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) 2010-2024, 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 2060 : InitWalSender(void)
279 : {
280 2060 : am_cascading_walsender = RecoveryInProgress();
281 :
282 : /* Create a per-walsender data structure in shared memory */
283 2060 : InitWalSenderSlot();
284 :
285 : /* need resource owner for e.g. basebackups */
286 2060 : 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 2060 : MarkPostmasterChildWalSender();
296 2060 : 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 2060 : if (MyDatabaseId == InvalidOid)
305 : {
306 : Assert(MyProc->xmin == InvalidTransactionId);
307 888 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
308 888 : MyProc->statusFlags |= PROC_AFFECTS_ALL_HORIZONS;
309 888 : ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
310 888 : LWLockRelease(ProcArrayLock);
311 : }
312 :
313 : /* Initialize empty timestamp buffer for lag tracking. */
314 2060 : lag_tracker = MemoryContextAllocZero(TopMemoryContext, sizeof(LagTracker));
315 2060 : }
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 98 : WalSndErrorCleanup(void)
326 : {
327 98 : LWLockReleaseAll();
328 98 : ConditionVariableCancelSleep();
329 98 : pgstat_report_wait_end();
330 :
331 98 : if (xlogreader != NULL && xlogreader->seg.ws_file >= 0)
332 14 : wal_segment_close(xlogreader);
333 :
334 98 : if (MyReplicationSlot != NULL)
335 34 : ReplicationSlotRelease();
336 :
337 98 : ReplicationSlotCleanup(false);
338 :
339 98 : 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 98 : if (!IsTransactionOrTransactionBlock())
347 96 : ReleaseAuxProcessResources(false);
348 :
349 98 : if (got_STOPPING || got_SIGUSR2)
350 0 : proc_exit(0);
351 :
352 : /* Revert back to startup state */
353 98 : WalSndSetState(WALSNDSTATE_STARTUP);
354 98 : }
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 1294 : IdentifySystem(void)
378 : {
379 : char sysid[32];
380 : char xloc[MAXFNAMELEN];
381 : XLogRecPtr logptr;
382 1294 : char *dbname = NULL;
383 : DestReceiver *dest;
384 : TupOutputState *tstate;
385 : TupleDesc tupdesc;
386 : Datum values[4];
387 1294 : 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 1294 : snprintf(sysid, sizeof(sysid), UINT64_FORMAT,
397 : GetSystemIdentifier());
398 :
399 1294 : am_cascading_walsender = RecoveryInProgress();
400 1294 : if (am_cascading_walsender)
401 112 : logptr = GetStandbyFlushRecPtr(&currTLI);
402 : else
403 1182 : logptr = GetFlushRecPtr(&currTLI);
404 :
405 1294 : snprintf(xloc, sizeof(xloc), "%X/%X", LSN_FORMAT_ARGS(logptr));
406 :
407 1294 : if (MyDatabaseId != InvalidOid)
408 : {
409 408 : MemoryContext cur = CurrentMemoryContext;
410 :
411 : /* syscache access needs a transaction env. */
412 408 : StartTransactionCommand();
413 408 : dbname = get_database_name(MyDatabaseId);
414 : /* copy dbname out of TX context */
415 408 : dbname = MemoryContextStrdup(cur, dbname);
416 408 : CommitTransactionCommand();
417 : }
418 :
419 1294 : dest = CreateDestReceiver(DestRemoteSimple);
420 :
421 : /* need a tuple descriptor representing four columns */
422 1294 : tupdesc = CreateTemplateTupleDesc(4);
423 1294 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "systemid",
424 : TEXTOID, -1, 0);
425 1294 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "timeline",
426 : INT8OID, -1, 0);
427 1294 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 3, "xlogpos",
428 : TEXTOID, -1, 0);
429 1294 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 4, "dbname",
430 : TEXTOID, -1, 0);
431 :
432 : /* prepare for projection of tuples */
433 1294 : tstate = begin_tup_output_tupdesc(dest, tupdesc, &TTSOpsVirtual);
434 :
435 : /* column 1: system identifier */
436 1294 : values[0] = CStringGetTextDatum(sysid);
437 :
438 : /* column 2: timeline */
439 1294 : values[1] = Int64GetDatum(currTLI);
440 :
441 : /* column 3: wal location */
442 1294 : values[2] = CStringGetTextDatum(xloc);
443 :
444 : /* column 4: database name, or NULL if none */
445 1294 : if (dbname)
446 408 : values[3] = CStringGetTextDatum(dbname);
447 : else
448 886 : nulls[3] = true;
449 :
450 : /* send it to dest */
451 1294 : do_tup_output(tstate, values, nulls);
452 :
453 1294 : end_tup_output(tstate);
454 1294 : }
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 20 : UploadManifest(void)
648 : {
649 : MemoryContext mcxt;
650 : IncrementalBackupInfo *ib;
651 20 : 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 20 : CurrentResourceOwner = AuxProcessResourceOwner;
662 :
663 : /* Prepare to read manifest data into a temporary context. */
664 20 : mcxt = AllocSetContextCreate(CurrentMemoryContext,
665 : "incremental backup information",
666 : ALLOCSET_DEFAULT_SIZES);
667 20 : ib = CreateIncrementalBackupInfo(mcxt);
668 :
669 : /* Send a CopyInResponse message */
670 20 : pq_beginmessage(&buf, PqMsg_CopyInResponse);
671 20 : pq_sendbyte(&buf, 0);
672 20 : pq_sendint16(&buf, 0);
673 20 : pq_endmessage_reuse(&buf);
674 20 : pq_flush();
675 :
676 : /* Receive packets from client until done. */
677 78 : while (HandleUploadManifestPacket(&buf, &offset, ib))
678 : ;
679 :
680 : /* Finish up manifest processing. */
681 18 : 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 18 : if (uploaded_manifest_mcxt != NULL)
692 0 : MemoryContextDelete(uploaded_manifest_mcxt);
693 18 : MemoryContextSetParent(mcxt, CacheMemoryContext);
694 18 : uploaded_manifest = ib;
695 18 : uploaded_manifest_mcxt = mcxt;
696 :
697 : /* clean up the resource owner we created */
698 18 : ReleaseAuxProcessResources(true);
699 18 : }
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 78 : HandleUploadManifestPacket(StringInfo buf, off_t *offset,
714 : IncrementalBackupInfo *ib)
715 : {
716 : int mtype;
717 : int maxmsglen;
718 :
719 78 : HOLD_CANCEL_INTERRUPTS();
720 :
721 78 : pq_startmsgread();
722 78 : mtype = pq_getbyte();
723 78 : 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 78 : switch (mtype)
729 : {
730 60 : case 'd': /* CopyData */
731 60 : maxmsglen = PQ_LARGE_MESSAGE_LIMIT;
732 60 : break;
733 18 : case 'c': /* CopyDone */
734 : case 'f': /* CopyFail */
735 : case 'H': /* Flush */
736 : case 'S': /* Sync */
737 18 : maxmsglen = PQ_SMALL_MESSAGE_LIMIT;
738 18 : 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 78 : 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 78 : RESUME_CANCEL_INTERRUPTS();
754 :
755 : /* Process the message */
756 78 : switch (mtype)
757 : {
758 60 : case 'd': /* CopyData */
759 60 : AppendIncrementalManifestData(ib, buf->data, buf->len);
760 58 : return true;
761 :
762 18 : case 'c': /* CopyDone */
763 18 : 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 24 : FlushPtr = GetStandbyFlushRecPtr(&FlushTLI);
839 : else
840 484 : 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 :
956 292 : replication_active = false;
957 292 : if (got_STOPPING)
958 0 : proc_exit(0);
959 292 : WalSndSetState(WALSNDSTATE_STARTUP);
960 :
961 : Assert(streamingDoneSending && streamingDoneReceiving);
962 : }
963 :
964 292 : if (cmd->slotname)
965 260 : ReplicationSlotRelease();
966 :
967 : /*
968 : * Copy is finished now. Send a single-row result set indicating the next
969 : * timeline.
970 : */
971 292 : 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 292 : EndReplicationCommand("START_STREAMING");
1010 292 : }
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 23260 : 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 23260 : flushptr = WalSndWaitForWal(targetPagePtr + reqLen);
1035 :
1036 : /* Fail if not enough (implies we are going to shut down) */
1037 22946 : if (flushptr < targetPagePtr + reqLen)
1038 380 : 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 22566 : am_cascading_walsender = RecoveryInProgress();
1049 :
1050 22566 : if (am_cascading_walsender)
1051 616 : GetXLogReplayRecPtr(&currTLI);
1052 : else
1053 21950 : currTLI = GetWALInsertionTimeLine();
1054 :
1055 22566 : XLogReadDetermineTimeline(state, targetPagePtr, reqLen, currTLI);
1056 22566 : sendTimeLineIsHistoric = (state->currTLI != currTLI);
1057 22566 : sendTimeLine = state->currTLI;
1058 22566 : sendTimeLineValidUpto = state->currTLIValidUntil;
1059 22566 : sendTimeLineNextTLI = state->nextTLI;
1060 :
1061 22566 : if (targetPagePtr + XLOG_BLCKSZ <= flushptr)
1062 19064 : count = XLOG_BLCKSZ; /* more than one block available */
1063 : else
1064 3502 : count = flushptr - targetPagePtr; /* part of the page available */
1065 :
1066 : /* now actually read the data, we know it's there */
1067 22566 : 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 22566 : XLByteToSeg(targetPagePtr, segno, state->segcxt.ws_segsize);
1085 22566 : CheckXLogRemoved(segno, state->seg.ws_tli);
1086 :
1087 22566 : return count;
1088 : }
1089 :
1090 : /*
1091 : * Process extra options given to CREATE_REPLICATION_SLOT.
1092 : */
1093 : static void
1094 852 : parseCreateReplSlotOptions(CreateReplicationSlotCmd *cmd,
1095 : bool *reserve_wal,
1096 : CRSSnapshotAction *snapshot_action,
1097 : bool *two_phase, bool *failover)
1098 : {
1099 : ListCell *lc;
1100 852 : bool snapshot_action_given = false;
1101 852 : bool reserve_wal_given = false;
1102 852 : bool two_phase_given = false;
1103 852 : bool failover_given = false;
1104 :
1105 : /* Parse options */
1106 1718 : foreach(lc, cmd->options)
1107 : {
1108 866 : DefElem *defel = (DefElem *) lfirst(lc);
1109 :
1110 866 : if (strcmp(defel->defname, "snapshot") == 0)
1111 : {
1112 : char *action;
1113 :
1114 586 : 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 586 : action = defGetString(defel);
1120 586 : snapshot_action_given = true;
1121 :
1122 586 : if (strcmp(action, "export") == 0)
1123 0 : *snapshot_action = CRS_EXPORT_SNAPSHOT;
1124 586 : else if (strcmp(action, "nothing") == 0)
1125 242 : *snapshot_action = CRS_NOEXPORT_SNAPSHOT;
1126 344 : else if (strcmp(action, "use") == 0)
1127 344 : *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 280 : else if (strcmp(defel->defname, "reserve_wal") == 0)
1135 : {
1136 264 : 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 264 : reserve_wal_given = true;
1142 264 : *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 852 : }
1166 :
1167 : /*
1168 : * Create a new replication slot.
1169 : */
1170 : static void
1171 852 : CreateReplicationSlot(CreateReplicationSlotCmd *cmd)
1172 : {
1173 852 : const char *snapshot_name = NULL;
1174 : char xloc[MAXFNAMELEN];
1175 : char *slot_name;
1176 852 : bool reserve_wal = false;
1177 852 : bool two_phase = false;
1178 852 : bool failover = false;
1179 852 : CRSSnapshotAction snapshot_action = CRS_EXPORT_SNAPSHOT;
1180 : DestReceiver *dest;
1181 : TupOutputState *tstate;
1182 : TupleDesc tupdesc;
1183 : Datum values[4];
1184 852 : bool nulls[4] = {0};
1185 :
1186 : Assert(!MyReplicationSlot);
1187 :
1188 852 : parseCreateReplSlotOptions(cmd, &reserve_wal, &snapshot_action, &two_phase,
1189 : &failover);
1190 :
1191 852 : if (cmd->kind == REPLICATION_KIND_PHYSICAL)
1192 : {
1193 266 : ReplicationSlotCreate(cmd->slotname, false,
1194 266 : cmd->temporary ? RS_TEMPORARY : RS_PERSISTENT,
1195 : false, false, false);
1196 :
1197 264 : if (reserve_wal)
1198 : {
1199 262 : ReplicationSlotReserveWal();
1200 :
1201 262 : ReplicationSlotMarkDirty();
1202 :
1203 : /* Write this slot to disk if it's a permanent one. */
1204 262 : if (!cmd->temporary)
1205 6 : ReplicationSlotSave();
1206 : }
1207 : }
1208 : else
1209 : {
1210 : LogicalDecodingContext *ctx;
1211 586 : bool need_full_snapshot = false;
1212 :
1213 : Assert(cmd->kind == REPLICATION_KIND_LOGICAL);
1214 :
1215 586 : 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 586 : ReplicationSlotCreate(cmd->slotname, true,
1225 586 : 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 586 : 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 586 : else if (snapshot_action == CRS_USE_SNAPSHOT)
1243 : {
1244 344 : 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 344 : 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 344 : 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 344 : 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 344 : 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 344 : need_full_snapshot = true;
1274 : }
1275 :
1276 586 : ctx = CreateInitDecodingContext(cmd->plugin, NIL, need_full_snapshot,
1277 : InvalidXLogRecPtr,
1278 586 : 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 586 : last_reply_timestamp = 0;
1292 :
1293 : /* build initial snapshot, might take a while */
1294 586 : 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 586 : if (snapshot_action == CRS_EXPORT_SNAPSHOT)
1303 : {
1304 0 : snapshot_name = SnapBuildExportSnapshot(ctx->snapshot_builder);
1305 : }
1306 586 : else if (snapshot_action == CRS_USE_SNAPSHOT)
1307 : {
1308 : Snapshot snap;
1309 :
1310 344 : snap = SnapBuildInitialSnapshot(ctx->snapshot_builder);
1311 344 : RestoreTransactionSnapshot(snap, MyProc);
1312 : }
1313 :
1314 : /* don't need the decoding context anymore */
1315 586 : FreeDecodingContext(ctx);
1316 :
1317 586 : if (!cmd->temporary)
1318 586 : ReplicationSlotPersist();
1319 : }
1320 :
1321 850 : snprintf(xloc, sizeof(xloc), "%X/%X",
1322 850 : LSN_FORMAT_ARGS(MyReplicationSlot->data.confirmed_flush));
1323 :
1324 850 : 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 850 : tupdesc = CreateTemplateTupleDesc(4);
1334 850 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "slot_name",
1335 : TEXTOID, -1, 0);
1336 850 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "consistent_point",
1337 : TEXTOID, -1, 0);
1338 850 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 3, "snapshot_name",
1339 : TEXTOID, -1, 0);
1340 850 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 4, "output_plugin",
1341 : TEXTOID, -1, 0);
1342 :
1343 : /* prepare for projection of tuples */
1344 850 : tstate = begin_tup_output_tupdesc(dest, tupdesc, &TTSOpsVirtual);
1345 :
1346 : /* slot_name */
1347 850 : slot_name = NameStr(MyReplicationSlot->data.name);
1348 850 : values[0] = CStringGetTextDatum(slot_name);
1349 :
1350 : /* consistent wal location */
1351 850 : values[1] = CStringGetTextDatum(xloc);
1352 :
1353 : /* snapshot name, or NULL if none */
1354 850 : if (snapshot_name != NULL)
1355 0 : values[2] = CStringGetTextDatum(snapshot_name);
1356 : else
1357 850 : nulls[2] = true;
1358 :
1359 : /* plugin, or NULL if none */
1360 850 : if (cmd->plugin != NULL)
1361 586 : values[3] = CStringGetTextDatum(cmd->plugin);
1362 : else
1363 264 : nulls[3] = true;
1364 :
1365 : /* send it to dest */
1366 850 : do_tup_output(tstate, values, nulls);
1367 850 : end_tup_output(tstate);
1368 :
1369 850 : ReplicationSlotRelease();
1370 850 : }
1371 :
1372 : /*
1373 : * Get rid of a replication slot that is no longer wanted.
1374 : */
1375 : static void
1376 462 : DropReplicationSlot(DropReplicationSlotCmd *cmd)
1377 : {
1378 462 : ReplicationSlotDrop(cmd->slotname, !cmd->wait);
1379 462 : }
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 692 : StartLogicalReplication(StartReplicationCmd *cmd)
1428 : {
1429 : StringInfoData buf;
1430 : QueryCompletion qc;
1431 :
1432 : /* make sure that our requirements are still fulfilled */
1433 692 : CheckLogicalDecodingRequirements();
1434 :
1435 : Assert(!MyReplicationSlot);
1436 :
1437 688 : 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 688 : 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 676 : logical_decoding_ctx =
1459 688 : CreateDecodingContext(cmd->startpoint, cmd->options, false,
1460 688 : XL_ROUTINE(.page_read = logical_read_xlog_page,
1461 : .segment_open = WalSndSegmentOpen,
1462 : .segment_close = wal_segment_close),
1463 : WalSndPrepareWrite, WalSndWriteData,
1464 : WalSndUpdateProgress);
1465 676 : xlogreader = logical_decoding_ctx->reader;
1466 :
1467 676 : WalSndSetState(WALSNDSTATE_CATCHUP);
1468 :
1469 : /* Send a CopyBothResponse message, and start streaming */
1470 676 : pq_beginmessage(&buf, PqMsg_CopyBothResponse);
1471 676 : pq_sendbyte(&buf, 0);
1472 676 : pq_sendint16(&buf, 0);
1473 676 : pq_endmessage(&buf);
1474 676 : pq_flush();
1475 :
1476 : /* Start reading WAL from the oldest required WAL. */
1477 676 : XLogBeginRead(logical_decoding_ctx->reader,
1478 676 : 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 676 : sentPtr = MyReplicationSlot->data.confirmed_flush;
1485 :
1486 : /* Also update the sent position status in shared memory */
1487 676 : SpinLockAcquire(&MyWalSnd->mutex);
1488 676 : MyWalSnd->sentPtr = MyReplicationSlot->data.restart_lsn;
1489 676 : SpinLockRelease(&MyWalSnd->mutex);
1490 :
1491 676 : replication_active = true;
1492 :
1493 676 : SyncRepInitConfig();
1494 :
1495 : /* Main loop of walsender */
1496 676 : WalSndLoop(XLogSendLogical);
1497 :
1498 340 : FreeDecodingContext(logical_decoding_ctx);
1499 340 : ReplicationSlotRelease();
1500 :
1501 340 : replication_active = false;
1502 340 : if (got_STOPPING)
1503 0 : proc_exit(0);
1504 340 : WalSndSetState(WALSNDSTATE_STARTUP);
1505 :
1506 : /* Get out of COPY mode (CommandComplete). */
1507 340 : SetQueryCompletion(&qc, CMDTAG_COPY, 0);
1508 340 : EndCommand(&qc, DestRemote, false);
1509 340 : }
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 369522 : 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 369522 : if (!last_write)
1524 722 : lsn = InvalidXLogRecPtr;
1525 :
1526 369522 : resetStringInfo(ctx->out);
1527 :
1528 369522 : pq_sendbyte(ctx->out, 'w');
1529 369522 : pq_sendint64(ctx->out, lsn); /* dataStart */
1530 369522 : 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 369522 : pq_sendint64(ctx->out, 0); /* sendtime */
1537 369522 : }
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 369522 : 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 369522 : resetStringInfo(&tmpbuf);
1558 369522 : now = GetCurrentTimestamp();
1559 369522 : pq_sendint64(&tmpbuf, now);
1560 369522 : memcpy(&ctx->out->data[1 + sizeof(int64) + sizeof(int64)],
1561 369522 : tmpbuf.data, sizeof(int64));
1562 :
1563 : /* output previously gathered data in a CopyData packet */
1564 369522 : pq_putmessage_noblock('d', ctx->out->data, ctx->out->len);
1565 :
1566 369522 : CHECK_FOR_INTERRUPTS();
1567 :
1568 : /* Try to flush pending output to the client */
1569 369522 : if (pq_flush_if_writable() != 0)
1570 2 : WalSndShutdown();
1571 :
1572 : /* Try taking fast path unless we get too close to walsender timeout. */
1573 369520 : if (now < TimestampTzPlusMilliseconds(last_reply_timestamp,
1574 369520 : wal_sender_timeout / 2) &&
1575 369520 : !pq_is_send_pending())
1576 : {
1577 369404 : return;
1578 : }
1579 :
1580 : /* If we have pending write here, go to slow path */
1581 116 : 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 284 : ProcessPendingWrites(void)
1590 : {
1591 : for (;;)
1592 168 : {
1593 : long sleeptime;
1594 :
1595 : /* Check for input from the client */
1596 284 : ProcessRepliesIfAny();
1597 :
1598 : /* die if timeout was reached */
1599 284 : WalSndCheckTimeOut();
1600 :
1601 : /* Send keepalive if the time has come */
1602 284 : WalSndKeepaliveIfNecessary();
1603 :
1604 284 : if (!pq_is_send_pending())
1605 116 : break;
1606 :
1607 168 : sleeptime = WalSndComputeSleeptime(GetCurrentTimestamp());
1608 :
1609 : /* Sleep until something happens or we time out */
1610 168 : WalSndWait(WL_SOCKET_WRITEABLE | WL_SOCKET_READABLE, sleeptime,
1611 : WAIT_EVENT_WAL_SENDER_WRITE_DATA);
1612 :
1613 : /* Clear any already-pending wakeups */
1614 168 : ResetLatch(MyLatch);
1615 :
1616 168 : CHECK_FOR_INTERRUPTS();
1617 :
1618 : /* Process any requests or signals received recently */
1619 168 : 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 168 : if (pq_flush_if_writable() != 0)
1628 0 : WalSndShutdown();
1629 : }
1630 :
1631 : /* reactivate latch so WalSndLoop knows to continue */
1632 116 : SetLatch(MyLatch);
1633 116 : }
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 4616 : WalSndUpdateProgress(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid,
1644 : bool skipped_xact)
1645 : {
1646 : static TimestampTz sendTime = 0;
1647 4616 : TimestampTz now = GetCurrentTimestamp();
1648 4616 : bool pending_writes = false;
1649 4616 : 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 4616 : if (end_xact && TimestampDifferenceExceeds(sendTime, now,
1661 : WALSND_LOGICAL_LAG_TRACK_INTERVAL_MS))
1662 : {
1663 356 : LagTrackerWrite(lsn, now);
1664 356 : 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 4616 : if (skipped_xact &&
1676 560 : SyncRepRequested() &&
1677 560 : ((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 4616 : if (pending_writes || (!end_xact &&
1698 3076 : now >= TimestampTzPlusMilliseconds(last_reply_timestamp,
1699 : wal_sender_timeout / 2)))
1700 0 : ProcessPendingWrites();
1701 4616 : }
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 9770 : 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 9770 : if (RecoveryInProgress())
1718 110 : return;
1719 :
1720 9660 : 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 22672 : NeedToWaitForStandbys(XLogRecPtr flushed_lsn, uint32 *wait_event)
1734 : {
1735 22672 : int elevel = got_STOPPING ? ERROR : WARNING;
1736 : bool failover_slot;
1737 :
1738 22672 : 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 22672 : 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 22664 : *wait_event = 0;
1752 22664 : 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 30756 : 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 30756 : if (target_lsn > flushed_lsn)
1770 : {
1771 8184 : *wait_event = WAIT_EVENT_WAL_SENDER_WAIT_FOR_WAL;
1772 8184 : return true;
1773 : }
1774 :
1775 : /* Check if the standby slots have caught up to the flushed position */
1776 22572 : 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 23260 : WalSndWaitForWal(XLogRecPtr loc)
1794 : {
1795 : int wakeEvents;
1796 23260 : 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 23260 : if (!XLogRecPtrIsInvalid(RecentFlushPtr) &&
1806 22326 : !NeedToWaitForWal(loc, RecentFlushPtr, &wait_event))
1807 19274 : 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 4858 : {
1816 8844 : bool wait_for_standby_at_stop = false;
1817 : long sleeptime;
1818 :
1819 : /* Clear any already-pending wakeups */
1820 8844 : ResetLatch(MyLatch);
1821 :
1822 8844 : CHECK_FOR_INTERRUPTS();
1823 :
1824 : /* Process any requests or signals received recently */
1825 8832 : 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 8832 : 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 8530 : if (got_STOPPING)
1841 100 : 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 8530 : if (wait_event != WAIT_EVENT_WAIT_FOR_STANDBY_CONFIRMATION)
1850 : {
1851 8522 : if (!RecoveryInProgress())
1852 7748 : RecentFlushPtr = GetFlushRecPtr(NULL);
1853 : else
1854 774 : 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 8530 : if (got_STOPPING)
1866 : {
1867 100 : if (NeedToWaitForStandbys(RecentFlushPtr, &wait_event))
1868 0 : wait_for_standby_at_stop = true;
1869 : else
1870 100 : 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 8430 : if (MyWalSnd->flush < sentPtr &&
1882 4474 : MyWalSnd->write < sentPtr &&
1883 3330 : !waiting_for_ping_response)
1884 3330 : 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 8430 : if (!wait_for_standby_at_stop &&
1891 8430 : !NeedToWaitForWal(loc, RecentFlushPtr, &wait_event))
1892 3290 : 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 5140 : WalSndCaughtUp = true;
1899 :
1900 : /*
1901 : * Try to flush any pending output to the client.
1902 : */
1903 5140 : 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 5140 : if (streamingDoneReceiving && streamingDoneSending &&
1912 282 : !pq_is_send_pending())
1913 282 : break;
1914 :
1915 : /* die if timeout was reached */
1916 4858 : WalSndCheckTimeOut();
1917 :
1918 : /* Send keepalive if the time has come */
1919 4858 : 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 4858 : sleeptime = WalSndComputeSleeptime(GetCurrentTimestamp());
1929 :
1930 4858 : wakeEvents = WL_SOCKET_READABLE;
1931 :
1932 4858 : if (pq_is_send_pending())
1933 0 : wakeEvents |= WL_SOCKET_WRITEABLE;
1934 :
1935 : Assert(wait_event != 0);
1936 :
1937 4858 : WalSndWait(wakeEvents, sleeptime, wait_event);
1938 : }
1939 :
1940 : /* reactivate latch so WalSndLoop knows to continue */
1941 3672 : SetLatch(MyLatch);
1942 3672 : 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 9354 : exec_replication_command(const char *cmd_string)
1953 : {
1954 : int parse_rc;
1955 : Node *cmd_node;
1956 : const char *cmdtag;
1957 : MemoryContext cmd_context;
1958 : MemoryContext old_context;
1959 :
1960 : /*
1961 : * If WAL sender has been told that shutdown is getting close, switch its
1962 : * status accordingly to handle the next replication commands correctly.
1963 : */
1964 9354 : if (got_STOPPING)
1965 0 : WalSndSetState(WALSNDSTATE_STOPPING);
1966 :
1967 : /*
1968 : * Throw error if in stopping mode. We need prevent commands that could
1969 : * generate WAL while the shutdown checkpoint is being written. To be
1970 : * safe, we just prohibit all new commands.
1971 : */
1972 9354 : if (MyWalSnd->state == WALSNDSTATE_STOPPING)
1973 0 : ereport(ERROR,
1974 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
1975 : errmsg("cannot execute new commands while WAL sender is in stopping mode")));
1976 :
1977 : /*
1978 : * CREATE_REPLICATION_SLOT ... LOGICAL exports a snapshot until the next
1979 : * command arrives. Clean up the old stuff if there's anything.
1980 : */
1981 9354 : SnapBuildClearExportedSnapshot();
1982 :
1983 9354 : CHECK_FOR_INTERRUPTS();
1984 :
1985 : /*
1986 : * Prepare to parse and execute the command.
1987 : */
1988 9354 : cmd_context = AllocSetContextCreate(CurrentMemoryContext,
1989 : "Replication command context",
1990 : ALLOCSET_DEFAULT_SIZES);
1991 9354 : old_context = MemoryContextSwitchTo(cmd_context);
1992 :
1993 9354 : replication_scanner_init(cmd_string);
1994 :
1995 : /*
1996 : * Is it a WalSender command?
1997 : */
1998 9354 : if (!replication_scanner_is_replication_command())
1999 : {
2000 : /* Nope; clean up and get out. */
2001 4030 : replication_scanner_finish();
2002 :
2003 4030 : MemoryContextSwitchTo(old_context);
2004 4030 : MemoryContextDelete(cmd_context);
2005 :
2006 : /* XXX this is a pretty random place to make this check */
2007 4030 : if (MyDatabaseId == InvalidOid)
2008 0 : ereport(ERROR,
2009 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2010 : errmsg("cannot execute SQL commands in WAL sender for physical replication")));
2011 :
2012 : /* Tell the caller that this wasn't a WalSender command. */
2013 4030 : return false;
2014 : }
2015 :
2016 : /*
2017 : * Looks like a WalSender command, so parse it.
2018 : */
2019 5324 : parse_rc = replication_yyparse();
2020 5324 : if (parse_rc != 0)
2021 0 : ereport(ERROR,
2022 : (errcode(ERRCODE_SYNTAX_ERROR),
2023 : errmsg_internal("replication command parser returned %d",
2024 : parse_rc)));
2025 5324 : replication_scanner_finish();
2026 :
2027 5324 : cmd_node = replication_parse_result;
2028 :
2029 : /*
2030 : * Report query to various monitoring facilities. For this purpose, we
2031 : * report replication commands just like SQL commands.
2032 : */
2033 5324 : debug_query_string = cmd_string;
2034 :
2035 5324 : pgstat_report_activity(STATE_RUNNING, cmd_string);
2036 :
2037 : /*
2038 : * Log replication command if log_replication_commands is enabled. Even
2039 : * when it's disabled, log the command with DEBUG1 level for backward
2040 : * compatibility.
2041 : */
2042 5324 : ereport(log_replication_commands ? LOG : DEBUG1,
2043 : (errmsg("received replication command: %s", cmd_string)));
2044 :
2045 : /*
2046 : * Disallow replication commands in aborted transaction blocks.
2047 : */
2048 5324 : if (IsAbortedTransactionBlockState())
2049 0 : ereport(ERROR,
2050 : (errcode(ERRCODE_IN_FAILED_SQL_TRANSACTION),
2051 : errmsg("current transaction is aborted, "
2052 : "commands ignored until end of transaction block")));
2053 :
2054 5324 : CHECK_FOR_INTERRUPTS();
2055 :
2056 : /*
2057 : * Allocate buffers that will be used for each outgoing and incoming
2058 : * message. We do this just once per command to reduce palloc overhead.
2059 : */
2060 5324 : initStringInfo(&output_message);
2061 5324 : initStringInfo(&reply_message);
2062 5324 : initStringInfo(&tmpbuf);
2063 :
2064 5324 : switch (cmd_node->type)
2065 : {
2066 1294 : case T_IdentifySystemCmd:
2067 1294 : cmdtag = "IDENTIFY_SYSTEM";
2068 1294 : set_ps_display(cmdtag);
2069 1294 : IdentifySystem();
2070 1294 : EndReplicationCommand(cmdtag);
2071 1294 : break;
2072 :
2073 12 : case T_ReadReplicationSlotCmd:
2074 12 : cmdtag = "READ_REPLICATION_SLOT";
2075 12 : set_ps_display(cmdtag);
2076 12 : ReadReplicationSlot((ReadReplicationSlotCmd *) cmd_node);
2077 10 : EndReplicationCommand(cmdtag);
2078 10 : break;
2079 :
2080 348 : case T_BaseBackupCmd:
2081 348 : cmdtag = "BASE_BACKUP";
2082 348 : set_ps_display(cmdtag);
2083 348 : PreventInTransactionBlock(true, cmdtag);
2084 348 : SendBaseBackup((BaseBackupCmd *) cmd_node, uploaded_manifest);
2085 292 : EndReplicationCommand(cmdtag);
2086 292 : break;
2087 :
2088 852 : case T_CreateReplicationSlotCmd:
2089 852 : cmdtag = "CREATE_REPLICATION_SLOT";
2090 852 : set_ps_display(cmdtag);
2091 852 : CreateReplicationSlot((CreateReplicationSlotCmd *) cmd_node);
2092 850 : EndReplicationCommand(cmdtag);
2093 850 : break;
2094 :
2095 462 : case T_DropReplicationSlotCmd:
2096 462 : cmdtag = "DROP_REPLICATION_SLOT";
2097 462 : set_ps_display(cmdtag);
2098 462 : DropReplicationSlot((DropReplicationSlotCmd *) cmd_node);
2099 462 : EndReplicationCommand(cmdtag);
2100 462 : break;
2101 :
2102 12 : case T_AlterReplicationSlotCmd:
2103 12 : cmdtag = "ALTER_REPLICATION_SLOT";
2104 12 : set_ps_display(cmdtag);
2105 12 : AlterReplicationSlot((AlterReplicationSlotCmd *) cmd_node);
2106 8 : EndReplicationCommand(cmdtag);
2107 8 : break;
2108 :
2109 1204 : case T_StartReplicationCmd:
2110 : {
2111 1204 : StartReplicationCmd *cmd = (StartReplicationCmd *) cmd_node;
2112 :
2113 1204 : cmdtag = "START_REPLICATION";
2114 1204 : set_ps_display(cmdtag);
2115 1204 : PreventInTransactionBlock(true, cmdtag);
2116 :
2117 1204 : if (cmd->kind == REPLICATION_KIND_PHYSICAL)
2118 512 : StartReplication(cmd);
2119 : else
2120 692 : StartLogicalReplication(cmd);
2121 :
2122 : /* dupe, but necessary per libpqrcv_endstreaming */
2123 632 : EndReplicationCommand(cmdtag);
2124 :
2125 : Assert(xlogreader != NULL);
2126 632 : break;
2127 : }
2128 :
2129 28 : case T_TimeLineHistoryCmd:
2130 28 : cmdtag = "TIMELINE_HISTORY";
2131 28 : set_ps_display(cmdtag);
2132 28 : PreventInTransactionBlock(true, cmdtag);
2133 28 : SendTimeLineHistory((TimeLineHistoryCmd *) cmd_node);
2134 28 : EndReplicationCommand(cmdtag);
2135 28 : break;
2136 :
2137 1092 : case T_VariableShowStmt:
2138 : {
2139 1092 : DestReceiver *dest = CreateDestReceiver(DestRemoteSimple);
2140 1092 : VariableShowStmt *n = (VariableShowStmt *) cmd_node;
2141 :
2142 1092 : cmdtag = "SHOW";
2143 1092 : set_ps_display(cmdtag);
2144 :
2145 : /* syscache access needs a transaction environment */
2146 1092 : StartTransactionCommand();
2147 1092 : GetPGVariable(n->name, dest);
2148 1092 : CommitTransactionCommand();
2149 1092 : EndReplicationCommand(cmdtag);
2150 : }
2151 1092 : break;
2152 :
2153 20 : case T_UploadManifestCmd:
2154 20 : cmdtag = "UPLOAD_MANIFEST";
2155 20 : set_ps_display(cmdtag);
2156 20 : PreventInTransactionBlock(true, cmdtag);
2157 20 : UploadManifest();
2158 18 : EndReplicationCommand(cmdtag);
2159 18 : break;
2160 :
2161 0 : default:
2162 0 : elog(ERROR, "unrecognized replication command node tag: %u",
2163 : cmd_node->type);
2164 : }
2165 :
2166 : /* done */
2167 4686 : MemoryContextSwitchTo(old_context);
2168 4686 : MemoryContextDelete(cmd_context);
2169 :
2170 : /*
2171 : * We need not update ps display or pg_stat_activity, because PostgresMain
2172 : * will reset those to "idle". But we must reset debug_query_string to
2173 : * ensure it doesn't become a dangling pointer.
2174 : */
2175 4686 : debug_query_string = NULL;
2176 :
2177 4686 : return true;
2178 : }
2179 :
2180 : /*
2181 : * Process any incoming messages while streaming. Also checks if the remote
2182 : * end has closed the connection.
2183 : */
2184 : static void
2185 1944524 : ProcessRepliesIfAny(void)
2186 : {
2187 : unsigned char firstchar;
2188 : int maxmsglen;
2189 : int r;
2190 1944524 : bool received = false;
2191 :
2192 1944524 : last_processing = GetCurrentTimestamp();
2193 :
2194 : /*
2195 : * If we already received a CopyDone from the frontend, any subsequent
2196 : * message is the beginning of a new command, and should be processed in
2197 : * the main processing loop.
2198 : */
2199 1944524 : while (!streamingDoneReceiving)
2200 : {
2201 2008420 : pq_startmsgread();
2202 2008420 : r = pq_getbyte_if_available(&firstchar);
2203 2008420 : if (r < 0)
2204 : {
2205 : /* unexpected error or EOF */
2206 36 : ereport(COMMERROR,
2207 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2208 : errmsg("unexpected EOF on standby connection")));
2209 36 : proc_exit(0);
2210 : }
2211 2008384 : if (r == 0)
2212 : {
2213 : /* no data available without blocking */
2214 1942772 : pq_endmsgread();
2215 1942772 : break;
2216 : }
2217 :
2218 : /* Validate message type and set packet size limit */
2219 65612 : switch (firstchar)
2220 : {
2221 64560 : case PqMsg_CopyData:
2222 64560 : maxmsglen = PQ_LARGE_MESSAGE_LIMIT;
2223 64560 : break;
2224 1052 : case PqMsg_CopyDone:
2225 : case PqMsg_Terminate:
2226 1052 : maxmsglen = PQ_SMALL_MESSAGE_LIMIT;
2227 1052 : break;
2228 0 : default:
2229 0 : ereport(FATAL,
2230 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2231 : errmsg("invalid standby message type \"%c\"",
2232 : firstchar)));
2233 : maxmsglen = 0; /* keep compiler quiet */
2234 : break;
2235 : }
2236 :
2237 : /* Read the message contents */
2238 65612 : resetStringInfo(&reply_message);
2239 65612 : if (pq_getmessage(&reply_message, maxmsglen))
2240 : {
2241 0 : ereport(COMMERROR,
2242 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2243 : errmsg("unexpected EOF on standby connection")));
2244 0 : proc_exit(0);
2245 : }
2246 :
2247 : /* ... and process it */
2248 65612 : switch (firstchar)
2249 : {
2250 : /*
2251 : * 'd' means a standby reply wrapped in a CopyData packet.
2252 : */
2253 64560 : case PqMsg_CopyData:
2254 64560 : ProcessStandbyMessage();
2255 64560 : received = true;
2256 64560 : break;
2257 :
2258 : /*
2259 : * CopyDone means the standby requested to finish streaming.
2260 : * Reply with CopyDone, if we had not sent that already.
2261 : */
2262 632 : case PqMsg_CopyDone:
2263 632 : if (!streamingDoneSending)
2264 : {
2265 606 : pq_putmessage_noblock('c', NULL, 0);
2266 606 : streamingDoneSending = true;
2267 : }
2268 :
2269 632 : streamingDoneReceiving = true;
2270 632 : received = true;
2271 632 : break;
2272 :
2273 : /*
2274 : * 'X' means that the standby is closing down the socket.
2275 : */
2276 420 : case PqMsg_Terminate:
2277 420 : proc_exit(0);
2278 :
2279 2009716 : default:
2280 : Assert(false); /* NOT REACHED */
2281 : }
2282 : }
2283 :
2284 : /*
2285 : * Save the last reply timestamp if we've received at least one reply.
2286 : */
2287 1944068 : if (received)
2288 : {
2289 18176 : last_reply_timestamp = last_processing;
2290 18176 : waiting_for_ping_response = false;
2291 : }
2292 1944068 : }
2293 :
2294 : /*
2295 : * Process a status update message received from standby.
2296 : */
2297 : static void
2298 64560 : ProcessStandbyMessage(void)
2299 : {
2300 : char msgtype;
2301 :
2302 : /*
2303 : * Check message type from the first byte.
2304 : */
2305 64560 : msgtype = pq_getmsgbyte(&reply_message);
2306 :
2307 64560 : switch (msgtype)
2308 : {
2309 64296 : case 'r':
2310 64296 : ProcessStandbyReplyMessage();
2311 64296 : break;
2312 :
2313 264 : case 'h':
2314 264 : ProcessStandbyHSFeedbackMessage();
2315 264 : break;
2316 :
2317 0 : default:
2318 0 : ereport(COMMERROR,
2319 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2320 : errmsg("unexpected message type \"%c\"", msgtype)));
2321 0 : proc_exit(0);
2322 : }
2323 64560 : }
2324 :
2325 : /*
2326 : * Remember that a walreceiver just confirmed receipt of lsn `lsn`.
2327 : */
2328 : static void
2329 19756 : PhysicalConfirmReceivedLocation(XLogRecPtr lsn)
2330 : {
2331 19756 : bool changed = false;
2332 19756 : ReplicationSlot *slot = MyReplicationSlot;
2333 :
2334 : Assert(lsn != InvalidXLogRecPtr);
2335 19756 : SpinLockAcquire(&slot->mutex);
2336 19756 : if (slot->data.restart_lsn != lsn)
2337 : {
2338 9768 : changed = true;
2339 9768 : slot->data.restart_lsn = lsn;
2340 : }
2341 19756 : SpinLockRelease(&slot->mutex);
2342 :
2343 19756 : if (changed)
2344 : {
2345 9768 : ReplicationSlotMarkDirty();
2346 9768 : ReplicationSlotsComputeRequiredLSN();
2347 9768 : PhysicalWakeupLogicalWalSnd();
2348 : }
2349 :
2350 : /*
2351 : * One could argue that the slot should be saved to disk now, but that'd
2352 : * be energy wasted - the worst thing lost information could cause here is
2353 : * to give wrong information in a statistics view - we'll just potentially
2354 : * be more conservative in removing files.
2355 : */
2356 19756 : }
2357 :
2358 : /*
2359 : * Regular reply from standby advising of WAL locations on standby server.
2360 : */
2361 : static void
2362 64296 : ProcessStandbyReplyMessage(void)
2363 : {
2364 : XLogRecPtr writePtr,
2365 : flushPtr,
2366 : applyPtr;
2367 : bool replyRequested;
2368 : TimeOffset writeLag,
2369 : flushLag,
2370 : applyLag;
2371 : bool clearLagTimes;
2372 : TimestampTz now;
2373 : TimestampTz replyTime;
2374 :
2375 : static bool fullyAppliedLastTime = false;
2376 :
2377 : /* the caller already consumed the msgtype byte */
2378 64296 : writePtr = pq_getmsgint64(&reply_message);
2379 64296 : flushPtr = pq_getmsgint64(&reply_message);
2380 64296 : applyPtr = pq_getmsgint64(&reply_message);
2381 64296 : replyTime = pq_getmsgint64(&reply_message);
2382 64296 : replyRequested = pq_getmsgbyte(&reply_message);
2383 :
2384 64296 : if (message_level_is_interesting(DEBUG2))
2385 : {
2386 : char *replyTimeStr;
2387 :
2388 : /* Copy because timestamptz_to_str returns a static buffer */
2389 572 : replyTimeStr = pstrdup(timestamptz_to_str(replyTime));
2390 :
2391 572 : elog(DEBUG2, "write %X/%X flush %X/%X apply %X/%X%s reply_time %s",
2392 : LSN_FORMAT_ARGS(writePtr),
2393 : LSN_FORMAT_ARGS(flushPtr),
2394 : LSN_FORMAT_ARGS(applyPtr),
2395 : replyRequested ? " (reply requested)" : "",
2396 : replyTimeStr);
2397 :
2398 572 : pfree(replyTimeStr);
2399 : }
2400 :
2401 : /* See if we can compute the round-trip lag for these positions. */
2402 64296 : now = GetCurrentTimestamp();
2403 64296 : writeLag = LagTrackerRead(SYNC_REP_WAIT_WRITE, writePtr, now);
2404 64296 : flushLag = LagTrackerRead(SYNC_REP_WAIT_FLUSH, flushPtr, now);
2405 64296 : applyLag = LagTrackerRead(SYNC_REP_WAIT_APPLY, applyPtr, now);
2406 :
2407 : /*
2408 : * If the standby reports that it has fully replayed the WAL in two
2409 : * consecutive reply messages, then the second such message must result
2410 : * from wal_receiver_status_interval expiring on the standby. This is a
2411 : * convenient time to forget the lag times measured when it last
2412 : * wrote/flushed/applied a WAL record, to avoid displaying stale lag data
2413 : * until more WAL traffic arrives.
2414 : */
2415 64296 : clearLagTimes = false;
2416 64296 : if (applyPtr == sentPtr)
2417 : {
2418 4116 : if (fullyAppliedLastTime)
2419 1736 : clearLagTimes = true;
2420 4116 : fullyAppliedLastTime = true;
2421 : }
2422 : else
2423 60180 : fullyAppliedLastTime = false;
2424 :
2425 : /* Send a reply if the standby requested one. */
2426 64296 : if (replyRequested)
2427 0 : WalSndKeepalive(false, InvalidXLogRecPtr);
2428 :
2429 : /*
2430 : * Update shared state for this WalSender process based on reply data from
2431 : * standby.
2432 : */
2433 : {
2434 64296 : WalSnd *walsnd = MyWalSnd;
2435 :
2436 64296 : SpinLockAcquire(&walsnd->mutex);
2437 64296 : walsnd->write = writePtr;
2438 64296 : walsnd->flush = flushPtr;
2439 64296 : walsnd->apply = applyPtr;
2440 64296 : if (writeLag != -1 || clearLagTimes)
2441 21026 : walsnd->writeLag = writeLag;
2442 64296 : if (flushLag != -1 || clearLagTimes)
2443 23442 : walsnd->flushLag = flushLag;
2444 64296 : if (applyLag != -1 || clearLagTimes)
2445 24680 : walsnd->applyLag = applyLag;
2446 64296 : walsnd->replyTime = replyTime;
2447 64296 : SpinLockRelease(&walsnd->mutex);
2448 : }
2449 :
2450 64296 : if (!am_cascading_walsender)
2451 63720 : SyncRepReleaseWaiters();
2452 :
2453 : /*
2454 : * Advance our local xmin horizon when the client confirmed a flush.
2455 : */
2456 64296 : if (MyReplicationSlot && flushPtr != InvalidXLogRecPtr)
2457 : {
2458 60902 : if (SlotIsLogical(MyReplicationSlot))
2459 41146 : LogicalConfirmReceivedLocation(flushPtr);
2460 : else
2461 19756 : PhysicalConfirmReceivedLocation(flushPtr);
2462 : }
2463 64296 : }
2464 :
2465 : /* compute new replication slot xmin horizon if needed */
2466 : static void
2467 122 : PhysicalReplicationSlotNewXmin(TransactionId feedbackXmin, TransactionId feedbackCatalogXmin)
2468 : {
2469 122 : bool changed = false;
2470 122 : ReplicationSlot *slot = MyReplicationSlot;
2471 :
2472 122 : SpinLockAcquire(&slot->mutex);
2473 122 : MyProc->xmin = InvalidTransactionId;
2474 :
2475 : /*
2476 : * For physical replication we don't need the interlock provided by xmin
2477 : * and effective_xmin since the consequences of a missed increase are
2478 : * limited to query cancellations, so set both at once.
2479 : */
2480 122 : if (!TransactionIdIsNormal(slot->data.xmin) ||
2481 60 : !TransactionIdIsNormal(feedbackXmin) ||
2482 60 : TransactionIdPrecedes(slot->data.xmin, feedbackXmin))
2483 : {
2484 80 : changed = true;
2485 80 : slot->data.xmin = feedbackXmin;
2486 80 : slot->effective_xmin = feedbackXmin;
2487 : }
2488 122 : if (!TransactionIdIsNormal(slot->data.catalog_xmin) ||
2489 30 : !TransactionIdIsNormal(feedbackCatalogXmin) ||
2490 30 : TransactionIdPrecedes(slot->data.catalog_xmin, feedbackCatalogXmin))
2491 : {
2492 94 : changed = true;
2493 94 : slot->data.catalog_xmin = feedbackCatalogXmin;
2494 94 : slot->effective_catalog_xmin = feedbackCatalogXmin;
2495 : }
2496 122 : SpinLockRelease(&slot->mutex);
2497 :
2498 122 : if (changed)
2499 : {
2500 100 : ReplicationSlotMarkDirty();
2501 100 : ReplicationSlotsComputeRequiredXmin(false);
2502 : }
2503 122 : }
2504 :
2505 : /*
2506 : * Check that the provided xmin/epoch are sane, that is, not in the future
2507 : * and not so far back as to be already wrapped around.
2508 : *
2509 : * Epoch of nextXid should be same as standby, or if the counter has
2510 : * wrapped, then one greater than standby.
2511 : *
2512 : * This check doesn't care about whether clog exists for these xids
2513 : * at all.
2514 : */
2515 : static bool
2516 124 : TransactionIdInRecentPast(TransactionId xid, uint32 epoch)
2517 : {
2518 : FullTransactionId nextFullXid;
2519 : TransactionId nextXid;
2520 : uint32 nextEpoch;
2521 :
2522 124 : nextFullXid = ReadNextFullTransactionId();
2523 124 : nextXid = XidFromFullTransactionId(nextFullXid);
2524 124 : nextEpoch = EpochFromFullTransactionId(nextFullXid);
2525 :
2526 124 : if (xid <= nextXid)
2527 : {
2528 124 : if (epoch != nextEpoch)
2529 0 : return false;
2530 : }
2531 : else
2532 : {
2533 0 : if (epoch + 1 != nextEpoch)
2534 0 : return false;
2535 : }
2536 :
2537 124 : if (!TransactionIdPrecedesOrEquals(xid, nextXid))
2538 0 : return false; /* epoch OK, but it's wrapped around */
2539 :
2540 124 : return true;
2541 : }
2542 :
2543 : /*
2544 : * Hot Standby feedback
2545 : */
2546 : static void
2547 264 : ProcessStandbyHSFeedbackMessage(void)
2548 : {
2549 : TransactionId feedbackXmin;
2550 : uint32 feedbackEpoch;
2551 : TransactionId feedbackCatalogXmin;
2552 : uint32 feedbackCatalogEpoch;
2553 : TimestampTz replyTime;
2554 :
2555 : /*
2556 : * Decipher the reply message. The caller already consumed the msgtype
2557 : * byte. See XLogWalRcvSendHSFeedback() in walreceiver.c for the creation
2558 : * of this message.
2559 : */
2560 264 : replyTime = pq_getmsgint64(&reply_message);
2561 264 : feedbackXmin = pq_getmsgint(&reply_message, 4);
2562 264 : feedbackEpoch = pq_getmsgint(&reply_message, 4);
2563 264 : feedbackCatalogXmin = pq_getmsgint(&reply_message, 4);
2564 264 : feedbackCatalogEpoch = pq_getmsgint(&reply_message, 4);
2565 :
2566 264 : if (message_level_is_interesting(DEBUG2))
2567 : {
2568 : char *replyTimeStr;
2569 :
2570 : /* Copy because timestamptz_to_str returns a static buffer */
2571 8 : replyTimeStr = pstrdup(timestamptz_to_str(replyTime));
2572 :
2573 8 : elog(DEBUG2, "hot standby feedback xmin %u epoch %u, catalog_xmin %u epoch %u reply_time %s",
2574 : feedbackXmin,
2575 : feedbackEpoch,
2576 : feedbackCatalogXmin,
2577 : feedbackCatalogEpoch,
2578 : replyTimeStr);
2579 :
2580 8 : pfree(replyTimeStr);
2581 : }
2582 :
2583 : /*
2584 : * Update shared state for this WalSender process based on reply data from
2585 : * standby.
2586 : */
2587 : {
2588 264 : WalSnd *walsnd = MyWalSnd;
2589 :
2590 264 : SpinLockAcquire(&walsnd->mutex);
2591 264 : walsnd->replyTime = replyTime;
2592 264 : SpinLockRelease(&walsnd->mutex);
2593 : }
2594 :
2595 : /*
2596 : * Unset WalSender's xmins if the feedback message values are invalid.
2597 : * This happens when the downstream turned hot_standby_feedback off.
2598 : */
2599 264 : if (!TransactionIdIsNormal(feedbackXmin)
2600 182 : && !TransactionIdIsNormal(feedbackCatalogXmin))
2601 : {
2602 182 : MyProc->xmin = InvalidTransactionId;
2603 182 : if (MyReplicationSlot != NULL)
2604 42 : PhysicalReplicationSlotNewXmin(feedbackXmin, feedbackCatalogXmin);
2605 182 : return;
2606 : }
2607 :
2608 : /*
2609 : * Check that the provided xmin/epoch are sane, that is, not in the future
2610 : * and not so far back as to be already wrapped around. Ignore if not.
2611 : */
2612 82 : if (TransactionIdIsNormal(feedbackXmin) &&
2613 82 : !TransactionIdInRecentPast(feedbackXmin, feedbackEpoch))
2614 0 : return;
2615 :
2616 82 : if (TransactionIdIsNormal(feedbackCatalogXmin) &&
2617 42 : !TransactionIdInRecentPast(feedbackCatalogXmin, feedbackCatalogEpoch))
2618 0 : return;
2619 :
2620 : /*
2621 : * Set the WalSender's xmin equal to the standby's requested xmin, so that
2622 : * the xmin will be taken into account by GetSnapshotData() /
2623 : * ComputeXidHorizons(). This will hold back the removal of dead rows and
2624 : * thereby prevent the generation of cleanup conflicts on the standby
2625 : * server.
2626 : *
2627 : * There is a small window for a race condition here: although we just
2628 : * checked that feedbackXmin precedes nextXid, the nextXid could have
2629 : * gotten advanced between our fetching it and applying the xmin below,
2630 : * perhaps far enough to make feedbackXmin wrap around. In that case the
2631 : * xmin we set here would be "in the future" and have no effect. No point
2632 : * in worrying about this since it's too late to save the desired data
2633 : * anyway. Assuming that the standby sends us an increasing sequence of
2634 : * xmins, this could only happen during the first reply cycle, else our
2635 : * own xmin would prevent nextXid from advancing so far.
2636 : *
2637 : * We don't bother taking the ProcArrayLock here. Setting the xmin field
2638 : * is assumed atomic, and there's no real need to prevent concurrent
2639 : * horizon determinations. (If we're moving our xmin forward, this is
2640 : * obviously safe, and if we're moving it backwards, well, the data is at
2641 : * risk already since a VACUUM could already have determined the horizon.)
2642 : *
2643 : * If we're using a replication slot we reserve the xmin via that,
2644 : * otherwise via the walsender's PGPROC entry. We can only track the
2645 : * catalog xmin separately when using a slot, so we store the least of the
2646 : * two provided when not using a slot.
2647 : *
2648 : * XXX: It might make sense to generalize the ephemeral slot concept and
2649 : * always use the slot mechanism to handle the feedback xmin.
2650 : */
2651 82 : if (MyReplicationSlot != NULL) /* XXX: persistency configurable? */
2652 80 : PhysicalReplicationSlotNewXmin(feedbackXmin, feedbackCatalogXmin);
2653 : else
2654 : {
2655 2 : if (TransactionIdIsNormal(feedbackCatalogXmin)
2656 0 : && TransactionIdPrecedes(feedbackCatalogXmin, feedbackXmin))
2657 0 : MyProc->xmin = feedbackCatalogXmin;
2658 : else
2659 2 : MyProc->xmin = feedbackXmin;
2660 : }
2661 : }
2662 :
2663 : /*
2664 : * Compute how long send/receive loops should sleep.
2665 : *
2666 : * If wal_sender_timeout is enabled we want to wake up in time to send
2667 : * keepalives and to abort the connection if wal_sender_timeout has been
2668 : * reached.
2669 : */
2670 : static long
2671 84074 : WalSndComputeSleeptime(TimestampTz now)
2672 : {
2673 84074 : long sleeptime = 10000; /* 10 s */
2674 :
2675 84074 : if (wal_sender_timeout > 0 && last_reply_timestamp > 0)
2676 : {
2677 : TimestampTz wakeup_time;
2678 :
2679 : /*
2680 : * At the latest stop sleeping once wal_sender_timeout has been
2681 : * reached.
2682 : */
2683 84028 : wakeup_time = TimestampTzPlusMilliseconds(last_reply_timestamp,
2684 : wal_sender_timeout);
2685 :
2686 : /*
2687 : * If no ping has been sent yet, wakeup when it's time to do so.
2688 : * WalSndKeepaliveIfNecessary() wants to send a keepalive once half of
2689 : * the timeout passed without a response.
2690 : */
2691 84028 : if (!waiting_for_ping_response)
2692 84006 : wakeup_time = TimestampTzPlusMilliseconds(last_reply_timestamp,
2693 : wal_sender_timeout / 2);
2694 :
2695 : /* Compute relative time until wakeup. */
2696 84028 : sleeptime = TimestampDifferenceMilliseconds(now, wakeup_time);
2697 : }
2698 :
2699 84074 : return sleeptime;
2700 : }
2701 :
2702 : /*
2703 : * Check whether there have been responses by the client within
2704 : * wal_sender_timeout and shutdown if not. Using last_processing as the
2705 : * reference point avoids counting server-side stalls against the client.
2706 : * However, a long server-side stall can make WalSndKeepaliveIfNecessary()
2707 : * postdate last_processing by more than wal_sender_timeout. If that happens,
2708 : * the client must reply almost immediately to avoid a timeout. This rarely
2709 : * affects the default configuration, under which clients spontaneously send a
2710 : * message every standby_message_timeout = wal_sender_timeout/6 = 10s. We
2711 : * could eliminate that problem by recognizing timeout expiration at
2712 : * wal_sender_timeout/2 after the keepalive.
2713 : */
2714 : static void
2715 1939372 : WalSndCheckTimeOut(void)
2716 : {
2717 : TimestampTz timeout;
2718 :
2719 : /* don't bail out if we're doing something that doesn't require timeouts */
2720 1939372 : if (last_reply_timestamp <= 0)
2721 46 : return;
2722 :
2723 1939326 : timeout = TimestampTzPlusMilliseconds(last_reply_timestamp,
2724 : wal_sender_timeout);
2725 :
2726 1939326 : if (wal_sender_timeout > 0 && last_processing >= timeout)
2727 : {
2728 : /*
2729 : * Since typically expiration of replication timeout means
2730 : * communication problem, we don't send the error message to the
2731 : * standby.
2732 : */
2733 0 : ereport(COMMERROR,
2734 : (errmsg("terminating walsender process due to replication timeout")));
2735 :
2736 0 : WalSndShutdown();
2737 : }
2738 : }
2739 :
2740 : /* Main loop of walsender process that streams the WAL over Copy messages. */
2741 : static void
2742 1184 : WalSndLoop(WalSndSendDataCallback send_data)
2743 : {
2744 : /*
2745 : * Initialize the last reply timestamp. That enables timeout processing
2746 : * from hereon.
2747 : */
2748 1184 : last_reply_timestamp = GetCurrentTimestamp();
2749 1184 : waiting_for_ping_response = false;
2750 :
2751 : /*
2752 : * Loop until we reach the end of this timeline or the client requests to
2753 : * stop streaming.
2754 : */
2755 : for (;;)
2756 : {
2757 : /* Clear any already-pending wakeups */
2758 1935414 : ResetLatch(MyLatch);
2759 :
2760 1935414 : CHECK_FOR_INTERRUPTS();
2761 :
2762 : /* Process any requests or signals received recently */
2763 1935408 : if (ConfigReloadPending)
2764 : {
2765 34 : ConfigReloadPending = false;
2766 34 : ProcessConfigFile(PGC_SIGHUP);
2767 34 : SyncRepInitConfig();
2768 : }
2769 :
2770 : /* Check for input from the client */
2771 1935408 : ProcessRepliesIfAny();
2772 :
2773 : /*
2774 : * If we have received CopyDone from the client, sent CopyDone
2775 : * ourselves, and the output buffer is empty, it's time to exit
2776 : * streaming.
2777 : */
2778 1935254 : if (streamingDoneReceiving && streamingDoneSending &&
2779 1012 : !pq_is_send_pending())
2780 632 : break;
2781 :
2782 : /*
2783 : * If we don't have any pending data in the output buffer, try to send
2784 : * some more. If there is some, we don't bother to call send_data
2785 : * again until we've flushed it ... but we'd better assume we are not
2786 : * caught up.
2787 : */
2788 1934622 : if (!pq_is_send_pending())
2789 1863782 : send_data();
2790 : else
2791 70840 : WalSndCaughtUp = false;
2792 :
2793 : /* Try to flush pending output to the client */
2794 1934298 : if (pq_flush_if_writable() != 0)
2795 0 : WalSndShutdown();
2796 :
2797 : /* If nothing remains to be sent right now ... */
2798 1934298 : if (WalSndCaughtUp && !pq_is_send_pending())
2799 : {
2800 : /*
2801 : * If we're in catchup state, move to streaming. This is an
2802 : * important state change for users to know about, since before
2803 : * this point data loss might occur if the primary dies and we
2804 : * need to failover to the standby. The state change is also
2805 : * important for synchronous replication, since commits that
2806 : * started to wait at that point might wait for some time.
2807 : */
2808 32316 : if (MyWalSnd->state == WALSNDSTATE_CATCHUP)
2809 : {
2810 1128 : ereport(DEBUG1,
2811 : (errmsg_internal("\"%s\" has now caught up with upstream server",
2812 : application_name)));
2813 1128 : WalSndSetState(WALSNDSTATE_STREAMING);
2814 : }
2815 :
2816 : /*
2817 : * When SIGUSR2 arrives, we send any outstanding logs up to the
2818 : * shutdown checkpoint record (i.e., the latest record), wait for
2819 : * them to be replicated to the standby, and exit. This may be a
2820 : * normal termination at shutdown, or a promotion, the walsender
2821 : * is not sure which.
2822 : */
2823 32316 : if (got_SIGUSR2)
2824 116 : WalSndDone(send_data);
2825 : }
2826 :
2827 : /* Check for replication timeout. */
2828 1934230 : WalSndCheckTimeOut();
2829 :
2830 : /* Send keepalive if the time has come */
2831 1934230 : WalSndKeepaliveIfNecessary();
2832 :
2833 : /*
2834 : * Block if we have unsent data. XXX For logical replication, let
2835 : * WalSndWaitForWal() handle any other blocking; idle receivers need
2836 : * its additional actions. For physical replication, also block if
2837 : * caught up; its send_data does not block.
2838 : */
2839 1934230 : if ((WalSndCaughtUp && send_data != XLogSendLogical &&
2840 1953826 : !streamingDoneSending) ||
2841 1925854 : pq_is_send_pending())
2842 : {
2843 : long sleeptime;
2844 : int wakeEvents;
2845 :
2846 79048 : if (!streamingDoneReceiving)
2847 78990 : wakeEvents = WL_SOCKET_READABLE;
2848 : else
2849 58 : wakeEvents = 0;
2850 :
2851 : /*
2852 : * Use fresh timestamp, not last_processing, to reduce the chance
2853 : * of reaching wal_sender_timeout before sending a keepalive.
2854 : */
2855 79048 : sleeptime = WalSndComputeSleeptime(GetCurrentTimestamp());
2856 :
2857 79048 : if (pq_is_send_pending())
2858 70742 : wakeEvents |= WL_SOCKET_WRITEABLE;
2859 :
2860 : /* Sleep until something happens or we time out */
2861 79048 : WalSndWait(wakeEvents, sleeptime, WAIT_EVENT_WAL_SENDER_MAIN);
2862 : }
2863 : }
2864 632 : }
2865 :
2866 : /* Initialize a per-walsender data structure for this walsender process */
2867 : static void
2868 2060 : InitWalSenderSlot(void)
2869 : {
2870 : int i;
2871 :
2872 : /*
2873 : * WalSndCtl should be set up already (we inherit this by fork() or
2874 : * EXEC_BACKEND mechanism from the postmaster).
2875 : */
2876 : Assert(WalSndCtl != NULL);
2877 : Assert(MyWalSnd == NULL);
2878 :
2879 : /*
2880 : * Find a free walsender slot and reserve it. This must not fail due to
2881 : * the prior check for free WAL senders in InitProcess().
2882 : */
2883 2972 : for (i = 0; i < max_wal_senders; i++)
2884 : {
2885 2972 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
2886 :
2887 2972 : SpinLockAcquire(&walsnd->mutex);
2888 :
2889 2972 : if (walsnd->pid != 0)
2890 : {
2891 912 : SpinLockRelease(&walsnd->mutex);
2892 912 : continue;
2893 : }
2894 : else
2895 : {
2896 : /*
2897 : * Found a free slot. Reserve it for us.
2898 : */
2899 2060 : walsnd->pid = MyProcPid;
2900 2060 : walsnd->state = WALSNDSTATE_STARTUP;
2901 2060 : walsnd->sentPtr = InvalidXLogRecPtr;
2902 2060 : walsnd->needreload = false;
2903 2060 : walsnd->write = InvalidXLogRecPtr;
2904 2060 : walsnd->flush = InvalidXLogRecPtr;
2905 2060 : walsnd->apply = InvalidXLogRecPtr;
2906 2060 : walsnd->writeLag = -1;
2907 2060 : walsnd->flushLag = -1;
2908 2060 : walsnd->applyLag = -1;
2909 2060 : walsnd->sync_standby_priority = 0;
2910 2060 : walsnd->replyTime = 0;
2911 :
2912 : /*
2913 : * The kind assignment is done here and not in StartReplication()
2914 : * and StartLogicalReplication(). Indeed, the logical walsender
2915 : * needs to read WAL records (like snapshot of running
2916 : * transactions) during the slot creation. So it needs to be woken
2917 : * up based on its kind.
2918 : *
2919 : * The kind assignment could also be done in StartReplication(),
2920 : * StartLogicalReplication() and CREATE_REPLICATION_SLOT but it
2921 : * seems better to set it on one place.
2922 : */
2923 2060 : if (MyDatabaseId == InvalidOid)
2924 888 : walsnd->kind = REPLICATION_KIND_PHYSICAL;
2925 : else
2926 1172 : walsnd->kind = REPLICATION_KIND_LOGICAL;
2927 :
2928 2060 : SpinLockRelease(&walsnd->mutex);
2929 : /* don't need the lock anymore */
2930 2060 : MyWalSnd = (WalSnd *) walsnd;
2931 :
2932 2060 : break;
2933 : }
2934 : }
2935 :
2936 : Assert(MyWalSnd != NULL);
2937 :
2938 : /* Arrange to clean up at walsender exit */
2939 2060 : on_shmem_exit(WalSndKill, 0);
2940 2060 : }
2941 :
2942 : /* Destroy the per-walsender data structure for this walsender process */
2943 : static void
2944 2060 : WalSndKill(int code, Datum arg)
2945 : {
2946 2060 : WalSnd *walsnd = MyWalSnd;
2947 :
2948 : Assert(walsnd != NULL);
2949 :
2950 2060 : MyWalSnd = NULL;
2951 :
2952 2060 : SpinLockAcquire(&walsnd->mutex);
2953 : /* Mark WalSnd struct as no longer being in use. */
2954 2060 : walsnd->pid = 0;
2955 2060 : SpinLockRelease(&walsnd->mutex);
2956 2060 : }
2957 :
2958 : /* XLogReaderRoutine->segment_open callback */
2959 : static void
2960 3130 : WalSndSegmentOpen(XLogReaderState *state, XLogSegNo nextSegNo,
2961 : TimeLineID *tli_p)
2962 : {
2963 : char path[MAXPGPATH];
2964 :
2965 : /*-------
2966 : * When reading from a historic timeline, and there is a timeline switch
2967 : * within this segment, read from the WAL segment belonging to the new
2968 : * timeline.
2969 : *
2970 : * For example, imagine that this server is currently on timeline 5, and
2971 : * we're streaming timeline 4. The switch from timeline 4 to 5 happened at
2972 : * 0/13002088. In pg_wal, we have these files:
2973 : *
2974 : * ...
2975 : * 000000040000000000000012
2976 : * 000000040000000000000013
2977 : * 000000050000000000000013
2978 : * 000000050000000000000014
2979 : * ...
2980 : *
2981 : * In this situation, when requested to send the WAL from segment 0x13, on
2982 : * timeline 4, we read the WAL from file 000000050000000000000013. Archive
2983 : * recovery prefers files from newer timelines, so if the segment was
2984 : * restored from the archive on this server, the file belonging to the old
2985 : * timeline, 000000040000000000000013, might not exist. Their contents are
2986 : * equal up to the switchpoint, because at a timeline switch, the used
2987 : * portion of the old segment is copied to the new file.
2988 : */
2989 3130 : *tli_p = sendTimeLine;
2990 3130 : if (sendTimeLineIsHistoric)
2991 : {
2992 : XLogSegNo endSegNo;
2993 :
2994 26 : XLByteToSeg(sendTimeLineValidUpto, endSegNo, state->segcxt.ws_segsize);
2995 26 : if (nextSegNo == endSegNo)
2996 20 : *tli_p = sendTimeLineNextTLI;
2997 : }
2998 :
2999 3130 : XLogFilePath(path, *tli_p, nextSegNo, state->segcxt.ws_segsize);
3000 3130 : state->seg.ws_file = BasicOpenFile(path, O_RDONLY | PG_BINARY);
3001 3130 : if (state->seg.ws_file >= 0)
3002 3126 : return;
3003 :
3004 : /*
3005 : * If the file is not found, assume it's because the standby asked for a
3006 : * too old WAL segment that has already been removed or recycled.
3007 : */
3008 4 : if (errno == ENOENT)
3009 : {
3010 : char xlogfname[MAXFNAMELEN];
3011 4 : int save_errno = errno;
3012 :
3013 4 : XLogFileName(xlogfname, *tli_p, nextSegNo, wal_segment_size);
3014 4 : errno = save_errno;
3015 4 : ereport(ERROR,
3016 : (errcode_for_file_access(),
3017 : errmsg("requested WAL segment %s has already been removed",
3018 : xlogfname)));
3019 : }
3020 : else
3021 0 : ereport(ERROR,
3022 : (errcode_for_file_access(),
3023 : errmsg("could not open file \"%s\": %m",
3024 : path)));
3025 : }
3026 :
3027 : /*
3028 : * Send out the WAL in its normal physical/stored form.
3029 : *
3030 : * Read up to MAX_SEND_SIZE bytes of WAL that's been flushed to disk,
3031 : * but not yet sent to the client, and buffer it in the libpq output
3032 : * buffer.
3033 : *
3034 : * If there is no unsent WAL remaining, WalSndCaughtUp is set to true,
3035 : * otherwise WalSndCaughtUp is set to false.
3036 : */
3037 : static void
3038 206002 : XLogSendPhysical(void)
3039 : {
3040 : XLogRecPtr SendRqstPtr;
3041 : XLogRecPtr startptr;
3042 : XLogRecPtr endptr;
3043 : Size nbytes;
3044 : XLogSegNo segno;
3045 : WALReadError errinfo;
3046 : Size rbytes;
3047 :
3048 : /* If requested switch the WAL sender to the stopping state. */
3049 206002 : if (got_STOPPING)
3050 860 : WalSndSetState(WALSNDSTATE_STOPPING);
3051 :
3052 206002 : if (streamingDoneSending)
3053 : {
3054 19570 : WalSndCaughtUp = true;
3055 25164 : return;
3056 : }
3057 :
3058 : /* Figure out how far we can safely send the WAL. */
3059 186432 : if (sendTimeLineIsHistoric)
3060 : {
3061 : /*
3062 : * Streaming an old timeline that's in this server's history, but is
3063 : * not the one we're currently inserting or replaying. It can be
3064 : * streamed up to the point where we switched off that timeline.
3065 : */
3066 332 : SendRqstPtr = sendTimeLineValidUpto;
3067 : }
3068 186100 : else if (am_cascading_walsender)
3069 : {
3070 : TimeLineID SendRqstTLI;
3071 :
3072 : /*
3073 : * Streaming the latest timeline on a standby.
3074 : *
3075 : * Attempt to send all WAL that has already been replayed, so that we
3076 : * know it's valid. If we're receiving WAL through streaming
3077 : * replication, it's also OK to send any WAL that has been received
3078 : * but not replayed.
3079 : *
3080 : * The timeline we're recovering from can change, or we can be
3081 : * promoted. In either case, the current timeline becomes historic. We
3082 : * need to detect that so that we don't try to stream past the point
3083 : * where we switched to another timeline. We check for promotion or
3084 : * timeline switch after calculating FlushPtr, to avoid a race
3085 : * condition: if the timeline becomes historic just after we checked
3086 : * that it was still current, it's still be OK to stream it up to the
3087 : * FlushPtr that was calculated before it became historic.
3088 : */
3089 1480 : bool becameHistoric = false;
3090 :
3091 1480 : SendRqstPtr = GetStandbyFlushRecPtr(&SendRqstTLI);
3092 :
3093 1480 : if (!RecoveryInProgress())
3094 : {
3095 : /* We have been promoted. */
3096 2 : SendRqstTLI = GetWALInsertionTimeLine();
3097 2 : am_cascading_walsender = false;
3098 2 : becameHistoric = true;
3099 : }
3100 : else
3101 : {
3102 : /*
3103 : * Still a cascading standby. But is the timeline we're sending
3104 : * still the one recovery is recovering from?
3105 : */
3106 1478 : if (sendTimeLine != SendRqstTLI)
3107 0 : becameHistoric = true;
3108 : }
3109 :
3110 1480 : if (becameHistoric)
3111 : {
3112 : /*
3113 : * The timeline we were sending has become historic. Read the
3114 : * timeline history file of the new timeline to see where exactly
3115 : * we forked off from the timeline we were sending.
3116 : */
3117 : List *history;
3118 :
3119 2 : history = readTimeLineHistory(SendRqstTLI);
3120 2 : sendTimeLineValidUpto = tliSwitchPoint(sendTimeLine, history, &sendTimeLineNextTLI);
3121 :
3122 : Assert(sendTimeLine < sendTimeLineNextTLI);
3123 2 : list_free_deep(history);
3124 :
3125 2 : sendTimeLineIsHistoric = true;
3126 :
3127 2 : SendRqstPtr = sendTimeLineValidUpto;
3128 : }
3129 : }
3130 : else
3131 : {
3132 : /*
3133 : * Streaming the current timeline on a primary.
3134 : *
3135 : * Attempt to send all data that's already been written out and
3136 : * fsync'd to disk. We cannot go further than what's been written out
3137 : * given the current implementation of WALRead(). And in any case
3138 : * it's unsafe to send WAL that is not securely down to disk on the
3139 : * primary: if the primary subsequently crashes and restarts, standbys
3140 : * must not have applied any WAL that got lost on the primary.
3141 : */
3142 184620 : SendRqstPtr = GetFlushRecPtr(NULL);
3143 : }
3144 :
3145 : /*
3146 : * Record the current system time as an approximation of the time at which
3147 : * this WAL location was written for the purposes of lag tracking.
3148 : *
3149 : * In theory we could make XLogFlush() record a time in shmem whenever WAL
3150 : * is flushed and we could get that time as well as the LSN when we call
3151 : * GetFlushRecPtr() above (and likewise for the cascading standby
3152 : * equivalent), but rather than putting any new code into the hot WAL path
3153 : * it seems good enough to capture the time here. We should reach this
3154 : * after XLogFlush() runs WalSndWakeupProcessRequests(), and although that
3155 : * may take some time, we read the WAL flush pointer and take the time
3156 : * very close to together here so that we'll get a later position if it is
3157 : * still moving.
3158 : *
3159 : * Because LagTrackerWrite ignores samples when the LSN hasn't advanced,
3160 : * this gives us a cheap approximation for the WAL flush time for this
3161 : * LSN.
3162 : *
3163 : * Note that the LSN is not necessarily the LSN for the data contained in
3164 : * the present message; it's the end of the WAL, which might be further
3165 : * ahead. All the lag tracking machinery cares about is finding out when
3166 : * that arbitrary LSN is eventually reported as written, flushed and
3167 : * applied, so that it can measure the elapsed time.
3168 : */
3169 186432 : LagTrackerWrite(SendRqstPtr, GetCurrentTimestamp());
3170 :
3171 : /*
3172 : * If this is a historic timeline and we've reached the point where we
3173 : * forked to the next timeline, stop streaming.
3174 : *
3175 : * Note: We might already have sent WAL > sendTimeLineValidUpto. The
3176 : * startup process will normally replay all WAL that has been received
3177 : * from the primary, before promoting, but if the WAL streaming is
3178 : * terminated at a WAL page boundary, the valid portion of the timeline
3179 : * might end in the middle of a WAL record. We might've already sent the
3180 : * first half of that partial WAL record to the cascading standby, so that
3181 : * sentPtr > sendTimeLineValidUpto. That's OK; the cascading standby can't
3182 : * replay the partial WAL record either, so it can still follow our
3183 : * timeline switch.
3184 : */
3185 186432 : if (sendTimeLineIsHistoric && sendTimeLineValidUpto <= sentPtr)
3186 : {
3187 : /* close the current file. */
3188 26 : if (xlogreader->seg.ws_file >= 0)
3189 26 : wal_segment_close(xlogreader);
3190 :
3191 : /* Send CopyDone */
3192 26 : pq_putmessage_noblock('c', NULL, 0);
3193 26 : streamingDoneSending = true;
3194 :
3195 26 : WalSndCaughtUp = true;
3196 :
3197 26 : elog(DEBUG1, "walsender reached end of timeline at %X/%X (sent up to %X/%X)",
3198 : LSN_FORMAT_ARGS(sendTimeLineValidUpto),
3199 : LSN_FORMAT_ARGS(sentPtr));
3200 26 : return;
3201 : }
3202 :
3203 : /* Do we have any work to do? */
3204 : Assert(sentPtr <= SendRqstPtr);
3205 186406 : if (SendRqstPtr <= sentPtr)
3206 : {
3207 5568 : WalSndCaughtUp = true;
3208 5568 : return;
3209 : }
3210 :
3211 : /*
3212 : * Figure out how much to send in one message. If there's no more than
3213 : * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
3214 : * MAX_SEND_SIZE bytes, but round back to logfile or page boundary.
3215 : *
3216 : * The rounding is not only for performance reasons. Walreceiver relies on
3217 : * the fact that we never split a WAL record across two messages. Since a
3218 : * long WAL record is split at page boundary into continuation records,
3219 : * page boundary is always a safe cut-off point. We also assume that
3220 : * SendRqstPtr never points to the middle of a WAL record.
3221 : */
3222 180838 : startptr = sentPtr;
3223 180838 : endptr = startptr;
3224 180838 : endptr += MAX_SEND_SIZE;
3225 :
3226 : /* if we went beyond SendRqstPtr, back off */
3227 180838 : if (SendRqstPtr <= endptr)
3228 : {
3229 2948 : endptr = SendRqstPtr;
3230 2948 : if (sendTimeLineIsHistoric)
3231 24 : WalSndCaughtUp = false;
3232 : else
3233 2924 : WalSndCaughtUp = true;
3234 : }
3235 : else
3236 : {
3237 : /* round down to page boundary. */
3238 177890 : endptr -= (endptr % XLOG_BLCKSZ);
3239 177890 : WalSndCaughtUp = false;
3240 : }
3241 :
3242 180838 : nbytes = endptr - startptr;
3243 : Assert(nbytes <= MAX_SEND_SIZE);
3244 :
3245 : /*
3246 : * OK to read and send the slice.
3247 : */
3248 180838 : resetStringInfo(&output_message);
3249 180838 : pq_sendbyte(&output_message, 'w');
3250 :
3251 180838 : pq_sendint64(&output_message, startptr); /* dataStart */
3252 180838 : pq_sendint64(&output_message, SendRqstPtr); /* walEnd */
3253 180838 : pq_sendint64(&output_message, 0); /* sendtime, filled in last */
3254 :
3255 : /*
3256 : * Read the log directly into the output buffer to avoid extra memcpy
3257 : * calls.
3258 : */
3259 180838 : enlargeStringInfo(&output_message, nbytes);
3260 :
3261 180838 : retry:
3262 : /* attempt to read WAL from WAL buffers first */
3263 180838 : rbytes = WALReadFromBuffers(&output_message.data[output_message.len],
3264 180838 : startptr, nbytes, xlogreader->seg.ws_tli);
3265 180838 : output_message.len += rbytes;
3266 180838 : startptr += rbytes;
3267 180838 : nbytes -= rbytes;
3268 :
3269 : /* now read the remaining WAL from WAL file */
3270 180838 : if (nbytes > 0 &&
3271 179138 : !WALRead(xlogreader,
3272 179142 : &output_message.data[output_message.len],
3273 : startptr,
3274 : nbytes,
3275 179142 : xlogreader->seg.ws_tli, /* Pass the current TLI because
3276 : * only WalSndSegmentOpen controls
3277 : * whether new TLI is needed. */
3278 : &errinfo))
3279 0 : WALReadRaiseError(&errinfo);
3280 :
3281 : /* See logical_read_xlog_page(). */
3282 180834 : XLByteToSeg(startptr, segno, xlogreader->segcxt.ws_segsize);
3283 180834 : CheckXLogRemoved(segno, xlogreader->seg.ws_tli);
3284 :
3285 : /*
3286 : * During recovery, the currently-open WAL file might be replaced with the
3287 : * file of the same name retrieved from archive. So we always need to
3288 : * check what we read was valid after reading into the buffer. If it's
3289 : * invalid, we try to open and read the file again.
3290 : */
3291 180834 : if (am_cascading_walsender)
3292 : {
3293 1092 : WalSnd *walsnd = MyWalSnd;
3294 : bool reload;
3295 :
3296 1092 : SpinLockAcquire(&walsnd->mutex);
3297 1092 : reload = walsnd->needreload;
3298 1092 : walsnd->needreload = false;
3299 1092 : SpinLockRelease(&walsnd->mutex);
3300 :
3301 1092 : if (reload && xlogreader->seg.ws_file >= 0)
3302 : {
3303 0 : wal_segment_close(xlogreader);
3304 :
3305 0 : goto retry;
3306 : }
3307 : }
3308 :
3309 180834 : output_message.len += nbytes;
3310 180834 : output_message.data[output_message.len] = '\0';
3311 :
3312 : /*
3313 : * Fill the send timestamp last, so that it is taken as late as possible.
3314 : */
3315 180834 : resetStringInfo(&tmpbuf);
3316 180834 : pq_sendint64(&tmpbuf, GetCurrentTimestamp());
3317 180834 : memcpy(&output_message.data[1 + sizeof(int64) + sizeof(int64)],
3318 180834 : tmpbuf.data, sizeof(int64));
3319 :
3320 180834 : pq_putmessage_noblock('d', output_message.data, output_message.len);
3321 :
3322 180834 : sentPtr = endptr;
3323 :
3324 : /* Update shared memory status */
3325 : {
3326 180834 : WalSnd *walsnd = MyWalSnd;
3327 :
3328 180834 : SpinLockAcquire(&walsnd->mutex);
3329 180834 : walsnd->sentPtr = sentPtr;
3330 180834 : SpinLockRelease(&walsnd->mutex);
3331 : }
3332 :
3333 : /* Report progress of XLOG streaming in PS display */
3334 180834 : if (update_process_title)
3335 : {
3336 : char activitymsg[50];
3337 :
3338 180834 : snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%X",
3339 180834 : LSN_FORMAT_ARGS(sentPtr));
3340 180834 : set_ps_display(activitymsg);
3341 : }
3342 : }
3343 :
3344 : /*
3345 : * Stream out logically decoded data.
3346 : */
3347 : static void
3348 1657896 : XLogSendLogical(void)
3349 : {
3350 : XLogRecord *record;
3351 : char *errm;
3352 :
3353 : /*
3354 : * We'll use the current flush point to determine whether we've caught up.
3355 : * This variable is static in order to cache it across calls. Caching is
3356 : * helpful because GetFlushRecPtr() needs to acquire a heavily-contended
3357 : * spinlock.
3358 : */
3359 : static XLogRecPtr flushPtr = InvalidXLogRecPtr;
3360 :
3361 : /*
3362 : * Don't know whether we've caught up yet. We'll set WalSndCaughtUp to
3363 : * true in WalSndWaitForWal, if we're actually waiting. We also set to
3364 : * true if XLogReadRecord() had to stop reading but WalSndWaitForWal
3365 : * didn't wait - i.e. when we're shutting down.
3366 : */
3367 1657896 : WalSndCaughtUp = false;
3368 :
3369 1657896 : record = XLogReadRecord(logical_decoding_ctx->reader, &errm);
3370 :
3371 : /* xlog record was invalid */
3372 1657582 : if (errm != NULL)
3373 0 : elog(ERROR, "could not find record while sending logically-decoded data: %s",
3374 : errm);
3375 :
3376 1657582 : if (record != NULL)
3377 : {
3378 : /*
3379 : * Note the lack of any call to LagTrackerWrite() which is handled by
3380 : * WalSndUpdateProgress which is called by output plugin through
3381 : * logical decoding write api.
3382 : */
3383 1657202 : LogicalDecodingProcessRecord(logical_decoding_ctx, logical_decoding_ctx->reader);
3384 :
3385 1657196 : sentPtr = logical_decoding_ctx->reader->EndRecPtr;
3386 : }
3387 :
3388 : /*
3389 : * If first time through in this session, initialize flushPtr. Otherwise,
3390 : * we only need to update flushPtr if EndRecPtr is past it.
3391 : */
3392 1657576 : if (flushPtr == InvalidXLogRecPtr ||
3393 1656940 : logical_decoding_ctx->reader->EndRecPtr >= flushPtr)
3394 : {
3395 4506 : if (am_cascading_walsender)
3396 120 : flushPtr = GetStandbyFlushRecPtr(NULL);
3397 : else
3398 4386 : flushPtr = GetFlushRecPtr(NULL);
3399 : }
3400 :
3401 : /* If EndRecPtr is still past our flushPtr, it means we caught up. */
3402 1657576 : if (logical_decoding_ctx->reader->EndRecPtr >= flushPtr)
3403 2848 : WalSndCaughtUp = true;
3404 :
3405 : /*
3406 : * If we're caught up and have been requested to stop, have WalSndLoop()
3407 : * terminate the connection in an orderly manner, after writing out all
3408 : * the pending data.
3409 : */
3410 1657576 : if (WalSndCaughtUp && got_STOPPING)
3411 102 : got_SIGUSR2 = true;
3412 :
3413 : /* Update shared memory status */
3414 : {
3415 1657576 : WalSnd *walsnd = MyWalSnd;
3416 :
3417 1657576 : SpinLockAcquire(&walsnd->mutex);
3418 1657576 : walsnd->sentPtr = sentPtr;
3419 1657576 : SpinLockRelease(&walsnd->mutex);
3420 : }
3421 1657576 : }
3422 :
3423 : /*
3424 : * Shutdown if the sender is caught up.
3425 : *
3426 : * NB: This should only be called when the shutdown signal has been received
3427 : * from postmaster.
3428 : *
3429 : * Note that if we determine that there's still more data to send, this
3430 : * function will return control to the caller.
3431 : */
3432 : static void
3433 116 : WalSndDone(WalSndSendDataCallback send_data)
3434 : {
3435 : XLogRecPtr replicatedPtr;
3436 :
3437 : /* ... let's just be real sure we're caught up ... */
3438 116 : send_data();
3439 :
3440 : /*
3441 : * To figure out whether all WAL has successfully been replicated, check
3442 : * flush location if valid, write otherwise. Tools like pg_receivewal will
3443 : * usually (unless in synchronous mode) return an invalid flush location.
3444 : */
3445 232 : replicatedPtr = XLogRecPtrIsInvalid(MyWalSnd->flush) ?
3446 116 : MyWalSnd->write : MyWalSnd->flush;
3447 :
3448 116 : if (WalSndCaughtUp && sentPtr == replicatedPtr &&
3449 68 : !pq_is_send_pending())
3450 : {
3451 : QueryCompletion qc;
3452 :
3453 : /* Inform the standby that XLOG streaming is done */
3454 68 : SetQueryCompletion(&qc, CMDTAG_COPY, 0);
3455 68 : EndCommand(&qc, DestRemote, false);
3456 68 : pq_flush();
3457 :
3458 68 : proc_exit(0);
3459 : }
3460 48 : if (!waiting_for_ping_response)
3461 16 : WalSndKeepalive(true, InvalidXLogRecPtr);
3462 48 : }
3463 :
3464 : /*
3465 : * Returns the latest point in WAL that has been safely flushed to disk.
3466 : * This should only be called when in recovery.
3467 : *
3468 : * This is called either by cascading walsender to find WAL position to be sent
3469 : * to a cascaded standby or by slot synchronization operation to validate remote
3470 : * slot's lsn before syncing it locally.
3471 : *
3472 : * As a side-effect, *tli is updated to the TLI of the last
3473 : * replayed WAL record.
3474 : */
3475 : XLogRecPtr
3476 1798 : GetStandbyFlushRecPtr(TimeLineID *tli)
3477 : {
3478 : XLogRecPtr replayPtr;
3479 : TimeLineID replayTLI;
3480 : XLogRecPtr receivePtr;
3481 : TimeLineID receiveTLI;
3482 : XLogRecPtr result;
3483 :
3484 : Assert(am_cascading_walsender || IsSyncingReplicationSlots());
3485 :
3486 : /*
3487 : * We can safely send what's already been replayed. Also, if walreceiver
3488 : * is streaming WAL from the same timeline, we can send anything that it
3489 : * has streamed, but hasn't been replayed yet.
3490 : */
3491 :
3492 1798 : receivePtr = GetWalRcvFlushRecPtr(NULL, &receiveTLI);
3493 1798 : replayPtr = GetXLogReplayRecPtr(&replayTLI);
3494 :
3495 1798 : if (tli)
3496 1616 : *tli = replayTLI;
3497 :
3498 1798 : result = replayPtr;
3499 1798 : if (receiveTLI == replayTLI && receivePtr > replayPtr)
3500 96 : result = receivePtr;
3501 :
3502 1798 : return result;
3503 : }
3504 :
3505 : /*
3506 : * Request walsenders to reload the currently-open WAL file
3507 : */
3508 : void
3509 42 : WalSndRqstFileReload(void)
3510 : {
3511 : int i;
3512 :
3513 438 : for (i = 0; i < max_wal_senders; i++)
3514 : {
3515 396 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3516 :
3517 396 : SpinLockAcquire(&walsnd->mutex);
3518 396 : if (walsnd->pid == 0)
3519 : {
3520 396 : SpinLockRelease(&walsnd->mutex);
3521 396 : continue;
3522 : }
3523 0 : walsnd->needreload = true;
3524 0 : SpinLockRelease(&walsnd->mutex);
3525 : }
3526 42 : }
3527 :
3528 : /*
3529 : * Handle PROCSIG_WALSND_INIT_STOPPING signal.
3530 : */
3531 : void
3532 68 : HandleWalSndInitStopping(void)
3533 : {
3534 : Assert(am_walsender);
3535 :
3536 : /*
3537 : * If replication has not yet started, die like with SIGTERM. If
3538 : * replication is active, only set a flag and wake up the main loop. It
3539 : * will send any outstanding WAL, wait for it to be replicated to the
3540 : * standby, and then exit gracefully.
3541 : */
3542 68 : if (!replication_active)
3543 0 : kill(MyProcPid, SIGTERM);
3544 : else
3545 68 : got_STOPPING = true;
3546 68 : }
3547 :
3548 : /*
3549 : * SIGUSR2: set flag to do a last cycle and shut down afterwards. The WAL
3550 : * sender should already have been switched to WALSNDSTATE_STOPPING at
3551 : * this point.
3552 : */
3553 : static void
3554 52 : WalSndLastCycleHandler(SIGNAL_ARGS)
3555 : {
3556 52 : got_SIGUSR2 = true;
3557 52 : SetLatch(MyLatch);
3558 52 : }
3559 :
3560 : /* Set up signal handlers */
3561 : void
3562 2060 : WalSndSignals(void)
3563 : {
3564 : /* Set up signal handlers */
3565 2060 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
3566 2060 : pqsignal(SIGINT, StatementCancelHandler); /* query cancel */
3567 2060 : pqsignal(SIGTERM, die); /* request shutdown */
3568 : /* SIGQUIT handler was already set up by InitPostmasterChild */
3569 2060 : InitializeTimeouts(); /* establishes SIGALRM handler */
3570 2060 : pqsignal(SIGPIPE, SIG_IGN);
3571 2060 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
3572 2060 : pqsignal(SIGUSR2, WalSndLastCycleHandler); /* request a last cycle and
3573 : * shutdown */
3574 :
3575 : /* Reset some signals that are accepted by postmaster but not here */
3576 2060 : pqsignal(SIGCHLD, SIG_DFL);
3577 2060 : }
3578 :
3579 : /* Report shared-memory space needed by WalSndShmemInit */
3580 : Size
3581 7338 : WalSndShmemSize(void)
3582 : {
3583 7338 : Size size = 0;
3584 :
3585 7338 : size = offsetof(WalSndCtlData, walsnds);
3586 7338 : size = add_size(size, mul_size(max_wal_senders, sizeof(WalSnd)));
3587 :
3588 7338 : return size;
3589 : }
3590 :
3591 : /* Allocate and initialize walsender-related shared memory */
3592 : void
3593 1902 : WalSndShmemInit(void)
3594 : {
3595 : bool found;
3596 : int i;
3597 :
3598 1902 : WalSndCtl = (WalSndCtlData *)
3599 1902 : ShmemInitStruct("Wal Sender Ctl", WalSndShmemSize(), &found);
3600 :
3601 1902 : if (!found)
3602 : {
3603 : /* First time through, so initialize */
3604 13374 : MemSet(WalSndCtl, 0, WalSndShmemSize());
3605 :
3606 7608 : for (i = 0; i < NUM_SYNC_REP_WAIT_MODE; i++)
3607 5706 : dlist_init(&(WalSndCtl->SyncRepQueue[i]));
3608 :
3609 14642 : for (i = 0; i < max_wal_senders; i++)
3610 : {
3611 12740 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3612 :
3613 12740 : SpinLockInit(&walsnd->mutex);
3614 : }
3615 :
3616 1902 : ConditionVariableInit(&WalSndCtl->wal_flush_cv);
3617 1902 : ConditionVariableInit(&WalSndCtl->wal_replay_cv);
3618 1902 : ConditionVariableInit(&WalSndCtl->wal_confirm_rcv_cv);
3619 : }
3620 1902 : }
3621 :
3622 : /*
3623 : * Wake up physical, logical or both kinds of walsenders
3624 : *
3625 : * The distinction between physical and logical walsenders is done, because:
3626 : * - physical walsenders can't send data until it's been flushed
3627 : * - logical walsenders on standby can't decode and send data until it's been
3628 : * applied
3629 : *
3630 : * For cascading replication we need to wake up physical walsenders separately
3631 : * from logical walsenders (see the comment before calling WalSndWakeup() in
3632 : * ApplyWalRecord() for more details).
3633 : *
3634 : * This will be called inside critical sections, so throwing an error is not
3635 : * advisable.
3636 : */
3637 : void
3638 5118470 : WalSndWakeup(bool physical, bool logical)
3639 : {
3640 : /*
3641 : * Wake up all the walsenders waiting on WAL being flushed or replayed
3642 : * respectively. Note that waiting walsender would have prepared to sleep
3643 : * on the CV (i.e., added itself to the CV's waitlist) in WalSndWait()
3644 : * before actually waiting.
3645 : */
3646 5118470 : if (physical)
3647 198466 : ConditionVariableBroadcast(&WalSndCtl->wal_flush_cv);
3648 :
3649 5118470 : if (logical)
3650 5106254 : ConditionVariableBroadcast(&WalSndCtl->wal_replay_cv);
3651 5118470 : }
3652 :
3653 : /*
3654 : * Wait for readiness on the FeBe socket, or a timeout. The mask should be
3655 : * composed of optional WL_SOCKET_WRITEABLE and WL_SOCKET_READABLE flags. Exit
3656 : * on postmaster death.
3657 : */
3658 : static void
3659 84074 : WalSndWait(uint32 socket_events, long timeout, uint32 wait_event)
3660 : {
3661 : WaitEvent event;
3662 :
3663 84074 : ModifyWaitEvent(FeBeWaitSet, FeBeWaitSetSocketPos, socket_events, NULL);
3664 :
3665 : /*
3666 : * We use a condition variable to efficiently wake up walsenders in
3667 : * WalSndWakeup().
3668 : *
3669 : * Every walsender prepares to sleep on a shared memory CV. Note that it
3670 : * just prepares to sleep on the CV (i.e., adds itself to the CV's
3671 : * waitlist), but does not actually wait on the CV (IOW, it never calls
3672 : * ConditionVariableSleep()). It still uses WaitEventSetWait() for
3673 : * waiting, because we also need to wait for socket events. The processes
3674 : * (startup process, walreceiver etc.) wanting to wake up walsenders use
3675 : * ConditionVariableBroadcast(), which in turn calls SetLatch(), helping
3676 : * walsenders come out of WaitEventSetWait().
3677 : *
3678 : * This approach is simple and efficient because, one doesn't have to loop
3679 : * through all the walsenders slots, with a spinlock acquisition and
3680 : * release for every iteration, just to wake up only the waiting
3681 : * walsenders. It makes WalSndWakeup() callers' life easy.
3682 : *
3683 : * XXX: A desirable future improvement would be to add support for CVs
3684 : * into WaitEventSetWait().
3685 : *
3686 : * And, we use separate shared memory CVs for physical and logical
3687 : * walsenders for selective wake ups, see WalSndWakeup() for more details.
3688 : *
3689 : * If the wait event is WAIT_FOR_STANDBY_CONFIRMATION, wait on another CV
3690 : * until awakened by physical walsenders after the walreceiver confirms
3691 : * the receipt of the LSN.
3692 : */
3693 84074 : if (wait_event == WAIT_EVENT_WAIT_FOR_STANDBY_CONFIRMATION)
3694 8 : ConditionVariablePrepareToSleep(&WalSndCtl->wal_confirm_rcv_cv);
3695 84066 : else if (MyWalSnd->kind == REPLICATION_KIND_PHYSICAL)
3696 79038 : ConditionVariablePrepareToSleep(&WalSndCtl->wal_flush_cv);
3697 5028 : else if (MyWalSnd->kind == REPLICATION_KIND_LOGICAL)
3698 5028 : ConditionVariablePrepareToSleep(&WalSndCtl->wal_replay_cv);
3699 :
3700 84074 : if (WaitEventSetWait(FeBeWaitSet, timeout, &event, 1, wait_event) == 1 &&
3701 84074 : (event.events & WL_POSTMASTER_DEATH))
3702 : {
3703 0 : ConditionVariableCancelSleep();
3704 0 : proc_exit(1);
3705 : }
3706 :
3707 84074 : ConditionVariableCancelSleep();
3708 84074 : }
3709 :
3710 : /*
3711 : * Signal all walsenders to move to stopping state.
3712 : *
3713 : * This will trigger walsenders to move to a state where no further WAL can be
3714 : * generated. See this file's header for details.
3715 : */
3716 : void
3717 1086 : WalSndInitStopping(void)
3718 : {
3719 : int i;
3720 :
3721 8602 : for (i = 0; i < max_wal_senders; i++)
3722 : {
3723 7516 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3724 : pid_t pid;
3725 :
3726 7516 : SpinLockAcquire(&walsnd->mutex);
3727 7516 : pid = walsnd->pid;
3728 7516 : SpinLockRelease(&walsnd->mutex);
3729 :
3730 7516 : if (pid == 0)
3731 7448 : continue;
3732 :
3733 68 : SendProcSignal(pid, PROCSIG_WALSND_INIT_STOPPING, INVALID_PROC_NUMBER);
3734 : }
3735 1086 : }
3736 :
3737 : /*
3738 : * Wait that all the WAL senders have quit or reached the stopping state. This
3739 : * is used by the checkpointer to control when the shutdown checkpoint can
3740 : * safely be performed.
3741 : */
3742 : void
3743 1150 : WalSndWaitStopping(void)
3744 : {
3745 : for (;;)
3746 64 : {
3747 : int i;
3748 1150 : bool all_stopped = true;
3749 :
3750 8668 : for (i = 0; i < max_wal_senders; i++)
3751 : {
3752 7582 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3753 :
3754 7582 : SpinLockAcquire(&walsnd->mutex);
3755 :
3756 7582 : if (walsnd->pid == 0)
3757 : {
3758 7466 : SpinLockRelease(&walsnd->mutex);
3759 7466 : continue;
3760 : }
3761 :
3762 116 : if (walsnd->state != WALSNDSTATE_STOPPING)
3763 : {
3764 64 : all_stopped = false;
3765 64 : SpinLockRelease(&walsnd->mutex);
3766 64 : break;
3767 : }
3768 52 : SpinLockRelease(&walsnd->mutex);
3769 : }
3770 :
3771 : /* safe to leave if confirmation is done for all WAL senders */
3772 1150 : if (all_stopped)
3773 1086 : return;
3774 :
3775 64 : pg_usleep(10000L); /* wait for 10 msec */
3776 : }
3777 : }
3778 :
3779 : /* Set state for current walsender (only called in walsender) */
3780 : void
3781 4220 : WalSndSetState(WalSndState state)
3782 : {
3783 4220 : WalSnd *walsnd = MyWalSnd;
3784 :
3785 : Assert(am_walsender);
3786 :
3787 4220 : if (walsnd->state == state)
3788 872 : return;
3789 :
3790 3348 : SpinLockAcquire(&walsnd->mutex);
3791 3348 : walsnd->state = state;
3792 3348 : SpinLockRelease(&walsnd->mutex);
3793 : }
3794 :
3795 : /*
3796 : * Return a string constant representing the state. This is used
3797 : * in system views, and should *not* be translated.
3798 : */
3799 : static const char *
3800 1522 : WalSndGetStateString(WalSndState state)
3801 : {
3802 1522 : switch (state)
3803 : {
3804 8 : case WALSNDSTATE_STARTUP:
3805 8 : return "startup";
3806 0 : case WALSNDSTATE_BACKUP:
3807 0 : return "backup";
3808 16 : case WALSNDSTATE_CATCHUP:
3809 16 : return "catchup";
3810 1498 : case WALSNDSTATE_STREAMING:
3811 1498 : return "streaming";
3812 0 : case WALSNDSTATE_STOPPING:
3813 0 : return "stopping";
3814 : }
3815 0 : return "UNKNOWN";
3816 : }
3817 :
3818 : static Interval *
3819 2428 : offset_to_interval(TimeOffset offset)
3820 : {
3821 2428 : Interval *result = palloc(sizeof(Interval));
3822 :
3823 2428 : result->month = 0;
3824 2428 : result->day = 0;
3825 2428 : result->time = offset;
3826 :
3827 2428 : return result;
3828 : }
3829 :
3830 : /*
3831 : * Returns activity of walsenders, including pids and xlog locations sent to
3832 : * standby servers.
3833 : */
3834 : Datum
3835 1296 : pg_stat_get_wal_senders(PG_FUNCTION_ARGS)
3836 : {
3837 : #define PG_STAT_GET_WAL_SENDERS_COLS 12
3838 1296 : ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
3839 : SyncRepStandbyData *sync_standbys;
3840 : int num_standbys;
3841 : int i;
3842 :
3843 1296 : InitMaterializedSRF(fcinfo, 0);
3844 :
3845 : /*
3846 : * Get the currently active synchronous standbys. This could be out of
3847 : * date before we're done, but we'll use the data anyway.
3848 : */
3849 1296 : num_standbys = SyncRepGetCandidateStandbys(&sync_standbys);
3850 :
3851 13868 : for (i = 0; i < max_wal_senders; i++)
3852 : {
3853 12572 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3854 : XLogRecPtr sent_ptr;
3855 : XLogRecPtr write;
3856 : XLogRecPtr flush;
3857 : XLogRecPtr apply;
3858 : TimeOffset writeLag;
3859 : TimeOffset flushLag;
3860 : TimeOffset applyLag;
3861 : int priority;
3862 : int pid;
3863 : WalSndState state;
3864 : TimestampTz replyTime;
3865 : bool is_sync_standby;
3866 : Datum values[PG_STAT_GET_WAL_SENDERS_COLS];
3867 12572 : bool nulls[PG_STAT_GET_WAL_SENDERS_COLS] = {0};
3868 : int j;
3869 :
3870 : /* Collect data from shared memory */
3871 12572 : SpinLockAcquire(&walsnd->mutex);
3872 12572 : if (walsnd->pid == 0)
3873 : {
3874 11050 : SpinLockRelease(&walsnd->mutex);
3875 11050 : continue;
3876 : }
3877 1522 : pid = walsnd->pid;
3878 1522 : sent_ptr = walsnd->sentPtr;
3879 1522 : state = walsnd->state;
3880 1522 : write = walsnd->write;
3881 1522 : flush = walsnd->flush;
3882 1522 : apply = walsnd->apply;
3883 1522 : writeLag = walsnd->writeLag;
3884 1522 : flushLag = walsnd->flushLag;
3885 1522 : applyLag = walsnd->applyLag;
3886 1522 : priority = walsnd->sync_standby_priority;
3887 1522 : replyTime = walsnd->replyTime;
3888 1522 : SpinLockRelease(&walsnd->mutex);
3889 :
3890 : /*
3891 : * Detect whether walsender is/was considered synchronous. We can
3892 : * provide some protection against stale data by checking the PID
3893 : * along with walsnd_index.
3894 : */
3895 1522 : is_sync_standby = false;
3896 1610 : for (j = 0; j < num_standbys; j++)
3897 : {
3898 148 : if (sync_standbys[j].walsnd_index == i &&
3899 60 : sync_standbys[j].pid == pid)
3900 : {
3901 60 : is_sync_standby = true;
3902 60 : break;
3903 : }
3904 : }
3905 :
3906 1522 : values[0] = Int32GetDatum(pid);
3907 :
3908 1522 : if (!has_privs_of_role(GetUserId(), ROLE_PG_READ_ALL_STATS))
3909 : {
3910 : /*
3911 : * Only superusers and roles with privileges of pg_read_all_stats
3912 : * can see details. Other users only get the pid value to know
3913 : * it's a walsender, but no details.
3914 : */
3915 0 : MemSet(&nulls[1], true, PG_STAT_GET_WAL_SENDERS_COLS - 1);
3916 : }
3917 : else
3918 : {
3919 1522 : values[1] = CStringGetTextDatum(WalSndGetStateString(state));
3920 :
3921 1522 : if (XLogRecPtrIsInvalid(sent_ptr))
3922 8 : nulls[2] = true;
3923 1522 : values[2] = LSNGetDatum(sent_ptr);
3924 :
3925 1522 : if (XLogRecPtrIsInvalid(write))
3926 14 : nulls[3] = true;
3927 1522 : values[3] = LSNGetDatum(write);
3928 :
3929 1522 : if (XLogRecPtrIsInvalid(flush))
3930 14 : nulls[4] = true;
3931 1522 : values[4] = LSNGetDatum(flush);
3932 :
3933 1522 : if (XLogRecPtrIsInvalid(apply))
3934 14 : nulls[5] = true;
3935 1522 : values[5] = LSNGetDatum(apply);
3936 :
3937 : /*
3938 : * Treat a standby such as a pg_basebackup background process
3939 : * which always returns an invalid flush location, as an
3940 : * asynchronous standby.
3941 : */
3942 1522 : priority = XLogRecPtrIsInvalid(flush) ? 0 : priority;
3943 :
3944 1522 : if (writeLag < 0)
3945 754 : nulls[6] = true;
3946 : else
3947 768 : values[6] = IntervalPGetDatum(offset_to_interval(writeLag));
3948 :
3949 1522 : if (flushLag < 0)
3950 630 : nulls[7] = true;
3951 : else
3952 892 : values[7] = IntervalPGetDatum(offset_to_interval(flushLag));
3953 :
3954 1522 : if (applyLag < 0)
3955 754 : nulls[8] = true;
3956 : else
3957 768 : values[8] = IntervalPGetDatum(offset_to_interval(applyLag));
3958 :
3959 1522 : values[9] = Int32GetDatum(priority);
3960 :
3961 : /*
3962 : * More easily understood version of standby state. This is purely
3963 : * informational.
3964 : *
3965 : * In quorum-based sync replication, the role of each standby
3966 : * listed in synchronous_standby_names can be changing very
3967 : * frequently. Any standbys considered as "sync" at one moment can
3968 : * be switched to "potential" ones at the next moment. So, it's
3969 : * basically useless to report "sync" or "potential" as their sync
3970 : * states. We report just "quorum" for them.
3971 : */
3972 1522 : if (priority == 0)
3973 1438 : values[10] = CStringGetTextDatum("async");
3974 84 : else if (is_sync_standby)
3975 60 : values[10] = SyncRepConfig->syncrep_method == SYNC_REP_PRIORITY ?
3976 60 : CStringGetTextDatum("sync") : CStringGetTextDatum("quorum");
3977 : else
3978 24 : values[10] = CStringGetTextDatum("potential");
3979 :
3980 1522 : if (replyTime == 0)
3981 8 : nulls[11] = true;
3982 : else
3983 1514 : values[11] = TimestampTzGetDatum(replyTime);
3984 : }
3985 :
3986 1522 : tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc,
3987 : values, nulls);
3988 : }
3989 :
3990 1296 : return (Datum) 0;
3991 : }
3992 :
3993 : /*
3994 : * Send a keepalive message to standby.
3995 : *
3996 : * If requestReply is set, the message requests the other party to send
3997 : * a message back to us, for heartbeat purposes. We also set a flag to
3998 : * let nearby code know that we're waiting for that response, to avoid
3999 : * repeated requests.
4000 : *
4001 : * writePtr is the location up to which the WAL is sent. It is essentially
4002 : * the same as sentPtr but in some cases, we need to send keep alive before
4003 : * sentPtr is updated like when skipping empty transactions.
4004 : */
4005 : static void
4006 3346 : WalSndKeepalive(bool requestReply, XLogRecPtr writePtr)
4007 : {
4008 3346 : elog(DEBUG2, "sending replication keepalive");
4009 :
4010 : /* construct the message... */
4011 3346 : resetStringInfo(&output_message);
4012 3346 : pq_sendbyte(&output_message, 'k');
4013 3346 : pq_sendint64(&output_message, XLogRecPtrIsInvalid(writePtr) ? sentPtr : writePtr);
4014 3346 : pq_sendint64(&output_message, GetCurrentTimestamp());
4015 3346 : pq_sendbyte(&output_message, requestReply ? 1 : 0);
4016 :
4017 : /* ... and send it wrapped in CopyData */
4018 3346 : pq_putmessage_noblock('d', output_message.data, output_message.len);
4019 :
4020 : /* Set local flag */
4021 3346 : if (requestReply)
4022 16 : waiting_for_ping_response = true;
4023 3346 : }
4024 :
4025 : /*
4026 : * Send keepalive message if too much time has elapsed.
4027 : */
4028 : static void
4029 1939372 : WalSndKeepaliveIfNecessary(void)
4030 : {
4031 : TimestampTz ping_time;
4032 :
4033 : /*
4034 : * Don't send keepalive messages if timeouts are globally disabled or
4035 : * we're doing something not partaking in timeouts.
4036 : */
4037 1939372 : if (wal_sender_timeout <= 0 || last_reply_timestamp <= 0)
4038 46 : return;
4039 :
4040 1939326 : if (waiting_for_ping_response)
4041 64 : return;
4042 :
4043 : /*
4044 : * If half of wal_sender_timeout has lapsed without receiving any reply
4045 : * from the standby, send a keep-alive message to the standby requesting
4046 : * an immediate reply.
4047 : */
4048 1939262 : ping_time = TimestampTzPlusMilliseconds(last_reply_timestamp,
4049 : wal_sender_timeout / 2);
4050 1939262 : if (last_processing >= ping_time)
4051 : {
4052 0 : WalSndKeepalive(true, InvalidXLogRecPtr);
4053 :
4054 : /* Try to flush pending output to the client */
4055 0 : if (pq_flush_if_writable() != 0)
4056 0 : WalSndShutdown();
4057 : }
4058 : }
4059 :
4060 : /*
4061 : * Record the end of the WAL and the time it was flushed locally, so that
4062 : * LagTrackerRead can compute the elapsed time (lag) when this WAL location is
4063 : * eventually reported to have been written, flushed and applied by the
4064 : * standby in a reply message.
4065 : */
4066 : static void
4067 186788 : LagTrackerWrite(XLogRecPtr lsn, TimestampTz local_flush_time)
4068 : {
4069 : bool buffer_full;
4070 : int new_write_head;
4071 : int i;
4072 :
4073 186788 : if (!am_walsender)
4074 0 : return;
4075 :
4076 : /*
4077 : * If the lsn hasn't advanced since last time, then do nothing. This way
4078 : * we only record a new sample when new WAL has been written.
4079 : */
4080 186788 : if (lag_tracker->last_lsn == lsn)
4081 179384 : return;
4082 7404 : lag_tracker->last_lsn = lsn;
4083 :
4084 : /*
4085 : * If advancing the write head of the circular buffer would crash into any
4086 : * of the read heads, then the buffer is full. In other words, the
4087 : * slowest reader (presumably apply) is the one that controls the release
4088 : * of space.
4089 : */
4090 7404 : new_write_head = (lag_tracker->write_head + 1) % LAG_TRACKER_BUFFER_SIZE;
4091 7404 : buffer_full = false;
4092 29616 : for (i = 0; i < NUM_SYNC_REP_WAIT_MODE; ++i)
4093 : {
4094 22212 : if (new_write_head == lag_tracker->read_heads[i])
4095 0 : buffer_full = true;
4096 : }
4097 :
4098 : /*
4099 : * If the buffer is full, for now we just rewind by one slot and overwrite
4100 : * the last sample, as a simple (if somewhat uneven) way to lower the
4101 : * sampling rate. There may be better adaptive compaction algorithms.
4102 : */
4103 7404 : if (buffer_full)
4104 : {
4105 0 : new_write_head = lag_tracker->write_head;
4106 0 : if (lag_tracker->write_head > 0)
4107 0 : lag_tracker->write_head--;
4108 : else
4109 0 : lag_tracker->write_head = LAG_TRACKER_BUFFER_SIZE - 1;
4110 : }
4111 :
4112 : /* Store a sample at the current write head position. */
4113 7404 : lag_tracker->buffer[lag_tracker->write_head].lsn = lsn;
4114 7404 : lag_tracker->buffer[lag_tracker->write_head].time = local_flush_time;
4115 7404 : lag_tracker->write_head = new_write_head;
4116 : }
4117 :
4118 : /*
4119 : * Find out how much time has elapsed between the moment WAL location 'lsn'
4120 : * (or the highest known earlier LSN) was flushed locally and the time 'now'.
4121 : * We have a separate read head for each of the reported LSN locations we
4122 : * receive in replies from standby; 'head' controls which read head is
4123 : * used. Whenever a read head crosses an LSN which was written into the
4124 : * lag buffer with LagTrackerWrite, we can use the associated timestamp to
4125 : * find out the time this LSN (or an earlier one) was flushed locally, and
4126 : * therefore compute the lag.
4127 : *
4128 : * Return -1 if no new sample data is available, and otherwise the elapsed
4129 : * time in microseconds.
4130 : */
4131 : static TimeOffset
4132 192888 : LagTrackerRead(int head, XLogRecPtr lsn, TimestampTz now)
4133 : {
4134 192888 : TimestampTz time = 0;
4135 :
4136 : /* Read all unread samples up to this LSN or end of buffer. */
4137 213254 : while (lag_tracker->read_heads[head] != lag_tracker->write_head &&
4138 77502 : lag_tracker->buffer[lag_tracker->read_heads[head]].lsn <= lsn)
4139 : {
4140 20366 : time = lag_tracker->buffer[lag_tracker->read_heads[head]].time;
4141 20366 : lag_tracker->last_read[head] =
4142 20366 : lag_tracker->buffer[lag_tracker->read_heads[head]];
4143 20366 : lag_tracker->read_heads[head] =
4144 20366 : (lag_tracker->read_heads[head] + 1) % LAG_TRACKER_BUFFER_SIZE;
4145 : }
4146 :
4147 : /*
4148 : * If the lag tracker is empty, that means the standby has processed
4149 : * everything we've ever sent so we should now clear 'last_read'. If we
4150 : * didn't do that, we'd risk using a stale and irrelevant sample for
4151 : * interpolation at the beginning of the next burst of WAL after a period
4152 : * of idleness.
4153 : */
4154 192888 : if (lag_tracker->read_heads[head] == lag_tracker->write_head)
4155 135752 : lag_tracker->last_read[head].time = 0;
4156 :
4157 192888 : if (time > now)
4158 : {
4159 : /* If the clock somehow went backwards, treat as not found. */
4160 0 : return -1;
4161 : }
4162 192888 : else if (time == 0)
4163 : {
4164 : /*
4165 : * We didn't cross a time. If there is a future sample that we
4166 : * haven't reached yet, and we've already reached at least one sample,
4167 : * let's interpolate the local flushed time. This is mainly useful
4168 : * for reporting a completely stuck apply position as having
4169 : * increasing lag, since otherwise we'd have to wait for it to
4170 : * eventually start moving again and cross one of our samples before
4171 : * we can show the lag increasing.
4172 : */
4173 181878 : if (lag_tracker->read_heads[head] == lag_tracker->write_head)
4174 : {
4175 : /* There are no future samples, so we can't interpolate. */
4176 128108 : return -1;
4177 : }
4178 53770 : else if (lag_tracker->last_read[head].time != 0)
4179 : {
4180 : /* We can interpolate between last_read and the next sample. */
4181 : double fraction;
4182 44302 : WalTimeSample prev = lag_tracker->last_read[head];
4183 44302 : WalTimeSample next = lag_tracker->buffer[lag_tracker->read_heads[head]];
4184 :
4185 44302 : if (lsn < prev.lsn)
4186 : {
4187 : /*
4188 : * Reported LSNs shouldn't normally go backwards, but it's
4189 : * possible when there is a timeline change. Treat as not
4190 : * found.
4191 : */
4192 0 : return -1;
4193 : }
4194 :
4195 : Assert(prev.lsn < next.lsn);
4196 :
4197 44302 : if (prev.time > next.time)
4198 : {
4199 : /* If the clock somehow went backwards, treat as not found. */
4200 0 : return -1;
4201 : }
4202 :
4203 : /* See how far we are between the previous and next samples. */
4204 44302 : fraction =
4205 44302 : (double) (lsn - prev.lsn) / (double) (next.lsn - prev.lsn);
4206 :
4207 : /* Scale the local flush time proportionally. */
4208 44302 : time = (TimestampTz)
4209 44302 : ((double) prev.time + (next.time - prev.time) * fraction);
4210 : }
4211 : else
4212 : {
4213 : /*
4214 : * We have only a future sample, implying that we were entirely
4215 : * caught up but and now there is a new burst of WAL and the
4216 : * standby hasn't processed the first sample yet. Until the
4217 : * standby reaches the future sample the best we can do is report
4218 : * the hypothetical lag if that sample were to be replayed now.
4219 : */
4220 9468 : time = lag_tracker->buffer[lag_tracker->read_heads[head]].time;
4221 : }
4222 : }
4223 :
4224 : /* Return the elapsed time since local flush time in microseconds. */
4225 : Assert(time != 0);
4226 64780 : return now - time;
4227 : }
|
