually read the data, we know it's there */
1067 29610 : if (!WALRead(state,
1068 : cur_page,
1069 : targetPagePtr,
1070 : count,
1071 : currTLI, /* Pass the current TLI because only
1072 : * WalSndSegmentOpen controls whether new TLI
1073 : * is needed. */
1074 : &errinfo))
1075 0 : WALReadRaiseError(&errinfo);
1076 :
1077 : /*
1078 : * After reading into the buffer, check that what we read was valid. We do
1079 : * this after reading, because even though the segment was present when we
1080 : * opened it, it might get recycled or removed while we read it. The
1081 : * read() succeeds in that case, but the data we tried to read might
1082 : * already have been overwritten with new WAL records.
1083 : */
1084 29610 : XLByteToSeg(targetPagePtr, segno, state->segcxt.ws_segsize);
1085 29610 : CheckXLogRemoved(segno, state->seg.ws_tli);
1086 :
1087 29610 : return count;
1088 : }
1089 :
1090 : /*
1091 : * Process extra options given to CREATE_REPLICATION_SLOT.
1092 : */
1093 : static void
1094 888 : parseCreateReplSlotOptions(CreateReplicationSlotCmd *cmd,
1095 : bool *reserve_wal,
1096 : CRSSnapshotAction *snapshot_action,
1097 : bool *two_phase, bool *failover)
1098 : {
1099 : ListCell *lc;
1100 888 : bool snapshot_action_given = false;
1101 888 : bool reserve_wal_given = false;
1102 888 : bool two_phase_given = false;
1103 888 : bool failover_given = false;
1104 :
1105 : /* Parse options */
1106 1790 : foreach(lc, cmd->options)
1107 : {
1108 902 : DefElem *defel = (DefElem *) lfirst(lc);
1109 :
1110 902 : if (strcmp(defel->defname, "snapshot") == 0)
1111 : {
1112 : char *action;
1113 :
1114 620 : if (snapshot_action_given || cmd->kind != REPLICATION_KIND_LOGICAL)
1115 0 : ereport(ERROR,
1116 : (errcode(ERRCODE_SYNTAX_ERROR),
1117 : errmsg("conflicting or redundant options")));
1118 :
1119 620 : action = defGetString(defel);
1120 620 : snapshot_action_given = true;
1121 :
1122 620 : if (strcmp(action, "export") == 0)
1123 0 : *snapshot_action = CRS_EXPORT_SNAPSHOT;
1124 620 : else if (strcmp(action, "nothing") == 0)
1125 252 : *snapshot_action = CRS_NOEXPORT_SNAPSHOT;
1126 368 : else if (strcmp(action, "use") == 0)
1127 368 : *snapshot_action = CRS_USE_SNAPSHOT;
1128 : else
1129 0 : ereport(ERROR,
1130 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1131 : errmsg("unrecognized value for CREATE_REPLICATION_SLOT option \"%s\": \"%s\"",
1132 : defel->defname, action)));
1133 : }
1134 282 : else if (strcmp(defel->defname, "reserve_wal") == 0)
1135 : {
1136 266 : if (reserve_wal_given || cmd->kind != REPLICATION_KIND_PHYSICAL)
1137 0 : ereport(ERROR,
1138 : (errcode(ERRCODE_SYNTAX_ERROR),
1139 : errmsg("conflicting or redundant options")));
1140 :
1141 266 : reserve_wal_given = true;
1142 266 : *reserve_wal = defGetBoolean(defel);
1143 : }
1144 16 : else if (strcmp(defel->defname, "two_phase") == 0)
1145 : {
1146 4 : if (two_phase_given || cmd->kind != REPLICATION_KIND_LOGICAL)
1147 0 : ereport(ERROR,
1148 : (errcode(ERRCODE_SYNTAX_ERROR),
1149 : errmsg("conflicting or redundant options")));
1150 4 : two_phase_given = true;
1151 4 : *two_phase = defGetBoolean(defel);
1152 : }
1153 12 : else if (strcmp(defel->defname, "failover") == 0)
1154 : {
1155 12 : if (failover_given || cmd->kind != REPLICATION_KIND_LOGICAL)
1156 0 : ereport(ERROR,
1157 : (errcode(ERRCODE_SYNTAX_ERROR),
1158 : errmsg("conflicting or redundant options")));
1159 12 : failover_given = true;
1160 12 : *failover = defGetBoolean(defel);
1161 : }
1162 : else
1163 0 : elog(ERROR, "unrecognized option: %s", defel->defname);
1164 : }
1165 888 : }
1166 :
1167 : /*
1168 : * Create a new replication slot.
1169 : */
1170 : static void
1171 888 : CreateReplicationSlot(CreateReplicationSlotCmd *cmd)
1172 : {
1173 888 : const char *snapshot_name = NULL;
1174 : char xloc[MAXFNAMELEN];
1175 : char *slot_name;
1176 888 : bool reserve_wal = false;
1177 888 : bool two_phase = false;
1178 888 : bool failover = false;
1179 888 : CRSSnapshotAction snapshot_action = CRS_EXPORT_SNAPSHOT;
1180 : DestReceiver *dest;
1181 : TupOutputState *tstate;
1182 : TupleDesc tupdesc;
1183 : Datum values[4];
1184 888 : bool nulls[4] = {0};
1185 :
1186 : Assert(!MyReplicationSlot);
1187 :
1188 888 : parseCreateReplSlotOptions(cmd, &reserve_wal, &snapshot_action, &two_phase,
1189 : &failover);
1190 :
1191 888 : if (cmd->kind == REPLICATION_KIND_PHYSICAL)
1192 : {
1193 268 : ReplicationSlotCreate(cmd->slotname, false,
1194 268 : cmd->temporary ? RS_TEMPORARY : RS_PERSISTENT,
1195 : false, false, false);
1196 :
1197 266 : if (reserve_wal)
1198 : {
1199 264 : ReplicationSlotReserveWal();
1200 :
1201 264 : ReplicationSlotMarkDirty();
1202 :
1203 : /* Write this slot to disk if it's a permanent one. */
1204 264 : if (!cmd->temporary)
1205 6 : ReplicationSlotSave();
1206 : }
1207 : }
1208 : else
1209 : {
1210 : LogicalDecodingContext *ctx;
1211 620 : bool need_full_snapshot = false;
1212 :
1213 : Assert(cmd->kind == REPLICATION_KIND_LOGICAL);
1214 :
1215 620 : CheckLogicalDecodingRequirements();
1216 :
1217 : /*
1218 : * Initially create persistent slot as ephemeral - that allows us to
1219 : * nicely handle errors during initialization because it'll get
1220 : * dropped if this transaction fails. We'll make it persistent at the
1221 : * end. Temporary slots can be created as temporary from beginning as
1222 : * they get dropped on error as well.
1223 : */
1224 620 : ReplicationSlotCreate(cmd->slotname, true,
1225 620 : cmd->temporary ? RS_TEMPORARY : RS_EPHEMERAL,
1226 : two_phase, failover, false);
1227 :
1228 : /*
1229 : * Do options check early so that we can bail before calling the
1230 : * DecodingContextFindStartpoint which can take long time.
1231 : */
1232 620 : if (snapshot_action == CRS_EXPORT_SNAPSHOT)
1233 : {
1234 0 : if (IsTransactionBlock())
1235 0 : ereport(ERROR,
1236 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1237 : (errmsg("%s must not be called inside a transaction",
1238 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'export')")));
1239 :
1240 0 : need_full_snapshot = true;
1241 : }
1242 620 : else if (snapshot_action == CRS_USE_SNAPSHOT)
1243 : {
1244 368 : if (!IsTransactionBlock())
1245 0 : ereport(ERROR,
1246 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1247 : (errmsg("%s must be called inside a transaction",
1248 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1249 :
1250 368 : if (XactIsoLevel != XACT_REPEATABLE_READ)
1251 0 : ereport(ERROR,
1252 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1253 : (errmsg("%s must be called in REPEATABLE READ isolation mode transaction",
1254 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1255 368 : if (!XactReadOnly)
1256 0 : ereport(ERROR,
1257 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1258 : (errmsg("%s must be called in a read-only transaction",
1259 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1260 :
1261 368 : if (FirstSnapshotSet)
1262 0 : ereport(ERROR,
1263 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1264 : (errmsg("%s must be called before any query",
1265 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1266 :
1267 368 : if (IsSubTransaction())
1268 0 : ereport(ERROR,
1269 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1270 : (errmsg("%s must not be called in a subtransaction",
1271 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1272 :
1273 368 : need_full_snapshot = true;
1274 : }
1275 :
1276 620 : ctx = CreateInitDecodingContext(cmd->plugin, NIL, need_full_snapshot,
1277 : InvalidXLogRecPtr,
1278 620 : XL_ROUTINE(.page_read = logical_read_xlog_page,
1279 : .segment_open = WalSndSegmentOpen,
1280 : .segment_close = wal_segment_close),
1281 : WalSndPrepareWrite, WalSndWriteData,
1282 : WalSndUpdateProgress);
1283 :
1284 : /*
1285 : * Signal that we don't need the timeout mechanism. We're just
1286 : * creating the replication slot and don't yet accept feedback
1287 : * messages or send keepalives. As we possibly need to wait for
1288 : * further WAL the walsender would otherwise possibly be killed too
1289 : * soon.
1290 : */
1291 620 : last_reply_timestamp = 0;
1292 :
1293 : /* build initial snapshot, might take a while */
1294 620 : DecodingContextFindStartpoint(ctx);
1295 :
1296 : /*
1297 : * Export or use the snapshot if we've been asked to do so.
1298 : *
1299 : * NB. We will convert the snapbuild.c kind of snapshot to normal
1300 : * snapshot when doing this.
1301 : */
1302 620 : if (snapshot_action == CRS_EXPORT_SNAPSHOT)
1303 : {
1304 0 : snapshot_name = SnapBuildExportSnapshot(ctx->snapshot_builder);
1305 : }
1306 620 : else if (snapshot_action == CRS_USE_SNAPSHOT)
1307 : {
1308 : Snapshot snap;
1309 :
1310 368 : snap = SnapBuildInitialSnapshot(ctx->snapshot_builder);
1311 368 : RestoreTransactionSnapshot(snap, MyProc);
1312 : }
1313 :
1314 : /* don't need the decoding context anymore */
1315 620 : FreeDecodingContext(ctx);
1316 :
1317 620 : if (!cmd->temporary)
1318 620 : ReplicationSlotPersist();
1319 : }
1320 :
1321 886 : snprintf(xloc, sizeof(xloc), "%X/%X",
1322 886 : LSN_FORMAT_ARGS(MyReplicationSlot->data.confirmed_flush));
1323 :
1324 886 : dest = CreateDestReceiver(DestRemoteSimple);
1325 :
1326 : /*----------
1327 : * Need a tuple descriptor representing four columns:
1328 : * - first field: the slot name
1329 : * - second field: LSN at which we became consistent
1330 : * - third field: exported snapshot's name
1331 : * - fourth field: output plugin
1332 : */
1333 886 : tupdesc = CreateTemplateTupleDesc(4);
1334 886 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "slot_name",
1335 : TEXTOID, -1, 0);
1336 886 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "consistent_point",
1337 : TEXTOID, -1, 0);
1338 886 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 3, "snapshot_name",
1339 : TEXTOID, -1, 0);
1340 886 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 4, "output_plugin",
1341 : TEXTOID, -1, 0);
1342 :
1343 : /* prepare for projection of tuples */
1344 886 : tstate = begin_tup_output_tupdesc(dest, tupdesc, &TTSOpsVirtual);
1345 :
1346 : /* slot_name */
1347 886 : slot_name = NameStr(MyReplicationSlot->data.name);
1348 886 : values[0] = CStringGetTextDatum(slot_name);
1349 :
1350 : /* consistent wal location */
1351 886 : values[1] = CStringGetTextDatum(xloc);
1352 :
1353 : /* snapshot name, or NULL if none */
1354 886 : if (snapshot_name != NULL)
1355 0 : values[2] = CStringGetTextDatum(snapshot_name);
1356 : else
1357 886 : nulls[2] = true;
1358 :
1359 : /* plugin, or NULL if none */
1360 886 : if (cmd->plugin != NULL)
1361 620 : values[3] = CStringGetTextDatum(cmd->plugin);
1362 : else
1363 266 : nulls[3] = true;
1364 :
1365 : /* send it to dest */
1366 886 : do_tup_output(tstate, values, nulls);
1367 886 : end_tup_output(tstate);
1368 :
1369 886 : ReplicationSlotRelease();
1370 886 : }
1371 :
1372 : /*
1373 : * Get rid of a replication slot that is no longer wanted.
1374 : */
1375 : static void
1376 494 : DropReplicationSlot(DropReplicationSlotCmd *cmd)
1377 : {
1378 494 : ReplicationSlotDrop(cmd->slotname, !cmd->wait);
1379 494 : }
1380 :
1381 : /*
1382 : * Change the definition of a replication slot.
1383 : */
1384 : static void
1385 12 : AlterReplicationSlot(AlterReplicationSlotCmd *cmd)
1386 : {
1387 12 : bool failover_given = false;
1388 12 : bool two_phase_given = false;
1389 : bool failover;
1390 : bool two_phase;
1391 :
1392 : /* Parse options */
1393 36 : foreach_ptr(DefElem, defel, cmd->options)
1394 : {
1395 12 : if (strcmp(defel->defname, "failover") == 0)
1396 : {
1397 10 : if (failover_given)
1398 0 : ereport(ERROR,
1399 : (errcode(ERRCODE_SYNTAX_ERROR),
1400 : errmsg("conflicting or redundant options")));
1401 10 : failover_given = true;
1402 10 : failover = defGetBoolean(defel);
1403 : }
1404 2 : else if (strcmp(defel->defname, "two_phase") == 0)
1405 : {
1406 2 : if (two_phase_given)
1407 0 : ereport(ERROR,
1408 : (errcode(ERRCODE_SYNTAX_ERROR),
1409 : errmsg("conflicting or redundant options")));
1410 2 : two_phase_given = true;
1411 2 : two_phase = defGetBoolean(defel);
1412 : }
1413 : else
1414 0 : elog(ERROR, "unrecognized option: %s", defel->defname);
1415 : }
1416 :
1417 12 : ReplicationSlotAlter(cmd->slotname,
1418 : failover_given ? &failover : NULL,
1419 : two_phase_given ? &two_phase : NULL);
1420 8 : }
1421 :
1422 : /*
1423 : * Load previously initiated logical slot and prepare for sending data (via
1424 : * WalSndLoop).
1425 : */
1426 : static void
1427 730 : StartLogicalReplication(StartReplicationCmd *cmd)
1428 : {
1429 : StringInfoData buf;
1430 : QueryCompletion qc;
1431 :
1432 : /* make sure that our requirements are still fulfilled */
1433 730 : CheckLogicalDecodingRequirements();
1434 :
1435 : Assert(!MyReplicationSlot);
1436 :
1437 726 : ReplicationSlotAcquire(cmd->slotname, true);
1438 :
1439 : /*
1440 : * Force a disconnect, so that the decoding code doesn't need to care
1441 : * about an eventual switch from running in recovery, to running in a
1442 : * normal environment. Client code is expected to handle reconnects.
1443 : */
1444 726 : if (am_cascading_walsender && !RecoveryInProgress())
1445 : {
1446 0 : ereport(LOG,
1447 : (errmsg("terminating walsender process after promotion")));
1448 0 : got_STOPPING = true;
1449 : }
1450 :
1451 : /*
1452 : * Create our decoding context, making it start at the previously ack'ed
1453 : * position.
1454 : *
1455 : * Do this before sending a CopyBothResponse message, so that any errors
1456 : * are reported early.
1457 : */
1458 714 : logical_decoding_ctx =
1459 726 : CreateDecodingContext(cmd->startpoint, cmd->options, false,
1460 726 : XL_ROUTINE(.page_read = logical_read_xlog_page,
1461 : .segment_open = WalSndSegmentOpen,
1462 : .segment_close = wal_segment_close),
1463 : WalSndPrepareWrite, WalSndWriteData,
1464 : WalSndUpdateProgress);
1465 714 : xlogreader = logical_decoding_ctx->reader;
1466 :
1467 714 : WalSndSetState(WALSNDSTATE_CATCHUP);
1468 :
1469 : /* Send a CopyBothResponse message, and start streaming */
1470 714 : pq_beginmessage(&buf, PqMsg_CopyBothResponse);
1471 714 : pq_sendbyte(&buf, 0);
1472 714 : pq_sendint16(&buf, 0);
1473 714 : pq_endmessage(&buf);
1474 714 : pq_flush();
1475 :
1476 : /* Start reading WAL from the oldest required WAL. */
1477 714 : XLogBeginRead(logical_decoding_ctx->reader,
1478 714 : MyReplicationSlot->data.restart_lsn);
1479 :
1480 : /*
1481 : * Report the location after which we'll send out further commits as the
1482 : * current sentPtr.
1483 : */
1484 714 : sentPtr = MyReplicationSlot->data.confirmed_flush;
1485 :
1486 : /* Also update the sent position status in shared memory */
1487 714 : SpinLockAcquire(&MyWalSnd->mutex);
1488 714 : MyWalSnd->sentPtr = MyReplicationSlot->data.restart_lsn;
1489 714 : SpinLockRelease(&MyWalSnd->mutex);
1490 :
1491 714 : replication_active = true;
1492 :
1493 714 : SyncRepInitConfig();
1494 :
1495 : /* Main loop of walsender */
1496 714 : WalSndLoop(XLogSendLogical);
1497 :
1498 362 : FreeDecodingContext(logical_decoding_ctx);
1499 362 : ReplicationSlotRelease();
1500 :
1501 362 : replication_active = false;
1502 362 : if (got_STOPPING)
1503 0 : proc_exit(0);
1504 362 : WalSndSetState(WALSNDSTATE_STARTUP);
1505 :
1506 : /* Get out of COPY mode (CommandComplete). */
1507 362 : SetQueryCompletion(&qc, CMDTAG_COPY, 0);
1508 362 : EndCommand(&qc, DestRemote, false);
1509 362 : }
1510 :
1511 : /*
1512 : * LogicalDecodingContext 'prepare_write' callback.
1513 : *
1514 : * Prepare a write into a StringInfo.
1515 : *
1516 : * Don't do anything lasting in here, it's quite possible that nothing will be done
1517 : * with the data.
1518 : */
1519 : static void
1520 369746 : WalSndPrepareWrite(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid, bool last_write)
1521 : {
1522 : /* can't have sync rep confused by sending the same LSN several times */
1523 369746 : if (!last_write)
1524 742 : lsn = InvalidXLogRecPtr;
1525 :
1526 369746 : resetStringInfo(ctx->out);
1527 :
1528 369746 : pq_sendbyte(ctx->out, 'w');
1529 369746 : pq_sendint64(ctx->out, lsn); /* dataStart */
1530 369746 : pq_sendint64(ctx->out, lsn); /* walEnd */
1531 :
1532 : /*
1533 : * Fill out the sendtime later, just as it's done in XLogSendPhysical, but
1534 : * reserve space here.
1535 : */
1536 369746 : pq_sendint64(ctx->out, 0); /* sendtime */
1537 369746 : }
1538 :
1539 : /*
1540 : * LogicalDecodingContext 'write' callback.
1541 : *
1542 : * Actually write out data previously prepared by WalSndPrepareWrite out to
1543 : * the network. Take as long as needed, but process replies from the other
1544 : * side and check timeouts during that.
1545 : */
1546 : static void
1547 369746 : WalSndWriteData(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid,
1548 : bool last_write)
1549 : {
1550 : TimestampTz now;
1551 :
1552 : /*
1553 : * Fill the send timestamp last, so that it is taken as late as possible.
1554 : * This is somewhat ugly, but the protocol is set as it's already used for
1555 : * several releases by streaming physical replication.
1556 : */
1557 369746 : resetStringInfo(&tmpbuf);
1558 369746 : now = GetCurrentTimestamp();
1559 369746 : pq_sendint64(&tmpbuf, now);
1560 369746 : memcpy(&ctx->out->data[1 + sizeof(int64) + sizeof(int64)],
1561 369746 : tmpbuf.data, sizeof(int64));
1562 :
1563 : /* output previously gathered data in a CopyData packet */
1564 369746 : pq_putmessage_noblock('d', ctx->out->data, ctx->out->len);
1565 :
1566 369746 : CHECK_FOR_INTERRUPTS();
1567 :
1568 : /* Try to flush pending output to the client */
1569 369746 : if (pq_flush_if_writable() != 0)
1570 2 : WalSndShutdown();
1571 :
1572 : /* Try taking fast path unless we get too close to walsender timeout. */
1573 369744 : if (now < TimestampTzPlusMilliseconds(last_reply_timestamp,
1574 369744 : wal_sender_timeout / 2) &&
1575 369744 : !pq_is_send_pending())
1576 : {
1577 369644 : return;
1578 : }
1579 :
1580 : /* If we have pending write here, go to slow path */
1581 100 : ProcessPendingWrites();
1582 : }
1583 :
1584 : /*
1585 : * Wait until there is no pending write. Also process replies from the other
1586 : * side and check timeouts during that.
1587 : */
1588 : static void
1589 240 : ProcessPendingWrites(void)
1590 : {
1591 : for (;;)
1592 140 : {
1593 : long sleeptime;
1594 :
1595 : /* Check for input from the client */
1596 240 : ProcessRepliesIfAny();
1597 :
1598 : /* die if timeout was reached */
1599 240 : WalSndCheckTimeOut();
1600 :
1601 : /* Send keepalive if the time has come */
1602 240 : WalSndKeepaliveIfNecessary();
1603 :
1604 240 : if (!pq_is_send_pending())
1605 100 : break;
1606 :
1607 140 : sleeptime = WalSndComputeSleeptime(GetCurrentTimestamp());
1608 :
1609 : /* Sleep until something happens or we time out */
1610 140 : WalSndWait(WL_SOCKET_WRITEABLE | WL_SOCKET_READABLE, sleeptime,
1611 : WAIT_EVENT_WAL_SENDER_WRITE_DATA);
1612 :
1613 : /* Clear any already-pending wakeups */
1614 140 : ResetLatch(MyLatch);
1615 :
1616 140 : CHECK_FOR_INTERRUPTS();
1617 :
1618 : /* Process any requests or signals received recently */
1619 140 : if (ConfigReloadPending)
1620 : {
1621 0 : ConfigReloadPending = false;
1622 0 : ProcessConfigFile(PGC_SIGHUP);
1623 0 : SyncRepInitConfig();
1624 : }
1625 :
1626 : /* Try to flush pending output to the client */
1627 140 : if (pq_flush_if_writable() != 0)
1628 0 : WalSndShutdown();
1629 : }
1630 :
1631 : /* reactivate latch so WalSndLoop knows to continue */
1632 100 : SetLatch(MyLatch);
1633 100 : }
1634 :
1635 : /*
1636 : * LogicalDecodingContext 'update_progress' callback.
1637 : *
1638 : * Write the current position to the lag tracker (see XLogSendPhysical).
1639 : *
1640 : * When skipping empty transactions, send a keepalive message if necessary.
1641 : */
1642 : static void
1643 4700 : WalSndUpdateProgress(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid,
1644 : bool skipped_xact)
1645 : {
1646 : static TimestampTz sendTime = 0;
1647 4700 : TimestampTz now = GetCurrentTimestamp();
1648 4700 : bool pending_writes = false;
1649 4700 : bool end_xact = ctx->end_xact;
1650 :
1651 : /*
1652 : * Track lag no more than once per WALSND_LOGICAL_LAG_TRACK_INTERVAL_MS to
1653 : * avoid flooding the lag tracker when we commit frequently.
1654 : *
1655 : * We don't have a mechanism to get the ack for any LSN other than end
1656 : * xact LSN from the downstream. So, we track lag only for end of
1657 : * transaction LSN.
1658 : */
1659 : #define WALSND_LOGICAL_LAG_TRACK_INTERVAL_MS 1000
1660 4700 : if (end_xact && TimestampDifferenceExceeds(sendTime, now,
1661 : WALSND_LOGICAL_LAG_TRACK_INTERVAL_MS))
1662 : {
1663 360 : LagTrackerWrite(lsn, now);
1664 360 : sendTime = now;
1665 : }
1666 :
1667 : /*
1668 : * When skipping empty transactions in synchronous replication, we send a
1669 : * keepalive message to avoid delaying such transactions.
1670 : *
1671 : * It is okay to check sync_standbys_defined flag without lock here as in
1672 : * the worst case we will just send an extra keepalive message when it is
1673 : * really not required.
1674 : */
1675 4700 : if (skipped_xact &&
1676 594 : SyncRepRequested() &&
1677 594 : ((volatile WalSndCtlData *) WalSndCtl)->sync_standbys_defined)
1678 : {
1679 0 : WalSndKeepalive(false, lsn);
1680 :
1681 : /* Try to flush pending output to the client */
1682 0 : if (pq_flush_if_writable() != 0)
1683 0 : WalSndShutdown();
1684 :
1685 : /* If we have pending write here, make sure it's actually flushed */
1686 0 : if (pq_is_send_pending())
1687 0 : pending_writes = true;
1688 : }
1689 :
1690 : /*
1691 : * Process pending writes if any or try to send a keepalive if required.
1692 : * We don't need to try sending keep alive messages at the transaction end
1693 : * as that will be done at a later point in time. This is required only
1694 : * for large transactions where we don't send any changes to the
1695 : * downstream and the receiver can timeout due to that.
1696 : */
1697 4700 : if (pending_writes || (!end_xact &&
1698 3076 : now >= TimestampTzPlusMilliseconds(last_reply_timestamp,
1699 : wal_sender_timeout / 2)))
1700 0 : ProcessPendingWrites();
1701 4700 : }
1702 :
1703 : /*
1704 : * Wake up the logical walsender processes with logical failover slots if the
1705 : * currently acquired physical slot is specified in synchronized_standby_slots GUC.
1706 : */
1707 : void
1708 8310 : PhysicalWakeupLogicalWalSnd(void)
1709 : {
1710 : Assert(MyReplicationSlot && SlotIsPhysical(MyReplicationSlot));
1711 :
1712 : /*
1713 : * If we are running in a standby, there is no need to wake up walsenders.
1714 : * This is because we do not support syncing slots to cascading standbys,
1715 : * so, there are no walsenders waiting for standbys to catch up.
1716 : */
1717 8310 : if (RecoveryInProgress())
1718 106 : return;
1719 :
1720 8204 : if (SlotExistsInSyncStandbySlots(NameStr(MyReplicationSlot->data.name)))
1721 10 : ConditionVariableBroadcast(&WalSndCtl->wal_confirm_rcv_cv);
1722 : }
1723 :
1724 : /*
1725 : * Returns true if not all standbys have caught up to the flushed position
1726 : * (flushed_lsn) when the current acquired slot is a logical failover
1727 : * slot and we are streaming; otherwise, returns false.
1728 : *
1729 : * If returning true, the function sets the appropriate wait event in
1730 : * wait_event; otherwise, wait_event is set to 0.
1731 : */
1732 : static bool
1733 36364 : NeedToWaitForStandbys(XLogRecPtr flushed_lsn, uint32 *wait_event)
1734 : {
1735 36364 : int elevel = got_STOPPING ? ERROR : WARNING;
1736 : bool failover_slot;
1737 :
1738 36364 : failover_slot = (replication_active && MyReplicationSlot->data.failover);
1739 :
1740 : /*
1741 : * Note that after receiving the shutdown signal, an ERROR is reported if
1742 : * any slots are dropped, invalidated, or inactive. This measure is taken
1743 : * to prevent the walsender from waiting indefinitely.
1744 : */
1745 36364 : if (failover_slot && !StandbySlotsHaveCaughtup(flushed_lsn, elevel))
1746 : {
1747 8 : *wait_event = WAIT_EVENT_WAIT_FOR_STANDBY_CONFIRMATION;
1748 8 : return true;
1749 : }
1750 :
1751 36356 : *wait_event = 0;
1752 36356 : return false;
1753 : }
1754 :
1755 : /*
1756 : * Returns true if we need to wait for WALs to be flushed to disk, or if not
1757 : * all standbys have caught up to the flushed position (flushed_lsn) when the
1758 : * current acquired slot is a logical failover slot and we are
1759 : * streaming; otherwise, returns false.
1760 : *
1761 : * If returning true, the function sets the appropriate wait event in
1762 : * wait_event; otherwise, wait_event is set to 0.
1763 : */
1764 : static bool
1765 44894 : NeedToWaitForWal(XLogRecPtr target_lsn, XLogRecPtr flushed_lsn,
1766 : uint32 *wait_event)
1767 : {
1768 : /* Check if we need to wait for WALs to be flushed to disk */
1769 44894 : if (target_lsn > flushed_lsn)
1770 : {
1771 15278 : *wait_event = WAIT_EVENT_WAL_SENDER_WAIT_FOR_WAL;
1772 15278 : return true;
1773 : }
1774 :
1775 : /* Check if the standby slots have caught up to the flushed position */
1776 29616 : return NeedToWaitForStandbys(flushed_lsn, wait_event);
1777 : }
1778 :
1779 : /*
1780 : * Wait till WAL < loc is flushed to disk so it can be safely sent to client.
1781 : *
1782 : * If the walsender holds a logical failover slot, we also wait for all the
1783 : * specified streaming replication standby servers to confirm receipt of WAL
1784 : * up to RecentFlushPtr. It is beneficial to wait here for the confirmation
1785 : * up to RecentFlushPtr rather than waiting before transmitting each change
1786 : * to logical subscribers, which is already covered by RecentFlushPtr.
1787 : *
1788 : * Returns end LSN of flushed WAL. Normally this will be >= loc, but if we
1789 : * detect a shutdown request (either from postmaster or client) we will return
1790 : * early, so caller must always check.
1791 : */
1792 : static XLogRecPtr
1793 37000 : WalSndWaitForWal(XLogRecPtr loc)
1794 : {
1795 : int wakeEvents;
1796 37000 : uint32 wait_event = 0;
1797 : static XLogRecPtr RecentFlushPtr = InvalidXLogRecPtr;
1798 :
1799 : /*
1800 : * Fast path to avoid acquiring the spinlock in case we already know we
1801 : * have enough WAL available and all the standby servers have confirmed
1802 : * receipt of WAL up to RecentFlushPtr. This is particularly interesting
1803 : * if we're far behind.
1804 : */
1805 37000 : if (!XLogRecPtrIsInvalid(RecentFlushPtr) &&
1806 36014 : !NeedToWaitForWal(loc, RecentFlushPtr, &wait_event))
1807 26148 : return RecentFlushPtr;
1808 :
1809 : /*
1810 : * Within the loop, we wait for the necessary WALs to be flushed to disk
1811 : * first, followed by waiting for standbys to catch up if there are enough
1812 : * WALs (see NeedToWaitForWal()) or upon receiving the shutdown signal.
1813 : */
1814 : for (;;)
1815 5102 : {
1816 15954 : bool wait_for_standby_at_stop = false;
1817 : long sleeptime;
1818 :
1819 : /* Clear any already-pending wakeups */
1820 15954 : ResetLatch(MyLatch);
1821 :
1822 15954 : CHECK_FOR_INTERRUPTS();
1823 :
1824 : /* Process any requests or signals received recently */
1825 15942 : if (ConfigReloadPending)
1826 : {
1827 14 : ConfigReloadPending = false;
1828 14 : ProcessConfigFile(PGC_SIGHUP);
1829 14 : SyncRepInitConfig();
1830 : }
1831 :
1832 : /* Check for input from the client */
1833 15942 : ProcessRepliesIfAny();
1834 :
1835 : /*
1836 : * If we're shutting down, trigger pending WAL to be written out,
1837 : * otherwise we'd possibly end up waiting for WAL that never gets
1838 : * written, because walwriter has shut down already.
1839 : */
1840 15628 : if (got_STOPPING)
1841 6748 : XLogBackgroundFlush();
1842 :
1843 : /*
1844 : * To avoid the scenario where standbys need to catch up to a newer
1845 : * WAL location in each iteration, we update our idea of the currently
1846 : * flushed position only if we are not waiting for standbys to catch
1847 : * up.
1848 : */
1849 15628 : if (wait_event != WAIT_EVENT_WAIT_FOR_STANDBY_CONFIRMATION)
1850 : {
1851 15620 : if (!RecoveryInProgress())
1852 14838 : RecentFlushPtr = GetFlushRecPtr(NULL);
1853 : else
1854 782 : RecentFlushPtr = GetXLogReplayRecPtr(NULL);
1855 : }
1856 :
1857 : /*
1858 : * If postmaster asked us to stop and the standby slots have caught up
1859 : * to the flushed position, don't wait anymore.
1860 : *
1861 : * It's important to do this check after the recomputation of
1862 : * RecentFlushPtr, so we can send all remaining data before shutting
1863 : * down.
1864 : */
1865 15628 : if (got_STOPPING)
1866 : {
1867 6748 : if (NeedToWaitForStandbys(RecentFlushPtr, &wait_event))
1868 0 : wait_for_standby_at_stop = true;
1869 : else
1870 6748 : break;
1871 : }
1872 :
1873 : /*
1874 : * We only send regular messages to the client for full decoded
1875 : * transactions, but a synchronous replication and walsender shutdown
1876 : * possibly are waiting for a later location. So, before sleeping, we
1877 : * send a ping containing the flush location. If the receiver is
1878 : * otherwise idle, this keepalive will trigger a reply. Processing the
1879 : * reply will update these MyWalSnd locations.
1880 : */
1881 8880 : if (MyWalSnd->flush < sentPtr &&
1882 4716 : MyWalSnd->write < sentPtr &&
1883 3504 : !waiting_for_ping_response)
1884 3504 : WalSndKeepalive(false, InvalidXLogRecPtr);
1885 :
1886 : /*
1887 : * Exit the loop if already caught up and doesn't need to wait for
1888 : * standby slots.
1889 : */
1890 8880 : if (!wait_for_standby_at_stop &&
1891 8880 : !NeedToWaitForWal(loc, RecentFlushPtr, &wait_event))
1892 3460 : break;
1893 :
1894 : /*
1895 : * Waiting for new WAL or waiting for standbys to catch up. Since we
1896 : * need to wait, we're now caught up.
1897 : */
1898 5420 : WalSndCaughtUp = true;
1899 :
1900 : /*
1901 : * Try to flush any pending output to the client.
1902 : */
1903 5420 : if (pq_flush_if_writable() != 0)
1904 0 : WalSndShutdown();
1905 :
1906 : /*
1907 : * If we have received CopyDone from the client, sent CopyDone
1908 : * ourselves, and the output buffer is empty, it's time to exit
1909 : * streaming, so fail the current WAL fetch request.
1910 : */
1911 5420 : if (streamingDoneReceiving && streamingDoneSending &&
1912 318 : !pq_is_send_pending())
1913 318 : break;
1914 :
1915 : /* die if timeout was reached */
1916 5102 : WalSndCheckTimeOut();
1917 :
1918 : /* Send keepalive if the time has come */
1919 5102 : WalSndKeepaliveIfNecessary();
1920 :
1921 : /*
1922 : * Sleep until something happens or we time out. Also wait for the
1923 : * socket becoming writable, if there's still pending output.
1924 : * Otherwise we might sit on sendable output data while waiting for
1925 : * new WAL to be generated. (But if we have nothing to send, we don't
1926 : * want to wake on socket-writable.)
1927 : */
1928 5102 : sleeptime = WalSndComputeSleeptime(GetCurrentTimestamp());
1929 :
1930 5102 : wakeEvents = WL_SOCKET_READABLE;
1931 :
1932 5102 : if (pq_is_send_pending())
1933 0 : wakeEvents |= WL_SOCKET_WRITEABLE;
1934 :
1935 : Assert(wait_event != 0);
1936 :
1937 5102 : WalSndWait(wakeEvents, sleeptime, wait_event);
1938 : }
1939 :
1940 : /* reactivate latch so WalSndLoop knows to continue */
1941 10526 : SetLatch(MyLatch);
1942 10526 : return RecentFlushPtr;
1943 : }
1944 :
1945 : /*
1946 : * Execute an incoming replication command.
1947 : *
1948 : * Returns true if the cmd_string was recognized as WalSender command, false
1949 : * if not.
1950 : */
1951 : bool
1952 9742 : exec_replication_command(const char *cmd_string)
1953 : {
1954 : yyscan_t scanner;
1955 : int parse_rc;
1956 : Node *cmd_node;
1957 : const char *cmdtag;
1958 : MemoryContext cmd_context;
1959 : MemoryContext old_context;
1960 :
1961 : /*
1962 : * If WAL sender has been told that shutdown is getting close, switch its
1963 : * status accordingly to handle the next replication commands correctly.
1964 : */
1965 9742 : if (got_STOPPING)
1966 0 : WalSndSetState(WALSNDSTATE_STOPPING);
1967 :
1968 : /*
1969 : * Throw error if in stopping mode. We need prevent commands that could
1970 : * generate WAL while the shutdown checkpoint is being written. To be
1971 : * safe, we just prohibit all new commands.
1972 : */
1973 9742 : if (MyWalSnd->state == WALSNDSTATE_STOPPING)
1974 0 : ereport(ERROR,
1975 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1976 : errmsg("cannot execute new commands while WAL sender is in stopping mode")));
1977 :
1978 : /*
1979 : * CREATE_REPLICATION_SLOT ... LOGICAL exports a snapshot until the next
1980 : * command arrives. Clean up the old stuff if there's anything.
1981 : */
1982 9742 : SnapBuildClearExportedSnapshot();
1983 :
1984 9742 : CHECK_FOR_INTERRUPTS();
1985 :
1986 : /*
1987 : * Prepare to parse and execute the command.
1988 : */
1989 9742 : cmd_context = AllocSetContextCreate(CurrentMemoryContext,
1990 : "Replication command context",
1991 : ALLOCSET_DEFAULT_SIZES);
1992 9742 : old_context = MemoryContextSwitchTo(cmd_context);
1993 :
1994 9742 : replication_scanner_init(cmd_string, &scanner);
1995 :
1996 : /*
1997 : * Is it a WalSender command?
1998 : */
1999 9742 : if (!replication_scanner_is_replication_command(scanner))
2000 : {
2001 : /* Nope; clean up and get out. */
2002 4264 : replication_scanner_finish(scanner);
2003 :
2004 4264 : MemoryContextSwitchTo(old_context);
2005 4264 : MemoryContextDelete(cmd_context);
2006 :
2007 : /* XXX this is a pretty random place to make this check */
2008 4264 : if (MyDatabaseId == InvalidOid)
2009 0 : ereport(ERROR,
2010 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2011 : errmsg("cannot execute SQL commands in WAL sender for physical replication")));
2012 :
2013 : /* Tell the caller that this wasn't a WalSender command. */
2014 4264 : return false;
2015 : }
2016 :
2017 : /*
2018 : * Looks like a WalSender command, so parse it.
2019 : */
2020 5478 : parse_rc = replication_yyparse(scanner);
2021 5478 : if (parse_rc != 0)
2022 0 : ereport(ERROR,
2023 : (errcode(ERRCODE_SYNTAX_ERROR),
2024 : errmsg_internal("replication command parser returned %d",
2025 : parse_rc)));
2026 5478 : replication_scanner_finish(scanner);
2027 :
2028 5478 : cmd_node = replication_parse_result;
2029 :
2030 : /*
2031 : * Report query to various monitoring facilities. For this purpose, we
2032 : * report replication commands just like SQL commands.
2033 : */
2034 5478 : debug_query_string = cmd_string;
2035 :
2036 5478 : pgstat_report_activity(STATE_RUNNING, cmd_string);
2037 :
2038 : /*
2039 : * Log replication command if log_replication_commands is enabled. Even
2040 : * when it's disabled, log the command with DEBUG1 level for backward
2041 : * compatibility.
2042 : */
2043 5478 : ereport(log_replication_commands ? LOG : DEBUG1,
2044 : (errmsg("received replication command: %s", cmd_string)));
2045 :
2046 : /*
2047 : * Disallow replication commands in aborted transaction blocks.
2048 : */
2049 5478 : if (IsAbortedTransactionBlockState())
2050 0 : ereport(ERROR,
2051 : (errcode(ERRCODE_IN_FAILED_SQL_TRANSACTION),
2052 : errmsg("current transaction is aborted, "
2053 : "commands ignored until end of transaction block")));
2054 :
2055 5478 : CHECK_FOR_INTERRUPTS();
2056 :
2057 : /*
2058 : * Allocate buffers that will be used for each outgoing and incoming
2059 : * message. We do this just once per command to reduce palloc overhead.
2060 : */
2061 5478 : initStringInfo(&output_message);
2062 5478 : initStringInfo(&reply_message);
2063 5478 : initStringInfo(&tmpbuf);
2064 :
2065 5478 : switch (cmd_node->type)
2066 : {
2067 1318 : case T_IdentifySystemCmd:
2068 1318 : cmdtag = "IDENTIFY_SYSTEM";
2069 1318 : set_ps_display(cmdtag);
2070 1318 : IdentifySystem();
2071 1318 : EndReplicationCommand(cmdtag);
2072 1318 : break;
2073 :
2074 12 : case T_ReadReplicationSlotCmd:
2075 12 : cmdtag = "READ_REPLICATION_SLOT";
2076 12 : set_ps_display(cmdtag);
2077 12 : ReadReplicationSlot((ReadReplicationSlotCmd *) cmd_node);
2078 10 : EndReplicationCommand(cmdtag);
2079 10 : break;
2080 :
2081 354 : case T_BaseBackupCmd:
2082 354 : cmdtag = "BASE_BACKUP";
2083 354 : set_ps_display(cmdtag);
2084 354 : PreventInTransactionBlock(true, cmdtag);
2085 354 : SendBaseBackup((BaseBackupCmd *) cmd_node, uploaded_manifest);
2086 296 : EndReplicationCommand(cmdtag);
2087 296 : break;
2088 :
2089 888 : case T_CreateReplicationSlotCmd:
2090 888 : cmdtag = "CREATE_REPLICATION_SLOT";
2091 888 : set_ps_display(cmdtag);
2092 888 : CreateReplicationSlot((CreateReplicationSlotCmd *) cmd_node);
2093 886 : EndReplicationCommand(cmdtag);
2094 886 : break;
2095 :
2096 494 : case T_DropReplicationSlotCmd:
2097 494 : cmdtag = "DROP_REPLICATION_SLOT";
2098 494 : set_ps_display(cmdtag);
2099 494 : DropReplicationSlot((DropReplicationSlotCmd *) cmd_node);
2100 494 : EndReplicationCommand(cmdtag);
2101 494 : break;
2102 :
2103 12 : case T_AlterReplicationSlotCmd:
2104 12 : cmdtag = "ALTER_REPLICATION_SLOT";
2105 12 : set_ps_display(cmdtag);
2106 12 : AlterReplicationSlot((AlterReplicationSlotCmd *) cmd_node);
2107 8 : EndReplicationCommand(cmdtag);
2108 8 : break;
2109 :
2110 1242 : case T_StartReplicationCmd:
2111 : {
2112 1242 : StartReplicationCmd *cmd = (StartReplicationCmd *) cmd_node;
2113 :
2114 1242 : cmdtag = "START_REPLICATION";
2115 1242 : set_ps_display(cmdtag);
2116 1242 : PreventInTransactionBlock(true, cmdtag);
2117 :
2118 1242 : if (cmd->kind == REPLICATION_KIND_PHYSICAL)
2119 512 : StartReplication(cmd);
2120 : else
2121 730 : StartLogicalReplication(cmd);
2122 :
2123 : /* dupe, but necessary per libpqrcv_endstreaming */
2124 656 : EndReplicationCommand(cmdtag);
2125 :
2126 : Assert(xlogreader != NULL);
2127 656 : break;
2128 : }
2129 :
2130 28 : case T_TimeLineHistoryCmd:
2131 28 : cmdtag = "TIMELINE_HISTORY";
2132 28 : set_ps_display(cmdtag);
2133 28 : PreventInTransactionBlock(true, cmdtag);
2134 28 : SendTimeLineHistory((TimeLineHistoryCmd *) cmd_node);
2135 28 : EndReplicationCommand(cmdtag);
2136 28 : break;
2137 :
2138 1108 : case T_VariableShowStmt:
2139 : {
2140 1108 : DestReceiver *dest = CreateDestReceiver(DestRemoteSimple);
2141 1108 : VariableShowStmt *n = (VariableShowStmt *) cmd_node;
2142 :
2143 1108 : cmdtag = "SHOW";
2144 1108 : set_ps_display(cmdtag);
2145 :
2146 : /* syscache access needs a transaction environment */
2147 1108 : StartTransactionCommand();
2148 1108 : GetPGVariable(n->name, dest);
2149 1108 : CommitTransactionCommand();
2150 1108 : EndReplicationCommand(cmdtag);
2151 : }
2152 1108 : break;
2153 :
2154 22 : case T_UploadManifestCmd:
2155 22 : cmdtag = "UPLOAD_MANIFEST";
2156 22 : set_ps_display(cmdtag);
2157 22 : PreventInTransactionBlock(true, cmdtag);
2158 22 : UploadManifest();
2159 20 : EndReplicationCommand(cmdtag);
2160 20 : break;
2161 :
2162 0 : default:
2163 0 : elog(ERROR, "unrecognized replication command node tag: %u",
2164 : cmd_node->type);
2165 : }
2166 :
2167 : /* done */
2168 4824 : MemoryContextSwitchTo(old_context);
2169 4824 : MemoryContextDelete(cmd_context);
2170 :
2171 : /*
2172 : * We need not update ps display or pg_stat_activity, because PostgresMain
2173 : * will reset those to "idle". But we must reset debug_query_string to
2174 : * ensure it doesn't become a dangling pointer.
2175 : */
2176 4824 : debug_query_string = NULL;
2177 :
2178 4824 : return true;
2179 : }
2180 :
2181 : /*
2182 : * Process any incoming messages while streaming. Also checks if the remote
2183 : * end has closed the connection.
2184 : */
2185 : static void
2186 2594066 : ProcessRepliesIfAny(void)
2187 : {
2188 : unsigned char firstchar;
2189 : int maxmsglen;
2190 : int r;
2191 2594066 : bool received = false;
2192 :
2193 2594066 : last_processing = GetCurrentTimestamp();
2194 :
2195 : /*
2196 : * If we already received a CopyDone from the frontend, any subsequent
2197 : * message is the beginning of a new command, and should be processed in
2198 : * the main processing loop.
2199 : */
2200 2594066 : while (!streamingDoneReceiving)
2201 : {
2202 2660708 : pq_startmsgread();
2203 2660708 : r = pq_getbyte_if_available(&firstchar);
2204 2660708 : if (r < 0)
2205 : {
2206 : /* unexpected error or EOF */
2207 34 : ereport(COMMERROR,
2208 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2209 : errmsg("unexpected EOF on standby connection")));
2210 34 : proc_exit(0);
2211 : }
2212 2660674 : if (r == 0)
2213 : {
2214 : /* no data available without blocking */
2215 2592276 : pq_endmsgread();
2216 2592276 : break;
2217 : }
2218 :
2219 : /* Validate message type and set packet size limit */
2220 68398 : switch (firstchar)
2221 : {
2222 67308 : case PqMsg_CopyData:
2223 67308 : maxmsglen = PQ_LARGE_MESSAGE_LIMIT;
2224 67308 : break;
2225 1090 : case PqMsg_CopyDone:
2226 : case PqMsg_Terminate:
2227 1090 : maxmsglen = PQ_SMALL_MESSAGE_LIMIT;
2228 1090 : break;
2229 0 : default:
2230 0 : ereport(FATAL,
2231 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2232 : errmsg("invalid standby message type \"%c\"",
2233 : firstchar)));
2234 : maxmsglen = 0; /* keep compiler quiet */
2235 : break;
2236 : }
2237 :
2238 : /* Read the message contents */
2239 68398 : resetStringInfo(&reply_message);
2240 68398 : if (pq_getmessage(&reply_message, maxmsglen))
2241 : {
2242 0 : ereport(COMMERROR,
2243 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2244 : errmsg("unexpected EOF on standby connection")));
2245 0 : proc_exit(0);
2246 : }
2247 :
2248 : /* ... and process it */
2249 68398 : switch (firstchar)
2250 : {
2251 : /*
2252 : * 'd' means a standby reply wrapped in a CopyData packet.
2253 : */
2254 67308 : case PqMsg_CopyData:
2255 67308 : ProcessStandbyMessage();
2256 67308 : received = true;
2257 67308 : break;
2258 :
2259 : /*
2260 : * CopyDone means the standby requested to finish streaming.
2261 : * Reply with CopyDone, if we had not sent that already.
2262 : */
2263 656 : case PqMsg_CopyDone:
2264 656 : if (!streamingDoneSending)
2265 : {
2266 630 : pq_putmessage_noblock('c', NULL, 0);
2267 630 : streamingDoneSending = true;
2268 : }
2269 :
2270 656 : streamingDoneReceiving = true;
2271 656 : received = true;
2272 656 : break;
2273 :
2274 : /*
2275 : * 'X' means that the standby is closing down the socket.
2276 : */
2277 434 : case PqMsg_Terminate:
2278 434 : proc_exit(0);
2279 :
2280 2662030 : default:
2281 : Assert(false); /* NOT REACHED */
2282 : }
2283 : }
2284 :
2285 : /*
2286 : * Save the last reply timestamp if we've received at least one reply.
2287 : */
2288 2593598 : if (received)
2289 : {
2290 18800 : last_reply_timestamp = last_processing;
2291 18800 : waiting_for_ping_response = false;
2292 : }
2293 2593598 : }
2294 :
2295 : /*
2296 : * Process a status update message received from standby.
2297 : */
2298 : static void
2299 67308 : ProcessStandbyMessage(void)
2300 : {
2301 : char msgtype;
2302 :
2303 : /*
2304 : * Check message type from the first byte.
2305 : */
2306 67308 : msgtype = pq_getmsgbyte(&reply_message);
2307 :
2308 67308 : switch (msgtype)
2309 : {
2310 67042 : case 'r':
2311 67042 : ProcessStandbyReplyMessage();
2312 67042 : break;
2313 :
2314 266 : case 'h':
2315 266 : ProcessStandbyHSFeedbackMessage();
2316 266 : break;
2317 :
2318 0 : default:
2319 0 : ereport(COMMERROR,
2320 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2321 : errmsg("unexpected message type \"%c\"", msgtype)));
2322 0 : proc_exit(0);
2323 : }
2324 67308 : }
2325 :
2326 : /*
2327 : * Remember that a walreceiver just confirmed receipt of lsn `lsn`.
2328 : */
2329 : static void
2330 16652 : PhysicalConfirmReceivedLocation(XLogRecPtr lsn)
2331 : {
2332 16652 : bool changed = false;
2333 16652 : ReplicationSlot *slot = MyReplicationSlot;
2334 :
2335 : Assert(lsn != InvalidXLogRecPtr);
2336 16652 : SpinLockAcquire(&slot->mutex);
2337 16652 : if (slot->data.restart_lsn != lsn)
2338 : {
2339 8308 : changed = true;
2340 8308 : slot->data.restart_lsn = lsn;
2341 : }
2342 16652 : SpinLockRelease(&slot->mutex);
2343 :
2344 16652 : if (changed)
2345 : {
2346 8308 : ReplicationSlotMarkDirty();
2347 8308 : ReplicationSlotsComputeRequiredLSN();
2348 8308 : PhysicalWakeupLogicalWalSnd();
2349 : }
2350 :
2351 : /*
2352 : * One could argue that the slot should be saved to disk now, but that'd
2353 : * be energy wasted - the worst thing lost information could cause here is
2354 : * to give wrong information in a statistics view - we'll just potentially
2355 : * be more conservative in removing files.
2356 : */
2357 16652 : }
2358 :
2359 : /*
2360 : * Regular reply from standby advising of WAL locations on standby server.
2361 : */
2362 : static void
2363 67042 : ProcessStandbyReplyMessage(void)
2364 : {
2365 : XLogRecPtr writePtr,
2366 : flushPtr,
2367 : applyPtr;
2368 : bool replyRequested;
2369 : TimeOffset writeLag,
2370 : flushLag,
2371 : applyLag;
2372 : bool clearLagTimes;
2373 : TimestampTz now;
2374 : TimestampTz replyTime;
2375 :
2376 : static bool fullyAppliedLastTime = false;
2377 :
2378 : /* the caller already consumed the msgtype byte */
2379 67042 : writePtr = pq_getmsgint64(&reply_message);
2380 67042 : flushPtr = pq_getmsgint64(&reply_message);
2381 67042 : applyPtr = pq_getmsgint64(&reply_message);
2382 67042 : replyTime = pq_getmsgint64(&reply_message);
2383 67042 : replyRequested = pq_getmsgbyte(&reply_message);
2384 :
2385 67042 : if (message_level_is_interesting(DEBUG2))
2386 : {
2387 : char *replyTimeStr;
2388 :
2389 : /* Copy because timestamptz_to_str returns a static buffer */
2390 574 : replyTimeStr = pstrdup(timestamptz_to_str(replyTime));
2391 :
2392 574 : elog(DEBUG2, "write %X/%X flush %X/%X apply %X/%X%s reply_time %s",
2393 : LSN_FORMAT_ARGS(writePtr),
2394 : LSN_FORMAT_ARGS(flushPtr),
2395 : LSN_FORMAT_ARGS(applyPtr),
2396 : replyRequested ? " (reply requested)" : "",
2397 : replyTimeStr);
2398 :
2399 574 : pfree(replyTimeStr);
2400 : }
2401 :
2402 : /* See if we can compute the round-trip lag for these positions. */
2403 67042 : now = GetCurrentTimestamp();
2404 67042 : writeLag = LagTrackerRead(SYNC_REP_WAIT_WRITE, writePtr, now);
2405 67042 : flushLag = LagTrackerRead(SYNC_REP_WAIT_FLUSH, flushPtr, now);
2406 67042 : applyLag = LagTrackerRead(SYNC_REP_WAIT_APPLY, applyPtr, now);
2407 :
2408 : /*
2409 : * If the standby reports that it has fully replayed the WAL in two
2410 : * consecutive reply messages, then the second such message must result
2411 : * from wal_receiver_status_interval expiring on the standby. This is a
2412 : * convenient time to forget the lag times measured when it last
2413 : * wrote/flushed/applied a WAL record, to avoid displaying stale lag data
2414 : * until more WAL traffic arrives.
2415 : */
2416 67042 : clearLagTimes = false;
2417 67042 : if (applyPtr == sentPtr)
2418 : {
2419 6126 : if (fullyAppliedLastTime)
2420 3798 : clearLagTimes = true;
2421 6126 : fullyAppliedLastTime = true;
2422 : }
2423 : else
2424 60916 : fullyAppliedLastTime = false;
2425 :
2426 : /* Send a reply if the standby requested one. */
2427 67042 : if (replyRequested)
2428 0 : WalSndKeepalive(false, InvalidXLogRecPtr);
2429 :
2430 : /*
2431 : * Update shared state for this WalSender process based on reply data from
2432 : * standby.
2433 : */
2434 : {
2435 67042 : WalSnd *walsnd = MyWalSnd;
2436 :
2437 67042 : SpinLockAcquire(&walsnd->mutex);
2438 67042 : walsnd->write = writePtr;
2439 67042 : walsnd->flush = flushPtr;
2440 67042 : walsnd->apply = applyPtr;
2441 67042 : if (writeLag != -1 || clearLagTimes)
2442 20334 : walsnd->writeLag = writeLag;
2443 67042 : if (flushLag != -1 || clearLagTimes)
2444 22802 : walsnd->flushLag = flushLag;
2445 67042 : if (applyLag != -1 || clearLagTimes)
2446 23804 : walsnd->applyLag = applyLag;
2447 67042 : walsnd->replyTime = replyTime;
2448 67042 : SpinLockRelease(&walsnd->mutex);
2449 : }
2450 :
2451 67042 : if (!am_cascading_walsender)
2452 66492 : SyncRepReleaseWaiters();
2453 :
2454 : /*
2455 : * Advance our local xmin horizon when the client confirmed a flush.
2456 : */
2457 67042 : if (MyReplicationSlot && flushPtr != InvalidXLogRecPtr)
2458 : {
2459 63472 : if (SlotIsLogical(MyReplicationSlot))
2460 46820 : LogicalConfirmReceivedLocation(flushPtr);
2461 : else
2462 16652 : PhysicalConfirmReceivedLocation(flushPtr);
2463 : }
2464 67042 : }
2465 :
2466 : /* compute new replication slot xmin horizon if needed */
2467 : static void
2468 122 : PhysicalReplicationSlotNewXmin(TransactionId feedbackXmin, TransactionId feedbackCatalogXmin)
2469 : {
2470 122 : bool changed = false;
2471 122 : ReplicationSlot *slot = MyReplicationSlot;
2472 :
2473 122 : SpinLockAcquire(&slot->mutex);
2474 122 : MyProc->xmin = InvalidTransactionId;
2475 :
2476 : /*
2477 : * For physical replication we don't need the interlock provided by xmin
2478 : * and effective_xmin since the consequences of a missed increase are
2479 : * limited to query cancellations, so set both at once.
2480 : */
2481 122 : if (!TransactionIdIsNormal(slot->data.xmin) ||
2482 60 : !TransactionIdIsNormal(feedbackXmin) ||
2483 60 : TransactionIdPrecedes(slot->data.xmin, feedbackXmin))
2484 : {
2485 80 : changed = true;
2486 80 : slot->data.xmin = feedbackXmin;
2487 80 : slot->effective_xmin = feedbackXmin;
2488 : }
2489 122 : if (!TransactionIdIsNormal(slot->data.catalog_xmin) ||
2490 30 : !TransactionIdIsNormal(feedbackCatalogXmin) ||
2491 30 : TransactionIdPrecedes(slot->data.catalog_xmin, feedbackCatalogXmin))
2492 : {
2493 94 : changed = true;
2494 94 : slot->data.catalog_xmin = feedbackCatalogXmin;
2495 94 : slot->effective_catalog_xmin = feedbackCatalogXmin;
2496 : }
2497 122 : SpinLockRelease(&slot->mutex);
2498 :
2499 122 : if (changed)
2500 : {
2501 100 : ReplicationSlotMarkDirty();
2502 100 : ReplicationSlotsComputeRequiredXmin(false);
2503 : }
2504 122 : }
2505 :
2506 : /*
2507 : * Check that the provided xmin/epoch are sane, that is, not in the future
2508 : * and not so far back as to be already wrapped around.
2509 : *
2510 : * Epoch of nextXid should be same as standby, or if the counter has
2511 : * wrapped, then one greater than standby.
2512 : *
2513 : * This check doesn't care about whether clog exists for these xids
2514 : * at all.
2515 : */
2516 : static bool
2517 124 : TransactionIdInRecentPast(TransactionId xid, uint32 epoch)
2518 : {
2519 : FullTransactionId nextFullXid;
2520 : TransactionId nextXid;
2521 : uint32 nextEpoch;
2522 :
2523 124 : nextFullXid = ReadNextFullTransactionId();
2524 124 : nextXid = XidFromFullTransactionId(nextFullXid);
2525 124 : nextEpoch = EpochFromFullTransactionId(nextFullXid);
2526 :
2527 124 : if (xid <= nextXid)
2528 : {
2529 124 : if (epoch != nextEpoch)
2530 0 : return false;
2531 : }
2532 : else
2533 : {
2534 0 : if (epoch + 1 != nextEpoch)
2535 0 : return false;
2536 : }
2537 :
2538 124 : if (!TransactionIdPrecedesOrEquals(xid, nextXid))
2539 0 : return false; /* epoch OK, but it's wrapped around */
2540 :
2541 124 : return true;
2542 : }
2543 :
2544 : /*
2545 : * Hot Standby feedback
2546 : */
2547 : static void
2548 266 : ProcessStandbyHSFeedbackMessage(void)
2549 : {
2550 : TransactionId feedbackXmin;
2551 : uint32 feedbackEpoch;
2552 : TransactionId feedbackCatalogXmin;
2553 : uint32 feedbackCatalogEpoch;
2554 : TimestampTz replyTime;
2555 :
2556 : /*
2557 : * Decipher the reply message. The caller already consumed the msgtype
2558 : * byte. See XLogWalRcvSendHSFeedback() in walreceiver.c for the creation
2559 : * of this message.
2560 : */
2561 266 : replyTime = pq_getmsgint64(&reply_message);
2562 266 : feedbackXmin = pq_getmsgint(&reply_message, 4);
2563 266 : feedbackEpoch = pq_getmsgint(&reply_message, 4);
2564 266 : feedbackCatalogXmin = pq_getmsgint(&reply_message, 4);
2565 266 : feedbackCatalogEpoch = pq_getmsgint(&reply_message, 4);
2566 :
2567 266 : if (message_level_is_interesting(DEBUG2))
2568 : {
2569 : char *replyTimeStr;
2570 :
2571 : /* Copy because timestamptz_to_str returns a static buffer */
2572 8 : replyTimeStr = pstrdup(timestamptz_to_str(replyTime));
2573 :
2574 8 : elog(DEBUG2, "hot standby feedback xmin %u epoch %u, catalog_xmin %u epoch %u reply_time %s",
2575 : feedbackXmin,
2576 : feedbackEpoch,
2577 : feedbackCatalogXmin,
2578 : feedbackCatalogEpoch,
2579 : replyTimeStr);
2580 :
2581 8 : pfree(replyTimeStr);
2582 : }
2583 :
2584 : /*
2585 : * Update shared state for this WalSender process based on reply data from
2586 : * standby.
2587 : */
2588 : {
2589 266 : WalSnd *walsnd = MyWalSnd;
2590 :
2591 266 : SpinLockAcquire(&walsnd->mutex);
2592 266 : walsnd->replyTime = replyTime;
2593 266 : SpinLockRelease(&walsnd->mutex);
2594 : }
2595 :
2596 : /*
2597 : * Unset WalSender's xmins if the feedback message values are invalid.
2598 : * This happens when the downstream turned hot_standby_feedback off.
2599 : */
2600 266 : if (!TransactionIdIsNormal(feedbackXmin)
2601 184 : && !TransactionIdIsNormal(feedbackCatalogXmin))
2602 : {
2603 184 : MyProc->xmin = InvalidTransactionId;
2604 184 : if (MyReplicationSlot != NULL)
2605 42 : PhysicalReplicationSlotNewXmin(feedbackXmin, feedbackCatalogXmin);
2606 184 : return;
2607 : }
2608 :
2609 : /*
2610 : * Check that the provided xmin/epoch are sane, that is, not in the future
2611 : * and not so far back as to be already wrapped around. Ignore if not.
2612 : */
2613 82 : if (TransactionIdIsNormal(feedbackXmin) &&
2614 82 : !TransactionIdInRecentPast(feedbackXmin, feedbackEpoch))
2615 0 : return;
2616 :
2617 82 : if (TransactionIdIsNormal(feedbackCatalogXmin) &&
2618 42 : !TransactionIdInRecentPast(feedbackCatalogXmin, feedbackCatalogEpoch))
2619 0 : return;
2620 :
2621 : /*
2622 : * Set the WalSender's xmin equal to the standby's requested xmin, so that
2623 : * the xmin will be taken into account by GetSnapshotData() /
2624 : * ComputeXidHorizons(). This will hold back the removal of dead rows and
2625 : * thereby prevent the generation of cleanup conflicts on the standby
2626 : * server.
2627 : *
2628 : * There is a small window for a race condition here: although we just
2629 : * checked that feedbackXmin precedes nextXid, the nextXid could have
2630 : * gotten advanced between our fetching it and applying the xmin below,
2631 : * perhaps far enough to make feedbackXmin wrap around. In that case the
2632 : * xmin we set here would be "in the future" and have no effect. No point
2633 : * in worrying about this since it's too late to save the desired data
2634 : * anyway. Assuming that the standby sends us an increasing sequence of
2635 : * xmins, this could only happen during the first reply cycle, else our
2636 : * own xmin would prevent nextXid from advancing so far.
2637 : *
2638 : * We don't bother taking the ProcArrayLock here. Setting the xmin field
2639 : * is assumed atomic, and there's no real need to prevent concurrent
2640 : * horizon determinations. (If we're moving our xmin forward, this is
2641 : * obviously safe, and if we're moving it backwards, well, the data is at
2642 : * risk already since a VACUUM could already have determined the horizon.)
2643 : *
2644 : * If we're using a replication slot we reserve the xmin via that,
2645 : * otherwise via the walsender's PGPROC entry. We can only track the
2646 : * catalog xmin separately when using a slot, so we store the least of the
2647 : * two provided when not using a slot.
2648 : *
2649 : * XXX: It might make sense to generalize the ephemeral slot concept and
2650 : * always use the slot mechanism to handle the feedback xmin.
2651 : */
2652 82 : if (MyReplicationSlot != NULL) /* XXX: persistency configurable? */
2653 80 : PhysicalReplicationSlotNewXmin(feedbackXmin, feedbackCatalogXmin);
2654 : else
2655 : {
2656 2 : if (TransactionIdIsNormal(feedbackCatalogXmin)
2657 0 : && TransactionIdPrecedes(feedbackCatalogXmin, feedbackXmin))
2658 0 : MyProc->xmin = feedbackCatalogXmin;
2659 : else
2660 2 : MyProc->xmin = feedbackXmin;
2661 : }
2662 : }
2663 :
2664 : /*
2665 : * Compute how long send/receive loops should sleep.
2666 : *
2667 : * If wal_sender_timeout is enabled we want to wake up in time to send
2668 : * keepalives and to abort the connection if wal_sender_timeout has been
2669 : * reached.
2670 : */
2671 : static long
2672 88524 : WalSndComputeSleeptime(TimestampTz now)
2673 : {
2674 88524 : long sleeptime = 10000; /* 10 s */
2675 :
2676 88524 : if (wal_sender_timeout > 0 && last_reply_timestamp > 0)
2677 : {
2678 : TimestampTz wakeup_time;
2679 :
2680 : /*
2681 : * At the latest stop sleeping once wal_sender_timeout has been
2682 : * reached.
2683 : */
2684 88478 : wakeup_time = TimestampTzPlusMilliseconds(last_reply_timestamp,
2685 : wal_sender_timeout);
2686 :
2687 : /*
2688 : * If no ping has been sent yet, wakeup when it's time to do so.
2689 : * WalSndKeepaliveIfNecessary() wants to send a keepalive once half of
2690 : * the timeout passed without a response.
2691 : */
2692 88478 : if (!waiting_for_ping_response)
2693 86566 : wakeup_time = TimestampTzPlusMilliseconds(last_reply_timestamp,
2694 : wal_sender_timeout / 2);
2695 :
2696 : /* Compute relative time until wakeup. */
2697 88478 : sleeptime = TimestampDifferenceMilliseconds(now, wakeup_time);
2698 : }
2699 :
2700 88524 : return sleeptime;
2701 : }
2702 :
2703 : /*
2704 : * Check whether there have been responses by the client within
2705 : * wal_sender_timeout and shutdown if not. Using last_processing as the
2706 : * reference point avoids counting server-side stalls against the client.
2707 : * However, a long server-side stall can make WalSndKeepaliveIfNecessary()
2708 : * postdate last_processing by more than wal_sender_timeout. If that happens,
2709 : * the client must reply almost immediately to avoid a timeout. This rarely
2710 : * affects the default configuration, under which clients spontaneously send a
2711 : * message every standby_message_timeout = wal_sender_timeout/6 = 10s. We
2712 : * could eliminate that problem by recognizing timeout expiration at
2713 : * wal_sender_timeout/2 after the keepalive.
2714 : */
2715 : static void
2716 2582010 : WalSndCheckTimeOut(void)
2717 : {
2718 : TimestampTz timeout;
2719 :
2720 : /* don't bail out if we're doing something that doesn't require timeouts */
2721 2582010 : if (last_reply_timestamp <= 0)
2722 46 : return;
2723 :
2724 2581964 : timeout = TimestampTzPlusMilliseconds(last_reply_timestamp,
2725 : wal_sender_timeout);
2726 :
2727 2581964 : if (wal_sender_timeout > 0 && last_processing >= timeout)
2728 : {
2729 : /*
2730 : * Since typically expiration of replication timeout means
2731 : * communication problem, we don't send the error message to the
2732 : * standby.
2733 : */
2734 0 : ereport(COMMERROR,
2735 : (errmsg("terminating walsender process due to replication timeout")));
2736 :
2737 0 : WalSndShutdown();
2738 : }
2739 : }
2740 :
2741 : /* Main loop of walsender process that streams the WAL over Copy messages. */
2742 : static void
2743 1222 : WalSndLoop(WalSndSendDataCallback send_data)
2744 : {
2745 : /*
2746 : * Initialize the last reply timestamp. That enables timeout processing
2747 : * from hereon.
2748 : */
2749 1222 : last_reply_timestamp = GetCurrentTimestamp();
2750 1222 : waiting_for_ping_response = false;
2751 :
2752 : /*
2753 : * Loop until we reach the end of this timeline or the client requests to
2754 : * stop streaming.
2755 : */
2756 : for (;;)
2757 : {
2758 : /* Clear any already-pending wakeups */
2759 2577890 : ResetLatch(MyLatch);
2760 :
2761 2577890 : CHECK_FOR_INTERRUPTS();
2762 :
2763 : /* Process any requests or signals received recently */
2764 2577884 : if (ConfigReloadPending)
2765 : {
2766 34 : ConfigReloadPending = false;
2767 34 : ProcessConfigFile(PGC_SIGHUP);
2768 34 : SyncRepInitConfig();
2769 : }
2770 :
2771 : /* Check for input from the client */
2772 2577884 : ProcessRepliesIfAny();
2773 :
2774 : /*
2775 : * If we have received CopyDone from the client, sent CopyDone
2776 : * ourselves, and the output buffer is empty, it's time to exit
2777 : * streaming.
2778 : */
2779 2577730 : if (streamingDoneReceiving && streamingDoneSending &&
2780 1004 : !pq_is_send_pending())
2781 656 : break;
2782 :
2783 : /*
2784 : * If we don't have any pending data in the output buffer, try to send
2785 : * some more. If there is some, we don't bother to call send_data
2786 : * again until we've flushed it ... but we'd better assume we are not
2787 : * caught up.
2788 : */
2789 2577074 : if (!pq_is_send_pending())
2790 2502080 : send_data();
2791 : else
2792 74994 : WalSndCaughtUp = false;
2793 :
2794 : /* Try to flush pending output to the client */
2795 2576736 : if (pq_flush_if_writable() != 0)
2796 0 : WalSndShutdown();
2797 :
2798 : /* If nothing remains to be sent right now ... */
2799 2576736 : if (WalSndCaughtUp && !pq_is_send_pending())
2800 : {
2801 : /*
2802 : * If we're in catchup state, move to streaming. This is an
2803 : * important state change for users to know about, since before
2804 : * this point data loss might occur if the primary dies and we
2805 : * need to failover to the standby. The state change is also
2806 : * important for synchronous replication, since commits that
2807 : * started to wait at that point might wait for some time.
2808 : */
2809 663194 : if (MyWalSnd->state == WALSNDSTATE_CATCHUP)
2810 : {
2811 1182 : ereport(DEBUG1,
2812 : (errmsg_internal("\"%s\" has now caught up with upstream server",
2813 : application_name)));
2814 1182 : WalSndSetState(WALSNDSTATE_STREAMING);
2815 : }
2816 :
2817 : /*
2818 : * When SIGUSR2 arrives, we send any outstanding logs up to the
2819 : * shutdown checkpoint record (i.e., the latest record), wait for
2820 : * them to be replicated to the standby, and exit. This may be a
2821 : * normal termination at shutdown, or a promotion, the walsender
2822 : * is not sure which.
2823 : */
2824 663194 : if (got_SIGUSR2)
2825 3446 : WalSndDone(send_data);
2826 : }
2827 :
2828 : /* Check for replication timeout. */
2829 2576668 : WalSndCheckTimeOut();
2830 :
2831 : /* Send keepalive if the time has come */
2832 2576668 : WalSndKeepaliveIfNecessary();
2833 :
2834 : /*
2835 : * Block if we have unsent data. XXX For logical replication, let
2836 : * WalSndWaitForWal() handle any other blocking; idle receivers need
2837 : * its additional actions. For physical replication, also block if
2838 : * caught up; its send_data does not block.
2839 : */
2840 2576668 : if ((WalSndCaughtUp && send_data != XLogSendLogical &&
2841 3223462 : !streamingDoneSending) ||
2842 2568208 : pq_is_send_pending())
2843 : {
2844 : long sleeptime;
2845 : int wakeEvents;
2846 :
2847 83282 : if (!streamingDoneReceiving)
2848 83246 : wakeEvents = WL_SOCKET_READABLE;
2849 : else
2850 36 : wakeEvents = 0;
2851 :
2852 : /*
2853 : * Use fresh timestamp, not last_processing, to reduce the chance
2854 : * of reaching wal_sender_timeout before sending a keepalive.
2855 : */
2856 83282 : sleeptime = WalSndComputeSleeptime(GetCurrentTimestamp());
2857 :
2858 83282 : if (pq_is_send_pending())
2859 74898 : wakeEvents |= WL_SOCKET_WRITEABLE;
2860 :
2861 : /* Sleep until something happens or we time out */
2862 83282 : WalSndWait(wakeEvents, sleeptime, WAIT_EVENT_WAL_SENDER_MAIN);
2863 : }
2864 : }
2865 656 : }
2866 :
2867 : /* Initialize a per-walsender data structure for this walsender process */
2868 : static void
2869 2124 : InitWalSenderSlot(void)
2870 : {
2871 : int i;
2872 :
2873 : /*
2874 : * WalSndCtl should be set up already (we inherit this by fork() or
2875 : * EXEC_BACKEND mechanism from the postmaster).
2876 : */
2877 : Assert(WalSndCtl != NULL);
2878 : Assert(MyWalSnd == NULL);
2879 :
2880 : /*
2881 : * Find a free walsender slot and reserve it. This must not fail due to
2882 : * the prior check for free WAL senders in InitProcess().
2883 : */
2884 3058 : for (i = 0; i < max_wal_senders; i++)
2885 : {
2886 3058 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
2887 :
2888 3058 : SpinLockAcquire(&walsnd->mutex);
2889 :
2890 3058 : if (walsnd->pid != 0)
2891 : {
2892 934 : SpinLockRelease(&walsnd->mutex);
2893 934 : continue;
2894 : }
2895 : else
2896 : {
2897 : /*
2898 : * Found a free slot. Reserve it for us.
2899 : */
2900 2124 : walsnd->pid = MyProcPid;
2901 2124 : walsnd->state = WALSNDSTATE_STARTUP;
2902 2124 : walsnd->sentPtr = InvalidXLogRecPtr;
2903 2124 : walsnd->needreload = false;
2904 2124 : walsnd->write = InvalidXLogRecPtr;
2905 2124 : walsnd->flush = InvalidXLogRecPtr;
2906 2124 : walsnd->apply = InvalidXLogRecPtr;
2907 2124 : walsnd->writeLag = -1;
2908 2124 : walsnd->flushLag = -1;
2909 2124 : walsnd->applyLag = -1;
2910 2124 : walsnd->sync_standby_priority = 0;
2911 2124 : walsnd->replyTime = 0;
2912 :
2913 : /*
2914 : * The kind assignment is done here and not in StartReplication()
2915 : * and StartLogicalReplication(). Indeed, the logical walsender
2916 : * needs to read WAL records (like snapshot of running
2917 : * transactions) during the slot creation. So it needs to be woken
2918 : * up based on its kind.
2919 : *
2920 : * The kind assignment could also be done in StartReplication(),
2921 : * StartLogicalReplication() and CREATE_REPLICATION_SLOT but it
2922 : * seems better to set it on one place.
2923 : */
2924 2124 : if (MyDatabaseId == InvalidOid)
2925 892 : walsnd->kind = REPLICATION_KIND_PHYSICAL;
2926 : else
2927 1232 : walsnd->kind = REPLICATION_KIND_LOGICAL;
2928 :
2929 2124 : SpinLockRelease(&walsnd->mutex);
2930 : /* don't need the lock anymore */
2931 2124 : MyWalSnd = (WalSnd *) walsnd;
2932 :
2933 2124 : break;
2934 : }
2935 : }
2936 :
2937 : Assert(MyWalSnd != NULL);
2938 :
2939 : /* Arrange to clean up at walsender exit */
2940 2124 : on_shmem_exit(WalSndKill, 0);
2941 2124 : }
2942 :
2943 : /* Destroy the per-walsender data structure for this walsender process */
2944 : static void
2945 2124 : WalSndKill(int code, Datum arg)
2946 : {
2947 2124 : WalSnd *walsnd = MyWalSnd;
2948 :
2949 : Assert(walsnd != NULL);
2950 :
2951 2124 : MyWalSnd = NULL;
2952 :
2953 2124 : SpinLockAcquire(&walsnd->mutex);
2954 : /* Mark WalSnd struct as no longer being in use. */
2955 2124 : walsnd->pid = 0;
2956 2124 : SpinLockRelease(&walsnd->mutex);
2957 2124 : }
2958 :
2959 : /* XLogReaderRoutine->segment_open callback */
2960 : static void
2961 9884 : WalSndSegmentOpen(XLogReaderState *state, XLogSegNo nextSegNo,
2962 : TimeLineID *tli_p)
2963 : {
2964 : char path[MAXPGPATH];
2965 :
2966 : /*-------
2967 : * When reading from a historic timeline, and there is a timeline switch
2968 : * within this segment, read from the WAL segment belonging to the new
2969 : * timeline.
2970 : *
2971 : * For example, imagine that this server is currently on timeline 5, and
2972 : * we're streaming timeline 4. The switch from timeline 4 to 5 happened at
2973 : * 0/13002088. In pg_wal, we have these files:
2974 : *
2975 : * ...
2976 : * 000000040000000000000012
2977 : * 000000040000000000000013
2978 : * 000000050000000000000013
2979 : * 000000050000000000000014
2980 : * ...
2981 : *
2982 : * In this situation, when requested to send the WAL from segment 0x13, on
2983 : * timeline 4, we read the WAL from file 000000050000000000000013. Archive
2984 : * recovery prefers files from newer timelines, so if the segment was
2985 : * restored from the archive on this server, the file belonging to the old
2986 : * timeline, 000000040000000000000013, might not exist. Their contents are
2987 : * equal up to the switchpoint, because at a timeline switch, the used
2988 : * portion of the old segment is copied to the new file.
2989 : */
2990 9884 : *tli_p = sendTimeLine;
2991 9884 : if (sendTimeLineIsHistoric)
2992 : {
2993 : XLogSegNo endSegNo;
2994 :
2995 26 : XLByteToSeg(sendTimeLineValidUpto, endSegNo, state->segcxt.ws_segsize);
2996 26 : if (nextSegNo == endSegNo)
2997 20 : *tli_p = sendTimeLineNextTLI;
2998 : }
2999 :
3000 9884 : XLogFilePath(path, *tli_p, nextSegNo, state->segcxt.ws_segsize);
3001 9884 : state->seg.ws_file = BasicOpenFile(path, O_RDONLY | PG_BINARY);
3002 9884 : if (state->seg.ws_file >= 0)
3003 9880 : return;
3004 :
3005 : /*
3006 : * If the file is not found, assume it's because the standby asked for a
3007 : * too old WAL segment that has already been removed or recycled.
3008 : */
3009 4 : if (errno == ENOENT)
3010 : {
3011 : char xlogfname[MAXFNAMELEN];
3012 4 : int save_errno = errno;
3013 :
3014 4 : XLogFileName(xlogfname, *tli_p, nextSegNo, wal_segment_size);
3015 4 : errno = save_errno;
3016 4 : ereport(ERROR,
3017 : (errcode_for_file_access(),
3018 : errmsg("requested WAL segment %s has already been removed",
3019 : xlogfname)));
3020 : }
3021 : else
3022 0 : ereport(ERROR,
3023 : (errcode_for_file_access(),
3024 : errmsg("could not open file \"%s\": %m",
3025 : path)));
3026 : }
3027 :
3028 : /*
3029 : * Send out the WAL in its normal physical/stored form.
3030 : *
3031 : * Read up to MAX_SEND_SIZE bytes of WAL that's been flushed to disk,
3032 : * but not yet sent to the client, and buffer it in the libpq output
3033 : * buffer.
3034 : *
3035 : * If there is no unsent WAL remaining, WalSndCaughtUp is set to true,
3036 : * otherwise WalSndCaughtUp is set to false.
3037 : */
3038 : static void
3039 834986 : XLogSendPhysical(void)
3040 : {
3041 : XLogRecPtr SendRqstPtr;
3042 : XLogRecPtr startptr;
3043 : XLogRecPtr endptr;
3044 : Size nbytes;
3045 : XLogSegNo segno;
3046 : WALReadError errinfo;
3047 : Size rbytes;
3048 :
3049 : /* If requested switch the WAL sender to the stopping state. */
3050 834986 : if (got_STOPPING)
3051 852 : WalSndSetState(WALSNDSTATE_STOPPING);
3052 :
3053 834986 : if (streamingDoneSending)
3054 : {
3055 646768 : WalSndCaughtUp = true;
3056 652410 : return;
3057 : }
3058 :
3059 : /* Figure out how far we can safely send the WAL. */
3060 188218 : if (sendTimeLineIsHistoric)
3061 : {
3062 : /*
3063 : * Streaming an old timeline that's in this server's history, but is
3064 : * not the one we're currently inserting or replaying. It can be
3065 : * streamed up to the point where we switched off that timeline.
3066 : */
3067 332 : SendRqstPtr = sendTimeLineValidUpto;
3068 : }
3069 187886 : else if (am_cascading_walsender)
3070 : {
3071 : TimeLineID SendRqstTLI;
3072 :
3073 : /*
3074 : * Streaming the latest timeline on a standby.
3075 : *
3076 : * Attempt to send all WAL that has already been replayed, so that we
3077 : * know it's valid. If we're receiving WAL through streaming
3078 : * replication, it's also OK to send any WAL that has been received
3079 : * but not replayed.
3080 : *
3081 : * The timeline we're recovering from can change, or we can be
3082 : * promoted. In either case, the current timeline becomes historic. We
3083 : * need to detect that so that we don't try to stream past the point
3084 : * where we switched to another timeline. We check for promotion or
3085 : * timeline switch after calculating FlushPtr, to avoid a race
3086 : * condition: if the timeline becomes historic just after we checked
3087 : * that it was still current, it's still be OK to stream it up to the
3088 : * FlushPtr that was calculated before it became historic.
3089 : */
3090 1474 : bool becameHistoric = false;
3091 :
3092 1474 : SendRqstPtr = GetStandbyFlushRecPtr(&SendRqstTLI);
3093 :
3094 1474 : if (!RecoveryInProgress())
3095 : {
3096 : /* We have been promoted. */
3097 2 : SendRqstTLI = GetWALInsertionTimeLine();
3098 2 : am_cascading_walsender = false;
3099 2 : becameHistoric = true;
3100 : }
3101 : else
3102 : {
3103 : /*
3104 : * Still a cascading standby. But is the timeline we're sending
3105 : * still the one recovery is recovering from?
3106 : */
3107 1472 : if (sendTimeLine != SendRqstTLI)
3108 0 : becameHistoric = true;
3109 : }
3110 :
3111 1474 : if (becameHistoric)
3112 : {
3113 : /*
3114 : * The timeline we were sending has become historic. Read the
3115 : * timeline history file of the new timeline to see where exactly
3116 : * we forked off from the timeline we were sending.
3117 : */
3118 : List *history;
3119 :
3120 2 : history = readTimeLineHistory(SendRqstTLI);
3121 2 : sendTimeLineValidUpto = tliSwitchPoint(sendTimeLine, history, &sendTimeLineNextTLI);
3122 :
3123 : Assert(sendTimeLine < sendTimeLineNextTLI);
3124 2 : list_free_deep(history);
3125 :
3126 2 : sendTimeLineIsHistoric = true;
3127 :
3128 2 : SendRqstPtr = sendTimeLineValidUpto;
3129 : }
3130 : }
3131 : else
3132 : {
3133 : /*
3134 : * Streaming the current timeline on a primary.
3135 : *
3136 : * Attempt to send all data that's already been written out and
3137 : * fsync'd to disk. We cannot go further than what's been written out
3138 : * given the current implementation of WALRead(). And in any case
3139 : * it's unsafe to send WAL that is not securely down to disk on the
3140 : * primary: if the primary subsequently crashes and restarts, standbys
3141 : * must not have applied any WAL that got lost on the primary.
3142 : */
3143 186412 : SendRqstPtr = GetFlushRecPtr(NULL);
3144 : }
3145 :
3146 : /*
3147 : * Record the current system time as an approximation of the time at which
3148 : * this WAL location was written for the purposes of lag tracking.
3149 : *
3150 : * In theory we could make XLogFlush() record a time in shmem whenever WAL
3151 : * is flushed and we could get that time as well as the LSN when we call
3152 : * GetFlushRecPtr() above (and likewise for the cascading standby
3153 : * equivalent), but rather than putting any new code into the hot WAL path
3154 : * it seems good enough to capture the time here. We should reach this
3155 : * after XLogFlush() runs WalSndWakeupProcessRequests(), and although that
3156 : * may take some time, we read the WAL flush pointer and take the time
3157 : * very close to together here so that we'll get a later position if it is
3158 : * still moving.
3159 : *
3160 : * Because LagTrackerWrite ignores samples when the LSN hasn't advanced,
3161 : * this gives us a cheap approximation for the WAL flush time for this
3162 : * LSN.
3163 : *
3164 : * Note that the LSN is not necessarily the LSN for the data contained in
3165 : * the present message; it's the end of the WAL, which might be further
3166 : * ahead. All the lag tracking machinery cares about is finding out when
3167 : * that arbitrary LSN is eventually reported as written, flushed and
3168 : * applied, so that it can measure the elapsed time.
3169 : */
3170 188218 : LagTrackerWrite(SendRqstPtr, GetCurrentTimestamp());
3171 :
3172 : /*
3173 : * If this is a historic timeline and we've reached the point where we
3174 : * forked to the next timeline, stop streaming.
3175 : *
3176 : * Note: We might already have sent WAL > sendTimeLineValidUpto. The
3177 : * startup process will normally replay all WAL that has been received
3178 : * from the primary, before promoting, but if the WAL streaming is
3179 : * terminated at a WAL page boundary, the valid portion of the timeline
3180 : * might end in the middle of a WAL record. We might've already sent the
3181 : * first half of that partial WAL record to the cascading standby, so that
3182 : * sentPtr > sendTimeLineValidUpto. That's OK; the cascading standby can't
3183 : * replay the partial WAL record either, so it can still follow our
3184 : * timeline switch.
3185 : */
3186 188218 : if (sendTimeLineIsHistoric && sendTimeLineValidUpto <= sentPtr)
3187 : {
3188 : /* close the current file. */
3189 26 : if (xlogreader->seg.ws_file >= 0)
3190 26 : wal_segment_close(xlogreader);
3191 :
3192 : /* Send CopyDone */
3193 26 : pq_putmessage_noblock('c', NULL, 0);
3194 26 : streamingDoneSending = true;
3195 :
3196 26 : WalSndCaughtUp = true;
3197 :
3198 26 : elog(DEBUG1, "walsender reached end of timeline at %X/%X (sent up to %X/%X)",
3199 : LSN_FORMAT_ARGS(sendTimeLineValidUpto),
3200 : LSN_FORMAT_ARGS(sentPtr));
3201 26 : return;
3202 : }
3203 :
3204 : /* Do we have any work to do? */
3205 : Assert(sentPtr <= SendRqstPtr);
3206 188192 : if (SendRqstPtr <= sentPtr)
3207 : {
3208 5616 : WalSndCaughtUp = true;
3209 5616 : return;
3210 : }
3211 :
3212 : /*
3213 : * Figure out how much to send in one message. If there's no more than
3214 : * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
3215 : * MAX_SEND_SIZE bytes, but round back to logfile or page boundary.
3216 : *
3217 : * The rounding is not only for performance reasons. Walreceiver relies on
3218 : * the fact that we never split a WAL record across two messages. Since a
3219 : * long WAL record is split at page boundary into continuation records,
3220 : * page boundary is always a safe cut-off point. We also assume that
3221 : * SendRqstPtr never points to the middle of a WAL record.
3222 : */
3223 182576 : startptr = sentPtr;
3224 182576 : endptr = startptr;
3225 182576 : endptr += MAX_SEND_SIZE;
3226 :
3227 : /* if we went beyond SendRqstPtr, back off */
3228 182576 : if (SendRqstPtr <= endptr)
3229 : {
3230 2990 : endptr = SendRqstPtr;
3231 2990 : if (sendTimeLineIsHistoric)
3232 24 : WalSndCaughtUp = false;
3233 : else
3234 2966 : WalSndCaughtUp = true;
3235 : }
3236 : else
3237 : {
3238 : /* round down to page boundary. */
3239 179586 : endptr -= (endptr % XLOG_BLCKSZ);
3240 179586 : WalSndCaughtUp = false;
3241 : }
3242 :
3243 182576 : nbytes = endptr - startptr;
3244 : Assert(nbytes <= MAX_SEND_SIZE);
3245 :
3246 : /*
3247 : * OK to read and send the slice.
3248 : */
3249 182576 : resetStringInfo(&output_message);
3250 182576 : pq_sendbyte(&output_message, 'w');
3251 :
3252 182576 : pq_sendint64(&output_message, startptr); /* dataStart */
3253 182576 : pq_sendint64(&output_message, SendRqstPtr); /* walEnd */
3254 182576 : pq_sendint64(&output_message, 0); /* sendtime, filled in last */
3255 :
3256 : /*
3257 : * Read the log directly into the output buffer to avoid extra memcpy
3258 : * calls.
3259 : */
3260 182576 : enlargeStringInfo(&output_message, nbytes);
3261 :
3262 182576 : retry:
3263 : /* attempt to read WAL from WAL buffers first */
3264 182576 : rbytes = WALReadFromBuffers(&output_message.data[output_message.len],
3265 182576 : startptr, nbytes, xlogreader->seg.ws_tli);
3266 182576 : output_message.len += rbytes;
3267 182576 : startptr += rbytes;
3268 182576 : nbytes -= rbytes;
3269 :
3270 : /* now read the remaining WAL from WAL file */
3271 182576 : if (nbytes > 0 &&
3272 180776 : !WALRead(xlogreader,
3273 180780 : &output_message.data[output_message.len],
3274 : startptr,
3275 : nbytes,
3276 180780 : xlogreader->seg.ws_tli, /* Pass the current TLI because
3277 : * only WalSndSegmentOpen controls
3278 : * whether new TLI is needed. */
3279 : &errinfo))
3280 0 : WALReadRaiseError(&errinfo);
3281 :
3282 : /* See logical_read_xlog_page(). */
3283 182572 : XLByteToSeg(startptr, segno, xlogreader->segcxt.ws_segsize);
3284 182572 : CheckXLogRemoved(segno, xlogreader->seg.ws_tli);
3285 :
3286 : /*
3287 : * During recovery, the currently-open WAL file might be replaced with the
3288 : * file of the same name retrieved from archive. So we always need to
3289 : * check what we read was valid after reading into the buffer. If it's
3290 : * invalid, we try to open and read the file again.
3291 : */
3292 182572 : if (am_cascading_walsender)
3293 : {
3294 1098 : WalSnd *walsnd = MyWalSnd;
3295 : bool reload;
3296 :
3297 1098 : SpinLockAcquire(&walsnd->mutex);
3298 1098 : reload = walsnd->needreload;
3299 1098 : walsnd->needreload = false;
3300 1098 : SpinLockRelease(&walsnd->mutex);
3301 :
3302 1098 : if (reload && xlogreader->seg.ws_file >= 0)
3303 : {
3304 0 : wal_segment_close(xlogreader);
3305 :
3306 0 : goto retry;
3307 : }
3308 : }
3309 :
3310 182572 : output_message.len += nbytes;
3311 182572 : output_message.data[output_message.len] = '\0';
3312 :
3313 : /*
3314 : * Fill the send timestamp last, so that it is taken as late as possible.
3315 : */
3316 182572 : resetStringInfo(&tmpbuf);
3317 182572 : pq_sendint64(&tmpbuf, GetCurrentTimestamp());
3318 182572 : memcpy(&output_message.data[1 + sizeof(int64) + sizeof(int64)],
3319 182572 : tmpbuf.data, sizeof(int64));
3320 :
3321 182572 : pq_putmessage_noblock('d', output_message.data, output_message.len);
3322 :
3323 182572 : sentPtr = endptr;
3324 :
3325 : /* Update shared memory status */
3326 : {
3327 182572 : WalSnd *walsnd = MyWalSnd;
3328 :
3329 182572 : SpinLockAcquire(&walsnd->mutex);
3330 182572 : walsnd->sentPtr = sentPtr;
3331 182572 : SpinLockRelease(&walsnd->mutex);
3332 : }
3333 :
3334 : /* Report progress of XLOG streaming in PS display */
3335 182572 : if (update_process_title)
3336 : {
3337 : char activitymsg[50];
3338 :
3339 182572 : snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%X",
3340 182572 : LSN_FORMAT_ARGS(sentPtr));
3341 182572 : set_ps_display(activitymsg);
3342 : }
3343 : }
3344 :
3345 : /*
3346 : * Stream out logically decoded data.
3347 : */
3348 : static void
3349 1670540 : XLogSendLogical(void)
3350 : {
3351 : XLogRecord *record;
3352 : char *errm;
3353 :
3354 : /*
3355 : * We'll use the current flush point to determine whether we've caught up.
3356 : * This variable is static in order to cache it across calls. Caching is
3357 : * helpful because GetFlushRecPtr() needs to acquire a heavily-contended
3358 : * spinlock.
3359 : */
3360 : static XLogRecPtr flushPtr = InvalidXLogRecPtr;
3361 :
3362 : /*
3363 : * Don't know whether we've caught up yet. We'll set WalSndCaughtUp to
3364 : * true in WalSndWaitForWal, if we're actually waiting. We also set to
3365 : * true if XLogReadRecord() had to stop reading but WalSndWaitForWal
3366 : * didn't wait - i.e. when we're shutting down.
3367 : */
3368 1670540 : WalSndCaughtUp = false;
3369 :
3370 1670540 : record = XLogReadRecord(logical_decoding_ctx->reader, &errm);
3371 :
3372 : /* xlog record was invalid */
3373 1670214 : if (errm != NULL)
3374 0 : elog(ERROR, "could not find record while sending logically-decoded data: %s",
3375 : errm);
3376 :
3377 1670214 : if (record != NULL)
3378 : {
3379 : /*
3380 : * Note the lack of any call to LagTrackerWrite() which is handled by
3381 : * WalSndUpdateProgress which is called by output plugin through
3382 : * logical decoding write api.
3383 : */
3384 1663150 : LogicalDecodingProcessRecord(logical_decoding_ctx, logical_decoding_ctx->reader);
3385 :
3386 1663142 : sentPtr = logical_decoding_ctx->reader->EndRecPtr;
3387 : }
3388 :
3389 : /*
3390 : * If first time through in this session, initialize flushPtr. Otherwise,
3391 : * we only need to update flushPtr if EndRecPtr is past it.
3392 : */
3393 1670206 : if (flushPtr == InvalidXLogRecPtr ||
3394 1669512 : logical_decoding_ctx->reader->EndRecPtr >= flushPtr)
3395 : {
3396 11436 : if (am_cascading_walsender)
3397 118 : flushPtr = GetStandbyFlushRecPtr(NULL);
3398 : else
3399 11318 : flushPtr = GetFlushRecPtr(NULL);
3400 : }
3401 :
3402 : /* If EndRecPtr is still past our flushPtr, it means we caught up. */
3403 1670206 : if (logical_decoding_ctx->reader->EndRecPtr >= flushPtr)
3404 9686 : WalSndCaughtUp = true;
3405 :
3406 : /*
3407 : * If we're caught up and have been requested to stop, have WalSndLoop()
3408 : * terminate the connection in an orderly manner, after writing out all
3409 : * the pending data.
3410 : */
3411 1670206 : if (WalSndCaughtUp && got_STOPPING)
3412 6750 : got_SIGUSR2 = true;
3413 :
3414 : /* Update shared memory status */
3415 : {
3416 1670206 : WalSnd *walsnd = MyWalSnd;
3417 :
3418 1670206 : SpinLockAcquire(&walsnd->mutex);
3419 1670206 : walsnd->sentPtr = sentPtr;
3420 1670206 : SpinLockRelease(&walsnd->mutex);
3421 : }
3422 1670206 : }
3423 :
3424 : /*
3425 : * Shutdown if the sender is caught up.
3426 : *
3427 : * NB: This should only be called when the shutdown signal has been received
3428 : * from postmaster.
3429 : *
3430 : * Note that if we determine that there's still more data to send, this
3431 : * function will return control to the caller.
3432 : */
3433 : static void
3434 3446 : WalSndDone(WalSndSendDataCallback send_data)
3435 : {
3436 : XLogRecPtr replicatedPtr;
3437 :
3438 : /* ... let's just be real sure we're caught up ... */
3439 3446 : send_data();
3440 :
3441 : /*
3442 : * To figure out whether all WAL has successfully been replicated, check
3443 : * flush location if valid, write otherwise. Tools like pg_receivewal will
3444 : * usually (unless in synchronous mode) return an invalid flush location.
3445 : */
3446 6892 : replicatedPtr = XLogRecPtrIsInvalid(MyWalSnd->flush) ?
3447 3446 : MyWalSnd->write : MyWalSnd->flush;
3448 :
3449 3446 : if (WalSndCaughtUp && sentPtr == replicatedPtr &&
3450 68 : !pq_is_send_pending())
3451 : {
3452 : QueryCompletion qc;
3453 :
3454 : /* Inform the standby that XLOG streaming is done */
3455 68 : SetQueryCompletion(&qc, CMDTAG_COPY, 0);
3456 68 : EndCommand(&qc, DestRemote, false);
3457 68 : pq_flush();
3458 :
3459 68 : proc_exit(0);
3460 : }
3461 3378 : if (!waiting_for_ping_response)
3462 1904 : WalSndKeepalive(true, InvalidXLogRecPtr);
3463 3378 : }
3464 :
3465 : /*
3466 : * Returns the latest point in WAL that has been safely flushed to disk.
3467 : * This should only be called when in recovery.
3468 : *
3469 : * This is called either by cascading walsender to find WAL position to be sent
3470 : * to a cascaded standby or by slot synchronization operation to validate remote
3471 : * slot's lsn before syncing it locally.
3472 : *
3473 : * As a side-effect, *tli is updated to the TLI of the last
3474 : * replayed WAL record.
3475 : */
3476 : XLogRecPtr
3477 1794 : GetStandbyFlushRecPtr(TimeLineID *tli)
3478 : {
3479 : XLogRecPtr replayPtr;
3480 : TimeLineID replayTLI;
3481 : XLogRecPtr receivePtr;
3482 : TimeLineID receiveTLI;
3483 : XLogRecPtr result;
3484 :
3485 : Assert(am_cascading_walsender || IsSyncingReplicationSlots());
3486 :
3487 : /*
3488 : * We can safely send what's already been replayed. Also, if walreceiver
3489 : * is streaming WAL from the same timeline, we can send anything that it
3490 : * has streamed, but hasn't been replayed yet.
3491 : */
3492 :
3493 1794 : receivePtr = GetWalRcvFlushRecPtr(NULL, &receiveTLI);
3494 1794 : replayPtr = GetXLogReplayRecPtr(&replayTLI);
3495 :
3496 1794 : if (tli)
3497 1614 : *tli = replayTLI;
3498 :
3499 1794 : result = replayPtr;
3500 1794 : if (receiveTLI == replayTLI && receivePtr > replayPtr)
3501 58 : result = receivePtr;
3502 :
3503 1794 : return result;
3504 : }
3505 :
3506 : /*
3507 : * Request walsenders to reload the currently-open WAL file
3508 : */
3509 : void
3510 40 : WalSndRqstFileReload(void)
3511 : {
3512 : int i;
3513 :
3514 416 : for (i = 0; i < max_wal_senders; i++)
3515 : {
3516 376 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3517 :
3518 376 : SpinLockAcquire(&walsnd->mutex);
3519 376 : if (walsnd->pid == 0)
3520 : {
3521 376 : SpinLockRelease(&walsnd->mutex);
3522 376 : continue;
3523 : }
3524 0 : walsnd->needreload = true;
3525 0 : SpinLockRelease(&walsnd->mutex);
3526 : }
3527 40 : }
3528 :
3529 : /*
3530 : * Handle PROCSIG_WALSND_INIT_STOPPING signal.
3531 : */
3532 : void
3533 68 : HandleWalSndInitStopping(void)
3534 : {
3535 : Assert(am_walsender);
3536 :
3537 : /*
3538 : * If replication has not yet started, die like with SIGTERM. If
3539 : * replication is active, only set a flag and wake up the main loop. It
3540 : * will send any outstanding WAL, wait for it to be replicated to the
3541 : * standby, and then exit gracefully.
3542 : */
3543 68 : if (!replication_active)
3544 0 : kill(MyProcPid, SIGTERM);
3545 : else
3546 68 : got_STOPPING = true;
3547 68 : }
3548 :
3549 : /*
3550 : * SIGUSR2: set flag to do a last cycle and shut down afterwards. The WAL
3551 : * sender should already have been switched to WALSNDSTATE_STOPPING at
3552 : * this point.
3553 : */
3554 : static void
3555 52 : WalSndLastCycleHandler(SIGNAL_ARGS)
3556 : {
3557 52 : got_SIGUSR2 = true;
3558 52 : SetLatch(MyLatch);
3559 52 : }
3560 :
3561 : /* Set up signal handlers */
3562 : void
3563 2124 : WalSndSignals(void)
3564 : {
3565 : /* Set up signal handlers */
3566 2124 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
3567 2124 : pqsignal(SIGINT, StatementCancelHandler); /* query cancel */
3568 2124 : pqsignal(SIGTERM, die); /* request shutdown */
3569 : /* SIGQUIT handler was already set up by InitPostmasterChild */
3570 2124 : InitializeTimeouts(); /* establishes SIGALRM handler */
3571 2124 : pqsignal(SIGPIPE, SIG_IGN);
3572 2124 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
3573 2124 : pqsignal(SIGUSR2, WalSndLastCycleHandler); /* request a last cycle and
3574 : * shutdown */
3575 :
3576 : /* Reset some signals that are accepted by postmaster but not here */
3577 2124 : pqsignal(SIGCHLD, SIG_DFL);
3578 2124 : }
3579 :
3580 : /* Report shared-memory space needed by WalSndShmemInit */
3581 : Size
3582 7402 : WalSndShmemSize(void)
3583 : {
3584 7402 : Size size = 0;
3585 :
3586 7402 : size = offsetof(WalSndCtlData, walsnds);
3587 7402 : size = add_size(size, mul_size(max_wal_senders, sizeof(WalSnd)));
3588 :
3589 7402 : return size;
3590 : }
3591 :
3592 : /* Allocate and initialize walsender-related shared memory */
3593 : void
3594 1918 : WalSndShmemInit(void)
3595 : {
3596 : bool found;
3597 : int i;
3598 :
3599 1918 : WalSndCtl = (WalSndCtlData *)
3600 1918 : ShmemInitStruct("Wal Sender Ctl", WalSndShmemSize(), &found);
3601 :
3602 1918 : if (!found)
3603 : {
3604 : /* First time through, so initialize */
3605 13558 : MemSet(WalSndCtl, 0, WalSndShmemSize());
3606 :
3607 7672 : for (i = 0; i < NUM_SYNC_REP_WAIT_MODE; i++)
3608 5754 : dlist_init(&(WalSndCtl->SyncRepQueue[i]));
3609 :
3610 14698 : for (i = 0; i < max_wal_senders; i++)
3611 : {
3612 12780 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3613 :
3614 12780 : SpinLockInit(&walsnd->mutex);
3615 : }
3616 :
3617 1918 : ConditionVariableInit(&WalSndCtl->wal_flush_cv);
3618 1918 : ConditionVariableInit(&WalSndCtl->wal_replay_cv);
3619 1918 : ConditionVariableInit(&WalSndCtl->wal_confirm_rcv_cv);
3620 : }
3621 1918 : }
3622 :
3623 : /*
3624 : * Wake up physical, logical or both kinds of walsenders
3625 : *
3626 : * The distinction between physical and logical walsenders is done, because:
3627 : * - physical walsenders can't send data until it's been flushed
3628 : * - logical walsenders on standby can't decode and send data until it's been
3629 : * applied
3630 : *
3631 : * For cascading replication we need to wake up physical walsenders separately
3632 : * from logical walsenders (see the comment before calling WalSndWakeup() in
3633 : * ApplyWalRecord() for more details).
3634 : *
3635 : * This will be called inside critical sections, so throwing an error is not
3636 : * advisable.
3637 : */
3638 : void
3639 5144782 : WalSndWakeup(bool physical, bool logical)
3640 : {
3641 : /*
3642 : * Wake up all the walsenders waiting on WAL being flushed or replayed
3643 : * respectively. Note that waiting walsender would have prepared to sleep
3644 : * on the CV (i.e., added itself to the CV's waitlist) in WalSndWait()
3645 : * before actually waiting.
3646 : */
3647 5144782 : if (physical)
3648 199070 : ConditionVariableBroadcast(&WalSndCtl->wal_flush_cv);
3649 :
3650 5144782 : if (logical)
3651 5133948 : ConditionVariableBroadcast(&WalSndCtl->wal_replay_cv);
3652 5144782 : }
3653 :
3654 : /*
3655 : * Wait for readiness on the FeBe socket, or a timeout. The mask should be
3656 : * composed of optional WL_SOCKET_WRITEABLE and WL_SOCKET_READABLE flags. Exit
3657 : * on postmaster death.
3658 : */
3659 : static void
3660 88524 : WalSndWait(uint32 socket_events, long timeout, uint32 wait_event)
3661 : {
3662 : WaitEvent event;
3663 :
3664 88524 : ModifyWaitEvent(FeBeWaitSet, FeBeWaitSetSocketPos, socket_events, NULL);
3665 :
3666 : /*
3667 : * We use a condition variable to efficiently wake up walsenders in
3668 : * WalSndWakeup().
3669 : *
3670 : * Every walsender prepares to sleep on a shared memory CV. Note that it
3671 : * just prepares to sleep on the CV (i.e., adds itself to the CV's
3672 : * waitlist), but does not actually wait on the CV (IOW, it never calls
3673 : * ConditionVariableSleep()). It still uses WaitEventSetWait() for
3674 : * waiting, because we also need to wait for socket events. The processes
3675 : * (startup process, walreceiver etc.) wanting to wake up walsenders use
3676 : * ConditionVariableBroadcast(), which in turn calls SetLatch(), helping
3677 : * walsenders come out of WaitEventSetWait().
3678 : *
3679 : * This approach is simple and efficient because, one doesn't have to loop
3680 : * through all the walsenders slots, with a spinlock acquisition and
3681 : * release for every iteration, just to wake up only the waiting
3682 : * walsenders. It makes WalSndWakeup() callers' life easy.
3683 : *
3684 : * XXX: A desirable future improvement would be to add support for CVs
3685 : * into WaitEventSetWait().
3686 : *
3687 : * And, we use separate shared memory CVs for physical and logical
3688 : * walsenders for selective wake ups, see WalSndWakeup() for more details.
3689 : *
3690 : * If the wait event is WAIT_FOR_STANDBY_CONFIRMATION, wait on another CV
3691 : * until awakened by physical walsenders after the walreceiver confirms
3692 : * the receipt of the LSN.
3693 : */
3694 88524 : if (wait_event == WAIT_EVENT_WAIT_FOR_STANDBY_CONFIRMATION)
3695 8 : ConditionVariablePrepareToSleep(&WalSndCtl->wal_confirm_rcv_cv);
3696 88516 : else if (MyWalSnd->kind == REPLICATION_KIND_PHYSICAL)
3697 81388 : ConditionVariablePrepareToSleep(&WalSndCtl->wal_flush_cv);
3698 7128 : else if (MyWalSnd->kind == REPLICATION_KIND_LOGICAL)
3699 7128 : ConditionVariablePrepareToSleep(&WalSndCtl->wal_replay_cv);
3700 :
3701 88524 : if (WaitEventSetWait(FeBeWaitSet, timeout, &event, 1, wait_event) == 1 &&
3702 88524 : (event.events & WL_POSTMASTER_DEATH))
3703 : {
3704 0 : ConditionVariableCancelSleep();
3705 0 : proc_exit(1);
3706 : }
3707 :
3708 88524 : ConditionVariableCancelSleep();
3709 88524 : }
3710 :
3711 : /*
3712 : * Signal all walsenders to move to stopping state.
3713 : *
3714 : * This will trigger walsenders to move to a state where no further WAL can be
3715 : * generated. See this file's header for details.
3716 : */
3717 : void
3718 1092 : WalSndInitStopping(void)
3719 : {
3720 : int i;
3721 :
3722 8628 : for (i = 0; i < max_wal_senders; i++)
3723 : {
3724 7536 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3725 : pid_t pid;
3726 :
3727 7536 : SpinLockAcquire(&walsnd->mutex);
3728 7536 : pid = walsnd->pid;
3729 7536 : SpinLockRelease(&walsnd->mutex);
3730 :
3731 7536 : if (pid == 0)
3732 7468 : continue;
3733 :
3734 68 : SendProcSignal(pid, PROCSIG_WALSND_INIT_STOPPING, INVALID_PROC_NUMBER);
3735 : }
3736 1092 : }
3737 :
3738 : /*
3739 : * Wait that all the WAL senders have quit or reached the stopping state. This
3740 : * is used by the checkpointer to control when the shutdown checkpoint can
3741 : * safely be performed.
3742 : */
3743 : void
3744 1160 : WalSndWaitStopping(void)
3745 : {
3746 : for (;;)
3747 68 : {
3748 : int i;
3749 1160 : bool all_stopped = true;
3750 :
3751 8702 : for (i = 0; i < max_wal_senders; i++)
3752 : {
3753 7610 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3754 :
3755 7610 : SpinLockAcquire(&walsnd->mutex);
3756 :
3757 7610 : if (walsnd->pid == 0)
3758 : {
3759 7490 : SpinLockRelease(&walsnd->mutex);
3760 7490 : continue;
3761 : }
3762 :
3763 120 : if (walsnd->state != WALSNDSTATE_STOPPING)
3764 : {
3765 68 : all_stopped = false;
3766 68 : SpinLockRelease(&walsnd->mutex);
3767 68 : break;
3768 : }
3769 52 : SpinLockRelease(&walsnd->mutex);
3770 : }
3771 :
3772 : /* safe to leave if confirmation is done for all WAL senders */
3773 1160 : if (all_stopped)
3774 1092 : return;
3775 :
3776 68 : pg_usleep(10000L); /* wait for 10 msec */
3777 : }
3778 : }
3779 :
3780 : /* Set state for current walsender (only called in walsender) */
3781 : void
3782 4338 : WalSndSetState(WalSndState state)
3783 : {
3784 4338 : WalSnd *walsnd = MyWalSnd;
3785 :
3786 : Assert(am_walsender);
3787 :
3788 4338 : if (walsnd->state == state)
3789 864 : return;
3790 :
3791 3474 : SpinLockAcquire(&walsnd->mutex);
3792 3474 : walsnd->state = state;
3793 3474 : SpinLockRelease(&walsnd->mutex);
3794 : }
3795 :
3796 : /*
3797 : * Return a string constant representing the state. This is used
3798 : * in system views, and should *not* be translated.
3799 : */
3800 : static const char *
3801 1560 : WalSndGetStateString(WalSndState state)
3802 : {
3803 1560 : switch (state)
3804 : {
3805 14 : case WALSNDSTATE_STARTUP:
3806 14 : return "startup";
3807 0 : case WALSNDSTATE_BACKUP:
3808 0 : return "backup";
3809 16 : case WALSNDSTATE_CATCHUP:
3810 16 : return "catchup";
3811 1530 : case WALSNDSTATE_STREAMING:
3812 1530 : return "streaming";
3813 0 : case WALSNDSTATE_STOPPING:
3814 0 : return "stopping";
3815 : }
3816 0 : return "UNKNOWN";
3817 : }
3818 :
3819 : static Interval *
3820 2426 : offset_to_interval(TimeOffset offset)
3821 : {
3822 2426 : Interval *result = palloc(sizeof(Interval));
3823 :
3824 2426 : result->month = 0;
3825 2426 : result->day = 0;
3826 2426 : result->time = offset;
3827 :
3828 2426 : return result;
3829 : }
3830 :
3831 : /*
3832 : * Returns activity of walsenders, including pids and xlog locations sent to
3833 : * standby servers.
3834 : */
3835 : Datum
3836 1328 : pg_stat_get_wal_senders(PG_FUNCTION_ARGS)
3837 : {
3838 : #define PG_STAT_GET_WAL_SENDERS_COLS 12
3839 1328 : ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
3840 : SyncRepStandbyData *sync_standbys;
3841 : int num_standbys;
3842 : int i;
3843 :
3844 1328 : InitMaterializedSRF(fcinfo, 0);
3845 :
3846 : /*
3847 : * Get the currently active synchronous standbys. This could be out of
3848 : * date before we're done, but we'll use the data anyway.
3849 : */
3850 1328 : num_standbys = SyncRepGetCandidateStandbys(&sync_standbys);
3851 :
3852 14220 : for (i = 0; i < max_wal_senders; i++)
3853 : {
3854 12892 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3855 : XLogRecPtr sent_ptr;
3856 : XLogRecPtr write;
3857 : XLogRecPtr flush;
3858 : XLogRecPtr apply;
3859 : TimeOffset writeLag;
3860 : TimeOffset flushLag;
3861 : TimeOffset applyLag;
3862 : int priority;
3863 : int pid;
3864 : WalSndState state;
3865 : TimestampTz replyTime;
3866 : bool is_sync_standby;
3867 : Datum values[PG_STAT_GET_WAL_SENDERS_COLS];
3868 12892 : bool nulls[PG_STAT_GET_WAL_SENDERS_COLS] = {0};
3869 : int j;
3870 :
3871 : /* Collect data from shared memory */
3872 12892 : SpinLockAcquire(&walsnd->mutex);
3873 12892 : if (walsnd->pid == 0)
3874 : {
3875 11332 : SpinLockRelease(&walsnd->mutex);
3876 11332 : continue;
3877 : }
3878 1560 : pid = walsnd->pid;
3879 1560 : sent_ptr = walsnd->sentPtr;
3880 1560 : state = walsnd->state;
3881 1560 : write = walsnd->write;
3882 1560 : flush = walsnd->flush;
3883 1560 : apply = walsnd->apply;
3884 1560 : writeLag = walsnd->writeLag;
3885 1560 : flushLag = walsnd->flushLag;
3886 1560 : applyLag = walsnd->applyLag;
3887 1560 : priority = walsnd->sync_standby_priority;
3888 1560 : replyTime = walsnd->replyTime;
3889 1560 : SpinLockRelease(&walsnd->mutex);
3890 :
3891 : /*
3892 : * Detect whether walsender is/was considered synchronous. We can
3893 : * provide some protection against stale data by checking the PID
3894 : * along with walsnd_index.
3895 : */
3896 1560 : is_sync_standby = false;
3897 1644 : for (j = 0; j < num_standbys; j++)
3898 : {
3899 142 : if (sync_standbys[j].walsnd_index == i &&
3900 58 : sync_standbys[j].pid == pid)
3901 : {
3902 58 : is_sync_standby = true;
3903 58 : break;
3904 : }
3905 : }
3906 :
3907 1560 : values[0] = Int32GetDatum(pid);
3908 :
3909 1560 : if (!has_privs_of_role(GetUserId(), ROLE_PG_READ_ALL_STATS))
3910 : {
3911 : /*
3912 : * Only superusers and roles with privileges of pg_read_all_stats
3913 : * can see details. Other users only get the pid value to know
3914 : * it's a walsender, but no details.
3915 : */
3916 0 : MemSet(&nulls[1], true, PG_STAT_GET_WAL_SENDERS_COLS - 1);
3917 : }
3918 : else
3919 : {
3920 1560 : values[1] = CStringGetTextDatum(WalSndGetStateString(state));
3921 :
3922 1560 : if (XLogRecPtrIsInvalid(sent_ptr))
3923 12 : nulls[2] = true;
3924 1560 : values[2] = LSNGetDatum(sent_ptr);
3925 :
3926 1560 : if (XLogRecPtrIsInvalid(write))
3927 18 : nulls[3] = true;
3928 1560 : values[3] = LSNGetDatum(write);
3929 :
3930 1560 : if (XLogRecPtrIsInvalid(flush))
3931 18 : nulls[4] = true;
3932 1560 : values[4] = LSNGetDatum(flush);
3933 :
3934 1560 : if (XLogRecPtrIsInvalid(apply))
3935 18 : nulls[5] = true;
3936 1560 : values[5] = LSNGetDatum(apply);
3937 :
3938 : /*
3939 : * Treat a standby such as a pg_basebackup background process
3940 : * which always returns an invalid flush location, as an
3941 : * asynchronous standby.
3942 : */
3943 1560 : priority = XLogRecPtrIsInvalid(flush) ? 0 : priority;
3944 :
3945 1560 : if (writeLag < 0)
3946 794 : nulls[6] = true;
3947 : else
3948 766 : values[6] = IntervalPGetDatum(offset_to_interval(writeLag));
3949 :
3950 1560 : if (flushLag < 0)
3951 666 : nulls[7] = true;
3952 : else
3953 894 : values[7] = IntervalPGetDatum(offset_to_interval(flushLag));
3954 :
3955 1560 : if (applyLag < 0)
3956 794 : nulls[8] = true;
3957 : else
3958 766 : values[8] = IntervalPGetDatum(offset_to_interval(applyLag));
3959 :
3960 1560 : values[9] = Int32GetDatum(priority);
3961 :
3962 : /*
3963 : * More easily understood version of standby state. This is purely
3964 : * informational.
3965 : *
3966 : * In quorum-based sync replication, the role of each standby
3967 : * listed in synchronous_standby_names can be changing very
3968 : * frequently. Any standbys considered as "sync" at one moment can
3969 : * be switched to "potential" ones at the next moment. So, it's
3970 : * basically useless to report "sync" or "potential" as their sync
3971 : * states. We report just "quorum" for them.
3972 : */
3973 1560 : if (priority == 0)
3974 1480 : values[10] = CStringGetTextDatum("async");
3975 80 : else if (is_sync_standby)
3976 58 : values[10] = SyncRepConfig->syncrep_method == SYNC_REP_PRIORITY ?
3977 58 : CStringGetTextDatum("sync") : CStringGetTextDatum("quorum");
3978 : else
3979 22 : values[10] = CStringGetTextDatum("potential");
3980 :
3981 1560 : if (replyTime == 0)
3982 12 : nulls[11] = true;
3983 : else
3984 1548 : values[11] = TimestampTzGetDatum(replyTime);
3985 : }
3986 :
3987 1560 : tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc,
3988 : values, nulls);
3989 : }
3990 :
3991 1328 : return (Datum) 0;
3992 : }
3993 :
3994 : /*
3995 : * Send a keepalive message to standby.
3996 : *
3997 : * If requestReply is set, the message requests the other party to send
3998 : * a message back to us, for heartbeat purposes. We also set a flag to
3999 : * let nearby code know that we're waiting for that response, to avoid
4000 : * repeated requests.
4001 : *
4002 : * writePtr is the location up to which the WAL is sent. It is essentially
4003 : * the same as sentPtr but in some cases, we need to send keep alive before
4004 : * sentPtr is updated like when skipping empty transactions.
4005 : */
4006 : static void
4007 5408 : WalSndKeepalive(bool requestReply, XLogRecPtr writePtr)
4008 : {
4009 5408 : elog(DEBUG2, "sending replication keepalive");
4010 :
4011 : /* construct the message... */
4012 5408 : resetStringInfo(&output_message);
4013 5408 : pq_sendbyte(&output_message, 'k');
4014 5408 : pq_sendint64(&output_message, XLogRecPtrIsInvalid(writePtr) ? sentPtr : writePtr);
4015 5408 : pq_sendint64(&output_message, GetCurrentTimestamp());
4016 5408 : pq_sendbyte(&output_message, requestReply ? 1 : 0);
4017 :
4018 : /* ... and send it wrapped in CopyData */
4019 5408 : pq_putmessage_noblock('d', output_message.data, output_message.len);
4020 :
4021 : /* Set local flag */
4022 5408 : if (requestReply)
4023 1904 : waiting_for_ping_response = true;
4024 5408 : }
4025 :
4026 : /*
4027 : * Send keepalive message if too much time has elapsed.
4028 : */
4029 : static void
4030 2582010 : WalSndKeepaliveIfNecessary(void)
4031 : {
4032 : TimestampTz ping_time;
4033 :
4034 : /*
4035 : * Don't send keepalive messages if timeouts are globally disabled or
4036 : * we're doing something not partaking in timeouts.
4037 : */
4038 2582010 : if (wal_sender_timeout <= 0 || last_reply_timestamp <= 0)
4039 46 : return;
4040 :
4041 2581964 : if (waiting_for_ping_response)
4042 5280 : return;
4043 :
4044 : /*
4045 : * If half of wal_sender_timeout has lapsed without receiving any reply
4046 : * from the standby, send a keep-alive message to the standby requesting
4047 : * an immediate reply.
4048 : */
4049 2576684 : ping_time = TimestampTzPlusMilliseconds(last_reply_timestamp,
4050 : wal_sender_timeout / 2);
4051 2576684 : if (last_processing >= ping_time)
4052 : {
4053 0 : WalSndKeepalive(true, InvalidXLogRecPtr);
4054 :
4055 : /* Try to flush pending output to the client */
4056 0 : if (pq_flush_if_writable() != 0)
4057 0 : WalSndShutdown();
4058 : }
4059 : }
4060 :
4061 : /*
4062 : * Record the end of the WAL and the time it was flushed locally, so that
4063 : * LagTrackerRead can compute the elapsed time (lag) when this WAL location is
4064 : * eventually reported to have been written, flushed and applied by the
4065 : * standby in a reply message.
4066 : */
4067 : static void
4068 188578 : LagTrackerWrite(XLogRecPtr lsn, TimestampTz local_flush_time)
4069 : {
4070 : bool buffer_full;
4071 : int new_write_head;
4072 : int i;
4073 :
4074 188578 : if (!am_walsender)
4075 0 : return;
4076 :
4077 : /*
4078 : * If the lsn hasn't advanced since last time, then do nothing. This way
4079 : * we only record a new sample when new WAL has been written.
4080 : */
4081 188578 : if (lag_tracker->last_lsn == lsn)
4082 181846 : return;
4083 6732 : lag_tracker->last_lsn = lsn;
4084 :
4085 : /*
4086 : * If advancing the write head of the circular buffer would crash into any
4087 : * of the read heads, then the buffer is full. In other words, the
4088 : * slowest reader (presumably apply) is the one that controls the release
4089 : * of space.
4090 : */
4091 6732 : new_write_head = (lag_tracker->write_head + 1) % LAG_TRACKER_BUFFER_SIZE;
4092 6732 : buffer_full = false;
4093 26928 : for (i = 0; i < NUM_SYNC_REP_WAIT_MODE; ++i)
4094 : {
4095 20196 : if (new_write_head == lag_tracker->read_heads[i])
4096 0 : buffer_full = true;
4097 : }
4098 :
4099 : /*
4100 : * If the buffer is full, for now we just rewind by one slot and overwrite
4101 : * the last sample, as a simple (if somewhat uneven) way to lower the
4102 : * sampling rate. There may be better adaptive compaction algorithms.
4103 : */
4104 6732 : if (buffer_full)
4105 : {
4106 0 : new_write_head = lag_tracker->write_head;
4107 0 : if (lag_tracker->write_head > 0)
4108 0 : lag_tracker->write_head--;
4109 : else
4110 0 : lag_tracker->write_head = LAG_TRACKER_BUFFER_SIZE - 1;
4111 : }
4112 :
4113 : /* Store a sample at the current write head position. */
4114 6732 : lag_tracker->buffer[lag_tracker->write_head].lsn = lsn;
4115 6732 : lag_tracker->buffer[lag_tracker->write_head].time = local_flush_time;
4116 6732 : lag_tracker->write_head = new_write_head;
4117 : }
4118 :
4119 : /*
4120 : * Find out how much time has elapsed between the moment WAL location 'lsn'
4121 : * (or the highest known earlier LSN) was flushed locally and the time 'now'.
4122 : * We have a separate read head for each of the reported LSN locations we
4123 : * receive in replies from standby; 'head' controls which read head is
4124 : * used. Whenever a read head crosses an LSN which was written into the
4125 : * lag buffer with LagTrackerWrite, we can use the associated timestamp to
4126 : * find out the time this LSN (or an earlier one) was flushed locally, and
4127 : * therefore compute the lag.
4128 : *
4129 : * Return -1 if no new sample data is available, and otherwise the elapsed
4130 : * time in microseconds.
4131 : */
4132 : static TimeOffset
4133 201126 : LagTrackerRead(int head, XLogRecPtr lsn, TimestampTz now)
4134 : {
4135 201126 : TimestampTz time = 0;
4136 :
4137 : /* Read all unread samples up to this LSN or end of buffer. */
4138 219376 : while (lag_tracker->read_heads[head] != lag_tracker->write_head &&
4139 68932 : lag_tracker->buffer[lag_tracker->read_heads[head]].lsn <= lsn)
4140 : {
4141 18250 : time = lag_tracker->buffer[lag_tracker->read_heads[head]].time;
4142 18250 : lag_tracker->last_read[head] =
4143 18250 : lag_tracker->buffer[lag_tracker->read_heads[head]];
4144 18250 : lag_tracker->read_heads[head] =
4145 18250 : (lag_tracker->read_heads[head] + 1) % LAG_TRACKER_BUFFER_SIZE;
4146 : }
4147 :
4148 : /*
4149 : * If the lag tracker is empty, that means the standby has processed
4150 : * everything we've ever sent so we should now clear 'last_read'. If we
4151 : * didn't do that, we'd risk using a stale and irrelevant sample for
4152 : * interpolation at the beginning of the next burst of WAL after a period
4153 : * of idleness.
4154 : */
4155 201126 : if (lag_tracker->read_heads[head] == lag_tracker->write_head)
4156 150444 : lag_tracker->last_read[head].time = 0;
4157 :
4158 201126 : if (time > now)
4159 : {
4160 : /* If the clock somehow went backwards, treat as not found. */
4161 0 : return -1;
4162 : }
4163 201126 : else if (time == 0)
4164 : {
4165 : /*
4166 : * We didn't cross a time. If there is a future sample that we
4167 : * haven't reached yet, and we've already reached at least one sample,
4168 : * let's interpolate the local flushed time. This is mainly useful
4169 : * for reporting a completely stuck apply position as having
4170 : * increasing lag, since otherwise we'd have to wait for it to
4171 : * eventually start moving again and cross one of our samples before
4172 : * we can show the lag increasing.
4173 : */
4174 190644 : if (lag_tracker->read_heads[head] == lag_tracker->write_head)
4175 : {
4176 : /* There are no future samples, so we can't interpolate. */
4177 142814 : return -1;
4178 : }
4179 47830 : else if (lag_tracker->last_read[head].time != 0)
4180 : {
4181 : /* We can interpolate between last_read and the next sample. */
4182 : double fraction;
4183 35424 : WalTimeSample prev = lag_tracker->last_read[head];
4184 35424 : WalTimeSample next = lag_tracker->buffer[lag_tracker->read_heads[head]];
4185 :
4186 35424 : if (lsn < prev.lsn)
4187 : {
4188 : /*
4189 : * Reported LSNs shouldn't normally go backwards, but it's
4190 : * possible when there is a timeline change. Treat as not
4191 : * found.
4192 : */
4193 0 : return -1;
4194 : }
4195 :
4196 : Assert(prev.lsn < next.lsn);
4197 :
4198 35424 : if (prev.time > next.time)
4199 : {
4200 : /* If the clock somehow went backwards, treat as not found. */
4201 0 : return -1;
4202 : }
4203 :
4204 : /* See how far we are between the previous and next samples. */
4205 35424 : fraction =
4206 35424 : (double) (lsn - prev.lsn) / (double) (next.lsn - prev.lsn);
4207 :
4208 : /* Scale the local flush time proportionally. */
4209 35424 : time = (TimestampTz)
4210 35424 : ((double) prev.time + (next.time - prev.time) * fraction);
4211 : }
4212 : else
4213 : {
4214 : /*
4215 : * We have only a future sample, implying that we were entirely
4216 : * caught up but and now there is a new burst of WAL and the
4217 : * standby hasn't processed the first sample yet. Until the
4218 : * standby reaches the future sample the best we can do is report
4219 : * the hypothetical lag if that sample were to be replayed now.
4220 : */
4221 12406 : time = lag_tracker->buffer[lag_tracker->read_heads[head]].time;
4222 : }
4223 : }
4224 :
4225 : /* Return the elapsed time since local flush time in microseconds. */
4226 : Assert(time != 0);
4227 58312 : return now - time;
4228 : }
|
