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 : * This waiting time can be limited by the wal_sender_shutdown_timeout
39 : * parameter.
40 : *
41 : *
42 : * Portions Copyright (c) 2010-2026, PostgreSQL Global Development Group
43 : *
44 : * IDENTIFICATION
45 : * src/backend/replication/walsender.c
46 : *
47 : *-------------------------------------------------------------------------
48 : */
49 : #include "postgres.h"
50 :
51 : #include <signal.h>
52 : #include <unistd.h>
53 :
54 : #include "access/timeline.h"
55 : #include "access/transam.h"
56 : #include "access/twophase.h"
57 : #include "access/xact.h"
58 : #include "access/xlog_internal.h"
59 : #include "access/xlogreader.h"
60 : #include "access/xlogrecovery.h"
61 : #include "access/xlogutils.h"
62 : #include "backup/basebackup.h"
63 : #include "backup/basebackup_incremental.h"
64 : #include "catalog/pg_authid.h"
65 : #include "catalog/pg_type.h"
66 : #include "commands/defrem.h"
67 : #include "funcapi.h"
68 : #include "libpq/libpq.h"
69 : #include "libpq/pqformat.h"
70 : #include "libpq/protocol.h"
71 : #include "miscadmin.h"
72 : #include "nodes/replnodes.h"
73 : #include "pgstat.h"
74 : #include "postmaster/interrupt.h"
75 : #include "replication/decode.h"
76 : #include "replication/logical.h"
77 : #include "replication/slotsync.h"
78 : #include "replication/slot.h"
79 : #include "replication/snapbuild.h"
80 : #include "replication/syncrep.h"
81 : #include "replication/walreceiver.h"
82 : #include "replication/walsender.h"
83 : #include "replication/walsender_private.h"
84 : #include "storage/condition_variable.h"
85 : #include "storage/aio_subsys.h"
86 : #include "storage/fd.h"
87 : #include "storage/ipc.h"
88 : #include "storage/pmsignal.h"
89 : #include "storage/proc.h"
90 : #include "storage/procarray.h"
91 : #include "storage/subsystems.h"
92 : #include "tcop/dest.h"
93 : #include "tcop/tcopprot.h"
94 : #include "utils/acl.h"
95 : #include "utils/builtins.h"
96 : #include "utils/guc.h"
97 : #include "utils/lsyscache.h"
98 : #include "utils/memutils.h"
99 : #include "utils/pg_lsn.h"
100 : #include "utils/pgstat_internal.h"
101 : #include "utils/ps_status.h"
102 : #include "utils/timeout.h"
103 : #include "utils/timestamp.h"
104 : #include "utils/wait_event.h"
105 :
106 : /* Minimum interval used by walsender for stats flushes, in ms */
107 : #define WALSENDER_STATS_FLUSH_INTERVAL 1000
108 :
109 : /*
110 : * Maximum data payload in a WAL data message. Must be >= XLOG_BLCKSZ.
111 : *
112 : * We don't have a good idea of what a good value would be; there's some
113 : * overhead per message in both walsender and walreceiver, but on the other
114 : * hand sending large batches makes walsender less responsive to signals
115 : * because signals are checked only between messages. 128kB (with
116 : * default 8k blocks) seems like a reasonable guess for now.
117 : */
118 : #define MAX_SEND_SIZE (XLOG_BLCKSZ * 16)
119 :
120 : /* Array of WalSnds in shared memory */
121 : WalSndCtlData *WalSndCtl = NULL;
122 :
123 : static void WalSndShmemRequest(void *arg);
124 : static void WalSndShmemInit(void *arg);
125 :
126 : const ShmemCallbacks WalSndShmemCallbacks = {
127 : .request_fn = WalSndShmemRequest,
128 : .init_fn = WalSndShmemInit,
129 : };
130 :
131 : /* My slot in the shared memory array */
132 : WalSnd *MyWalSnd = NULL;
133 :
134 : /* Global state */
135 : bool am_walsender = false; /* Am I a walsender process? */
136 : bool am_cascading_walsender = false; /* Am I cascading WAL to another
137 : * standby? */
138 : bool am_db_walsender = false; /* Connected to a database? */
139 :
140 : /* GUC variables */
141 : int max_wal_senders = 10; /* the maximum number of concurrent
142 : * walsenders */
143 : int wal_sender_timeout = 60 * 1000; /* maximum time to send one WAL
144 : * data message */
145 :
146 : int wal_sender_shutdown_timeout = -1; /* maximum time to wait during
147 : * shutdown for WAL
148 : * replication */
149 :
150 : bool log_replication_commands = false;
151 :
152 : /*
153 : * State for WalSndWakeupRequest
154 : */
155 : bool wake_wal_senders = false;
156 :
157 : /*
158 : * xlogreader used for replication. Note that a WAL sender doing physical
159 : * replication does not need xlogreader to read WAL, but it needs one to
160 : * keep a state of its work.
161 : */
162 : static XLogReaderState *xlogreader = NULL;
163 :
164 : /*
165 : * If the UPLOAD_MANIFEST command is used to provide a backup manifest in
166 : * preparation for an incremental backup, uploaded_manifest will be point
167 : * to an object containing information about its contexts, and
168 : * uploaded_manifest_mcxt will point to the memory context that contains
169 : * that object and all of its subordinate data. Otherwise, both values will
170 : * be NULL.
171 : */
172 : static IncrementalBackupInfo *uploaded_manifest = NULL;
173 : static MemoryContext uploaded_manifest_mcxt = NULL;
174 :
175 : /*
176 : * These variables keep track of the state of the timeline we're currently
177 : * sending. sendTimeLine identifies the timeline. If sendTimeLineIsHistoric,
178 : * the timeline is not the latest timeline on this server, and the server's
179 : * history forked off from that timeline at sendTimeLineValidUpto.
180 : */
181 : static TimeLineID sendTimeLine = 0;
182 : static TimeLineID sendTimeLineNextTLI = 0;
183 : static bool sendTimeLineIsHistoric = false;
184 : static XLogRecPtr sendTimeLineValidUpto = InvalidXLogRecPtr;
185 :
186 : /*
187 : * How far have we sent WAL already? This is also advertised in
188 : * MyWalSnd->sentPtr. (Actually, this is the next WAL location to send.)
189 : */
190 : static XLogRecPtr sentPtr = InvalidXLogRecPtr;
191 :
192 : /* Buffers for constructing outgoing messages and processing reply messages. */
193 : static StringInfoData output_message;
194 : static StringInfoData reply_message;
195 : static StringInfoData tmpbuf;
196 :
197 : /* Timestamp of last ProcessRepliesIfAny(). */
198 : static TimestampTz last_processing = 0;
199 :
200 : /*
201 : * Timestamp of last ProcessRepliesIfAny() that saw a reply from the
202 : * standby. Set to 0 if wal_sender_timeout doesn't need to be active.
203 : */
204 : static TimestampTz last_reply_timestamp = 0;
205 :
206 : /* Have we sent a heartbeat message asking for reply, since last reply? */
207 : static bool waiting_for_ping_response = false;
208 :
209 : /* Timestamp when walsender received the shutdown request */
210 : static TimestampTz shutdown_request_timestamp = 0;
211 :
212 : /*
213 : * Set after queueing the CommandComplete message that ends WAL streaming
214 : * during shutdown. This prevents WalSndDone() and WalSndDoneImmediate()
215 : * from queueing the same message twice.
216 : */
217 : static bool shutdown_stream_done_queued = false;
218 :
219 : /*
220 : * While streaming WAL in Copy mode, streamingDoneSending is set to true
221 : * after we have sent CopyDone. We should not send any more CopyData messages
222 : * after that. streamingDoneReceiving is set to true when we receive CopyDone
223 : * from the other end. When both become true, it's time to exit Copy mode.
224 : */
225 : static bool streamingDoneSending;
226 : static bool streamingDoneReceiving;
227 :
228 : /* Are we there yet? */
229 : static bool WalSndCaughtUp = false;
230 :
231 : /* Flags set by signal handlers for later service in main loop */
232 : static volatile sig_atomic_t got_SIGUSR2 = false;
233 : static volatile sig_atomic_t got_STOPPING = false;
234 :
235 : /*
236 : * This is set while we are streaming. When not set
237 : * PROCSIG_WALSND_INIT_STOPPING signal will be handled like SIGTERM. When set,
238 : * the main loop is responsible for checking got_STOPPING and terminating when
239 : * it's set (after streaming any remaining WAL).
240 : */
241 : static volatile sig_atomic_t replication_active = false;
242 :
243 : static LogicalDecodingContext *logical_decoding_ctx = NULL;
244 :
245 : /* A sample associating a WAL location with the time it was written. */
246 : typedef struct
247 : {
248 : XLogRecPtr lsn;
249 : TimestampTz time;
250 : } WalTimeSample;
251 :
252 : /* The size of our buffer of time samples. */
253 : #define LAG_TRACKER_BUFFER_SIZE 8192
254 :
255 : /* A mechanism for tracking replication lag. */
256 : typedef struct
257 : {
258 : XLogRecPtr last_lsn;
259 : WalTimeSample buffer[LAG_TRACKER_BUFFER_SIZE];
260 : int write_head;
261 : int read_heads[NUM_SYNC_REP_WAIT_MODE];
262 : WalTimeSample last_read[NUM_SYNC_REP_WAIT_MODE];
263 :
264 : /*
265 : * Overflow entries for read heads that collide with the write head.
266 : *
267 : * When the cyclic buffer fills (write head is about to collide with a
268 : * read head), we save that read head's current sample here and mark it as
269 : * using overflow (read_heads[i] = -1). This allows the write head to
270 : * continue advancing while the overflowed mode continues lag computation
271 : * using the saved sample.
272 : *
273 : * Once the standby's reported LSN advances past the overflow entry's LSN,
274 : * we transition back to normal buffer-based tracking.
275 : */
276 : WalTimeSample overflowed[NUM_SYNC_REP_WAIT_MODE];
277 : } LagTracker;
278 :
279 : static LagTracker *lag_tracker;
280 :
281 : /* Signal handlers */
282 : static void WalSndLastCycleHandler(SIGNAL_ARGS);
283 :
284 : /* Prototypes for private functions */
285 : typedef void (*WalSndSendDataCallback) (void);
286 : static void WalSndLoop(WalSndSendDataCallback send_data);
287 : static void InitWalSenderSlot(void);
288 : static void WalSndKill(int code, Datum arg);
289 : pg_noreturn static void WalSndShutdown(void);
290 : static void XLogSendPhysical(void);
291 : static void XLogSendLogical(void);
292 : pg_noreturn static void WalSndDoneImmediate(void);
293 : static void WalSndDone(WalSndSendDataCallback send_data);
294 : static void IdentifySystem(void);
295 : static void UploadManifest(void);
296 : static bool HandleUploadManifestPacket(StringInfo buf, off_t *offset,
297 : IncrementalBackupInfo *ib);
298 : static void ReadReplicationSlot(ReadReplicationSlotCmd *cmd);
299 : static void CreateReplicationSlot(CreateReplicationSlotCmd *cmd);
300 : static void DropReplicationSlot(DropReplicationSlotCmd *cmd);
301 : static void StartReplication(StartReplicationCmd *cmd);
302 : static void StartLogicalReplication(StartReplicationCmd *cmd);
303 : static void ProcessStandbyMessage(void);
304 : static void ProcessStandbyReplyMessage(void);
305 : static void ProcessStandbyHSFeedbackMessage(void);
306 : static void ProcessStandbyPSRequestMessage(void);
307 : static void ProcessRepliesIfAny(void);
308 : static void ProcessPendingWrites(void);
309 : static void WalSndKeepalive(bool requestReply, XLogRecPtr writePtr);
310 : static void WalSndKeepaliveIfNecessary(void);
311 : static void WalSndCheckTimeOut(void);
312 : static void WalSndCheckShutdownTimeout(void);
313 : static long WalSndComputeSleeptime(TimestampTz now);
314 : static void WalSndWait(uint32 socket_events, long timeout, uint32 wait_event);
315 : static void WalSndPrepareWrite(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid, bool last_write);
316 : static void WalSndWriteData(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid, bool last_write);
317 : static void WalSndUpdateProgress(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid,
318 : bool skipped_xact);
319 : static XLogRecPtr WalSndWaitForWal(XLogRecPtr loc);
320 : static void LagTrackerWrite(XLogRecPtr lsn, TimestampTz local_flush_time);
321 : static TimeOffset LagTrackerRead(int head, XLogRecPtr lsn, TimestampTz now);
322 : static bool TransactionIdInRecentPast(TransactionId xid, uint32 epoch);
323 :
324 : static void WalSndSegmentOpen(XLogReaderState *state, XLogSegNo nextSegNo,
325 : TimeLineID *tli_p);
326 :
327 :
328 : /* Initialize walsender process before entering the main command loop */
329 : void
330 1300 : InitWalSender(void)
331 : {
332 1300 : am_cascading_walsender = RecoveryInProgress();
333 :
334 : /* Create a per-walsender data structure in shared memory */
335 1300 : InitWalSenderSlot();
336 :
337 : /* need resource owner for e.g. basebackups */
338 1300 : CreateAuxProcessResourceOwner();
339 :
340 : /*
341 : * Let postmaster know that we're a WAL sender. Once we've declared us as
342 : * a WAL sender process, postmaster will let us outlive the bgwriter and
343 : * kill us last in the shutdown sequence, so we get a chance to stream all
344 : * remaining WAL at shutdown, including the shutdown checkpoint. Note that
345 : * there's no going back, and we mustn't write any WAL records after this.
346 : */
347 1300 : MarkPostmasterChildWalSender();
348 1300 : SendPostmasterSignal(PMSIGNAL_ADVANCE_STATE_MACHINE);
349 :
350 : /*
351 : * If the client didn't specify a database to connect to, show in PGPROC
352 : * that our advertised xmin should affect vacuum horizons in all
353 : * databases. This allows physical replication clients to send hot
354 : * standby feedback that will delay vacuum cleanup in all databases.
355 : */
356 1300 : if (MyDatabaseId == InvalidOid)
357 : {
358 : Assert(MyProc->xmin == InvalidTransactionId);
359 503 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
360 503 : MyProc->statusFlags |= PROC_AFFECTS_ALL_HORIZONS;
361 503 : ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
362 503 : LWLockRelease(ProcArrayLock);
363 : }
364 :
365 : /* Initialize empty timestamp buffer for lag tracking. */
366 1300 : lag_tracker = MemoryContextAllocZero(TopMemoryContext, sizeof(LagTracker));
367 1300 : }
368 :
369 : /*
370 : * Clean up after an error.
371 : *
372 : * WAL sender processes don't use transactions like regular backends do.
373 : * This function does any cleanup required after an error in a WAL sender
374 : * process, similar to what transaction abort does in a regular backend.
375 : */
376 : void
377 51 : WalSndErrorCleanup(void)
378 : {
379 51 : LWLockReleaseAll();
380 51 : ConditionVariableCancelSleep();
381 51 : pgstat_report_wait_end();
382 51 : pgaio_error_cleanup();
383 :
384 51 : if (xlogreader != NULL && xlogreader->seg.ws_file >= 0)
385 7 : wal_segment_close(xlogreader);
386 :
387 51 : if (MyReplicationSlot != NULL)
388 16 : ReplicationSlotRelease();
389 :
390 51 : ReplicationSlotCleanup(false);
391 :
392 51 : replication_active = false;
393 :
394 : /*
395 : * If there is a transaction in progress, it will clean up our
396 : * ResourceOwner, but if a replication command set up a resource owner
397 : * without a transaction, we've got to clean that up now.
398 : */
399 51 : if (!IsTransactionOrTransactionBlock())
400 50 : ReleaseAuxProcessResources(false);
401 :
402 51 : if (got_STOPPING || got_SIGUSR2)
403 1 : proc_exit(0);
404 :
405 : /* Revert back to startup state */
406 50 : WalSndSetState(WALSNDSTATE_STARTUP);
407 50 : }
408 :
409 : /*
410 : * Handle a client's connection abort in an orderly manner.
411 : */
412 : static void
413 15 : WalSndShutdown(void)
414 : {
415 : /*
416 : * Reset whereToSendOutput to prevent ereport from attempting to send any
417 : * more messages to the standby.
418 : */
419 15 : if (whereToSendOutput == DestRemote)
420 15 : whereToSendOutput = DestNone;
421 :
422 15 : proc_exit(0);
423 : }
424 :
425 : /*
426 : * Handle the IDENTIFY_SYSTEM command.
427 : */
428 : static void
429 817 : IdentifySystem(void)
430 : {
431 : char sysid[32];
432 : char xloc[MAXFNAMELEN];
433 : XLogRecPtr logptr;
434 817 : char *dbname = NULL;
435 : DestReceiver *dest;
436 : TupOutputState *tstate;
437 : TupleDesc tupdesc;
438 : Datum values[4];
439 817 : bool nulls[4] = {0};
440 : TimeLineID currTLI;
441 :
442 : /*
443 : * Reply with a result set with one row, four columns. First col is system
444 : * ID, second is timeline ID, third is current xlog location and the
445 : * fourth contains the database name if we are connected to one.
446 : */
447 :
448 817 : snprintf(sysid, sizeof(sysid), UINT64_FORMAT,
449 : GetSystemIdentifier());
450 :
451 817 : am_cascading_walsender = RecoveryInProgress();
452 817 : if (am_cascading_walsender)
453 66 : logptr = GetStandbyFlushRecPtr(&currTLI);
454 : else
455 751 : logptr = GetFlushRecPtr(&currTLI);
456 :
457 817 : snprintf(xloc, sizeof(xloc), "%X/%08X", LSN_FORMAT_ARGS(logptr));
458 :
459 817 : if (MyDatabaseId != InvalidOid)
460 : {
461 301 : MemoryContext cur = CurrentMemoryContext;
462 :
463 : /* syscache access needs a transaction env. */
464 301 : StartTransactionCommand();
465 301 : dbname = get_database_name(MyDatabaseId);
466 : /* copy dbname out of TX context */
467 301 : dbname = MemoryContextStrdup(cur, dbname);
468 301 : CommitTransactionCommand();
469 : }
470 :
471 817 : dest = CreateDestReceiver(DestRemoteSimple);
472 :
473 : /* need a tuple descriptor representing four columns */
474 817 : tupdesc = CreateTemplateTupleDesc(4);
475 817 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "systemid",
476 : TEXTOID, -1, 0);
477 817 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "timeline",
478 : INT8OID, -1, 0);
479 817 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 3, "xlogpos",
480 : TEXTOID, -1, 0);
481 817 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 4, "dbname",
482 : TEXTOID, -1, 0);
483 817 : TupleDescFinalize(tupdesc);
484 :
485 : /* prepare for projection of tuples */
486 817 : tstate = begin_tup_output_tupdesc(dest, tupdesc, &TTSOpsVirtual);
487 :
488 : /* column 1: system identifier */
489 817 : values[0] = CStringGetTextDatum(sysid);
490 :
491 : /* column 2: timeline */
492 817 : values[1] = Int64GetDatum(currTLI);
493 :
494 : /* column 3: wal location */
495 817 : values[2] = CStringGetTextDatum(xloc);
496 :
497 : /* column 4: database name, or NULL if none */
498 817 : if (dbname)
499 301 : values[3] = CStringGetTextDatum(dbname);
500 : else
501 516 : nulls[3] = true;
502 :
503 : /* send it to dest */
504 817 : do_tup_output(tstate, values, nulls);
505 :
506 817 : end_tup_output(tstate);
507 817 : }
508 :
509 : /* Handle READ_REPLICATION_SLOT command */
510 : static void
511 6 : ReadReplicationSlot(ReadReplicationSlotCmd *cmd)
512 : {
513 : #define READ_REPLICATION_SLOT_COLS 3
514 : ReplicationSlot *slot;
515 : DestReceiver *dest;
516 : TupOutputState *tstate;
517 : TupleDesc tupdesc;
518 6 : Datum values[READ_REPLICATION_SLOT_COLS] = {0};
519 : bool nulls[READ_REPLICATION_SLOT_COLS];
520 :
521 6 : tupdesc = CreateTemplateTupleDesc(READ_REPLICATION_SLOT_COLS);
522 6 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "slot_type",
523 : TEXTOID, -1, 0);
524 6 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "restart_lsn",
525 : TEXTOID, -1, 0);
526 : /* TimeLineID is unsigned, so int4 is not wide enough. */
527 6 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 3, "restart_tli",
528 : INT8OID, -1, 0);
529 6 : TupleDescFinalize(tupdesc);
530 :
531 6 : memset(nulls, true, READ_REPLICATION_SLOT_COLS * sizeof(bool));
532 :
533 6 : LWLockAcquire(ReplicationSlotControlLock, LW_SHARED);
534 6 : slot = SearchNamedReplicationSlot(cmd->slotname, false);
535 6 : if (slot == NULL || !slot->in_use)
536 : {
537 2 : LWLockRelease(ReplicationSlotControlLock);
538 : }
539 : else
540 : {
541 : ReplicationSlot slot_contents;
542 4 : int i = 0;
543 :
544 : /* Copy slot contents while holding spinlock */
545 4 : SpinLockAcquire(&slot->mutex);
546 4 : slot_contents = *slot;
547 4 : SpinLockRelease(&slot->mutex);
548 4 : LWLockRelease(ReplicationSlotControlLock);
549 :
550 4 : if (OidIsValid(slot_contents.data.database))
551 1 : ereport(ERROR,
552 : errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
553 : errmsg("cannot use %s with a logical replication slot",
554 : "READ_REPLICATION_SLOT"));
555 :
556 : /* slot type */
557 3 : values[i] = CStringGetTextDatum("physical");
558 3 : nulls[i] = false;
559 3 : i++;
560 :
561 : /* start LSN */
562 3 : if (XLogRecPtrIsValid(slot_contents.data.restart_lsn))
563 : {
564 : char xloc[64];
565 :
566 3 : snprintf(xloc, sizeof(xloc), "%X/%08X",
567 3 : LSN_FORMAT_ARGS(slot_contents.data.restart_lsn));
568 3 : values[i] = CStringGetTextDatum(xloc);
569 3 : nulls[i] = false;
570 : }
571 3 : i++;
572 :
573 : /* timeline this WAL was produced on */
574 3 : if (XLogRecPtrIsValid(slot_contents.data.restart_lsn))
575 : {
576 : TimeLineID slots_position_timeline;
577 : TimeLineID current_timeline;
578 3 : List *timeline_history = NIL;
579 :
580 : /*
581 : * While in recovery, use as timeline the currently-replaying one
582 : * to get the LSN position's history.
583 : */
584 3 : if (RecoveryInProgress())
585 0 : (void) GetXLogReplayRecPtr(¤t_timeline);
586 : else
587 3 : current_timeline = GetWALInsertionTimeLine();
588 :
589 3 : timeline_history = readTimeLineHistory(current_timeline);
590 3 : slots_position_timeline = tliOfPointInHistory(slot_contents.data.restart_lsn,
591 : timeline_history);
592 3 : values[i] = Int64GetDatum((int64) slots_position_timeline);
593 3 : nulls[i] = false;
594 : }
595 3 : i++;
596 :
597 : Assert(i == READ_REPLICATION_SLOT_COLS);
598 : }
599 :
600 5 : dest = CreateDestReceiver(DestRemoteSimple);
601 5 : tstate = begin_tup_output_tupdesc(dest, tupdesc, &TTSOpsVirtual);
602 5 : do_tup_output(tstate, values, nulls);
603 5 : end_tup_output(tstate);
604 5 : }
605 :
606 :
607 : /*
608 : * Handle TIMELINE_HISTORY command.
609 : */
610 : static void
611 15 : SendTimeLineHistory(TimeLineHistoryCmd *cmd)
612 : {
613 : DestReceiver *dest;
614 : TupleDesc tupdesc;
615 : StringInfoData buf;
616 : char histfname[MAXFNAMELEN];
617 : char path[MAXPGPATH];
618 : int fd;
619 : off_t histfilelen;
620 : off_t bytesleft;
621 : Size len;
622 :
623 15 : dest = CreateDestReceiver(DestRemoteSimple);
624 :
625 : /*
626 : * Reply with a result set with one row, and two columns. The first col is
627 : * the name of the history file, 2nd is the contents.
628 : */
629 15 : tupdesc = CreateTemplateTupleDesc(2);
630 15 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "filename", TEXTOID, -1, 0);
631 15 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "content", TEXTOID, -1, 0);
632 15 : TupleDescFinalize(tupdesc);
633 :
634 15 : TLHistoryFileName(histfname, cmd->timeline);
635 15 : TLHistoryFilePath(path, cmd->timeline);
636 :
637 : /* Send a RowDescription message */
638 15 : dest->rStartup(dest, CMD_SELECT, tupdesc);
639 :
640 : /* Send a DataRow message */
641 15 : pq_beginmessage(&buf, PqMsg_DataRow);
642 15 : pq_sendint16(&buf, 2); /* # of columns */
643 15 : len = strlen(histfname);
644 15 : pq_sendint32(&buf, len); /* col1 len */
645 15 : pq_sendbytes(&buf, histfname, len);
646 :
647 15 : fd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
648 15 : if (fd < 0)
649 0 : ereport(ERROR,
650 : (errcode_for_file_access(),
651 : errmsg("could not open file \"%s\": %m", path)));
652 :
653 : /* Determine file length and send it to client */
654 15 : histfilelen = lseek(fd, 0, SEEK_END);
655 15 : if (histfilelen < 0)
656 0 : ereport(ERROR,
657 : (errcode_for_file_access(),
658 : errmsg("could not seek to end of file \"%s\": %m", path)));
659 15 : if (lseek(fd, 0, SEEK_SET) != 0)
660 0 : ereport(ERROR,
661 : (errcode_for_file_access(),
662 : errmsg("could not seek to beginning of file \"%s\": %m", path)));
663 :
664 15 : pq_sendint32(&buf, histfilelen); /* col2 len */
665 :
666 15 : bytesleft = histfilelen;
667 30 : while (bytesleft > 0)
668 : {
669 : PGAlignedBlock rbuf;
670 : int nread;
671 :
672 15 : pgstat_report_wait_start(WAIT_EVENT_WALSENDER_TIMELINE_HISTORY_READ);
673 15 : nread = read(fd, rbuf.data, sizeof(rbuf));
674 15 : pgstat_report_wait_end();
675 15 : if (nread < 0)
676 0 : ereport(ERROR,
677 : (errcode_for_file_access(),
678 : errmsg("could not read file \"%s\": %m",
679 : path)));
680 15 : else if (nread == 0)
681 0 : ereport(ERROR,
682 : (errcode(ERRCODE_DATA_CORRUPTED),
683 : errmsg("could not read file \"%s\": read %d of %zu",
684 : path, nread, (Size) bytesleft)));
685 :
686 15 : pq_sendbytes(&buf, rbuf.data, nread);
687 15 : bytesleft -= nread;
688 : }
689 :
690 15 : if (CloseTransientFile(fd) != 0)
691 0 : ereport(ERROR,
692 : (errcode_for_file_access(),
693 : errmsg("could not close file \"%s\": %m", path)));
694 :
695 15 : pq_endmessage(&buf);
696 15 : }
697 :
698 : /*
699 : * Handle UPLOAD_MANIFEST command.
700 : */
701 : static void
702 13 : UploadManifest(void)
703 : {
704 : MemoryContext mcxt;
705 : IncrementalBackupInfo *ib;
706 13 : off_t offset = 0;
707 : StringInfoData buf;
708 :
709 : /*
710 : * parsing the manifest will use the cryptohash stuff, which requires a
711 : * resource owner
712 : */
713 : Assert(AuxProcessResourceOwner != NULL);
714 : Assert(CurrentResourceOwner == AuxProcessResourceOwner ||
715 : CurrentResourceOwner == NULL);
716 13 : CurrentResourceOwner = AuxProcessResourceOwner;
717 :
718 : /* Prepare to read manifest data into a temporary context. */
719 13 : mcxt = AllocSetContextCreate(CurrentMemoryContext,
720 : "incremental backup information",
721 : ALLOCSET_DEFAULT_SIZES);
722 13 : ib = CreateIncrementalBackupInfo(mcxt);
723 :
724 : /* Send a CopyInResponse message */
725 13 : pq_beginmessage(&buf, PqMsg_CopyInResponse);
726 13 : pq_sendbyte(&buf, 0);
727 13 : pq_sendint16(&buf, 0);
728 13 : pq_endmessage_reuse(&buf);
729 13 : pq_flush();
730 :
731 : /* Receive packets from client until done. */
732 52 : while (HandleUploadManifestPacket(&buf, &offset, ib))
733 : ;
734 :
735 : /* Finish up manifest processing. */
736 12 : FinalizeIncrementalManifest(ib);
737 :
738 : /*
739 : * Discard any old manifest information and arrange to preserve the new
740 : * information we just got.
741 : *
742 : * We assume that MemoryContextDelete and MemoryContextSetParent won't
743 : * fail, and thus we shouldn't end up bailing out of here in such a way as
744 : * to leave dangling pointers.
745 : */
746 12 : if (uploaded_manifest_mcxt != NULL)
747 0 : MemoryContextDelete(uploaded_manifest_mcxt);
748 12 : MemoryContextSetParent(mcxt, CacheMemoryContext);
749 12 : uploaded_manifest = ib;
750 12 : uploaded_manifest_mcxt = mcxt;
751 :
752 : /* clean up the resource owner we created */
753 12 : ReleaseAuxProcessResources(true);
754 12 : }
755 :
756 : /*
757 : * Process one packet received during the handling of an UPLOAD_MANIFEST
758 : * operation.
759 : *
760 : * 'buf' is scratch space. This function expects it to be initialized, doesn't
761 : * care what the current contents are, and may override them with completely
762 : * new contents.
763 : *
764 : * The return value is true if the caller should continue processing
765 : * additional packets and false if the UPLOAD_MANIFEST operation is complete.
766 : */
767 : static bool
768 52 : HandleUploadManifestPacket(StringInfo buf, off_t *offset,
769 : IncrementalBackupInfo *ib)
770 : {
771 : int mtype;
772 : int maxmsglen;
773 :
774 52 : HOLD_CANCEL_INTERRUPTS();
775 :
776 52 : pq_startmsgread();
777 52 : mtype = pq_getbyte();
778 52 : if (mtype == EOF)
779 0 : ereport(ERROR,
780 : (errcode(ERRCODE_CONNECTION_FAILURE),
781 : errmsg("unexpected EOF on client connection with an open transaction")));
782 :
783 52 : switch (mtype)
784 : {
785 40 : case PqMsg_CopyData:
786 40 : maxmsglen = PQ_LARGE_MESSAGE_LIMIT;
787 40 : break;
788 12 : case PqMsg_CopyDone:
789 : case PqMsg_CopyFail:
790 : case PqMsg_Flush:
791 : case PqMsg_Sync:
792 12 : maxmsglen = PQ_SMALL_MESSAGE_LIMIT;
793 12 : break;
794 0 : default:
795 0 : ereport(ERROR,
796 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
797 : errmsg("unexpected message type 0x%02X during COPY from stdin",
798 : mtype)));
799 : maxmsglen = 0; /* keep compiler quiet */
800 : break;
801 : }
802 :
803 : /* Now collect the message body */
804 52 : if (pq_getmessage(buf, maxmsglen))
805 0 : ereport(ERROR,
806 : (errcode(ERRCODE_CONNECTION_FAILURE),
807 : errmsg("unexpected EOF on client connection with an open transaction")));
808 52 : RESUME_CANCEL_INTERRUPTS();
809 :
810 : /* Process the message */
811 52 : switch (mtype)
812 : {
813 40 : case PqMsg_CopyData:
814 40 : AppendIncrementalManifestData(ib, buf->data, buf->len);
815 39 : return true;
816 :
817 12 : case PqMsg_CopyDone:
818 12 : return false;
819 :
820 0 : case PqMsg_Sync:
821 : case PqMsg_Flush:
822 : /* Ignore these while in CopyOut mode as we do elsewhere. */
823 0 : return true;
824 :
825 0 : case PqMsg_CopyFail:
826 0 : ereport(ERROR,
827 : (errcode(ERRCODE_QUERY_CANCELED),
828 : errmsg("COPY from stdin failed: %s",
829 : pq_getmsgstring(buf))));
830 : }
831 :
832 : /* Not reached. */
833 : Assert(false);
834 0 : return false;
835 : }
836 :
837 : /*
838 : * Handle START_REPLICATION command.
839 : *
840 : * At the moment, this never returns, but an ereport(ERROR) will take us back
841 : * to the main loop.
842 : */
843 : static void
844 303 : StartReplication(StartReplicationCmd *cmd)
845 : {
846 : StringInfoData buf;
847 : XLogRecPtr FlushPtr;
848 : TimeLineID FlushTLI;
849 :
850 : /* create xlogreader for physical replication */
851 303 : xlogreader =
852 303 : XLogReaderAllocate(wal_segment_size, NULL,
853 303 : XL_ROUTINE(.segment_open = WalSndSegmentOpen,
854 : .segment_close = wal_segment_close),
855 : NULL);
856 :
857 303 : if (!xlogreader)
858 0 : ereport(ERROR,
859 : (errcode(ERRCODE_OUT_OF_MEMORY),
860 : errmsg("out of memory"),
861 : errdetail("Failed while allocating a WAL reading processor.")));
862 :
863 : /*
864 : * We assume here that we're logging enough information in the WAL for
865 : * log-shipping, since this is checked in PostmasterMain().
866 : *
867 : * NOTE: wal_level can only change at shutdown, so in most cases it is
868 : * difficult for there to be WAL data that we can still see that was
869 : * written at wal_level='minimal'.
870 : */
871 :
872 303 : if (cmd->slotname)
873 : {
874 200 : ReplicationSlotAcquire(cmd->slotname, true, true);
875 198 : if (SlotIsLogical(MyReplicationSlot))
876 0 : ereport(ERROR,
877 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
878 : errmsg("cannot use a logical replication slot for physical replication")));
879 :
880 : /*
881 : * We don't need to verify the slot's restart_lsn here; instead we
882 : * rely on the caller requesting the starting point to use. If the
883 : * WAL segment doesn't exist, we'll fail later.
884 : */
885 : }
886 :
887 : /*
888 : * Select the timeline. If it was given explicitly by the client, use
889 : * that. Otherwise use the timeline of the last replayed record.
890 : */
891 301 : am_cascading_walsender = RecoveryInProgress();
892 301 : if (am_cascading_walsender)
893 16 : FlushPtr = GetStandbyFlushRecPtr(&FlushTLI);
894 : else
895 285 : FlushPtr = GetFlushRecPtr(&FlushTLI);
896 :
897 301 : if (cmd->timeline != 0)
898 : {
899 : XLogRecPtr switchpoint;
900 :
901 300 : sendTimeLine = cmd->timeline;
902 300 : if (sendTimeLine == FlushTLI)
903 : {
904 291 : sendTimeLineIsHistoric = false;
905 291 : sendTimeLineValidUpto = InvalidXLogRecPtr;
906 : }
907 : else
908 : {
909 : List *timeLineHistory;
910 :
911 9 : sendTimeLineIsHistoric = true;
912 :
913 : /*
914 : * Check that the timeline the client requested exists, and the
915 : * requested start location is on that timeline.
916 : */
917 9 : timeLineHistory = readTimeLineHistory(FlushTLI);
918 9 : switchpoint = tliSwitchPoint(cmd->timeline, timeLineHistory,
919 : &sendTimeLineNextTLI);
920 9 : list_free_deep(timeLineHistory);
921 :
922 : /*
923 : * Found the requested timeline in the history. Check that
924 : * requested startpoint is on that timeline in our history.
925 : *
926 : * This is quite loose on purpose. We only check that we didn't
927 : * fork off the requested timeline before the switchpoint. We
928 : * don't check that we switched *to* it before the requested
929 : * starting point. This is because the client can legitimately
930 : * request to start replication from the beginning of the WAL
931 : * segment that contains switchpoint, but on the new timeline, so
932 : * that it doesn't end up with a partial segment. If you ask for
933 : * too old a starting point, you'll get an error later when we
934 : * fail to find the requested WAL segment in pg_wal.
935 : *
936 : * XXX: we could be more strict here and only allow a startpoint
937 : * that's older than the switchpoint, if it's still in the same
938 : * WAL segment.
939 : */
940 9 : if (XLogRecPtrIsValid(switchpoint) &&
941 9 : switchpoint < cmd->startpoint)
942 : {
943 0 : ereport(ERROR,
944 : errmsg("requested starting point %X/%08X on timeline %u is not in this server's history",
945 : LSN_FORMAT_ARGS(cmd->startpoint),
946 : cmd->timeline),
947 : errdetail("This server's history forked from timeline %u at %X/%08X.",
948 : cmd->timeline,
949 : LSN_FORMAT_ARGS(switchpoint)));
950 : }
951 9 : sendTimeLineValidUpto = switchpoint;
952 : }
953 : }
954 : else
955 : {
956 1 : sendTimeLine = FlushTLI;
957 1 : sendTimeLineValidUpto = InvalidXLogRecPtr;
958 1 : sendTimeLineIsHistoric = false;
959 : }
960 :
961 301 : streamingDoneSending = streamingDoneReceiving = false;
962 :
963 : /* If there is nothing to stream, don't even enter COPY mode */
964 301 : if (!sendTimeLineIsHistoric || cmd->startpoint < sendTimeLineValidUpto)
965 : {
966 : /*
967 : * When we first start replication the standby will be behind the
968 : * primary. For some applications, for example synchronous
969 : * replication, it is important to have a clear state for this initial
970 : * catchup mode, so we can trigger actions when we change streaming
971 : * state later. We may stay in this state for a long time, which is
972 : * exactly why we want to be able to monitor whether or not we are
973 : * still here.
974 : */
975 301 : WalSndSetState(WALSNDSTATE_CATCHUP);
976 :
977 : /* Send a CopyBothResponse message, and start streaming */
978 301 : pq_beginmessage(&buf, PqMsg_CopyBothResponse);
979 301 : pq_sendbyte(&buf, 0);
980 301 : pq_sendint16(&buf, 0);
981 301 : pq_endmessage(&buf);
982 301 : pq_flush();
983 :
984 : /*
985 : * Don't allow a request to stream from a future point in WAL that
986 : * hasn't been flushed to disk in this server yet.
987 : */
988 301 : if (FlushPtr < cmd->startpoint)
989 : {
990 0 : ereport(ERROR,
991 : errmsg("requested starting point %X/%08X is ahead of the WAL flush position of this server %X/%08X",
992 : LSN_FORMAT_ARGS(cmd->startpoint),
993 : LSN_FORMAT_ARGS(FlushPtr)));
994 : }
995 :
996 : /* Start streaming from the requested point */
997 301 : sentPtr = cmd->startpoint;
998 :
999 : /* Initialize shared memory status, too */
1000 301 : SpinLockAcquire(&MyWalSnd->mutex);
1001 301 : MyWalSnd->sentPtr = sentPtr;
1002 301 : SpinLockRelease(&MyWalSnd->mutex);
1003 :
1004 301 : SyncRepInitConfig();
1005 :
1006 : /* Main loop of walsender */
1007 301 : replication_active = true;
1008 :
1009 301 : WalSndLoop(XLogSendPhysical);
1010 :
1011 163 : replication_active = false;
1012 163 : if (got_STOPPING)
1013 0 : proc_exit(0);
1014 163 : WalSndSetState(WALSNDSTATE_STARTUP);
1015 :
1016 : Assert(streamingDoneSending && streamingDoneReceiving);
1017 : }
1018 :
1019 163 : if (cmd->slotname)
1020 148 : ReplicationSlotRelease();
1021 :
1022 : /*
1023 : * Copy is finished now. Send a single-row result set indicating the next
1024 : * timeline.
1025 : */
1026 163 : if (sendTimeLineIsHistoric)
1027 : {
1028 : char startpos_str[8 + 1 + 8 + 1];
1029 : DestReceiver *dest;
1030 : TupOutputState *tstate;
1031 : TupleDesc tupdesc;
1032 : Datum values[2];
1033 11 : bool nulls[2] = {0};
1034 :
1035 11 : snprintf(startpos_str, sizeof(startpos_str), "%X/%08X",
1036 11 : LSN_FORMAT_ARGS(sendTimeLineValidUpto));
1037 :
1038 11 : dest = CreateDestReceiver(DestRemoteSimple);
1039 :
1040 : /*
1041 : * Need a tuple descriptor representing two columns. int8 may seem
1042 : * like a surprising data type for this, but in theory int4 would not
1043 : * be wide enough for this, as TimeLineID is unsigned.
1044 : */
1045 11 : tupdesc = CreateTemplateTupleDesc(2);
1046 11 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "next_tli",
1047 : INT8OID, -1, 0);
1048 11 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "next_tli_startpos",
1049 : TEXTOID, -1, 0);
1050 11 : TupleDescFinalize(tupdesc);
1051 :
1052 : /* prepare for projection of tuple */
1053 11 : tstate = begin_tup_output_tupdesc(dest, tupdesc, &TTSOpsVirtual);
1054 :
1055 11 : values[0] = Int64GetDatum((int64) sendTimeLineNextTLI);
1056 11 : values[1] = CStringGetTextDatum(startpos_str);
1057 :
1058 : /* send it to dest */
1059 11 : do_tup_output(tstate, values, nulls);
1060 :
1061 11 : end_tup_output(tstate);
1062 : }
1063 :
1064 : /* Send CommandComplete message */
1065 163 : EndReplicationCommand("START_STREAMING");
1066 163 : }
1067 :
1068 : /*
1069 : * XLogReaderRoutine->page_read callback for logical decoding contexts, as a
1070 : * walsender process.
1071 : *
1072 : * Inside the walsender we can do better than read_local_xlog_page,
1073 : * which has to do a plain sleep/busy loop, because the walsender's latch gets
1074 : * set every time WAL is flushed.
1075 : */
1076 : static int
1077 39176 : logical_read_xlog_page(XLogReaderState *state, XLogRecPtr targetPagePtr, int reqLen,
1078 : XLogRecPtr targetRecPtr, char *cur_page)
1079 : {
1080 : XLogRecPtr flushptr;
1081 : int count;
1082 : WALReadError errinfo;
1083 : XLogSegNo segno;
1084 : TimeLineID currTLI;
1085 :
1086 : /*
1087 : * Make sure we have enough WAL available before retrieving the current
1088 : * timeline.
1089 : */
1090 39176 : flushptr = WalSndWaitForWal(targetPagePtr + reqLen);
1091 :
1092 : /* Fail if not enough (implies we are going to shut down) */
1093 38955 : if (flushptr < targetPagePtr + reqLen)
1094 12621 : return -1;
1095 :
1096 : /*
1097 : * Since logical decoding is also permitted on a standby server, we need
1098 : * to check if the server is in recovery to decide how to get the current
1099 : * timeline ID (so that it also covers the promotion or timeline change
1100 : * cases). We must determine am_cascading_walsender after waiting for the
1101 : * required WAL so that it is correct when the walsender wakes up after a
1102 : * promotion.
1103 : */
1104 26334 : am_cascading_walsender = RecoveryInProgress();
1105 :
1106 26334 : if (am_cascading_walsender)
1107 1964 : GetXLogReplayRecPtr(&currTLI);
1108 : else
1109 24370 : currTLI = GetWALInsertionTimeLine();
1110 :
1111 26334 : XLogReadDetermineTimeline(state, targetPagePtr, reqLen, currTLI);
1112 26334 : sendTimeLineIsHistoric = (state->currTLI != currTLI);
1113 26334 : sendTimeLine = state->currTLI;
1114 26334 : sendTimeLineValidUpto = state->currTLIValidUntil;
1115 26334 : sendTimeLineNextTLI = state->nextTLI;
1116 :
1117 26334 : if (targetPagePtr + XLOG_BLCKSZ <= flushptr)
1118 24283 : count = XLOG_BLCKSZ; /* more than one block available */
1119 : else
1120 2051 : count = flushptr - targetPagePtr; /* part of the page available */
1121 :
1122 : /* now actually read the data, we know it's there */
1123 26334 : if (!WALRead(state,
1124 : cur_page,
1125 : targetPagePtr,
1126 : count,
1127 : currTLI, /* Pass the current TLI because only
1128 : * WalSndSegmentOpen controls whether new TLI
1129 : * is needed. */
1130 : &errinfo))
1131 0 : WALReadRaiseError(&errinfo);
1132 :
1133 : /*
1134 : * After reading into the buffer, check that what we read was valid. We do
1135 : * this after reading, because even though the segment was present when we
1136 : * opened it, it might get recycled or removed while we read it. The
1137 : * read() succeeds in that case, but the data we tried to read might
1138 : * already have been overwritten with new WAL records.
1139 : */
1140 26334 : XLByteToSeg(targetPagePtr, segno, state->segcxt.ws_segsize);
1141 26334 : CheckXLogRemoved(segno, state->seg.ws_tli);
1142 :
1143 26334 : return count;
1144 : }
1145 :
1146 : /*
1147 : * Process extra options given to CREATE_REPLICATION_SLOT.
1148 : */
1149 : static void
1150 516 : parseCreateReplSlotOptions(CreateReplicationSlotCmd *cmd,
1151 : bool *reserve_wal,
1152 : CRSSnapshotAction *snapshot_action,
1153 : bool *two_phase, bool *failover)
1154 : {
1155 : ListCell *lc;
1156 516 : bool snapshot_action_given = false;
1157 516 : bool reserve_wal_given = false;
1158 516 : bool two_phase_given = false;
1159 516 : bool failover_given = false;
1160 :
1161 : /* Parse options */
1162 1045 : foreach(lc, cmd->options)
1163 : {
1164 529 : DefElem *defel = (DefElem *) lfirst(lc);
1165 :
1166 529 : if (strcmp(defel->defname, "snapshot") == 0)
1167 : {
1168 : char *action;
1169 :
1170 364 : if (snapshot_action_given || cmd->kind != REPLICATION_KIND_LOGICAL)
1171 0 : ereport(ERROR,
1172 : (errcode(ERRCODE_SYNTAX_ERROR),
1173 : errmsg("conflicting or redundant options")));
1174 :
1175 364 : action = defGetString(defel);
1176 364 : snapshot_action_given = true;
1177 :
1178 364 : if (strcmp(action, "export") == 0)
1179 1 : *snapshot_action = CRS_EXPORT_SNAPSHOT;
1180 363 : else if (strcmp(action, "nothing") == 0)
1181 149 : *snapshot_action = CRS_NOEXPORT_SNAPSHOT;
1182 214 : else if (strcmp(action, "use") == 0)
1183 214 : *snapshot_action = CRS_USE_SNAPSHOT;
1184 : else
1185 0 : ereport(ERROR,
1186 : (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1187 : errmsg("unrecognized value for %s option \"%s\": \"%s\"",
1188 : "CREATE_REPLICATION_SLOT", defel->defname, action)));
1189 : }
1190 165 : else if (strcmp(defel->defname, "reserve_wal") == 0)
1191 : {
1192 151 : if (reserve_wal_given || cmd->kind != REPLICATION_KIND_PHYSICAL)
1193 0 : ereport(ERROR,
1194 : (errcode(ERRCODE_SYNTAX_ERROR),
1195 : errmsg("conflicting or redundant options")));
1196 :
1197 151 : reserve_wal_given = true;
1198 151 : *reserve_wal = defGetBoolean(defel);
1199 : }
1200 14 : else if (strcmp(defel->defname, "two_phase") == 0)
1201 : {
1202 2 : if (two_phase_given || cmd->kind != REPLICATION_KIND_LOGICAL)
1203 0 : ereport(ERROR,
1204 : (errcode(ERRCODE_SYNTAX_ERROR),
1205 : errmsg("conflicting or redundant options")));
1206 2 : two_phase_given = true;
1207 2 : *two_phase = defGetBoolean(defel);
1208 : }
1209 12 : else if (strcmp(defel->defname, "failover") == 0)
1210 : {
1211 12 : if (failover_given || cmd->kind != REPLICATION_KIND_LOGICAL)
1212 0 : ereport(ERROR,
1213 : (errcode(ERRCODE_SYNTAX_ERROR),
1214 : errmsg("conflicting or redundant options")));
1215 12 : failover_given = true;
1216 12 : *failover = defGetBoolean(defel);
1217 : }
1218 : else
1219 0 : elog(ERROR, "unrecognized option: %s", defel->defname);
1220 : }
1221 516 : }
1222 :
1223 : /*
1224 : * Create a new replication slot.
1225 : */
1226 : static void
1227 516 : CreateReplicationSlot(CreateReplicationSlotCmd *cmd)
1228 : {
1229 516 : const char *snapshot_name = NULL;
1230 : char xloc[MAXFNAMELEN];
1231 : char *slot_name;
1232 516 : bool reserve_wal = false;
1233 516 : bool two_phase = false;
1234 516 : bool failover = false;
1235 516 : CRSSnapshotAction snapshot_action = CRS_EXPORT_SNAPSHOT;
1236 : DestReceiver *dest;
1237 : TupOutputState *tstate;
1238 : TupleDesc tupdesc;
1239 : Datum values[4];
1240 516 : bool nulls[4] = {0};
1241 :
1242 : Assert(!MyReplicationSlot);
1243 :
1244 516 : parseCreateReplSlotOptions(cmd, &reserve_wal, &snapshot_action, &two_phase,
1245 : &failover);
1246 :
1247 516 : if (cmd->kind == REPLICATION_KIND_PHYSICAL)
1248 : {
1249 152 : ReplicationSlotCreate(cmd->slotname, false,
1250 152 : cmd->temporary ? RS_TEMPORARY : RS_PERSISTENT,
1251 : false, false, false, false);
1252 :
1253 151 : if (reserve_wal)
1254 : {
1255 150 : ReplicationSlotReserveWal();
1256 :
1257 150 : ReplicationSlotMarkDirty();
1258 :
1259 : /* Write this slot to disk if it's a permanent one. */
1260 150 : if (!cmd->temporary)
1261 4 : ReplicationSlotSave();
1262 : }
1263 : }
1264 : else
1265 : {
1266 : LogicalDecodingContext *ctx;
1267 364 : bool need_full_snapshot = false;
1268 :
1269 : Assert(cmd->kind == REPLICATION_KIND_LOGICAL);
1270 :
1271 364 : CheckLogicalDecodingRequirements(false);
1272 :
1273 : /*
1274 : * Initially create persistent slot as ephemeral - that allows us to
1275 : * nicely handle errors during initialization because it'll get
1276 : * dropped if this transaction fails. We'll make it persistent at the
1277 : * end. Temporary slots can be created as temporary from beginning as
1278 : * they get dropped on error as well.
1279 : */
1280 364 : ReplicationSlotCreate(cmd->slotname, true,
1281 364 : cmd->temporary ? RS_TEMPORARY : RS_EPHEMERAL,
1282 : two_phase, false, failover, false);
1283 :
1284 : /*
1285 : * Do options check early so that we can bail before calling the
1286 : * DecodingContextFindStartpoint which can take long time.
1287 : */
1288 364 : if (snapshot_action == CRS_EXPORT_SNAPSHOT)
1289 : {
1290 1 : if (IsTransactionBlock())
1291 0 : ereport(ERROR,
1292 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1293 : (errmsg("%s must not be called inside a transaction",
1294 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'export')")));
1295 :
1296 1 : need_full_snapshot = true;
1297 : }
1298 363 : else if (snapshot_action == CRS_USE_SNAPSHOT)
1299 : {
1300 214 : if (!IsTransactionBlock())
1301 0 : ereport(ERROR,
1302 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1303 : (errmsg("%s must be called inside a transaction",
1304 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1305 :
1306 214 : if (XactIsoLevel != XACT_REPEATABLE_READ)
1307 0 : ereport(ERROR,
1308 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1309 : (errmsg("%s must be called in REPEATABLE READ isolation mode transaction",
1310 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1311 214 : if (!XactReadOnly)
1312 0 : ereport(ERROR,
1313 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1314 : (errmsg("%s must be called in a read-only transaction",
1315 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1316 :
1317 214 : if (FirstSnapshotSet)
1318 0 : ereport(ERROR,
1319 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1320 : (errmsg("%s must be called before any query",
1321 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1322 :
1323 214 : if (IsSubTransaction())
1324 0 : ereport(ERROR,
1325 : /*- translator: %s is a CREATE_REPLICATION_SLOT statement */
1326 : (errmsg("%s must not be called in a subtransaction",
1327 : "CREATE_REPLICATION_SLOT ... (SNAPSHOT 'use')")));
1328 :
1329 214 : need_full_snapshot = true;
1330 : }
1331 :
1332 : /*
1333 : * Ensure the logical decoding is enabled before initializing the
1334 : * logical decoding context.
1335 : */
1336 364 : EnsureLogicalDecodingEnabled();
1337 : Assert(IsLogicalDecodingEnabled());
1338 :
1339 364 : ctx = CreateInitDecodingContext(cmd->plugin, NIL, need_full_snapshot,
1340 : false,
1341 : InvalidXLogRecPtr,
1342 364 : XL_ROUTINE(.page_read = logical_read_xlog_page,
1343 : .segment_open = WalSndSegmentOpen,
1344 : .segment_close = wal_segment_close),
1345 : WalSndPrepareWrite, WalSndWriteData,
1346 : WalSndUpdateProgress);
1347 :
1348 : /*
1349 : * Signal that we don't need the timeout mechanism. We're just
1350 : * creating the replication slot and don't yet accept feedback
1351 : * messages or send keepalives. As we possibly need to wait for
1352 : * further WAL the walsender would otherwise possibly be killed too
1353 : * soon.
1354 : */
1355 364 : last_reply_timestamp = 0;
1356 :
1357 : /* build initial snapshot, might take a while */
1358 364 : DecodingContextFindStartpoint(ctx);
1359 :
1360 : /*
1361 : * Export or use the snapshot if we've been asked to do so.
1362 : *
1363 : * NB. We will convert the snapbuild.c kind of snapshot to normal
1364 : * snapshot when doing this.
1365 : */
1366 364 : if (snapshot_action == CRS_EXPORT_SNAPSHOT)
1367 : {
1368 1 : snapshot_name = SnapBuildExportSnapshot(ctx->snapshot_builder);
1369 : }
1370 363 : else if (snapshot_action == CRS_USE_SNAPSHOT)
1371 : {
1372 : Snapshot snap;
1373 :
1374 214 : snap = SnapBuildInitialSnapshot(ctx->snapshot_builder);
1375 214 : RestoreTransactionSnapshot(snap, MyProc);
1376 : }
1377 :
1378 : /* don't need the decoding context anymore */
1379 364 : FreeDecodingContext(ctx);
1380 :
1381 364 : if (!cmd->temporary)
1382 364 : ReplicationSlotPersist();
1383 : }
1384 :
1385 515 : snprintf(xloc, sizeof(xloc), "%X/%08X",
1386 515 : LSN_FORMAT_ARGS(MyReplicationSlot->data.confirmed_flush));
1387 :
1388 515 : dest = CreateDestReceiver(DestRemoteSimple);
1389 :
1390 : /*----------
1391 : * Need a tuple descriptor representing four columns:
1392 : * - first field: the slot name
1393 : * - second field: LSN at which we became consistent
1394 : * - third field: exported snapshot's name
1395 : * - fourth field: output plugin
1396 : */
1397 515 : tupdesc = CreateTemplateTupleDesc(4);
1398 515 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 1, "slot_name",
1399 : TEXTOID, -1, 0);
1400 515 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 2, "consistent_point",
1401 : TEXTOID, -1, 0);
1402 515 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 3, "snapshot_name",
1403 : TEXTOID, -1, 0);
1404 515 : TupleDescInitBuiltinEntry(tupdesc, (AttrNumber) 4, "output_plugin",
1405 : TEXTOID, -1, 0);
1406 515 : TupleDescFinalize(tupdesc);
1407 :
1408 : /* prepare for projection of tuples */
1409 515 : tstate = begin_tup_output_tupdesc(dest, tupdesc, &TTSOpsVirtual);
1410 :
1411 : /* slot_name */
1412 515 : slot_name = NameStr(MyReplicationSlot->data.name);
1413 515 : values[0] = CStringGetTextDatum(slot_name);
1414 :
1415 : /* consistent wal location */
1416 515 : values[1] = CStringGetTextDatum(xloc);
1417 :
1418 : /* snapshot name, or NULL if none */
1419 515 : if (snapshot_name != NULL)
1420 1 : values[2] = CStringGetTextDatum(snapshot_name);
1421 : else
1422 514 : nulls[2] = true;
1423 :
1424 : /* plugin, or NULL if none */
1425 515 : if (cmd->plugin != NULL)
1426 364 : values[3] = CStringGetTextDatum(cmd->plugin);
1427 : else
1428 151 : nulls[3] = true;
1429 :
1430 : /* send it to dest */
1431 515 : do_tup_output(tstate, values, nulls);
1432 515 : end_tup_output(tstate);
1433 :
1434 515 : ReplicationSlotRelease();
1435 515 : }
1436 :
1437 : /*
1438 : * Get rid of a replication slot that is no longer wanted.
1439 : */
1440 : static void
1441 296 : DropReplicationSlot(DropReplicationSlotCmd *cmd)
1442 : {
1443 296 : ReplicationSlotDrop(cmd->slotname, !cmd->wait);
1444 295 : }
1445 :
1446 : /*
1447 : * Change the definition of a replication slot.
1448 : */
1449 : static void
1450 7 : AlterReplicationSlot(AlterReplicationSlotCmd *cmd)
1451 : {
1452 7 : bool failover_given = false;
1453 7 : bool two_phase_given = false;
1454 : bool failover;
1455 : bool two_phase;
1456 :
1457 : /* Parse options */
1458 21 : foreach_ptr(DefElem, defel, cmd->options)
1459 : {
1460 7 : if (strcmp(defel->defname, "failover") == 0)
1461 : {
1462 6 : if (failover_given)
1463 0 : ereport(ERROR,
1464 : (errcode(ERRCODE_SYNTAX_ERROR),
1465 : errmsg("conflicting or redundant options")));
1466 6 : failover_given = true;
1467 6 : failover = defGetBoolean(defel);
1468 : }
1469 1 : else if (strcmp(defel->defname, "two_phase") == 0)
1470 : {
1471 1 : if (two_phase_given)
1472 0 : ereport(ERROR,
1473 : (errcode(ERRCODE_SYNTAX_ERROR),
1474 : errmsg("conflicting or redundant options")));
1475 1 : two_phase_given = true;
1476 1 : two_phase = defGetBoolean(defel);
1477 : }
1478 : else
1479 0 : elog(ERROR, "unrecognized option: %s", defel->defname);
1480 : }
1481 :
1482 7 : ReplicationSlotAlter(cmd->slotname,
1483 : failover_given ? &failover : NULL,
1484 : two_phase_given ? &two_phase : NULL);
1485 5 : }
1486 :
1487 : /*
1488 : * Load previously initiated logical slot and prepare for sending data (via
1489 : * WalSndLoop).
1490 : */
1491 : static void
1492 473 : StartLogicalReplication(StartReplicationCmd *cmd)
1493 : {
1494 : StringInfoData buf;
1495 : QueryCompletion qc;
1496 :
1497 : /* make sure that our requirements are still fulfilled */
1498 473 : CheckLogicalDecodingRequirements(false);
1499 :
1500 : Assert(!MyReplicationSlot);
1501 :
1502 471 : ReplicationSlotAcquire(cmd->slotname, true, true);
1503 :
1504 : /*
1505 : * Force a disconnect, so that the decoding code doesn't need to care
1506 : * about an eventual switch from running in recovery, to running in a
1507 : * normal environment. Client code is expected to handle reconnects.
1508 : */
1509 466 : if (am_cascading_walsender && !RecoveryInProgress())
1510 : {
1511 0 : ereport(LOG,
1512 : (errmsg("terminating walsender process after promotion")));
1513 0 : got_STOPPING = true;
1514 : }
1515 :
1516 : /*
1517 : * Create our decoding context, making it start at the previously ack'ed
1518 : * position.
1519 : *
1520 : * Do this before sending a CopyBothResponse message, so that any errors
1521 : * are reported early.
1522 : */
1523 465 : logical_decoding_ctx =
1524 466 : CreateDecodingContext(cmd->startpoint, cmd->options, false,
1525 466 : XL_ROUTINE(.page_read = logical_read_xlog_page,
1526 : .segment_open = WalSndSegmentOpen,
1527 : .segment_close = wal_segment_close),
1528 : WalSndPrepareWrite, WalSndWriteData,
1529 : WalSndUpdateProgress);
1530 465 : xlogreader = logical_decoding_ctx->reader;
1531 :
1532 465 : WalSndSetState(WALSNDSTATE_CATCHUP);
1533 :
1534 : /* Send a CopyBothResponse message, and start streaming */
1535 465 : pq_beginmessage(&buf, PqMsg_CopyBothResponse);
1536 465 : pq_sendbyte(&buf, 0);
1537 465 : pq_sendint16(&buf, 0);
1538 465 : pq_endmessage(&buf);
1539 465 : pq_flush();
1540 :
1541 : /* Start reading WAL from the oldest required WAL. */
1542 465 : XLogBeginRead(logical_decoding_ctx->reader,
1543 465 : MyReplicationSlot->data.restart_lsn);
1544 :
1545 : /*
1546 : * Report the location after which we'll send out further commits as the
1547 : * current sentPtr.
1548 : */
1549 465 : sentPtr = MyReplicationSlot->data.confirmed_flush;
1550 :
1551 : /* Also update the sent position status in shared memory */
1552 465 : SpinLockAcquire(&MyWalSnd->mutex);
1553 465 : MyWalSnd->sentPtr = MyReplicationSlot->data.restart_lsn;
1554 465 : SpinLockRelease(&MyWalSnd->mutex);
1555 :
1556 465 : replication_active = true;
1557 :
1558 465 : SyncRepInitConfig();
1559 :
1560 : /* Main loop of walsender */
1561 465 : WalSndLoop(XLogSendLogical);
1562 :
1563 208 : FreeDecodingContext(logical_decoding_ctx);
1564 208 : ReplicationSlotRelease();
1565 :
1566 208 : replication_active = false;
1567 208 : if (got_STOPPING)
1568 0 : proc_exit(0);
1569 208 : WalSndSetState(WALSNDSTATE_STARTUP);
1570 :
1571 : /* Get out of COPY mode (CommandComplete). */
1572 208 : SetQueryCompletion(&qc, CMDTAG_COPY, 0);
1573 208 : EndCommand(&qc, DestRemote, false);
1574 208 : }
1575 :
1576 : /*
1577 : * LogicalDecodingContext 'prepare_write' callback.
1578 : *
1579 : * Prepare a write into a StringInfo.
1580 : *
1581 : * Don't do anything lasting in here, it's quite possible that nothing will be done
1582 : * with the data.
1583 : */
1584 : static void
1585 205642 : WalSndPrepareWrite(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid, bool last_write)
1586 : {
1587 : /* can't have sync rep confused by sending the same LSN several times */
1588 205642 : if (!last_write)
1589 445 : lsn = InvalidXLogRecPtr;
1590 :
1591 205642 : resetStringInfo(ctx->out);
1592 :
1593 205642 : pq_sendbyte(ctx->out, PqReplMsg_WALData);
1594 205642 : pq_sendint64(ctx->out, lsn); /* dataStart */
1595 205642 : pq_sendint64(ctx->out, lsn); /* walEnd */
1596 :
1597 : /*
1598 : * Fill out the sendtime later, just as it's done in XLogSendPhysical, but
1599 : * reserve space here.
1600 : */
1601 205642 : pq_sendint64(ctx->out, 0); /* sendtime */
1602 205642 : }
1603 :
1604 : /*
1605 : * LogicalDecodingContext 'write' callback.
1606 : *
1607 : * Actually write out data previously prepared by WalSndPrepareWrite out to
1608 : * the network. Take as long as needed, but process replies from the other
1609 : * side and check timeouts during that.
1610 : */
1611 : static void
1612 205642 : WalSndWriteData(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid,
1613 : bool last_write)
1614 : {
1615 : TimestampTz now;
1616 :
1617 : /*
1618 : * Fill the send timestamp last, so that it is taken as late as possible.
1619 : * This is somewhat ugly, but the protocol is set as it's already used for
1620 : * several releases by streaming physical replication.
1621 : */
1622 205642 : resetStringInfo(&tmpbuf);
1623 205642 : now = GetCurrentTimestamp();
1624 205642 : pq_sendint64(&tmpbuf, now);
1625 205642 : memcpy(&ctx->out->data[1 + sizeof(int64) + sizeof(int64)],
1626 205642 : tmpbuf.data, sizeof(int64));
1627 :
1628 : /* output previously gathered data in a CopyData packet */
1629 205642 : pq_putmessage_noblock(PqMsg_CopyData, ctx->out->data, ctx->out->len);
1630 :
1631 205642 : CHECK_FOR_INTERRUPTS();
1632 :
1633 : /* Try to flush pending output to the client */
1634 205642 : if (pq_flush_if_writable() != 0)
1635 15 : WalSndShutdown();
1636 :
1637 : /* Try taking fast path unless we get too close to walsender timeout. */
1638 205627 : if (now < TimestampTzPlusMilliseconds(last_reply_timestamp,
1639 205627 : wal_sender_timeout / 2) &&
1640 205627 : !pq_is_send_pending())
1641 : {
1642 205064 : return;
1643 : }
1644 :
1645 : /* If we have pending write here, go to slow path */
1646 563 : ProcessPendingWrites();
1647 : }
1648 :
1649 : /*
1650 : * Handle configuration reload.
1651 : *
1652 : * Process the pending configuration file reload and reinitializes synchronous
1653 : * replication settings. Also releases any waiters that may now be satisfied due
1654 : * to changes in synchronous replication requirements.
1655 : */
1656 : static void
1657 1176593 : WalSndHandleConfigReload(void)
1658 : {
1659 1176593 : if (!ConfigReloadPending)
1660 1176556 : return;
1661 :
1662 37 : ConfigReloadPending = false;
1663 37 : ProcessConfigFile(PGC_SIGHUP);
1664 37 : SyncRepInitConfig();
1665 :
1666 : /*
1667 : * Recheck and release any now-satisfied waiters after config reload
1668 : * changes synchronous replication requirements (e.g., reducing the number
1669 : * of sync standbys or changing the standby names).
1670 : */
1671 37 : if (!am_cascading_walsender)
1672 34 : SyncRepReleaseWaiters();
1673 : }
1674 :
1675 : /*
1676 : * Wait until there is no pending write. Also process replies from the other
1677 : * side and check timeouts during that.
1678 : */
1679 : static void
1680 563 : ProcessPendingWrites(void)
1681 : {
1682 : for (;;)
1683 783 : {
1684 : long sleeptime;
1685 :
1686 : /* Check for input from the client */
1687 1346 : ProcessRepliesIfAny();
1688 :
1689 : /* die if timeout was reached */
1690 1346 : WalSndCheckTimeOut();
1691 :
1692 : /*
1693 : * During shutdown, die if the shutdown timeout expires. Call this
1694 : * before WalSndComputeSleeptime() so the timeout is considered when
1695 : * computing sleep time.
1696 : */
1697 1346 : WalSndCheckShutdownTimeout();
1698 :
1699 : /* Send keepalive if the time has come */
1700 1345 : WalSndKeepaliveIfNecessary();
1701 :
1702 1345 : if (!pq_is_send_pending())
1703 562 : break;
1704 :
1705 783 : sleeptime = WalSndComputeSleeptime(GetCurrentTimestamp());
1706 :
1707 : /* Sleep until something happens or we time out */
1708 783 : WalSndWait(WL_SOCKET_WRITEABLE | WL_SOCKET_READABLE, sleeptime,
1709 : WAIT_EVENT_WAL_SENDER_WRITE_DATA);
1710 :
1711 : /* Clear any already-pending wakeups */
1712 783 : ResetLatch(MyLatch);
1713 :
1714 783 : CHECK_FOR_INTERRUPTS();
1715 :
1716 : /* Process any requests or signals received recently */
1717 783 : WalSndHandleConfigReload();
1718 :
1719 : /* Try to flush pending output to the client */
1720 783 : if (pq_flush_if_writable() != 0)
1721 0 : WalSndShutdown();
1722 : }
1723 :
1724 : /* reactivate latch so WalSndLoop knows to continue */
1725 562 : SetLatch(MyLatch);
1726 562 : }
1727 :
1728 : /*
1729 : * LogicalDecodingContext 'update_progress' callback.
1730 : *
1731 : * Write the current position to the lag tracker (see XLogSendPhysical).
1732 : *
1733 : * When skipping empty transactions, send a keepalive message if necessary.
1734 : */
1735 : static void
1736 3149 : WalSndUpdateProgress(LogicalDecodingContext *ctx, XLogRecPtr lsn, TransactionId xid,
1737 : bool skipped_xact)
1738 : {
1739 : static TimestampTz sendTime = 0;
1740 3149 : TimestampTz now = GetCurrentTimestamp();
1741 3149 : bool pending_writes = false;
1742 3149 : bool end_xact = ctx->end_xact;
1743 :
1744 : /*
1745 : * Track lag no more than once per WALSND_LOGICAL_LAG_TRACK_INTERVAL_MS to
1746 : * avoid flooding the lag tracker when we commit frequently.
1747 : *
1748 : * We don't have a mechanism to get the ack for any LSN other than end
1749 : * xact LSN from the downstream. So, we track lag only for end of
1750 : * transaction LSN.
1751 : */
1752 : #define WALSND_LOGICAL_LAG_TRACK_INTERVAL_MS 1000
1753 3149 : if (end_xact && TimestampDifferenceExceeds(sendTime, now,
1754 : WALSND_LOGICAL_LAG_TRACK_INTERVAL_MS))
1755 : {
1756 352 : LagTrackerWrite(lsn, now);
1757 352 : sendTime = now;
1758 : }
1759 :
1760 : /*
1761 : * When skipping empty transactions in synchronous replication, we send a
1762 : * keepalive message to avoid delaying such transactions.
1763 : *
1764 : * It is okay to check sync_standbys_status without lock here as in the
1765 : * worst case we will just send an extra keepalive message when it is
1766 : * really not required.
1767 : */
1768 3149 : if (skipped_xact &&
1769 810 : SyncRepRequested() &&
1770 810 : (((volatile WalSndCtlData *) WalSndCtl)->sync_standbys_status & SYNC_STANDBY_DEFINED))
1771 : {
1772 0 : WalSndKeepalive(false, lsn);
1773 :
1774 : /* Try to flush pending output to the client */
1775 0 : if (pq_flush_if_writable() != 0)
1776 0 : WalSndShutdown();
1777 :
1778 : /* If we have pending write here, make sure it's actually flushed */
1779 0 : if (pq_is_send_pending())
1780 0 : pending_writes = true;
1781 : }
1782 :
1783 : /*
1784 : * Process pending writes if any or try to send a keepalive if required.
1785 : * We don't need to try sending keep alive messages at the transaction end
1786 : * as that will be done at a later point in time. This is required only
1787 : * for large transactions where we don't send any changes to the
1788 : * downstream and the receiver can timeout due to that.
1789 : */
1790 3149 : if (pending_writes || (!end_xact &&
1791 1744 : now >= TimestampTzPlusMilliseconds(last_reply_timestamp,
1792 : wal_sender_timeout / 2)))
1793 0 : ProcessPendingWrites();
1794 3149 : }
1795 :
1796 : /*
1797 : * Wake up the logical walsender processes with logical failover slots if the
1798 : * currently acquired physical slot is specified in synchronized_standby_slots GUC.
1799 : */
1800 : void
1801 39825 : PhysicalWakeupLogicalWalSnd(void)
1802 : {
1803 : Assert(MyReplicationSlot && SlotIsPhysical(MyReplicationSlot));
1804 :
1805 : /*
1806 : * If we are running in a standby, there is no need to wake up walsenders.
1807 : * This is because we do not support syncing slots to cascading standbys,
1808 : * so, there are no walsenders waiting for standbys to catch up.
1809 : */
1810 39825 : if (RecoveryInProgress())
1811 47 : return;
1812 :
1813 39778 : if (SlotExistsInSyncStandbySlots(NameStr(MyReplicationSlot->data.name)))
1814 9 : ConditionVariableBroadcast(&WalSndCtl->wal_confirm_rcv_cv);
1815 : }
1816 :
1817 : /*
1818 : * Returns true if not all standbys have caught up to the flushed position
1819 : * (flushed_lsn) when the current acquired slot is a logical failover
1820 : * slot and we are streaming; otherwise, returns false.
1821 : *
1822 : * If returning true, the function sets the appropriate wait event in
1823 : * wait_event; otherwise, wait_event is set to 0.
1824 : */
1825 : static bool
1826 38799 : NeedToWaitForStandbys(XLogRecPtr flushed_lsn, uint32 *wait_event)
1827 : {
1828 38799 : int elevel = got_STOPPING ? ERROR : WARNING;
1829 : bool failover_slot;
1830 :
1831 38799 : failover_slot = (replication_active && MyReplicationSlot->data.failover);
1832 :
1833 : /*
1834 : * Note that after receiving the shutdown signal, an ERROR is reported if
1835 : * any slots are dropped, invalidated, or inactive. This measure is taken
1836 : * to prevent the walsender from waiting indefinitely.
1837 : */
1838 38799 : if (failover_slot && !StandbySlotsHaveCaughtup(flushed_lsn, elevel))
1839 : {
1840 13 : *wait_event = WAIT_EVENT_WAIT_FOR_STANDBY_CONFIRMATION;
1841 13 : return true;
1842 : }
1843 :
1844 38785 : *wait_event = 0;
1845 38785 : return false;
1846 : }
1847 :
1848 : /*
1849 : * Returns true if we need to wait for WALs to be flushed to disk, or if not
1850 : * all standbys have caught up to the flushed position (flushed_lsn) when the
1851 : * current acquired slot is a logical failover slot and we are
1852 : * streaming; otherwise, returns false.
1853 : *
1854 : * If returning true, the function sets the appropriate wait event in
1855 : * wait_event; otherwise, wait_event is set to 0.
1856 : */
1857 : static bool
1858 50581 : NeedToWaitForWal(XLogRecPtr target_lsn, XLogRecPtr flushed_lsn,
1859 : uint32 *wait_event)
1860 : {
1861 : /* Check if we need to wait for WALs to be flushed to disk */
1862 50581 : if (target_lsn > flushed_lsn)
1863 : {
1864 24236 : *wait_event = WAIT_EVENT_WAL_SENDER_WAIT_FOR_WAL;
1865 24236 : return true;
1866 : }
1867 :
1868 : /* Check if the standby slots have caught up to the flushed position */
1869 26345 : return NeedToWaitForStandbys(flushed_lsn, wait_event);
1870 : }
1871 :
1872 : /*
1873 : * Wait till WAL < loc is flushed to disk so it can be safely sent to client.
1874 : *
1875 : * If the walsender holds a logical failover slot, we also wait for all the
1876 : * specified streaming replication standby servers to confirm receipt of WAL
1877 : * up to RecentFlushPtr. It is beneficial to wait here for the confirmation
1878 : * up to RecentFlushPtr rather than waiting before transmitting each change
1879 : * to logical subscribers, which is already covered by RecentFlushPtr.
1880 : *
1881 : * Returns end LSN of flushed WAL. Normally this will be >= loc, but if we
1882 : * detect a shutdown request (either from postmaster or client) we will return
1883 : * early, so caller must always check.
1884 : */
1885 : static XLogRecPtr
1886 39176 : WalSndWaitForWal(XLogRecPtr loc)
1887 : {
1888 : int wakeEvents;
1889 39176 : uint32 wait_event = 0;
1890 : static XLogRecPtr RecentFlushPtr = InvalidXLogRecPtr;
1891 39176 : TimestampTz last_flush = 0;
1892 :
1893 : /*
1894 : * Fast path to avoid acquiring the spinlock in case we already know we
1895 : * have enough WAL available and all the standby servers have confirmed
1896 : * receipt of WAL up to RecentFlushPtr. This is particularly interesting
1897 : * if we're far behind.
1898 : */
1899 39176 : if (XLogRecPtrIsValid(RecentFlushPtr) &&
1900 38546 : !NeedToWaitForWal(loc, RecentFlushPtr, &wait_event))
1901 24388 : return RecentFlushPtr;
1902 :
1903 : /*
1904 : * Within the loop, we wait for the necessary WALs to be flushed to disk
1905 : * first, followed by waiting for standbys to catch up if there are enough
1906 : * WALs (see NeedToWaitForWal()) or upon receiving the shutdown signal.
1907 : */
1908 : for (;;)
1909 9921 : {
1910 24709 : bool wait_for_standby_at_stop = false;
1911 : long sleeptime;
1912 : TimestampTz now;
1913 :
1914 : /* Clear any already-pending wakeups */
1915 24709 : ResetLatch(MyLatch);
1916 :
1917 24709 : CHECK_FOR_INTERRUPTS();
1918 :
1919 : /* Process any requests or signals received recently */
1920 24702 : WalSndHandleConfigReload();
1921 :
1922 : /* Check for input from the client */
1923 24702 : ProcessRepliesIfAny();
1924 :
1925 : /*
1926 : * If we're shutting down, trigger pending WAL to be written out,
1927 : * otherwise we'd possibly end up waiting for WAL that never gets
1928 : * written, because walwriter has shut down already.
1929 : *
1930 : * Note that GetXLogInsertEndRecPtr() is used to obtain the WAL flush
1931 : * request location instead of GetXLogInsertRecPtr(). Because if the
1932 : * last WAL record ends at a page boundary, GetXLogInsertRecPtr() can
1933 : * return an LSN pointing past the page header, which may cause
1934 : * XLogFlush() to report an error.
1935 : */
1936 24489 : if (got_STOPPING && !RecoveryInProgress())
1937 10654 : XLogFlush(GetXLogInsertEndRecPtr());
1938 :
1939 : /*
1940 : * To avoid the scenario where standbys need to catch up to a newer
1941 : * WAL location in each iteration, we update our idea of the currently
1942 : * flushed position only if we are not waiting for standbys to catch
1943 : * up.
1944 : */
1945 24489 : if (wait_event != WAIT_EVENT_WAIT_FOR_STANDBY_CONFIRMATION)
1946 : {
1947 24476 : if (!RecoveryInProgress())
1948 22481 : RecentFlushPtr = GetFlushRecPtr(NULL);
1949 : else
1950 1995 : RecentFlushPtr = GetXLogReplayRecPtr(NULL);
1951 : }
1952 :
1953 : /*
1954 : * If postmaster asked us to stop and the standby slots have caught up
1955 : * to the flushed position, don't wait anymore.
1956 : *
1957 : * It's important to do this check after the recomputation of
1958 : * RecentFlushPtr, so we can send all remaining data before shutting
1959 : * down.
1960 : */
1961 24489 : if (got_STOPPING)
1962 : {
1963 12454 : if (NeedToWaitForStandbys(RecentFlushPtr, &wait_event))
1964 1 : wait_for_standby_at_stop = true;
1965 : else
1966 12452 : break;
1967 : }
1968 :
1969 : /*
1970 : * We only send regular messages to the client for full decoded
1971 : * transactions, but a synchronous replication and walsender shutdown
1972 : * possibly are waiting for a later location. So, before sleeping, we
1973 : * send a ping containing the flush location. If the receiver is
1974 : * otherwise idle, this keepalive will trigger a reply. Processing the
1975 : * reply will update these MyWalSnd locations.
1976 : */
1977 12036 : if (MyWalSnd->flush < sentPtr &&
1978 2553 : MyWalSnd->write < sentPtr &&
1979 1962 : !waiting_for_ping_response)
1980 1962 : WalSndKeepalive(false, InvalidXLogRecPtr);
1981 :
1982 : /*
1983 : * Exit the loop if already caught up and doesn't need to wait for
1984 : * standby slots.
1985 : */
1986 12036 : if (!wait_for_standby_at_stop &&
1987 12035 : !NeedToWaitForWal(loc, RecentFlushPtr, &wait_event))
1988 1945 : break;
1989 :
1990 : /*
1991 : * Waiting for new WAL or waiting for standbys to catch up. Since we
1992 : * need to wait, we're now caught up.
1993 : */
1994 10091 : WalSndCaughtUp = true;
1995 :
1996 : /*
1997 : * Try to flush any pending output to the client.
1998 : */
1999 10091 : if (pq_flush_if_writable() != 0)
2000 0 : WalSndShutdown();
2001 :
2002 : /*
2003 : * If we have received CopyDone from the client, sent CopyDone
2004 : * ourselves, and the output buffer is empty, it's time to exit
2005 : * streaming, so fail the current WAL fetch request.
2006 : */
2007 10091 : if (streamingDoneReceiving && streamingDoneSending &&
2008 170 : !pq_is_send_pending())
2009 170 : break;
2010 :
2011 : /* die if timeout was reached */
2012 9921 : WalSndCheckTimeOut();
2013 :
2014 : /*
2015 : * During shutdown, die if the shutdown timeout expires. Call this
2016 : * before WalSndComputeSleeptime() so the timeout is considered when
2017 : * computing sleep time.
2018 : */
2019 9921 : WalSndCheckShutdownTimeout();
2020 :
2021 : /* Send keepalive if the time has come */
2022 9921 : WalSndKeepaliveIfNecessary();
2023 :
2024 : /*
2025 : * Sleep until something happens or we time out. Also wait for the
2026 : * socket becoming writable, if there's still pending output.
2027 : * Otherwise we might sit on sendable output data while waiting for
2028 : * new WAL to be generated. (But if we have nothing to send, we don't
2029 : * want to wake on socket-writable.)
2030 : */
2031 9921 : now = GetCurrentTimestamp();
2032 9921 : sleeptime = WalSndComputeSleeptime(now);
2033 :
2034 9921 : wakeEvents = WL_SOCKET_READABLE;
2035 :
2036 9921 : if (pq_is_send_pending())
2037 0 : wakeEvents |= WL_SOCKET_WRITEABLE;
2038 :
2039 : Assert(wait_event != 0);
2040 :
2041 : /* Report IO statistics, if needed */
2042 9921 : if (TimestampDifferenceExceeds(last_flush, now,
2043 : WALSENDER_STATS_FLUSH_INTERVAL))
2044 : {
2045 1640 : pgstat_flush_io(false);
2046 1640 : (void) pgstat_flush_backend(false, PGSTAT_BACKEND_FLUSH_IO);
2047 1640 : last_flush = now;
2048 : }
2049 :
2050 9921 : WalSndWait(wakeEvents, sleeptime, wait_event);
2051 : }
2052 :
2053 : /* reactivate latch so WalSndLoop knows to continue */
2054 14567 : SetLatch(MyLatch);
2055 14567 : return RecentFlushPtr;
2056 : }
2057 :
2058 : /*
2059 : * Execute an incoming replication command.
2060 : *
2061 : * Returns true if the cmd_string was recognized as WalSender command, false
2062 : * if not.
2063 : */
2064 : bool
2065 5880 : exec_replication_command(const char *cmd_string)
2066 : {
2067 : yyscan_t scanner;
2068 : int parse_rc;
2069 : Node *cmd_node;
2070 : const char *cmdtag;
2071 5880 : MemoryContext old_context = CurrentMemoryContext;
2072 :
2073 : /* We save and re-use the cmd_context across calls */
2074 : static MemoryContext cmd_context = NULL;
2075 :
2076 : /*
2077 : * If WAL sender has been told that shutdown is getting close, switch its
2078 : * status accordingly to handle the next replication commands correctly.
2079 : */
2080 5880 : if (got_STOPPING)
2081 0 : WalSndSetState(WALSNDSTATE_STOPPING);
2082 :
2083 : /*
2084 : * Throw error if in stopping mode. We need prevent commands that could
2085 : * generate WAL while the shutdown checkpoint is being written. To be
2086 : * safe, we just prohibit all new commands.
2087 : */
2088 5880 : if (MyWalSnd->state == WALSNDSTATE_STOPPING)
2089 0 : ereport(ERROR,
2090 : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
2091 : errmsg("cannot execute new commands while WAL sender is in stopping mode")));
2092 :
2093 : /*
2094 : * CREATE_REPLICATION_SLOT ... LOGICAL exports a snapshot until the next
2095 : * command arrives. Clean up the old stuff if there's anything.
2096 : */
2097 5880 : SnapBuildClearExportedSnapshot();
2098 :
2099 5880 : CHECK_FOR_INTERRUPTS();
2100 :
2101 : /*
2102 : * Prepare to parse and execute the command.
2103 : *
2104 : * Because replication command execution can involve beginning or ending
2105 : * transactions, we need a working context that will survive that, so we
2106 : * make it a child of TopMemoryContext. That in turn creates a hazard of
2107 : * long-lived memory leaks if we lose track of the working context. We
2108 : * deal with that by creating it only once per walsender, and resetting it
2109 : * for each new command. (Normally this reset is a no-op, but if the
2110 : * prior exec_replication_command call failed with an error, it won't be.)
2111 : *
2112 : * This is subtler than it looks. The transactions we manage can extend
2113 : * across replication commands, indeed SnapBuildClearExportedSnapshot
2114 : * might have just ended one. Because transaction exit will revert to the
2115 : * memory context that was current at transaction start, we need to be
2116 : * sure that that context is still valid. That motivates re-using the
2117 : * same cmd_context rather than making a new one each time.
2118 : */
2119 5880 : if (cmd_context == NULL)
2120 1300 : cmd_context = AllocSetContextCreate(TopMemoryContext,
2121 : "Replication command context",
2122 : ALLOCSET_DEFAULT_SIZES);
2123 : else
2124 4580 : MemoryContextReset(cmd_context);
2125 :
2126 5880 : MemoryContextSwitchTo(cmd_context);
2127 :
2128 5880 : replication_scanner_init(cmd_string, &scanner);
2129 :
2130 : /*
2131 : * Is it a WalSender command?
2132 : */
2133 5880 : if (!replication_scanner_is_replication_command(scanner))
2134 : {
2135 : /* Nope; clean up and get out. */
2136 2613 : replication_scanner_finish(scanner);
2137 :
2138 2613 : MemoryContextSwitchTo(old_context);
2139 2613 : MemoryContextReset(cmd_context);
2140 :
2141 : /* XXX this is a pretty random place to make this check */
2142 2613 : if (MyDatabaseId == InvalidOid)
2143 0 : ereport(ERROR,
2144 : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2145 : errmsg("cannot execute SQL commands in WAL sender for physical replication")));
2146 :
2147 : /* Tell the caller that this wasn't a WalSender command. */
2148 2613 : return false;
2149 : }
2150 :
2151 : /*
2152 : * Looks like a WalSender command, so parse it.
2153 : */
2154 3267 : parse_rc = replication_yyparse(&cmd_node, scanner);
2155 3267 : if (parse_rc != 0)
2156 0 : ereport(ERROR,
2157 : (errcode(ERRCODE_SYNTAX_ERROR),
2158 : errmsg_internal("replication command parser returned %d",
2159 : parse_rc)));
2160 3267 : replication_scanner_finish(scanner);
2161 :
2162 : /*
2163 : * Report query to various monitoring facilities. For this purpose, we
2164 : * report replication commands just like SQL commands.
2165 : */
2166 3267 : debug_query_string = cmd_string;
2167 :
2168 3267 : pgstat_report_activity(STATE_RUNNING, cmd_string);
2169 :
2170 : /*
2171 : * Log replication command if log_replication_commands is enabled. Even
2172 : * when it's disabled, log the command with DEBUG1 level for backward
2173 : * compatibility.
2174 : */
2175 3267 : ereport(log_replication_commands ? LOG : DEBUG1,
2176 : (errmsg("received replication command: %s", cmd_string)));
2177 :
2178 : /*
2179 : * Disallow replication commands in aborted transaction blocks.
2180 : */
2181 3267 : if (IsAbortedTransactionBlockState())
2182 0 : ereport(ERROR,
2183 : (errcode(ERRCODE_IN_FAILED_SQL_TRANSACTION),
2184 : errmsg("current transaction is aborted, "
2185 : "commands ignored until end of transaction block")));
2186 :
2187 3267 : CHECK_FOR_INTERRUPTS();
2188 :
2189 : /*
2190 : * Allocate buffers that will be used for each outgoing and incoming
2191 : * message. We do this just once per command to reduce palloc overhead.
2192 : */
2193 3267 : initStringInfo(&output_message);
2194 3267 : initStringInfo(&reply_message);
2195 3267 : initStringInfo(&tmpbuf);
2196 :
2197 3267 : switch (cmd_node->type)
2198 : {
2199 817 : case T_IdentifySystemCmd:
2200 817 : cmdtag = "IDENTIFY_SYSTEM";
2201 817 : set_ps_display(cmdtag);
2202 817 : IdentifySystem();
2203 817 : EndReplicationCommand(cmdtag);
2204 817 : break;
2205 :
2206 6 : case T_ReadReplicationSlotCmd:
2207 6 : cmdtag = "READ_REPLICATION_SLOT";
2208 6 : set_ps_display(cmdtag);
2209 6 : ReadReplicationSlot((ReadReplicationSlotCmd *) cmd_node);
2210 5 : EndReplicationCommand(cmdtag);
2211 5 : break;
2212 :
2213 198 : case T_BaseBackupCmd:
2214 198 : cmdtag = "BASE_BACKUP";
2215 198 : set_ps_display(cmdtag);
2216 198 : PreventInTransactionBlock(true, cmdtag);
2217 198 : SendBaseBackup((BaseBackupCmd *) cmd_node, uploaded_manifest);
2218 170 : EndReplicationCommand(cmdtag);
2219 170 : break;
2220 :
2221 516 : case T_CreateReplicationSlotCmd:
2222 516 : cmdtag = "CREATE_REPLICATION_SLOT";
2223 516 : set_ps_display(cmdtag);
2224 516 : CreateReplicationSlot((CreateReplicationSlotCmd *) cmd_node);
2225 515 : EndReplicationCommand(cmdtag);
2226 515 : break;
2227 :
2228 296 : case T_DropReplicationSlotCmd:
2229 296 : cmdtag = "DROP_REPLICATION_SLOT";
2230 296 : set_ps_display(cmdtag);
2231 296 : DropReplicationSlot((DropReplicationSlotCmd *) cmd_node);
2232 295 : EndReplicationCommand(cmdtag);
2233 295 : break;
2234 :
2235 7 : case T_AlterReplicationSlotCmd:
2236 7 : cmdtag = "ALTER_REPLICATION_SLOT";
2237 7 : set_ps_display(cmdtag);
2238 7 : AlterReplicationSlot((AlterReplicationSlotCmd *) cmd_node);
2239 5 : EndReplicationCommand(cmdtag);
2240 5 : break;
2241 :
2242 776 : case T_StartReplicationCmd:
2243 : {
2244 776 : StartReplicationCmd *cmd = (StartReplicationCmd *) cmd_node;
2245 :
2246 776 : cmdtag = "START_REPLICATION";
2247 776 : set_ps_display(cmdtag);
2248 776 : PreventInTransactionBlock(true, cmdtag);
2249 :
2250 776 : if (cmd->kind == REPLICATION_KIND_PHYSICAL)
2251 303 : StartReplication(cmd);
2252 : else
2253 473 : StartLogicalReplication(cmd);
2254 :
2255 : /* dupe, but necessary per libpqrcv_endstreaming */
2256 371 : EndReplicationCommand(cmdtag);
2257 :
2258 : Assert(xlogreader != NULL);
2259 371 : break;
2260 : }
2261 :
2262 15 : case T_TimeLineHistoryCmd:
2263 15 : cmdtag = "TIMELINE_HISTORY";
2264 15 : set_ps_display(cmdtag);
2265 15 : PreventInTransactionBlock(true, cmdtag);
2266 15 : SendTimeLineHistory((TimeLineHistoryCmd *) cmd_node);
2267 15 : EndReplicationCommand(cmdtag);
2268 15 : break;
2269 :
2270 623 : case T_VariableShowStmt:
2271 : {
2272 623 : DestReceiver *dest = CreateDestReceiver(DestRemoteSimple);
2273 623 : VariableShowStmt *n = (VariableShowStmt *) cmd_node;
2274 :
2275 623 : cmdtag = "SHOW";
2276 623 : set_ps_display(cmdtag);
2277 :
2278 : /* syscache access needs a transaction environment */
2279 623 : StartTransactionCommand();
2280 623 : GetPGVariable(n->name, dest);
2281 623 : CommitTransactionCommand();
2282 623 : EndReplicationCommand(cmdtag);
2283 : }
2284 623 : break;
2285 :
2286 13 : case T_UploadManifestCmd:
2287 13 : cmdtag = "UPLOAD_MANIFEST";
2288 13 : set_ps_display(cmdtag);
2289 13 : PreventInTransactionBlock(true, cmdtag);
2290 13 : UploadManifest();
2291 12 : EndReplicationCommand(cmdtag);
2292 12 : break;
2293 :
2294 0 : default:
2295 0 : elog(ERROR, "unrecognized replication command node tag: %u",
2296 : cmd_node->type);
2297 : }
2298 :
2299 : /*
2300 : * Done. Revert to caller's memory context, and clean out the cmd_context
2301 : * to recover memory right away.
2302 : */
2303 2828 : MemoryContextSwitchTo(old_context);
2304 2828 : MemoryContextReset(cmd_context);
2305 :
2306 : /*
2307 : * We need not update ps display or pg_stat_activity, because PostgresMain
2308 : * will reset those to "idle". But we must reset debug_query_string to
2309 : * ensure it doesn't become a dangling pointer.
2310 : */
2311 2828 : debug_query_string = NULL;
2312 :
2313 2828 : return true;
2314 : }
2315 :
2316 : /*
2317 : * Process any incoming messages while streaming. Also checks if the remote
2318 : * end has closed the connection.
2319 : */
2320 : static void
2321 1177156 : ProcessRepliesIfAny(void)
2322 : {
2323 : unsigned char firstchar;
2324 : int maxmsglen;
2325 : int r;
2326 1177156 : bool received = false;
2327 :
2328 1177156 : last_processing = GetCurrentTimestamp();
2329 :
2330 : /*
2331 : * If we already received a CopyDone from the frontend, any subsequent
2332 : * message is the beginning of a new command, and should be processed in
2333 : * the main processing loop.
2334 : */
2335 2472665 : while (!streamingDoneReceiving)
2336 : {
2337 1294757 : pq_startmsgread();
2338 1294757 : r = pq_getbyte_if_available(&firstchar);
2339 1294757 : if (r < 0)
2340 : {
2341 : /* unexpected error or EOF */
2342 12 : ereport(COMMERROR,
2343 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2344 : errmsg("unexpected EOF on standby connection")));
2345 12 : proc_exit(0);
2346 : }
2347 1294745 : if (r == 0)
2348 : {
2349 : /* no data available without blocking */
2350 1176083 : pq_endmsgread();
2351 1176083 : break;
2352 : }
2353 :
2354 : /* Validate message type and set packet size limit */
2355 118662 : switch (firstchar)
2356 : {
2357 117982 : case PqMsg_CopyData:
2358 117982 : maxmsglen = PQ_LARGE_MESSAGE_LIMIT;
2359 117982 : break;
2360 680 : case PqMsg_CopyDone:
2361 : case PqMsg_Terminate:
2362 680 : maxmsglen = PQ_SMALL_MESSAGE_LIMIT;
2363 680 : break;
2364 0 : default:
2365 0 : ereport(FATAL,
2366 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2367 : errmsg("invalid standby message type \"%c\"",
2368 : firstchar)));
2369 : maxmsglen = 0; /* keep compiler quiet */
2370 : break;
2371 : }
2372 :
2373 : /* Read the message contents */
2374 118662 : resetStringInfo(&reply_message);
2375 118662 : if (pq_getmessage(&reply_message, maxmsglen))
2376 : {
2377 0 : ereport(COMMERROR,
2378 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2379 : errmsg("unexpected EOF on standby connection")));
2380 0 : proc_exit(0);
2381 : }
2382 :
2383 : /* ... and process it */
2384 118662 : switch (firstchar)
2385 : {
2386 : /*
2387 : * PqMsg_CopyData means a standby reply wrapped in a CopyData
2388 : * packet.
2389 : */
2390 117982 : case PqMsg_CopyData:
2391 117982 : ProcessStandbyMessage();
2392 117982 : received = true;
2393 117982 : break;
2394 :
2395 : /*
2396 : * PqMsg_CopyDone means the standby requested to finish
2397 : * streaming. Reply with CopyDone, if we had not sent that
2398 : * already.
2399 : */
2400 371 : case PqMsg_CopyDone:
2401 371 : if (!streamingDoneSending)
2402 : {
2403 360 : pq_putmessage_noblock(PqMsg_CopyDone, NULL, 0);
2404 360 : streamingDoneSending = true;
2405 : }
2406 :
2407 371 : streamingDoneReceiving = true;
2408 371 : received = true;
2409 371 : break;
2410 :
2411 : /*
2412 : * PqMsg_Terminate means that the standby is closing down the
2413 : * socket.
2414 : */
2415 309 : case PqMsg_Terminate:
2416 309 : proc_exit(0);
2417 :
2418 118353 : default:
2419 : Assert(false); /* NOT REACHED */
2420 : }
2421 : }
2422 :
2423 : /*
2424 : * Save the last reply timestamp if we've received at least one reply.
2425 : */
2426 1176835 : if (received)
2427 : {
2428 64663 : last_reply_timestamp = last_processing;
2429 64663 : waiting_for_ping_response = false;
2430 : }
2431 1176835 : }
2432 :
2433 : /*
2434 : * Process a status update message received from standby.
2435 : */
2436 : static void
2437 117982 : ProcessStandbyMessage(void)
2438 : {
2439 : char msgtype;
2440 :
2441 : /*
2442 : * Check message type from the first byte.
2443 : */
2444 117982 : msgtype = pq_getmsgbyte(&reply_message);
2445 :
2446 117982 : switch (msgtype)
2447 : {
2448 111009 : case PqReplMsg_StandbyStatusUpdate:
2449 111009 : ProcessStandbyReplyMessage();
2450 111009 : break;
2451 :
2452 155 : case PqReplMsg_HotStandbyFeedback:
2453 155 : ProcessStandbyHSFeedbackMessage();
2454 155 : break;
2455 :
2456 6818 : case PqReplMsg_PrimaryStatusRequest:
2457 6818 : ProcessStandbyPSRequestMessage();
2458 6818 : break;
2459 :
2460 0 : default:
2461 0 : ereport(COMMERROR,
2462 : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2463 : errmsg("unexpected message type \"%c\"", msgtype)));
2464 0 : proc_exit(0);
2465 : }
2466 117982 : }
2467 :
2468 : /*
2469 : * Remember that a walreceiver just confirmed receipt of lsn `lsn`.
2470 : */
2471 : static void
2472 88427 : PhysicalConfirmReceivedLocation(XLogRecPtr lsn)
2473 : {
2474 88427 : bool changed = false;
2475 88427 : ReplicationSlot *slot = MyReplicationSlot;
2476 :
2477 : Assert(XLogRecPtrIsValid(lsn));
2478 88427 : SpinLockAcquire(&slot->mutex);
2479 88427 : if (slot->data.restart_lsn != lsn)
2480 : {
2481 39818 : changed = true;
2482 39818 : slot->data.restart_lsn = lsn;
2483 : }
2484 88427 : SpinLockRelease(&slot->mutex);
2485 :
2486 88427 : if (changed)
2487 : {
2488 39818 : ReplicationSlotMarkDirty();
2489 39818 : ReplicationSlotsComputeRequiredLSN();
2490 39818 : PhysicalWakeupLogicalWalSnd();
2491 : }
2492 :
2493 : /*
2494 : * One could argue that the slot should be saved to disk now, but that'd
2495 : * be energy wasted - the worst thing lost information could cause here is
2496 : * to give wrong information in a statistics view - we'll just potentially
2497 : * be more conservative in removing files.
2498 : */
2499 88427 : }
2500 :
2501 : /*
2502 : * Regular reply from standby advising of WAL locations on standby server.
2503 : */
2504 : static void
2505 111009 : ProcessStandbyReplyMessage(void)
2506 : {
2507 : XLogRecPtr writePtr,
2508 : flushPtr,
2509 : applyPtr;
2510 : bool replyRequested;
2511 : TimeOffset writeLag,
2512 : flushLag,
2513 : applyLag;
2514 : bool clearLagTimes;
2515 : TimestampTz now;
2516 : TimestampTz replyTime;
2517 :
2518 : static XLogRecPtr prevWritePtr = InvalidXLogRecPtr;
2519 : static XLogRecPtr prevFlushPtr = InvalidXLogRecPtr;
2520 : static XLogRecPtr prevApplyPtr = InvalidXLogRecPtr;
2521 :
2522 : /* the caller already consumed the msgtype byte */
2523 111009 : writePtr = pq_getmsgint64(&reply_message);
2524 111009 : flushPtr = pq_getmsgint64(&reply_message);
2525 111009 : applyPtr = pq_getmsgint64(&reply_message);
2526 111009 : replyTime = pq_getmsgint64(&reply_message);
2527 111009 : replyRequested = pq_getmsgbyte(&reply_message);
2528 :
2529 111009 : if (message_level_is_interesting(DEBUG2))
2530 : {
2531 : char *replyTimeStr;
2532 :
2533 : /* Copy because timestamptz_to_str returns a static buffer */
2534 624 : replyTimeStr = pstrdup(timestamptz_to_str(replyTime));
2535 :
2536 624 : elog(DEBUG2, "write %X/%08X flush %X/%08X apply %X/%08X%s reply_time %s",
2537 : LSN_FORMAT_ARGS(writePtr),
2538 : LSN_FORMAT_ARGS(flushPtr),
2539 : LSN_FORMAT_ARGS(applyPtr),
2540 : replyRequested ? " (reply requested)" : "",
2541 : replyTimeStr);
2542 :
2543 624 : pfree(replyTimeStr);
2544 : }
2545 :
2546 : /* See if we can compute the round-trip lag for these positions. */
2547 111009 : now = GetCurrentTimestamp();
2548 111009 : writeLag = LagTrackerRead(SYNC_REP_WAIT_WRITE, writePtr, now);
2549 111009 : flushLag = LagTrackerRead(SYNC_REP_WAIT_FLUSH, flushPtr, now);
2550 111009 : applyLag = LagTrackerRead(SYNC_REP_WAIT_APPLY, applyPtr, now);
2551 :
2552 : /*
2553 : * If the standby reports that it has fully replayed the WAL, and the
2554 : * write/flush/apply positions remain unchanged across two consecutive
2555 : * reply messages, forget the lag times measured when it last
2556 : * wrote/flushed/applied a WAL record.
2557 : *
2558 : * The second message with unchanged positions typically results from
2559 : * wal_receiver_status_interval expiring on the standby, so lag values are
2560 : * usually cleared after that interval when there is no activity. This
2561 : * avoids displaying stale lag data until more WAL traffic arrives.
2562 : */
2563 10794 : clearLagTimes = (applyPtr == sentPtr && flushPtr == sentPtr &&
2564 129711 : writePtr == prevWritePtr && flushPtr == prevFlushPtr &&
2565 7908 : applyPtr == prevApplyPtr);
2566 :
2567 111009 : prevWritePtr = writePtr;
2568 111009 : prevFlushPtr = flushPtr;
2569 111009 : prevApplyPtr = applyPtr;
2570 :
2571 : /* Send a reply if the standby requested one. */
2572 111009 : if (replyRequested)
2573 0 : WalSndKeepalive(false, InvalidXLogRecPtr);
2574 :
2575 : /*
2576 : * Update shared state for this WalSender process based on reply data from
2577 : * standby.
2578 : */
2579 : {
2580 111009 : WalSnd *walsnd = MyWalSnd;
2581 :
2582 111009 : SpinLockAcquire(&walsnd->mutex);
2583 111009 : walsnd->write = writePtr;
2584 111009 : walsnd->flush = flushPtr;
2585 111009 : walsnd->apply = applyPtr;
2586 111009 : if (writeLag != -1 || clearLagTimes)
2587 65761 : walsnd->writeLag = writeLag;
2588 111009 : if (flushLag != -1 || clearLagTimes)
2589 84788 : walsnd->flushLag = flushLag;
2590 111009 : if (applyLag != -1 || clearLagTimes)
2591 91213 : walsnd->applyLag = applyLag;
2592 111009 : walsnd->replyTime = replyTime;
2593 111009 : SpinLockRelease(&walsnd->mutex);
2594 : }
2595 :
2596 111009 : if (!am_cascading_walsender)
2597 110727 : SyncRepReleaseWaiters();
2598 :
2599 : /*
2600 : * Advance our local xmin horizon when the client confirmed a flush.
2601 : */
2602 111009 : if (MyReplicationSlot && XLogRecPtrIsValid(flushPtr))
2603 : {
2604 107712 : if (SlotIsLogical(MyReplicationSlot))
2605 19285 : LogicalConfirmReceivedLocation(flushPtr);
2606 : else
2607 88427 : PhysicalConfirmReceivedLocation(flushPtr);
2608 : }
2609 111009 : }
2610 :
2611 : /* compute new replication slot xmin horizon if needed */
2612 : static void
2613 67 : PhysicalReplicationSlotNewXmin(TransactionId feedbackXmin, TransactionId feedbackCatalogXmin)
2614 : {
2615 67 : bool changed = false;
2616 67 : ReplicationSlot *slot = MyReplicationSlot;
2617 :
2618 67 : SpinLockAcquire(&slot->mutex);
2619 67 : MyProc->xmin = InvalidTransactionId;
2620 :
2621 : /*
2622 : * For physical replication we don't need the interlock provided by xmin
2623 : * and effective_xmin since the consequences of a missed increase are
2624 : * limited to query cancellations, so set both at once.
2625 : */
2626 67 : if (!TransactionIdIsNormal(slot->data.xmin) ||
2627 31 : !TransactionIdIsNormal(feedbackXmin) ||
2628 31 : TransactionIdPrecedes(slot->data.xmin, feedbackXmin))
2629 : {
2630 47 : changed = true;
2631 47 : slot->data.xmin = feedbackXmin;
2632 47 : slot->effective_xmin = feedbackXmin;
2633 : }
2634 67 : if (!TransactionIdIsNormal(slot->data.catalog_xmin) ||
2635 16 : !TransactionIdIsNormal(feedbackCatalogXmin) ||
2636 16 : TransactionIdPrecedes(slot->data.catalog_xmin, feedbackCatalogXmin))
2637 : {
2638 52 : changed = true;
2639 52 : slot->data.catalog_xmin = feedbackCatalogXmin;
2640 52 : slot->effective_catalog_xmin = feedbackCatalogXmin;
2641 : }
2642 67 : SpinLockRelease(&slot->mutex);
2643 :
2644 67 : if (changed)
2645 : {
2646 57 : ReplicationSlotMarkDirty();
2647 57 : ReplicationSlotsComputeRequiredXmin(false);
2648 : }
2649 67 : }
2650 :
2651 : /*
2652 : * Check that the provided xmin/epoch are sane, that is, not in the future
2653 : * and not so far back as to be already wrapped around.
2654 : *
2655 : * Epoch of nextXid should be same as standby, or if the counter has
2656 : * wrapped, then one greater than standby.
2657 : *
2658 : * This check doesn't care about whether clog exists for these xids
2659 : * at all.
2660 : */
2661 : static bool
2662 68 : TransactionIdInRecentPast(TransactionId xid, uint32 epoch)
2663 : {
2664 : FullTransactionId nextFullXid;
2665 : TransactionId nextXid;
2666 : uint32 nextEpoch;
2667 :
2668 68 : nextFullXid = ReadNextFullTransactionId();
2669 68 : nextXid = XidFromFullTransactionId(nextFullXid);
2670 68 : nextEpoch = EpochFromFullTransactionId(nextFullXid);
2671 :
2672 68 : if (xid <= nextXid)
2673 : {
2674 68 : if (epoch != nextEpoch)
2675 0 : return false;
2676 : }
2677 : else
2678 : {
2679 0 : if (epoch + 1 != nextEpoch)
2680 0 : return false;
2681 : }
2682 :
2683 68 : if (!TransactionIdPrecedesOrEquals(xid, nextXid))
2684 0 : return false; /* epoch OK, but it's wrapped around */
2685 :
2686 68 : return true;
2687 : }
2688 :
2689 : /*
2690 : * Hot Standby feedback
2691 : */
2692 : static void
2693 155 : ProcessStandbyHSFeedbackMessage(void)
2694 : {
2695 : TransactionId feedbackXmin;
2696 : uint32 feedbackEpoch;
2697 : TransactionId feedbackCatalogXmin;
2698 : uint32 feedbackCatalogEpoch;
2699 : TimestampTz replyTime;
2700 :
2701 : /*
2702 : * Decipher the reply message. The caller already consumed the msgtype
2703 : * byte. See XLogWalRcvSendHSFeedback() in walreceiver.c for the creation
2704 : * of this message.
2705 : */
2706 155 : replyTime = pq_getmsgint64(&reply_message);
2707 155 : feedbackXmin = pq_getmsgint(&reply_message, 4);
2708 155 : feedbackEpoch = pq_getmsgint(&reply_message, 4);
2709 155 : feedbackCatalogXmin = pq_getmsgint(&reply_message, 4);
2710 155 : feedbackCatalogEpoch = pq_getmsgint(&reply_message, 4);
2711 :
2712 155 : if (message_level_is_interesting(DEBUG2))
2713 : {
2714 : char *replyTimeStr;
2715 :
2716 : /* Copy because timestamptz_to_str returns a static buffer */
2717 4 : replyTimeStr = pstrdup(timestamptz_to_str(replyTime));
2718 :
2719 4 : elog(DEBUG2, "hot standby feedback xmin %u epoch %u, catalog_xmin %u epoch %u reply_time %s",
2720 : feedbackXmin,
2721 : feedbackEpoch,
2722 : feedbackCatalogXmin,
2723 : feedbackCatalogEpoch,
2724 : replyTimeStr);
2725 :
2726 4 : pfree(replyTimeStr);
2727 : }
2728 :
2729 : /*
2730 : * Update shared state for this WalSender process based on reply data from
2731 : * standby.
2732 : */
2733 : {
2734 155 : WalSnd *walsnd = MyWalSnd;
2735 :
2736 155 : SpinLockAcquire(&walsnd->mutex);
2737 155 : walsnd->replyTime = replyTime;
2738 155 : SpinLockRelease(&walsnd->mutex);
2739 : }
2740 :
2741 : /*
2742 : * Unset WalSender's xmins if the feedback message values are invalid.
2743 : * This happens when the downstream turned hot_standby_feedback off.
2744 : */
2745 155 : if (!TransactionIdIsNormal(feedbackXmin)
2746 109 : && !TransactionIdIsNormal(feedbackCatalogXmin))
2747 : {
2748 109 : MyProc->xmin = InvalidTransactionId;
2749 109 : if (MyReplicationSlot != NULL)
2750 25 : PhysicalReplicationSlotNewXmin(feedbackXmin, feedbackCatalogXmin);
2751 109 : return;
2752 : }
2753 :
2754 : /*
2755 : * Check that the provided xmin/epoch are sane, that is, not in the future
2756 : * and not so far back as to be already wrapped around. Ignore if not.
2757 : */
2758 46 : if (TransactionIdIsNormal(feedbackXmin) &&
2759 46 : !TransactionIdInRecentPast(feedbackXmin, feedbackEpoch))
2760 0 : return;
2761 :
2762 46 : if (TransactionIdIsNormal(feedbackCatalogXmin) &&
2763 22 : !TransactionIdInRecentPast(feedbackCatalogXmin, feedbackCatalogEpoch))
2764 0 : return;
2765 :
2766 : /*
2767 : * Set the WalSender's xmin equal to the standby's requested xmin, so that
2768 : * the xmin will be taken into account by GetSnapshotData() /
2769 : * ComputeXidHorizons(). This will hold back the removal of dead rows and
2770 : * thereby prevent the generation of cleanup conflicts on the standby
2771 : * server.
2772 : *
2773 : * There is a small window for a race condition here: although we just
2774 : * checked that feedbackXmin precedes nextXid, the nextXid could have
2775 : * gotten advanced between our fetching it and applying the xmin below,
2776 : * perhaps far enough to make feedbackXmin wrap around. In that case the
2777 : * xmin we set here would be "in the future" and have no effect. No point
2778 : * in worrying about this since it's too late to save the desired data
2779 : * anyway. Assuming that the standby sends us an increasing sequence of
2780 : * xmins, this could only happen during the first reply cycle, else our
2781 : * own xmin would prevent nextXid from advancing so far.
2782 : *
2783 : * We don't bother taking the ProcArrayLock here. Setting the xmin field
2784 : * is assumed atomic, and there's no real need to prevent concurrent
2785 : * horizon determinations. (If we're moving our xmin forward, this is
2786 : * obviously safe, and if we're moving it backwards, well, the data is at
2787 : * risk already since a VACUUM could already have determined the horizon.)
2788 : *
2789 : * If we're using a replication slot we reserve the xmin via that,
2790 : * otherwise via the walsender's PGPROC entry. We can only track the
2791 : * catalog xmin separately when using a slot, so we store the least of the
2792 : * two provided when not using a slot.
2793 : *
2794 : * XXX: It might make sense to generalize the ephemeral slot concept and
2795 : * always use the slot mechanism to handle the feedback xmin.
2796 : */
2797 46 : if (MyReplicationSlot != NULL) /* XXX: persistency configurable? */
2798 42 : PhysicalReplicationSlotNewXmin(feedbackXmin, feedbackCatalogXmin);
2799 : else
2800 : {
2801 4 : if (TransactionIdIsNormal(feedbackCatalogXmin)
2802 0 : && TransactionIdPrecedes(feedbackCatalogXmin, feedbackXmin))
2803 0 : MyProc->xmin = feedbackCatalogXmin;
2804 : else
2805 4 : MyProc->xmin = feedbackXmin;
2806 : }
2807 : }
2808 :
2809 : /*
2810 : * Process the request for a primary status update message.
2811 : */
2812 : static void
2813 6818 : ProcessStandbyPSRequestMessage(void)
2814 : {
2815 6818 : XLogRecPtr lsn = InvalidXLogRecPtr;
2816 : TransactionId oldestXidInCommit;
2817 : TransactionId oldestGXidInCommit;
2818 : FullTransactionId nextFullXid;
2819 : FullTransactionId fullOldestXidInCommit;
2820 6818 : WalSnd *walsnd = MyWalSnd;
2821 : TimestampTz replyTime;
2822 :
2823 : /*
2824 : * This shouldn't happen because we don't support getting primary status
2825 : * message from standby.
2826 : */
2827 6818 : if (RecoveryInProgress())
2828 0 : elog(ERROR, "the primary status is unavailable during recovery");
2829 :
2830 6818 : replyTime = pq_getmsgint64(&reply_message);
2831 :
2832 : /*
2833 : * Update shared state for this WalSender process based on reply data from
2834 : * standby.
2835 : */
2836 6818 : SpinLockAcquire(&walsnd->mutex);
2837 6818 : walsnd->replyTime = replyTime;
2838 6818 : SpinLockRelease(&walsnd->mutex);
2839 :
2840 : /*
2841 : * Consider transactions in the current database, as only these are the
2842 : * ones replicated.
2843 : */
2844 6818 : oldestXidInCommit = GetOldestActiveTransactionId(true, false);
2845 6818 : oldestGXidInCommit = TwoPhaseGetOldestXidInCommit();
2846 :
2847 : /*
2848 : * Update the oldest xid for standby transmission if an older prepared
2849 : * transaction exists and is currently in commit phase.
2850 : */
2851 13279 : if (TransactionIdIsValid(oldestGXidInCommit) &&
2852 6461 : TransactionIdPrecedes(oldestGXidInCommit, oldestXidInCommit))
2853 6461 : oldestXidInCommit = oldestGXidInCommit;
2854 :
2855 6818 : nextFullXid = ReadNextFullTransactionId();
2856 6818 : fullOldestXidInCommit = FullTransactionIdFromAllowableAt(nextFullXid,
2857 : oldestXidInCommit);
2858 6818 : lsn = GetXLogWriteRecPtr();
2859 :
2860 6818 : elog(DEBUG2, "sending primary status");
2861 :
2862 : /* construct the message... */
2863 6818 : resetStringInfo(&output_message);
2864 6818 : pq_sendbyte(&output_message, PqReplMsg_PrimaryStatusUpdate);
2865 6818 : pq_sendint64(&output_message, lsn);
2866 6818 : pq_sendint64(&output_message, (int64) U64FromFullTransactionId(fullOldestXidInCommit));
2867 6818 : pq_sendint64(&output_message, (int64) U64FromFullTransactionId(nextFullXid));
2868 6818 : pq_sendint64(&output_message, GetCurrentTimestamp());
2869 :
2870 : /* ... and send it wrapped in CopyData */
2871 6818 : pq_putmessage_noblock(PqMsg_CopyData, output_message.data, output_message.len);
2872 6818 : }
2873 :
2874 : /*
2875 : * Compute how long send/receive loops should sleep.
2876 : *
2877 : * If wal_sender_timeout is enabled we want to wake up in time to send
2878 : * keepalives and to abort the connection if wal_sender_timeout has been
2879 : * reached.
2880 : *
2881 : * If wal_sender_shutdown_timeout is enabled, during shutdown, we want to
2882 : * wake up in time to exit when it expires.
2883 : */
2884 : static long
2885 100663 : WalSndComputeSleeptime(TimestampTz now)
2886 : {
2887 : TimestampTz wakeup_time;
2888 100663 : long sleeptime = 10000; /* 10 s */
2889 :
2890 100663 : if (wal_sender_timeout > 0 && last_reply_timestamp > 0)
2891 : {
2892 : /*
2893 : * At the latest stop sleeping once wal_sender_timeout has been
2894 : * reached.
2895 : */
2896 100593 : wakeup_time = TimestampTzPlusMilliseconds(last_reply_timestamp,
2897 : wal_sender_timeout);
2898 :
2899 : /*
2900 : * If no ping has been sent yet, wakeup when it's time to do so.
2901 : * WalSndKeepaliveIfNecessary() wants to send a keepalive once half of
2902 : * the timeout passed without a response.
2903 : */
2904 100593 : if (!waiting_for_ping_response)
2905 98379 : wakeup_time = TimestampTzPlusMilliseconds(last_reply_timestamp,
2906 : wal_sender_timeout / 2);
2907 :
2908 : /* Compute relative time until wakeup. */
2909 100593 : sleeptime = TimestampDifferenceMilliseconds(now, wakeup_time);
2910 : }
2911 :
2912 100663 : if (shutdown_request_timestamp != 0 && wal_sender_shutdown_timeout > 0)
2913 : {
2914 : long shutdown_sleeptime;
2915 :
2916 4 : wakeup_time = TimestampTzPlusMilliseconds(shutdown_request_timestamp,
2917 : wal_sender_shutdown_timeout);
2918 :
2919 4 : shutdown_sleeptime = TimestampDifferenceMilliseconds(now, wakeup_time);
2920 :
2921 : /* Choose the earliest wakeup. */
2922 4 : if (shutdown_sleeptime < sleeptime)
2923 4 : sleeptime = shutdown_sleeptime;
2924 : }
2925 :
2926 100663 : return sleeptime;
2927 : }
2928 :
2929 : /*
2930 : * Check whether there have been responses by the client within
2931 : * wal_sender_timeout and shutdown if not. Using last_processing as the
2932 : * reference point avoids counting server-side stalls against the client.
2933 : * However, a long server-side stall can make WalSndKeepaliveIfNecessary()
2934 : * postdate last_processing by more than wal_sender_timeout. If that happens,
2935 : * the client must reply almost immediately to avoid a timeout. This rarely
2936 : * affects the default configuration, under which clients spontaneously send a
2937 : * message every standby_message_timeout = wal_sender_timeout/6 = 10s. We
2938 : * could eliminate that problem by recognizing timeout expiration at
2939 : * wal_sender_timeout/2 after the keepalive.
2940 : */
2941 : static void
2942 1161638 : WalSndCheckTimeOut(void)
2943 : {
2944 : TimestampTz timeout;
2945 :
2946 : /* don't bail out if we're doing something that doesn't require timeouts */
2947 1161638 : if (last_reply_timestamp <= 0)
2948 27 : return;
2949 :
2950 1161611 : timeout = TimestampTzPlusMilliseconds(last_reply_timestamp,
2951 : wal_sender_timeout);
2952 :
2953 1161611 : if (wal_sender_timeout > 0 && last_processing >= timeout)
2954 : {
2955 : /*
2956 : * Since typically expiration of replication timeout means
2957 : * communication problem, we don't send the error message to the
2958 : * standby.
2959 : */
2960 0 : ereport(COMMERROR,
2961 : (errmsg("terminating walsender process due to replication timeout")));
2962 :
2963 0 : WalSndShutdown();
2964 : }
2965 : }
2966 :
2967 : /*
2968 : * Check whether the walsender process should terminate due to the expiration
2969 : * of wal_sender_shutdown_timeout after the receipt of a shutdown request.
2970 : */
2971 : static void
2972 1161724 : WalSndCheckShutdownTimeout(void)
2973 : {
2974 : TimestampTz now;
2975 :
2976 : /* Do nothing if shutdown has not been requested yet */
2977 1161724 : if (!(got_STOPPING || got_SIGUSR2))
2978 1152168 : return;
2979 :
2980 : /* Terminate immediately if the timeout is set to 0 */
2981 9556 : if (wal_sender_shutdown_timeout == 0)
2982 0 : WalSndDoneImmediate();
2983 :
2984 : /*
2985 : * Record the shutdown request timestamp even if
2986 : * wal_sender_shutdown_timeout is disabled (-1), since the setting may
2987 : * change during shutdown and the timestamp will be needed in that case.
2988 : */
2989 9556 : if (shutdown_request_timestamp == 0)
2990 : {
2991 47 : shutdown_request_timestamp = GetCurrentTimestamp();
2992 47 : return;
2993 : }
2994 :
2995 : /* Do not check the timeout if it's disabled */
2996 9509 : if (wal_sender_shutdown_timeout == -1)
2997 9115 : return;
2998 :
2999 : /* Terminate immediately if the timeout expires */
3000 394 : now = GetCurrentTimestamp();
3001 394 : if (TimestampDifferenceExceeds(shutdown_request_timestamp, now,
3002 : wal_sender_shutdown_timeout))
3003 3 : WalSndDoneImmediate();
3004 : }
3005 :
3006 : /* Main loop of walsender process that streams the WAL over Copy messages. */
3007 : static void
3008 767 : WalSndLoop(WalSndSendDataCallback send_data)
3009 : {
3010 767 : TimestampTz last_flush = 0;
3011 :
3012 : /*
3013 : * Initialize the last reply timestamp. That enables timeout processing
3014 : * from hereon.
3015 : */
3016 767 : last_reply_timestamp = GetCurrentTimestamp();
3017 767 : waiting_for_ping_response = false;
3018 :
3019 : /*
3020 : * Loop until we reach the end of this timeline or the client requests to
3021 : * stop streaming.
3022 : */
3023 : for (;;)
3024 : {
3025 : /* Clear any already-pending wakeups */
3026 1151126 : ResetLatch(MyLatch);
3027 :
3028 1151126 : CHECK_FOR_INTERRUPTS();
3029 :
3030 : /* Process any requests or signals received recently */
3031 1151123 : WalSndHandleConfigReload();
3032 :
3033 : /* Check for input from the client */
3034 1151123 : ProcessRepliesIfAny();
3035 :
3036 : /*
3037 : * If we have received CopyDone from the client, sent CopyDone
3038 : * ourselves, and the output buffer is empty, it's time to exit
3039 : * streaming.
3040 : */
3041 1151015 : if (streamingDoneReceiving && streamingDoneSending &&
3042 582 : !pq_is_send_pending())
3043 371 : break;
3044 :
3045 : /*
3046 : * If we don't have any pending data in the output buffer, try to send
3047 : * some more. If there is some, we don't bother to call send_data
3048 : * again until we've flushed it ... but we'd better assume we are not
3049 : * caught up.
3050 : */
3051 1150644 : if (!pq_is_send_pending())
3052 1109063 : send_data();
3053 : else
3054 41581 : WalSndCaughtUp = false;
3055 :
3056 : /* Try to flush pending output to the client */
3057 1150404 : if (pq_flush_if_writable() != 0)
3058 0 : WalSndShutdown();
3059 :
3060 : /* If nothing remains to be sent right now ... */
3061 1150404 : if (WalSndCaughtUp && !pq_is_send_pending())
3062 : {
3063 : /*
3064 : * If we're in catchup state, move to streaming. This is an
3065 : * important state change for users to know about, since before
3066 : * this point data loss might occur if the primary dies and we
3067 : * need to failover to the standby. The state change is also
3068 : * important for synchronous replication, since commits that
3069 : * started to wait at that point might wait for some time.
3070 : */
3071 93096 : if (MyWalSnd->state == WALSNDSTATE_CATCHUP)
3072 : {
3073 711 : ereport(DEBUG1,
3074 : (errmsg_internal("\"%s\" has now caught up with upstream server",
3075 : application_name)));
3076 711 : WalSndSetState(WALSNDSTATE_STREAMING);
3077 : }
3078 :
3079 : /*
3080 : * When SIGUSR2 arrives, we send any outstanding logs up to the
3081 : * shutdown checkpoint record (i.e., the latest record), wait for
3082 : * them to be replicated to the standby, and exit. This may be a
3083 : * normal termination at shutdown, or a promotion, the walsender
3084 : * is not sure which.
3085 : */
3086 93096 : if (got_SIGUSR2)
3087 7354 : WalSndDone(send_data);
3088 : }
3089 :
3090 : /* Check for replication timeout. */
3091 1150361 : WalSndCheckTimeOut();
3092 :
3093 : /*
3094 : * During shutdown, die if the shutdown timeout expires. Call this
3095 : * before WalSndComputeSleeptime() so the timeout is considered when
3096 : * computing sleep time.
3097 : */
3098 1150361 : WalSndCheckShutdownTimeout();
3099 :
3100 : /* Send keepalive if the time has come */
3101 1150359 : WalSndKeepaliveIfNecessary();
3102 :
3103 : /*
3104 : * Block if we have unsent data. XXX For logical replication, let
3105 : * WalSndWaitForWal() handle any other blocking; idle receivers need
3106 : * its additional actions. For physical replication, also block if
3107 : * caught up; its send_data does not block.
3108 : *
3109 : * The IO statistics are reported in WalSndWaitForWal() for the
3110 : * logical WAL senders.
3111 : */
3112 1150359 : if ((WalSndCaughtUp && send_data != XLogSendLogical &&
3113 1185587 : !streamingDoneSending) ||
3114 1099785 : pq_is_send_pending())
3115 : {
3116 : long sleeptime;
3117 : int wakeEvents;
3118 : TimestampTz now;
3119 :
3120 89906 : if (!streamingDoneReceiving)
3121 89885 : wakeEvents = WL_SOCKET_READABLE;
3122 : else
3123 21 : wakeEvents = 0;
3124 :
3125 : /*
3126 : * Use fresh timestamp, not last_processing, to reduce the chance
3127 : * of reaching wal_sender_timeout before sending a keepalive.
3128 : */
3129 89906 : now = GetCurrentTimestamp();
3130 89906 : sleeptime = WalSndComputeSleeptime(now);
3131 :
3132 89906 : if (pq_is_send_pending())
3133 41522 : wakeEvents |= WL_SOCKET_WRITEABLE;
3134 :
3135 : /* Report IO statistics, if needed */
3136 89906 : if (TimestampDifferenceExceeds(last_flush, now,
3137 : WALSENDER_STATS_FLUSH_INTERVAL))
3138 : {
3139 598 : pgstat_flush_io(false);
3140 598 : (void) pgstat_flush_backend(false, PGSTAT_BACKEND_FLUSH_IO);
3141 598 : last_flush = now;
3142 : }
3143 :
3144 : /* Sleep until something happens or we time out */
3145 89906 : WalSndWait(wakeEvents, sleeptime, WAIT_EVENT_WAL_SENDER_MAIN);
3146 : }
3147 : }
3148 371 : }
3149 :
3150 : /* Initialize a per-walsender data structure for this walsender process */
3151 : static void
3152 1301 : InitWalSenderSlot(void)
3153 : {
3154 : int i;
3155 :
3156 : /*
3157 : * WalSndCtl should be set up already (we inherit this by fork() or
3158 : * EXEC_BACKEND mechanism from the postmaster).
3159 : */
3160 : Assert(WalSndCtl != NULL);
3161 : Assert(MyWalSnd == NULL);
3162 :
3163 : /*
3164 : * Find a free walsender slot and reserve it. This must not fail due to
3165 : * the prior check for free WAL senders in InitProcess().
3166 : */
3167 1921 : for (i = 0; i < max_wal_senders; i++)
3168 : {
3169 1921 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3170 :
3171 1921 : SpinLockAcquire(&walsnd->mutex);
3172 :
3173 1921 : if (walsnd->pid != 0)
3174 : {
3175 620 : SpinLockRelease(&walsnd->mutex);
3176 620 : continue;
3177 : }
3178 : else
3179 : {
3180 : /*
3181 : * Found a free slot. Reserve it for us.
3182 : */
3183 1301 : walsnd->pid = MyProcPid;
3184 1301 : walsnd->state = WALSNDSTATE_STARTUP;
3185 1301 : walsnd->sentPtr = InvalidXLogRecPtr;
3186 1301 : walsnd->needreload = false;
3187 1301 : walsnd->write = InvalidXLogRecPtr;
3188 1301 : walsnd->flush = InvalidXLogRecPtr;
3189 1301 : walsnd->apply = InvalidXLogRecPtr;
3190 1301 : walsnd->writeLag = -1;
3191 1301 : walsnd->flushLag = -1;
3192 1301 : walsnd->applyLag = -1;
3193 1301 : walsnd->sync_standby_priority = 0;
3194 1301 : walsnd->replyTime = 0;
3195 :
3196 : /*
3197 : * The kind assignment is done here and not in StartReplication()
3198 : * and StartLogicalReplication(). Indeed, the logical walsender
3199 : * needs to read WAL records (like snapshot of running
3200 : * transactions) during the slot creation. So it needs to be woken
3201 : * up based on its kind.
3202 : *
3203 : * The kind assignment could also be done in StartReplication(),
3204 : * StartLogicalReplication() and CREATE_REPLICATION_SLOT but it
3205 : * seems better to set it on one place.
3206 : */
3207 1301 : if (MyDatabaseId == InvalidOid)
3208 503 : walsnd->kind = REPLICATION_KIND_PHYSICAL;
3209 : else
3210 798 : walsnd->kind = REPLICATION_KIND_LOGICAL;
3211 :
3212 1301 : SpinLockRelease(&walsnd->mutex);
3213 : /* don't need the lock anymore */
3214 1301 : MyWalSnd = walsnd;
3215 :
3216 1301 : break;
3217 : }
3218 : }
3219 :
3220 : Assert(MyWalSnd != NULL);
3221 :
3222 : /* Arrange to clean up at walsender exit */
3223 1301 : on_shmem_exit(WalSndKill, 0);
3224 1301 : }
3225 :
3226 : /* Destroy the per-walsender data structure for this walsender process */
3227 : static void
3228 1301 : WalSndKill(int code, Datum arg)
3229 : {
3230 1301 : WalSnd *walsnd = MyWalSnd;
3231 :
3232 : Assert(walsnd != NULL);
3233 :
3234 1301 : MyWalSnd = NULL;
3235 :
3236 1301 : SpinLockAcquire(&walsnd->mutex);
3237 : /* Mark WalSnd struct as no longer being in use. */
3238 1301 : walsnd->pid = 0;
3239 1301 : SpinLockRelease(&walsnd->mutex);
3240 1301 : }
3241 :
3242 : /* XLogReaderRoutine->segment_open callback */
3243 : static void
3244 14836 : WalSndSegmentOpen(XLogReaderState *state, XLogSegNo nextSegNo,
3245 : TimeLineID *tli_p)
3246 : {
3247 : char path[MAXPGPATH];
3248 :
3249 : /*-------
3250 : * When reading from a historic timeline, and there is a timeline switch
3251 : * within this segment, read from the WAL segment belonging to the new
3252 : * timeline.
3253 : *
3254 : * For example, imagine that this server is currently on timeline 5, and
3255 : * we're streaming timeline 4. The switch from timeline 4 to 5 happened at
3256 : * 0/13002088. In pg_wal, we have these files:
3257 : *
3258 : * ...
3259 : * 000000040000000000000012
3260 : * 000000040000000000000013
3261 : * 000000050000000000000013
3262 : * 000000050000000000000014
3263 : * ...
3264 : *
3265 : * In this situation, when requested to send the WAL from segment 0x13, on
3266 : * timeline 4, we read the WAL from file 000000050000000000000013. Archive
3267 : * recovery prefers files from newer timelines, so if the segment was
3268 : * restored from the archive on this server, the file belonging to the old
3269 : * timeline, 000000040000000000000013, might not exist. Their contents are
3270 : * equal up to the switchpoint, because at a timeline switch, the used
3271 : * portion of the old segment is copied to the new file.
3272 : */
3273 14836 : *tli_p = sendTimeLine;
3274 14836 : if (sendTimeLineIsHistoric)
3275 : {
3276 : XLogSegNo endSegNo;
3277 :
3278 586 : XLByteToSeg(sendTimeLineValidUpto, endSegNo, state->segcxt.ws_segsize);
3279 586 : if (nextSegNo == endSegNo)
3280 8 : *tli_p = sendTimeLineNextTLI;
3281 : }
3282 :
3283 14836 : XLogFilePath(path, *tli_p, nextSegNo, state->segcxt.ws_segsize);
3284 14836 : state->seg.ws_file = BasicOpenFile(path, O_RDONLY | PG_BINARY);
3285 14836 : if (state->seg.ws_file >= 0)
3286 14835 : return;
3287 :
3288 : /*
3289 : * If the file is not found, assume it's because the standby asked for a
3290 : * too old WAL segment that has already been removed or recycled.
3291 : */
3292 1 : if (errno == ENOENT)
3293 : {
3294 : char xlogfname[MAXFNAMELEN];
3295 1 : int save_errno = errno;
3296 :
3297 1 : XLogFileName(xlogfname, *tli_p, nextSegNo, wal_segment_size);
3298 1 : errno = save_errno;
3299 1 : ereport(ERROR,
3300 : (errcode_for_file_access(),
3301 : errmsg("requested WAL segment %s has already been removed",
3302 : xlogfname)));
3303 : }
3304 : else
3305 0 : ereport(ERROR,
3306 : (errcode_for_file_access(),
3307 : errmsg("could not open file \"%s\": %m",
3308 : path)));
3309 : }
3310 :
3311 : /*
3312 : * Send out the WAL in its normal physical/stored form.
3313 : *
3314 : * Read up to MAX_SEND_SIZE bytes of WAL that's been flushed to disk,
3315 : * but not yet sent to the client, and buffer it in the libpq output
3316 : * buffer.
3317 : *
3318 : * If there is no unsent WAL remaining, WalSndCaughtUp is set to true,
3319 : * otherwise WalSndCaughtUp is set to false.
3320 : */
3321 : static void
3322 174936 : XLogSendPhysical(void)
3323 : {
3324 : XLogRecPtr SendRqstPtr;
3325 : XLogRecPtr startptr;
3326 : XLogRecPtr endptr;
3327 : Size nbytes;
3328 : XLogSegNo segno;
3329 : WALReadError errinfo;
3330 : Size rbytes;
3331 :
3332 : /* If requested switch the WAL sender to the stopping state. */
3333 174936 : if (got_STOPPING)
3334 2580 : WalSndSetState(WALSNDSTATE_STOPPING);
3335 :
3336 174936 : if (streamingDoneSending)
3337 : {
3338 35217 : WalSndCaughtUp = true;
3339 67764 : return;
3340 : }
3341 :
3342 : /* Figure out how far we can safely send the WAL. */
3343 139719 : if (sendTimeLineIsHistoric)
3344 : {
3345 : /*
3346 : * Streaming an old timeline that's in this server's history, but is
3347 : * not the one we're currently inserting or replaying. It can be
3348 : * streamed up to the point where we switched off that timeline.
3349 : */
3350 32 : SendRqstPtr = sendTimeLineValidUpto;
3351 : }
3352 139687 : else if (am_cascading_walsender)
3353 : {
3354 : TimeLineID SendRqstTLI;
3355 :
3356 : /*
3357 : * Streaming the latest timeline on a standby.
3358 : *
3359 : * Attempt to send all WAL that has already been replayed, so that we
3360 : * know it's valid. If we're receiving WAL through streaming
3361 : * replication, it's also OK to send any WAL that has been received
3362 : * but not replayed.
3363 : *
3364 : * The timeline we're recovering from can change, or we can be
3365 : * promoted. In either case, the current timeline becomes historic. We
3366 : * need to detect that so that we don't try to stream past the point
3367 : * where we switched to another timeline. We check for promotion or
3368 : * timeline switch after calculating FlushPtr, to avoid a race
3369 : * condition: if the timeline becomes historic just after we checked
3370 : * that it was still current, it's still be OK to stream it up to the
3371 : * FlushPtr that was calculated before it became historic.
3372 : */
3373 971 : bool becameHistoric = false;
3374 :
3375 971 : SendRqstPtr = GetStandbyFlushRecPtr(&SendRqstTLI);
3376 :
3377 971 : if (!RecoveryInProgress())
3378 : {
3379 : /* We have been promoted. */
3380 2 : SendRqstTLI = GetWALInsertionTimeLine();
3381 2 : am_cascading_walsender = false;
3382 2 : becameHistoric = true;
3383 : }
3384 : else
3385 : {
3386 : /*
3387 : * Still a cascading standby. But is the timeline we're sending
3388 : * still the one recovery is recovering from?
3389 : */
3390 969 : if (sendTimeLine != SendRqstTLI)
3391 0 : becameHistoric = true;
3392 : }
3393 :
3394 971 : if (becameHistoric)
3395 : {
3396 : /*
3397 : * The timeline we were sending has become historic. Read the
3398 : * timeline history file of the new timeline to see where exactly
3399 : * we forked off from the timeline we were sending.
3400 : */
3401 : List *history;
3402 :
3403 2 : history = readTimeLineHistory(SendRqstTLI);
3404 2 : sendTimeLineValidUpto = tliSwitchPoint(sendTimeLine, history, &sendTimeLineNextTLI);
3405 :
3406 : Assert(sendTimeLine < sendTimeLineNextTLI);
3407 2 : list_free_deep(history);
3408 :
3409 2 : sendTimeLineIsHistoric = true;
3410 :
3411 2 : SendRqstPtr = sendTimeLineValidUpto;
3412 : }
3413 : }
3414 : else
3415 : {
3416 : /*
3417 : * Streaming the current timeline on a primary.
3418 : *
3419 : * Attempt to send all data that's already been written out and
3420 : * fsync'd to disk. We cannot go further than what's been written out
3421 : * given the current implementation of WALRead(). And in any case
3422 : * it's unsafe to send WAL that is not securely down to disk on the
3423 : * primary: if the primary subsequently crashes and restarts, standbys
3424 : * must not have applied any WAL that got lost on the primary.
3425 : */
3426 138716 : SendRqstPtr = GetFlushRecPtr(NULL);
3427 : }
3428 :
3429 : /*
3430 : * Record the current system time as an approximation of the time at which
3431 : * this WAL location was written for the purposes of lag tracking.
3432 : *
3433 : * In theory we could make XLogFlush() record a time in shmem whenever WAL
3434 : * is flushed and we could get that time as well as the LSN when we call
3435 : * GetFlushRecPtr() above (and likewise for the cascading standby
3436 : * equivalent), but rather than putting any new code into the hot WAL path
3437 : * it seems good enough to capture the time here. We should reach this
3438 : * after XLogFlush() runs WalSndWakeupProcessRequests(), and although that
3439 : * may take some time, we read the WAL flush pointer and take the time
3440 : * very close to together here so that we'll get a later position if it is
3441 : * still moving.
3442 : *
3443 : * Because LagTrackerWrite ignores samples when the LSN hasn't advanced,
3444 : * this gives us a cheap approximation for the WAL flush time for this
3445 : * LSN.
3446 : *
3447 : * Note that the LSN is not necessarily the LSN for the data contained in
3448 : * the present message; it's the end of the WAL, which might be further
3449 : * ahead. All the lag tracking machinery cares about is finding out when
3450 : * that arbitrary LSN is eventually reported as written, flushed and
3451 : * applied, so that it can measure the elapsed time.
3452 : */
3453 139719 : LagTrackerWrite(SendRqstPtr, GetCurrentTimestamp());
3454 :
3455 : /*
3456 : * If this is a historic timeline and we've reached the point where we
3457 : * forked to the next timeline, stop streaming.
3458 : *
3459 : * Note: We might already have sent WAL > sendTimeLineValidUpto. The
3460 : * startup process will normally replay all WAL that has been received
3461 : * from the primary, before promoting, but if the WAL streaming is
3462 : * terminated at a WAL page boundary, the valid portion of the timeline
3463 : * might end in the middle of a WAL record. We might've already sent the
3464 : * first half of that partial WAL record to the cascading standby, so that
3465 : * sentPtr > sendTimeLineValidUpto. That's OK; the cascading standby can't
3466 : * replay the partial WAL record either, so it can still follow our
3467 : * timeline switch.
3468 : */
3469 139719 : if (sendTimeLineIsHistoric && sendTimeLineValidUpto <= sentPtr)
3470 : {
3471 : /* close the current file. */
3472 11 : if (xlogreader->seg.ws_file >= 0)
3473 11 : wal_segment_close(xlogreader);
3474 :
3475 : /* Send CopyDone */
3476 11 : pq_putmessage_noblock(PqMsg_CopyDone, NULL, 0);
3477 11 : streamingDoneSending = true;
3478 :
3479 11 : WalSndCaughtUp = true;
3480 :
3481 11 : elog(DEBUG1, "walsender reached end of timeline at %X/%08X (sent up to %X/%08X)",
3482 : LSN_FORMAT_ARGS(sendTimeLineValidUpto),
3483 : LSN_FORMAT_ARGS(sentPtr));
3484 11 : return;
3485 : }
3486 :
3487 : /* Do we have any work to do? */
3488 : Assert(sentPtr <= SendRqstPtr);
3489 139708 : if (SendRqstPtr <= sentPtr)
3490 : {
3491 32536 : WalSndCaughtUp = true;
3492 32536 : return;
3493 : }
3494 :
3495 : /*
3496 : * Figure out how much to send in one message. If there's no more than
3497 : * MAX_SEND_SIZE bytes to send, send everything. Otherwise send
3498 : * MAX_SEND_SIZE bytes, but round back to logfile or page boundary.
3499 : *
3500 : * The rounding is not only for performance reasons. Walreceiver relies on
3501 : * the fact that we never split a WAL record across two messages. Since a
3502 : * long WAL record is split at page boundary into continuation records,
3503 : * page boundary is always a safe cut-off point. We also assume that
3504 : * SendRqstPtr never points to the middle of a WAL record.
3505 : */
3506 107172 : startptr = sentPtr;
3507 107172 : endptr = startptr;
3508 107172 : endptr += MAX_SEND_SIZE;
3509 :
3510 : /* if we went beyond SendRqstPtr, back off */
3511 107172 : if (SendRqstPtr <= endptr)
3512 : {
3513 19207 : endptr = SendRqstPtr;
3514 19207 : if (sendTimeLineIsHistoric)
3515 9 : WalSndCaughtUp = false;
3516 : else
3517 19198 : WalSndCaughtUp = true;
3518 : }
3519 : else
3520 : {
3521 : /* round down to page boundary. */
3522 87965 : endptr -= (endptr % XLOG_BLCKSZ);
3523 87965 : WalSndCaughtUp = false;
3524 : }
3525 :
3526 107172 : nbytes = endptr - startptr;
3527 : Assert(nbytes <= MAX_SEND_SIZE);
3528 :
3529 : /*
3530 : * OK to read and send the slice.
3531 : */
3532 107172 : resetStringInfo(&output_message);
3533 107172 : pq_sendbyte(&output_message, PqReplMsg_WALData);
3534 :
3535 107172 : pq_sendint64(&output_message, startptr); /* dataStart */
3536 107172 : pq_sendint64(&output_message, SendRqstPtr); /* walEnd */
3537 107172 : pq_sendint64(&output_message, 0); /* sendtime, filled in last */
3538 :
3539 : /*
3540 : * Read the log directly into the output buffer to avoid extra memcpy
3541 : * calls.
3542 : */
3543 107172 : enlargeStringInfo(&output_message, nbytes);
3544 :
3545 107172 : retry:
3546 : /* attempt to read WAL from WAL buffers first */
3547 107172 : rbytes = WALReadFromBuffers(&output_message.data[output_message.len],
3548 107172 : startptr, nbytes, xlogreader->seg.ws_tli);
3549 107172 : output_message.len += rbytes;
3550 107172 : startptr += rbytes;
3551 107172 : nbytes -= rbytes;
3552 :
3553 : /* now read the remaining WAL from WAL file */
3554 107172 : if (nbytes > 0 &&
3555 98290 : !WALRead(xlogreader,
3556 98291 : &output_message.data[output_message.len],
3557 : startptr,
3558 : nbytes,
3559 98291 : xlogreader->seg.ws_tli, /* Pass the current TLI because
3560 : * only WalSndSegmentOpen controls
3561 : * whether new TLI is needed. */
3562 : &errinfo))
3563 0 : WALReadRaiseError(&errinfo);
3564 :
3565 : /* See logical_read_xlog_page(). */
3566 107171 : XLByteToSeg(startptr, segno, xlogreader->segcxt.ws_segsize);
3567 107171 : CheckXLogRemoved(segno, xlogreader->seg.ws_tli);
3568 :
3569 : /*
3570 : * During recovery, the currently-open WAL file might be replaced with the
3571 : * file of the same name retrieved from archive. So we always need to
3572 : * check what we read was valid after reading into the buffer. If it's
3573 : * invalid, we try to open and read the file again.
3574 : */
3575 107171 : if (am_cascading_walsender)
3576 : {
3577 775 : WalSnd *walsnd = MyWalSnd;
3578 : bool reload;
3579 :
3580 775 : SpinLockAcquire(&walsnd->mutex);
3581 775 : reload = walsnd->needreload;
3582 775 : walsnd->needreload = false;
3583 775 : SpinLockRelease(&walsnd->mutex);
3584 :
3585 775 : if (reload && xlogreader->seg.ws_file >= 0)
3586 : {
3587 0 : wal_segment_close(xlogreader);
3588 :
3589 0 : goto retry;
3590 : }
3591 : }
3592 :
3593 107171 : output_message.len += nbytes;
3594 107171 : output_message.data[output_message.len] = '\0';
3595 :
3596 : /*
3597 : * Fill the send timestamp last, so that it is taken as late as possible.
3598 : */
3599 107171 : resetStringInfo(&tmpbuf);
3600 107171 : pq_sendint64(&tmpbuf, GetCurrentTimestamp());
3601 107171 : memcpy(&output_message.data[1 + sizeof(int64) + sizeof(int64)],
3602 107171 : tmpbuf.data, sizeof(int64));
3603 :
3604 107171 : pq_putmessage_noblock(PqMsg_CopyData, output_message.data, output_message.len);
3605 :
3606 107171 : sentPtr = endptr;
3607 :
3608 : /* Update shared memory status */
3609 : {
3610 107171 : WalSnd *walsnd = MyWalSnd;
3611 :
3612 107171 : SpinLockAcquire(&walsnd->mutex);
3613 107171 : walsnd->sentPtr = sentPtr;
3614 107171 : SpinLockRelease(&walsnd->mutex);
3615 : }
3616 :
3617 : /* Report progress of XLOG streaming in PS display */
3618 107171 : if (update_process_title)
3619 : {
3620 : char activitymsg[50];
3621 :
3622 107171 : snprintf(activitymsg, sizeof(activitymsg), "streaming %X/%08X",
3623 107171 : LSN_FORMAT_ARGS(sentPtr));
3624 107171 : set_ps_display(activitymsg);
3625 : }
3626 : }
3627 :
3628 : /*
3629 : * Stream out logically decoded data.
3630 : */
3631 : static void
3632 941481 : XLogSendLogical(void)
3633 : {
3634 : XLogRecord *record;
3635 : char *errm;
3636 :
3637 : /*
3638 : * We'll use the current flush point to determine whether we've caught up.
3639 : * This variable is static in order to cache it across calls. Caching is
3640 : * helpful because GetFlushRecPtr() needs to acquire a heavily-contended
3641 : * spinlock.
3642 : */
3643 : static XLogRecPtr flushPtr = InvalidXLogRecPtr;
3644 :
3645 : /*
3646 : * Don't know whether we've caught up yet. We'll set WalSndCaughtUp to
3647 : * true in WalSndWaitForWal, if we're actually waiting. We also set to
3648 : * true if XLogReadRecord() had to stop reading but WalSndWaitForWal
3649 : * didn't wait - i.e. when we're shutting down.
3650 : */
3651 941481 : WalSndCaughtUp = false;
3652 :
3653 941481 : record = XLogReadRecord(logical_decoding_ctx->reader, &errm);
3654 :
3655 : /* xlog record was invalid */
3656 941259 : if (errm != NULL)
3657 0 : elog(ERROR, "could not find record while sending logically-decoded data: %s",
3658 : errm);
3659 :
3660 941259 : if (record != NULL)
3661 : {
3662 : /*
3663 : * Note the lack of any call to LagTrackerWrite() which is handled by
3664 : * WalSndUpdateProgress which is called by output plugin through
3665 : * logical decoding write api.
3666 : */
3667 928638 : LogicalDecodingProcessRecord(logical_decoding_ctx, logical_decoding_ctx->reader);
3668 :
3669 928621 : sentPtr = logical_decoding_ctx->reader->EndRecPtr;
3670 : }
3671 :
3672 : /*
3673 : * If first time through in this session, initialize flushPtr. Otherwise,
3674 : * we only need to update flushPtr if EndRecPtr is past it.
3675 : */
3676 941242 : if (!XLogRecPtrIsValid(flushPtr) ||
3677 940811 : logical_decoding_ctx->reader->EndRecPtr >= flushPtr)
3678 : {
3679 : /*
3680 : * For cascading logical WAL senders, we use the replay LSN instead of
3681 : * the flush LSN, since logical decoding on a standby only processes
3682 : * WAL that has been replayed. This distinction becomes particularly
3683 : * important during shutdown, as new WAL is no longer replayed and the
3684 : * last replayed LSN marks the furthest point up to which decoding can
3685 : * proceed.
3686 : */
3687 15357 : if (am_cascading_walsender)
3688 1894 : flushPtr = GetXLogReplayRecPtr(NULL);
3689 : else
3690 13463 : flushPtr = GetFlushRecPtr(NULL);
3691 : }
3692 :
3693 : /* If EndRecPtr is still past our flushPtr, it means we caught up. */
3694 941242 : if (logical_decoding_ctx->reader->EndRecPtr >= flushPtr)
3695 14211 : WalSndCaughtUp = true;
3696 :
3697 : /*
3698 : * If we're caught up and have been requested to stop, have WalSndLoop()
3699 : * terminate the connection in an orderly manner, after writing out all
3700 : * the pending data.
3701 : */
3702 941242 : if (WalSndCaughtUp && got_STOPPING)
3703 12453 : got_SIGUSR2 = true;
3704 :
3705 : /* Update shared memory status */
3706 : {
3707 941242 : WalSnd *walsnd = MyWalSnd;
3708 :
3709 941242 : SpinLockAcquire(&walsnd->mutex);
3710 941242 : walsnd->sentPtr = sentPtr;
3711 941242 : SpinLockRelease(&walsnd->mutex);
3712 : }
3713 941242 : }
3714 :
3715 : /*
3716 : * Forced shutdown of walsender if wal_sender_shutdown_timeout has expired.
3717 : */
3718 : static void
3719 3 : WalSndDoneImmediate(void)
3720 : {
3721 3 : WalSndState state = MyWalSnd->state;
3722 :
3723 3 : if ((state == WALSNDSTATE_CATCHUP ||
3724 1 : state == WALSNDSTATE_STREAMING ||
3725 3 : state == WALSNDSTATE_STOPPING) &&
3726 3 : !shutdown_stream_done_queued)
3727 : {
3728 : QueryCompletion qc;
3729 :
3730 : /* Try to inform receiver that XLOG streaming is done */
3731 3 : SetQueryCompletion(&qc, CMDTAG_COPY, 0);
3732 3 : EndCommandExtended(&qc, DestRemote, false, true);
3733 3 : shutdown_stream_done_queued = true;
3734 :
3735 : /*
3736 : * Note that the output buffer may be full during the forced shutdown
3737 : * of walsender. If pq_flush() is called at that time, the walsender
3738 : * process will be stuck. Therefore, call pq_flush_if_writable()
3739 : * instead. Successful reception of the done message with the
3740 : * walsender forced into a shutdown is not guaranteed.
3741 : */
3742 3 : pq_flush_if_writable();
3743 : }
3744 :
3745 : /*
3746 : * Prevent ereport from attempting to send any more messages to the
3747 : * standby. Otherwise, it can cause the process to get stuck if the output
3748 : * buffers are full.
3749 : */
3750 3 : if (whereToSendOutput == DestRemote)
3751 3 : whereToSendOutput = DestNone;
3752 :
3753 3 : ereport(WARNING,
3754 : (errmsg("terminating walsender process due to replication shutdown timeout"),
3755 : errdetail("Walsender process might have been terminated before all WAL data was replicated to the receiver.")));
3756 :
3757 3 : proc_exit(0);
3758 : }
3759 :
3760 : /*
3761 : * Shutdown if the sender is caught up.
3762 : *
3763 : * NB: This should only be called when the shutdown signal has been received
3764 : * from postmaster.
3765 : *
3766 : * Note that if we determine that there's still more data to send, this
3767 : * function will return control to the caller.
3768 : */
3769 : static void
3770 7354 : WalSndDone(WalSndSendDataCallback send_data)
3771 : {
3772 : XLogRecPtr replicatedPtr;
3773 :
3774 : /* ... let's just be real sure we're caught up ... */
3775 7354 : send_data();
3776 :
3777 : /*
3778 : * To figure out whether all WAL has successfully been replicated, check
3779 : * flush location if valid, write otherwise. Tools like pg_receivewal will
3780 : * usually (unless in synchronous mode) return an invalid flush location.
3781 : */
3782 14708 : replicatedPtr = XLogRecPtrIsValid(MyWalSnd->flush) ?
3783 7354 : MyWalSnd->flush : MyWalSnd->write;
3784 :
3785 7354 : if (WalSndCaughtUp && sentPtr == replicatedPtr &&
3786 43 : !pq_is_send_pending())
3787 : {
3788 : QueryCompletion qc;
3789 :
3790 : Assert(!shutdown_stream_done_queued);
3791 :
3792 : /* Inform the standby that XLOG streaming is done */
3793 43 : SetQueryCompletion(&qc, CMDTAG_COPY, 0);
3794 43 : EndCommandExtended(&qc, DestRemote, false, true);
3795 43 : shutdown_stream_done_queued = true;
3796 :
3797 : /*
3798 : * Reset last_reply_timestamp so subsequent WalSndComputeSleeptime()
3799 : * calls ignore wal_sender_timeout during shutdown.
3800 : */
3801 43 : last_reply_timestamp = 0;
3802 :
3803 : /*
3804 : * Do not call pq_flush() here, since it can block indefinitely while
3805 : * waiting for the socket to become writable, preventing
3806 : * wal_sender_shutdown_timeout from being enforced. Instead, use the
3807 : * walsender nonblocking flush path so the shutdown timeout continues
3808 : * to be checked while the send buffer drains.
3809 : */
3810 : for (;;)
3811 43 : {
3812 : long sleeptime;
3813 :
3814 : /*
3815 : * During shutdown, die if the shutdown timeout expires. Call this
3816 : * before WalSndComputeSleeptime() so the timeout is considered
3817 : * when computing sleep time.
3818 : */
3819 86 : WalSndCheckShutdownTimeout();
3820 :
3821 86 : if (!pq_is_send_pending())
3822 43 : break;
3823 :
3824 43 : sleeptime = WalSndComputeSleeptime(GetCurrentTimestamp());
3825 :
3826 : /* Sleep until something happens or we time out */
3827 43 : WalSndWait(WL_SOCKET_WRITEABLE, sleeptime,
3828 : WAIT_EVENT_WAL_SENDER_WRITE_DATA);
3829 :
3830 : /* Clear any already-pending wakeups */
3831 43 : ResetLatch(MyLatch);
3832 :
3833 43 : CHECK_FOR_INTERRUPTS();
3834 :
3835 : /* Try to flush pending output to the client */
3836 43 : if (pq_flush_if_writable() != 0)
3837 0 : WalSndShutdown();
3838 : }
3839 :
3840 43 : proc_exit(0);
3841 : }
3842 7311 : if (!waiting_for_ping_response)
3843 1668 : WalSndKeepalive(true, InvalidXLogRecPtr);
3844 7311 : }
3845 :
3846 : /*
3847 : * Returns the latest point in WAL that has been safely flushed to disk.
3848 : * This should only be called when in recovery.
3849 : *
3850 : * This is called either by cascading walsender to find WAL position to be sent
3851 : * to a cascaded standby or by slot synchronization operation to validate remote
3852 : * slot's lsn before syncing it locally.
3853 : *
3854 : * As a side-effect, *tli is updated to the TLI of the last
3855 : * replayed WAL record.
3856 : */
3857 : XLogRecPtr
3858 1113 : GetStandbyFlushRecPtr(TimeLineID *tli)
3859 : {
3860 : XLogRecPtr replayPtr;
3861 : TimeLineID replayTLI;
3862 : XLogRecPtr receivePtr;
3863 : TimeLineID receiveTLI;
3864 : XLogRecPtr result;
3865 :
3866 : Assert(am_cascading_walsender || IsSyncingReplicationSlots());
3867 :
3868 : /*
3869 : * We can safely send what's already been replayed. Also, if walreceiver
3870 : * is streaming WAL from the same timeline, we can send anything that it
3871 : * has streamed, but hasn't been replayed yet.
3872 : */
3873 :
3874 1113 : receivePtr = GetWalRcvFlushRecPtr(NULL, &receiveTLI);
3875 1113 : replayPtr = GetXLogReplayRecPtr(&replayTLI);
3876 :
3877 1113 : if (tli)
3878 1053 : *tli = replayTLI;
3879 :
3880 1113 : result = replayPtr;
3881 1113 : if (receiveTLI == replayTLI && receivePtr > replayPtr)
3882 148 : result = receivePtr;
3883 :
3884 1113 : return result;
3885 : }
3886 :
3887 : /*
3888 : * Request walsenders to reload the currently-open WAL file
3889 : */
3890 : void
3891 30 : WalSndRqstFileReload(void)
3892 : {
3893 : int i;
3894 :
3895 306 : for (i = 0; i < max_wal_senders; i++)
3896 : {
3897 276 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3898 :
3899 276 : SpinLockAcquire(&walsnd->mutex);
3900 276 : if (walsnd->pid == 0)
3901 : {
3902 276 : SpinLockRelease(&walsnd->mutex);
3903 276 : continue;
3904 : }
3905 0 : walsnd->needreload = true;
3906 0 : SpinLockRelease(&walsnd->mutex);
3907 : }
3908 30 : }
3909 :
3910 : /*
3911 : * Handle PROCSIG_WALSND_INIT_STOPPING signal.
3912 : */
3913 : void
3914 47 : HandleWalSndInitStopping(void)
3915 : {
3916 : Assert(am_walsender);
3917 :
3918 : /*
3919 : * If replication has not yet started, die like with SIGTERM. If
3920 : * replication is active, only set a flag and wake up the main loop. It
3921 : * will send any outstanding WAL, wait for it to be replicated to the
3922 : * standby, and then exit gracefully.
3923 : */
3924 47 : if (!replication_active)
3925 0 : kill(MyProcPid, SIGTERM);
3926 : else
3927 47 : got_STOPPING = true;
3928 :
3929 : /* latch will be set by procsignal_sigusr1_handler */
3930 47 : }
3931 :
3932 : /*
3933 : * SIGUSR2: set flag to do a last cycle and shut down afterwards. The WAL
3934 : * sender should already have been switched to WALSNDSTATE_STOPPING at
3935 : * this point.
3936 : */
3937 : static void
3938 47 : WalSndLastCycleHandler(SIGNAL_ARGS)
3939 : {
3940 47 : got_SIGUSR2 = true;
3941 47 : SetLatch(MyLatch);
3942 47 : }
3943 :
3944 : /* Set up signal handlers */
3945 : void
3946 1301 : WalSndSignals(void)
3947 : {
3948 : /* Set up signal handlers */
3949 1301 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
3950 1301 : pqsignal(SIGINT, StatementCancelHandler); /* query cancel */
3951 1301 : pqsignal(SIGTERM, die); /* request shutdown */
3952 : /* SIGQUIT handler was already set up by InitPostmasterChild */
3953 1301 : InitializeTimeouts(); /* establishes SIGALRM handler */
3954 1301 : pqsignal(SIGPIPE, PG_SIG_IGN);
3955 1301 : pqsignal(SIGUSR1, procsignal_sigusr1_handler);
3956 1301 : pqsignal(SIGUSR2, WalSndLastCycleHandler); /* request a last cycle and
3957 : * shutdown */
3958 :
3959 : /* Reset some signals that are accepted by postmaster but not here */
3960 1301 : pqsignal(SIGCHLD, PG_SIG_DFL);
3961 1301 : }
3962 :
3963 : /* Register shared-memory space needed by walsender */
3964 : static void
3965 1248 : WalSndShmemRequest(void *arg)
3966 : {
3967 : Size size;
3968 :
3969 1248 : size = offsetof(WalSndCtlData, walsnds);
3970 1248 : size = add_size(size, mul_size(max_wal_senders, sizeof(WalSnd)));
3971 1248 : ShmemRequestStruct(.name = "Wal Sender Ctl",
3972 : .size = size,
3973 : .ptr = (void **) &WalSndCtl,
3974 : );
3975 1248 : }
3976 :
3977 : /* Initialize walsender-related shared memory */
3978 : static void
3979 1245 : WalSndShmemInit(void *arg)
3980 : {
3981 4980 : for (int i = 0; i < NUM_SYNC_REP_WAIT_MODE; i++)
3982 3735 : dlist_init(&(WalSndCtl->SyncRepQueue[i]));
3983 :
3984 9163 : for (int i = 0; i < max_wal_senders; i++)
3985 : {
3986 7918 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
3987 :
3988 7918 : SpinLockInit(&walsnd->mutex);
3989 : }
3990 :
3991 1245 : ConditionVariableInit(&WalSndCtl->wal_flush_cv);
3992 1245 : ConditionVariableInit(&WalSndCtl->wal_replay_cv);
3993 1245 : ConditionVariableInit(&WalSndCtl->wal_confirm_rcv_cv);
3994 1245 : }
3995 :
3996 : /*
3997 : * Wake up physical, logical or both kinds of walsenders
3998 : *
3999 : * The distinction between physical and logical walsenders is done, because:
4000 : * - physical walsenders can't send data until it's been flushed
4001 : * - logical walsenders on standby can't decode and send data until it's been
4002 : * applied
4003 : *
4004 : * For cascading replication we need to wake up physical walsenders separately
4005 : * from logical walsenders (see the comment before calling WalSndWakeup() in
4006 : * ApplyWalRecord() for more details).
4007 : *
4008 : * This will be called inside critical sections, so throwing an error is not
4009 : * advisable.
4010 : */
4011 : void
4012 2846003 : WalSndWakeup(bool physical, bool logical)
4013 : {
4014 : /*
4015 : * Wake up all the walsenders waiting on WAL being flushed or replayed
4016 : * respectively. Note that waiting walsender would have prepared to sleep
4017 : * on the CV (i.e., added itself to the CV's waitlist) in WalSndWait()
4018 : * before actually waiting.
4019 : */
4020 2846003 : if (physical)
4021 154933 : ConditionVariableBroadcast(&WalSndCtl->wal_flush_cv);
4022 :
4023 2846003 : if (logical)
4024 2803886 : ConditionVariableBroadcast(&WalSndCtl->wal_replay_cv);
4025 2846003 : }
4026 :
4027 : /*
4028 : * Wait for readiness on the FeBe socket, or a timeout. The mask should be
4029 : * composed of optional WL_SOCKET_WRITEABLE and WL_SOCKET_READABLE flags. Exit
4030 : * on postmaster death.
4031 : */
4032 : static void
4033 100663 : WalSndWait(uint32 socket_events, long timeout, uint32 wait_event)
4034 : {
4035 : WaitEvent event;
4036 :
4037 100663 : ModifyWaitEvent(FeBeWaitSet, FeBeWaitSetSocketPos, socket_events, NULL);
4038 :
4039 : /*
4040 : * We use a condition variable to efficiently wake up walsenders in
4041 : * WalSndWakeup().
4042 : *
4043 : * Every walsender prepares to sleep on a shared memory CV. Note that it
4044 : * just prepares to sleep on the CV (i.e., adds itself to the CV's
4045 : * waitlist), but does not actually wait on the CV (IOW, it never calls
4046 : * ConditionVariableSleep()). It still uses WaitEventSetWait() for
4047 : * waiting, because we also need to wait for socket events. The processes
4048 : * (startup process, walreceiver etc.) wanting to wake up walsenders use
4049 : * ConditionVariableBroadcast(), which in turn calls SetLatch(), helping
4050 : * walsenders come out of WaitEventSetWait().
4051 : *
4052 : * This approach is simple and efficient because, one doesn't have to loop
4053 : * through all the walsenders slots, with a spinlock acquisition and
4054 : * release for every iteration, just to wake up only the waiting
4055 : * walsenders. It makes WalSndWakeup() callers' life easy.
4056 : *
4057 : * XXX: A desirable future improvement would be to add support for CVs
4058 : * into WaitEventSetWait().
4059 : *
4060 : * And, we use separate shared memory CVs for physical and logical
4061 : * walsenders for selective wake ups, see WalSndWakeup() for more details.
4062 : *
4063 : * If the wait event is WAIT_FOR_STANDBY_CONFIRMATION, wait on another CV
4064 : * until awakened by physical walsenders after the walreceiver confirms
4065 : * the receipt of the LSN.
4066 : */
4067 100663 : if (wait_event == WAIT_EVENT_WAIT_FOR_STANDBY_CONFIRMATION)
4068 13 : ConditionVariablePrepareToSleep(&WalSndCtl->wal_confirm_rcv_cv);
4069 100650 : else if (MyWalSnd->kind == REPLICATION_KIND_PHYSICAL)
4070 88819 : ConditionVariablePrepareToSleep(&WalSndCtl->wal_flush_cv);
4071 11831 : else if (MyWalSnd->kind == REPLICATION_KIND_LOGICAL)
4072 11831 : ConditionVariablePrepareToSleep(&WalSndCtl->wal_replay_cv);
4073 :
4074 100663 : if (WaitEventSetWait(FeBeWaitSet, timeout, &event, 1, wait_event) == 1 &&
4075 100660 : (event.events & WL_POSTMASTER_DEATH))
4076 : {
4077 0 : ConditionVariableCancelSleep();
4078 0 : proc_exit(1);
4079 : }
4080 :
4081 100663 : ConditionVariableCancelSleep();
4082 100663 : }
4083 :
4084 : /*
4085 : * Signal all walsenders to move to stopping state.
4086 : *
4087 : * This will trigger walsenders to move to a state where no further WAL can be
4088 : * generated. See this file's header for details.
4089 : */
4090 : void
4091 773 : WalSndInitStopping(void)
4092 : {
4093 : int i;
4094 :
4095 5715 : for (i = 0; i < max_wal_senders; i++)
4096 : {
4097 4942 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
4098 : pid_t pid;
4099 :
4100 4942 : SpinLockAcquire(&walsnd->mutex);
4101 4942 : pid = walsnd->pid;
4102 4942 : SpinLockRelease(&walsnd->mutex);
4103 :
4104 4942 : if (pid == 0)
4105 4895 : continue;
4106 :
4107 47 : SendProcSignal(pid, PROCSIG_WALSND_INIT_STOPPING, INVALID_PROC_NUMBER);
4108 : }
4109 773 : }
4110 :
4111 : /*
4112 : * Wait that all the WAL senders have quit or reached the stopping state. This
4113 : * is used by the checkpointer to control when the shutdown checkpoint can
4114 : * safely be performed.
4115 : */
4116 : void
4117 773 : WalSndWaitStopping(void)
4118 : {
4119 : for (;;)
4120 56 : {
4121 : int i;
4122 829 : bool all_stopped = true;
4123 :
4124 5775 : for (i = 0; i < max_wal_senders; i++)
4125 : {
4126 5002 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
4127 :
4128 5002 : SpinLockAcquire(&walsnd->mutex);
4129 :
4130 5002 : if (walsnd->pid == 0)
4131 : {
4132 4912 : SpinLockRelease(&walsnd->mutex);
4133 4912 : continue;
4134 : }
4135 :
4136 90 : if (walsnd->state != WALSNDSTATE_STOPPING)
4137 : {
4138 56 : all_stopped = false;
4139 56 : SpinLockRelease(&walsnd->mutex);
4140 56 : break;
4141 : }
4142 34 : SpinLockRelease(&walsnd->mutex);
4143 : }
4144 :
4145 : /* safe to leave if confirmation is done for all WAL senders */
4146 829 : if (all_stopped)
4147 773 : return;
4148 :
4149 56 : pg_usleep(10000L); /* wait for 10 msec */
4150 : }
4151 : }
4152 :
4153 : /* Set state for current walsender (only called in walsender) */
4154 : void
4155 4662 : WalSndSetState(WalSndState state)
4156 : {
4157 4662 : WalSnd *walsnd = MyWalSnd;
4158 :
4159 : Assert(am_walsender);
4160 :
4161 4662 : if (walsnd->state == state)
4162 2581 : return;
4163 :
4164 2081 : SpinLockAcquire(&walsnd->mutex);
4165 2081 : walsnd->state = state;
4166 2081 : SpinLockRelease(&walsnd->mutex);
4167 : }
4168 :
4169 : /*
4170 : * Return a string constant representing the state. This is used
4171 : * in system views, and should *not* be translated.
4172 : */
4173 : static const char *
4174 609 : WalSndGetStateString(WalSndState state)
4175 : {
4176 609 : switch (state)
4177 : {
4178 1 : case WALSNDSTATE_STARTUP:
4179 1 : return "startup";
4180 0 : case WALSNDSTATE_BACKUP:
4181 0 : return "backup";
4182 5 : case WALSNDSTATE_CATCHUP:
4183 5 : return "catchup";
4184 603 : case WALSNDSTATE_STREAMING:
4185 603 : return "streaming";
4186 0 : case WALSNDSTATE_STOPPING:
4187 0 : return "stopping";
4188 : }
4189 0 : return "UNKNOWN";
4190 : }
4191 :
4192 : static Interval *
4193 1431 : offset_to_interval(TimeOffset offset)
4194 : {
4195 1431 : Interval *result = palloc_object(Interval);
4196 :
4197 1431 : result->month = 0;
4198 1431 : result->day = 0;
4199 1431 : result->time = offset;
4200 :
4201 1431 : return result;
4202 : }
4203 :
4204 : /*
4205 : * Returns activity of walsenders, including pids and xlog locations sent to
4206 : * standby servers.
4207 : */
4208 : Datum
4209 474 : pg_stat_get_wal_senders(PG_FUNCTION_ARGS)
4210 : {
4211 : #define PG_STAT_GET_WAL_SENDERS_COLS 12
4212 474 : ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
4213 : SyncRepStandbyData *sync_standbys;
4214 : int num_standbys;
4215 : int i;
4216 :
4217 474 : InitMaterializedSRF(fcinfo, 0);
4218 :
4219 : /*
4220 : * Get the currently active synchronous standbys. This could be out of
4221 : * date before we're done, but we'll use the data anyway.
4222 : */
4223 474 : num_standbys = SyncRepGetCandidateStandbys(&sync_standbys);
4224 :
4225 5202 : for (i = 0; i < max_wal_senders; i++)
4226 : {
4227 4728 : WalSnd *walsnd = &WalSndCtl->walsnds[i];
4228 : XLogRecPtr sent_ptr;
4229 : XLogRecPtr write;
4230 : XLogRecPtr flush;
4231 : XLogRecPtr apply;
4232 : TimeOffset writeLag;
4233 : TimeOffset flushLag;
4234 : TimeOffset applyLag;
4235 : int priority;
4236 : int pid;
4237 : WalSndState state;
4238 : TimestampTz replyTime;
4239 : bool is_sync_standby;
4240 : Datum values[PG_STAT_GET_WAL_SENDERS_COLS];
4241 4728 : bool nulls[PG_STAT_GET_WAL_SENDERS_COLS] = {0};
4242 : int j;
4243 :
4244 : /* Collect data from shared memory */
4245 4728 : SpinLockAcquire(&walsnd->mutex);
4246 4728 : if (walsnd->pid == 0)
4247 : {
4248 4119 : SpinLockRelease(&walsnd->mutex);
4249 4119 : continue;
4250 : }
4251 609 : pid = walsnd->pid;
4252 609 : sent_ptr = walsnd->sentPtr;
4253 609 : state = walsnd->state;
4254 609 : write = walsnd->write;
4255 609 : flush = walsnd->flush;
4256 609 : apply = walsnd->apply;
4257 609 : writeLag = walsnd->writeLag;
4258 609 : flushLag = walsnd->flushLag;
4259 609 : applyLag = walsnd->applyLag;
4260 609 : priority = walsnd->sync_standby_priority;
4261 609 : replyTime = walsnd->replyTime;
4262 609 : SpinLockRelease(&walsnd->mutex);
4263 :
4264 : /*
4265 : * Detect whether walsender is/was considered synchronous. We can
4266 : * provide some protection against stale data by checking the PID
4267 : * along with walsnd_index.
4268 : */
4269 609 : is_sync_standby = false;
4270 656 : for (j = 0; j < num_standbys; j++)
4271 : {
4272 76 : if (sync_standbys[j].walsnd_index == i &&
4273 29 : sync_standbys[j].pid == pid)
4274 : {
4275 29 : is_sync_standby = true;
4276 29 : break;
4277 : }
4278 : }
4279 :
4280 609 : values[0] = Int32GetDatum(pid);
4281 :
4282 609 : if (!has_privs_of_role(GetUserId(), ROLE_PG_READ_ALL_STATS))
4283 : {
4284 : /*
4285 : * Only superusers and roles with privileges of pg_read_all_stats
4286 : * can see details. Other users only get the pid value to know
4287 : * it's a walsender, but no details.
4288 : */
4289 0 : MemSet(&nulls[1], true, PG_STAT_GET_WAL_SENDERS_COLS - 1);
4290 : }
4291 : else
4292 : {
4293 609 : values[1] = CStringGetTextDatum(WalSndGetStateString(state));
4294 :
4295 609 : if (!XLogRecPtrIsValid(sent_ptr))
4296 1 : nulls[2] = true;
4297 609 : values[2] = LSNGetDatum(sent_ptr);
4298 :
4299 609 : if (!XLogRecPtrIsValid(write))
4300 1 : nulls[3] = true;
4301 609 : values[3] = LSNGetDatum(write);
4302 :
4303 609 : if (!XLogRecPtrIsValid(flush))
4304 1 : nulls[4] = true;
4305 609 : values[4] = LSNGetDatum(flush);
4306 :
4307 609 : if (!XLogRecPtrIsValid(apply))
4308 1 : nulls[5] = true;
4309 609 : values[5] = LSNGetDatum(apply);
4310 :
4311 : /*
4312 : * Treat a standby such as a pg_basebackup background process
4313 : * which always returns an invalid flush location, as an
4314 : * asynchronous standby.
4315 : */
4316 609 : priority = XLogRecPtrIsValid(flush) ? priority : 0;
4317 :
4318 609 : if (writeLag < 0)
4319 132 : nulls[6] = true;
4320 : else
4321 477 : values[6] = IntervalPGetDatum(offset_to_interval(writeLag));
4322 :
4323 609 : if (flushLag < 0)
4324 132 : nulls[7] = true;
4325 : else
4326 477 : values[7] = IntervalPGetDatum(offset_to_interval(flushLag));
4327 :
4328 609 : if (applyLag < 0)
4329 132 : nulls[8] = true;
4330 : else
4331 477 : values[8] = IntervalPGetDatum(offset_to_interval(applyLag));
4332 :
4333 609 : values[9] = Int32GetDatum(priority);
4334 :
4335 : /*
4336 : * More easily understood version of standby state. This is purely
4337 : * informational.
4338 : *
4339 : * In quorum-based sync replication, the role of each standby
4340 : * listed in synchronous_standby_names can be changing very
4341 : * frequently. Any standbys considered as "sync" at one moment can
4342 : * be switched to "potential" ones at the next moment. So, it's
4343 : * basically useless to report "sync" or "potential" as their sync
4344 : * states. We report just "quorum" for them.
4345 : */
4346 609 : if (priority == 0)
4347 569 : values[10] = CStringGetTextDatum("async");
4348 40 : else if (is_sync_standby)
4349 29 : values[10] = SyncRepConfig->syncrep_method == SYNC_REP_PRIORITY ?
4350 29 : CStringGetTextDatum("sync") : CStringGetTextDatum("quorum");
4351 : else
4352 11 : values[10] = CStringGetTextDatum("potential");
4353 :
4354 609 : if (replyTime == 0)
4355 1 : nulls[11] = true;
4356 : else
4357 608 : values[11] = TimestampTzGetDatum(replyTime);
4358 : }
4359 :
4360 609 : tuplestore_putvalues(rsinfo->setResult, rsinfo->setDesc,
4361 : values, nulls);
4362 : }
4363 :
4364 474 : return (Datum) 0;
4365 : }
4366 :
4367 : /*
4368 : * Send a keepalive message to standby.
4369 : *
4370 : * If requestReply is set, the message requests the other party to send
4371 : * a message back to us, for heartbeat purposes. We also set a flag to
4372 : * let nearby code know that we're waiting for that response, to avoid
4373 : * repeated requests.
4374 : *
4375 : * writePtr is the location up to which the WAL is sent. It is essentially
4376 : * the same as sentPtr but in some cases, we need to send keep alive before
4377 : * sentPtr is updated like when skipping empty transactions.
4378 : */
4379 : static void
4380 3636 : WalSndKeepalive(bool requestReply, XLogRecPtr writePtr)
4381 : {
4382 3636 : elog(DEBUG2, "sending replication keepalive");
4383 :
4384 : /* construct the message... */
4385 3636 : resetStringInfo(&output_message);
4386 3636 : pq_sendbyte(&output_message, PqReplMsg_Keepalive);
4387 3636 : pq_sendint64(&output_message, XLogRecPtrIsValid(writePtr) ? writePtr : sentPtr);
4388 3636 : pq_sendint64(&output_message, GetCurrentTimestamp());
4389 3636 : pq_sendbyte(&output_message, requestReply ? 1 : 0);
4390 :
4391 : /* ... and send it wrapped in CopyData */
4392 3636 : pq_putmessage_noblock(PqMsg_CopyData, output_message.data, output_message.len);
4393 :
4394 : /* Set local flag */
4395 3636 : if (requestReply)
4396 1668 : waiting_for_ping_response = true;
4397 3636 : }
4398 :
4399 : /*
4400 : * Send keepalive message if too much time has elapsed.
4401 : */
4402 : static void
4403 1161635 : WalSndKeepaliveIfNecessary(void)
4404 : {
4405 : TimestampTz ping_time;
4406 :
4407 : /*
4408 : * Don't send keepalive messages if timeouts are globally disabled or
4409 : * we're doing something not partaking in timeouts.
4410 : */
4411 1161635 : if (wal_sender_timeout <= 0 || last_reply_timestamp <= 0)
4412 27 : return;
4413 :
4414 1161608 : if (waiting_for_ping_response)
4415 8975 : return;
4416 :
4417 : /*
4418 : * If half of wal_sender_timeout has lapsed without receiving any reply
4419 : * from the standby, send a keep-alive message to the standby requesting
4420 : * an immediate reply.
4421 : */
4422 1152633 : ping_time = TimestampTzPlusMilliseconds(last_reply_timestamp,
4423 : wal_sender_timeout / 2);
4424 1152633 : if (last_processing >= ping_time)
4425 : {
4426 0 : WalSndKeepalive(true, InvalidXLogRecPtr);
4427 :
4428 : /* Try to flush pending output to the client */
4429 0 : if (pq_flush_if_writable() != 0)
4430 0 : WalSndShutdown();
4431 : }
4432 : }
4433 :
4434 : /*
4435 : * Record the end of the WAL and the time it was flushed locally, so that
4436 : * LagTrackerRead can compute the elapsed time (lag) when this WAL location is
4437 : * eventually reported to have been written, flushed and applied by the
4438 : * standby in a reply message.
4439 : */
4440 : static void
4441 140074 : LagTrackerWrite(XLogRecPtr lsn, TimestampTz local_flush_time)
4442 : {
4443 : int new_write_head;
4444 : int i;
4445 :
4446 140074 : if (!am_walsender)
4447 0 : return;
4448 :
4449 : /*
4450 : * If the lsn hasn't advanced since last time, then do nothing. This way
4451 : * we only record a new sample when new WAL has been written.
4452 : */
4453 140074 : if (lag_tracker->last_lsn == lsn)
4454 116979 : return;
4455 23095 : lag_tracker->last_lsn = lsn;
4456 :
4457 : /*
4458 : * If advancing the write head of the circular buffer would crash into any
4459 : * of the read heads, then the buffer is full. In other words, the
4460 : * slowest reader (presumably apply) is the one that controls the release
4461 : * of space.
4462 : */
4463 23095 : new_write_head = (lag_tracker->write_head + 1) % LAG_TRACKER_BUFFER_SIZE;
4464 92380 : for (i = 0; i < NUM_SYNC_REP_WAIT_MODE; ++i)
4465 : {
4466 : /*
4467 : * If the buffer is full, move the slowest reader to a separate
4468 : * overflow entry and free its space in the buffer so the write head
4469 : * can advance.
4470 : */
4471 69285 : if (new_write_head == lag_tracker->read_heads[i])
4472 : {
4473 0 : lag_tracker->overflowed[i] =
4474 0 : lag_tracker->buffer[lag_tracker->read_heads[i]];
4475 0 : lag_tracker->read_heads[i] = -1;
4476 : }
4477 : }
4478 :
4479 : /* Store a sample at the current write head position. */
4480 23095 : lag_tracker->buffer[lag_tracker->write_head].lsn = lsn;
4481 23095 : lag_tracker->buffer[lag_tracker->write_head].time = local_flush_time;
4482 23095 : lag_tracker->write_head = new_write_head;
4483 : }
4484 :
4485 : /*
4486 : * Find out how much time has elapsed between the moment WAL location 'lsn'
4487 : * (or the highest known earlier LSN) was flushed locally and the time 'now'.
4488 : * We have a separate read head for each of the reported LSN locations we
4489 : * receive in replies from standby; 'head' controls which read head is
4490 : * used. Whenever a read head crosses an LSN which was written into the
4491 : * lag buffer with LagTrackerWrite, we can use the associated timestamp to
4492 : * find out the time this LSN (or an earlier one) was flushed locally, and
4493 : * therefore compute the lag.
4494 : *
4495 : * Return -1 if no new sample data is available, and otherwise the elapsed
4496 : * time in microseconds.
4497 : */
4498 : static TimeOffset
4499 333048 : LagTrackerRead(int head, XLogRecPtr lsn, TimestampTz now)
4500 : {
4501 333048 : TimestampTz time = 0;
4502 :
4503 : /*
4504 : * If 'lsn' has not passed the WAL position stored in the overflow entry,
4505 : * return the elapsed time (in microseconds) since the saved local flush
4506 : * time. If the flush time is in the future (due to clock drift), return
4507 : * -1 to treat as no valid sample.
4508 : *
4509 : * Otherwise, switch back to using the buffer to control the read head and
4510 : * compute the elapsed time. The read head is then reset to point to the
4511 : * oldest entry in the buffer.
4512 : */
4513 333048 : if (lag_tracker->read_heads[head] == -1)
4514 : {
4515 0 : if (lag_tracker->overflowed[head].lsn > lsn)
4516 0 : return (now >= lag_tracker->overflowed[head].time) ?
4517 0 : now - lag_tracker->overflowed[head].time : -1;
4518 :
4519 0 : time = lag_tracker->overflowed[head].time;
4520 0 : lag_tracker->last_read[head] = lag_tracker->overflowed[head];
4521 0 : lag_tracker->read_heads[head] =
4522 0 : (lag_tracker->write_head + 1) % LAG_TRACKER_BUFFER_SIZE;
4523 : }
4524 :
4525 : /* Read all unread samples up to this LSN or end of buffer. */
4526 401153 : while (lag_tracker->read_heads[head] != lag_tracker->write_head &&
4527 264841 : lag_tracker->buffer[lag_tracker->read_heads[head]].lsn <= lsn)
4528 : {
4529 68105 : time = lag_tracker->buffer[lag_tracker->read_heads[head]].time;
4530 68105 : lag_tracker->last_read[head] =
4531 68105 : lag_tracker->buffer[lag_tracker->read_heads[head]];
4532 68105 : lag_tracker->read_heads[head] =
4533 68105 : (lag_tracker->read_heads[head] + 1) % LAG_TRACKER_BUFFER_SIZE;
4534 : }
4535 :
4536 : /*
4537 : * If the lag tracker is empty, that means the standby has processed
4538 : * everything we've ever sent so we should now clear 'last_read'. If we
4539 : * didn't do that, we'd risk using a stale and irrelevant sample for
4540 : * interpolation at the beginning of the next burst of WAL after a period
4541 : * of idleness.
4542 : */
4543 333048 : if (lag_tracker->read_heads[head] == lag_tracker->write_head)
4544 136312 : lag_tracker->last_read[head].time = 0;
4545 :
4546 333048 : if (time > now)
4547 : {
4548 : /* If the clock somehow went backwards, treat as not found. */
4549 0 : return -1;
4550 : }
4551 333048 : else if (time == 0)
4552 : {
4553 : /*
4554 : * We didn't cross a time. If there is a future sample that we
4555 : * haven't reached yet, and we've already reached at least one sample,
4556 : * let's interpolate the local flushed time. This is mainly useful
4557 : * for reporting a completely stuck apply position as having
4558 : * increasing lag, since otherwise we'd have to wait for it to
4559 : * eventually start moving again and cross one of our samples before
4560 : * we can show the lag increasing.
4561 : */
4562 278963 : if (lag_tracker->read_heads[head] == lag_tracker->write_head)
4563 : {
4564 : /* There are no future samples, so we can't interpolate. */
4565 91292 : return -1;
4566 : }
4567 187671 : else if (lag_tracker->last_read[head].time != 0)
4568 : {
4569 : /* We can interpolate between last_read and the next sample. */
4570 : double fraction;
4571 73775 : WalTimeSample prev = lag_tracker->last_read[head];
4572 73775 : WalTimeSample next = lag_tracker->buffer[lag_tracker->read_heads[head]];
4573 :
4574 73775 : if (lsn < prev.lsn)
4575 : {
4576 : /*
4577 : * Reported LSNs shouldn't normally go backwards, but it's
4578 : * possible when there is a timeline change. Treat as not
4579 : * found.
4580 : */
4581 0 : return -1;
4582 : }
4583 :
4584 : Assert(prev.lsn < next.lsn);
4585 :
4586 73775 : if (prev.time > next.time)
4587 : {
4588 : /* If the clock somehow went backwards, treat as not found. */
4589 0 : return -1;
4590 : }
4591 :
4592 : /* See how far we are between the previous and next samples. */
4593 73775 : fraction =
4594 73775 : (double) (lsn - prev.lsn) / (double) (next.lsn - prev.lsn);
4595 :
4596 : /* Scale the local flush time proportionally. */
4597 73775 : time = (TimestampTz)
4598 73775 : ((double) prev.time + (next.time - prev.time) * fraction);
4599 : }
4600 : else
4601 : {
4602 : /*
4603 : * We have only a future sample, implying that we were entirely
4604 : * caught up but and now there is a new burst of WAL and the
4605 : * standby hasn't processed the first sample yet. Until the
4606 : * standby reaches the future sample the best we can do is report
4607 : * the hypothetical lag if that sample were to be replayed now.
4608 : */
4609 113896 : time = lag_tracker->buffer[lag_tracker->read_heads[head]].time;
4610 : }
4611 : }
4612 :
4613 : /* Return the elapsed time since local flush time in microseconds. */
4614 : Assert(time != 0);
4615 241756 : return now - time;
4616 : }
|
