Branch data Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : * worker.c
3 : : * PostgreSQL logical replication worker (apply)
4 : : *
5 : : * Copyright (c) 2016-2026, PostgreSQL Global Development Group
6 : : *
7 : : * IDENTIFICATION
8 : : * src/backend/replication/logical/worker.c
9 : : *
10 : : * NOTES
11 : : * This file contains the worker which applies logical changes as they come
12 : : * from remote logical replication stream.
13 : : *
14 : : * The main worker (apply) is started by logical replication worker
15 : : * launcher for every enabled subscription in a database. It uses
16 : : * walsender protocol to communicate with publisher.
17 : : *
18 : : * This module includes server facing code and shares libpqwalreceiver
19 : : * module with walreceiver for providing the libpq specific functionality.
20 : : *
21 : : *
22 : : * STREAMED TRANSACTIONS
23 : : * ---------------------
24 : : * Streamed transactions (large transactions exceeding a memory limit on the
25 : : * upstream) are applied using one of two approaches:
26 : : *
27 : : * 1) Write to temporary files and apply when the final commit arrives
28 : : *
29 : : * This approach is used when the user has set the subscription's streaming
30 : : * option as on.
31 : : *
32 : : * Unlike the regular (non-streamed) case, handling streamed transactions has
33 : : * to handle aborts of both the toplevel transaction and subtransactions. This
34 : : * is achieved by tracking offsets for subtransactions, which is then used
35 : : * to truncate the file with serialized changes.
36 : : *
37 : : * The files are placed in tmp file directory by default, and the filenames
38 : : * include both the XID of the toplevel transaction and OID of the
39 : : * subscription. This is necessary so that different workers processing a
40 : : * remote transaction with the same XID doesn't interfere.
41 : : *
42 : : * We use BufFiles instead of using normal temporary files because (a) the
43 : : * BufFile infrastructure supports temporary files that exceed the OS file size
44 : : * limit, (b) provides a way for automatic clean up on the error and (c) provides
45 : : * a way to survive these files across local transactions and allow to open and
46 : : * close at stream start and close. We decided to use FileSet
47 : : * infrastructure as without that it deletes the files on the closure of the
48 : : * file and if we decide to keep stream files open across the start/stop stream
49 : : * then it will consume a lot of memory (more than 8K for each BufFile and
50 : : * there could be multiple such BufFiles as the subscriber could receive
51 : : * multiple start/stop streams for different transactions before getting the
52 : : * commit). Moreover, if we don't use FileSet then we also need to invent
53 : : * a new way to pass filenames to BufFile APIs so that we are allowed to open
54 : : * the file we desired across multiple stream-open calls for the same
55 : : * transaction.
56 : : *
57 : : * 2) Parallel apply workers.
58 : : *
59 : : * This approach is used when the user has set the subscription's streaming
60 : : * option as parallel. See logical/applyparallelworker.c for information about
61 : : * this approach.
62 : : *
63 : : * TWO_PHASE TRANSACTIONS
64 : : * ----------------------
65 : : * Two phase transactions are replayed at prepare and then committed or
66 : : * rolled back at commit prepared and rollback prepared respectively. It is
67 : : * possible to have a prepared transaction that arrives at the apply worker
68 : : * when the tablesync is busy doing the initial copy. In this case, the apply
69 : : * worker skips all the prepared operations [e.g. inserts] while the tablesync
70 : : * is still busy (see the condition of should_apply_changes_for_rel). The
71 : : * tablesync worker might not get such a prepared transaction because say it
72 : : * was prior to the initial consistent point but might have got some later
73 : : * commits. Now, the tablesync worker will exit without doing anything for the
74 : : * prepared transaction skipped by the apply worker as the sync location for it
75 : : * will be already ahead of the apply worker's current location. This would lead
76 : : * to an "empty prepare", because later when the apply worker does the commit
77 : : * prepare, there is nothing in it (the inserts were skipped earlier).
78 : : *
79 : : * To avoid this, and similar prepare confusions the subscription's two_phase
80 : : * commit is enabled only after the initial sync is over. The two_phase option
81 : : * has been implemented as a tri-state with values DISABLED, PENDING, and
82 : : * ENABLED.
83 : : *
84 : : * Even if the user specifies they want a subscription with two_phase = on,
85 : : * internally it will start with a tri-state of PENDING which only becomes
86 : : * ENABLED after all tablesync initializations are completed - i.e. when all
87 : : * tablesync workers have reached their READY state. In other words, the value
88 : : * PENDING is only a temporary state for subscription start-up.
89 : : *
90 : : * Until the two_phase is properly available (ENABLED) the subscription will
91 : : * behave as if two_phase = off. When the apply worker detects that all
92 : : * tablesyncs have become READY (while the tri-state was PENDING) it will
93 : : * restart the apply worker process. This happens in
94 : : * ProcessSyncingTablesForApply.
95 : : *
96 : : * When the (re-started) apply worker finds that all tablesyncs are READY for a
97 : : * two_phase tri-state of PENDING it start streaming messages with the
98 : : * two_phase option which in turn enables the decoding of two-phase commits at
99 : : * the publisher. Then, it updates the tri-state value from PENDING to ENABLED.
100 : : * Now, it is possible that during the time we have not enabled two_phase, the
101 : : * publisher (replication server) would have skipped some prepares but we
102 : : * ensure that such prepares are sent along with commit prepare, see
103 : : * ReorderBufferFinishPrepared.
104 : : *
105 : : * If the subscription has no tables then a two_phase tri-state PENDING is
106 : : * left unchanged. This lets the user still do an ALTER SUBSCRIPTION REFRESH
107 : : * PUBLICATION which might otherwise be disallowed (see below).
108 : : *
109 : : * If ever a user needs to be aware of the tri-state value, they can fetch it
110 : : * from the pg_subscription catalog (see column subtwophasestate).
111 : : *
112 : : * Finally, to avoid problems mentioned in previous paragraphs from any
113 : : * subsequent (not READY) tablesyncs (need to toggle two_phase option from 'on'
114 : : * to 'off' and then again back to 'on') there is a restriction for
115 : : * ALTER SUBSCRIPTION REFRESH PUBLICATION. This command is not permitted when
116 : : * the two_phase tri-state is ENABLED, except when copy_data = false.
117 : : *
118 : : * We can get prepare of the same GID more than once for the genuine cases
119 : : * where we have defined multiple subscriptions for publications on the same
120 : : * server and prepared transaction has operations on tables subscribed to those
121 : : * subscriptions. For such cases, if we use the GID sent by publisher one of
122 : : * the prepares will be successful and others will fail, in which case the
123 : : * server will send them again. Now, this can lead to a deadlock if user has
124 : : * set synchronous_standby_names for all the subscriptions on subscriber. To
125 : : * avoid such deadlocks, we generate a unique GID (consisting of the
126 : : * subscription oid and the xid of the prepared transaction) for each prepare
127 : : * transaction on the subscriber.
128 : : *
129 : : * FAILOVER
130 : : * ----------------------
131 : : * The logical slot on the primary can be synced to the standby by specifying
132 : : * failover = true when creating the subscription. Enabling failover allows us
133 : : * to smoothly transition to the promoted standby, ensuring that we can
134 : : * subscribe to the new primary without losing any data.
135 : : *
136 : : * RETAIN DEAD TUPLES
137 : : * ----------------------
138 : : * Each apply worker that enabled retain_dead_tuples option maintains a
139 : : * non-removable transaction ID (oldest_nonremovable_xid) in shared memory to
140 : : * prevent dead rows from being removed prematurely when the apply worker still
141 : : * needs them to detect update_deleted conflicts. Additionally, this helps to
142 : : * retain the required commit_ts module information, which further helps to
143 : : * detect update_origin_differs and delete_origin_differs conflicts reliably, as
144 : : * otherwise, vacuum freeze could remove the required information.
145 : : *
146 : : * The logical replication launcher manages an internal replication slot named
147 : : * "pg_conflict_detection". It asynchronously aggregates the non-removable
148 : : * transaction ID from all apply workers to determine the appropriate xmin for
149 : : * the slot, thereby retaining necessary tuples.
150 : : *
151 : : * The non-removable transaction ID in the apply worker is advanced to the
152 : : * oldest running transaction ID once all concurrent transactions on the
153 : : * publisher have been applied and flushed locally. The process involves:
154 : : *
155 : : * - RDT_GET_CANDIDATE_XID:
156 : : * Call GetOldestActiveTransactionId() to take oldestRunningXid as the
157 : : * candidate xid.
158 : : *
159 : : * - RDT_REQUEST_PUBLISHER_STATUS:
160 : : * Send a message to the walsender requesting the publisher status, which
161 : : * includes the latest WAL write position and information about transactions
162 : : * that are in the commit phase.
163 : : *
164 : : * - RDT_WAIT_FOR_PUBLISHER_STATUS:
165 : : * Wait for the status from the walsender. After receiving the first status,
166 : : * do not proceed if there are concurrent remote transactions that are still
167 : : * in the commit phase. These transactions might have been assigned an
168 : : * earlier commit timestamp but have not yet written the commit WAL record.
169 : : * Continue to request the publisher status (RDT_REQUEST_PUBLISHER_STATUS)
170 : : * until all these transactions have completed.
171 : : *
172 : : * - RDT_WAIT_FOR_LOCAL_FLUSH:
173 : : * Advance the non-removable transaction ID if the current flush location has
174 : : * reached or surpassed the last received WAL position.
175 : : *
176 : : * - RDT_STOP_CONFLICT_INFO_RETENTION:
177 : : * This phase is required only when max_retention_duration is defined. We
178 : : * enter this phase if the wait time in either the
179 : : * RDT_WAIT_FOR_PUBLISHER_STATUS or RDT_WAIT_FOR_LOCAL_FLUSH phase exceeds
180 : : * configured max_retention_duration. In this phase,
181 : : * pg_subscription.subretentionactive is updated to false within a new
182 : : * transaction, and oldest_nonremovable_xid is set to InvalidTransactionId.
183 : : *
184 : : * - RDT_RESUME_CONFLICT_INFO_RETENTION:
185 : : * This phase is required only when max_retention_duration is defined. We
186 : : * enter this phase if the retention was previously stopped, and the time
187 : : * required to advance the non-removable transaction ID in the
188 : : * RDT_WAIT_FOR_LOCAL_FLUSH phase has decreased to within acceptable limits
189 : : * (or if max_retention_duration is set to 0). During this phase,
190 : : * pg_subscription.subretentionactive is updated to true within a new
191 : : * transaction, and the worker will be restarted.
192 : : *
193 : : * The overall state progression is: GET_CANDIDATE_XID ->
194 : : * REQUEST_PUBLISHER_STATUS -> WAIT_FOR_PUBLISHER_STATUS -> (loop to
195 : : * REQUEST_PUBLISHER_STATUS till concurrent remote transactions end) ->
196 : : * WAIT_FOR_LOCAL_FLUSH -> loop back to GET_CANDIDATE_XID.
197 : : *
198 : : * Retaining the dead tuples for this period is sufficient for ensuring
199 : : * eventual consistency using last-update-wins strategy, as dead tuples are
200 : : * useful for detecting conflicts only during the application of concurrent
201 : : * transactions from remote nodes. After applying and flushing all remote
202 : : * transactions that occurred concurrently with the tuple DELETE, any
203 : : * subsequent UPDATE from a remote node should have a later timestamp. In such
204 : : * cases, it is acceptable to detect an update_missing scenario and convert the
205 : : * UPDATE to an INSERT when applying it. But, for concurrent remote
206 : : * transactions with earlier timestamps than the DELETE, detecting
207 : : * update_deleted is necessary, as the UPDATEs in remote transactions should be
208 : : * ignored if their timestamp is earlier than that of the dead tuples.
209 : : *
210 : : * Note that advancing the non-removable transaction ID is not supported if the
211 : : * publisher is also a physical standby. This is because the logical walsender
212 : : * on the standby can only get the WAL replay position but there may be more
213 : : * WALs that are being replicated from the primary and those WALs could have
214 : : * earlier commit timestamp.
215 : : *
216 : : * Similarly, when the publisher has subscribed to another publisher,
217 : : * information necessary for conflict detection cannot be retained for
218 : : * changes from origins other than the publisher. This is because publisher
219 : : * lacks the information on concurrent transactions of other publishers to
220 : : * which it subscribes. As the information on concurrent transactions is
221 : : * unavailable beyond subscriber's immediate publishers, the non-removable
222 : : * transaction ID might be advanced prematurely before changes from other
223 : : * origins have been fully applied.
224 : : *
225 : : * XXX Retaining information for changes from other origins might be possible
226 : : * by requesting the subscription on that origin to enable retain_dead_tuples
227 : : * and fetching the conflict detection slot.xmin along with the publisher's
228 : : * status. In the RDT_WAIT_FOR_PUBLISHER_STATUS phase, the apply worker could
229 : : * wait for the remote slot's xmin to reach the oldest active transaction ID,
230 : : * ensuring that all transactions from other origins have been applied on the
231 : : * publisher, thereby getting the latest WAL position that includes all
232 : : * concurrent changes. However, this approach may impact performance, so it
233 : : * might not worth the effort.
234 : : *
235 : : * XXX It seems feasible to get the latest commit's WAL location from the
236 : : * publisher and wait till that is applied. However, we can't do that
237 : : * because commit timestamps can regress as a commit with a later LSN is not
238 : : * guaranteed to have a later timestamp than those with earlier LSNs. Having
239 : : * said that, even if that is possible, it won't improve performance much as
240 : : * the apply always lag and moves slowly as compared with the transactions
241 : : * on the publisher.
242 : : *-------------------------------------------------------------------------
243 : : */
244 : :
245 : : #include "postgres.h"
246 : :
247 : : #include <sys/stat.h>
248 : : #include <unistd.h>
249 : :
250 : : #include "access/genam.h"
251 : : #include "access/commit_ts.h"
252 : : #include "access/table.h"
253 : : #include "access/tableam.h"
254 : : #include "access/tupconvert.h"
255 : : #include "access/twophase.h"
256 : : #include "access/xact.h"
257 : : #include "catalog/indexing.h"
258 : : #include "catalog/pg_inherits.h"
259 : : #include "catalog/pg_subscription.h"
260 : : #include "catalog/pg_subscription_rel.h"
261 : : #include "commands/subscriptioncmds.h"
262 : : #include "commands/tablecmds.h"
263 : : #include "commands/trigger.h"
264 : : #include "executor/executor.h"
265 : : #include "executor/execPartition.h"
266 : : #include "libpq/pqformat.h"
267 : : #include "miscadmin.h"
268 : : #include "optimizer/optimizer.h"
269 : : #include "parser/parse_relation.h"
270 : : #include "pgstat.h"
271 : : #include "port/pg_bitutils.h"
272 : : #include "postmaster/bgworker.h"
273 : : #include "postmaster/interrupt.h"
274 : : #include "postmaster/walwriter.h"
275 : : #include "replication/conflict.h"
276 : : #include "replication/logicallauncher.h"
277 : : #include "replication/logicalproto.h"
278 : : #include "replication/logicalrelation.h"
279 : : #include "replication/logicalworker.h"
280 : : #include "replication/origin.h"
281 : : #include "replication/slot.h"
282 : : #include "replication/walreceiver.h"
283 : : #include "replication/worker_internal.h"
284 : : #include "rewrite/rewriteHandler.h"
285 : : #include "storage/buffile.h"
286 : : #include "storage/ipc.h"
287 : : #include "storage/latch.h"
288 : : #include "storage/lmgr.h"
289 : : #include "storage/procarray.h"
290 : : #include "tcop/tcopprot.h"
291 : : #include "utils/acl.h"
292 : : #include "utils/guc.h"
293 : : #include "utils/inval.h"
294 : : #include "utils/lsyscache.h"
295 : : #include "utils/memutils.h"
296 : : #include "utils/pg_lsn.h"
297 : : #include "utils/rel.h"
298 : : #include "utils/rls.h"
299 : : #include "utils/snapmgr.h"
300 : : #include "utils/syscache.h"
301 : : #include "utils/usercontext.h"
302 : : #include "utils/wait_event.h"
303 : :
304 : : #define NAPTIME_PER_CYCLE 1000 /* max sleep time between cycles (1s) */
305 : :
306 : : typedef struct FlushPosition
307 : : {
308 : : dlist_node node;
309 : : XLogRecPtr local_end;
310 : : XLogRecPtr remote_end;
311 : : } FlushPosition;
312 : :
313 : : static dlist_head lsn_mapping = DLIST_STATIC_INIT(lsn_mapping);
314 : :
315 : : typedef struct ApplyExecutionData
316 : : {
317 : : EState *estate; /* executor state, used to track resources */
318 : :
319 : : LogicalRepRelMapEntry *targetRel; /* replication target rel */
320 : : ResultRelInfo *targetRelInfo; /* ResultRelInfo for same */
321 : :
322 : : /* These fields are used when the target relation is partitioned: */
323 : : ModifyTableState *mtstate; /* dummy ModifyTable state */
324 : : PartitionTupleRouting *proute; /* partition routing info */
325 : : } ApplyExecutionData;
326 : :
327 : : /*
328 : : * Context describing the remote transaction whose changes are currently
329 : : * being applied, and the change within it.
330 : : *
331 : : * The remote transaction information (remote_xid and finish_lsn) is set when
332 : : * the transaction's changes begin to be applied. finish_lsn is invalid when
333 : : * the final LSN of the remote transaction is not yet known (e.g. while
334 : : * streaming an in-progress transaction).
335 : : *
336 : : * The remaining fields describe the individual change being applied and are
337 : : * used only for error context reporting.
338 : : */
339 : : typedef struct ApplyRemoteCtx
340 : : {
341 : : LogicalRepMsgType command; /* 0 if invalid */
342 : : LogicalRepRelMapEntry *rel;
343 : :
344 : : /* Remote node information */
345 : : int remote_attnum; /* -1 if invalid */
346 : : TransactionId remote_xid;
347 : : XLogRecPtr finish_lsn;
348 : : char *origin_name;
349 : : } ApplyRemoteCtx;
350 : :
351 : : /*
352 : : * The action to be taken for the changes in the transaction.
353 : : *
354 : : * TRANS_LEADER_APPLY:
355 : : * This action means that we are in the leader apply worker or table sync
356 : : * worker. The changes of the transaction are either directly applied or
357 : : * are read from temporary files (for streaming transactions) and then
358 : : * applied by the worker.
359 : : *
360 : : * TRANS_LEADER_SERIALIZE:
361 : : * This action means that we are in the leader apply worker or table sync
362 : : * worker. Changes are written to temporary files and then applied when the
363 : : * final commit arrives.
364 : : *
365 : : * TRANS_LEADER_SEND_TO_PARALLEL:
366 : : * This action means that we are in the leader apply worker and need to send
367 : : * the changes to the parallel apply worker.
368 : : *
369 : : * TRANS_LEADER_PARTIAL_SERIALIZE:
370 : : * This action means that we are in the leader apply worker and have sent some
371 : : * changes directly to the parallel apply worker and the remaining changes are
372 : : * serialized to a file, due to timeout while sending data. The parallel apply
373 : : * worker will apply these serialized changes when the final commit arrives.
374 : : *
375 : : * We can't use TRANS_LEADER_SERIALIZE for this case because, in addition to
376 : : * serializing changes, the leader worker also needs to serialize the
377 : : * STREAM_XXX message to a file, and wait for the parallel apply worker to
378 : : * finish the transaction when processing the transaction finish command. So
379 : : * this new action was introduced to keep the code and logic clear.
380 : : *
381 : : * TRANS_PARALLEL_APPLY:
382 : : * This action means that we are in the parallel apply worker and changes of
383 : : * the transaction are applied directly by the worker.
384 : : */
385 : : typedef enum
386 : : {
387 : : /* The action for non-streaming transactions. */
388 : : TRANS_LEADER_APPLY,
389 : :
390 : : /* Actions for streaming transactions. */
391 : : TRANS_LEADER_SERIALIZE,
392 : : TRANS_LEADER_SEND_TO_PARALLEL,
393 : : TRANS_LEADER_PARTIAL_SERIALIZE,
394 : : TRANS_PARALLEL_APPLY,
395 : : } TransApplyAction;
396 : :
397 : : /*
398 : : * The phases involved in advancing the non-removable transaction ID.
399 : : *
400 : : * See comments atop worker.c for details of the transition between these
401 : : * phases.
402 : : */
403 : : typedef enum
404 : : {
405 : : RDT_GET_CANDIDATE_XID,
406 : : RDT_REQUEST_PUBLISHER_STATUS,
407 : : RDT_WAIT_FOR_PUBLISHER_STATUS,
408 : : RDT_WAIT_FOR_LOCAL_FLUSH,
409 : : RDT_STOP_CONFLICT_INFO_RETENTION,
410 : : RDT_RESUME_CONFLICT_INFO_RETENTION,
411 : : } RetainDeadTuplesPhase;
412 : :
413 : : /*
414 : : * Critical information for managing phase transitions within the
415 : : * RetainDeadTuplesPhase.
416 : : */
417 : : typedef struct RetainDeadTuplesData
418 : : {
419 : : RetainDeadTuplesPhase phase; /* current phase */
420 : : XLogRecPtr remote_lsn; /* WAL write position on the publisher */
421 : :
422 : : /*
423 : : * Oldest transaction ID that was in the commit phase on the publisher.
424 : : * Use FullTransactionId to prevent issues with transaction ID wraparound,
425 : : * where a new remote_oldestxid could falsely appear to originate from the
426 : : * past and block advancement.
427 : : */
428 : : FullTransactionId remote_oldestxid;
429 : :
430 : : /*
431 : : * Next transaction ID to be assigned on the publisher. Use
432 : : * FullTransactionId for consistency and to allow straightforward
433 : : * comparisons with remote_oldestxid.
434 : : */
435 : : FullTransactionId remote_nextxid;
436 : :
437 : : TimestampTz reply_time; /* when the publisher responds with status */
438 : :
439 : : /*
440 : : * Publisher transaction ID that must be awaited to complete before
441 : : * entering the final phase (RDT_WAIT_FOR_LOCAL_FLUSH). Use
442 : : * FullTransactionId for the same reason as remote_nextxid.
443 : : */
444 : : FullTransactionId remote_wait_for;
445 : :
446 : : TransactionId candidate_xid; /* candidate for the non-removable
447 : : * transaction ID */
448 : : TimestampTz flushpos_update_time; /* when the remote flush position was
449 : : * updated in final phase
450 : : * (RDT_WAIT_FOR_LOCAL_FLUSH) */
451 : :
452 : : long table_sync_wait_time; /* time spent waiting for table sync
453 : : * to finish */
454 : :
455 : : /*
456 : : * The following fields are used to determine the timing for the next
457 : : * round of transaction ID advancement.
458 : : */
459 : : TimestampTz last_recv_time; /* when the last message was received */
460 : : TimestampTz candidate_xid_time; /* when the candidate_xid is decided */
461 : : int xid_advance_interval; /* how much time (ms) to wait before
462 : : * attempting to advance the
463 : : * non-removable transaction ID */
464 : : } RetainDeadTuplesData;
465 : :
466 : : /*
467 : : * The minimum (100ms) and maximum (3 minutes) intervals for advancing
468 : : * non-removable transaction IDs. The maximum interval is a bit arbitrary but
469 : : * is sufficient to not cause any undue network traffic.
470 : : */
471 : : #define MIN_XID_ADVANCE_INTERVAL 100
472 : : #define MAX_XID_ADVANCE_INTERVAL 180000
473 : :
474 : : /* Context of the remote transaction being applied */
475 : : static ApplyRemoteCtx remote_ctx =
476 : : {
477 : : .command = 0,
478 : : .rel = NULL,
479 : : .remote_attnum = -1,
480 : : .remote_xid = InvalidTransactionId,
481 : : .finish_lsn = InvalidXLogRecPtr,
482 : : .origin_name = NULL,
483 : : };
484 : :
485 : : ErrorContextCallback *apply_error_context_stack = NULL;
486 : :
487 : : MemoryContext ApplyMessageContext = NULL;
488 : : MemoryContext ApplyContext = NULL;
489 : :
490 : : /* per stream context for streaming transactions */
491 : : static MemoryContext LogicalStreamingContext = NULL;
492 : :
493 : : WalReceiverConn *LogRepWorkerWalRcvConn = NULL;
494 : :
495 : : Subscription *MySubscription = NULL;
496 : : char *MySubscriptionConninfo = NULL;
497 : : static bool MySubscriptionValid = false;
498 : :
499 : : static List *on_commit_wakeup_workers_subids = NIL;
500 : :
501 : : bool in_remote_transaction = false;
502 : :
503 : : /* fields valid only when processing streamed transaction */
504 : : static bool in_streamed_transaction = false;
505 : :
506 : : static TransactionId stream_xid = InvalidTransactionId;
507 : :
508 : : /*
509 : : * The number of changes applied by parallel apply worker during one streaming
510 : : * block.
511 : : */
512 : : static uint32 parallel_stream_nchanges = 0;
513 : :
514 : : /* Are we initializing an apply worker? */
515 : : bool InitializingApplyWorker = false;
516 : :
517 : : /*
518 : : * We enable skipping all data modification changes (INSERT, UPDATE, etc.) for
519 : : * the subscription if the remote transaction's finish LSN matches the subskiplsn.
520 : : * Once we start skipping changes, we don't stop it until we skip all changes of
521 : : * the transaction even if pg_subscription is updated and MySubscription->skiplsn
522 : : * gets changed or reset during that. Also, in streaming transaction cases (streaming = on),
523 : : * we don't skip receiving and spooling the changes since we decide whether or not
524 : : * to skip applying the changes when starting to apply changes. The subskiplsn is
525 : : * cleared after successfully skipping the transaction or applying non-empty
526 : : * transaction. The latter prevents the mistakenly specified subskiplsn from
527 : : * being left. Note that we cannot skip the streaming transactions when using
528 : : * parallel apply workers because we cannot get the finish LSN before applying
529 : : * the changes. So, we don't start parallel apply worker when finish LSN is set
530 : : * by the user.
531 : : */
532 : : static XLogRecPtr skip_xact_finish_lsn = InvalidXLogRecPtr;
533 : : #define is_skipping_changes() (unlikely(XLogRecPtrIsValid(skip_xact_finish_lsn)))
534 : :
535 : : /* BufFile handle of the current streaming file */
536 : : static BufFile *stream_fd = NULL;
537 : :
538 : : /*
539 : : * The remote WAL position that has been applied and flushed locally. We record
540 : : * and use this information both while sending feedback to the server and
541 : : * advancing oldest_nonremovable_xid.
542 : : */
543 : : static XLogRecPtr last_flushpos = InvalidXLogRecPtr;
544 : :
545 : : typedef struct SubXactInfo
546 : : {
547 : : TransactionId xid; /* XID of the subxact */
548 : : int fileno; /* file number in the buffile */
549 : : pgoff_t offset; /* offset in the file */
550 : : } SubXactInfo;
551 : :
552 : : /* Sub-transaction data for the current streaming transaction */
553 : : typedef struct ApplySubXactData
554 : : {
555 : : uint32 nsubxacts; /* number of sub-transactions */
556 : : uint32 nsubxacts_max; /* current capacity of subxacts */
557 : : TransactionId subxact_last; /* xid of the last sub-transaction */
558 : : SubXactInfo *subxacts; /* sub-xact offset in changes file */
559 : : } ApplySubXactData;
560 : :
561 : : static ApplySubXactData subxact_data = {0, 0, InvalidTransactionId, NULL};
562 : :
563 : : static inline void subxact_filename(char *path, Oid subid, TransactionId xid);
564 : : static inline void changes_filename(char *path, Oid subid, TransactionId xid);
565 : :
566 : : /*
567 : : * Information about subtransactions of a given toplevel transaction.
568 : : */
569 : : static void subxact_info_write(Oid subid, TransactionId xid);
570 : : static void subxact_info_read(Oid subid, TransactionId xid);
571 : : static void subxact_info_add(TransactionId xid);
572 : : static inline void cleanup_subxact_info(void);
573 : :
574 : : /*
575 : : * Serialize and deserialize changes for a toplevel transaction.
576 : : */
577 : : static void stream_open_file(Oid subid, TransactionId xid,
578 : : bool first_segment);
579 : : static void stream_write_change(char action, StringInfo s);
580 : : static void stream_open_and_write_change(TransactionId xid, char action, StringInfo s);
581 : : static void stream_close_file(void);
582 : :
583 : : static void send_feedback(XLogRecPtr recvpos, bool force, bool requestReply);
584 : :
585 : : static void maybe_advance_nonremovable_xid(RetainDeadTuplesData *rdt_data,
586 : : bool status_received);
587 : : static bool can_advance_nonremovable_xid(RetainDeadTuplesData *rdt_data);
588 : : static void process_rdt_phase_transition(RetainDeadTuplesData *rdt_data,
589 : : bool status_received);
590 : : static void get_candidate_xid(RetainDeadTuplesData *rdt_data);
591 : : static void request_publisher_status(RetainDeadTuplesData *rdt_data);
592 : : static void wait_for_publisher_status(RetainDeadTuplesData *rdt_data,
593 : : bool status_received);
594 : : static void wait_for_local_flush(RetainDeadTuplesData *rdt_data);
595 : : static bool should_stop_conflict_info_retention(RetainDeadTuplesData *rdt_data);
596 : : static void stop_conflict_info_retention(RetainDeadTuplesData *rdt_data);
597 : : static void resume_conflict_info_retention(RetainDeadTuplesData *rdt_data);
598 : : static bool update_retention_status(bool active);
599 : : static void reset_retention_data_fields(RetainDeadTuplesData *rdt_data);
600 : : static void adjust_xid_advance_interval(RetainDeadTuplesData *rdt_data,
601 : : bool new_xid_found);
602 : :
603 : : static void apply_worker_exit(void);
604 : :
605 : : static void apply_handle_commit_internal(LogicalRepCommitData *commit_data);
606 : : static void apply_handle_insert_internal(ApplyExecutionData *edata,
607 : : ResultRelInfo *relinfo,
608 : : TupleTableSlot *remoteslot);
609 : : static void apply_handle_update_internal(ApplyExecutionData *edata,
610 : : ResultRelInfo *relinfo,
611 : : TupleTableSlot *remoteslot,
612 : : LogicalRepTupleData *newtup,
613 : : Oid localindexoid);
614 : : static void apply_handle_delete_internal(ApplyExecutionData *edata,
615 : : ResultRelInfo *relinfo,
616 : : TupleTableSlot *remoteslot,
617 : : Oid localindexoid);
618 : : static bool FindReplTupleInLocalRel(ApplyExecutionData *edata, Relation localrel,
619 : : LogicalRepRelation *remoterel,
620 : : Oid localidxoid,
621 : : TupleTableSlot *remoteslot,
622 : : TupleTableSlot **localslot);
623 : : static bool FindDeletedTupleInLocalRel(Relation localrel,
624 : : Oid localidxoid,
625 : : TupleTableSlot *remoteslot,
626 : : TransactionId *delete_xid,
627 : : ReplOriginId *delete_origin,
628 : : TimestampTz *delete_time);
629 : : static void apply_handle_tuple_routing(ApplyExecutionData *edata,
630 : : TupleTableSlot *remoteslot,
631 : : LogicalRepTupleData *newtup,
632 : : CmdType operation);
633 : :
634 : : /* Functions for skipping changes */
635 : : static void maybe_start_skipping_changes(XLogRecPtr finish_lsn);
636 : : static void stop_skipping_changes(void);
637 : : static void clear_subscription_skip_lsn(XLogRecPtr finish_lsn);
638 : :
639 : : /* Functions to maintain the context of the remote transaction being applied */
640 : : static inline void set_remote_transaction_info(TransactionId xid, XLogRecPtr lsn);
641 : : static inline void reset_apply_remote_context(void);
642 : :
643 : : static TransApplyAction get_transaction_apply_action(TransactionId xid,
644 : : ParallelApplyWorkerInfo **winfo);
645 : :
646 : : static void set_wal_receiver_timeout(void);
647 : :
648 : : static void on_exit_clear_xact_state(int code, Datum arg);
649 : :
650 : : /*
651 : : * Form the origin name for the subscription.
652 : : *
653 : : * This is a common function for tablesync and other workers. Tablesync workers
654 : : * must pass a valid relid. Other callers must pass relid = InvalidOid.
655 : : *
656 : : * Return the name in the supplied buffer.
657 : : */
658 : : void
659 : 1554 : ReplicationOriginNameForLogicalRep(Oid suboid, Oid relid,
660 : : char *originname, Size szoriginname)
661 : : {
662 [ + + ]: 1554 : if (OidIsValid(relid))
663 : : {
664 : : /* Replication origin name for tablesync workers. */
665 : 816 : snprintf(originname, szoriginname, "pg_%u_%u", suboid, relid);
666 : : }
667 : : else
668 : : {
669 : : /* Replication origin name for non-tablesync workers. */
670 : 738 : snprintf(originname, szoriginname, "pg_%u", suboid);
671 : : }
672 : 1554 : }
673 : :
674 : : /*
675 : : * Should this worker apply changes for given relation.
676 : : *
677 : : * This is mainly needed for initial relation data sync as that runs in
678 : : * separate worker process running in parallel and we need some way to skip
679 : : * changes coming to the leader apply worker during the sync of a table.
680 : : *
681 : : * Note we need to do smaller or equals comparison for SYNCDONE state because
682 : : * it might hold position of end of initial slot consistent point WAL
683 : : * record + 1 (ie start of next record) and next record can be COMMIT of
684 : : * transaction we are now processing (which is what we set the finish LSN of
685 : : * the remote transaction context to in apply_handle_begin).
686 : : *
687 : : * Note that for streaming transactions that are being applied in the parallel
688 : : * apply worker, we disallow applying changes if the target table in the
689 : : * subscription is not in the READY state, because we cannot decide whether to
690 : : * apply the change as we won't know the finish LSN of the transaction by
691 : : * that time.
692 : : *
693 : : * We already checked this in pa_can_start() before assigning the
694 : : * streaming transaction to the parallel worker, but it also needs to be
695 : : * checked here because if the user executes ALTER SUBSCRIPTION ... REFRESH
696 : : * PUBLICATION in parallel, the new table can be added to pg_subscription_rel
697 : : * while applying this transaction.
698 : : */
699 : : static bool
700 : 171858 : should_apply_changes_for_rel(LogicalRepRelMapEntry *rel)
701 : : {
702 [ + + + - : 171858 : switch (MyLogicalRepWorker->type)
- - ]
703 : : {
704 : 3010 : case WORKERTYPE_TABLESYNC:
705 : 3010 : return MyLogicalRepWorker->relid == rel->localreloid;
706 : :
707 : 68367 : case WORKERTYPE_PARALLEL_APPLY:
708 : : /* We don't synchronize rel's that are in unknown state. */
709 [ - + ]: 68367 : if (rel->state != SUBREL_STATE_READY &&
710 [ # # ]: 0 : rel->state != SUBREL_STATE_UNKNOWN)
711 [ # # ]: 0 : ereport(ERROR,
712 : : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
713 : : errmsg("logical replication parallel apply worker for subscription \"%s\" will stop",
714 : : MySubscription->name),
715 : : errdetail("Cannot handle streamed replication transactions using parallel apply workers until all tables have been synchronized.")));
716 : :
717 : 68367 : return rel->state == SUBREL_STATE_READY;
718 : :
719 : 100481 : case WORKERTYPE_APPLY:
720 [ + + ]: 100545 : return (rel->state == SUBREL_STATE_READY ||
721 [ + + ]: 64 : (rel->state == SUBREL_STATE_SYNCDONE &&
722 [ + - ]: 5 : rel->statelsn <= remote_ctx.finish_lsn));
723 : :
724 : 0 : case WORKERTYPE_SEQUENCESYNC:
725 : : /* Should never happen. */
726 [ # # ]: 0 : elog(ERROR, "sequence synchronization worker is not expected to apply changes");
727 : : break;
728 : :
729 : 0 : case WORKERTYPE_UNKNOWN:
730 : : /* Should never happen. */
731 [ # # ]: 0 : elog(ERROR, "Unknown worker type");
732 : : }
733 : :
734 : 0 : return false; /* dummy for compiler */
735 : : }
736 : :
737 : : /*
738 : : * Begin one step (one INSERT, UPDATE, etc) of a replication transaction.
739 : : *
740 : : * Start a transaction, if this is the first step (else we keep using the
741 : : * existing transaction).
742 : : * Also provide a global snapshot and ensure we run in ApplyMessageContext.
743 : : */
744 : : static void
745 : 172325 : begin_replication_step(void)
746 : : {
747 : 172325 : SetCurrentStatementStartTimestamp();
748 : :
749 [ + + ]: 172325 : if (!IsTransactionState())
750 : : {
751 : 1024 : StartTransactionCommand();
752 : 1024 : maybe_reread_subscription();
753 : : }
754 : :
755 : 172323 : PushActiveSnapshot(GetTransactionSnapshot());
756 : :
757 : 172323 : MemoryContextSwitchTo(ApplyMessageContext);
758 : 172323 : }
759 : :
760 : : /*
761 : : * Finish up one step of a replication transaction.
762 : : * Callers of begin_replication_step() must also call this.
763 : : *
764 : : * We don't close out the transaction here, but we should increment
765 : : * the command counter to make the effects of this step visible.
766 : : */
767 : : static void
768 : 172247 : end_replication_step(void)
769 : : {
770 : 172247 : PopActiveSnapshot();
771 : :
772 : 172247 : CommandCounterIncrement();
773 : 172247 : }
774 : :
775 : : /*
776 : : * Handle streamed transactions for both the leader apply worker and the
777 : : * parallel apply workers.
778 : : *
779 : : * In the streaming case (receiving a block of the streamed transaction), for
780 : : * serialize mode, simply redirect it to a file for the proper toplevel
781 : : * transaction, and for parallel mode, the leader apply worker will send the
782 : : * changes to parallel apply workers and the parallel apply worker will define
783 : : * savepoints if needed. (LOGICAL_REP_MSG_RELATION or LOGICAL_REP_MSG_TYPE
784 : : * messages will be applied by both leader apply worker and parallel apply
785 : : * workers).
786 : : *
787 : : * Returns true for streamed transactions (when the change is either serialized
788 : : * to file or sent to parallel apply worker), false otherwise (regular mode or
789 : : * needs to be processed by parallel apply worker).
790 : : *
791 : : * Exception: If the message being processed is LOGICAL_REP_MSG_RELATION
792 : : * or LOGICAL_REP_MSG_TYPE, return false even if the message needs to be sent
793 : : * to a parallel apply worker.
794 : : */
795 : : static bool
796 : 348248 : handle_streamed_transaction(LogicalRepMsgType action, StringInfo s)
797 : : {
798 : : TransactionId current_xid;
799 : : ParallelApplyWorkerInfo *winfo;
800 : : TransApplyAction apply_action;
801 : : StringInfoData original_msg;
802 : :
803 : 348248 : apply_action = get_transaction_apply_action(stream_xid, &winfo);
804 : :
805 : : /* not in streaming mode */
806 [ + + ]: 348248 : if (apply_action == TRANS_LEADER_APPLY)
807 : 103942 : return false;
808 : :
809 : : Assert(TransactionIdIsValid(stream_xid));
810 : :
811 : : /*
812 : : * The parallel apply worker needs the xid in this message to decide
813 : : * whether to define a savepoint, so save the original message that has
814 : : * not moved the cursor after the xid. We will serialize this message to a
815 : : * file in PARTIAL_SERIALIZE mode.
816 : : */
817 : 244306 : original_msg = *s;
818 : :
819 : : /*
820 : : * We should have received XID of the subxact as the first part of the
821 : : * message, so extract it.
822 : : */
823 : 244306 : current_xid = pq_getmsgint(s, 4);
824 : :
825 [ - + ]: 244306 : if (!TransactionIdIsValid(current_xid))
826 [ # # ]: 0 : ereport(ERROR,
827 : : (errcode(ERRCODE_PROTOCOL_VIOLATION),
828 : : errmsg_internal("invalid transaction ID in streamed replication transaction")));
829 : :
830 [ + + + + : 244306 : switch (apply_action)
- ]
831 : : {
832 : 102513 : case TRANS_LEADER_SERIALIZE:
833 : : Assert(stream_fd);
834 : :
835 : : /* Add the new subxact to the array (unless already there). */
836 : 102513 : subxact_info_add(current_xid);
837 : :
838 : : /* Write the change to the current file */
839 : 102513 : stream_write_change(action, s);
840 : 102513 : return true;
841 : :
842 : 68390 : case TRANS_LEADER_SEND_TO_PARALLEL:
843 : : Assert(winfo);
844 : :
845 : : /*
846 : : * XXX The publisher side doesn't always send relation/type update
847 : : * messages after the streaming transaction, so also update the
848 : : * relation/type in leader apply worker. See function
849 : : * cleanup_rel_sync_cache.
850 : : */
851 [ + - ]: 68390 : if (pa_send_data(winfo, s->len, s->data))
852 [ + + + - ]: 68390 : return (action != LOGICAL_REP_MSG_RELATION &&
853 : : action != LOGICAL_REP_MSG_TYPE);
854 : :
855 : : /*
856 : : * Switch to serialize mode when we are not able to send the
857 : : * change to parallel apply worker.
858 : : */
859 : 0 : pa_switch_to_partial_serialize(winfo, false);
860 : :
861 : : pg_fallthrough;
862 : 5006 : case TRANS_LEADER_PARTIAL_SERIALIZE:
863 : 5006 : stream_write_change(action, &original_msg);
864 : :
865 : : /* Same reason as TRANS_LEADER_SEND_TO_PARALLEL case. */
866 [ + + + - ]: 5006 : return (action != LOGICAL_REP_MSG_RELATION &&
867 : : action != LOGICAL_REP_MSG_TYPE);
868 : :
869 : 68397 : case TRANS_PARALLEL_APPLY:
870 : 68397 : parallel_stream_nchanges += 1;
871 : :
872 : : /* Define a savepoint for a subxact if needed. */
873 : 68397 : pa_start_subtrans(current_xid, stream_xid);
874 : 68397 : return false;
875 : :
876 : 0 : default:
877 [ # # ]: 0 : elog(ERROR, "unexpected apply action: %d", (int) apply_action);
878 : : return false; /* silence compiler warning */
879 : : }
880 : : }
881 : :
882 : : /*
883 : : * Executor state preparation for evaluation of constraint expressions,
884 : : * indexes and triggers for the specified relation.
885 : : *
886 : : * Note that the caller must open and close any indexes to be updated.
887 : : */
888 : : static ApplyExecutionData *
889 : 171769 : create_edata_for_relation(LogicalRepRelMapEntry *rel)
890 : : {
891 : : ApplyExecutionData *edata;
892 : : EState *estate;
893 : : RangeTblEntry *rte;
894 : 171769 : List *perminfos = NIL;
895 : : ResultRelInfo *resultRelInfo;
896 : :
897 : 171769 : edata = palloc0_object(ApplyExecutionData);
898 : 171769 : edata->targetRel = rel;
899 : :
900 : 171769 : edata->estate = estate = CreateExecutorState();
901 : :
902 : 171769 : rte = makeNode(RangeTblEntry);
903 : 171769 : rte->rtekind = RTE_RELATION;
904 : 171769 : rte->relid = RelationGetRelid(rel->localrel);
905 : 171769 : rte->relkind = rel->localrel->rd_rel->relkind;
906 : 171769 : rte->rellockmode = AccessShareLock;
907 : :
908 : 171769 : addRTEPermissionInfo(&perminfos, rte);
909 : :
910 : 171769 : ExecInitRangeTable(estate, list_make1(rte), perminfos,
911 : : bms_make_singleton(1));
912 : :
913 : 171769 : edata->targetRelInfo = resultRelInfo = makeNode(ResultRelInfo);
914 : :
915 : : /*
916 : : * Use Relation opened by logicalrep_rel_open() instead of opening it
917 : : * again.
918 : : */
919 : 171769 : InitResultRelInfo(resultRelInfo, rel->localrel, 1, NULL, 0);
920 : :
921 : : /*
922 : : * We put the ResultRelInfo in the es_opened_result_relations list, even
923 : : * though we don't populate the es_result_relations array. That's a bit
924 : : * bogus, but it's enough to make ExecGetTriggerResultRel() find them.
925 : : *
926 : : * ExecOpenIndices() is not called here either, each execution path doing
927 : : * an apply operation being responsible for that.
928 : : */
929 : 171769 : estate->es_opened_result_relations =
930 : 171769 : lappend(estate->es_opened_result_relations, resultRelInfo);
931 : :
932 : 171769 : estate->es_output_cid = GetCurrentCommandId(true);
933 : :
934 : : /* Prepare to catch AFTER triggers. */
935 : 171769 : AfterTriggerBeginQuery();
936 : :
937 : : /* other fields of edata remain NULL for now */
938 : :
939 : 171769 : return edata;
940 : : }
941 : :
942 : : /*
943 : : * Finish any operations related to the executor state created by
944 : : * create_edata_for_relation().
945 : : */
946 : : static void
947 : 171708 : finish_edata(ApplyExecutionData *edata)
948 : : {
949 : 171708 : EState *estate = edata->estate;
950 : :
951 : : /* Handle any queued AFTER triggers. */
952 : 171708 : AfterTriggerEndQuery(estate);
953 : :
954 : : /* Shut down tuple routing, if any was done. */
955 [ + + ]: 171708 : if (edata->proute)
956 : 74 : ExecCleanupTupleRouting(edata->mtstate, edata->proute);
957 : :
958 : : /*
959 : : * Close relations opened specifically for trigger targets. It might seem
960 : : * that we should call ExecCloseResultRelations() here, but we
961 : : * intentionally don't as that would close the rel we added to
962 : : * es_opened_result_relations above, which is wrong because we took no
963 : : * corresponding refcount. ExecCleanupTupleRouting() closes relations
964 : : * opened for tuple routing, while ExecCloseTrigTargetRelations() closes
965 : : * any relations we opened for AFTER triggers.
966 : : */
967 : 171708 : ExecCloseTrigTargetRelations(estate);
968 : :
969 : 171708 : ExecResetTupleTable(estate->es_tupleTable, false);
970 : 171708 : FreeExecutorState(estate);
971 : 171708 : pfree(edata);
972 : 171708 : }
973 : :
974 : : /*
975 : : * Executes default values for columns for which we can't map to remote
976 : : * relation columns.
977 : : *
978 : : * This allows us to support tables which have more columns on the downstream
979 : : * than on the upstream.
980 : : */
981 : : static void
982 : 99489 : slot_fill_defaults(LogicalRepRelMapEntry *rel, EState *estate,
983 : : TupleTableSlot *slot)
984 : : {
985 : 99489 : TupleDesc desc = RelationGetDescr(rel->localrel);
986 : 99489 : int num_phys_attrs = desc->natts;
987 : : int i;
988 : : int attnum,
989 : 99489 : num_defaults = 0;
990 : : int *defmap;
991 : : ExprState **defexprs;
992 : : ExprContext *econtext;
993 : :
994 [ + - ]: 99489 : econtext = GetPerTupleExprContext(estate);
995 : :
996 : : /* We got all the data via replication, no need to evaluate anything. */
997 [ + + ]: 99489 : if (num_phys_attrs == rel->remoterel.natts)
998 : 59344 : return;
999 : :
1000 : 40145 : defmap = palloc_array(int, num_phys_attrs);
1001 : 40145 : defexprs = palloc_array(ExprState *, num_phys_attrs);
1002 : :
1003 : : Assert(rel->attrmap->maplen == num_phys_attrs);
1004 [ + + ]: 210663 : for (attnum = 0; attnum < num_phys_attrs; attnum++)
1005 : : {
1006 : 170518 : CompactAttribute *cattr = TupleDescCompactAttr(desc, attnum);
1007 : : Expr *defexpr;
1008 : :
1009 [ + - + + ]: 170518 : if (cattr->attisdropped || cattr->attgenerated)
1010 : 9 : continue;
1011 : :
1012 [ + + ]: 170509 : if (rel->attrmap->attnums[attnum] >= 0)
1013 : 92268 : continue;
1014 : :
1015 : 78241 : defexpr = (Expr *) build_column_default(rel->localrel, attnum + 1);
1016 : :
1017 [ + + ]: 78241 : if (defexpr != NULL)
1018 : : {
1019 : : /* Run the expression through planner */
1020 : 70131 : defexpr = expression_planner(defexpr);
1021 : :
1022 : : /* Initialize executable expression in copycontext */
1023 : 70131 : defexprs[num_defaults] = ExecInitExpr(defexpr, NULL);
1024 : 70131 : defmap[num_defaults] = attnum;
1025 : 70131 : num_defaults++;
1026 : : }
1027 : : }
1028 : :
1029 [ + + ]: 110276 : for (i = 0; i < num_defaults; i++)
1030 : 70131 : slot->tts_values[defmap[i]] =
1031 : 70131 : ExecEvalExpr(defexprs[i], econtext, &slot->tts_isnull[defmap[i]]);
1032 : : }
1033 : :
1034 : : /*
1035 : : * Store tuple data into slot.
1036 : : *
1037 : : * Incoming data can be either text or binary format.
1038 : : */
1039 : : static void
1040 : 171787 : slot_store_data(TupleTableSlot *slot, LogicalRepRelMapEntry *rel,
1041 : : LogicalRepTupleData *tupleData)
1042 : : {
1043 : 171787 : int natts = slot->tts_tupleDescriptor->natts;
1044 : : int i;
1045 : :
1046 : 171787 : ExecClearTuple(slot);
1047 : :
1048 : : /* Call the "in" function for each non-dropped, non-null attribute */
1049 : : Assert(natts == rel->attrmap->maplen);
1050 [ + + ]: 705290 : for (i = 0; i < natts; i++)
1051 : : {
1052 : 533503 : Form_pg_attribute att = TupleDescAttr(slot->tts_tupleDescriptor, i);
1053 : 533503 : int remoteattnum = rel->attrmap->attnums[i];
1054 : :
1055 [ + + + + ]: 533503 : if (!att->attisdropped && remoteattnum >= 0)
1056 : 326586 : {
1057 : : StringInfo colvalue;
1058 : :
1059 [ - + ]: 326586 : if (remoteattnum >= tupleData->ncols)
1060 [ # # ]: 0 : ereport(ERROR,
1061 : : (errcode(ERRCODE_PROTOCOL_VIOLATION),
1062 : : errmsg("logical replication column %d not found in tuple: only %d column(s) received",
1063 : : remoteattnum + 1, tupleData->ncols)));
1064 : :
1065 : 326586 : colvalue = &tupleData->colvalues[remoteattnum];
1066 : :
1067 : : /* Set attnum for error callback */
1068 : 326586 : remote_ctx.remote_attnum = remoteattnum;
1069 : :
1070 [ + + ]: 326586 : if (tupleData->colstatus[remoteattnum] == LOGICALREP_COLUMN_TEXT)
1071 : : {
1072 : : Oid typinput;
1073 : : Oid typioparam;
1074 : :
1075 : 166262 : getTypeInputInfo(att->atttypid, &typinput, &typioparam);
1076 : 332524 : slot->tts_values[i] =
1077 : 166262 : OidInputFunctionCall(typinput, colvalue->data,
1078 : : typioparam, att->atttypmod);
1079 : 166262 : slot->tts_isnull[i] = false;
1080 : : }
1081 [ + + ]: 160324 : else if (tupleData->colstatus[remoteattnum] == LOGICALREP_COLUMN_BINARY)
1082 : : {
1083 : : Oid typreceive;
1084 : : Oid typioparam;
1085 : :
1086 : : /*
1087 : : * In some code paths we may be asked to re-parse the same
1088 : : * tuple data. Reset the StringInfo's cursor so that works.
1089 : : */
1090 : 109980 : colvalue->cursor = 0;
1091 : :
1092 : 109980 : getTypeBinaryInputInfo(att->atttypid, &typreceive, &typioparam);
1093 : 219960 : slot->tts_values[i] =
1094 : 109980 : OidReceiveFunctionCall(typreceive, colvalue,
1095 : : typioparam, att->atttypmod);
1096 : :
1097 : : /* Trouble if it didn't eat the whole buffer */
1098 [ - + ]: 109980 : if (colvalue->cursor != colvalue->len)
1099 [ # # ]: 0 : ereport(ERROR,
1100 : : (errcode(ERRCODE_INVALID_BINARY_REPRESENTATION),
1101 : : errmsg("incorrect binary data format in logical replication column %d",
1102 : : remoteattnum + 1)));
1103 : 109980 : slot->tts_isnull[i] = false;
1104 : : }
1105 : : else
1106 : : {
1107 : : /*
1108 : : * NULL value from remote. (We don't expect to see
1109 : : * LOGICALREP_COLUMN_UNCHANGED here, but if we do, treat it as
1110 : : * NULL.)
1111 : : */
1112 : 50344 : slot->tts_values[i] = (Datum) 0;
1113 : 50344 : slot->tts_isnull[i] = true;
1114 : : }
1115 : :
1116 : : /* Reset attnum for error callback */
1117 : 326586 : remote_ctx.remote_attnum = -1;
1118 : : }
1119 : : else
1120 : : {
1121 : : /*
1122 : : * We assign NULL to dropped attributes and missing values
1123 : : * (missing values should be later filled using
1124 : : * slot_fill_defaults).
1125 : : */
1126 : 206917 : slot->tts_values[i] = (Datum) 0;
1127 : 206917 : slot->tts_isnull[i] = true;
1128 : : }
1129 : : }
1130 : :
1131 : 171787 : ExecStoreVirtualTuple(slot);
1132 : 171787 : }
1133 : :
1134 : : /*
1135 : : * Replace updated columns with data from the LogicalRepTupleData struct.
1136 : : * This is somewhat similar to heap_modify_tuple but also calls the type
1137 : : * input functions on the user data.
1138 : : *
1139 : : * "slot" is filled with a copy of the tuple in "srcslot", replacing
1140 : : * columns provided in "tupleData" and leaving others as-is.
1141 : : *
1142 : : * Caution: unreplaced pass-by-ref columns in "slot" will point into the
1143 : : * storage for "srcslot". This is OK for current usage, but someday we may
1144 : : * need to materialize "slot" at the end to make it independent of "srcslot".
1145 : : */
1146 : : static void
1147 : 31931 : slot_modify_data(TupleTableSlot *slot, TupleTableSlot *srcslot,
1148 : : LogicalRepRelMapEntry *rel,
1149 : : LogicalRepTupleData *tupleData)
1150 : : {
1151 : 31931 : int natts = slot->tts_tupleDescriptor->natts;
1152 : : int i;
1153 : :
1154 : : /* We'll fill "slot" with a virtual tuple, so we must start with ... */
1155 : 31931 : ExecClearTuple(slot);
1156 : :
1157 : : /*
1158 : : * Copy all the column data from srcslot, so that we'll have valid values
1159 : : * for unreplaced columns.
1160 : : */
1161 : : Assert(natts == srcslot->tts_tupleDescriptor->natts);
1162 : 31931 : slot_getallattrs(srcslot);
1163 : 31931 : memcpy(slot->tts_values, srcslot->tts_values, natts * sizeof(Datum));
1164 : 31931 : memcpy(slot->tts_isnull, srcslot->tts_isnull, natts * sizeof(bool));
1165 : :
1166 : : /* Call the "in" function for each replaced attribute */
1167 : : Assert(natts == rel->attrmap->maplen);
1168 [ + + ]: 159307 : for (i = 0; i < natts; i++)
1169 : : {
1170 : 127376 : Form_pg_attribute att = TupleDescAttr(slot->tts_tupleDescriptor, i);
1171 : 127376 : int remoteattnum = rel->attrmap->attnums[i];
1172 : :
1173 [ + + ]: 127376 : if (remoteattnum < 0)
1174 : 58519 : continue;
1175 : :
1176 [ - + ]: 68857 : if (remoteattnum >= tupleData->ncols)
1177 [ # # ]: 0 : ereport(ERROR,
1178 : : (errcode(ERRCODE_PROTOCOL_VIOLATION),
1179 : : errmsg("logical replication column %d not found in tuple: only %d column(s) received",
1180 : : remoteattnum + 1, tupleData->ncols)));
1181 : :
1182 [ + - ]: 68857 : if (tupleData->colstatus[remoteattnum] != LOGICALREP_COLUMN_UNCHANGED)
1183 : : {
1184 : 68857 : StringInfo colvalue = &tupleData->colvalues[remoteattnum];
1185 : :
1186 : : /* Set attnum for error callback */
1187 : 68857 : remote_ctx.remote_attnum = remoteattnum;
1188 : :
1189 [ + + ]: 68857 : if (tupleData->colstatus[remoteattnum] == LOGICALREP_COLUMN_TEXT)
1190 : : {
1191 : : Oid typinput;
1192 : : Oid typioparam;
1193 : :
1194 : 25453 : getTypeInputInfo(att->atttypid, &typinput, &typioparam);
1195 : 50906 : slot->tts_values[i] =
1196 : 25453 : OidInputFunctionCall(typinput, colvalue->data,
1197 : : typioparam, att->atttypmod);
1198 : 25453 : slot->tts_isnull[i] = false;
1199 : : }
1200 [ + + ]: 43404 : else if (tupleData->colstatus[remoteattnum] == LOGICALREP_COLUMN_BINARY)
1201 : : {
1202 : : Oid typreceive;
1203 : : Oid typioparam;
1204 : :
1205 : : /*
1206 : : * In some code paths we may be asked to re-parse the same
1207 : : * tuple data. Reset the StringInfo's cursor so that works.
1208 : : */
1209 : 43356 : colvalue->cursor = 0;
1210 : :
1211 : 43356 : getTypeBinaryInputInfo(att->atttypid, &typreceive, &typioparam);
1212 : 86712 : slot->tts_values[i] =
1213 : 43356 : OidReceiveFunctionCall(typreceive, colvalue,
1214 : : typioparam, att->atttypmod);
1215 : :
1216 : : /* Trouble if it didn't eat the whole buffer */
1217 [ - + ]: 43356 : if (colvalue->cursor != colvalue->len)
1218 [ # # ]: 0 : ereport(ERROR,
1219 : : (errcode(ERRCODE_INVALID_BINARY_REPRESENTATION),
1220 : : errmsg("incorrect binary data format in logical replication column %d",
1221 : : remoteattnum + 1)));
1222 : 43356 : slot->tts_isnull[i] = false;
1223 : : }
1224 : : else
1225 : : {
1226 : : /* must be LOGICALREP_COLUMN_NULL */
1227 : 48 : slot->tts_values[i] = (Datum) 0;
1228 : 48 : slot->tts_isnull[i] = true;
1229 : : }
1230 : :
1231 : : /* Reset attnum for error callback */
1232 : 68857 : remote_ctx.remote_attnum = -1;
1233 : : }
1234 : : }
1235 : :
1236 : : /* And finally, declare that "slot" contains a valid virtual tuple */
1237 : 31931 : ExecStoreVirtualTuple(slot);
1238 : 31931 : }
1239 : :
1240 : : /*
1241 : : * Handle BEGIN message.
1242 : : */
1243 : : static void
1244 : 544 : apply_handle_begin(StringInfo s)
1245 : : {
1246 : : LogicalRepBeginData begin_data;
1247 : :
1248 : : /* There must not be an active streaming transaction. */
1249 : : Assert(!TransactionIdIsValid(stream_xid));
1250 : :
1251 : 544 : logicalrep_read_begin(s, &begin_data);
1252 : 544 : set_remote_transaction_info(begin_data.xid, begin_data.final_lsn);
1253 : :
1254 : 544 : maybe_start_skipping_changes(begin_data.final_lsn);
1255 : :
1256 : 544 : in_remote_transaction = true;
1257 : :
1258 : 544 : pgstat_report_activity(STATE_RUNNING, NULL);
1259 : 544 : }
1260 : :
1261 : : /*
1262 : : * Handle COMMIT message.
1263 : : *
1264 : : * TODO, support tracking of multiple origins
1265 : : */
1266 : : static void
1267 : 468 : apply_handle_commit(StringInfo s)
1268 : : {
1269 : : LogicalRepCommitData commit_data;
1270 : :
1271 : 468 : logicalrep_read_commit(s, &commit_data);
1272 : :
1273 [ - + ]: 468 : if (commit_data.commit_lsn != remote_ctx.finish_lsn)
1274 [ # # ]: 0 : ereport(ERROR,
1275 : : (errcode(ERRCODE_PROTOCOL_VIOLATION),
1276 : : errmsg_internal("incorrect commit LSN %X/%08X in commit message (expected %X/%08X)",
1277 : : LSN_FORMAT_ARGS(commit_data.commit_lsn),
1278 : : LSN_FORMAT_ARGS(remote_ctx.finish_lsn))));
1279 : :
1280 : 468 : apply_handle_commit_internal(&commit_data);
1281 : :
1282 : : /*
1283 : : * Process any tables that are being synchronized in parallel, as well as
1284 : : * any newly added tables or sequences.
1285 : : */
1286 : 468 : ProcessSyncingRelations(commit_data.end_lsn);
1287 : :
1288 : 466 : pgstat_report_activity(STATE_IDLE, NULL);
1289 : 466 : reset_apply_remote_context();
1290 : 466 : }
1291 : :
1292 : : /*
1293 : : * Handle BEGIN PREPARE message.
1294 : : */
1295 : : static void
1296 : 18 : apply_handle_begin_prepare(StringInfo s)
1297 : : {
1298 : : LogicalRepPreparedTxnData begin_data;
1299 : :
1300 : : /* Tablesync should never receive prepare. */
1301 [ - + ]: 18 : if (am_tablesync_worker())
1302 [ # # ]: 0 : ereport(ERROR,
1303 : : (errcode(ERRCODE_PROTOCOL_VIOLATION),
1304 : : errmsg_internal("tablesync worker received a BEGIN PREPARE message")));
1305 : :
1306 : : /* There must not be an active streaming transaction. */
1307 : : Assert(!TransactionIdIsValid(stream_xid));
1308 : :
1309 : 18 : logicalrep_read_begin_prepare(s, &begin_data);
1310 : 18 : set_remote_transaction_info(begin_data.xid, begin_data.prepare_lsn);
1311 : :
1312 : 18 : maybe_start_skipping_changes(begin_data.prepare_lsn);
1313 : :
1314 : 18 : in_remote_transaction = true;
1315 : :
1316 : 18 : pgstat_report_activity(STATE_RUNNING, NULL);
1317 : 18 : }
1318 : :
1319 : : /*
1320 : : * Common function to prepare the GID.
1321 : : */
1322 : : static void
1323 : 27 : apply_handle_prepare_internal(LogicalRepPreparedTxnData *prepare_data)
1324 : : {
1325 : : char gid[GIDSIZE];
1326 : :
1327 : : /*
1328 : : * Compute unique GID for two_phase transactions. We don't use GID of
1329 : : * prepared transaction sent by server as that can lead to deadlock when
1330 : : * we have multiple subscriptions from same node point to publications on
1331 : : * the same node. See comments atop worker.c
1332 : : */
1333 : 27 : TwoPhaseTransactionGid(MySubscription->oid, prepare_data->xid,
1334 : : gid, sizeof(gid));
1335 : :
1336 : : /*
1337 : : * BeginTransactionBlock is necessary to balance the EndTransactionBlock
1338 : : * called within the PrepareTransactionBlock below.
1339 : : */
1340 [ + - ]: 27 : if (!IsTransactionBlock())
1341 : : {
1342 : 27 : BeginTransactionBlock();
1343 : 27 : CommitTransactionCommand(); /* Completes the preceding Begin command. */
1344 : : }
1345 : :
1346 : : /*
1347 : : * Update origin state so we can restart streaming from correct position
1348 : : * in case of crash.
1349 : : */
1350 : 27 : replorigin_xact_state.origin_lsn = prepare_data->end_lsn;
1351 : 27 : replorigin_xact_state.origin_timestamp = prepare_data->prepare_time;
1352 : :
1353 : 27 : PrepareTransactionBlock(gid);
1354 : 27 : }
1355 : :
1356 : : /*
1357 : : * Handle PREPARE message.
1358 : : */
1359 : : static void
1360 : 17 : apply_handle_prepare(StringInfo s)
1361 : : {
1362 : : LogicalRepPreparedTxnData prepare_data;
1363 : :
1364 : 17 : logicalrep_read_prepare(s, &prepare_data);
1365 : :
1366 [ - + ]: 17 : if (prepare_data.prepare_lsn != remote_ctx.finish_lsn)
1367 [ # # ]: 0 : ereport(ERROR,
1368 : : (errcode(ERRCODE_PROTOCOL_VIOLATION),
1369 : : errmsg_internal("incorrect prepare LSN %X/%08X in prepare message (expected %X/%08X)",
1370 : : LSN_FORMAT_ARGS(prepare_data.prepare_lsn),
1371 : : LSN_FORMAT_ARGS(remote_ctx.finish_lsn))));
1372 : :
1373 : : /*
1374 : : * Unlike commit, here, we always prepare the transaction even though no
1375 : : * change has happened in this transaction or all changes are skipped. It
1376 : : * is done this way because at commit prepared time, we won't know whether
1377 : : * we have skipped preparing a transaction because of those reasons.
1378 : : *
1379 : : * XXX, We can optimize such that at commit prepared time, we first check
1380 : : * whether we have prepared the transaction or not but that doesn't seem
1381 : : * worthwhile because such cases shouldn't be common.
1382 : : */
1383 : 17 : begin_replication_step();
1384 : :
1385 : 17 : apply_handle_prepare_internal(&prepare_data);
1386 : :
1387 : 17 : end_replication_step();
1388 : 17 : CommitTransactionCommand();
1389 : 15 : pgstat_report_stat(false);
1390 : :
1391 : : /*
1392 : : * It is okay not to set the local_end LSN for the prepare because we
1393 : : * always flush the prepare record. So, we can send the acknowledgment of
1394 : : * the remote_end LSN as soon as prepare is finished.
1395 : : *
1396 : : * XXX For the sake of consistency with commit, we could have set it with
1397 : : * the LSN of prepare but as of now we don't track that value similar to
1398 : : * XactLastCommitEnd, and adding it for this purpose doesn't seems worth
1399 : : * it.
1400 : : */
1401 : 15 : store_flush_position(prepare_data.end_lsn, InvalidXLogRecPtr);
1402 : :
1403 : 15 : in_remote_transaction = false;
1404 : :
1405 : : /*
1406 : : * Process any tables that are being synchronized in parallel, as well as
1407 : : * any newly added tables or sequences.
1408 : : */
1409 : 15 : ProcessSyncingRelations(prepare_data.end_lsn);
1410 : :
1411 : : /*
1412 : : * Since we have already prepared the transaction, in a case where the
1413 : : * server crashes before clearing the subskiplsn, it will be left but the
1414 : : * transaction won't be resent. But that's okay because it's a rare case
1415 : : * and the subskiplsn will be cleared when finishing the next transaction.
1416 : : */
1417 : 15 : stop_skipping_changes();
1418 : 15 : clear_subscription_skip_lsn(prepare_data.prepare_lsn);
1419 : :
1420 : 15 : pgstat_report_activity(STATE_IDLE, NULL);
1421 : 15 : reset_apply_remote_context();
1422 : 15 : }
1423 : :
1424 : : /*
1425 : : * Handle a COMMIT PREPARED of a previously PREPARED transaction.
1426 : : *
1427 : : * Note that we don't need to wait here if the transaction was prepared in a
1428 : : * parallel apply worker. In that case, we have already waited for the prepare
1429 : : * to finish in apply_handle_stream_prepare() which will ensure all the
1430 : : * operations in that transaction have happened in the subscriber, so no
1431 : : * concurrent transaction can cause deadlock or transaction dependency issues.
1432 : : */
1433 : : static void
1434 : 22 : apply_handle_commit_prepared(StringInfo s)
1435 : : {
1436 : : LogicalRepCommitPreparedTxnData prepare_data;
1437 : : char gid[GIDSIZE];
1438 : :
1439 : 22 : logicalrep_read_commit_prepared(s, &prepare_data);
1440 : 22 : set_remote_transaction_info(prepare_data.xid, prepare_data.commit_lsn);
1441 : :
1442 : : /* Compute GID for two_phase transactions. */
1443 : 22 : TwoPhaseTransactionGid(MySubscription->oid, prepare_data.xid,
1444 : : gid, sizeof(gid));
1445 : :
1446 : : /* There is no transaction when COMMIT PREPARED is called */
1447 : 22 : begin_replication_step();
1448 : :
1449 : : /*
1450 : : * Update origin state so we can restart streaming from correct position
1451 : : * in case of crash.
1452 : : */
1453 : 22 : replorigin_xact_state.origin_lsn = prepare_data.end_lsn;
1454 : 22 : replorigin_xact_state.origin_timestamp = prepare_data.commit_time;
1455 : :
1456 : 22 : FinishPreparedTransaction(gid, true);
1457 : 22 : end_replication_step();
1458 : 22 : CommitTransactionCommand();
1459 : 22 : pgstat_report_stat(false);
1460 : :
1461 : 22 : store_flush_position(prepare_data.end_lsn, XactLastCommitEnd);
1462 : 22 : in_remote_transaction = false;
1463 : :
1464 : : /*
1465 : : * Process any tables that are being synchronized in parallel, as well as
1466 : : * any newly added tables or sequences.
1467 : : */
1468 : 22 : ProcessSyncingRelations(prepare_data.end_lsn);
1469 : :
1470 : 22 : clear_subscription_skip_lsn(prepare_data.end_lsn);
1471 : :
1472 : 22 : pgstat_report_activity(STATE_IDLE, NULL);
1473 : 22 : reset_apply_remote_context();
1474 : 22 : }
1475 : :
1476 : : /*
1477 : : * Handle a ROLLBACK PREPARED of a previously PREPARED TRANSACTION.
1478 : : *
1479 : : * Note that we don't need to wait here if the transaction was prepared in a
1480 : : * parallel apply worker. In that case, we have already waited for the prepare
1481 : : * to finish in apply_handle_stream_prepare() which will ensure all the
1482 : : * operations in that transaction have happened in the subscriber, so no
1483 : : * concurrent transaction can cause deadlock or transaction dependency issues.
1484 : : */
1485 : : static void
1486 : 5 : apply_handle_rollback_prepared(StringInfo s)
1487 : : {
1488 : : LogicalRepRollbackPreparedTxnData rollback_data;
1489 : : char gid[GIDSIZE];
1490 : :
1491 : 5 : logicalrep_read_rollback_prepared(s, &rollback_data);
1492 : 5 : set_remote_transaction_info(rollback_data.xid, rollback_data.rollback_end_lsn);
1493 : :
1494 : : /* Compute GID for two_phase transactions. */
1495 : 5 : TwoPhaseTransactionGid(MySubscription->oid, rollback_data.xid,
1496 : : gid, sizeof(gid));
1497 : :
1498 : : /*
1499 : : * It is possible that we haven't received prepare because it occurred
1500 : : * before walsender reached a consistent point or the two_phase was still
1501 : : * not enabled by that time, so in such cases, we need to skip rollback
1502 : : * prepared.
1503 : : */
1504 [ + - ]: 5 : if (LookupGXact(gid, rollback_data.prepare_end_lsn,
1505 : : rollback_data.prepare_time))
1506 : : {
1507 : : /*
1508 : : * Update origin state so we can restart streaming from correct
1509 : : * position in case of crash.
1510 : : */
1511 : 5 : replorigin_xact_state.origin_lsn = rollback_data.rollback_end_lsn;
1512 : 5 : replorigin_xact_state.origin_timestamp = rollback_data.rollback_time;
1513 : :
1514 : : /* There is no transaction when ABORT/ROLLBACK PREPARED is called */
1515 : 5 : begin_replication_step();
1516 : 5 : FinishPreparedTransaction(gid, false);
1517 : 5 : end_replication_step();
1518 : 5 : CommitTransactionCommand();
1519 : :
1520 : 5 : clear_subscription_skip_lsn(rollback_data.rollback_end_lsn);
1521 : : }
1522 : :
1523 : 5 : pgstat_report_stat(false);
1524 : :
1525 : : /*
1526 : : * It is okay not to set the local_end LSN for the rollback of prepared
1527 : : * transaction because we always flush the WAL record for it. See
1528 : : * apply_handle_prepare.
1529 : : */
1530 : 5 : store_flush_position(rollback_data.rollback_end_lsn, InvalidXLogRecPtr);
1531 : 5 : in_remote_transaction = false;
1532 : :
1533 : : /*
1534 : : * Process any tables that are being synchronized in parallel, as well as
1535 : : * any newly added tables or sequences.
1536 : : */
1537 : 5 : ProcessSyncingRelations(rollback_data.rollback_end_lsn);
1538 : :
1539 : 5 : pgstat_report_activity(STATE_IDLE, NULL);
1540 : 5 : reset_apply_remote_context();
1541 : 5 : }
1542 : :
1543 : : /*
1544 : : * Handle STREAM PREPARE.
1545 : : */
1546 : : static void
1547 : 15 : apply_handle_stream_prepare(StringInfo s)
1548 : : {
1549 : : LogicalRepPreparedTxnData prepare_data;
1550 : : ParallelApplyWorkerInfo *winfo;
1551 : : TransApplyAction apply_action;
1552 : :
1553 : : /* Save the message before it is consumed. */
1554 : 15 : StringInfoData original_msg = *s;
1555 : :
1556 [ - + ]: 15 : if (in_streamed_transaction)
1557 [ # # ]: 0 : ereport(ERROR,
1558 : : (errcode(ERRCODE_PROTOCOL_VIOLATION),
1559 : : errmsg_internal("STREAM PREPARE message without STREAM STOP")));
1560 : :
1561 : : /* Tablesync should never receive prepare. */
1562 [ - + ]: 15 : if (am_tablesync_worker())
1563 [ # # ]: 0 : ereport(ERROR,
1564 : : (errcode(ERRCODE_PROTOCOL_VIOLATION),
1565 : : errmsg_internal("tablesync worker received a STREAM PREPARE message")));
1566 : :
1567 : 15 : logicalrep_read_stream_prepare(s, &prepare_data);
1568 : 15 : set_remote_transaction_info(prepare_data.xid, prepare_data.prepare_lsn);
1569 : :
1570 : 15 : apply_action = get_transaction_apply_action(prepare_data.xid, &winfo);
1571 : :
1572 [ + + + + : 15 : switch (apply_action)
- ]
1573 : : {
1574 : 5 : case TRANS_LEADER_APPLY:
1575 : :
1576 : : /*
1577 : : * The transaction has been serialized to file, so replay all the
1578 : : * spooled operations.
1579 : : */
1580 : 5 : apply_spooled_messages(MyLogicalRepWorker->stream_fileset,
1581 : : prepare_data.xid, prepare_data.prepare_lsn);
1582 : :
1583 : : /* Mark the transaction as prepared. */
1584 : 5 : apply_handle_prepare_internal(&prepare_data);
1585 : :
1586 : 5 : CommitTransactionCommand();
1587 : :
1588 : : /*
1589 : : * It is okay not to set the local_end LSN for the prepare because
1590 : : * we always flush the prepare record. See apply_handle_prepare.
1591 : : */
1592 : 5 : store_flush_position(prepare_data.end_lsn, InvalidXLogRecPtr);
1593 : :
1594 : 5 : in_remote_transaction = false;
1595 : :
1596 : : /* Unlink the files with serialized changes and subxact info. */
1597 : 5 : stream_cleanup_files(MyLogicalRepWorker->subid, prepare_data.xid);
1598 : :
1599 [ - + ]: 5 : elog(DEBUG1, "finished processing the STREAM PREPARE command");
1600 : 5 : break;
1601 : :
1602 : 4 : case TRANS_LEADER_SEND_TO_PARALLEL:
1603 : : Assert(winfo);
1604 : :
1605 [ + - ]: 4 : if (pa_send_data(winfo, s->len, s->data))
1606 : : {
1607 : : /* Finish processing the streaming transaction. */
1608 : 4 : pa_xact_finish(winfo, prepare_data.end_lsn);
1609 : 3 : break;
1610 : : }
1611 : :
1612 : : /*
1613 : : * Switch to serialize mode when we are not able to send the
1614 : : * change to parallel apply worker.
1615 : : */
1616 : 0 : pa_switch_to_partial_serialize(winfo, true);
1617 : :
1618 : : pg_fallthrough;
1619 : 1 : case TRANS_LEADER_PARTIAL_SERIALIZE:
1620 : : Assert(winfo);
1621 : :
1622 : 1 : stream_open_and_write_change(prepare_data.xid,
1623 : : LOGICAL_REP_MSG_STREAM_PREPARE,
1624 : : &original_msg);
1625 : :
1626 : 1 : pa_set_fileset_state(winfo->shared, FS_SERIALIZE_DONE);
1627 : :
1628 : : /* Finish processing the streaming transaction. */
1629 : 1 : pa_xact_finish(winfo, prepare_data.end_lsn);
1630 : 1 : break;
1631 : :
1632 : 5 : case TRANS_PARALLEL_APPLY:
1633 : :
1634 : : /*
1635 : : * If the parallel apply worker is applying spooled messages then
1636 : : * close the file before preparing.
1637 : : */
1638 [ + + ]: 5 : if (stream_fd)
1639 : 1 : stream_close_file();
1640 : :
1641 : 5 : begin_replication_step();
1642 : :
1643 : : /* Mark the transaction as prepared. */
1644 : 5 : apply_handle_prepare_internal(&prepare_data);
1645 : :
1646 : 5 : end_replication_step();
1647 : :
1648 : 5 : CommitTransactionCommand();
1649 : :
1650 : : /*
1651 : : * It is okay not to set the local_end LSN for the prepare because
1652 : : * we always flush the prepare record. See apply_handle_prepare.
1653 : : */
1654 : 4 : MyParallelShared->last_commit_end = InvalidXLogRecPtr;
1655 : :
1656 : 4 : pa_set_xact_state(MyParallelShared, PARALLEL_TRANS_FINISHED);
1657 : 4 : pa_unlock_transaction(MyParallelShared->xid, AccessExclusiveLock);
1658 : :
1659 : 4 : pa_reset_subtrans();
1660 : :
1661 [ + + ]: 4 : elog(DEBUG1, "finished processing the STREAM PREPARE command");
1662 : 4 : break;
1663 : :
1664 : 0 : default:
1665 [ # # ]: 0 : elog(ERROR, "unexpected apply action: %d", (int) apply_action);
1666 : : break;
1667 : : }
1668 : :
1669 : 13 : pgstat_report_stat(false);
1670 : :
1671 : : /*
1672 : : * Process any tables that are being synchronized in parallel, as well as
1673 : : * any newly added tables or sequences.
1674 : : */
1675 : 13 : ProcessSyncingRelations(prepare_data.end_lsn);
1676 : :
1677 : : /*
1678 : : * Similar to prepare case, the subskiplsn could be left in a case of
1679 : : * server crash but it's okay. See the comments in apply_handle_prepare().
1680 : : */
1681 : 13 : stop_skipping_changes();
1682 : 13 : clear_subscription_skip_lsn(prepare_data.prepare_lsn);
1683 : :
1684 : 13 : pgstat_report_activity(STATE_IDLE, NULL);
1685 : :
1686 : 13 : reset_apply_remote_context();
1687 : 13 : }
1688 : :
1689 : : /*
1690 : : * Handle ORIGIN message.
1691 : : *
1692 : : * TODO, support tracking of multiple origins
1693 : : */
1694 : : static void
1695 : 7 : apply_handle_origin(StringInfo s)
1696 : : {
1697 : : /*
1698 : : * ORIGIN message can only come inside streaming transaction or inside
1699 : : * remote transaction and before any actual writes.
1700 : : */
1701 [ + + ]: 7 : if (!in_streamed_transaction &&
1702 [ + - - + ]: 10 : (!in_remote_transaction ||
1703 [ - - ]: 5 : (IsTransactionState() && !am_tablesync_worker())))
1704 [ # # ]: 0 : ereport(ERROR,
1705 : : (errcode(ERRCODE_PROTOCOL_VIOLATION),
1706 : : errmsg_internal("ORIGIN message sent out of order")));
1707 : 7 : }
1708 : :
1709 : : /*
1710 : : * Initialize fileset (if not already done).
1711 : : *
1712 : : * Create a new file when first_segment is true, otherwise open the existing
1713 : : * file.
1714 : : */
1715 : : void
1716 : 363 : stream_start_internal(TransactionId xid, bool first_segment)
1717 : : {
1718 : 363 : begin_replication_step();
1719 : :
1720 : : /*
1721 : : * Initialize the worker's stream_fileset if we haven't yet. This will be
1722 : : * used for the entire duration of the worker so create it in a permanent
1723 : : * context. We create this on the very first streaming message from any
1724 : : * transaction and then use it for this and other streaming transactions.
1725 : : * Now, we could create a fileset at the start of the worker as well but
1726 : : * then we won't be sure that it will ever be used.
1727 : : */
1728 [ + + ]: 363 : if (!MyLogicalRepWorker->stream_fileset)
1729 : : {
1730 : : MemoryContext oldctx;
1731 : :
1732 : 14 : oldctx = MemoryContextSwitchTo(ApplyContext);
1733 : :
1734 : 14 : MyLogicalRepWorker->stream_fileset = palloc_object(FileSet);
1735 : 14 : FileSetInit(MyLogicalRepWorker->stream_fileset);
1736 : :
1737 : 14 : MemoryContextSwitchTo(oldctx);
1738 : : }
1739 : :
1740 : : /* Open the spool file for this transaction. */
1741 : 363 : stream_open_file(MyLogicalRepWorker->subid, xid, first_segment);
1742 : :
1743 : : /* If this is not the first segment, open existing subxact file. */
1744 [ + + ]: 363 : if (!first_segment)
1745 : 331 : subxact_info_read(MyLogicalRepWorker->subid, xid);
1746 : :
1747 : 363 : end_replication_step();
1748 : 363 : }
1749 : :
1750 : : /*
1751 : : * Handle STREAM START message.
1752 : : */
1753 : : static void
1754 : 843 : apply_handle_stream_start(StringInfo s)
1755 : : {
1756 : : bool first_segment;
1757 : : ParallelApplyWorkerInfo *winfo;
1758 : : TransApplyAction apply_action;
1759 : :
1760 : : /* Save the message before it is consumed. */
1761 : 843 : StringInfoData original_msg = *s;
1762 : :
1763 [ - + ]: 843 : if (in_streamed_transaction)
1764 [ # # ]: 0 : ereport(ERROR,
1765 : : (errcode(ERRCODE_PROTOCOL_VIOLATION),
1766 : : errmsg_internal("duplicate STREAM START message")));
1767 : :
1768 : : /* There must not be an active streaming transaction. */
1769 : : Assert(!TransactionIdIsValid(stream_xid));
1770 : :
1771 : : /* notify handle methods we're processing a remote transaction */
1772 : 843 : in_streamed_transaction = true;
1773 : :
1774 : : /* extract XID of the top-level transaction */
1775 : 843 : stream_xid = logicalrep_read_stream_start(s, &first_segment);
1776 : :
1777 [ - + ]: 843 : if (!TransactionIdIsValid(stream_xid))
1778 [ # # ]: 0 : ereport(ERROR,
1779 : : (errcode(ERRCODE_PROTOCOL_VIOLATION),
1780 : : errmsg_internal("invalid transaction ID in streamed replication transaction")));
1781 : :
1782 : : /*
1783 : : * The final LSN of the streamed transaction is known only when its commit
1784 : : * record arrives.
1785 : : */
1786 : 843 : set_remote_transaction_info(stream_xid, InvalidXLogRecPtr);
1787 : :
1788 : : /* Try to allocate a worker for the streaming transaction. */
1789 [ + + ]: 843 : if (first_segment)
1790 : 86 : pa_allocate_worker(stream_xid);
1791 : :
1792 : 843 : apply_action = get_transaction_apply_action(stream_xid, &winfo);
1793 : :
1794 [ + + + + : 843 : switch (apply_action)
- ]
1795 : : {
1796 : 343 : case TRANS_LEADER_SERIALIZE:
1797 : :
1798 : : /*
1799 : : * Function stream_start_internal starts a transaction. This
1800 : : * transaction will be committed on the stream stop unless it is a
1801 : : * tablesync worker in which case it will be committed after
1802 : : * processing all the messages. We need this transaction for
1803 : : * handling the BufFile, used for serializing the streaming data
1804 : : * and subxact info.
1805 : : */
1806 : 343 : stream_start_internal(stream_xid, first_segment);
1807 : 343 : break;
1808 : :
1809 : 244 : case TRANS_LEADER_SEND_TO_PARALLEL:
1810 : : Assert(winfo);
1811 : :
1812 : : /*
1813 : : * Once we start serializing the changes, the parallel apply
1814 : : * worker will wait for the leader to release the stream lock
1815 : : * until the end of the transaction. So, we don't need to release
1816 : : * the lock or increment the stream count in that case.
1817 : : */
1818 [ + + ]: 244 : if (pa_send_data(winfo, s->len, s->data))
1819 : : {
1820 : : /*
1821 : : * Unlock the shared object lock so that the parallel apply
1822 : : * worker can continue to receive changes.
1823 : : */
1824 [ + + ]: 240 : if (!first_segment)
1825 : 215 : pa_unlock_stream(winfo->shared->xid, AccessExclusiveLock);
1826 : :
1827 : : /*
1828 : : * Increment the number of streaming blocks waiting to be
1829 : : * processed by parallel apply worker.
1830 : : */
1831 : 240 : pg_atomic_add_fetch_u32(&winfo->shared->pending_stream_count, 1);
1832 : :
1833 : : /* Cache the parallel apply worker for this transaction. */
1834 : 240 : pa_set_stream_apply_worker(winfo);
1835 : 240 : break;
1836 : : }
1837 : :
1838 : : /*
1839 : : * Switch to serialize mode when we are not able to send the
1840 : : * change to parallel apply worker.
1841 : : */
1842 : 4 : pa_switch_to_partial_serialize(winfo, !first_segment);
1843 : :
1844 : : pg_fallthrough;
1845 : 15 : case TRANS_LEADER_PARTIAL_SERIALIZE:
1846 : : Assert(winfo);
1847 : :
1848 : : /*
1849 : : * Open the spool file unless it was already opened when switching
1850 : : * to serialize mode. The transaction started in
1851 : : * stream_start_internal will be committed on the stream stop.
1852 : : */
1853 [ + + ]: 15 : if (apply_action != TRANS_LEADER_SEND_TO_PARALLEL)
1854 : 11 : stream_start_internal(stream_xid, first_segment);
1855 : :
1856 : 15 : stream_write_change(LOGICAL_REP_MSG_STREAM_START, &original_msg);
1857 : :
1858 : : /* Cache the parallel apply worker for this transaction. */
1859 : 15 : pa_set_stream_apply_worker(winfo);
1860 : 15 : break;
1861 : :
1862 : 245 : case TRANS_PARALLEL_APPLY:
1863 [ + + ]: 245 : if (first_segment)
1864 : : {
1865 : : /* Hold the lock until the end of the transaction. */
1866 : 29 : pa_lock_transaction(MyParallelShared->xid, AccessExclusiveLock);
1867 : 29 : pa_set_xact_state(MyParallelShared, PARALLEL_TRANS_STARTED);
1868 : :
1869 : : /*
1870 : : * Signal the leader apply worker, as it may be waiting for
1871 : : * us.
1872 : : */
1873 : 29 : logicalrep_worker_wakeup(WORKERTYPE_APPLY,
1874 : 29 : MyLogicalRepWorker->subid, InvalidOid);
1875 : : }
1876 : :
1877 : 245 : parallel_stream_nchanges = 0;
1878 : 245 : break;
1879 : :
1880 : 0 : default:
1881 [ # # ]: 0 : elog(ERROR, "unexpected apply action: %d", (int) apply_action);
1882 : : break;
1883 : : }
1884 : :
1885 : 843 : pgstat_report_activity(STATE_RUNNING, NULL);
1886 : 843 : }
1887 : :
1888 : : /*
1889 : : * Update the information about subxacts and close the file.
1890 : : *
1891 : : * This function should be called when the stream_start_internal function has
1892 : : * been called.
1893 : : */
1894 : : void
1895 : 363 : stream_stop_internal(TransactionId xid)
1896 : : {
1897 : : /*
1898 : : * Serialize information about subxacts for the toplevel transaction, then
1899 : : * close the stream messages spool file.
1900 : : */
1901 : 363 : subxact_info_write(MyLogicalRepWorker->subid, xid);
1902 : 363 : stream_close_file();
1903 : :
1904 : : /* We must be in a valid transaction state */
1905 : : Assert(IsTransactionState());
1906 : :
1907 : : /* Commit the per-stream transaction */
1908 : 363 : CommitTransactionCommand();
1909 : :
1910 : : /* Reset per-stream context */
1911 : 363 : MemoryContextReset(LogicalStreamingContext);
1912 : 363 : }
1913 : :
1914 : : /*
1915 : : * Handle STREAM STOP message.
1916 : : */
1917 : : static void
1918 : 842 : apply_handle_stream_stop(StringInfo s)
1919 : : {
1920 : : ParallelApplyWorkerInfo *winfo;
1921 : : TransApplyAction apply_action;
1922 : :
1923 [ - + ]: 842 : if (!in_streamed_transaction)
1924 [ # # ]: 0 : ereport(ERROR,
1925 : : (errcode(ERRCODE_PROTOCOL_VIOLATION),
1926 : : errmsg_internal("STREAM STOP message without STREAM START")));
1927 : :
1928 : 842 : apply_action = get_transaction_apply_action(stream_xid, &winfo);
1929 : :
1930 [ + + + + : 842 : switch (apply_action)
- ]
1931 : : {
1932 : 343 : case TRANS_LEADER_SERIALIZE:
1933 : 343 : stream_stop_internal(stream_xid);
1934 : 343 : break;
1935 : :
1936 : 240 : case TRANS_LEADER_SEND_TO_PARALLEL:
1937 : : Assert(winfo);
1938 : :
1939 : : /*
1940 : : * Lock before sending the STREAM_STOP message so that the leader
1941 : : * can hold the lock first and the parallel apply worker will wait
1942 : : * for leader to release the lock. See Locking Considerations atop
1943 : : * applyparallelworker.c.
1944 : : */
1945 : 240 : pa_lock_stream(winfo->shared->xid, AccessExclusiveLock);
1946 : :
1947 [ + - ]: 240 : if (pa_send_data(winfo, s->len, s->data))
1948 : : {
1949 : 240 : pa_set_stream_apply_worker(NULL);
1950 : 240 : break;
1951 : : }
1952 : :
1953 : : /*
1954 : : * Switch to serialize mode when we are not able to send the
1955 : : * change to parallel apply worker.
1956 : : */
1957 : 0 : pa_switch_to_partial_serialize(winfo, true);
1958 : :
1959 : : pg_fallthrough;
1960 : 15 : case TRANS_LEADER_PARTIAL_SERIALIZE:
1961 : 15 : stream_write_change(LOGICAL_REP_MSG_STREAM_STOP, s);
1962 : 15 : stream_stop_internal(stream_xid);
1963 : 15 : pa_set_stream_apply_worker(NULL);
1964 : 15 : break;
1965 : :
1966 : 244 : case TRANS_PARALLEL_APPLY:
1967 [ + + ]: 244 : elog(DEBUG1, "applied %u changes in the streaming chunk",
1968 : : parallel_stream_nchanges);
1969 : :
1970 : : /*
1971 : : * By the time parallel apply worker is processing the changes in
1972 : : * the current streaming block, the leader apply worker may have
1973 : : * sent multiple streaming blocks. This can lead to parallel apply
1974 : : * worker start waiting even when there are more chunk of streams
1975 : : * in the queue. So, try to lock only if there is no message left
1976 : : * in the queue. See Locking Considerations atop
1977 : : * applyparallelworker.c.
1978 : : *
1979 : : * Note that here we have a race condition where we can start
1980 : : * waiting even when there are pending streaming chunks. This can
1981 : : * happen if the leader sends another streaming block and acquires
1982 : : * the stream lock again after the parallel apply worker checks
1983 : : * that there is no pending streaming block and before it actually
1984 : : * starts waiting on a lock. We can handle this case by not
1985 : : * allowing the leader to increment the stream block count during
1986 : : * the time parallel apply worker acquires the lock but it is not
1987 : : * clear whether that is worth the complexity.
1988 : : *
1989 : : * Now, if this missed chunk contains rollback to savepoint, then
1990 : : * there is a risk of deadlock which probably shouldn't happen
1991 : : * after restart.
1992 : : */
1993 : 244 : pa_decr_and_wait_stream_block();
1994 : 242 : break;
1995 : :
1996 : 0 : default:
1997 [ # # ]: 0 : elog(ERROR, "unexpected apply action: %d", (int) apply_action);
1998 : : break;
1999 : : }
2000 : :
2001 : 840 : in_streamed_transaction = false;
2002 : 840 : stream_xid = InvalidTransactionId;
2003 : :
2004 : : /*
2005 : : * The parallel apply worker could be in a transaction in which case we
2006 : : * need to report the state as STATE_IDLEINTRANSACTION.
2007 : : */
2008 [ + + ]: 840 : if (IsTransactionOrTransactionBlock())
2009 : 242 : pgstat_report_activity(STATE_IDLEINTRANSACTION, NULL);
2010 : : else
2011 : 598 : pgstat_report_activity(STATE_IDLE, NULL);
2012 : :
2013 : 840 : reset_apply_remote_context();
2014 : 840 : }
2015 : :
2016 : : /*
2017 : : * Helper function to handle STREAM ABORT message when the transaction was
2018 : : * serialized to file.
2019 : : */
2020 : : static void
2021 : 14 : stream_abort_internal(TransactionId xid, TransactionId subxid)
2022 : : {
2023 : : /*
2024 : : * If the two XIDs are the same, it's in fact abort of toplevel xact, so
2025 : : * just delete the files with serialized info.
2026 : : */
2027 [ + + ]: 14 : if (xid == subxid)
2028 : 1 : stream_cleanup_files(MyLogicalRepWorker->subid, xid);
2029 : : else
2030 : : {
2031 : : /*
2032 : : * OK, so it's a subxact. We need to read the subxact file for the
2033 : : * toplevel transaction, determine the offset tracked for the subxact,
2034 : : * and truncate the file with changes. We also remove the subxacts
2035 : : * with higher offsets (or rather higher XIDs).
2036 : : *
2037 : : * We intentionally scan the array from the tail, because we're likely
2038 : : * aborting a change for the most recent subtransactions.
2039 : : *
2040 : : * We can't use the binary search here as subxact XIDs won't
2041 : : * necessarily arrive in sorted order, consider the case where we have
2042 : : * released the savepoint for multiple subtransactions and then
2043 : : * performed rollback to savepoint for one of the earlier
2044 : : * sub-transaction.
2045 : : */
2046 : : int64 i;
2047 : : int64 subidx;
2048 : : BufFile *fd;
2049 : 13 : bool found = false;
2050 : : char path[MAXPGPATH];
2051 : :
2052 : 13 : subidx = -1;
2053 : 13 : begin_replication_step();
2054 : 13 : subxact_info_read(MyLogicalRepWorker->subid, xid);
2055 : :
2056 [ + + ]: 15 : for (i = subxact_data.nsubxacts; i > 0; i--)
2057 : : {
2058 [ + + ]: 11 : if (subxact_data.subxacts[i - 1].xid == subxid)
2059 : : {
2060 : 9 : subidx = (i - 1);
2061 : 9 : found = true;
2062 : 9 : break;
2063 : : }
2064 : : }
2065 : :
2066 : : /*
2067 : : * If it's an empty sub-transaction then we will not find the subxid
2068 : : * here so just cleanup the subxact info and return.
2069 : : */
2070 [ + + ]: 13 : if (!found)
2071 : : {
2072 : : /* Cleanup the subxact info */
2073 : 4 : cleanup_subxact_info();
2074 : 4 : end_replication_step();
2075 : 4 : CommitTransactionCommand();
2076 : 4 : return;
2077 : : }
2078 : :
2079 : : /* open the changes file */
2080 : 9 : changes_filename(path, MyLogicalRepWorker->subid, xid);
2081 : 9 : fd = BufFileOpenFileSet(MyLogicalRepWorker->stream_fileset, path,
2082 : : O_RDWR, false);
2083 : :
2084 : : /* OK, truncate the file at the right offset */
2085 : 9 : BufFileTruncateFileSet(fd, subxact_data.subxacts[subidx].fileno,
2086 : 9 : subxact_data.subxacts[subidx].offset);
2087 : 9 : BufFileClose(fd);
2088 : :
2089 : : /* discard the subxacts added later */
2090 : 9 : subxact_data.nsubxacts = subidx;
2091 : :
2092 : : /* write the updated subxact list */
2093 : 9 : subxact_info_write(MyLogicalRepWorker->subid, xid);
2094 : :
2095 : 9 : end_replication_step();
2096 : 9 : CommitTransactionCommand();
2097 : : }
2098 : : }
2099 : :
2100 : : /*
2101 : : * Handle STREAM ABORT message.
2102 : : */
2103 : : static void
2104 : 38 : apply_handle_stream_abort(StringInfo s)
2105 : : {
2106 : : TransactionId xid;
2107 : : TransactionId subxid;
2108 : : LogicalRepStreamAbortData abort_data;
2109 : : ParallelApplyWorkerInfo *winfo;
2110 : : TransApplyAction apply_action;
2111 : :
2112 : : /* Save the message before it is consumed. */
2113 : 38 : StringInfoData original_msg = *s;
2114 : : bool toplevel_xact;
2115 : :
2116 [ - + ]: 38 : if (in_streamed_transaction)
2117 [ # # ]: 0 : ereport(ERROR,
2118 : : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2119 : : errmsg_internal("STREAM ABORT message without STREAM STOP")));
2120 : :
2121 : : /* We receive abort information only when we can apply in parallel. */
2122 : 38 : logicalrep_read_stream_abort(s, &abort_data,
2123 : 38 : MyLogicalRepWorker->parallel_apply);
2124 : :
2125 : 38 : xid = abort_data.xid;
2126 : 38 : subxid = abort_data.subxid;
2127 : 38 : toplevel_xact = (xid == subxid);
2128 : :
2129 : : /*
2130 : : * Record the xid of the (sub)transaction being aborted, so that the error
2131 : : * context names whatever failed. Note this is the top-level xid itself
2132 : : * when a top-level transaction aborts, and a subxid only when a
2133 : : * subtransaction rolls back. See set_remote_transaction_info().
2134 : : */
2135 : 38 : set_remote_transaction_info(subxid, abort_data.abort_lsn);
2136 : :
2137 : 38 : apply_action = get_transaction_apply_action(xid, &winfo);
2138 : :
2139 [ + + + + : 38 : switch (apply_action)
- ]
2140 : : {
2141 : 14 : case TRANS_LEADER_APPLY:
2142 : :
2143 : : /*
2144 : : * We are in the leader apply worker and the transaction has been
2145 : : * serialized to file.
2146 : : */
2147 : 14 : stream_abort_internal(xid, subxid);
2148 : :
2149 [ - + ]: 14 : elog(DEBUG1, "finished processing the STREAM ABORT command");
2150 : 14 : break;
2151 : :
2152 : 10 : case TRANS_LEADER_SEND_TO_PARALLEL:
2153 : : Assert(winfo);
2154 : :
2155 : : /*
2156 : : * For the case of aborting the subtransaction, we increment the
2157 : : * number of streaming blocks and take the lock again before
2158 : : * sending the STREAM_ABORT to ensure that the parallel apply
2159 : : * worker will wait on the lock for the next set of changes after
2160 : : * processing the STREAM_ABORT message if it is not already
2161 : : * waiting for STREAM_STOP message.
2162 : : *
2163 : : * It is important to perform this locking before sending the
2164 : : * STREAM_ABORT message so that the leader can hold the lock first
2165 : : * and the parallel apply worker will wait for the leader to
2166 : : * release the lock. This is the same as what we do in
2167 : : * apply_handle_stream_stop. See Locking Considerations atop
2168 : : * applyparallelworker.c.
2169 : : */
2170 [ + + ]: 10 : if (!toplevel_xact)
2171 : : {
2172 : 9 : pa_unlock_stream(xid, AccessExclusiveLock);
2173 : 9 : pg_atomic_add_fetch_u32(&winfo->shared->pending_stream_count, 1);
2174 : 9 : pa_lock_stream(xid, AccessExclusiveLock);
2175 : : }
2176 : :
2177 [ + - ]: 10 : if (pa_send_data(winfo, s->len, s->data))
2178 : : {
2179 : : /*
2180 : : * Unlike STREAM_COMMIT and STREAM_PREPARE, we don't need to
2181 : : * wait here for the parallel apply worker to finish as that
2182 : : * is not required to maintain the commit order and won't have
2183 : : * the risk of failures due to transaction dependencies and
2184 : : * deadlocks. However, it is possible that before the parallel
2185 : : * worker finishes and we clear the worker info, the xid
2186 : : * wraparound happens on the upstream and a new transaction
2187 : : * with the same xid can appear and that can lead to duplicate
2188 : : * entries in ParallelApplyTxnHash. Yet another problem could
2189 : : * be that we may have serialized the changes in partial
2190 : : * serialize mode and the file containing xact changes may
2191 : : * already exist, and after xid wraparound trying to create
2192 : : * the file for the same xid can lead to an error. To avoid
2193 : : * these problems, we decide to wait for the aborts to finish.
2194 : : *
2195 : : * Note, it is okay to not update the flush location position
2196 : : * for aborts as in worst case that means such a transaction
2197 : : * won't be sent again after restart.
2198 : : */
2199 [ + + ]: 10 : if (toplevel_xact)
2200 : 1 : pa_xact_finish(winfo, InvalidXLogRecPtr);
2201 : :
2202 : 10 : break;
2203 : : }
2204 : :
2205 : : /*
2206 : : * Switch to serialize mode when we are not able to send the
2207 : : * change to parallel apply worker.
2208 : : */
2209 : 0 : pa_switch_to_partial_serialize(winfo, true);
2210 : :
2211 : : pg_fallthrough;
2212 : 2 : case TRANS_LEADER_PARTIAL_SERIALIZE:
2213 : : Assert(winfo);
2214 : :
2215 : : /*
2216 : : * Parallel apply worker might have applied some changes, so write
2217 : : * the STREAM_ABORT message so that it can rollback the
2218 : : * subtransaction if needed.
2219 : : */
2220 : 2 : stream_open_and_write_change(xid, LOGICAL_REP_MSG_STREAM_ABORT,
2221 : : &original_msg);
2222 : :
2223 [ + + ]: 2 : if (toplevel_xact)
2224 : : {
2225 : 1 : pa_set_fileset_state(winfo->shared, FS_SERIALIZE_DONE);
2226 : 1 : pa_xact_finish(winfo, InvalidXLogRecPtr);
2227 : : }
2228 : 2 : break;
2229 : :
2230 : 12 : case TRANS_PARALLEL_APPLY:
2231 : :
2232 : : /*
2233 : : * If the parallel apply worker is applying spooled messages then
2234 : : * close the file before aborting.
2235 : : */
2236 [ + + + + ]: 12 : if (toplevel_xact && stream_fd)
2237 : 1 : stream_close_file();
2238 : :
2239 : 12 : pa_stream_abort(&abort_data);
2240 : :
2241 : : /*
2242 : : * We need to wait after processing rollback to savepoint for the
2243 : : * next set of changes.
2244 : : *
2245 : : * We have a race condition here due to which we can start waiting
2246 : : * here when there are more chunk of streams in the queue. See
2247 : : * apply_handle_stream_stop.
2248 : : */
2249 [ + + ]: 12 : if (!toplevel_xact)
2250 : 10 : pa_decr_and_wait_stream_block();
2251 : :
2252 [ + + ]: 12 : elog(DEBUG1, "finished processing the STREAM ABORT command");
2253 : 12 : break;
2254 : :
2255 : 0 : default:
2256 [ # # ]: 0 : elog(ERROR, "unexpected apply action: %d", (int) apply_action);
2257 : : break;
2258 : : }
2259 : :
2260 : 38 : reset_apply_remote_context();
2261 : 38 : }
2262 : :
2263 : : /*
2264 : : * Ensure that the passed location is fileset's end.
2265 : : */
2266 : : static void
2267 : 4 : ensure_last_message(FileSet *stream_fileset, TransactionId xid, int fileno,
2268 : : pgoff_t offset)
2269 : : {
2270 : : char path[MAXPGPATH];
2271 : : BufFile *fd;
2272 : : int last_fileno;
2273 : : pgoff_t last_offset;
2274 : :
2275 : : Assert(!IsTransactionState());
2276 : :
2277 : 4 : begin_replication_step();
2278 : :
2279 : 4 : changes_filename(path, MyLogicalRepWorker->subid, xid);
2280 : :
2281 : 4 : fd = BufFileOpenFileSet(stream_fileset, path, O_RDONLY, false);
2282 : :
2283 : 4 : BufFileSeek(fd, 0, 0, SEEK_END);
2284 : 4 : BufFileTell(fd, &last_fileno, &last_offset);
2285 : :
2286 : 4 : BufFileClose(fd);
2287 : :
2288 : 4 : end_replication_step();
2289 : :
2290 [ + - - + ]: 4 : if (last_fileno != fileno || last_offset != offset)
2291 [ # # ]: 0 : elog(ERROR, "unexpected message left in streaming transaction's changes file \"%s\"",
2292 : : path);
2293 : 4 : }
2294 : :
2295 : : /*
2296 : : * Common spoolfile processing.
2297 : : */
2298 : : void
2299 : 31 : apply_spooled_messages(FileSet *stream_fileset, TransactionId xid,
2300 : : XLogRecPtr lsn)
2301 : : {
2302 : : int nchanges;
2303 : : char path[MAXPGPATH];
2304 : 31 : char *buffer = NULL;
2305 : : MemoryContext oldcxt;
2306 : : ResourceOwner oldowner;
2307 : : int fileno;
2308 : : pgoff_t offset;
2309 : :
2310 [ + + ]: 31 : if (!am_parallel_apply_worker())
2311 : 27 : maybe_start_skipping_changes(lsn);
2312 : :
2313 : : /* Make sure we have an open transaction */
2314 : 31 : begin_replication_step();
2315 : :
2316 : : /*
2317 : : * Allocate file handle and memory required to process all the messages in
2318 : : * TopTransactionContext to avoid them getting reset after each message is
2319 : : * processed.
2320 : : */
2321 : 31 : oldcxt = MemoryContextSwitchTo(TopTransactionContext);
2322 : :
2323 : : /* Open the spool file for the committed/prepared transaction */
2324 : 31 : changes_filename(path, MyLogicalRepWorker->subid, xid);
2325 [ - + ]: 31 : elog(DEBUG1, "replaying changes from file \"%s\"", path);
2326 : :
2327 : : /*
2328 : : * Make sure the file is owned by the toplevel transaction so that the
2329 : : * file will not be accidentally closed when aborting a subtransaction.
2330 : : */
2331 : 31 : oldowner = CurrentResourceOwner;
2332 : 31 : CurrentResourceOwner = TopTransactionResourceOwner;
2333 : :
2334 : 31 : stream_fd = BufFileOpenFileSet(stream_fileset, path, O_RDONLY, false);
2335 : :
2336 : 31 : CurrentResourceOwner = oldowner;
2337 : :
2338 : 31 : buffer = palloc(BLCKSZ);
2339 : :
2340 : 31 : MemoryContextSwitchTo(oldcxt);
2341 : :
2342 : : /*
2343 : : * Make sure the handle apply_dispatch methods are aware we're in a remote
2344 : : * transaction.
2345 : : */
2346 : 31 : in_remote_transaction = true;
2347 : 31 : pgstat_report_activity(STATE_RUNNING, NULL);
2348 : :
2349 : 31 : end_replication_step();
2350 : :
2351 : : /*
2352 : : * Read the entries one by one and pass them through the same logic as in
2353 : : * apply_dispatch.
2354 : : */
2355 : 31 : nchanges = 0;
2356 : : while (true)
2357 : 88470 : {
2358 : : StringInfoData s2;
2359 : : size_t nbytes;
2360 : : int len;
2361 : :
2362 [ - + ]: 88501 : CHECK_FOR_INTERRUPTS();
2363 : :
2364 : : /* read length of the on-disk record */
2365 : 88501 : nbytes = BufFileReadMaybeEOF(stream_fd, &len, sizeof(len), true);
2366 : :
2367 : : /* have we reached end of the file? */
2368 [ + + ]: 88501 : if (nbytes == 0)
2369 : 26 : break;
2370 : :
2371 : : /* do we have a correct length? */
2372 [ - + ]: 88475 : if (len <= 0)
2373 [ # # ]: 0 : elog(ERROR, "incorrect length %d in streaming transaction's changes file \"%s\"",
2374 : : len, path);
2375 : :
2376 : : /* make sure we have sufficiently large buffer */
2377 : 88475 : buffer = repalloc(buffer, len);
2378 : :
2379 : : /* and finally read the data into the buffer */
2380 : 88475 : BufFileReadExact(stream_fd, buffer, len);
2381 : :
2382 : 88475 : BufFileTell(stream_fd, &fileno, &offset);
2383 : :
2384 : : /* init a stringinfo using the buffer and call apply_dispatch */
2385 : 88475 : initReadOnlyStringInfo(&s2, buffer, len);
2386 : :
2387 : : /* Ensure we are reading the data into our memory context. */
2388 : 88475 : oldcxt = MemoryContextSwitchTo(ApplyMessageContext);
2389 : :
2390 : 88475 : apply_dispatch(&s2);
2391 : :
2392 : 88474 : MemoryContextReset(ApplyMessageContext);
2393 : :
2394 : 88474 : MemoryContextSwitchTo(oldcxt);
2395 : :
2396 : 88474 : nchanges++;
2397 : :
2398 : : /*
2399 : : * It is possible the file has been closed because we have processed
2400 : : * the transaction end message like stream_commit in which case that
2401 : : * must be the last message.
2402 : : */
2403 [ + + ]: 88474 : if (!stream_fd)
2404 : : {
2405 : 4 : ensure_last_message(stream_fileset, xid, fileno, offset);
2406 : 4 : break;
2407 : : }
2408 : :
2409 [ + + ]: 88470 : if (nchanges % 1000 == 0)
2410 [ - + ]: 83 : elog(DEBUG1, "replayed %d changes from file \"%s\"",
2411 : : nchanges, path);
2412 : : }
2413 : :
2414 [ + + ]: 30 : if (stream_fd)
2415 : 26 : stream_close_file();
2416 : :
2417 [ - + ]: 30 : elog(DEBUG1, "replayed %d (all) changes from file \"%s\"",
2418 : : nchanges, path);
2419 : :
2420 : 30 : return;
2421 : : }
2422 : :
2423 : : /*
2424 : : * Handle STREAM COMMIT message.
2425 : : */
2426 : : static void
2427 : 61 : apply_handle_stream_commit(StringInfo s)
2428 : : {
2429 : : TransactionId xid;
2430 : : LogicalRepCommitData commit_data;
2431 : : ParallelApplyWorkerInfo *winfo;
2432 : : TransApplyAction apply_action;
2433 : :
2434 : : /* Save the message before it is consumed. */
2435 : 61 : StringInfoData original_msg = *s;
2436 : :
2437 [ - + ]: 61 : if (in_streamed_transaction)
2438 [ # # ]: 0 : ereport(ERROR,
2439 : : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2440 : : errmsg_internal("STREAM COMMIT message without STREAM STOP")));
2441 : :
2442 : 61 : xid = logicalrep_read_stream_commit(s, &commit_data);
2443 : 61 : set_remote_transaction_info(xid, commit_data.commit_lsn);
2444 : :
2445 : 61 : apply_action = get_transaction_apply_action(xid, &winfo);
2446 : :
2447 [ + + + + : 61 : switch (apply_action)
- ]
2448 : : {
2449 : 22 : case TRANS_LEADER_APPLY:
2450 : :
2451 : : /*
2452 : : * The transaction has been serialized to file, so replay all the
2453 : : * spooled operations.
2454 : : */
2455 : 22 : apply_spooled_messages(MyLogicalRepWorker->stream_fileset, xid,
2456 : : commit_data.commit_lsn);
2457 : :
2458 : 21 : apply_handle_commit_internal(&commit_data);
2459 : :
2460 : : /* Unlink the files with serialized changes and subxact info. */
2461 : 21 : stream_cleanup_files(MyLogicalRepWorker->subid, xid);
2462 : :
2463 [ - + ]: 21 : elog(DEBUG1, "finished processing the STREAM COMMIT command");
2464 : 21 : break;
2465 : :
2466 : 18 : case TRANS_LEADER_SEND_TO_PARALLEL:
2467 : : Assert(winfo);
2468 : :
2469 [ + - ]: 18 : if (pa_send_data(winfo, s->len, s->data))
2470 : : {
2471 : : /* Finish processing the streaming transaction. */
2472 : 18 : pa_xact_finish(winfo, commit_data.end_lsn);
2473 : 17 : break;
2474 : : }
2475 : :
2476 : : /*
2477 : : * Switch to serialize mode when we are not able to send the
2478 : : * change to parallel apply worker.
2479 : : */
2480 : 0 : pa_switch_to_partial_serialize(winfo, true);
2481 : :
2482 : : pg_fallthrough;
2483 : 2 : case TRANS_LEADER_PARTIAL_SERIALIZE:
2484 : : Assert(winfo);
2485 : :
2486 : 2 : stream_open_and_write_change(xid, LOGICAL_REP_MSG_STREAM_COMMIT,
2487 : : &original_msg);
2488 : :
2489 : 2 : pa_set_fileset_state(winfo->shared, FS_SERIALIZE_DONE);
2490 : :
2491 : : /* Finish processing the streaming transaction. */
2492 : 2 : pa_xact_finish(winfo, commit_data.end_lsn);
2493 : 2 : break;
2494 : :
2495 : 19 : case TRANS_PARALLEL_APPLY:
2496 : :
2497 : : /*
2498 : : * If the parallel apply worker is applying spooled messages then
2499 : : * close the file before committing.
2500 : : */
2501 [ + + ]: 19 : if (stream_fd)
2502 : 2 : stream_close_file();
2503 : :
2504 : 19 : apply_handle_commit_internal(&commit_data);
2505 : :
2506 : 19 : MyParallelShared->last_commit_end = XactLastCommitEnd;
2507 : :
2508 : : /*
2509 : : * It is important to set the transaction state as finished before
2510 : : * releasing the lock. See pa_wait_for_xact_finish.
2511 : : */
2512 : 19 : pa_set_xact_state(MyParallelShared, PARALLEL_TRANS_FINISHED);
2513 : 19 : pa_unlock_transaction(xid, AccessExclusiveLock);
2514 : :
2515 : 19 : pa_reset_subtrans();
2516 : :
2517 [ + + ]: 19 : elog(DEBUG1, "finished processing the STREAM COMMIT command");
2518 : 19 : break;
2519 : :
2520 : 0 : default:
2521 [ # # ]: 0 : elog(ERROR, "unexpected apply action: %d", (int) apply_action);
2522 : : break;
2523 : : }
2524 : :
2525 : : /*
2526 : : * Process any tables that are being synchronized in parallel, as well as
2527 : : * any newly added tables or sequences.
2528 : : */
2529 : 59 : ProcessSyncingRelations(commit_data.end_lsn);
2530 : :
2531 : 59 : pgstat_report_activity(STATE_IDLE, NULL);
2532 : :
2533 : 59 : reset_apply_remote_context();
2534 : 59 : }
2535 : :
2536 : : /*
2537 : : * Helper function for apply_handle_commit and apply_handle_stream_commit.
2538 : : */
2539 : : static void
2540 : 508 : apply_handle_commit_internal(LogicalRepCommitData *commit_data)
2541 : : {
2542 [ + + ]: 508 : if (is_skipping_changes())
2543 : : {
2544 : 2 : stop_skipping_changes();
2545 : :
2546 : : /*
2547 : : * Start a new transaction to clear the subskiplsn, if not started
2548 : : * yet.
2549 : : */
2550 [ + + ]: 2 : if (!IsTransactionState())
2551 : 1 : StartTransactionCommand();
2552 : : }
2553 : :
2554 [ + - ]: 508 : if (IsTransactionState())
2555 : : {
2556 : : /*
2557 : : * The transaction is either non-empty or skipped, so we clear the
2558 : : * subskiplsn.
2559 : : */
2560 : 508 : clear_subscription_skip_lsn(commit_data->commit_lsn);
2561 : :
2562 : : /*
2563 : : * Update origin state so we can restart streaming from correct
2564 : : * position in case of crash.
2565 : : */
2566 : 508 : replorigin_xact_state.origin_lsn = commit_data->end_lsn;
2567 : 508 : replorigin_xact_state.origin_timestamp = commit_data->committime;
2568 : :
2569 : 508 : CommitTransactionCommand();
2570 : :
2571 [ + + ]: 508 : if (IsTransactionBlock())
2572 : : {
2573 : 4 : EndTransactionBlock(false);
2574 : 4 : CommitTransactionCommand();
2575 : : }
2576 : :
2577 : 508 : pgstat_report_stat(false);
2578 : :
2579 : 508 : store_flush_position(commit_data->end_lsn, XactLastCommitEnd);
2580 : : }
2581 : : else
2582 : : {
2583 : : /* Process any invalidation messages that might have accumulated. */
2584 : 0 : AcceptInvalidationMessages();
2585 : 0 : maybe_reread_subscription();
2586 : : }
2587 : :
2588 : 508 : in_remote_transaction = false;
2589 : 508 : }
2590 : :
2591 : : /*
2592 : : * Handle RELATION message.
2593 : : *
2594 : : * Note we don't do validation against local schema here. The validation
2595 : : * against local schema is postponed until first change for given relation
2596 : : * comes as we only care about it when applying changes for it anyway and we
2597 : : * do less locking this way.
2598 : : */
2599 : : static void
2600 : 522 : apply_handle_relation(StringInfo s)
2601 : : {
2602 : : LogicalRepRelation *rel;
2603 : :
2604 [ + + ]: 522 : if (handle_streamed_transaction(LOGICAL_REP_MSG_RELATION, s))
2605 : 36 : return;
2606 : :
2607 : 486 : rel = logicalrep_read_rel(s);
2608 : 486 : logicalrep_relmap_update(rel);
2609 : :
2610 : : /* Also reset all entries in the partition map that refer to remoterel. */
2611 : 486 : logicalrep_partmap_reset_relmap(rel);
2612 : : }
2613 : :
2614 : : /*
2615 : : * Handle TYPE message.
2616 : : *
2617 : : * This implementation pays no attention to TYPE messages; we expect the user
2618 : : * to have set things up so that the incoming data is acceptable to the input
2619 : : * functions for the locally subscribed tables. Hence, we just read and
2620 : : * discard the message.
2621 : : */
2622 : : static void
2623 : 18 : apply_handle_type(StringInfo s)
2624 : : {
2625 : : LogicalRepTyp typ;
2626 : :
2627 [ - + ]: 18 : if (handle_streamed_transaction(LOGICAL_REP_MSG_TYPE, s))
2628 : 0 : return;
2629 : :
2630 : 18 : logicalrep_read_typ(s, &typ);
2631 : : }
2632 : :
2633 : : /*
2634 : : * Check that we (the subscription owner) have sufficient privileges on the
2635 : : * target relation to perform the given operation.
2636 : : */
2637 : : static void
2638 : 244055 : TargetPrivilegesCheck(Relation rel, AclMode mode)
2639 : : {
2640 : : Oid relid;
2641 : : AclResult aclresult;
2642 : :
2643 : 244055 : relid = RelationGetRelid(rel);
2644 : 244055 : aclresult = pg_class_aclcheck(relid, GetUserId(), mode);
2645 [ + + ]: 244055 : if (aclresult != ACLCHECK_OK)
2646 : 12 : aclcheck_error(aclresult,
2647 : 12 : get_relkind_objtype(rel->rd_rel->relkind),
2648 : 12 : get_rel_name(relid));
2649 : :
2650 : : /*
2651 : : * We lack the infrastructure to honor RLS policies. It might be possible
2652 : : * to add such infrastructure here, but tablesync workers lack it, too, so
2653 : : * we don't bother. RLS does not ordinarily apply to TRUNCATE commands,
2654 : : * but it seems dangerous to replicate a TRUNCATE and then refuse to
2655 : : * replicate subsequent INSERTs, so we forbid all commands the same.
2656 : : */
2657 [ + + ]: 244043 : if (check_enable_rls(relid, InvalidOid, false) == RLS_ENABLED)
2658 [ + - ]: 4 : ereport(ERROR,
2659 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
2660 : : errmsg("user \"%s\" cannot replicate into relation with row-level security enabled: \"%s\"",
2661 : : GetUserNameFromId(GetUserId(), true),
2662 : : RelationGetRelationName(rel))));
2663 : 244039 : }
2664 : :
2665 : : /*
2666 : : * Handle INSERT message.
2667 : : */
2668 : :
2669 : : static void
2670 : 209572 : apply_handle_insert(StringInfo s)
2671 : : {
2672 : : LogicalRepRelMapEntry *rel;
2673 : : LogicalRepTupleData newtup;
2674 : : LogicalRepRelId relid;
2675 : : UserContext ucxt;
2676 : : ApplyExecutionData *edata;
2677 : : EState *estate;
2678 : : TupleTableSlot *remoteslot;
2679 : : MemoryContext oldctx;
2680 : : bool run_as_owner;
2681 : :
2682 : : /*
2683 : : * Quick return if we are skipping data modification changes or handling
2684 : : * streamed transactions.
2685 : : */
2686 [ + + + + ]: 409143 : if (is_skipping_changes() ||
2687 : 199571 : handle_streamed_transaction(LOGICAL_REP_MSG_INSERT, s))
2688 : 110068 : return;
2689 : :
2690 : 99563 : begin_replication_step();
2691 : :
2692 : 99562 : relid = logicalrep_read_insert(s, &newtup);
2693 : 99562 : rel = logicalrep_rel_open(relid, RowExclusiveLock);
2694 [ + + ]: 99548 : if (!should_apply_changes_for_rel(rel))
2695 : : {
2696 : : /*
2697 : : * The relation can't become interesting in the middle of the
2698 : : * transaction so it's safe to unlock it.
2699 : : */
2700 : 59 : logicalrep_rel_close(rel, RowExclusiveLock);
2701 : 59 : end_replication_step();
2702 : 59 : return;
2703 : : }
2704 : :
2705 : : /*
2706 : : * Make sure that any user-supplied code runs as the table owner, unless
2707 : : * the user has opted out of that behavior.
2708 : : */
2709 : 99489 : run_as_owner = MySubscription->runasowner;
2710 [ + + ]: 99489 : if (!run_as_owner)
2711 : 99479 : SwitchToUntrustedUser(rel->localrel->rd_rel->relowner, &ucxt);
2712 : :
2713 : : /* Set relation for error callback */
2714 : 99489 : remote_ctx.rel = rel;
2715 : :
2716 : : /* Initialize the executor state. */
2717 : 99489 : edata = create_edata_for_relation(rel);
2718 : 99489 : estate = edata->estate;
2719 : 99489 : remoteslot = ExecInitExtraTupleSlot(estate,
2720 : 99489 : RelationGetDescr(rel->localrel),
2721 : : &TTSOpsVirtual);
2722 : :
2723 : : /* Process and store remote tuple in the slot */
2724 [ - + ]: 99489 : oldctx = MemoryContextSwitchTo(GetPerTupleMemoryContext(estate));
2725 : 99489 : slot_store_data(remoteslot, rel, &newtup);
2726 : 99489 : slot_fill_defaults(rel, estate, remoteslot);
2727 : 99489 : MemoryContextSwitchTo(oldctx);
2728 : :
2729 : : /* For a partitioned table, insert the tuple into a partition. */
2730 [ + + ]: 99489 : if (rel->localrel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
2731 : 77 : apply_handle_tuple_routing(edata,
2732 : : remoteslot, NULL, CMD_INSERT);
2733 : : else
2734 : : {
2735 : 99412 : ResultRelInfo *relinfo = edata->targetRelInfo;
2736 : :
2737 : 99412 : ExecOpenIndices(relinfo, false);
2738 : 99412 : apply_handle_insert_internal(edata, relinfo, remoteslot);
2739 : 99393 : ExecCloseIndices(relinfo);
2740 : : }
2741 : :
2742 : 99437 : finish_edata(edata);
2743 : :
2744 : : /* Reset relation for error callback */
2745 : 99437 : remote_ctx.rel = NULL;
2746 : :
2747 [ + + ]: 99437 : if (!run_as_owner)
2748 : 99432 : RestoreUserContext(&ucxt);
2749 : :
2750 : 99437 : logicalrep_rel_close(rel, NoLock);
2751 : :
2752 : 99437 : end_replication_step();
2753 : : }
2754 : :
2755 : : /*
2756 : : * Workhorse for apply_handle_insert()
2757 : : * relinfo is for the relation we're actually inserting into
2758 : : * (could be a child partition of edata->targetRelInfo)
2759 : : */
2760 : : static void
2761 : 99490 : apply_handle_insert_internal(ApplyExecutionData *edata,
2762 : : ResultRelInfo *relinfo,
2763 : : TupleTableSlot *remoteslot)
2764 : : {
2765 : 99490 : EState *estate = edata->estate;
2766 : :
2767 : : /* Caller should have opened indexes already. */
2768 : : Assert(relinfo->ri_IndexRelationDescs != NULL ||
2769 : : !relinfo->ri_RelationDesc->rd_rel->relhasindex ||
2770 : : RelationGetIndexList(relinfo->ri_RelationDesc) == NIL);
2771 : :
2772 : : /* Caller will not have done this bit. */
2773 : : Assert(relinfo->ri_onConflictArbiterIndexes == NIL);
2774 : 99490 : InitConflictIndexes(relinfo);
2775 : :
2776 : : /* Do the insert. */
2777 : 99490 : TargetPrivilegesCheck(relinfo->ri_RelationDesc, ACL_INSERT);
2778 : 99481 : ExecSimpleRelationInsert(relinfo, estate, remoteslot);
2779 : 99438 : }
2780 : :
2781 : : /*
2782 : : * Check if the logical replication relation is updatable and throw
2783 : : * appropriate error if it isn't.
2784 : : */
2785 : : static void
2786 : 72311 : check_relation_updatable(LogicalRepRelMapEntry *rel)
2787 : : {
2788 : : /*
2789 : : * For partitioned tables, we only need to care if the target partition is
2790 : : * updatable (aka has PK or RI defined for it).
2791 : : */
2792 [ + + ]: 72311 : if (rel->localrel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
2793 : 30 : return;
2794 : :
2795 : : /* Updatable, no error. */
2796 [ + - ]: 72281 : if (rel->updatable)
2797 : 72281 : return;
2798 : :
2799 : : /*
2800 : : * We are in error mode so it's fine this is somewhat slow. It's better to
2801 : : * give user correct error.
2802 : : */
2803 [ # # ]: 0 : if (OidIsValid(GetRelationIdentityOrPK(rel->localrel)))
2804 : : {
2805 [ # # ]: 0 : ereport(ERROR,
2806 : : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
2807 : : errmsg("publisher did not send replica identity column "
2808 : : "expected by the logical replication target relation \"%s.%s\"",
2809 : : rel->remoterel.nspname, rel->remoterel.relname)));
2810 : : }
2811 : :
2812 [ # # ]: 0 : ereport(ERROR,
2813 : : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
2814 : : errmsg("logical replication target relation \"%s.%s\" has "
2815 : : "neither REPLICA IDENTITY index nor PRIMARY "
2816 : : "KEY and published relation does not have "
2817 : : "REPLICA IDENTITY FULL",
2818 : : rel->remoterel.nspname, rel->remoterel.relname)));
2819 : : }
2820 : :
2821 : : /*
2822 : : * Handle UPDATE message.
2823 : : *
2824 : : * TODO: FDW support
2825 : : */
2826 : : static void
2827 : 66178 : apply_handle_update(StringInfo s)
2828 : : {
2829 : : LogicalRepRelMapEntry *rel;
2830 : : LogicalRepRelId relid;
2831 : : UserContext ucxt;
2832 : : ApplyExecutionData *edata;
2833 : : EState *estate;
2834 : : LogicalRepTupleData oldtup;
2835 : : LogicalRepTupleData newtup;
2836 : : bool has_oldtup;
2837 : : TupleTableSlot *remoteslot;
2838 : : RTEPermissionInfo *target_perminfo;
2839 : : MemoryContext oldctx;
2840 : : bool run_as_owner;
2841 : :
2842 : : /*
2843 : : * Quick return if we are skipping data modification changes or handling
2844 : : * streamed transactions.
2845 : : */
2846 [ + + + + ]: 132353 : if (is_skipping_changes() ||
2847 : 66175 : handle_streamed_transaction(LOGICAL_REP_MSG_UPDATE, s))
2848 : 34223 : return;
2849 : :
2850 : 31955 : begin_replication_step();
2851 : :
2852 : 31954 : relid = logicalrep_read_update(s, &has_oldtup, &oldtup,
2853 : : &newtup);
2854 : 31954 : rel = logicalrep_rel_open(relid, RowExclusiveLock);
2855 [ - + ]: 31954 : if (!should_apply_changes_for_rel(rel))
2856 : : {
2857 : : /*
2858 : : * The relation can't become interesting in the middle of the
2859 : : * transaction so it's safe to unlock it.
2860 : : */
2861 : 0 : logicalrep_rel_close(rel, RowExclusiveLock);
2862 : 0 : end_replication_step();
2863 : 0 : return;
2864 : : }
2865 : :
2866 : : /* Set relation for error callback */
2867 : 31954 : remote_ctx.rel = rel;
2868 : :
2869 : : /* Check if we can do the update. */
2870 : 31954 : check_relation_updatable(rel);
2871 : :
2872 : : /*
2873 : : * Make sure that any user-supplied code runs as the table owner, unless
2874 : : * the user has opted out of that behavior.
2875 : : */
2876 : 31954 : run_as_owner = MySubscription->runasowner;
2877 [ + + ]: 31954 : if (!run_as_owner)
2878 : 31950 : SwitchToUntrustedUser(rel->localrel->rd_rel->relowner, &ucxt);
2879 : :
2880 : : /* Initialize the executor state. */
2881 : 31953 : edata = create_edata_for_relation(rel);
2882 : 31953 : estate = edata->estate;
2883 : 31953 : remoteslot = ExecInitExtraTupleSlot(estate,
2884 : 31953 : RelationGetDescr(rel->localrel),
2885 : : &TTSOpsVirtual);
2886 : :
2887 : : /*
2888 : : * Populate updatedCols so that per-column triggers can fire, and so
2889 : : * executor can correctly pass down indexUnchanged hint. This could
2890 : : * include more columns than were actually changed on the publisher
2891 : : * because the logical replication protocol doesn't contain that
2892 : : * information. But it would for example exclude columns that only exist
2893 : : * on the subscriber, since we are not touching those.
2894 : : */
2895 : 31953 : target_perminfo = list_nth(estate->es_rteperminfos, 0);
2896 [ + + ]: 159370 : for (int i = 0; i < remoteslot->tts_tupleDescriptor->natts; i++)
2897 : : {
2898 : 127417 : CompactAttribute *att = TupleDescCompactAttr(remoteslot->tts_tupleDescriptor, i);
2899 : 127417 : int remoteattnum = rel->attrmap->attnums[i];
2900 : :
2901 [ + + + + ]: 127417 : if (!att->attisdropped && remoteattnum >= 0)
2902 : : {
2903 [ - + ]: 68900 : if (remoteattnum >= newtup.ncols)
2904 [ # # ]: 0 : ereport(ERROR,
2905 : : (errcode(ERRCODE_PROTOCOL_VIOLATION),
2906 : : errmsg("logical replication column %d not found in tuple: only %d column(s) received",
2907 : : remoteattnum + 1, newtup.ncols)));
2908 : :
2909 [ + - ]: 68900 : if (newtup.colstatus[remoteattnum] != LOGICALREP_COLUMN_UNCHANGED)
2910 : 68900 : target_perminfo->updatedCols =
2911 : 68900 : bms_add_member(target_perminfo->updatedCols,
2912 : : i + 1 - FirstLowInvalidHeapAttributeNumber);
2913 : : }
2914 : : }
2915 : :
2916 : : /* Build the search tuple. */
2917 [ - + ]: 31953 : oldctx = MemoryContextSwitchTo(GetPerTupleMemoryContext(estate));
2918 : 31953 : slot_store_data(remoteslot, rel,
2919 [ + + ]: 31953 : has_oldtup ? &oldtup : &newtup);
2920 : 31953 : MemoryContextSwitchTo(oldctx);
2921 : :
2922 : : /* For a partitioned table, apply update to correct partition. */
2923 [ + + ]: 31953 : if (rel->localrel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
2924 : 13 : apply_handle_tuple_routing(edata,
2925 : : remoteslot, &newtup, CMD_UPDATE);
2926 : : else
2927 : 31940 : apply_handle_update_internal(edata, edata->targetRelInfo,
2928 : : remoteslot, &newtup, rel->localindexoid);
2929 : :
2930 : 31944 : finish_edata(edata);
2931 : :
2932 : : /* Reset relation for error callback */
2933 : 31944 : remote_ctx.rel = NULL;
2934 : :
2935 [ + + ]: 31944 : if (!run_as_owner)
2936 : 31942 : RestoreUserContext(&ucxt);
2937 : :
2938 : 31944 : logicalrep_rel_close(rel, NoLock);
2939 : :
2940 : 31944 : end_replication_step();
2941 : : }
2942 : :
2943 : : /*
2944 : : * Workhorse for apply_handle_update()
2945 : : * relinfo is for the relation we're actually updating in
2946 : : * (could be a child partition of edata->targetRelInfo)
2947 : : */
2948 : : static void
2949 : 31940 : apply_handle_update_internal(ApplyExecutionData *edata,
2950 : : ResultRelInfo *relinfo,
2951 : : TupleTableSlot *remoteslot,
2952 : : LogicalRepTupleData *newtup,
2953 : : Oid localindexoid)
2954 : : {
2955 : 31940 : EState *estate = edata->estate;
2956 : 31940 : LogicalRepRelMapEntry *relmapentry = edata->targetRel;
2957 : 31940 : Relation localrel = relinfo->ri_RelationDesc;
2958 : : EPQState epqstate;
2959 : 31940 : TupleTableSlot *localslot = NULL;
2960 : 31940 : ConflictTupleInfo conflicttuple = {0};
2961 : : bool found;
2962 : : MemoryContext oldctx;
2963 : :
2964 : 31940 : EvalPlanQualInit(&epqstate, estate, NULL, NIL, -1, NIL);
2965 : 31940 : ExecOpenIndices(relinfo, false);
2966 : :
2967 : 31940 : found = FindReplTupleInLocalRel(edata, localrel,
2968 : : &relmapentry->remoterel,
2969 : : localindexoid,
2970 : : remoteslot, &localslot);
2971 : :
2972 : : /*
2973 : : * Tuple found.
2974 : : *
2975 : : * Note this will fail if there are other conflicting unique indexes.
2976 : : */
2977 [ + + ]: 31933 : if (found)
2978 : : {
2979 : : /*
2980 : : * Report the conflict if the tuple was modified by a different
2981 : : * origin.
2982 : : */
2983 [ + + ]: 31920 : if (GetTupleTransactionInfo(localslot, &conflicttuple.xmin,
2984 : 2 : &conflicttuple.origin, &conflicttuple.ts) &&
2985 [ + - ]: 2 : conflicttuple.origin != replorigin_xact_state.origin)
2986 : : {
2987 : : TupleTableSlot *newslot;
2988 : :
2989 : : /* Store the new tuple for conflict reporting */
2990 : 2 : newslot = table_slot_create(localrel, &estate->es_tupleTable);
2991 : 2 : slot_store_data(newslot, relmapentry, newtup);
2992 : :
2993 : 2 : conflicttuple.slot = localslot;
2994 : :
2995 : 2 : ReportApplyConflict(estate, relinfo, LOG, CT_UPDATE_ORIGIN_DIFFERS,
2996 : : remoteslot, newslot,
2997 : : list_make1(&conflicttuple));
2998 : : }
2999 : :
3000 : : /* Process and store remote tuple in the slot */
3001 [ + - ]: 31920 : oldctx = MemoryContextSwitchTo(GetPerTupleMemoryContext(estate));
3002 : 31920 : slot_modify_data(remoteslot, localslot, relmapentry, newtup);
3003 : 31920 : MemoryContextSwitchTo(oldctx);
3004 : :
3005 : 31920 : EvalPlanQualSetSlot(&epqstate, remoteslot);
3006 : :
3007 : 31920 : InitConflictIndexes(relinfo);
3008 : :
3009 : : /* Do the actual update. */
3010 : 31920 : TargetPrivilegesCheck(relinfo->ri_RelationDesc, ACL_UPDATE);
3011 : 31920 : ExecSimpleRelationUpdate(relinfo, estate, &epqstate, localslot,
3012 : : remoteslot);
3013 : : }
3014 : : else
3015 : : {
3016 : : ConflictType type;
3017 : 13 : TupleTableSlot *newslot = localslot;
3018 : :
3019 : : /*
3020 : : * Detecting whether the tuple was recently deleted or never existed
3021 : : * is crucial to avoid misleading the user during conflict handling.
3022 : : */
3023 [ + + ]: 13 : if (FindDeletedTupleInLocalRel(localrel, localindexoid, remoteslot,
3024 : : &conflicttuple.xmin,
3025 : : &conflicttuple.origin,
3026 : 3 : &conflicttuple.ts) &&
3027 [ + - ]: 3 : conflicttuple.origin != replorigin_xact_state.origin)
3028 : 3 : type = CT_UPDATE_DELETED;
3029 : : else
3030 : 10 : type = CT_UPDATE_MISSING;
3031 : :
3032 : : /* Store the new tuple for conflict reporting */
3033 : 13 : slot_store_data(newslot, relmapentry, newtup);
3034 : :
3035 : : /*
3036 : : * The tuple to be updated could not be found or was deleted. Do
3037 : : * nothing except for emitting a log message.
3038 : : */
3039 : 13 : ReportApplyConflict(estate, relinfo, LOG, type, remoteslot, newslot,
3040 : : list_make1(&conflicttuple));
3041 : : }
3042 : :
3043 : : /* Cleanup. */
3044 : 31931 : ExecCloseIndices(relinfo);
3045 : 31931 : EvalPlanQualEnd(&epqstate);
3046 : 31931 : }
3047 : :
3048 : : /*
3049 : : * Handle DELETE message.
3050 : : *
3051 : : * TODO: FDW support
3052 : : */
3053 : : static void
3054 : 81942 : apply_handle_delete(StringInfo s)
3055 : : {
3056 : : LogicalRepRelMapEntry *rel;
3057 : : LogicalRepTupleData oldtup;
3058 : : LogicalRepRelId relid;
3059 : : UserContext ucxt;
3060 : : ApplyExecutionData *edata;
3061 : : EState *estate;
3062 : : TupleTableSlot *remoteslot;
3063 : : MemoryContext oldctx;
3064 : : bool run_as_owner;
3065 : :
3066 : : /*
3067 : : * Quick return if we are skipping data modification changes or handling
3068 : : * streamed transactions.
3069 : : */
3070 [ + - + + ]: 163884 : if (is_skipping_changes() ||
3071 : 81942 : handle_streamed_transaction(LOGICAL_REP_MSG_DELETE, s))
3072 : 41615 : return;
3073 : :
3074 : 40327 : begin_replication_step();
3075 : :
3076 : 40327 : relid = logicalrep_read_delete(s, &oldtup);
3077 : 40327 : rel = logicalrep_rel_open(relid, RowExclusiveLock);
3078 [ - + ]: 40327 : if (!should_apply_changes_for_rel(rel))
3079 : : {
3080 : : /*
3081 : : * The relation can't become interesting in the middle of the
3082 : : * transaction so it's safe to unlock it.
3083 : : */
3084 : 0 : logicalrep_rel_close(rel, RowExclusiveLock);
3085 : 0 : end_replication_step();
3086 : 0 : return;
3087 : : }
3088 : :
3089 : : /* Set relation for error callback */
3090 : 40327 : remote_ctx.rel = rel;
3091 : :
3092 : : /* Check if we can do the delete. */
3093 : 40327 : check_relation_updatable(rel);
3094 : :
3095 : : /*
3096 : : * Make sure that any user-supplied code runs as the table owner, unless
3097 : : * the user has opted out of that behavior.
3098 : : */
3099 : 40327 : run_as_owner = MySubscription->runasowner;
3100 [ + + ]: 40327 : if (!run_as_owner)
3101 : 40325 : SwitchToUntrustedUser(rel->localrel->rd_rel->relowner, &ucxt);
3102 : :
3103 : : /* Initialize the executor state. */
3104 : 40327 : edata = create_edata_for_relation(rel);
3105 : 40327 : estate = edata->estate;
3106 : 40327 : remoteslot = ExecInitExtraTupleSlot(estate,
3107 : 40327 : RelationGetDescr(rel->localrel),
3108 : : &TTSOpsVirtual);
3109 : :
3110 : : /* Build the search tuple. */
3111 [ - + ]: 40327 : oldctx = MemoryContextSwitchTo(GetPerTupleMemoryContext(estate));
3112 : 40327 : slot_store_data(remoteslot, rel, &oldtup);
3113 : 40327 : MemoryContextSwitchTo(oldctx);
3114 : :
3115 : : /* For a partitioned table, apply delete to correct partition. */
3116 [ + + ]: 40327 : if (rel->localrel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
3117 : 17 : apply_handle_tuple_routing(edata,
3118 : : remoteslot, NULL, CMD_DELETE);
3119 : : else
3120 : : {
3121 : 40310 : ResultRelInfo *relinfo = edata->targetRelInfo;
3122 : :
3123 : 40310 : ExecOpenIndices(relinfo, false);
3124 : 40310 : apply_handle_delete_internal(edata, relinfo,
3125 : : remoteslot, rel->localindexoid);
3126 : 40310 : ExecCloseIndices(relinfo);
3127 : : }
3128 : :
3129 : 40327 : finish_edata(edata);
3130 : :
3131 : : /* Reset relation for error callback */
3132 : 40327 : remote_ctx.rel = NULL;
3133 : :
3134 [ + + ]: 40327 : if (!run_as_owner)
3135 : 40325 : RestoreUserContext(&ucxt);
3136 : :
3137 : 40327 : logicalrep_rel_close(rel, NoLock);
3138 : :
3139 : 40327 : end_replication_step();
3140 : : }
3141 : :
3142 : : /*
3143 : : * Workhorse for apply_handle_delete()
3144 : : * relinfo is for the relation we're actually deleting from
3145 : : * (could be a child partition of edata->targetRelInfo)
3146 : : */
3147 : : static void
3148 : 40327 : apply_handle_delete_internal(ApplyExecutionData *edata,
3149 : : ResultRelInfo *relinfo,
3150 : : TupleTableSlot *remoteslot,
3151 : : Oid localindexoid)
3152 : : {
3153 : 40327 : EState *estate = edata->estate;
3154 : 40327 : Relation localrel = relinfo->ri_RelationDesc;
3155 : 40327 : LogicalRepRelation *remoterel = &edata->targetRel->remoterel;
3156 : : EPQState epqstate;
3157 : : TupleTableSlot *localslot;
3158 : 40327 : ConflictTupleInfo conflicttuple = {0};
3159 : : bool found;
3160 : :
3161 : 40327 : EvalPlanQualInit(&epqstate, estate, NULL, NIL, -1, NIL);
3162 : :
3163 : : /* Caller should have opened indexes already. */
3164 : : Assert(relinfo->ri_IndexRelationDescs != NULL ||
3165 : : !localrel->rd_rel->relhasindex ||
3166 : : RelationGetIndexList(localrel) == NIL);
3167 : :
3168 : 40327 : found = FindReplTupleInLocalRel(edata, localrel, remoterel, localindexoid,
3169 : : remoteslot, &localslot);
3170 : :
3171 : : /* If found delete it. */
3172 [ + + ]: 40327 : if (found)
3173 : : {
3174 : : /*
3175 : : * Report the conflict if the tuple was modified by a different
3176 : : * origin.
3177 : : */
3178 [ + + ]: 40318 : if (GetTupleTransactionInfo(localslot, &conflicttuple.xmin,
3179 : 6 : &conflicttuple.origin, &conflicttuple.ts) &&
3180 [ + + ]: 6 : conflicttuple.origin != replorigin_xact_state.origin)
3181 : : {
3182 : 5 : conflicttuple.slot = localslot;
3183 : 5 : ReportApplyConflict(estate, relinfo, LOG, CT_DELETE_ORIGIN_DIFFERS,
3184 : : remoteslot, NULL,
3185 : : list_make1(&conflicttuple));
3186 : : }
3187 : :
3188 : 40318 : EvalPlanQualSetSlot(&epqstate, localslot);
3189 : :
3190 : : /* Do the actual delete. */
3191 : 40318 : TargetPrivilegesCheck(relinfo->ri_RelationDesc, ACL_DELETE);
3192 : 40318 : ExecSimpleRelationDelete(relinfo, estate, &epqstate, localslot);
3193 : : }
3194 : : else
3195 : : {
3196 : : /*
3197 : : * The tuple to be deleted could not be found. Do nothing except for
3198 : : * emitting a log message.
3199 : : */
3200 : 9 : ReportApplyConflict(estate, relinfo, LOG, CT_DELETE_MISSING,
3201 : : remoteslot, NULL, list_make1(&conflicttuple));
3202 : : }
3203 : :
3204 : : /* Cleanup. */
3205 : 40327 : EvalPlanQualEnd(&epqstate);
3206 : 40327 : }
3207 : :
3208 : : /*
3209 : : * Try to find a tuple received from the publication side (in 'remoteslot') in
3210 : : * the corresponding local relation using either replica identity index,
3211 : : * primary key, index or if needed, sequential scan.
3212 : : *
3213 : : * Local tuple, if found, is returned in '*localslot'.
3214 : : */
3215 : : static bool
3216 : 72280 : FindReplTupleInLocalRel(ApplyExecutionData *edata, Relation localrel,
3217 : : LogicalRepRelation *remoterel,
3218 : : Oid localidxoid,
3219 : : TupleTableSlot *remoteslot,
3220 : : TupleTableSlot **localslot)
3221 : : {
3222 : 72280 : EState *estate = edata->estate;
3223 : : bool found;
3224 : :
3225 : : /*
3226 : : * Regardless of the top-level operation, we're performing a read here, so
3227 : : * check for SELECT privileges.
3228 : : */
3229 : 72280 : TargetPrivilegesCheck(localrel, ACL_SELECT);
3230 : :
3231 : 72273 : *localslot = table_slot_create(localrel, &estate->es_tupleTable);
3232 : :
3233 : : Assert(OidIsValid(localidxoid) ||
3234 : : (remoterel->replident == REPLICA_IDENTITY_FULL));
3235 : :
3236 [ + + ]: 72273 : if (OidIsValid(localidxoid))
3237 : : {
3238 : : #ifdef USE_ASSERT_CHECKING
3239 : : Relation idxrel = index_open(localidxoid, AccessShareLock);
3240 : :
3241 : : /* Index must be PK, RI, or usable for REPLICA IDENTITY FULL tables */
3242 : : Assert(GetRelationIdentityOrPK(localrel) == localidxoid ||
3243 : : (remoterel->replident == REPLICA_IDENTITY_FULL &&
3244 : : IsIndexUsableForReplicaIdentityFull(idxrel,
3245 : : edata->targetRel->attrmap)));
3246 : : index_close(idxrel, AccessShareLock);
3247 : : #endif
3248 : :
3249 : 72119 : found = RelationFindReplTupleByIndex(localrel, localidxoid,
3250 : : LockTupleExclusive,
3251 : : remoteslot, *localslot);
3252 : : }
3253 : : else
3254 : 154 : found = RelationFindReplTupleSeq(localrel, LockTupleExclusive,
3255 : : remoteslot, *localslot);
3256 : :
3257 : 72273 : return found;
3258 : : }
3259 : :
3260 : : /*
3261 : : * Determine whether the index can reliably locate the deleted tuple in the
3262 : : * local relation.
3263 : : *
3264 : : * An index may exclude deleted tuples if it was re-indexed or re-created during
3265 : : * change application. Therefore, an index is considered usable only if the
3266 : : * conflict detection slot.xmin (conflict_detection_xmin) is greater than the
3267 : : * index tuple's xmin. This ensures that any tuples deleted prior to the index
3268 : : * creation or re-indexing are not relevant for conflict detection in the
3269 : : * current apply worker.
3270 : : *
3271 : : * Note that indexes may also be excluded if they were modified by other DDL
3272 : : * operations, such as ALTER INDEX. However, this is acceptable, as the
3273 : : * likelihood of such DDL changes coinciding with the need to scan dead
3274 : : * tuples for the update_deleted is low.
3275 : : */
3276 : : static bool
3277 : 1 : IsIndexUsableForFindingDeletedTuple(Oid localindexoid,
3278 : : TransactionId conflict_detection_xmin)
3279 : : {
3280 : : HeapTuple index_tuple;
3281 : : TransactionId index_xmin;
3282 : :
3283 : 1 : index_tuple = SearchSysCache1(INDEXRELID, ObjectIdGetDatum(localindexoid));
3284 : :
3285 [ - + ]: 1 : if (!HeapTupleIsValid(index_tuple)) /* should not happen */
3286 [ # # ]: 0 : elog(ERROR, "cache lookup failed for index %u", localindexoid);
3287 : :
3288 : : /*
3289 : : * No need to check for a frozen transaction ID, as
3290 : : * TransactionIdPrecedes() manages it internally, treating it as falling
3291 : : * behind the conflict_detection_xmin.
3292 : : */
3293 : 1 : index_xmin = HeapTupleHeaderGetXmin(index_tuple->t_data);
3294 : :
3295 : 1 : ReleaseSysCache(index_tuple);
3296 : :
3297 : 1 : return TransactionIdPrecedes(index_xmin, conflict_detection_xmin);
3298 : : }
3299 : :
3300 : : /*
3301 : : * Attempts to locate a deleted tuple in the local relation that matches the
3302 : : * values of the tuple received from the publication side (in 'remoteslot').
3303 : : * The search is performed using either the replica identity index, primary
3304 : : * key, other available index, or a sequential scan if necessary.
3305 : : *
3306 : : * Returns true if the deleted tuple is found. If found, the transaction ID,
3307 : : * origin, and commit timestamp of the deletion are stored in '*delete_xid',
3308 : : * '*delete_origin', and '*delete_time' respectively.
3309 : : */
3310 : : static bool
3311 : 15 : FindDeletedTupleInLocalRel(Relation localrel, Oid localidxoid,
3312 : : TupleTableSlot *remoteslot,
3313 : : TransactionId *delete_xid, ReplOriginId *delete_origin,
3314 : : TimestampTz *delete_time)
3315 : : {
3316 : : TransactionId oldestxmin;
3317 : :
3318 : : /*
3319 : : * Return false if either dead tuples are not retained or commit timestamp
3320 : : * data is not available.
3321 : : */
3322 [ + + - + ]: 15 : if (!MySubscription->retaindeadtuples || !track_commit_timestamp)
3323 : 12 : return false;
3324 : :
3325 : : /*
3326 : : * For conflict detection, we use the leader worker's
3327 : : * oldest_nonremovable_xid value instead of invoking
3328 : : * GetOldestNonRemovableTransactionId() or using the conflict detection
3329 : : * slot's xmin. The oldest_nonremovable_xid acts as a threshold to
3330 : : * identify tuples that were recently deleted. These deleted tuples are no
3331 : : * longer visible to concurrent transactions. However, if a remote update
3332 : : * matches such a tuple, we log an update_deleted conflict.
3333 : : *
3334 : : * While GetOldestNonRemovableTransactionId() and slot.xmin may return
3335 : : * transaction IDs older than oldest_nonremovable_xid, for our current
3336 : : * purpose, it is acceptable to treat tuples deleted by transactions prior
3337 : : * to oldest_nonremovable_xid as update_missing conflicts.
3338 : : */
3339 [ + - ]: 3 : if (am_leader_apply_worker())
3340 : : {
3341 : 3 : oldestxmin = MyLogicalRepWorker->oldest_nonremovable_xid;
3342 : : }
3343 : : else
3344 : : {
3345 : : LogicalRepWorker *leader;
3346 : :
3347 : : /*
3348 : : * Obtain the information from the leader apply worker as only the
3349 : : * leader manages oldest_nonremovable_xid (see
3350 : : * maybe_advance_nonremovable_xid() for details).
3351 : : */
3352 : 0 : LWLockAcquire(LogicalRepWorkerLock, LW_SHARED);
3353 : 0 : leader = logicalrep_worker_find(WORKERTYPE_APPLY,
3354 : 0 : MyLogicalRepWorker->subid, InvalidOid,
3355 : : false);
3356 [ # # ]: 0 : if (!leader)
3357 : : {
3358 [ # # ]: 0 : ereport(ERROR,
3359 : : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
3360 : : errmsg("could not detect conflict as the leader apply worker has exited")));
3361 : : }
3362 : :
3363 : 0 : SpinLockAcquire(&leader->relmutex);
3364 : 0 : oldestxmin = leader->oldest_nonremovable_xid;
3365 : 0 : SpinLockRelease(&leader->relmutex);
3366 : 0 : LWLockRelease(LogicalRepWorkerLock);
3367 : : }
3368 : :
3369 : : /*
3370 : : * Return false if the leader apply worker has stopped retaining
3371 : : * information for detecting conflicts. This implies that update_deleted
3372 : : * can no longer be reliably detected.
3373 : : */
3374 [ - + ]: 3 : if (!TransactionIdIsValid(oldestxmin))
3375 : 0 : return false;
3376 : :
3377 [ + + + - ]: 4 : if (OidIsValid(localidxoid) &&
3378 : 1 : IsIndexUsableForFindingDeletedTuple(localidxoid, oldestxmin))
3379 : 1 : return RelationFindDeletedTupleInfoByIndex(localrel, localidxoid,
3380 : : remoteslot, oldestxmin,
3381 : : delete_xid, delete_origin,
3382 : : delete_time);
3383 : : else
3384 : 2 : return RelationFindDeletedTupleInfoSeq(localrel, remoteslot,
3385 : : oldestxmin, delete_xid,
3386 : : delete_origin, delete_time);
3387 : : }
3388 : :
3389 : : /*
3390 : : * This handles insert, update, delete on a partitioned table.
3391 : : */
3392 : : static void
3393 : 107 : apply_handle_tuple_routing(ApplyExecutionData *edata,
3394 : : TupleTableSlot *remoteslot,
3395 : : LogicalRepTupleData *newtup,
3396 : : CmdType operation)
3397 : : {
3398 : 107 : EState *estate = edata->estate;
3399 : 107 : LogicalRepRelMapEntry *relmapentry = edata->targetRel;
3400 : 107 : ResultRelInfo *relinfo = edata->targetRelInfo;
3401 : 107 : Relation parentrel = relinfo->ri_RelationDesc;
3402 : : ModifyTableState *mtstate;
3403 : : PartitionTupleRouting *proute;
3404 : : ResultRelInfo *partrelinfo;
3405 : : Relation partrel;
3406 : : TupleTableSlot *remoteslot_part;
3407 : : TupleConversionMap *map;
3408 : : MemoryContext oldctx;
3409 : 107 : LogicalRepRelMapEntry *part_entry = NULL;
3410 : 107 : AttrMap *attrmap = NULL;
3411 : :
3412 : : /* ModifyTableState is needed for ExecFindPartition(). */
3413 : 107 : edata->mtstate = mtstate = makeNode(ModifyTableState);
3414 : 107 : mtstate->ps.plan = NULL;
3415 : 107 : mtstate->ps.state = estate;
3416 : 107 : mtstate->operation = operation;
3417 : 107 : mtstate->resultRelInfo = relinfo;
3418 : :
3419 : : /* ... as is PartitionTupleRouting. */
3420 : 107 : edata->proute = proute = ExecSetupPartitionTupleRouting(estate, parentrel);
3421 : :
3422 : : /*
3423 : : * Find the partition to which the "search tuple" belongs.
3424 : : */
3425 : : Assert(remoteslot != NULL);
3426 [ + - ]: 107 : oldctx = MemoryContextSwitchTo(GetPerTupleMemoryContext(estate));
3427 : 107 : partrelinfo = ExecFindPartition(mtstate, relinfo, proute,
3428 : : remoteslot, estate);
3429 : : Assert(partrelinfo != NULL);
3430 : 107 : partrel = partrelinfo->ri_RelationDesc;
3431 : :
3432 : : /*
3433 : : * Check for supported relkind. We need this since partitions might be of
3434 : : * unsupported relkinds; and the set of partitions can change, so checking
3435 : : * at CREATE/ALTER SUBSCRIPTION would be insufficient.
3436 : : */
3437 : 107 : CheckSubscriptionRelkind(partrel->rd_rel->relkind,
3438 : 107 : relmapentry->remoterel.relkind,
3439 : 107 : get_namespace_name(RelationGetNamespace(partrel)),
3440 : 107 : RelationGetRelationName(partrel));
3441 : :
3442 : : /*
3443 : : * To perform any of the operations below, the tuple must match the
3444 : : * partition's rowtype. Convert if needed or just copy, using a dedicated
3445 : : * slot to store the tuple in any case.
3446 : : */
3447 : 107 : remoteslot_part = partrelinfo->ri_PartitionTupleSlot;
3448 [ + + ]: 107 : if (remoteslot_part == NULL)
3449 : 74 : remoteslot_part = table_slot_create(partrel, &estate->es_tupleTable);
3450 : 107 : map = ExecGetRootToChildMap(partrelinfo, estate);
3451 [ + + ]: 107 : if (map != NULL)
3452 : : {
3453 : 33 : attrmap = map->attrMap;
3454 : 33 : remoteslot_part = execute_attr_map_slot(attrmap, remoteslot,
3455 : : remoteslot_part);
3456 : : }
3457 : : else
3458 : : {
3459 : 74 : remoteslot_part = ExecCopySlot(remoteslot_part, remoteslot);
3460 : 74 : slot_getallattrs(remoteslot_part);
3461 : : }
3462 : 107 : MemoryContextSwitchTo(oldctx);
3463 : :
3464 : : /* Check if we can do the update or delete on the leaf partition. */
3465 [ + + + + ]: 107 : if (operation == CMD_UPDATE || operation == CMD_DELETE)
3466 : : {
3467 : 30 : part_entry = logicalrep_partition_open(relmapentry, partrel,
3468 : : attrmap);
3469 : 30 : check_relation_updatable(part_entry);
3470 : : }
3471 : :
3472 [ + + + - ]: 107 : switch (operation)
3473 : : {
3474 : 77 : case CMD_INSERT:
3475 : 77 : apply_handle_insert_internal(edata, partrelinfo,
3476 : : remoteslot_part);
3477 : 44 : break;
3478 : :
3479 : 17 : case CMD_DELETE:
3480 : 17 : apply_handle_delete_internal(edata, partrelinfo,
3481 : : remoteslot_part,
3482 : : part_entry->localindexoid);
3483 : 17 : break;
3484 : :
3485 : 13 : case CMD_UPDATE:
3486 : :
3487 : : /*
3488 : : * For UPDATE, depending on whether or not the updated tuple
3489 : : * satisfies the partition's constraint, perform a simple UPDATE
3490 : : * of the partition or move the updated tuple into a different
3491 : : * suitable partition.
3492 : : */
3493 : : {
3494 : : TupleTableSlot *localslot;
3495 : : ResultRelInfo *partrelinfo_new;
3496 : : Relation partrel_new;
3497 : : bool found;
3498 : : EPQState epqstate;
3499 : 13 : ConflictTupleInfo conflicttuple = {0};
3500 : :
3501 : : /* Get the matching local tuple from the partition. */
3502 : 13 : found = FindReplTupleInLocalRel(edata, partrel,
3503 : : &part_entry->remoterel,
3504 : : part_entry->localindexoid,
3505 : : remoteslot_part, &localslot);
3506 [ + + ]: 13 : if (!found)
3507 : : {
3508 : : ConflictType type;
3509 : 2 : TupleTableSlot *newslot = localslot;
3510 : :
3511 : : /*
3512 : : * Detecting whether the tuple was recently deleted or
3513 : : * never existed is crucial to avoid misleading the user
3514 : : * during conflict handling.
3515 : : */
3516 [ - + ]: 2 : if (FindDeletedTupleInLocalRel(partrel,
3517 : : part_entry->localindexoid,
3518 : : remoteslot_part,
3519 : : &conflicttuple.xmin,
3520 : : &conflicttuple.origin,
3521 : 0 : &conflicttuple.ts) &&
3522 [ # # ]: 0 : conflicttuple.origin != replorigin_xact_state.origin)
3523 : 0 : type = CT_UPDATE_DELETED;
3524 : : else
3525 : 2 : type = CT_UPDATE_MISSING;
3526 : :
3527 : : /* Store the new tuple for conflict reporting */
3528 : 2 : slot_store_data(newslot, part_entry, newtup);
3529 : :
3530 : : /*
3531 : : * The tuple to be updated could not be found or was
3532 : : * deleted. Do nothing except for emitting a log message.
3533 : : */
3534 : 2 : ReportApplyConflict(estate, partrelinfo, LOG,
3535 : : type, remoteslot_part, newslot,
3536 : : list_make1(&conflicttuple));
3537 : :
3538 : 2 : return;
3539 : : }
3540 : :
3541 : : /*
3542 : : * Report the conflict if the tuple was modified by a
3543 : : * different origin.
3544 : : */
3545 [ + + ]: 11 : if (GetTupleTransactionInfo(localslot, &conflicttuple.xmin,
3546 : : &conflicttuple.origin,
3547 : 1 : &conflicttuple.ts) &&
3548 [ + - ]: 1 : conflicttuple.origin != replorigin_xact_state.origin)
3549 : : {
3550 : : TupleTableSlot *newslot;
3551 : :
3552 : : /* Store the new tuple for conflict reporting */
3553 : 1 : newslot = table_slot_create(partrel, &estate->es_tupleTable);
3554 : 1 : slot_store_data(newslot, part_entry, newtup);
3555 : :
3556 : 1 : conflicttuple.slot = localslot;
3557 : :
3558 : 1 : ReportApplyConflict(estate, partrelinfo, LOG, CT_UPDATE_ORIGIN_DIFFERS,
3559 : : remoteslot_part, newslot,
3560 : : list_make1(&conflicttuple));
3561 : : }
3562 : :
3563 : : /*
3564 : : * Apply the update to the local tuple, putting the result in
3565 : : * remoteslot_part.
3566 : : */
3567 [ + - ]: 11 : oldctx = MemoryContextSwitchTo(GetPerTupleMemoryContext(estate));
3568 : 11 : slot_modify_data(remoteslot_part, localslot, part_entry,
3569 : : newtup);
3570 : 11 : MemoryContextSwitchTo(oldctx);
3571 : :
3572 : 11 : EvalPlanQualInit(&epqstate, estate, NULL, NIL, -1, NIL);
3573 : :
3574 : : /*
3575 : : * Does the updated tuple still satisfy the current
3576 : : * partition's constraint?
3577 : : */
3578 [ + - + + ]: 22 : if (!partrel->rd_rel->relispartition ||
3579 : 11 : ExecPartitionCheck(partrelinfo, remoteslot_part, estate,
3580 : : false))
3581 : : {
3582 : : /*
3583 : : * Yes, so simply UPDATE the partition. We don't call
3584 : : * apply_handle_update_internal() here, which would
3585 : : * normally do the following work, to avoid repeating some
3586 : : * work already done above to find the local tuple in the
3587 : : * partition.
3588 : : */
3589 : 10 : InitConflictIndexes(partrelinfo);
3590 : :
3591 : 10 : EvalPlanQualSetSlot(&epqstate, remoteslot_part);
3592 : 10 : TargetPrivilegesCheck(partrelinfo->ri_RelationDesc,
3593 : : ACL_UPDATE);
3594 : 10 : ExecSimpleRelationUpdate(partrelinfo, estate, &epqstate,
3595 : : localslot, remoteslot_part);
3596 : : }
3597 : : else
3598 : : {
3599 : : /* Move the tuple into the new partition. */
3600 : :
3601 : : /*
3602 : : * New partition will be found using tuple routing, which
3603 : : * can only occur via the parent table. We might need to
3604 : : * convert the tuple to the parent's rowtype. Note that
3605 : : * this is the tuple found in the partition, not the
3606 : : * original search tuple received by this function.
3607 : : */
3608 [ + - ]: 1 : if (map)
3609 : : {
3610 : : TupleConversionMap *PartitionToRootMap =
3611 : 1 : convert_tuples_by_name(RelationGetDescr(partrel),
3612 : : RelationGetDescr(parentrel));
3613 : :
3614 : : remoteslot =
3615 : 1 : execute_attr_map_slot(PartitionToRootMap->attrMap,
3616 : : remoteslot_part, remoteslot);
3617 : : }
3618 : : else
3619 : : {
3620 : 0 : remoteslot = ExecCopySlot(remoteslot, remoteslot_part);
3621 : 0 : slot_getallattrs(remoteslot);
3622 : : }
3623 : :
3624 : : /* Find the new partition. */
3625 [ + - ]: 1 : oldctx = MemoryContextSwitchTo(GetPerTupleMemoryContext(estate));
3626 : 1 : partrelinfo_new = ExecFindPartition(mtstate, relinfo,
3627 : : proute, remoteslot,
3628 : : estate);
3629 : 1 : MemoryContextSwitchTo(oldctx);
3630 : : Assert(partrelinfo_new != partrelinfo);
3631 : 1 : partrel_new = partrelinfo_new->ri_RelationDesc;
3632 : :
3633 : : /* Check that new partition also has supported relkind. */
3634 : 1 : CheckSubscriptionRelkind(partrel_new->rd_rel->relkind,
3635 : 1 : relmapentry->remoterel.relkind,
3636 : 1 : get_namespace_name(RelationGetNamespace(partrel_new)),
3637 : 1 : RelationGetRelationName(partrel_new));
3638 : :
3639 : : /* DELETE old tuple found in the old partition. */
3640 : 1 : EvalPlanQualSetSlot(&epqstate, localslot);
3641 : 1 : TargetPrivilegesCheck(partrelinfo->ri_RelationDesc, ACL_DELETE);
3642 : 1 : ExecSimpleRelationDelete(partrelinfo, estate, &epqstate, localslot);
3643 : :
3644 : : /* INSERT new tuple into the new partition. */
3645 : :
3646 : : /*
3647 : : * Convert the replacement tuple to match the destination
3648 : : * partition rowtype.
3649 : : */
3650 [ + - ]: 1 : oldctx = MemoryContextSwitchTo(GetPerTupleMemoryContext(estate));
3651 : 1 : remoteslot_part = partrelinfo_new->ri_PartitionTupleSlot;
3652 [ + - ]: 1 : if (remoteslot_part == NULL)
3653 : 1 : remoteslot_part = table_slot_create(partrel_new,
3654 : : &estate->es_tupleTable);
3655 : 1 : map = ExecGetRootToChildMap(partrelinfo_new, estate);
3656 [ - + ]: 1 : if (map != NULL)
3657 : : {
3658 : 0 : remoteslot_part = execute_attr_map_slot(map->attrMap,
3659 : : remoteslot,
3660 : : remoteslot_part);
3661 : : }
3662 : : else
3663 : : {
3664 : 1 : remoteslot_part = ExecCopySlot(remoteslot_part,
3665 : : remoteslot);
3666 : 1 : slot_getallattrs(remoteslot);
3667 : : }
3668 : 1 : MemoryContextSwitchTo(oldctx);
3669 : 1 : apply_handle_insert_internal(edata, partrelinfo_new,
3670 : : remoteslot_part);
3671 : : }
3672 : :
3673 : 11 : EvalPlanQualEnd(&epqstate);
3674 : : }
3675 : 11 : break;
3676 : :
3677 : 0 : default:
3678 [ # # ]: 0 : elog(ERROR, "unrecognized CmdType: %d", (int) operation);
3679 : : break;
3680 : : }
3681 : : }
3682 : :
3683 : : /*
3684 : : * Handle TRUNCATE message.
3685 : : *
3686 : : * TODO: FDW support
3687 : : */
3688 : : static void
3689 : 20 : apply_handle_truncate(StringInfo s)
3690 : : {
3691 : 20 : bool cascade = false;
3692 : 20 : bool restart_seqs = false;
3693 : 20 : List *remote_relids = NIL;
3694 : 20 : List *remote_rels = NIL;
3695 : 20 : List *rels = NIL;
3696 : 20 : List *part_rels = NIL;
3697 : 20 : List *relids = NIL;
3698 : 20 : List *relids_logged = NIL;
3699 : : ListCell *lc;
3700 : 20 : LOCKMODE lockmode = AccessExclusiveLock;
3701 : :
3702 : : /*
3703 : : * Quick return if we are skipping data modification changes or handling
3704 : : * streamed transactions.
3705 : : */
3706 [ + - - + ]: 40 : if (is_skipping_changes() ||
3707 : 20 : handle_streamed_transaction(LOGICAL_REP_MSG_TRUNCATE, s))
3708 : 0 : return;
3709 : :
3710 : 20 : begin_replication_step();
3711 : :
3712 : 20 : remote_relids = logicalrep_read_truncate(s, &cascade, &restart_seqs);
3713 : :
3714 [ + - + + : 49 : foreach(lc, remote_relids)
+ + ]
3715 : : {
3716 : 29 : LogicalRepRelId relid = lfirst_oid(lc);
3717 : : LogicalRepRelMapEntry *rel;
3718 : :
3719 : 29 : rel = logicalrep_rel_open(relid, lockmode);
3720 [ - + ]: 29 : if (!should_apply_changes_for_rel(rel))
3721 : : {
3722 : : /*
3723 : : * The relation can't become interesting in the middle of the
3724 : : * transaction so it's safe to unlock it.
3725 : : */
3726 : 0 : logicalrep_rel_close(rel, lockmode);
3727 : 0 : continue;
3728 : : }
3729 : :
3730 : 29 : remote_rels = lappend(remote_rels, rel);
3731 : 29 : TargetPrivilegesCheck(rel->localrel, ACL_TRUNCATE);
3732 : 29 : rels = lappend(rels, rel->localrel);
3733 : 29 : relids = lappend_oid(relids, rel->localreloid);
3734 [ + + - + : 29 : if (RelationIsLogicallyLogged(rel->localrel))
+ - - + -
- - - + -
+ - ]
3735 : 1 : relids_logged = lappend_oid(relids_logged, rel->localreloid);
3736 : :
3737 : : /*
3738 : : * Truncate partitions if we got a message to truncate a partitioned
3739 : : * table.
3740 : : */
3741 [ + + ]: 29 : if (rel->localrel->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
3742 : : {
3743 : : ListCell *child;
3744 : 4 : List *children = find_all_inheritors(rel->localreloid,
3745 : : lockmode,
3746 : : NULL);
3747 : :
3748 [ + - + + : 15 : foreach(child, children)
+ + ]
3749 : : {
3750 : 11 : Oid childrelid = lfirst_oid(child);
3751 : : Relation childrel;
3752 : :
3753 [ + + ]: 11 : if (list_member_oid(relids, childrelid))
3754 : 4 : continue;
3755 : :
3756 : : /* find_all_inheritors already got lock */
3757 : 7 : childrel = table_open(childrelid, NoLock);
3758 : :
3759 : : /*
3760 : : * Ignore temp tables of other backends. See similar code in
3761 : : * ExecuteTruncate().
3762 : : */
3763 [ - + - - ]: 7 : if (RELATION_IS_OTHER_TEMP(childrel))
3764 : : {
3765 : 0 : table_close(childrel, lockmode);
3766 : 0 : continue;
3767 : : }
3768 : :
3769 : 7 : TargetPrivilegesCheck(childrel, ACL_TRUNCATE);
3770 : 7 : rels = lappend(rels, childrel);
3771 : 7 : part_rels = lappend(part_rels, childrel);
3772 : 7 : relids = lappend_oid(relids, childrelid);
3773 : : /* Log this relation only if needed for logical decoding */
3774 [ + - - + : 7 : if (RelationIsLogicallyLogged(childrel))
- - - - -
- - - - -
- - ]
3775 : 0 : relids_logged = lappend_oid(relids_logged, childrelid);
3776 : : }
3777 : : }
3778 : : }
3779 : :
3780 : : /*
3781 : : * Even if we used CASCADE on the upstream primary we explicitly default
3782 : : * to replaying changes without further cascading. This might be later
3783 : : * changeable with a user specified option.
3784 : : *
3785 : : * MySubscription->runasowner tells us whether we want to execute
3786 : : * replication actions as the subscription owner; the last argument to
3787 : : * TruncateGuts tells it whether we want to switch to the table owner.
3788 : : * Those are exactly opposite conditions.
3789 : : */
3790 : 20 : ExecuteTruncateGuts(rels,
3791 : : relids,
3792 : : relids_logged,
3793 : : DROP_RESTRICT,
3794 : : restart_seqs,
3795 : 20 : !MySubscription->runasowner);
3796 [ + - + + : 49 : foreach(lc, remote_rels)
+ + ]
3797 : : {
3798 : 29 : LogicalRepRelMapEntry *rel = lfirst(lc);
3799 : :
3800 : 29 : logicalrep_rel_close(rel, NoLock);
3801 : : }
3802 [ + + + + : 27 : foreach(lc, part_rels)
+ + ]
3803 : : {
3804 : 7 : Relation rel = lfirst(lc);
3805 : :
3806 : 7 : table_close(rel, NoLock);
3807 : : }
3808 : :
3809 : 20 : end_replication_step();
3810 : : }
3811 : :
3812 : :
3813 : : /*
3814 : : * Logical replication protocol message dispatcher.
3815 : : */
3816 : : void
3817 : 361132 : apply_dispatch(StringInfo s)
3818 : : {
3819 : 361132 : LogicalRepMsgType action = pq_getmsgbyte(s);
3820 : : LogicalRepMsgType saved_command;
3821 : :
3822 : : /*
3823 : : * Set the current command being applied. Since this function can be
3824 : : * called recursively when applying spooled changes, save the current
3825 : : * command.
3826 : : */
3827 : 361132 : saved_command = remote_ctx.command;
3828 : 361132 : remote_ctx.command = action;
3829 : :
3830 [ + + + + : 361132 : switch (action)
+ + + + +
- + + + +
+ + + + +
- ]
3831 : : {
3832 : 544 : case LOGICAL_REP_MSG_BEGIN:
3833 : 544 : apply_handle_begin(s);
3834 : 544 : break;
3835 : :
3836 : 468 : case LOGICAL_REP_MSG_COMMIT:
3837 : 468 : apply_handle_commit(s);
3838 : 466 : break;
3839 : :
3840 : 209572 : case LOGICAL_REP_MSG_INSERT:
3841 : 209572 : apply_handle_insert(s);
3842 : 209505 : break;
3843 : :
3844 : 66178 : case LOGICAL_REP_MSG_UPDATE:
3845 : 66178 : apply_handle_update(s);
3846 : 66167 : break;
3847 : :
3848 : 81942 : case LOGICAL_REP_MSG_DELETE:
3849 : 81942 : apply_handle_delete(s);
3850 : 81942 : break;
3851 : :
3852 : 20 : case LOGICAL_REP_MSG_TRUNCATE:
3853 : 20 : apply_handle_truncate(s);
3854 : 20 : break;
3855 : :
3856 : 522 : case LOGICAL_REP_MSG_RELATION:
3857 : 522 : apply_handle_relation(s);
3858 : 522 : break;
3859 : :
3860 : 18 : case LOGICAL_REP_MSG_TYPE:
3861 : 18 : apply_handle_type(s);
3862 : 18 : break;
3863 : :
3864 : 7 : case LOGICAL_REP_MSG_ORIGIN:
3865 : 7 : apply_handle_origin(s);
3866 : 7 : break;
3867 : :
3868 : 0 : case LOGICAL_REP_MSG_MESSAGE:
3869 : :
3870 : : /*
3871 : : * Logical replication does not use generic logical messages yet.
3872 : : * Although, it could be used by other applications that use this
3873 : : * output plugin.
3874 : : */
3875 : 0 : break;
3876 : :
3877 : 843 : case LOGICAL_REP_MSG_STREAM_START:
3878 : 843 : apply_handle_stream_start(s);
3879 : 843 : break;
3880 : :
3881 : 842 : case LOGICAL_REP_MSG_STREAM_STOP:
3882 : 842 : apply_handle_stream_stop(s);
3883 : 840 : break;
3884 : :
3885 : 38 : case LOGICAL_REP_MSG_STREAM_ABORT:
3886 : 38 : apply_handle_stream_abort(s);
3887 : 38 : break;
3888 : :
3889 : 61 : case LOGICAL_REP_MSG_STREAM_COMMIT:
3890 : 61 : apply_handle_stream_commit(s);
3891 : 59 : break;
3892 : :
3893 : 18 : case LOGICAL_REP_MSG_BEGIN_PREPARE:
3894 : 18 : apply_handle_begin_prepare(s);
3895 : 18 : break;
3896 : :
3897 : 17 : case LOGICAL_REP_MSG_PREPARE:
3898 : 17 : apply_handle_prepare(s);
3899 : 15 : break;
3900 : :
3901 : 22 : case LOGICAL_REP_MSG_COMMIT_PREPARED:
3902 : 22 : apply_handle_commit_prepared(s);
3903 : 22 : break;
3904 : :
3905 : 5 : case LOGICAL_REP_MSG_ROLLBACK_PREPARED:
3906 : 5 : apply_handle_rollback_prepared(s);
3907 : 5 : break;
3908 : :
3909 : 15 : case LOGICAL_REP_MSG_STREAM_PREPARE:
3910 : 15 : apply_handle_stream_prepare(s);
3911 : 13 : break;
3912 : :
3913 : 0 : default:
3914 [ # # ]: 0 : ereport(ERROR,
3915 : : (errcode(ERRCODE_PROTOCOL_VIOLATION),
3916 : : errmsg("invalid logical replication message type \"??? (%d)\"", action)));
3917 : : }
3918 : :
3919 : : /* Reset the current command */
3920 : 361044 : remote_ctx.command = saved_command;
3921 : 361044 : }
3922 : :
3923 : : /*
3924 : : * Figure out which write/flush positions to report to the walsender process.
3925 : : *
3926 : : * We can't simply report back the last LSN the walsender sent us because the
3927 : : * local transaction might not yet be flushed to disk locally. Instead we
3928 : : * build a list that associates local with remote LSNs for every commit. When
3929 : : * reporting back the flush position to the sender we iterate that list and
3930 : : * check which entries on it are already locally flushed. Those we can report
3931 : : * as having been flushed.
3932 : : *
3933 : : * The have_pending_txes is true if there are outstanding transactions that
3934 : : * need to be flushed.
3935 : : */
3936 : : static void
3937 : 33591 : get_flush_position(XLogRecPtr *write, XLogRecPtr *flush,
3938 : : bool *have_pending_txes)
3939 : : {
3940 : : dlist_mutable_iter iter;
3941 : 33591 : XLogRecPtr local_flush = GetFlushRecPtr(NULL);
3942 : :
3943 : 33591 : *write = InvalidXLogRecPtr;
3944 : 33591 : *flush = InvalidXLogRecPtr;
3945 : :
3946 [ + - + + ]: 34134 : dlist_foreach_modify(iter, &lsn_mapping)
3947 : : {
3948 : 4324 : FlushPosition *pos =
3949 : : dlist_container(FlushPosition, node, iter.cur);
3950 : :
3951 : 4324 : *write = pos->remote_end;
3952 : :
3953 [ + + ]: 4324 : if (pos->local_end <= local_flush)
3954 : : {
3955 : 543 : *flush = pos->remote_end;
3956 : 543 : dlist_delete(iter.cur);
3957 : 543 : pfree(pos);
3958 : : }
3959 : : else
3960 : : {
3961 : : /*
3962 : : * Don't want to uselessly iterate over the rest of the list which
3963 : : * could potentially be long. Instead get the last element and
3964 : : * grab the write position from there.
3965 : : */
3966 : 3781 : pos = dlist_tail_element(FlushPosition, node,
3967 : : &lsn_mapping);
3968 : 3781 : *write = pos->remote_end;
3969 : 3781 : *have_pending_txes = true;
3970 : 3781 : return;
3971 : : }
3972 : : }
3973 : :
3974 : 29810 : *have_pending_txes = !dlist_is_empty(&lsn_mapping);
3975 : : }
3976 : :
3977 : : /*
3978 : : * Store current remote/local lsn pair in the tracking list.
3979 : : */
3980 : : void
3981 : 578 : store_flush_position(XLogRecPtr remote_lsn, XLogRecPtr local_lsn)
3982 : : {
3983 : : FlushPosition *flushpos;
3984 : :
3985 : : /*
3986 : : * Skip for parallel apply workers, because the lsn_mapping is maintained
3987 : : * by the leader apply worker.
3988 : : */
3989 [ + + ]: 578 : if (am_parallel_apply_worker())
3990 : 19 : return;
3991 : :
3992 : : /* Need to do this in permanent context */
3993 : 559 : MemoryContextSwitchTo(ApplyContext);
3994 : :
3995 : : /* Track commit lsn */
3996 : 559 : flushpos = palloc_object(FlushPosition);
3997 : 559 : flushpos->local_end = local_lsn;
3998 : 559 : flushpos->remote_end = remote_lsn;
3999 : :
4000 : 559 : dlist_push_tail(&lsn_mapping, &flushpos->node);
4001 : 559 : MemoryContextSwitchTo(ApplyMessageContext);
4002 : : }
4003 : :
4004 : :
4005 : : /* Update statistics of the worker. */
4006 : : static void
4007 : 220800 : UpdateWorkerStats(XLogRecPtr last_lsn, TimestampTz send_time, bool reply)
4008 : : {
4009 : 220800 : MyLogicalRepWorker->last_lsn = last_lsn;
4010 : 220800 : MyLogicalRepWorker->last_send_time = send_time;
4011 : 220800 : MyLogicalRepWorker->last_recv_time = GetCurrentTimestamp();
4012 [ + + ]: 220800 : if (reply)
4013 : : {
4014 : 1918 : MyLogicalRepWorker->reply_lsn = last_lsn;
4015 : 1918 : MyLogicalRepWorker->reply_time = send_time;
4016 : : }
4017 : 220800 : }
4018 : :
4019 : : /*
4020 : : * Apply main loop.
4021 : : */
4022 : : static void
4023 : 470 : LogicalRepApplyLoop(XLogRecPtr last_received)
4024 : : {
4025 : 470 : TimestampTz last_recv_timestamp = GetCurrentTimestamp();
4026 : 470 : bool ping_sent = false;
4027 : : TimeLineID tli;
4028 : : ErrorContextCallback errcallback;
4029 : 470 : RetainDeadTuplesData rdt_data = {0};
4030 : :
4031 : : /*
4032 : : * Init the ApplyMessageContext which we clean up after each replication
4033 : : * protocol message.
4034 : : */
4035 : 470 : ApplyMessageContext = AllocSetContextCreate(ApplyContext,
4036 : : "ApplyMessageContext",
4037 : : ALLOCSET_DEFAULT_SIZES);
4038 : :
4039 : : /*
4040 : : * This memory context is used for per-stream data when the streaming mode
4041 : : * is enabled. This context is reset on each stream stop.
4042 : : */
4043 : 470 : LogicalStreamingContext = AllocSetContextCreate(ApplyContext,
4044 : : "LogicalStreamingContext",
4045 : : ALLOCSET_DEFAULT_SIZES);
4046 : :
4047 : : /* mark as idle, before starting to loop */
4048 : 470 : pgstat_report_activity(STATE_IDLE, NULL);
4049 : :
4050 : : /*
4051 : : * Push apply error context callback. Fields will be filled while applying
4052 : : * a change.
4053 : : */
4054 : 470 : errcallback.callback = apply_error_callback;
4055 : 470 : errcallback.previous = error_context_stack;
4056 : 470 : error_context_stack = &errcallback;
4057 : 470 : apply_error_context_stack = error_context_stack;
4058 : :
4059 : : /* This outer loop iterates once per wait. */
4060 : : for (;;)
4061 : 30853 : {
4062 : 31323 : pgsocket fd = PGINVALID_SOCKET;
4063 : : int rc;
4064 : : int len;
4065 : 31323 : char *buf = NULL;
4066 : 31323 : bool endofstream = false;
4067 : : long wait_time;
4068 : :
4069 [ - + ]: 31323 : CHECK_FOR_INTERRUPTS();
4070 : :
4071 : 31323 : MemoryContextSwitchTo(ApplyMessageContext);
4072 : :
4073 : 31323 : len = walrcv_receive(LogRepWorkerWalRcvConn, &buf, &fd);
4074 : :
4075 [ + + ]: 31305 : if (len != 0)
4076 : : {
4077 : : /* Loop to process all available data (without blocking). */
4078 : : for (;;)
4079 : : {
4080 [ - + ]: 250718 : CHECK_FOR_INTERRUPTS();
4081 : :
4082 [ + + ]: 250718 : if (len == 0)
4083 : : {
4084 : 29903 : break;
4085 : : }
4086 [ + + ]: 220815 : else if (len < 0)
4087 : : {
4088 [ + - ]: 14 : ereport(LOG,
4089 : : (errmsg("data stream from publisher has ended")));
4090 : 14 : endofstream = true;
4091 : 14 : break;
4092 : : }
4093 : : else
4094 : : {
4095 : : int c;
4096 : : StringInfoData s;
4097 : :
4098 [ - + ]: 220801 : if (ConfigReloadPending)
4099 : : {
4100 : 0 : ConfigReloadPending = false;
4101 : 0 : ProcessConfigFile(PGC_SIGHUP);
4102 : : }
4103 : :
4104 : : /* Reset timeout. */
4105 : 220801 : last_recv_timestamp = GetCurrentTimestamp();
4106 : 220801 : ping_sent = false;
4107 : :
4108 : 220801 : rdt_data.last_recv_time = last_recv_timestamp;
4109 : :
4110 : : /* Ensure we are reading the data into our memory context. */
4111 : 220801 : MemoryContextSwitchTo(ApplyMessageContext);
4112 : :
4113 : 220801 : initReadOnlyStringInfo(&s, buf, len);
4114 : :
4115 : 220801 : c = pq_getmsgbyte(&s);
4116 : :
4117 [ + + ]: 220801 : if (c == PqReplMsg_WALData)
4118 : : {
4119 : : XLogRecPtr start_lsn;
4120 : : XLogRecPtr end_lsn;
4121 : : TimestampTz send_time;
4122 : :
4123 : 208776 : start_lsn = pq_getmsgint64(&s);
4124 : 208776 : end_lsn = pq_getmsgint64(&s);
4125 : 208776 : send_time = pq_getmsgint64(&s);
4126 : :
4127 [ + + ]: 208776 : if (last_received < start_lsn)
4128 : 165222 : last_received = start_lsn;
4129 : :
4130 [ - + ]: 208776 : if (last_received < end_lsn)
4131 : 0 : last_received = end_lsn;
4132 : :
4133 : 208776 : UpdateWorkerStats(last_received, send_time, false);
4134 : :
4135 : 208776 : apply_dispatch(&s);
4136 : :
4137 : 208693 : maybe_advance_nonremovable_xid(&rdt_data, false);
4138 : : }
4139 [ + + ]: 12025 : else if (c == PqReplMsg_Keepalive)
4140 : : {
4141 : : XLogRecPtr end_lsn;
4142 : : TimestampTz timestamp;
4143 : : bool reply_requested;
4144 : :
4145 : 1919 : end_lsn = pq_getmsgint64(&s);
4146 : 1919 : timestamp = pq_getmsgint64(&s);
4147 : 1919 : reply_requested = pq_getmsgbyte(&s);
4148 : :
4149 [ + + ]: 1919 : if (last_received < end_lsn)
4150 : 1086 : last_received = end_lsn;
4151 : :
4152 : 1919 : send_feedback(last_received, reply_requested, false);
4153 : :
4154 : 1918 : maybe_advance_nonremovable_xid(&rdt_data, false);
4155 : :
4156 : 1918 : UpdateWorkerStats(last_received, timestamp, true);
4157 : : }
4158 [ + - ]: 10106 : else if (c == PqReplMsg_PrimaryStatusUpdate)
4159 : : {
4160 : 10106 : rdt_data.remote_lsn = pq_getmsgint64(&s);
4161 : 10106 : rdt_data.remote_oldestxid = FullTransactionIdFromU64((uint64) pq_getmsgint64(&s));
4162 : 10106 : rdt_data.remote_nextxid = FullTransactionIdFromU64((uint64) pq_getmsgint64(&s));
4163 : 10106 : rdt_data.reply_time = pq_getmsgint64(&s);
4164 : :
4165 : : /*
4166 : : * This should never happen, see
4167 : : * ProcessStandbyPSRequestMessage. But if it happens
4168 : : * due to a bug, we don't want to proceed as it can
4169 : : * incorrectly advance oldest_nonremovable_xid.
4170 : : */
4171 [ - + ]: 10106 : if (!XLogRecPtrIsValid(rdt_data.remote_lsn))
4172 [ # # ]: 0 : elog(ERROR, "cannot get the latest WAL position from the publisher");
4173 : :
4174 : 10106 : maybe_advance_nonremovable_xid(&rdt_data, true);
4175 : :
4176 : 10106 : UpdateWorkerStats(last_received, rdt_data.reply_time, false);
4177 : : }
4178 : : /* other message types are purposefully ignored */
4179 : :
4180 : 220717 : MemoryContextReset(ApplyMessageContext);
4181 : : }
4182 : :
4183 : 220717 : len = walrcv_receive(LogRepWorkerWalRcvConn, &buf, &fd);
4184 : : }
4185 : : }
4186 : :
4187 : : /* confirm all writes so far */
4188 : 31220 : send_feedback(last_received, false, false);
4189 : :
4190 : : /* Reset the timestamp if no message was received */
4191 : 31220 : rdt_data.last_recv_time = 0;
4192 : :
4193 : 31220 : maybe_advance_nonremovable_xid(&rdt_data, false);
4194 : :
4195 [ + + + + ]: 31219 : if (!in_remote_transaction && !in_streamed_transaction)
4196 : : {
4197 : : /*
4198 : : * If we didn't get any transactions for a while there might be
4199 : : * unconsumed invalidation messages in the queue, consume them
4200 : : * now.
4201 : : */
4202 : 8047 : AcceptInvalidationMessages();
4203 : 8047 : maybe_reread_subscription();
4204 : :
4205 : : /*
4206 : : * Process any relations that are being synchronized in parallel
4207 : : * and any newly added tables or sequences.
4208 : : */
4209 : 8002 : ProcessSyncingRelations(last_received);
4210 : : }
4211 : :
4212 : : /* Cleanup the memory. */
4213 : 30970 : MemoryContextReset(ApplyMessageContext);
4214 : 30970 : MemoryContextSwitchTo(TopMemoryContext);
4215 : :
4216 : : /* Check if we need to exit the streaming loop. */
4217 [ + + ]: 30970 : if (endofstream)
4218 : 14 : break;
4219 : :
4220 : : /*
4221 : : * Wait for more data or latch. If we have unflushed transactions,
4222 : : * wake up after WalWriterDelay to see if they've been flushed yet (in
4223 : : * which case we should send a feedback message). Otherwise, there's
4224 : : * no particular urgency about waking up unless we get data or a
4225 : : * signal.
4226 : : */
4227 [ + + ]: 30956 : if (!dlist_is_empty(&lsn_mapping))
4228 : 3305 : wait_time = WalWriterDelay;
4229 : : else
4230 : 27651 : wait_time = NAPTIME_PER_CYCLE;
4231 : :
4232 : : /*
4233 : : * Ensure to wake up when it's possible to advance the non-removable
4234 : : * transaction ID, or when the retention duration may have exceeded
4235 : : * max_retention_duration.
4236 : : */
4237 [ + + ]: 30956 : if (MySubscription->retentionactive)
4238 : : {
4239 [ + + ]: 5128 : if (rdt_data.phase == RDT_GET_CANDIDATE_XID &&
4240 [ + - ]: 133 : rdt_data.xid_advance_interval)
4241 : 133 : wait_time = Min(wait_time, rdt_data.xid_advance_interval);
4242 [ + + ]: 4995 : else if (MySubscription->maxretention > 0)
4243 : 1 : wait_time = Min(wait_time, MySubscription->maxretention);
4244 : : }
4245 : :
4246 : 30956 : rc = WaitLatchOrSocket(MyLatch,
4247 : : WL_SOCKET_READABLE | WL_LATCH_SET |
4248 : : WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
4249 : : fd, wait_time,
4250 : : WAIT_EVENT_LOGICAL_APPLY_MAIN);
4251 : :
4252 [ + + ]: 30956 : if (rc & WL_LATCH_SET)
4253 : : {
4254 : 806 : ResetLatch(MyLatch);
4255 [ + + ]: 806 : CHECK_FOR_INTERRUPTS();
4256 : : }
4257 : :
4258 [ + + ]: 30853 : if (ConfigReloadPending)
4259 : : {
4260 : 12 : ConfigReloadPending = false;
4261 : 12 : ProcessConfigFile(PGC_SIGHUP);
4262 : : }
4263 : :
4264 [ + + ]: 30853 : if (rc & WL_TIMEOUT)
4265 : : {
4266 : : /*
4267 : : * We didn't receive anything new. If we haven't heard anything
4268 : : * from the server for more than wal_receiver_timeout / 2, ping
4269 : : * the server. Also, if it's been longer than
4270 : : * wal_receiver_status_interval since the last update we sent,
4271 : : * send a status update to the primary anyway, to report any
4272 : : * progress in applying WAL.
4273 : : */
4274 : 438 : bool requestReply = false;
4275 : :
4276 : : /*
4277 : : * Check if time since last receive from primary has reached the
4278 : : * configured limit.
4279 : : */
4280 [ + - ]: 438 : if (wal_receiver_timeout > 0)
4281 : : {
4282 : 438 : TimestampTz now = GetCurrentTimestamp();
4283 : : TimestampTz timeout;
4284 : :
4285 : 438 : timeout =
4286 : 438 : TimestampTzPlusMilliseconds(last_recv_timestamp,
4287 : : wal_receiver_timeout);
4288 : :
4289 [ - + ]: 438 : if (now >= timeout)
4290 [ # # ]: 0 : ereport(ERROR,
4291 : : (errcode(ERRCODE_CONNECTION_FAILURE),
4292 : : errmsg("terminating logical replication worker due to timeout")));
4293 : :
4294 : : /* Check to see if it's time for a ping. */
4295 [ + - ]: 438 : if (!ping_sent)
4296 : : {
4297 : 438 : timeout = TimestampTzPlusMilliseconds(last_recv_timestamp,
4298 : : (wal_receiver_timeout / 2));
4299 [ - + ]: 438 : if (now >= timeout)
4300 : : {
4301 : 0 : requestReply = true;
4302 : 0 : ping_sent = true;
4303 : : }
4304 : : }
4305 : : }
4306 : :
4307 : 438 : send_feedback(last_received, requestReply, requestReply);
4308 : :
4309 : 438 : maybe_advance_nonremovable_xid(&rdt_data, false);
4310 : :
4311 : : /*
4312 : : * Force reporting to ensure long idle periods don't lead to
4313 : : * arbitrarily delayed stats. Stats can only be reported outside
4314 : : * of (implicit or explicit) transactions. That shouldn't lead to
4315 : : * stats being delayed for long, because transactions are either
4316 : : * sent as a whole on commit or streamed. Streamed transactions
4317 : : * are spilled to disk and applied on commit.
4318 : : */
4319 [ + - ]: 438 : if (!IsTransactionState())
4320 : 438 : pgstat_report_stat(true);
4321 : : }
4322 : : }
4323 : :
4324 : : /* Pop the error context stack */
4325 : 14 : error_context_stack = errcallback.previous;
4326 : 14 : apply_error_context_stack = error_context_stack;
4327 : :
4328 : : /* All done */
4329 : 14 : walrcv_endstreaming(LogRepWorkerWalRcvConn, &tli);
4330 : 0 : }
4331 : :
4332 : : /*
4333 : : * Send a Standby Status Update message to server.
4334 : : *
4335 : : * 'recvpos' is the latest LSN we've received data to, force is set if we need
4336 : : * to send a response to avoid timeouts.
4337 : : */
4338 : : static void
4339 : 33577 : send_feedback(XLogRecPtr recvpos, bool force, bool requestReply)
4340 : : {
4341 : : static StringInfo reply_message = NULL;
4342 : : static TimestampTz send_time = 0;
4343 : :
4344 : : static XLogRecPtr last_recvpos = InvalidXLogRecPtr;
4345 : : static XLogRecPtr last_writepos = InvalidXLogRecPtr;
4346 : :
4347 : : XLogRecPtr writepos;
4348 : : XLogRecPtr flushpos;
4349 : : TimestampTz now;
4350 : : bool have_pending_txes;
4351 : :
4352 : : /*
4353 : : * If the user doesn't want status to be reported to the publisher, be
4354 : : * sure to exit before doing anything at all.
4355 : : */
4356 [ + + - + ]: 33577 : if (!force && wal_receiver_status_interval <= 0)
4357 : 15229 : return;
4358 : :
4359 : : /* It's legal to not pass a recvpos */
4360 [ - + ]: 33577 : if (recvpos < last_recvpos)
4361 : 0 : recvpos = last_recvpos;
4362 : :
4363 : 33577 : get_flush_position(&writepos, &flushpos, &have_pending_txes);
4364 : :
4365 : : /*
4366 : : * No outstanding transactions to flush, we can report the latest received
4367 : : * position. This is important for synchronous replication.
4368 : : */
4369 [ + + ]: 33577 : if (!have_pending_txes)
4370 : 29799 : flushpos = writepos = recvpos;
4371 : :
4372 [ - + ]: 33577 : if (writepos < last_writepos)
4373 : 0 : writepos = last_writepos;
4374 : :
4375 [ + + ]: 33577 : if (flushpos < last_flushpos)
4376 : 3724 : flushpos = last_flushpos;
4377 : :
4378 : 33577 : now = GetCurrentTimestamp();
4379 : :
4380 : : /* if we've already reported everything we're good */
4381 [ + + ]: 33577 : if (!force &&
4382 [ + + ]: 33574 : writepos == last_writepos &&
4383 [ + + ]: 15554 : flushpos == last_flushpos &&
4384 [ + + ]: 15323 : !TimestampDifferenceExceeds(send_time, now,
4385 : : wal_receiver_status_interval * 1000))
4386 : 15229 : return;
4387 : 18348 : send_time = now;
4388 : :
4389 [ + + ]: 18348 : if (!reply_message)
4390 : : {
4391 : 470 : MemoryContext oldctx = MemoryContextSwitchTo(ApplyContext);
4392 : :
4393 : 470 : reply_message = makeStringInfo();
4394 : 470 : MemoryContextSwitchTo(oldctx);
4395 : : }
4396 : : else
4397 : 17878 : resetStringInfo(reply_message);
4398 : :
4399 : 18348 : pq_sendbyte(reply_message, PqReplMsg_StandbyStatusUpdate);
4400 : 18348 : pq_sendint64(reply_message, recvpos); /* write */
4401 : 18348 : pq_sendint64(reply_message, flushpos); /* flush */
4402 : 18348 : pq_sendint64(reply_message, writepos); /* apply */
4403 : 18348 : pq_sendint64(reply_message, now); /* sendTime */
4404 : 18348 : pq_sendbyte(reply_message, requestReply); /* replyRequested */
4405 : :
4406 [ + + ]: 18348 : elog(DEBUG2, "sending feedback (force %d) to recv %X/%08X, write %X/%08X, flush %X/%08X",
4407 : : force,
4408 : : LSN_FORMAT_ARGS(recvpos),
4409 : : LSN_FORMAT_ARGS(writepos),
4410 : : LSN_FORMAT_ARGS(flushpos));
4411 : :
4412 : 18348 : walrcv_send(LogRepWorkerWalRcvConn,
4413 : : reply_message->data, reply_message->len);
4414 : :
4415 [ + + ]: 18347 : if (recvpos > last_recvpos)
4416 : 18021 : last_recvpos = recvpos;
4417 [ + + ]: 18347 : if (writepos > last_writepos)
4418 : 18022 : last_writepos = writepos;
4419 [ + + ]: 18347 : if (flushpos > last_flushpos)
4420 : 17862 : last_flushpos = flushpos;
4421 : : }
4422 : :
4423 : : /*
4424 : : * Attempt to advance the non-removable transaction ID.
4425 : : *
4426 : : * See comments atop worker.c for details.
4427 : : */
4428 : : static void
4429 : 252375 : maybe_advance_nonremovable_xid(RetainDeadTuplesData *rdt_data,
4430 : : bool status_received)
4431 : : {
4432 [ + + ]: 252375 : if (!can_advance_nonremovable_xid(rdt_data))
4433 : 236886 : return;
4434 : :
4435 : 15489 : process_rdt_phase_transition(rdt_data, status_received);
4436 : : }
4437 : :
4438 : : /*
4439 : : * Preliminary check to determine if advancing the non-removable transaction ID
4440 : : * is allowed.
4441 : : */
4442 : : static bool
4443 : 252375 : can_advance_nonremovable_xid(RetainDeadTuplesData *rdt_data)
4444 : : {
4445 : : /*
4446 : : * It is sufficient to manage non-removable transaction ID for a
4447 : : * subscription by the main apply worker to detect update_deleted reliably
4448 : : * even for table sync or parallel apply workers.
4449 : : */
4450 [ + + ]: 252375 : if (!am_leader_apply_worker())
4451 : 3397 : return false;
4452 : :
4453 : : /* No need to advance if retaining dead tuples is not required */
4454 [ + + ]: 248978 : if (!MySubscription->retaindeadtuples)
4455 : 233489 : return false;
4456 : :
4457 : 15489 : return true;
4458 : : }
4459 : :
4460 : : /*
4461 : : * Process phase transitions during the non-removable transaction ID
4462 : : * advancement. See comments atop worker.c for details of the transition.
4463 : : */
4464 : : static void
4465 : 25678 : process_rdt_phase_transition(RetainDeadTuplesData *rdt_data,
4466 : : bool status_received)
4467 : : {
4468 [ + + + + : 25678 : switch (rdt_data->phase)
+ + - ]
4469 : : {
4470 : 382 : case RDT_GET_CANDIDATE_XID:
4471 : 382 : get_candidate_xid(rdt_data);
4472 : 382 : break;
4473 : 10107 : case RDT_REQUEST_PUBLISHER_STATUS:
4474 : 10107 : request_publisher_status(rdt_data);
4475 : 10107 : break;
4476 : 15081 : case RDT_WAIT_FOR_PUBLISHER_STATUS:
4477 : 15081 : wait_for_publisher_status(rdt_data, status_received);
4478 : 15081 : break;
4479 : 106 : case RDT_WAIT_FOR_LOCAL_FLUSH:
4480 : 106 : wait_for_local_flush(rdt_data);
4481 : 106 : break;
4482 : 1 : case RDT_STOP_CONFLICT_INFO_RETENTION:
4483 : 1 : stop_conflict_info_retention(rdt_data);
4484 : 1 : break;
4485 : 1 : case RDT_RESUME_CONFLICT_INFO_RETENTION:
4486 : 1 : resume_conflict_info_retention(rdt_data);
4487 : 0 : break;
4488 : : }
4489 : 25677 : }
4490 : :
4491 : : /*
4492 : : * Workhorse for the RDT_GET_CANDIDATE_XID phase.
4493 : : */
4494 : : static void
4495 : 382 : get_candidate_xid(RetainDeadTuplesData *rdt_data)
4496 : : {
4497 : : TransactionId oldest_running_xid;
4498 : : TimestampTz now;
4499 : :
4500 : : /*
4501 : : * Use last_recv_time when applying changes in the loop to avoid
4502 : : * unnecessary system time retrieval. If last_recv_time is not available,
4503 : : * obtain the current timestamp.
4504 : : */
4505 [ + + ]: 382 : now = rdt_data->last_recv_time ? rdt_data->last_recv_time : GetCurrentTimestamp();
4506 : :
4507 : : /*
4508 : : * Compute the candidate_xid and request the publisher status at most once
4509 : : * per xid_advance_interval. Refer to adjust_xid_advance_interval() for
4510 : : * details on how this value is dynamically adjusted. This is to avoid
4511 : : * using CPU and network resources without making much progress.
4512 : : */
4513 [ + + ]: 382 : if (!TimestampDifferenceExceeds(rdt_data->candidate_xid_time, now,
4514 : : rdt_data->xid_advance_interval))
4515 : 291 : return;
4516 : :
4517 : : /*
4518 : : * Immediately update the timer, even if the function returns later
4519 : : * without setting candidate_xid due to inactivity on the subscriber. This
4520 : : * avoids frequent calls to GetOldestActiveTransactionId.
4521 : : */
4522 : 91 : rdt_data->candidate_xid_time = now;
4523 : :
4524 : : /*
4525 : : * Consider transactions in the current database, as only dead tuples from
4526 : : * this database are required for conflict detection.
4527 : : */
4528 : 91 : oldest_running_xid = GetOldestActiveTransactionId(false, false);
4529 : :
4530 : : /*
4531 : : * Oldest active transaction ID (oldest_running_xid) can't be behind any
4532 : : * of its previously computed value.
4533 : : */
4534 : : Assert(TransactionIdPrecedesOrEquals(MyLogicalRepWorker->oldest_nonremovable_xid,
4535 : : oldest_running_xid));
4536 : :
4537 : : /* Return if the oldest_nonremovable_xid cannot be advanced */
4538 [ + + ]: 91 : if (TransactionIdEquals(MyLogicalRepWorker->oldest_nonremovable_xid,
4539 : : oldest_running_xid))
4540 : : {
4541 : 48 : adjust_xid_advance_interval(rdt_data, false);
4542 : 48 : return;
4543 : : }
4544 : :
4545 : 43 : adjust_xid_advance_interval(rdt_data, true);
4546 : :
4547 : 43 : rdt_data->candidate_xid = oldest_running_xid;
4548 : 43 : rdt_data->phase = RDT_REQUEST_PUBLISHER_STATUS;
4549 : :
4550 : : /* process the next phase */
4551 : 43 : process_rdt_phase_transition(rdt_data, false);
4552 : : }
4553 : :
4554 : : /*
4555 : : * Workhorse for the RDT_REQUEST_PUBLISHER_STATUS phase.
4556 : : */
4557 : : static void
4558 : 10107 : request_publisher_status(RetainDeadTuplesData *rdt_data)
4559 : : {
4560 : : static StringInfo request_message = NULL;
4561 : :
4562 [ + + ]: 10107 : if (!request_message)
4563 : : {
4564 : 11 : MemoryContext oldctx = MemoryContextSwitchTo(ApplyContext);
4565 : :
4566 : 11 : request_message = makeStringInfo();
4567 : 11 : MemoryContextSwitchTo(oldctx);
4568 : : }
4569 : : else
4570 : 10096 : resetStringInfo(request_message);
4571 : :
4572 : : /*
4573 : : * Send the current time to update the remote walsender's latest reply
4574 : : * message received time.
4575 : : */
4576 : 10107 : pq_sendbyte(request_message, PqReplMsg_PrimaryStatusRequest);
4577 : 10107 : pq_sendint64(request_message, GetCurrentTimestamp());
4578 : :
4579 [ + + ]: 10107 : elog(DEBUG2, "sending publisher status request message");
4580 : :
4581 : : /* Send a request for the publisher status */
4582 : 10107 : walrcv_send(LogRepWorkerWalRcvConn,
4583 : : request_message->data, request_message->len);
4584 : :
4585 : 10107 : rdt_data->phase = RDT_WAIT_FOR_PUBLISHER_STATUS;
4586 : :
4587 : : /*
4588 : : * Skip calling maybe_advance_nonremovable_xid() since further transition
4589 : : * is possible only once we receive the publisher status message.
4590 : : */
4591 : 10107 : }
4592 : :
4593 : : /*
4594 : : * Workhorse for the RDT_WAIT_FOR_PUBLISHER_STATUS phase.
4595 : : */
4596 : : static void
4597 : 15081 : wait_for_publisher_status(RetainDeadTuplesData *rdt_data,
4598 : : bool status_received)
4599 : : {
4600 : : /*
4601 : : * Return if we have requested but not yet received the publisher status.
4602 : : */
4603 [ + + ]: 15081 : if (!status_received)
4604 : 4975 : return;
4605 : :
4606 : : /*
4607 : : * We don't need to maintain oldest_nonremovable_xid if we decide to stop
4608 : : * retaining conflict information for this worker.
4609 : : */
4610 [ - + ]: 10106 : if (should_stop_conflict_info_retention(rdt_data))
4611 : : {
4612 : 0 : rdt_data->phase = RDT_STOP_CONFLICT_INFO_RETENTION;
4613 : 0 : return;
4614 : : }
4615 : :
4616 [ + + ]: 10106 : if (!FullTransactionIdIsValid(rdt_data->remote_wait_for))
4617 : 42 : rdt_data->remote_wait_for = rdt_data->remote_nextxid;
4618 : :
4619 : : /*
4620 : : * Check if all remote concurrent transactions that were active at the
4621 : : * first status request have now completed. If completed, proceed to the
4622 : : * next phase; otherwise, continue checking the publisher status until
4623 : : * these transactions finish.
4624 : : *
4625 : : * It's possible that transactions in the commit phase during the last
4626 : : * cycle have now finished committing, but remote_oldestxid remains older
4627 : : * than remote_wait_for. This can happen if some old transaction came in
4628 : : * the commit phase when we requested status in this cycle. We do not
4629 : : * handle this case explicitly as it's rare and the benefit doesn't
4630 : : * justify the required complexity. Tracking would require either caching
4631 : : * all xids at the publisher or sending them to subscribers. The condition
4632 : : * will resolve naturally once the remaining transactions are finished.
4633 : : *
4634 : : * Directly advancing the non-removable transaction ID is possible if
4635 : : * there are no activities on the publisher since the last advancement
4636 : : * cycle. However, it requires maintaining two fields, last_remote_nextxid
4637 : : * and last_remote_lsn, within the structure for comparison with the
4638 : : * current cycle's values. Considering the minimal cost of continuing in
4639 : : * RDT_WAIT_FOR_LOCAL_FLUSH without awaiting changes, we opted not to
4640 : : * advance the transaction ID here.
4641 : : */
4642 [ + + ]: 10106 : if (FullTransactionIdPrecedesOrEquals(rdt_data->remote_wait_for,
4643 : : rdt_data->remote_oldestxid))
4644 : 42 : rdt_data->phase = RDT_WAIT_FOR_LOCAL_FLUSH;
4645 : : else
4646 : 10064 : rdt_data->phase = RDT_REQUEST_PUBLISHER_STATUS;
4647 : :
4648 : : /* process the next phase */
4649 : 10106 : process_rdt_phase_transition(rdt_data, false);
4650 : : }
4651 : :
4652 : : /*
4653 : : * Workhorse for the RDT_WAIT_FOR_LOCAL_FLUSH phase.
4654 : : */
4655 : : static void
4656 : 106 : wait_for_local_flush(RetainDeadTuplesData *rdt_data)
4657 : : {
4658 : : Assert(XLogRecPtrIsValid(rdt_data->remote_lsn) &&
4659 : : TransactionIdIsValid(rdt_data->candidate_xid));
4660 : :
4661 : : /*
4662 : : * We expect the publisher and subscriber clocks to be in sync using time
4663 : : * sync service like NTP. Otherwise, we will advance this worker's
4664 : : * oldest_nonremovable_xid prematurely, leading to the removal of rows
4665 : : * required to detect update_deleted reliably. This check primarily
4666 : : * addresses scenarios where the publisher's clock falls behind; if the
4667 : : * publisher's clock is ahead, subsequent transactions will naturally bear
4668 : : * later commit timestamps, conforming to the design outlined atop
4669 : : * worker.c.
4670 : : *
4671 : : * XXX Consider waiting for the publisher's clock to catch up with the
4672 : : * subscriber's before proceeding to the next phase.
4673 : : */
4674 [ - + ]: 106 : if (TimestampDifferenceExceeds(rdt_data->reply_time,
4675 : : rdt_data->candidate_xid_time, 0))
4676 [ # # ]: 0 : ereport(ERROR,
4677 : : errmsg_internal("oldest_nonremovable_xid transaction ID could be advanced prematurely"),
4678 : : errdetail_internal("The clock on the publisher is behind that of the subscriber."));
4679 : :
4680 : : /*
4681 : : * Do not attempt to advance the non-removable transaction ID when table
4682 : : * sync is in progress. During this time, changes from a single
4683 : : * transaction may be applied by multiple table sync workers corresponding
4684 : : * to the target tables. So, it's necessary for all table sync workers to
4685 : : * apply and flush the corresponding changes before advancing the
4686 : : * transaction ID, otherwise, dead tuples that are still needed for
4687 : : * conflict detection in table sync workers could be removed prematurely.
4688 : : * However, confirming the apply and flush progress across all table sync
4689 : : * workers is complex and not worth the effort, so we simply return if not
4690 : : * all tables are in the READY state.
4691 : : *
4692 : : * Advancing the transaction ID is necessary even when no tables are
4693 : : * currently subscribed, to avoid retaining dead tuples unnecessarily.
4694 : : * While it might seem safe to skip all phases and directly assign
4695 : : * candidate_xid to oldest_nonremovable_xid during the
4696 : : * RDT_GET_CANDIDATE_XID phase in such cases, this is unsafe. If users
4697 : : * concurrently add tables to the subscription, the apply worker may not
4698 : : * process invalidations in time. Consequently,
4699 : : * HasSubscriptionTablesCached() might miss the new tables, leading to
4700 : : * premature advancement of oldest_nonremovable_xid.
4701 : : *
4702 : : * Performing the check during RDT_WAIT_FOR_LOCAL_FLUSH is safe, as
4703 : : * invalidations are guaranteed to be processed before applying changes
4704 : : * from newly added tables while waiting for the local flush to reach
4705 : : * remote_lsn.
4706 : : *
4707 : : * Additionally, even if we check for subscription tables during
4708 : : * RDT_GET_CANDIDATE_XID, they might be dropped before reaching
4709 : : * RDT_WAIT_FOR_LOCAL_FLUSH. Therefore, it's still necessary to verify
4710 : : * subscription tables at this stage to prevent unnecessary tuple
4711 : : * retention.
4712 : : */
4713 [ + + + + ]: 106 : if (HasSubscriptionTablesCached() && !AllTablesyncsReady())
4714 : : {
4715 : : TimestampTz now;
4716 : :
4717 : 10 : now = rdt_data->last_recv_time
4718 [ + + ]: 5 : ? rdt_data->last_recv_time : GetCurrentTimestamp();
4719 : :
4720 : : /*
4721 : : * Record the time spent waiting for table sync, it is needed for the
4722 : : * timeout check in should_stop_conflict_info_retention().
4723 : : */
4724 : 5 : rdt_data->table_sync_wait_time =
4725 : 5 : TimestampDifferenceMilliseconds(rdt_data->candidate_xid_time, now);
4726 : :
4727 : 5 : return;
4728 : : }
4729 : :
4730 : : /*
4731 : : * We don't need to maintain oldest_nonremovable_xid if we decide to stop
4732 : : * retaining conflict information for this worker.
4733 : : */
4734 [ + + ]: 101 : if (should_stop_conflict_info_retention(rdt_data))
4735 : : {
4736 : 1 : rdt_data->phase = RDT_STOP_CONFLICT_INFO_RETENTION;
4737 : 1 : return;
4738 : : }
4739 : :
4740 : : /*
4741 : : * Update and check the remote flush position if we are applying changes
4742 : : * in a loop. This is done at most once per WalWriterDelay to avoid
4743 : : * performing costly operations in get_flush_position() too frequently
4744 : : * during change application.
4745 : : */
4746 [ + + + + : 139 : if (last_flushpos < rdt_data->remote_lsn && rdt_data->last_recv_time &&
+ + ]
4747 : 39 : TimestampDifferenceExceeds(rdt_data->flushpos_update_time,
4748 : : rdt_data->last_recv_time, WalWriterDelay))
4749 : : {
4750 : : XLogRecPtr writepos;
4751 : : XLogRecPtr flushpos;
4752 : : bool have_pending_txes;
4753 : :
4754 : : /* Fetch the latest remote flush position */
4755 : 14 : get_flush_position(&writepos, &flushpos, &have_pending_txes);
4756 : :
4757 [ - + ]: 14 : if (flushpos > last_flushpos)
4758 : 0 : last_flushpos = flushpos;
4759 : :
4760 : 14 : rdt_data->flushpos_update_time = rdt_data->last_recv_time;
4761 : : }
4762 : :
4763 : : /* Return to wait for the changes to be applied */
4764 [ + + ]: 100 : if (last_flushpos < rdt_data->remote_lsn)
4765 : 59 : return;
4766 : :
4767 : : /*
4768 : : * Reaching this point implies should_stop_conflict_info_retention()
4769 : : * returned false earlier, meaning that the most recent duration for
4770 : : * advancing the non-removable transaction ID is within the
4771 : : * max_retention_duration or max_retention_duration is set to 0.
4772 : : *
4773 : : * Therefore, if conflict info retention was previously stopped due to a
4774 : : * timeout, it is now safe to resume retention.
4775 : : */
4776 [ + + ]: 41 : if (!MySubscription->retentionactive)
4777 : : {
4778 : 1 : rdt_data->phase = RDT_RESUME_CONFLICT_INFO_RETENTION;
4779 : 1 : return;
4780 : : }
4781 : :
4782 : : /*
4783 : : * Reaching here means the remote WAL position has been received, and all
4784 : : * transactions up to that position on the publisher have been applied and
4785 : : * flushed locally. So, we can advance the non-removable transaction ID.
4786 : : */
4787 : 40 : SpinLockAcquire(&MyLogicalRepWorker->relmutex);
4788 : 40 : MyLogicalRepWorker->oldest_nonremovable_xid = rdt_data->candidate_xid;
4789 : 40 : SpinLockRelease(&MyLogicalRepWorker->relmutex);
4790 : :
4791 [ + + ]: 40 : elog(DEBUG2, "confirmed flush up to remote lsn %X/%08X: new oldest_nonremovable_xid %u",
4792 : : LSN_FORMAT_ARGS(rdt_data->remote_lsn),
4793 : : rdt_data->candidate_xid);
4794 : :
4795 : : /* Notify launcher to update the xmin of the conflict slot */
4796 : 40 : ApplyLauncherWakeup();
4797 : :
4798 : 40 : reset_retention_data_fields(rdt_data);
4799 : :
4800 : : /* process the next phase */
4801 : 40 : process_rdt_phase_transition(rdt_data, false);
4802 : : }
4803 : :
4804 : : /*
4805 : : * Check whether conflict information retention should be stopped due to
4806 : : * exceeding the maximum wait time (max_retention_duration).
4807 : : *
4808 : : * If retention should be stopped, return true. Otherwise, return false.
4809 : : */
4810 : : static bool
4811 : 10207 : should_stop_conflict_info_retention(RetainDeadTuplesData *rdt_data)
4812 : : {
4813 : : TimestampTz now;
4814 : :
4815 : : Assert(TransactionIdIsValid(rdt_data->candidate_xid));
4816 : : Assert(rdt_data->phase == RDT_WAIT_FOR_PUBLISHER_STATUS ||
4817 : : rdt_data->phase == RDT_WAIT_FOR_LOCAL_FLUSH);
4818 : :
4819 [ + + ]: 10207 : if (!MySubscription->maxretention)
4820 : 10206 : return false;
4821 : :
4822 : : /*
4823 : : * Use last_recv_time when applying changes in the loop to avoid
4824 : : * unnecessary system time retrieval. If last_recv_time is not available,
4825 : : * obtain the current timestamp.
4826 : : */
4827 [ - + ]: 1 : now = rdt_data->last_recv_time ? rdt_data->last_recv_time : GetCurrentTimestamp();
4828 : :
4829 : : /*
4830 : : * Return early if the wait time has not exceeded the configured maximum
4831 : : * (max_retention_duration). Time spent waiting for table synchronization
4832 : : * is excluded from this calculation, as it occurs infrequently.
4833 : : */
4834 [ - + ]: 1 : if (!TimestampDifferenceExceeds(rdt_data->candidate_xid_time, now,
4835 : 1 : MySubscription->maxretention +
4836 : 1 : rdt_data->table_sync_wait_time))
4837 : 0 : return false;
4838 : :
4839 : 1 : return true;
4840 : : }
4841 : :
4842 : : /*
4843 : : * Workhorse for the RDT_STOP_CONFLICT_INFO_RETENTION phase.
4844 : : */
4845 : : static void
4846 : 1 : stop_conflict_info_retention(RetainDeadTuplesData *rdt_data)
4847 : : {
4848 : : /* Stop retention if not yet */
4849 [ + - ]: 1 : if (MySubscription->retentionactive)
4850 : : {
4851 : : /*
4852 : : * If the retention status cannot be updated (e.g., due to active
4853 : : * transaction), skip further processing to avoid inconsistent
4854 : : * retention behavior.
4855 : : */
4856 [ - + ]: 1 : if (!update_retention_status(false))
4857 : 0 : return;
4858 : :
4859 : 1 : SpinLockAcquire(&MyLogicalRepWorker->relmutex);
4860 : 1 : MyLogicalRepWorker->oldest_nonremovable_xid = InvalidTransactionId;
4861 : 1 : SpinLockRelease(&MyLogicalRepWorker->relmutex);
4862 : :
4863 [ + - ]: 1 : ereport(LOG,
4864 : : errmsg("logical replication worker for subscription \"%s\" has stopped retaining the information for detecting conflicts",
4865 : : MySubscription->name),
4866 : : errdetail("Retention is stopped because the apply process has not caught up with the publisher within the configured max_retention_duration."));
4867 : : }
4868 : :
4869 : : Assert(!TransactionIdIsValid(MyLogicalRepWorker->oldest_nonremovable_xid));
4870 : :
4871 : : /*
4872 : : * If retention has been stopped, reset to the initial phase to retry
4873 : : * resuming retention. This reset is required to recalculate the current
4874 : : * wait time and resume retention if the time falls within
4875 : : * max_retention_duration.
4876 : : */
4877 : 1 : reset_retention_data_fields(rdt_data);
4878 : : }
4879 : :
4880 : : /*
4881 : : * Workhorse for the RDT_RESUME_CONFLICT_INFO_RETENTION phase.
4882 : : */
4883 : : static void
4884 : 1 : resume_conflict_info_retention(RetainDeadTuplesData *rdt_data)
4885 : : {
4886 : : /* We can't resume retention without updating retention status. */
4887 [ - + ]: 1 : if (!update_retention_status(true))
4888 : 0 : return;
4889 : :
4890 [ + - - + ]: 1 : ereport(LOG,
4891 : : errmsg("logical replication worker for subscription \"%s\" will resume retaining the information for detecting conflicts",
4892 : : MySubscription->name),
4893 : : MySubscription->maxretention
4894 : : ? errdetail("Retention is re-enabled because the apply process has caught up with the publisher within the configured max_retention_duration.")
4895 : : : errdetail("Retention is re-enabled because max_retention_duration has been set to unlimited."));
4896 : :
4897 : : /*
4898 : : * Restart the worker to let the launcher initialize
4899 : : * oldest_nonremovable_xid at startup.
4900 : : *
4901 : : * While it's technically possible to derive this value on-the-fly using
4902 : : * the conflict detection slot's xmin, doing so risks a race condition:
4903 : : * the launcher might clean slot.xmin just after retention resumes. This
4904 : : * would make oldest_nonremovable_xid unreliable, especially during xid
4905 : : * wraparound.
4906 : : *
4907 : : * Although this can be prevented by introducing heavy weight locking, the
4908 : : * complexity it will bring doesn't seem worthwhile given how rarely
4909 : : * retention is resumed.
4910 : : */
4911 : 1 : apply_worker_exit();
4912 : : }
4913 : :
4914 : : /*
4915 : : * Updates pg_subscription.subretentionactive to the given value within a
4916 : : * new transaction.
4917 : : *
4918 : : * If already inside an active transaction, skips the update and returns
4919 : : * false.
4920 : : *
4921 : : * Returns true if the update is successfully performed.
4922 : : */
4923 : : static bool
4924 : 2 : update_retention_status(bool active)
4925 : : {
4926 : : /*
4927 : : * Do not update the catalog during an active transaction. The transaction
4928 : : * may be started during change application, leading to a possible
4929 : : * rollback of catalog updates if the application fails subsequently.
4930 : : */
4931 [ - + ]: 2 : if (IsTransactionState())
4932 : 0 : return false;
4933 : :
4934 : 2 : StartTransactionCommand();
4935 : :
4936 : : /*
4937 : : * Updating pg_subscription might involve TOAST table access, so ensure we
4938 : : * have a valid snapshot.
4939 : : */
4940 : 2 : PushActiveSnapshot(GetTransactionSnapshot());
4941 : :
4942 : : /* Update pg_subscription.subretentionactive */
4943 : 2 : UpdateDeadTupleRetentionStatus(MySubscription->oid, active);
4944 : :
4945 : 2 : PopActiveSnapshot();
4946 : 2 : CommitTransactionCommand();
4947 : :
4948 : : /* Notify launcher to update the conflict slot */
4949 : 2 : ApplyLauncherWakeup();
4950 : :
4951 : 2 : MySubscription->retentionactive = active;
4952 : :
4953 : 2 : return true;
4954 : : }
4955 : :
4956 : : /*
4957 : : * Reset all data fields of RetainDeadTuplesData except those used to
4958 : : * determine the timing for the next round of transaction ID advancement. We
4959 : : * can even use flushpos_update_time in the next round to decide whether to get
4960 : : * the latest flush position.
4961 : : */
4962 : : static void
4963 : 41 : reset_retention_data_fields(RetainDeadTuplesData *rdt_data)
4964 : : {
4965 : 41 : rdt_data->phase = RDT_GET_CANDIDATE_XID;
4966 : 41 : rdt_data->remote_lsn = InvalidXLogRecPtr;
4967 : 41 : rdt_data->remote_oldestxid = InvalidFullTransactionId;
4968 : 41 : rdt_data->remote_nextxid = InvalidFullTransactionId;
4969 : 41 : rdt_data->reply_time = 0;
4970 : 41 : rdt_data->remote_wait_for = InvalidFullTransactionId;
4971 : 41 : rdt_data->candidate_xid = InvalidTransactionId;
4972 : 41 : rdt_data->table_sync_wait_time = 0;
4973 : 41 : }
4974 : :
4975 : : /*
4976 : : * Adjust the interval for advancing non-removable transaction IDs.
4977 : : *
4978 : : * If there is no activity on the node or retention has been stopped, we
4979 : : * progressively double the interval used to advance non-removable transaction
4980 : : * ID. This helps conserve CPU and network resources when there's little benefit
4981 : : * to frequent updates.
4982 : : *
4983 : : * The interval is capped by the lowest of the following:
4984 : : * - wal_receiver_status_interval (if set and retention is active),
4985 : : * - a default maximum of 3 minutes,
4986 : : * - max_retention_duration (if retention is active).
4987 : : *
4988 : : * This ensures the interval never exceeds the retention boundary, even if other
4989 : : * limits are higher. Once activity resumes on the node and the retention is
4990 : : * active, the interval is reset to lesser of 100ms and max_retention_duration,
4991 : : * allowing timely advancement of non-removable transaction ID.
4992 : : *
4993 : : * XXX The use of wal_receiver_status_interval is a bit arbitrary so we can
4994 : : * consider the other interval or a separate GUC if the need arises.
4995 : : */
4996 : : static void
4997 : 91 : adjust_xid_advance_interval(RetainDeadTuplesData *rdt_data, bool new_xid_found)
4998 : : {
4999 [ + + + + ]: 91 : if (rdt_data->xid_advance_interval && !new_xid_found)
5000 : 43 : {
5001 : 43 : int max_interval = wal_receiver_status_interval
5002 : 86 : ? wal_receiver_status_interval * 1000
5003 [ + - ]: 43 : : MAX_XID_ADVANCE_INTERVAL;
5004 : :
5005 : : /*
5006 : : * No new transaction ID has been assigned since the last check, so
5007 : : * double the interval, but not beyond the maximum allowable value.
5008 : : */
5009 : 43 : rdt_data->xid_advance_interval = Min(rdt_data->xid_advance_interval * 2,
5010 : : max_interval);
5011 : : }
5012 [ + + ]: 48 : else if (rdt_data->xid_advance_interval &&
5013 [ + + ]: 35 : !MySubscription->retentionactive)
5014 : : {
5015 : : /*
5016 : : * Retention has been stopped, so double the interval-capped at a
5017 : : * maximum of 3 minutes. The wal_receiver_status_interval is
5018 : : * intentionally not used as an upper bound, since the likelihood of
5019 : : * retention resuming is lower than that of general activity resuming.
5020 : : */
5021 : 1 : rdt_data->xid_advance_interval = Min(rdt_data->xid_advance_interval * 2,
5022 : : MAX_XID_ADVANCE_INTERVAL);
5023 : : }
5024 : : else
5025 : : {
5026 : : /*
5027 : : * A new transaction ID was found or the interval is not yet
5028 : : * initialized, so set the interval to the minimum value.
5029 : : */
5030 : 47 : rdt_data->xid_advance_interval = MIN_XID_ADVANCE_INTERVAL;
5031 : : }
5032 : :
5033 : : /*
5034 : : * Ensure the wait time remains within the maximum retention time limit
5035 : : * when retention is active. Skip this cap when maxretention is zero,
5036 : : * which means unlimited retention (no timeout).
5037 : : */
5038 [ + + - + ]: 91 : if (MySubscription->retentionactive && MySubscription->maxretention > 0)
5039 : 0 : rdt_data->xid_advance_interval = Min(rdt_data->xid_advance_interval,
5040 : : MySubscription->maxretention);
5041 : 91 : }
5042 : :
5043 : : /*
5044 : : * Exit routine for apply workers due to subscription parameter changes.
5045 : : */
5046 : : static void
5047 : 48 : apply_worker_exit(void)
5048 : : {
5049 [ - + ]: 48 : if (am_parallel_apply_worker())
5050 : : {
5051 : : /*
5052 : : * Don't stop the parallel apply worker as the leader will detect the
5053 : : * subscription parameter change and restart logical replication later
5054 : : * anyway. This also prevents the leader from reporting errors when
5055 : : * trying to communicate with a stopped parallel apply worker, which
5056 : : * would accidentally disable subscriptions if disable_on_error was
5057 : : * set.
5058 : : */
5059 : 0 : return;
5060 : : }
5061 : :
5062 : : /*
5063 : : * Reset the last-start time for this apply worker so that the launcher
5064 : : * will restart it without waiting for wal_retrieve_retry_interval if the
5065 : : * subscription is still active, and so that we won't leak that hash table
5066 : : * entry if it isn't.
5067 : : */
5068 [ + - ]: 48 : if (am_leader_apply_worker())
5069 : 48 : ApplyLauncherForgetWorkerStartTime(MyLogicalRepWorker->subid);
5070 : :
5071 : 48 : proc_exit(0);
5072 : : }
5073 : :
5074 : : /*
5075 : : * Reread subscription info if needed.
5076 : : *
5077 : : * For significant changes, we react by exiting the current process; a new
5078 : : * one will be launched afterwards if needed.
5079 : : */
5080 : : void
5081 : 9158 : maybe_reread_subscription(void)
5082 : : {
5083 : : Subscription *newsub;
5084 : : char *old_conninfo;
5085 : : char *new_conninfo;
5086 : 9158 : bool started_tx = false;
5087 : :
5088 : : /* When cache state is valid there is nothing to do here. */
5089 [ + + ]: 9158 : if (MySubscriptionValid)
5090 : 9063 : return;
5091 : :
5092 : : /* This function might be called inside or outside of transaction. */
5093 [ + + ]: 95 : if (!IsTransactionState())
5094 : : {
5095 : 90 : StartTransactionCommand();
5096 : 90 : started_tx = true;
5097 : : }
5098 : :
5099 : 95 : newsub = GetSubscription(MyLogicalRepWorker->subid, true);
5100 : :
5101 [ + - ]: 95 : if (newsub)
5102 : : {
5103 : 95 : MemoryContextSetParent(newsub->cxt, ApplyContext);
5104 : : }
5105 : : else
5106 : : {
5107 : : /*
5108 : : * Exit if the subscription was removed. This normally should not
5109 : : * happen as the worker gets killed during DROP SUBSCRIPTION.
5110 : : */
5111 [ # # ]: 0 : ereport(LOG,
5112 : : (errmsg("logical replication worker for subscription \"%s\" will stop because the subscription was removed",
5113 : : MySubscription->name)));
5114 : :
5115 : : /* Ensure we remove no-longer-useful entry for worker's start time */
5116 [ # # ]: 0 : if (am_leader_apply_worker())
5117 : 0 : ApplyLauncherForgetWorkerStartTime(MyLogicalRepWorker->subid);
5118 : :
5119 : 0 : proc_exit(0);
5120 : : }
5121 : :
5122 : : /* Exit if the subscription was disabled. */
5123 [ + + ]: 95 : if (!newsub->enabled)
5124 : : {
5125 [ + - ]: 14 : ereport(LOG,
5126 : : (errmsg("logical replication worker for subscription \"%s\" will stop because the subscription was disabled",
5127 : : MySubscription->name)));
5128 : :
5129 : 14 : apply_worker_exit();
5130 : : }
5131 : :
5132 : : /*
5133 : : * May raise error, so build conninfo after checking that the subscription
5134 : : * is enabled. Allocated in transaction context; must be copied to
5135 : : * ApplyContext when we set MySubscriptionConninfo.
5136 : : */
5137 : 81 : new_conninfo = SubscriptionConninfo(newsub);
5138 : :
5139 : : /* !slotname should never happen when enabled is true. */
5140 : : Assert(newsub->slotname);
5141 : :
5142 : : /* two-phase cannot be altered while the worker is running */
5143 : : Assert(newsub->twophasestate == MySubscription->twophasestate);
5144 : :
5145 : : /*
5146 : : * Exit if any parameter that affects the remote connection was changed.
5147 : : * The launcher will start a new worker but note that the parallel apply
5148 : : * worker won't restart if the streaming option's value is changed from
5149 : : * 'parallel' to any other value or the server decides not to stream the
5150 : : * in-progress transaction.
5151 : : */
5152 [ + + ]: 81 : if (strcmp(new_conninfo, MySubscriptionConninfo) != 0 ||
5153 [ + + ]: 76 : strcmp(newsub->name, MySubscription->name) != 0 ||
5154 [ + - ]: 75 : strcmp(newsub->slotname, MySubscription->slotname) != 0 ||
5155 [ + + ]: 75 : newsub->binary != MySubscription->binary ||
5156 [ + + ]: 69 : newsub->stream != MySubscription->stream ||
5157 [ + - ]: 64 : newsub->passwordrequired != MySubscription->passwordrequired ||
5158 [ + + ]: 64 : strcmp(newsub->origin, MySubscription->origin) != 0 ||
5159 [ + + ]: 62 : newsub->owner != MySubscription->owner ||
5160 [ + + ]: 61 : !equal(newsub->publications, MySubscription->publications))
5161 : : {
5162 [ - + ]: 29 : if (am_parallel_apply_worker())
5163 [ # # ]: 0 : ereport(LOG,
5164 : : (errmsg("logical replication parallel apply worker for subscription \"%s\" will stop because of a parameter change",
5165 : : MySubscription->name)));
5166 : : else
5167 [ + - ]: 29 : ereport(LOG,
5168 : : (errmsg("logical replication worker for subscription \"%s\" will restart because of a parameter change",
5169 : : MySubscription->name)));
5170 : :
5171 : 29 : apply_worker_exit();
5172 : : }
5173 : :
5174 : : /*
5175 : : * Exit if the subscription owner's superuser privileges have been
5176 : : * revoked.
5177 : : */
5178 [ + + + + ]: 52 : if (!newsub->ownersuperuser && MySubscription->ownersuperuser)
5179 : : {
5180 [ - + ]: 4 : if (am_parallel_apply_worker())
5181 [ # # ]: 0 : ereport(LOG,
5182 : : errmsg("logical replication parallel apply worker for subscription \"%s\" will stop because the subscription owner's superuser privileges have been revoked",
5183 : : MySubscription->name));
5184 : : else
5185 [ + - ]: 4 : ereport(LOG,
5186 : : errmsg("logical replication worker for subscription \"%s\" will restart because the subscription owner's superuser privileges have been revoked",
5187 : : MySubscription->name));
5188 : :
5189 : 4 : apply_worker_exit();
5190 : : }
5191 : :
5192 : : /* Check for other changes that should never happen too. */
5193 [ - + ]: 48 : if (newsub->dbid != MySubscription->dbid)
5194 : : {
5195 [ # # ]: 0 : elog(ERROR, "subscription %u changed unexpectedly",
5196 : : MyLogicalRepWorker->subid);
5197 : : }
5198 : :
5199 : : /* Clean old subscription info and switch to new one. */
5200 : 48 : MemoryContextDelete(MySubscription->cxt);
5201 : 48 : MySubscription = newsub;
5202 : :
5203 : : /* copy to ApplyContext and update MySubscriptionConninfo */
5204 : 48 : old_conninfo = MySubscriptionConninfo;
5205 : 48 : MySubscriptionConninfo = MemoryContextStrdup(ApplyContext, new_conninfo);
5206 : 48 : pfree(old_conninfo);
5207 : :
5208 : : /* Change synchronous commit according to the user's wishes */
5209 : 48 : SetConfigOption("synchronous_commit", MySubscription->synccommit,
5210 : : PGC_BACKEND, PGC_S_OVERRIDE);
5211 : :
5212 : : /* Change wal_receiver_timeout according to the user's wishes */
5213 : 48 : set_wal_receiver_timeout();
5214 : :
5215 [ + + ]: 48 : if (started_tx)
5216 : 45 : CommitTransactionCommand();
5217 : :
5218 : 48 : MySubscriptionValid = true;
5219 : : }
5220 : :
5221 : : /*
5222 : : * Change wal_receiver_timeout to MySubscription->walrcvtimeout.
5223 : : */
5224 : : static void
5225 : 620 : set_wal_receiver_timeout(void)
5226 : : {
5227 : : bool parsed;
5228 : : int val;
5229 : 620 : int prev_timeout = wal_receiver_timeout;
5230 : :
5231 : : /*
5232 : : * Set the wal_receiver_timeout GUC to MySubscription->walrcvtimeout,
5233 : : * which comes from the subscription's wal_receiver_timeout option. If the
5234 : : * value is -1, reset the GUC to its default, meaning it will inherit from
5235 : : * the server config, command line, or role/database settings.
5236 : : */
5237 : 620 : parsed = parse_int(MySubscription->walrcvtimeout, &val, 0, NULL);
5238 [ + - + - ]: 620 : if (parsed && val == -1)
5239 : 620 : SetConfigOption("wal_receiver_timeout", NULL,
5240 : : PGC_BACKEND, PGC_S_SESSION);
5241 : : else
5242 : 0 : SetConfigOption("wal_receiver_timeout", MySubscription->walrcvtimeout,
5243 : : PGC_BACKEND, PGC_S_SESSION);
5244 : :
5245 : : /*
5246 : : * Log the wal_receiver_timeout setting (in milliseconds) as a debug
5247 : : * message when it changes, to verify it was set correctly.
5248 : : */
5249 [ - + ]: 620 : if (prev_timeout != wal_receiver_timeout)
5250 [ # # ]: 0 : elog(DEBUG1, "logical replication worker for subscription \"%s\" wal_receiver_timeout: %d ms",
5251 : : MySubscription->name, wal_receiver_timeout);
5252 : 620 : }
5253 : :
5254 : : /*
5255 : : * Callback from subscription syscache invalidation. Also needed for server or
5256 : : * user mapping invalidation, which can change the connection information for
5257 : : * subscriptions that connect using a server object.
5258 : : */
5259 : : static void
5260 : 100 : subscription_change_cb(Datum arg, SysCacheIdentifier cacheid, uint32 hashvalue)
5261 : : {
5262 : 100 : MySubscriptionValid = false;
5263 : 100 : }
5264 : :
5265 : : /*
5266 : : * subxact_info_write
5267 : : * Store information about subxacts for a toplevel transaction.
5268 : : *
5269 : : * For each subxact we store offset of its first change in the main file.
5270 : : * The file is always over-written as a whole.
5271 : : *
5272 : : * XXX We should only store subxacts that were not aborted yet.
5273 : : */
5274 : : static void
5275 : 372 : subxact_info_write(Oid subid, TransactionId xid)
5276 : : {
5277 : : char path[MAXPGPATH];
5278 : : Size len;
5279 : : BufFile *fd;
5280 : :
5281 : : Assert(TransactionIdIsValid(xid));
5282 : :
5283 : : /* construct the subxact filename */
5284 : 372 : subxact_filename(path, subid, xid);
5285 : :
5286 : : /* Delete the subxacts file, if exists. */
5287 [ + + ]: 372 : if (subxact_data.nsubxacts == 0)
5288 : : {
5289 : 290 : cleanup_subxact_info();
5290 : 290 : BufFileDeleteFileSet(MyLogicalRepWorker->stream_fileset, path, true);
5291 : :
5292 : 290 : return;
5293 : : }
5294 : :
5295 : : /*
5296 : : * Create the subxact file if it not already created, otherwise open the
5297 : : * existing file.
5298 : : */
5299 : 82 : fd = BufFileOpenFileSet(MyLogicalRepWorker->stream_fileset, path, O_RDWR,
5300 : : true);
5301 [ + + ]: 82 : if (fd == NULL)
5302 : 8 : fd = BufFileCreateFileSet(MyLogicalRepWorker->stream_fileset, path);
5303 : :
5304 : 82 : len = sizeof(SubXactInfo) * subxact_data.nsubxacts;
5305 : :
5306 : : /* Write the subxact count and subxact info */
5307 : 82 : BufFileWrite(fd, &subxact_data.nsubxacts, sizeof(subxact_data.nsubxacts));
5308 : 82 : BufFileWrite(fd, subxact_data.subxacts, len);
5309 : :
5310 : 82 : BufFileClose(fd);
5311 : :
5312 : : /* free the memory allocated for subxact info */
5313 : 82 : cleanup_subxact_info();
5314 : : }
5315 : :
5316 : : /*
5317 : : * subxact_info_read
5318 : : * Restore information about subxacts of a streamed transaction.
5319 : : *
5320 : : * Read information about subxacts into the structure subxact_data that can be
5321 : : * used later.
5322 : : */
5323 : : static void
5324 : 344 : subxact_info_read(Oid subid, TransactionId xid)
5325 : : {
5326 : : char path[MAXPGPATH];
5327 : : Size len;
5328 : : BufFile *fd;
5329 : : MemoryContext oldctx;
5330 : :
5331 : : Assert(!subxact_data.subxacts);
5332 : : Assert(subxact_data.nsubxacts == 0);
5333 : : Assert(subxact_data.nsubxacts_max == 0);
5334 : :
5335 : : /*
5336 : : * If the subxact file doesn't exist that means we don't have any subxact
5337 : : * info.
5338 : : */
5339 : 344 : subxact_filename(path, subid, xid);
5340 : 344 : fd = BufFileOpenFileSet(MyLogicalRepWorker->stream_fileset, path, O_RDONLY,
5341 : : true);
5342 [ + + ]: 344 : if (fd == NULL)
5343 : 265 : return;
5344 : :
5345 : : /* read number of subxact items */
5346 : 79 : BufFileReadExact(fd, &subxact_data.nsubxacts, sizeof(subxact_data.nsubxacts));
5347 : :
5348 : 79 : len = sizeof(SubXactInfo) * subxact_data.nsubxacts;
5349 : :
5350 : : /* we keep the maximum as a power of 2 */
5351 : 79 : subxact_data.nsubxacts_max = 1 << pg_ceil_log2_32(subxact_data.nsubxacts);
5352 : :
5353 : : /*
5354 : : * Allocate subxact information in the logical streaming context. We need
5355 : : * this information during the complete stream so that we can add the sub
5356 : : * transaction info to this. On stream stop we will flush this information
5357 : : * to the subxact file and reset the logical streaming context.
5358 : : */
5359 : 79 : oldctx = MemoryContextSwitchTo(LogicalStreamingContext);
5360 : 79 : subxact_data.subxacts = palloc_array(SubXactInfo,
5361 : : subxact_data.nsubxacts_max);
5362 : 79 : MemoryContextSwitchTo(oldctx);
5363 : :
5364 [ + - ]: 79 : if (len > 0)
5365 : 79 : BufFileReadExact(fd, subxact_data.subxacts, len);
5366 : :
5367 : 79 : BufFileClose(fd);
5368 : : }
5369 : :
5370 : : /*
5371 : : * subxact_info_add
5372 : : * Add information about a subxact (offset in the main file).
5373 : : */
5374 : : static void
5375 : 102513 : subxact_info_add(TransactionId xid)
5376 : : {
5377 : 102513 : SubXactInfo *subxacts = subxact_data.subxacts;
5378 : : int64 i;
5379 : :
5380 : : /* We must have a valid top level stream xid and a stream fd. */
5381 : : Assert(TransactionIdIsValid(stream_xid));
5382 : : Assert(stream_fd != NULL);
5383 : :
5384 : : /*
5385 : : * If the XID matches the toplevel transaction, we don't want to add it.
5386 : : */
5387 [ + + ]: 102513 : if (stream_xid == xid)
5388 : 92389 : return;
5389 : :
5390 : : /*
5391 : : * In most cases we're checking the same subxact as we've already seen in
5392 : : * the last call, so make sure to ignore it (this change comes later).
5393 : : */
5394 [ + + ]: 10124 : if (subxact_data.subxact_last == xid)
5395 : 10048 : return;
5396 : :
5397 : : /* OK, remember we're processing this XID. */
5398 : 76 : subxact_data.subxact_last = xid;
5399 : :
5400 : : /*
5401 : : * Check if the transaction is already present in the array of subxact. We
5402 : : * intentionally scan the array from the tail, because we're likely adding
5403 : : * a change for the most recent subtransactions.
5404 : : *
5405 : : * XXX Can we rely on the subxact XIDs arriving in sorted order? That
5406 : : * would allow us to use binary search here.
5407 : : */
5408 [ + + ]: 95 : for (i = subxact_data.nsubxacts; i > 0; i--)
5409 : : {
5410 : : /* found, so we're done */
5411 [ + + ]: 76 : if (subxacts[i - 1].xid == xid)
5412 : 57 : return;
5413 : : }
5414 : :
5415 : : /* This is a new subxact, so we need to add it to the array. */
5416 [ + + ]: 19 : if (subxact_data.nsubxacts == 0)
5417 : : {
5418 : : MemoryContext oldctx;
5419 : :
5420 : 8 : subxact_data.nsubxacts_max = 128;
5421 : :
5422 : : /*
5423 : : * Allocate this memory for subxacts in per-stream context, see
5424 : : * subxact_info_read.
5425 : : */
5426 : 8 : oldctx = MemoryContextSwitchTo(LogicalStreamingContext);
5427 : 8 : subxacts = palloc_array(SubXactInfo, subxact_data.nsubxacts_max);
5428 : 8 : MemoryContextSwitchTo(oldctx);
5429 : : }
5430 [ + + ]: 11 : else if (subxact_data.nsubxacts == subxact_data.nsubxacts_max)
5431 : : {
5432 : 10 : subxact_data.nsubxacts_max *= 2;
5433 : 10 : subxacts = repalloc_array(subxacts, SubXactInfo,
5434 : : subxact_data.nsubxacts_max);
5435 : : }
5436 : :
5437 : 19 : subxacts[subxact_data.nsubxacts].xid = xid;
5438 : :
5439 : : /*
5440 : : * Get the current offset of the stream file and store it as offset of
5441 : : * this subxact.
5442 : : */
5443 : 19 : BufFileTell(stream_fd,
5444 : 19 : &subxacts[subxact_data.nsubxacts].fileno,
5445 : 19 : &subxacts[subxact_data.nsubxacts].offset);
5446 : :
5447 : 19 : subxact_data.nsubxacts++;
5448 : 19 : subxact_data.subxacts = subxacts;
5449 : : }
5450 : :
5451 : : /* format filename for file containing the info about subxacts */
5452 : : static inline void
5453 : 747 : subxact_filename(char *path, Oid subid, TransactionId xid)
5454 : : {
5455 : 747 : snprintf(path, MAXPGPATH, "%u-%u.subxacts", subid, xid);
5456 : 747 : }
5457 : :
5458 : : /* format filename for file containing serialized changes */
5459 : : static inline void
5460 : 438 : changes_filename(char *path, Oid subid, TransactionId xid)
5461 : : {
5462 : 438 : snprintf(path, MAXPGPATH, "%u-%u.changes", subid, xid);
5463 : 438 : }
5464 : :
5465 : : /*
5466 : : * stream_cleanup_files
5467 : : * Cleanup files for a subscription / toplevel transaction.
5468 : : *
5469 : : * Remove files with serialized changes and subxact info for a particular
5470 : : * toplevel transaction. Each subscription has a separate set of files
5471 : : * for any toplevel transaction.
5472 : : */
5473 : : void
5474 : 31 : stream_cleanup_files(Oid subid, TransactionId xid)
5475 : : {
5476 : : char path[MAXPGPATH];
5477 : :
5478 : : /* Delete the changes file. */
5479 : 31 : changes_filename(path, subid, xid);
5480 : 31 : BufFileDeleteFileSet(MyLogicalRepWorker->stream_fileset, path, false);
5481 : :
5482 : : /* Delete the subxact file, if it exists. */
5483 : 31 : subxact_filename(path, subid, xid);
5484 : 31 : BufFileDeleteFileSet(MyLogicalRepWorker->stream_fileset, path, true);
5485 : 31 : }
5486 : :
5487 : : /*
5488 : : * stream_open_file
5489 : : * Open a file that we'll use to serialize changes for a toplevel
5490 : : * transaction.
5491 : : *
5492 : : * Open a file for streamed changes from a toplevel transaction identified
5493 : : * by stream_xid (global variable). If it's the first chunk of streamed
5494 : : * changes for this transaction, create the buffile, otherwise open the
5495 : : * previously created file.
5496 : : */
5497 : : static void
5498 : 363 : stream_open_file(Oid subid, TransactionId xid, bool first_segment)
5499 : : {
5500 : : char path[MAXPGPATH];
5501 : : MemoryContext oldcxt;
5502 : :
5503 : : Assert(OidIsValid(subid));
5504 : : Assert(TransactionIdIsValid(xid));
5505 : : Assert(stream_fd == NULL);
5506 : :
5507 : :
5508 : 363 : changes_filename(path, subid, xid);
5509 [ - + ]: 363 : elog(DEBUG1, "opening file \"%s\" for streamed changes", path);
5510 : :
5511 : : /*
5512 : : * Create/open the buffiles under the logical streaming context so that we
5513 : : * have those files until stream stop.
5514 : : */
5515 : 363 : oldcxt = MemoryContextSwitchTo(LogicalStreamingContext);
5516 : :
5517 : : /*
5518 : : * If this is the first streamed segment, create the changes file.
5519 : : * Otherwise, just open the file for writing, in append mode.
5520 : : */
5521 [ + + ]: 363 : if (first_segment)
5522 : 32 : stream_fd = BufFileCreateFileSet(MyLogicalRepWorker->stream_fileset,
5523 : : path);
5524 : : else
5525 : : {
5526 : : /*
5527 : : * Open the file and seek to the end of the file because we always
5528 : : * append the changes file.
5529 : : */
5530 : 331 : stream_fd = BufFileOpenFileSet(MyLogicalRepWorker->stream_fileset,
5531 : : path, O_RDWR, false);
5532 : 331 : BufFileSeek(stream_fd, 0, 0, SEEK_END);
5533 : : }
5534 : :
5535 : 363 : MemoryContextSwitchTo(oldcxt);
5536 : 363 : }
5537 : :
5538 : : /*
5539 : : * stream_close_file
5540 : : * Close the currently open file with streamed changes.
5541 : : */
5542 : : static void
5543 : 393 : stream_close_file(void)
5544 : : {
5545 : : Assert(stream_fd != NULL);
5546 : :
5547 : 393 : BufFileClose(stream_fd);
5548 : :
5549 : 393 : stream_fd = NULL;
5550 : 393 : }
5551 : :
5552 : : /*
5553 : : * stream_write_change
5554 : : * Serialize a change to a file for the current toplevel transaction.
5555 : : *
5556 : : * The change is serialized in a simple format, with length (not including
5557 : : * the length), action code (identifying the message type) and message
5558 : : * contents (without the subxact TransactionId value).
5559 : : */
5560 : : static void
5561 : 107554 : stream_write_change(char action, StringInfo s)
5562 : : {
5563 : : int len;
5564 : :
5565 : : Assert(stream_fd != NULL);
5566 : :
5567 : : /* total on-disk size, including the action type character */
5568 : 107554 : len = (s->len - s->cursor) + sizeof(char);
5569 : :
5570 : : /* first write the size */
5571 : 107554 : BufFileWrite(stream_fd, &len, sizeof(len));
5572 : :
5573 : : /* then the action */
5574 : 107554 : BufFileWrite(stream_fd, &action, sizeof(action));
5575 : :
5576 : : /* and finally the remaining part of the buffer (after the XID) */
5577 : 107554 : len = (s->len - s->cursor);
5578 : :
5579 : 107554 : BufFileWrite(stream_fd, &s->data[s->cursor], len);
5580 : 107554 : }
5581 : :
5582 : : /*
5583 : : * stream_open_and_write_change
5584 : : * Serialize a message to a file for the given transaction.
5585 : : *
5586 : : * This function is similar to stream_write_change except that it will open the
5587 : : * target file if not already before writing the message and close the file at
5588 : : * the end.
5589 : : */
5590 : : static void
5591 : 5 : stream_open_and_write_change(TransactionId xid, char action, StringInfo s)
5592 : : {
5593 : : Assert(!in_streamed_transaction);
5594 : :
5595 [ + - ]: 5 : if (!stream_fd)
5596 : 5 : stream_start_internal(xid, false);
5597 : :
5598 : 5 : stream_write_change(action, s);
5599 : 5 : stream_stop_internal(xid);
5600 : 5 : }
5601 : :
5602 : : /*
5603 : : * Sets streaming options including replication slot name and origin start
5604 : : * position. Workers need these options for logical replication.
5605 : : */
5606 : : void
5607 : 470 : set_stream_options(WalRcvStreamOptions *options,
5608 : : char *slotname,
5609 : : XLogRecPtr *origin_startpos)
5610 : : {
5611 : : int server_version;
5612 : :
5613 : 470 : options->logical = true;
5614 : 470 : options->startpoint = *origin_startpos;
5615 : 470 : options->slotname = slotname;
5616 : :
5617 : 470 : server_version = walrcv_server_version(LogRepWorkerWalRcvConn);
5618 : 470 : options->proto.logical.proto_version =
5619 [ - + - - : 470 : server_version >= 160000 ? LOGICALREP_PROTO_STREAM_PARALLEL_VERSION_NUM :
- - ]
5620 : : server_version >= 150000 ? LOGICALREP_PROTO_TWOPHASE_VERSION_NUM :
5621 : : server_version >= 140000 ? LOGICALREP_PROTO_STREAM_VERSION_NUM :
5622 : : LOGICALREP_PROTO_VERSION_NUM;
5623 : :
5624 : 470 : options->proto.logical.publication_names = MySubscription->publications;
5625 : 470 : options->proto.logical.binary = MySubscription->binary;
5626 : :
5627 : : /*
5628 : : * Assign the appropriate option value for streaming option according to
5629 : : * the 'streaming' mode and the publisher's ability to support that mode.
5630 : : */
5631 [ + - ]: 470 : if (server_version >= 160000 &&
5632 [ + + ]: 470 : MySubscription->stream == LOGICALREP_STREAM_PARALLEL)
5633 : : {
5634 : 436 : options->proto.logical.streaming_str = "parallel";
5635 : 436 : MyLogicalRepWorker->parallel_apply = true;
5636 : : }
5637 [ + - ]: 34 : else if (server_version >= 140000 &&
5638 [ + + ]: 34 : MySubscription->stream != LOGICALREP_STREAM_OFF)
5639 : : {
5640 : 26 : options->proto.logical.streaming_str = "on";
5641 : 26 : MyLogicalRepWorker->parallel_apply = false;
5642 : : }
5643 : : else
5644 : : {
5645 : 8 : options->proto.logical.streaming_str = NULL;
5646 : 8 : MyLogicalRepWorker->parallel_apply = false;
5647 : : }
5648 : :
5649 : 470 : options->proto.logical.twophase = false;
5650 : 470 : options->proto.logical.origin = pstrdup(MySubscription->origin);
5651 : 470 : }
5652 : :
5653 : : /*
5654 : : * Cleanup the memory for subxacts and reset the related variables.
5655 : : */
5656 : : static inline void
5657 : 376 : cleanup_subxact_info(void)
5658 : : {
5659 [ + + ]: 376 : if (subxact_data.subxacts)
5660 : 87 : pfree(subxact_data.subxacts);
5661 : :
5662 : 376 : subxact_data.subxacts = NULL;
5663 : 376 : subxact_data.subxact_last = InvalidTransactionId;
5664 : 376 : subxact_data.nsubxacts = 0;
5665 : 376 : subxact_data.nsubxacts_max = 0;
5666 : 376 : }
5667 : :
5668 : : /*
5669 : : * Common function to run the apply loop with error handling. Disable the
5670 : : * subscription, if necessary.
5671 : : *
5672 : : * Note that we don't handle FATAL errors which are probably because
5673 : : * of system resource error and are not repeatable.
5674 : : */
5675 : : void
5676 : 470 : start_apply(XLogRecPtr origin_startpos)
5677 : : {
5678 [ + + ]: 470 : PG_TRY();
5679 : : {
5680 : 470 : LogicalRepApplyLoop(origin_startpos);
5681 : : }
5682 : 113 : PG_CATCH();
5683 : : {
5684 : : /*
5685 : : * Reset the origin state to prevent the advancement of origin
5686 : : * progress if we fail to apply. Otherwise, this will result in
5687 : : * transaction loss as that transaction won't be sent again by the
5688 : : * server.
5689 : : */
5690 : 113 : replorigin_xact_clear(true);
5691 : :
5692 [ + + ]: 113 : if (MySubscription->disableonerr)
5693 : 3 : DisableSubscriptionAndExit();
5694 : : else
5695 : : {
5696 : : /*
5697 : : * Report the worker failed while applying changes. Abort the
5698 : : * current transaction so that the stats message is sent in an
5699 : : * idle state.
5700 : : */
5701 : 110 : AbortOutOfAnyTransaction();
5702 : 110 : pgstat_report_subscription_error(MySubscription->oid);
5703 : :
5704 : 110 : PG_RE_THROW();
5705 : : }
5706 : : }
5707 [ # # ]: 0 : PG_END_TRY();
5708 : 0 : }
5709 : :
5710 : : /*
5711 : : * Runs the leader apply worker.
5712 : : *
5713 : : * It sets up replication origin, streaming options and then starts streaming.
5714 : : */
5715 : : static void
5716 : 329 : run_apply_worker(void)
5717 : : {
5718 : : char originname[NAMEDATALEN];
5719 : 329 : XLogRecPtr origin_startpos = InvalidXLogRecPtr;
5720 : 329 : char *slotname = NULL;
5721 : : WalRcvStreamOptions options;
5722 : : ReplOriginId originid;
5723 : : TimeLineID startpointTLI;
5724 : : char *err;
5725 : : bool must_use_password;
5726 : :
5727 : 329 : slotname = MySubscription->slotname;
5728 : :
5729 : : /*
5730 : : * This shouldn't happen if the subscription is enabled, but guard against
5731 : : * DDL bugs or manual catalog changes. (libpqwalreceiver will crash if
5732 : : * slot is NULL.)
5733 : : */
5734 [ - + ]: 329 : if (!slotname)
5735 [ # # ]: 0 : ereport(ERROR,
5736 : : (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
5737 : : errmsg("subscription has no replication slot set")));
5738 : :
5739 : : /* Setup replication origin tracking. */
5740 : 329 : ReplicationOriginNameForLogicalRep(MySubscription->oid, InvalidOid,
5741 : : originname, sizeof(originname));
5742 : 329 : StartTransactionCommand();
5743 : 329 : originid = replorigin_by_name(originname, true);
5744 [ - + ]: 329 : if (!OidIsValid(originid))
5745 : 0 : originid = replorigin_create(originname);
5746 : 329 : replorigin_session_setup(originid, 0);
5747 : 329 : replorigin_xact_state.origin = originid;
5748 : 329 : origin_startpos = replorigin_session_get_progress(false);
5749 : 329 : CommitTransactionCommand();
5750 : :
5751 : : /* Is the use of a password mandatory? */
5752 [ + + ]: 629 : must_use_password = MySubscription->passwordrequired &&
5753 [ + + ]: 300 : !MySubscription->ownersuperuser;
5754 : :
5755 : 329 : LogRepWorkerWalRcvConn = walrcv_connect(MySubscriptionConninfo, true,
5756 : : true, must_use_password,
5757 : : MySubscription->name, &err);
5758 : :
5759 [ + + ]: 317 : if (LogRepWorkerWalRcvConn == NULL)
5760 [ + - ]: 46 : ereport(ERROR,
5761 : : (errcode(ERRCODE_CONNECTION_FAILURE),
5762 : : errmsg("apply worker for subscription \"%s\" could not connect to the publisher: %s",
5763 : : MySubscription->name, err)));
5764 : :
5765 : : /*
5766 : : * We don't really use the output identify_system for anything but it does
5767 : : * some initializations on the upstream so let's still call it.
5768 : : */
5769 : 271 : (void) walrcv_identify_system(LogRepWorkerWalRcvConn, &startpointTLI, NULL);
5770 : :
5771 : : /*
5772 : : * If retain_dead_tuples is enabled, verify that the publisher is
5773 : : * suitable, that is, it runs a version that supports the feature and is
5774 : : * not in recovery. This is the authoritative check. Although the same
5775 : : * validation is performed opportunistically at DDL time, the publisher's
5776 : : * version or recovery status may have changed since then, for example
5777 : : * after a failover.
5778 : : */
5779 [ + + ]: 271 : if (MySubscription->retaindeadtuples)
5780 : : {
5781 : 13 : StartTransactionCommand();
5782 : 13 : CheckPubDeadTupleRetention(LogRepWorkerWalRcvConn);
5783 : 13 : CommitTransactionCommand();
5784 : : }
5785 : :
5786 : 271 : set_apply_error_context_origin(originname);
5787 : :
5788 : 271 : set_stream_options(&options, slotname, &origin_startpos);
5789 : :
5790 : : /*
5791 : : * Even when the two_phase mode is requested by the user, it remains as
5792 : : * the tri-state PENDING until all tablesyncs have reached READY state.
5793 : : * Only then, can it become ENABLED.
5794 : : *
5795 : : * Note: If the subscription has no tables then leave the state as
5796 : : * PENDING, which allows ALTER SUBSCRIPTION ... REFRESH PUBLICATION to
5797 : : * work.
5798 : : */
5799 [ + + + + ]: 285 : if (MySubscription->twophasestate == LOGICALREP_TWOPHASE_STATE_PENDING &&
5800 : 14 : AllTablesyncsReady())
5801 : : {
5802 : : /* Start streaming with two_phase enabled */
5803 : 7 : options.proto.logical.twophase = true;
5804 : 7 : walrcv_startstreaming(LogRepWorkerWalRcvConn, &options);
5805 : :
5806 : 7 : StartTransactionCommand();
5807 : :
5808 : : /*
5809 : : * Updating pg_subscription might involve TOAST table access, so
5810 : : * ensure we have a valid snapshot.
5811 : : */
5812 : 7 : PushActiveSnapshot(GetTransactionSnapshot());
5813 : :
5814 : 7 : UpdateTwoPhaseState(MySubscription->oid, LOGICALREP_TWOPHASE_STATE_ENABLED);
5815 : 7 : MySubscription->twophasestate = LOGICALREP_TWOPHASE_STATE_ENABLED;
5816 : 7 : PopActiveSnapshot();
5817 : 7 : CommitTransactionCommand();
5818 : : }
5819 : : else
5820 : : {
5821 : 264 : walrcv_startstreaming(LogRepWorkerWalRcvConn, &options);
5822 : : }
5823 : :
5824 [ + + + + : 271 : ereport(DEBUG1,
+ - + - ]
5825 : : (errmsg_internal("logical replication apply worker for subscription \"%s\" two_phase is %s",
5826 : : MySubscription->name,
5827 : : MySubscription->twophasestate == LOGICALREP_TWOPHASE_STATE_DISABLED ? "DISABLED" :
5828 : : MySubscription->twophasestate == LOGICALREP_TWOPHASE_STATE_PENDING ? "PENDING" :
5829 : : MySubscription->twophasestate == LOGICALREP_TWOPHASE_STATE_ENABLED ? "ENABLED" :
5830 : : "?")));
5831 : :
5832 : : /* Run the main loop. */
5833 : 271 : start_apply(origin_startpos);
5834 : 0 : }
5835 : :
5836 : : /*
5837 : : * Common initialization for leader apply worker, parallel apply worker,
5838 : : * tablesync worker and sequencesync worker.
5839 : : *
5840 : : * Initialize the database connection, in-memory subscription and necessary
5841 : : * config options.
5842 : : */
5843 : : void
5844 : 650 : InitializeLogRepWorker(void)
5845 : : {
5846 : : /* Run as replica session replication role. */
5847 : 650 : SetConfigOption("session_replication_role", "replica",
5848 : : PGC_SUSET, PGC_S_OVERRIDE);
5849 : :
5850 : : /* Connect to our database. */
5851 : 650 : BackgroundWorkerInitializeConnectionByOid(MyLogicalRepWorker->dbid,
5852 : 650 : MyLogicalRepWorker->userid,
5853 : : 0);
5854 : :
5855 : : /*
5856 : : * Set always-secure search path, so malicious users can't redirect user
5857 : : * code (e.g. pg_index.indexprs).
5858 : : */
5859 : 642 : SetConfigOption("search_path", "", PGC_SUSET, PGC_S_OVERRIDE);
5860 : :
5861 : : /*
5862 : : * Ignore default_transaction_read_only for logical replication workers,
5863 : : * as they need to be able to modify subscriber-side state regardless of
5864 : : * that setting.
5865 : : */
5866 : 642 : SetConfigOption("default_transaction_read_only", "off", PGC_SUSET,
5867 : : PGC_S_OVERRIDE);
5868 : :
5869 : 642 : ApplyContext = AllocSetContextCreate(TopMemoryContext,
5870 : : "ApplyContext",
5871 : : ALLOCSET_DEFAULT_SIZES);
5872 : :
5873 : 642 : StartTransactionCommand();
5874 : :
5875 : : /*
5876 : : * Lock the subscription to prevent it from being concurrently dropped,
5877 : : * then re-verify its existence. After the initialization, the worker will
5878 : : * be terminated gracefully if the subscription is dropped.
5879 : : */
5880 : 642 : LockSharedObject(SubscriptionRelationId, MyLogicalRepWorker->subid, 0,
5881 : : AccessShareLock);
5882 : :
5883 : 641 : MySubscription = GetSubscription(MyLogicalRepWorker->subid, true);
5884 : :
5885 [ + + ]: 641 : if (MySubscription)
5886 : : {
5887 : 572 : MemoryContextSetParent(MySubscription->cxt, ApplyContext);
5888 : : }
5889 : : else
5890 : : {
5891 [ + - ]: 69 : ereport(LOG,
5892 : : (errmsg("logical replication worker for subscription %u will not start because the subscription was removed during startup",
5893 : : MyLogicalRepWorker->subid)));
5894 : :
5895 : : /* Ensure we remove no-longer-useful entry for worker's start time */
5896 [ + - ]: 69 : if (am_leader_apply_worker())
5897 : 69 : ApplyLauncherForgetWorkerStartTime(MyLogicalRepWorker->subid);
5898 : :
5899 : 69 : proc_exit(0);
5900 : : }
5901 : :
5902 [ - + ]: 572 : if (!MySubscription->enabled)
5903 : : {
5904 [ # # ]: 0 : ereport(LOG,
5905 : : (errmsg("logical replication worker for subscription \"%s\" will not start because the subscription was disabled during startup",
5906 : : MySubscription->name)));
5907 : :
5908 : 0 : apply_worker_exit();
5909 : : }
5910 : :
5911 : : /*
5912 : : * May raise error for server-based subscriptions, so build conninfo after
5913 : : * checking that the subscription is enabled. Build in transaction context
5914 : : * and copy to ApplyContext.
5915 : : */
5916 : 572 : MySubscriptionConninfo =
5917 : 572 : MemoryContextStrdup(ApplyContext,
5918 : 572 : SubscriptionConninfo(MySubscription));
5919 : :
5920 : 572 : MySubscriptionValid = true;
5921 : :
5922 : : /*
5923 : : * Restart the worker if retain_dead_tuples was enabled during startup.
5924 : : *
5925 : : * At this point, the replication slot used for conflict detection might
5926 : : * not exist yet, or could be dropped soon if the launcher perceives
5927 : : * retain_dead_tuples as disabled. To avoid unnecessary tracking of
5928 : : * oldest_nonremovable_xid when the slot is absent or at risk of being
5929 : : * dropped, a restart is initiated.
5930 : : *
5931 : : * The oldest_nonremovable_xid should be initialized only when the
5932 : : * subscription's retention is active before launching the worker. See
5933 : : * logicalrep_worker_launch.
5934 : : */
5935 [ + + ]: 572 : if (am_leader_apply_worker() &&
5936 [ + + ]: 329 : MySubscription->retaindeadtuples &&
5937 [ + - ]: 17 : MySubscription->retentionactive &&
5938 [ - + ]: 17 : !TransactionIdIsValid(MyLogicalRepWorker->oldest_nonremovable_xid))
5939 : : {
5940 [ # # ]: 0 : ereport(LOG,
5941 : : errmsg("logical replication worker for subscription \"%s\" will restart because the option %s was enabled during startup",
5942 : : MySubscription->name, "retain_dead_tuples"));
5943 : :
5944 : 0 : apply_worker_exit();
5945 : : }
5946 : :
5947 : : /* Setup synchronous commit according to the user's wishes */
5948 : 572 : SetConfigOption("synchronous_commit", MySubscription->synccommit,
5949 : : PGC_BACKEND, PGC_S_OVERRIDE);
5950 : :
5951 : : /* Change wal_receiver_timeout according to the user's wishes */
5952 : 572 : set_wal_receiver_timeout();
5953 : :
5954 : : /*
5955 : : * Keep us informed about subscription or role changes. Note that the
5956 : : * role's superuser privilege can be revoked.
5957 : : */
5958 : 572 : CacheRegisterSyscacheCallback(SUBSCRIPTIONOID,
5959 : : subscription_change_cb,
5960 : : (Datum) 0);
5961 : : /* Changes to foreign servers may affect subscriptions using SERVER. */
5962 : 572 : CacheRegisterSyscacheCallback(FOREIGNSERVEROID,
5963 : : subscription_change_cb,
5964 : : (Datum) 0);
5965 : : /* Changes to user mappings may affect subscriptions using SERVER. */
5966 : 572 : CacheRegisterSyscacheCallback(USERMAPPINGOID,
5967 : : subscription_change_cb,
5968 : : (Datum) 0);
5969 : :
5970 : : /*
5971 : : * Changes to FDW connection_function may affect subscriptions using
5972 : : * SERVER.
5973 : : */
5974 : 572 : CacheRegisterSyscacheCallback(FOREIGNDATAWRAPPEROID,
5975 : : subscription_change_cb,
5976 : : (Datum) 0);
5977 : :
5978 : 572 : CacheRegisterSyscacheCallback(AUTHOID,
5979 : : subscription_change_cb,
5980 : : (Datum) 0);
5981 : :
5982 [ + + ]: 572 : if (am_tablesync_worker())
5983 [ + - ]: 216 : ereport(LOG,
5984 : : errmsg("logical replication table synchronization worker for subscription \"%s\", table \"%s\" has started",
5985 : : MySubscription->name,
5986 : : get_rel_name(MyLogicalRepWorker->relid)));
5987 [ + + ]: 356 : else if (am_sequencesync_worker())
5988 [ + - ]: 15 : ereport(LOG,
5989 : : errmsg("logical replication sequence synchronization worker for subscription \"%s\" has started",
5990 : : MySubscription->name));
5991 : : else
5992 [ + - ]: 341 : ereport(LOG,
5993 : : errmsg("logical replication apply worker for subscription \"%s\" has started",
5994 : : MySubscription->name));
5995 : :
5996 : 572 : CommitTransactionCommand();
5997 : :
5998 : : /*
5999 : : * Register a callback to reset the origin state before aborting any
6000 : : * pending transaction during shutdown (see ShutdownPostgres()). This will
6001 : : * avoid origin advancement for an incomplete transaction which could
6002 : : * otherwise lead to its loss as such a transaction won't be sent by the
6003 : : * server again.
6004 : : *
6005 : : * Note that even a LOG or DEBUG statement placed after setting the origin
6006 : : * state may process a shutdown signal before committing the current apply
6007 : : * operation. So, it is important to register such a callback here.
6008 : : *
6009 : : * Register this callback here to ensure that all types of logical
6010 : : * replication workers that set up origins and apply remote transactions
6011 : : * are protected.
6012 : : */
6013 : 572 : before_shmem_exit(on_exit_clear_xact_state, (Datum) 0);
6014 : 572 : }
6015 : :
6016 : : /*
6017 : : * Callback on exit to clear transaction-level replication origin state.
6018 : : */
6019 : : static void
6020 : 572 : on_exit_clear_xact_state(int code, Datum arg)
6021 : : {
6022 : 572 : replorigin_xact_clear(true);
6023 : 572 : }
6024 : :
6025 : : /*
6026 : : * Common function to setup the leader apply, tablesync and sequencesync worker.
6027 : : */
6028 : : void
6029 : 638 : SetupApplyOrSyncWorker(int worker_slot)
6030 : : {
6031 : : /* Attach to slot */
6032 : 638 : logicalrep_worker_attach(worker_slot);
6033 : :
6034 : : Assert(am_tablesync_worker() || am_sequencesync_worker() || am_leader_apply_worker());
6035 : :
6036 : : /* Setup signal handling */
6037 : 638 : pqsignal(SIGHUP, SignalHandlerForConfigReload);
6038 : 638 : BackgroundWorkerUnblockSignals();
6039 : :
6040 : : /*
6041 : : * We don't currently need any ResourceOwner in a walreceiver process, but
6042 : : * if we did, we could call CreateAuxProcessResourceOwner here.
6043 : : */
6044 : :
6045 : : /* Initialise stats to a sanish value */
6046 [ + + ]: 638 : if (am_sequencesync_worker())
6047 : : {
6048 : 15 : MyLogicalRepWorker->last_send_time =
6049 : 15 : MyLogicalRepWorker->last_recv_time =
6050 : 15 : MyLogicalRepWorker->reply_time = 0;
6051 : : }
6052 : : else
6053 : : {
6054 : 623 : MyLogicalRepWorker->last_send_time =
6055 : 623 : MyLogicalRepWorker->last_recv_time =
6056 : 623 : MyLogicalRepWorker->reply_time = GetCurrentTimestamp();
6057 : : }
6058 : :
6059 : : /* Load the libpq-specific functions */
6060 : 638 : load_file("libpqwalreceiver", false);
6061 : :
6062 : 638 : InitializeLogRepWorker();
6063 : :
6064 : : /*
6065 : : * Setup callback for syscache so that we know when something changes in
6066 : : * the subscription relation state.
6067 : : */
6068 : 560 : CacheRegisterSyscacheCallback(SUBSCRIPTIONRELMAP,
6069 : : InvalidateSyncingRelStates,
6070 : : (Datum) 0);
6071 : 560 : }
6072 : :
6073 : : /* Logical Replication Apply worker entry point */
6074 : : void
6075 : 406 : ApplyWorkerMain(Datum main_arg)
6076 : : {
6077 : 406 : int worker_slot = DatumGetInt32(main_arg);
6078 : :
6079 : 406 : InitializingApplyWorker = true;
6080 : :
6081 : 406 : SetupApplyOrSyncWorker(worker_slot);
6082 : :
6083 : 329 : InitializingApplyWorker = false;
6084 : :
6085 : 329 : run_apply_worker();
6086 : :
6087 : 0 : proc_exit(0);
6088 : : }
6089 : :
6090 : : /*
6091 : : * After error recovery, disable the subscription in a new transaction
6092 : : * and exit cleanly.
6093 : : */
6094 : : void
6095 : 4 : DisableSubscriptionAndExit(void)
6096 : : {
6097 : : /*
6098 : : * Emit the error message, and recover from the error state to an idle
6099 : : * state
6100 : : */
6101 : 4 : HOLD_INTERRUPTS();
6102 : :
6103 : 4 : EmitErrorReport();
6104 : 4 : AbortOutOfAnyTransaction();
6105 : 4 : FlushErrorState();
6106 : :
6107 : 4 : RESUME_INTERRUPTS();
6108 : :
6109 : : /*
6110 : : * Report the worker failed during sequence synchronization, table
6111 : : * synchronization, or apply.
6112 : : */
6113 : 4 : pgstat_report_subscription_error(MyLogicalRepWorker->subid);
6114 : :
6115 : : /* Disable the subscription */
6116 : 4 : StartTransactionCommand();
6117 : :
6118 : : /*
6119 : : * Updating pg_subscription might involve TOAST table access, so ensure we
6120 : : * have a valid snapshot.
6121 : : */
6122 : 4 : PushActiveSnapshot(GetTransactionSnapshot());
6123 : :
6124 : 4 : DisableSubscription(MySubscription->oid);
6125 : 4 : PopActiveSnapshot();
6126 : 4 : CommitTransactionCommand();
6127 : :
6128 : : /* Ensure we remove no-longer-useful entry for worker's start time */
6129 [ + + ]: 4 : if (am_leader_apply_worker())
6130 : 3 : ApplyLauncherForgetWorkerStartTime(MyLogicalRepWorker->subid);
6131 : :
6132 : : /* Notify the subscription has been disabled and exit */
6133 [ + - ]: 4 : ereport(LOG,
6134 : : errmsg("subscription \"%s\" has been disabled because of an error",
6135 : : MySubscription->name));
6136 : :
6137 : : /*
6138 : : * Skip the track_commit_timestamp check when disabling the worker due to
6139 : : * an error, as verifying commit timestamps is unnecessary in this
6140 : : * context.
6141 : : */
6142 : 4 : CheckSubDeadTupleRetention(false, true, WARNING,
6143 : 4 : MySubscription->retaindeadtuples,
6144 : 4 : MySubscription->retentionactive, false);
6145 : :
6146 : 4 : proc_exit(0);
6147 : : }
6148 : :
6149 : : /*
6150 : : * Is current process a logical replication worker?
6151 : : */
6152 : : bool
6153 : 2801 : IsLogicalWorker(void)
6154 : : {
6155 : 2801 : return MyLogicalRepWorker != NULL;
6156 : : }
6157 : :
6158 : : /*
6159 : : * Is current process a logical replication parallel apply worker?
6160 : : */
6161 : : bool
6162 : 2023 : IsLogicalParallelApplyWorker(void)
6163 : : {
6164 [ + + + - ]: 2023 : return IsLogicalWorker() && am_parallel_apply_worker();
6165 : : }
6166 : :
6167 : : /*
6168 : : * Start skipping changes of the transaction if the given LSN matches the
6169 : : * LSN specified by subscription's skiplsn.
6170 : : */
6171 : : static void
6172 : 589 : maybe_start_skipping_changes(XLogRecPtr finish_lsn)
6173 : : {
6174 : : Assert(!is_skipping_changes());
6175 : : Assert(!in_remote_transaction);
6176 : : Assert(!in_streamed_transaction);
6177 : :
6178 : : /*
6179 : : * Quick return if it's not requested to skip this transaction. This
6180 : : * function is called for every remote transaction and we assume that
6181 : : * skipping the transaction is not used often.
6182 : : */
6183 [ + + - + : 589 : if (likely(!XLogRecPtrIsValid(MySubscription->skiplsn) ||
+ + ]
6184 : : MySubscription->skiplsn != finish_lsn))
6185 : 586 : return;
6186 : :
6187 : : /* Start skipping all changes of this transaction */
6188 : 3 : skip_xact_finish_lsn = finish_lsn;
6189 : :
6190 [ + - ]: 3 : ereport(LOG,
6191 : : errmsg("logical replication starts skipping transaction at LSN %X/%08X",
6192 : : LSN_FORMAT_ARGS(skip_xact_finish_lsn)));
6193 : : }
6194 : :
6195 : : /*
6196 : : * Stop skipping changes by resetting skip_xact_finish_lsn if enabled.
6197 : : */
6198 : : static void
6199 : 30 : stop_skipping_changes(void)
6200 : : {
6201 [ + + ]: 30 : if (!is_skipping_changes())
6202 : 27 : return;
6203 : :
6204 [ + - ]: 3 : ereport(LOG,
6205 : : errmsg("logical replication completed skipping transaction at LSN %X/%08X",
6206 : : LSN_FORMAT_ARGS(skip_xact_finish_lsn)));
6207 : :
6208 : : /* Stop skipping changes */
6209 : 3 : skip_xact_finish_lsn = InvalidXLogRecPtr;
6210 : : }
6211 : :
6212 : : /*
6213 : : * Clear subskiplsn of pg_subscription catalog.
6214 : : *
6215 : : * finish_lsn is the transaction's finish LSN that is used to check if the
6216 : : * subskiplsn matches it. If not matched, we raise a warning when clearing the
6217 : : * subskiplsn in order to inform users for cases e.g., where the user mistakenly
6218 : : * specified the wrong subskiplsn.
6219 : : */
6220 : : static void
6221 : 563 : clear_subscription_skip_lsn(XLogRecPtr finish_lsn)
6222 : : {
6223 : : Relation rel;
6224 : : Form_pg_subscription subform;
6225 : : HeapTuple tup;
6226 : 563 : XLogRecPtr myskiplsn = MySubscription->skiplsn;
6227 : 563 : bool started_tx = false;
6228 : :
6229 [ + + - + ]: 563 : if (likely(!XLogRecPtrIsValid(myskiplsn)) || am_parallel_apply_worker())
6230 : 560 : return;
6231 : :
6232 [ + + ]: 3 : if (!IsTransactionState())
6233 : : {
6234 : 1 : StartTransactionCommand();
6235 : 1 : started_tx = true;
6236 : : }
6237 : :
6238 : : /*
6239 : : * Updating pg_subscription might involve TOAST table access, so ensure we
6240 : : * have a valid snapshot.
6241 : : */
6242 : 3 : PushActiveSnapshot(GetTransactionSnapshot());
6243 : :
6244 : : /*
6245 : : * Protect subskiplsn of pg_subscription from being concurrently updated
6246 : : * while clearing it.
6247 : : */
6248 : 3 : LockSharedObject(SubscriptionRelationId, MySubscription->oid, 0,
6249 : : AccessShareLock);
6250 : :
6251 : 3 : rel = table_open(SubscriptionRelationId, RowExclusiveLock);
6252 : :
6253 : : /* Fetch the existing tuple. */
6254 : 3 : tup = SearchSysCacheCopy1(SUBSCRIPTIONOID,
6255 : : ObjectIdGetDatum(MySubscription->oid));
6256 : :
6257 [ - + ]: 3 : if (!HeapTupleIsValid(tup))
6258 [ # # ]: 0 : elog(ERROR, "subscription \"%s\" does not exist", MySubscription->name);
6259 : :
6260 : 3 : subform = (Form_pg_subscription) GETSTRUCT(tup);
6261 : :
6262 : : /*
6263 : : * Clear the subskiplsn. If the user has already changed subskiplsn before
6264 : : * clearing it we don't update the catalog and the replication origin
6265 : : * state won't get advanced. So in the worst case, if the server crashes
6266 : : * before sending an acknowledgment of the flush position the transaction
6267 : : * will be sent again and the user needs to set subskiplsn again. We can
6268 : : * reduce the possibility by logging a replication origin WAL record to
6269 : : * advance the origin LSN instead but there is no way to advance the
6270 : : * origin timestamp and it doesn't seem to be worth doing anything about
6271 : : * it since it's a very rare case.
6272 : : */
6273 [ + - ]: 3 : if (subform->subskiplsn == myskiplsn)
6274 : : {
6275 : : bool nulls[Natts_pg_subscription];
6276 : : bool replaces[Natts_pg_subscription];
6277 : : Datum values[Natts_pg_subscription];
6278 : :
6279 : 3 : memset(values, 0, sizeof(values));
6280 : 3 : memset(nulls, false, sizeof(nulls));
6281 : 3 : memset(replaces, false, sizeof(replaces));
6282 : :
6283 : : /* reset subskiplsn */
6284 : 3 : values[Anum_pg_subscription_subskiplsn - 1] = LSNGetDatum(InvalidXLogRecPtr);
6285 : 3 : replaces[Anum_pg_subscription_subskiplsn - 1] = true;
6286 : :
6287 : 3 : tup = heap_modify_tuple(tup, RelationGetDescr(rel), values, nulls,
6288 : : replaces);
6289 : 3 : CatalogTupleUpdate(rel, &tup->t_self, tup);
6290 : :
6291 [ - + ]: 3 : if (myskiplsn != finish_lsn)
6292 [ # # ]: 0 : ereport(WARNING,
6293 : : errmsg("skip-LSN of subscription \"%s\" cleared", MySubscription->name),
6294 : : errdetail("Remote transaction's finish WAL location (LSN) %X/%08X did not match skip-LSN %X/%08X.",
6295 : : LSN_FORMAT_ARGS(finish_lsn),
6296 : : LSN_FORMAT_ARGS(myskiplsn)));
6297 : : }
6298 : :
6299 : 3 : heap_freetuple(tup);
6300 : 3 : table_close(rel, NoLock);
6301 : :
6302 : 3 : PopActiveSnapshot();
6303 : :
6304 [ + + ]: 3 : if (started_tx)
6305 : 1 : CommitTransactionCommand();
6306 : : }
6307 : :
6308 : : /* Error callback to give more context info about the change being applied */
6309 : : void
6310 : 10704 : apply_error_callback(void *arg)
6311 : : {
6312 : 10704 : ApplyRemoteCtx *ctx = &remote_ctx;
6313 : :
6314 [ + + ]: 10704 : if (ctx->command == 0)
6315 : 10273 : return;
6316 : :
6317 : : Assert(ctx->origin_name);
6318 : :
6319 [ + + ]: 431 : if (ctx->rel == NULL)
6320 : : {
6321 [ - + ]: 334 : if (!TransactionIdIsValid(ctx->remote_xid))
6322 : 0 : errcontext("processing remote data for replication origin \"%s\" during message type \"%s\"",
6323 : : ctx->origin_name,
6324 : : logicalrep_message_type(ctx->command));
6325 [ + + ]: 334 : else if (!XLogRecPtrIsValid(ctx->finish_lsn))
6326 : 262 : errcontext("processing remote data for replication origin \"%s\" during message type \"%s\" in transaction %u",
6327 : : ctx->origin_name,
6328 : : logicalrep_message_type(ctx->command),
6329 : : ctx->remote_xid);
6330 : : else
6331 : 144 : errcontext("processing remote data for replication origin \"%s\" during message type \"%s\" in transaction %u, finished at %X/%08X",
6332 : : ctx->origin_name,
6333 : : logicalrep_message_type(ctx->command),
6334 : : ctx->remote_xid,
6335 : 72 : LSN_FORMAT_ARGS(ctx->finish_lsn));
6336 : : }
6337 : : else
6338 : : {
6339 [ + - ]: 97 : if (ctx->remote_attnum < 0)
6340 : : {
6341 [ + + ]: 97 : if (!XLogRecPtrIsValid(ctx->finish_lsn))
6342 : 4 : errcontext("processing remote data for replication origin \"%s\" during message type \"%s\" for replication target relation \"%s.%s\" in transaction %u",
6343 : : ctx->origin_name,
6344 : : logicalrep_message_type(ctx->command),
6345 : 2 : ctx->rel->remoterel.nspname,
6346 : 2 : ctx->rel->remoterel.relname,
6347 : : ctx->remote_xid);
6348 : : else
6349 : 190 : errcontext("processing remote data for replication origin \"%s\" during message type \"%s\" for replication target relation \"%s.%s\" in transaction %u, finished at %X/%08X",
6350 : : ctx->origin_name,
6351 : : logicalrep_message_type(ctx->command),
6352 : 95 : ctx->rel->remoterel.nspname,
6353 : 95 : ctx->rel->remoterel.relname,
6354 : : ctx->remote_xid,
6355 : 95 : LSN_FORMAT_ARGS(ctx->finish_lsn));
6356 : : }
6357 : : else
6358 : : {
6359 [ # # ]: 0 : if (!XLogRecPtrIsValid(ctx->finish_lsn))
6360 : 0 : errcontext("processing remote data for replication origin \"%s\" during message type \"%s\" for replication target relation \"%s.%s\" column \"%s\" in transaction %u",
6361 : : ctx->origin_name,
6362 : : logicalrep_message_type(ctx->command),
6363 : 0 : ctx->rel->remoterel.nspname,
6364 : 0 : ctx->rel->remoterel.relname,
6365 : 0 : ctx->rel->remoterel.attnames[ctx->remote_attnum],
6366 : : ctx->remote_xid);
6367 : : else
6368 : 0 : errcontext("processing remote data for replication origin \"%s\" during message type \"%s\" for replication target relation \"%s.%s\" column \"%s\" in transaction %u, finished at %X/%08X",
6369 : : ctx->origin_name,
6370 : : logicalrep_message_type(ctx->command),
6371 : 0 : ctx->rel->remoterel.nspname,
6372 : 0 : ctx->rel->remoterel.relname,
6373 : 0 : ctx->rel->remoterel.attnames[ctx->remote_attnum],
6374 : : ctx->remote_xid,
6375 : 0 : LSN_FORMAT_ARGS(ctx->finish_lsn));
6376 : : }
6377 : : }
6378 : : }
6379 : :
6380 : : /*
6381 : : * Set information identifying the remote transaction currently being
6382 : : * applied, kept for the duration of that transaction.
6383 : : *
6384 : : * This must be called for every message type that begins or resumes applying
6385 : : * a remote transaction's changes (BEGIN, BEGIN PREPARE, STREAM START, STREAM
6386 : : * COMMIT, STREAM PREPARE), since interleaved transactions (possible only for
6387 : : * streaming) would otherwise leave stale values from whichever transaction
6388 : : * last called this.
6389 : : *
6390 : : * Callers normally pass the top-level transaction's xid. The exception is a
6391 : : * STREAM ABORT, which passes the xid of the (sub)transaction being aborted so
6392 : : * that the error context names whatever failed; that is a subxid only when a
6393 : : * subtransaction rolls back, and the top-level xid otherwise. Nothing else
6394 : : * observes a subxid recorded this way, because no change is applied between a
6395 : : * STREAM ABORT and the STREAM START or STREAM COMMIT/PREPARE that follows it,
6396 : : * and each of those calls this again with the transaction's own values.
6397 : : */
6398 : : static inline void
6399 : 3004 : set_remote_transaction_info(TransactionId xid, XLogRecPtr lsn)
6400 : : {
6401 : 3004 : remote_ctx.remote_xid = xid;
6402 : 3004 : remote_ctx.finish_lsn = lsn;
6403 : 3004 : }
6404 : :
6405 : : /* Reset all information of the remote transaction context */
6406 : : static inline void
6407 : 1458 : reset_apply_remote_context(void)
6408 : : {
6409 : 1458 : remote_ctx.command = 0;
6410 : 1458 : remote_ctx.rel = NULL;
6411 : 1458 : remote_ctx.remote_attnum = -1;
6412 : 1458 : set_remote_transaction_info(InvalidTransactionId, InvalidXLogRecPtr);
6413 : 1458 : }
6414 : :
6415 : : /*
6416 : : * Request wakeup of the workers for the given subscription OID
6417 : : * at commit of the current transaction.
6418 : : *
6419 : : * This is used to ensure that the workers process assorted changes
6420 : : * as soon as possible.
6421 : : */
6422 : : void
6423 : 395 : LogicalRepWorkersWakeupAtCommit(Oid subid)
6424 : : {
6425 : : MemoryContext oldcxt;
6426 : :
6427 : 395 : oldcxt = MemoryContextSwitchTo(TopTransactionContext);
6428 : 395 : on_commit_wakeup_workers_subids =
6429 : 395 : list_append_unique_oid(on_commit_wakeup_workers_subids, subid);
6430 : 395 : MemoryContextSwitchTo(oldcxt);
6431 : 395 : }
6432 : :
6433 : : /*
6434 : : * Wake up the workers of any subscriptions that were changed in this xact.
6435 : : */
6436 : : void
6437 : 661335 : AtEOXact_LogicalRepWorkers(bool isCommit)
6438 : : {
6439 [ + + + + ]: 661335 : if (isCommit && on_commit_wakeup_workers_subids != NIL)
6440 : : {
6441 : : ListCell *lc;
6442 : :
6443 : 381 : LWLockAcquire(LogicalRepWorkerLock, LW_SHARED);
6444 [ + - + + : 762 : foreach(lc, on_commit_wakeup_workers_subids)
+ + ]
6445 : : {
6446 : 381 : Oid subid = lfirst_oid(lc);
6447 : : List *workers;
6448 : : ListCell *lc2;
6449 : :
6450 : 381 : workers = logicalrep_workers_find(subid, true, false);
6451 [ + + + + : 460 : foreach(lc2, workers)
+ + ]
6452 : : {
6453 : 79 : LogicalRepWorker *worker = (LogicalRepWorker *) lfirst(lc2);
6454 : :
6455 : 79 : logicalrep_worker_wakeup_ptr(worker);
6456 : : }
6457 : : }
6458 : 381 : LWLockRelease(LogicalRepWorkerLock);
6459 : : }
6460 : :
6461 : : /* The List storage will be reclaimed automatically in xact cleanup. */
6462 : 661335 : on_commit_wakeup_workers_subids = NIL;
6463 : 661335 : }
6464 : :
6465 : : /*
6466 : : * Allocate the origin name in long-lived context for error context message.
6467 : : */
6468 : : void
6469 : 482 : set_apply_error_context_origin(char *originname)
6470 : : {
6471 : 482 : remote_ctx.origin_name = MemoryContextStrdup(ApplyContext, originname);
6472 : 482 : }
6473 : :
6474 : : /*
6475 : : * Return the action to be taken for the given transaction. See
6476 : : * TransApplyAction for information on each of the actions.
6477 : : *
6478 : : * *winfo is assigned to the destination parallel worker info when the leader
6479 : : * apply worker has to pass all the transaction's changes to the parallel
6480 : : * apply worker.
6481 : : */
6482 : : static TransApplyAction
6483 : 350047 : get_transaction_apply_action(TransactionId xid, ParallelApplyWorkerInfo **winfo)
6484 : : {
6485 : 350047 : *winfo = NULL;
6486 : :
6487 [ + + ]: 350047 : if (am_parallel_apply_worker())
6488 : : {
6489 : 68922 : return TRANS_PARALLEL_APPLY;
6490 : : }
6491 : :
6492 : : /*
6493 : : * If we are processing this transaction using a parallel apply worker
6494 : : * then either we send the changes to the parallel worker or if the worker
6495 : : * is busy then serialize the changes to the file which will later be
6496 : : * processed by the parallel worker.
6497 : : */
6498 : 281125 : *winfo = pa_find_worker(xid);
6499 : :
6500 [ + + + + ]: 281125 : if (*winfo && (*winfo)->serialize_changes)
6501 : : {
6502 : 5037 : return TRANS_LEADER_PARTIAL_SERIALIZE;
6503 : : }
6504 [ + + ]: 276088 : else if (*winfo)
6505 : : {
6506 : 68906 : return TRANS_LEADER_SEND_TO_PARALLEL;
6507 : : }
6508 : :
6509 : : /*
6510 : : * If there is no parallel worker involved to process this transaction
6511 : : * then we either directly apply the change or serialize it to a file
6512 : : * which will later be applied when the transaction finish message is
6513 : : * processed.
6514 : : */
6515 [ + + ]: 207182 : else if (in_streamed_transaction)
6516 : : {
6517 : 103199 : return TRANS_LEADER_SERIALIZE;
6518 : : }
6519 : : else
6520 : : {
6521 : 103983 : return TRANS_LEADER_APPLY;
6522 : : }
6523 : : }
|
