1 /*-------------------------------------------------------------------------
2 *
3 * connection.c
4 * Connection management functions for postgres_fdw
5 *
6 * Portions Copyright (c) 2012-2019, PostgreSQL Global Development Group
7 *
8 * IDENTIFICATION
9 * contrib/postgres_fdw/connection.c
10 *
11 *-------------------------------------------------------------------------
12 */
13 #include "postgres.h"
14
15 #include "postgres_fdw.h"
16
17 #include "access/htup_details.h"
18 #include "access/xact.h"
19 #include "catalog/pg_user_mapping.h"
20 #include "mb/pg_wchar.h"
21 #include "miscadmin.h"
22 #include "pgstat.h"
23 #include "storage/latch.h"
24 #include "utils/hsearch.h"
25 #include "utils/inval.h"
26 #include "utils/memutils.h"
27 #include "utils/syscache.h"
28
29
30 /*
31 * Connection cache hash table entry
32 *
33 * The lookup key in this hash table is the user mapping OID. We use just one
get_os_thread_id()34 * connection per user mapping ID, which ensures that all the scans use the
35 * same snapshot during a query. Using the user mapping OID rather than
36 * the foreign server OID + user OID avoids creating multiple connections when
37 * the public user mapping applies to all user OIDs.
38 *
39 * The "conn" pointer can be NULL if we don't currently have a live connection.
40 * When we do have a connection, xact_depth tracks the current depth of
41 * transactions and subtransactions open on the remote side. We need to issue
42 * commands at the same nesting depth on the remote as we're executing at
43 * ourselves, so that rolling back a subtransaction will kill the right
44 * queries and not the wrong ones.
45 */
46 typedef Oid ConnCacheKey;
47
48 typedef struct ConnCacheEntry
49 {
50 ConnCacheKey key; /* hash key (must be first) */
51 PGconn *conn; /* connection to foreign server, or NULL */
52 /* Remaining fields are invalid when conn is NULL: */
53 int xact_depth; /* 0 = no xact open, 1 = main xact open, 2 =
54 * one level of subxact open, etc */
55 bool have_prep_stmt; /* have we prepared any stmts in this xact? */
56 bool have_error; /* have any subxacts aborted in this xact? */
57 bool changing_xact_state; /* xact state change in process */
58 bool invalidated; /* true if reconnect is pending */
59 uint32 server_hashvalue; /* hash value of foreign server OID */
60 uint32 mapping_hashvalue; /* hash value of user mapping OID */
61 } ConnCacheEntry;
62
63 /*
64 * Connection cache (initialized on first use)
65 */
66 static HTAB *ConnectionHash = NULL;
67
68 /* for assigning cursor numbers and prepared statement numbers */
69 static unsigned int cursor_number = 0;
70 static unsigned int prep_stmt_number = 0;
71
72 /* tracks whether any work is needed in callback functions */
73 static bool xact_got_connection = false;
74
75 /* prototypes of private functions */
76 static PGconn *connect_pg_server(ForeignServer *server, UserMapping *user);
77 static void disconnect_pg_server(ConnCacheEntry *entry);
thread_func(void * arg)78 static void check_conn_params(const char **keywords, const char **values, UserMapping *user);
79 static void configure_remote_session(PGconn *conn);
80 static void do_sql_command(PGconn *conn, const char *sql);
81 static void begin_remote_xact(ConnCacheEntry *entry);
82 static void pgfdw_xact_callback(XactEvent event, void *arg);
83 static void pgfdw_subxact_callback(SubXactEvent event,
84 SubTransactionId mySubid,
85 SubTransactionId parentSubid,
86 void *arg);
87 static void pgfdw_inval_callback(Datum arg, int cacheid, uint32 hashvalue);
88 static void pgfdw_reject_incomplete_xact_state_change(ConnCacheEntry *entry);
89 static bool pgfdw_cancel_query(PGconn *conn);
90 static bool pgfdw_exec_cleanup_query(PGconn *conn, const char *query,
91 bool ignore_errors);
92 static bool pgfdw_get_cleanup_result(PGconn *conn, TimestampTz endtime,
93 PGresult **result);
94
95
96 /*
97 * Get a PGconn which can be used to execute queries on the remote PostgreSQL
98 * server with the user's authorization. A new connection is established
99 * if we don't already have a suitable one, and a transaction is opened at
100 * the right subtransaction nesting depth if we didn't do that already.
101 *
102 * will_prep_stmt must be true if caller intends to create any prepared
103 * statements. Since those don't go away automatically at transaction end
104 * (not even on error), we need this flag to cue manual cleanup.
105 */
106 PGconn *
107 GetConnection(UserMapping *user, bool will_prep_stmt)
108 {
109 bool found;
110 ConnCacheEntry *entry;
111 ConnCacheKey key;
112
113 /* First time through, initialize connection cache hashtable */
114 if (ConnectionHash == NULL)
115 {
116 HASHCTL ctl;
117
118 MemSet(&ctl, 0, sizeof(ctl));
119 ctl.keysize = sizeof(ConnCacheKey);
120 ctl.entrysize = sizeof(ConnCacheEntry);
121 /* allocate ConnectionHash in the cache context */
122 ctl.hcxt = CacheMemoryContext;
123 ConnectionHash = hash_create("postgres_fdw connections", 8,
124 &ctl,
125 HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
126
127 /*
128 * Register some callback functions that manage connection cleanup.
129 * This should be done just once in each backend.
130 */
131 RegisterXactCallback(pgfdw_xact_callback, NULL);
132 RegisterSubXactCallback(pgfdw_subxact_callback, NULL);
133 CacheRegisterSyscacheCallback(FOREIGNSERVEROID,
134 pgfdw_inval_callback, (Datum) 0);
135 CacheRegisterSyscacheCallback(USERMAPPINGOID,
136 pgfdw_inval_callback, (Datum) 0);
137 }
138
139 /* Set flag that we did GetConnection during the current transaction */
140 xact_got_connection = true;
141
142 /* Create hash key for the entry. Assume no pad bytes in key struct */
143 key = user->umid;
144
145 /*
146 * Find or create cached entry for requested connection.
147 */
148 entry = hash_search(ConnectionHash, &key, HASH_ENTER, &found);
149 if (!found)
150 {
151 /*
152 * We need only clear "conn" here; remaining fields will be filled
153 * later when "conn" is set.
main(int argc,char ** argv)154 */
155 entry->conn = NULL;
156 }
157
158 /* Reject further use of connections which failed abort cleanup. */
159 pgfdw_reject_incomplete_xact_state_change(entry);
160
161 /*
162 * If the connection needs to be remade due to invalidation, disconnect as
163 * soon as we're out of all transactions.
164 */
165 if (entry->conn != NULL && entry->invalidated && entry->xact_depth == 0)
166 {
167 elog(DEBUG3, "closing connection %p for option changes to take effect",
168 entry->conn);
169 disconnect_pg_server(entry);
170 }
171
172 /*
173 * We don't check the health of cached connection here, because it would
174 * require some overhead. Broken connection will be detected when the
175 * connection is actually used.
176 */
177
178 /*
179 * If cache entry doesn't have a connection, we have to establish a new
180 * connection. (If connect_pg_server throws an error, the cache entry
181 * will remain in a valid empty state, ie conn == NULL.)
182 */
183 if (entry->conn == NULL)
184 {
185 ForeignServer *server = GetForeignServer(user->serverid);
186
187 /* Reset all transient state fields, to be sure all are clean */
188 entry->xact_depth = 0;
189 entry->have_prep_stmt = false;
190 entry->have_error = false;
191 entry->changing_xact_state = false;
192 entry->invalidated = false;
193 entry->server_hashvalue =
194 GetSysCacheHashValue1(FOREIGNSERVEROID,
195 ObjectIdGetDatum(server->serverid));
196 entry->mapping_hashvalue =
197 GetSysCacheHashValue1(USERMAPPINGOID,
198 ObjectIdGetDatum(user->umid));
199
200 /* Now try to make the connection */
201 entry->conn = connect_pg_server(server, user);
202
203 elog(DEBUG3, "new postgres_fdw connection %p for server \"%s\" (user mapping oid %u, userid %u)",
204 entry->conn, server->servername, user->umid, user->userid);
205 }
206
207 /*
208 * Start a new transaction or subtransaction if needed.
209 */
210 begin_remote_xact(entry);
211
212 /* Remember if caller will prepare statements */
213 entry->have_prep_stmt |= will_prep_stmt;
214
215 return entry->conn;
216 }
217
218 /*
219 * Connect to remote server using specified server and user mapping properties.
220 */
221 static PGconn *
222 connect_pg_server(ForeignServer *server, UserMapping *user)
223 {
224 PGconn *volatile conn = NULL;
225
226 /*
227 * Use PG_TRY block to ensure closing connection on error.
228 */
229 PG_TRY();
230 {
231 const char **keywords;
232 const char **values;
233 int n;
234
235 /*
236 * Construct connection params from generic options of ForeignServer
237 * and UserMapping. (Some of them might not be libpq options, in
238 * which case we'll just waste a few array slots.) Add 3 extra slots
239 * for fallback_application_name, client_encoding, end marker.
240 */
241 n = list_length(server->options) + list_length(user->options) + 3;
242 keywords = (const char **) palloc(n * sizeof(char *));
243 values = (const char **) palloc(n * sizeof(char *));
244
245 n = 0;
246 n += ExtractConnectionOptions(server->options,
247 keywords + n, values + n);
248 n += ExtractConnectionOptions(user->options,
249 keywords + n, values + n);
250
251 /* Use "postgres_fdw" as fallback_application_name. */
252 keywords[n] = "fallback_application_name";
253 values[n] = "postgres_fdw";
254 n++;
255
256 /* Set client_encoding so that libpq can convert encoding properly. */
257 keywords[n] = "client_encoding";
258 values[n] = GetDatabaseEncodingName();
259 n++;
260
261 keywords[n] = values[n] = NULL;
262
263 /* verify connection parameters and make connection */
264 check_conn_params(keywords, values, user);
265
266 conn = PQconnectdbParams(keywords, values, false);
267 if (!conn || PQstatus(conn) != CONNECTION_OK)
268 ereport(ERROR,
269 (errcode(ERRCODE_SQLCLIENT_UNABLE_TO_ESTABLISH_SQLCONNECTION),
270 errmsg("could not connect to server \"%s\"",
271 server->servername),
272 errdetail_internal("%s", pchomp(PQerrorMessage(conn)))));
273
274 /*
275 * Check that non-superuser has used password to establish connection;
276 * otherwise, he's piggybacking on the postgres server's user
277 * identity. See also dblink_security_check() in contrib/dblink.
278 */
279 if (!superuser_arg(user->userid) && !PQconnectionUsedPassword(conn))
280 ereport(ERROR,
281 (errcode(ERRCODE_S_R_E_PROHIBITED_SQL_STATEMENT_ATTEMPTED),
282 errmsg("password is required"),
283 errdetail("Non-superuser cannot connect if the server does not request a password."),
284 errhint("Target server's authentication method must be changed.")));
285
286 /* Prepare new session for use */
287 configure_remote_session(conn);
288
289 pfree(keywords);
290 pfree(values);
291 }
292 PG_CATCH();
293 {
294 /* Release PGconn data structure if we managed to create one */
295 if (conn)
296 PQfinish(conn);
297 PG_RE_THROW();
298 }
299 PG_END_TRY();
300
301 return conn;
302 }
303
304 /*
305 * Disconnect any open connection for a connection cache entry.
306 */
307 static void
308 disconnect_pg_server(ConnCacheEntry *entry)
309 {
310 if (entry->conn != NULL)
311 {
312 PQfinish(entry->conn);
313 entry->conn = NULL;
314 }
315 }
316
317 /*
318 * For non-superusers, insist that the connstr specify a password. This
319 * prevents a password from being picked up from .pgpass, a service file,
320 * the environment, etc. We don't want the postgres user's passwords
321 * to be accessible to non-superusers. (See also dblink_connstr_check in
322 * contrib/dblink.)
323 */
324 static void
325 check_conn_params(const char **keywords, const char **values, UserMapping *user)
326 {
327 int i;
328
329 /* no check required if superuser */
330 if (superuser_arg(user->userid))
331 return;
332
333 /* ok if params contain a non-empty password */
334 for (i = 0; keywords[i] != NULL; i++)
335 {
336 if (strcmp(keywords[i], "password") == 0 && values[i][0] != '\0')
337 return;
338 }
339
340 ereport(ERROR,
341 (errcode(ERRCODE_S_R_E_PROHIBITED_SQL_STATEMENT_ATTEMPTED),
342 errmsg("password is required"),
343 errdetail("Non-superusers must provide a password in the user mapping.")));
344 }
345
346 /*
347 * Issue SET commands to make sure remote session is configured properly.
348 *
349 * We do this just once at connection, assuming nothing will change the
350 * values later. Since we'll never send volatile function calls to the
351 * remote, there shouldn't be any way to break this assumption from our end.
352 * It's possible to think of ways to break it at the remote end, eg making
353 * a foreign table point to a view that includes a set_config call ---
354 * but once you admit the possibility of a malicious view definition,
355 * there are any number of ways to break things.
356 */
357 static void
358 configure_remote_session(PGconn *conn)
359 {
360 int remoteversion = PQserverVersion(conn);
361
362 /* Force the search path to contain only pg_catalog (see deparse.c) */
363 do_sql_command(conn, "SET search_path = pg_catalog");
364
365 /*
366 * Set remote timezone; this is basically just cosmetic, since all
367 * transmitted and returned timestamptzs should specify a zone explicitly
368 * anyway. However it makes the regression test outputs more predictable.
369 *
370 * We don't risk setting remote zone equal to ours, since the remote
371 * server might use a different timezone database. Instead, use UTC
372 * (quoted, because very old servers are picky about case).
373 */
374 do_sql_command(conn, "SET timezone = 'UTC'");
375
376 /*
377 * Set values needed to ensure unambiguous data output from remote. (This
378 * logic should match what pg_dump does. See also set_transmission_modes
379 * in postgres_fdw.c.)
380 */
381 do_sql_command(conn, "SET datestyle = ISO");
382 if (remoteversion >= 80400)
383 do_sql_command(conn, "SET intervalstyle = postgres");
384 if (remoteversion >= 90000)
385 do_sql_command(conn, "SET extra_float_digits = 3");
386 else
387 do_sql_command(conn, "SET extra_float_digits = 2");
388 }
389
390 /*
391 * Convenience subroutine to issue a non-data-returning SQL command to remote
392 */
393 static void
394 do_sql_command(PGconn *conn, const char *sql)
395 {
396 PGresult *res;
397
398 if (!PQsendQuery(conn, sql))
399 pgfdw_report_error(ERROR, NULL, conn, false, sql);
400 res = pgfdw_get_result(conn, sql);
401 if (PQresultStatus(res) != PGRES_COMMAND_OK)
402 pgfdw_report_error(ERROR, res, conn, true, sql);
403 PQclear(res);
404 }
405
406 /*
407 * Start remote transaction or subtransaction, if needed.
408 *
409 * Note that we always use at least REPEATABLE READ in the remote session.
410 * This is so that, if a query initiates multiple scans of the same or
411 * different foreign tables, we will get snapshot-consistent results from
412 * those scans. A disadvantage is that we can't provide sane emulation of
413 * READ COMMITTED behavior --- it would be nice if we had some other way to
414 * control which remote queries share a snapshot.
415 */
416 static void
417 begin_remote_xact(ConnCacheEntry *entry)
418 {
419 int curlevel = GetCurrentTransactionNestLevel();
420
421 /* Start main transaction if we haven't yet */
422 if (entry->xact_depth <= 0)
423 {
424 const char *sql;
425
426 elog(DEBUG3, "starting remote transaction on connection %p",
427 entry->conn);
428
429 if (IsolationIsSerializable())
430 sql = "START TRANSACTION ISOLATION LEVEL SERIALIZABLE";
431 else
432 sql = "START TRANSACTION ISOLATION LEVEL REPEATABLE READ";
433 entry->changing_xact_state = true;
434 do_sql_command(entry->conn, sql);
435 entry->xact_depth = 1;
436 entry->changing_xact_state = false;
437 }
438
439 /*
440 * If we're in a subtransaction, stack up savepoints to match our level.
441 * This ensures we can rollback just the desired effects when a
442 * subtransaction aborts.
443 */
444 while (entry->xact_depth < curlevel)
445 {
446 char sql[64];
447
448 snprintf(sql, sizeof(sql), "SAVEPOINT s%d", entry->xact_depth + 1);
449 entry->changing_xact_state = true;
450 do_sql_command(entry->conn, sql);
451 entry->xact_depth++;
452 entry->changing_xact_state = false;
453 }
454 }
455
456 /*
457 * Release connection reference count created by calling GetConnection.
458 */
459 void
460 ReleaseConnection(PGconn *conn)
461 {
462 /*
463 * Currently, we don't actually track connection references because all
464 * cleanup is managed on a transaction or subtransaction basis instead. So
465 * there's nothing to do here.
466 */
467 }
468
469 /*
470 * Assign a "unique" number for a cursor.
471 *
472 * These really only need to be unique per connection within a transaction.
473 * For the moment we ignore the per-connection point and assign them across
474 * all connections in the transaction, but we ask for the connection to be
475 * supplied in case we want to refine that.
476 *
477 * Note that even if wraparound happens in a very long transaction, actual
478 * collisions are highly improbable; just be sure to use %u not %d to print.
479 */
480 unsigned int
481 GetCursorNumber(PGconn *conn)
482 {
483 return ++cursor_number;
484 }
485
486 /*
487 * Assign a "unique" number for a prepared statement.
488 *
489 * This works much like GetCursorNumber, except that we never reset the counter
490 * within a session. That's because we can't be 100% sure we've gotten rid
491 * of all prepared statements on all connections, and it's not really worth
492 * increasing the risk of prepared-statement name collisions by resetting.
493 */
494 unsigned int
495 GetPrepStmtNumber(PGconn *conn)
496 {
497 return ++prep_stmt_number;
498 }
499
500 /*
501 * Submit a query and wait for the result.
502 *
503 * This function is interruptible by signals.
504 *
505 * Caller is responsible for the error handling on the result.
506 */
507 PGresult *
508 pgfdw_exec_query(PGconn *conn, const char *query)
509 {
510 /*
511 * Submit a query. Since we don't use non-blocking mode, this also can
512 * block. But its risk is relatively small, so we ignore that for now.
513 */
514 if (!PQsendQuery(conn, query))
515 pgfdw_report_error(ERROR, NULL, conn, false, query);
516
517 /* Wait for the result. */
518 return pgfdw_get_result(conn, query);
519 }
520
521 /*
522 * Wait for the result from a prior asynchronous execution function call.
523 *
524 * This function offers quick responsiveness by checking for any interruptions.
525 *
526 * This function emulates PQexec()'s behavior of returning the last result
527 * when there are many.
528 *
529 * Caller is responsible for the error handling on the result.
530 */
531 PGresult *
532 pgfdw_get_result(PGconn *conn, const char *query)
533 {
534 PGresult *volatile last_res = NULL;
535
536 /* In what follows, do not leak any PGresults on an error. */
537 PG_TRY();
538 {
539 for (;;)
540 {
541 PGresult *res;
542
543 while (PQisBusy(conn))
544 {
545 int wc;
546
547 /* Sleep until there's something to do */
548 wc = WaitLatchOrSocket(MyLatch,
549 WL_LATCH_SET | WL_SOCKET_READABLE |
550 WL_EXIT_ON_PM_DEATH,
551 PQsocket(conn),
552 -1L, PG_WAIT_EXTENSION);
553 ResetLatch(MyLatch);
554
555 CHECK_FOR_INTERRUPTS();
556
557 /* Data available in socket? */
558 if (wc & WL_SOCKET_READABLE)
559 {
560 if (!PQconsumeInput(conn))
561 pgfdw_report_error(ERROR, NULL, conn, false, query);
562 }
563 }
564
565 res = PQgetResult(conn);
566 if (res == NULL)
567 break; /* query is complete */
568
569 PQclear(last_res);
570 last_res = res;
571 }
572 }
573 PG_CATCH();
574 {
575 PQclear(last_res);
576 PG_RE_THROW();
577 }
578 PG_END_TRY();
579
580 return last_res;
581 }
582
583 /*
584 * Report an error we got from the remote server.
585 *
586 * elevel: error level to use (typically ERROR, but might be less)
587 * res: PGresult containing the error
588 * conn: connection we did the query on
589 * clear: if true, PQclear the result (otherwise caller will handle it)
590 * sql: NULL, or text of remote command we tried to execute
591 *
592 * Note: callers that choose not to throw ERROR for a remote error are
593 * responsible for making sure that the associated ConnCacheEntry gets
594 * marked with have_error = true.
595 */
596 void
597 pgfdw_report_error(int elevel, PGresult *res, PGconn *conn,
598 bool clear, const char *sql)
599 {
600 /* If requested, PGresult must be released before leaving this function. */
601 PG_TRY();
602 {
603 char *diag_sqlstate = PQresultErrorField(res, PG_DIAG_SQLSTATE);
604 char *message_primary = PQresultErrorField(res, PG_DIAG_MESSAGE_PRIMARY);
605 char *message_detail = PQresultErrorField(res, PG_DIAG_MESSAGE_DETAIL);
606 char *message_hint = PQresultErrorField(res, PG_DIAG_MESSAGE_HINT);
607 char *message_context = PQresultErrorField(res, PG_DIAG_CONTEXT);
608 int sqlstate;
609
610 if (diag_sqlstate)
611 sqlstate = MAKE_SQLSTATE(diag_sqlstate[0],
612 diag_sqlstate[1],
613 diag_sqlstate[2],
614 diag_sqlstate[3],
615 diag_sqlstate[4]);
616 else
617 sqlstate = ERRCODE_CONNECTION_FAILURE;
618
619 /*
620 * If we don't get a message from the PGresult, try the PGconn. This
621 * is needed because for connection-level failures, PQexec may just
622 * return NULL, not a PGresult at all.
623 */
624 if (message_primary == NULL)
625 message_primary = pchomp(PQerrorMessage(conn));
626
627 ereport(elevel,
628 (errcode(sqlstate),
629 message_primary ? errmsg_internal("%s", message_primary) :
630 errmsg("could not obtain message string for remote error"),
631 message_detail ? errdetail_internal("%s", message_detail) : 0,
632 message_hint ? errhint("%s", message_hint) : 0,
633 message_context ? errcontext("%s", message_context) : 0,
634 sql ? errcontext("remote SQL command: %s", sql) : 0));
635 }
636 PG_CATCH();
637 {
638 if (clear)
639 PQclear(res);
640 PG_RE_THROW();
641 }
642 PG_END_TRY();
643 if (clear)
644 PQclear(res);
645 }
646
647 /*
648 * pgfdw_xact_callback --- cleanup at main-transaction end.
649 *
650 * This runs just late enough that it must not enter user-defined code
651 * locally. (Entering such code on the remote side is fine. Its remote
652 * COMMIT TRANSACTION may run deferred triggers.)
653 */
654 static void
655 pgfdw_xact_callback(XactEvent event, void *arg)
656 {
657 HASH_SEQ_STATUS scan;
658 ConnCacheEntry *entry;
659
660 /* Quick exit if no connections were touched in this transaction. */
661 if (!xact_got_connection)
662 return;
663
664 /*
665 * Scan all connection cache entries to find open remote transactions, and
666 * close them.
667 */
668 hash_seq_init(&scan, ConnectionHash);
669 while ((entry = (ConnCacheEntry *) hash_seq_search(&scan)))
670 {
671 PGresult *res;
672
673 /* Ignore cache entry if no open connection right now */
674 if (entry->conn == NULL)
675 continue;
676
677 /* If it has an open remote transaction, try to close it */
678 if (entry->xact_depth > 0)
679 {
680 bool abort_cleanup_failure = false;
681
682 elog(DEBUG3, "closing remote transaction on connection %p",
683 entry->conn);
684
685 switch (event)
686 {
687 case XACT_EVENT_PARALLEL_PRE_COMMIT:
688 case XACT_EVENT_PRE_COMMIT:
689
690 /*
691 * If abort cleanup previously failed for this connection,
692 * we can't issue any more commands against it.
693 */
694 pgfdw_reject_incomplete_xact_state_change(entry);
695
696 /* Commit all remote transactions during pre-commit */
697 entry->changing_xact_state = true;
698 do_sql_command(entry->conn, "COMMIT TRANSACTION");
699 entry->changing_xact_state = false;
700
701 /*
702 * If there were any errors in subtransactions, and we
703 * made prepared statements, do a DEALLOCATE ALL to make
704 * sure we get rid of all prepared statements. This is
705 * annoying and not terribly bulletproof, but it's
706 * probably not worth trying harder.
707 *
708 * DEALLOCATE ALL only exists in 8.3 and later, so this
709 * constrains how old a server postgres_fdw can
710 * communicate with. We intentionally ignore errors in
711 * the DEALLOCATE, so that we can hobble along to some
712 * extent with older servers (leaking prepared statements
713 * as we go; but we don't really support update operations
714 * pre-8.3 anyway).
715 */
716 if (entry->have_prep_stmt && entry->have_error)
717 {
718 res = PQexec(entry->conn, "DEALLOCATE ALL");
719 PQclear(res);
720 }
721 entry->have_prep_stmt = false;
722 entry->have_error = false;
723 break;
724 case XACT_EVENT_PRE_PREPARE:
725
726 /*
727 * We disallow any remote transactions, since it's not
728 * very reasonable to hold them open until the prepared
729 * transaction is committed. For the moment, throw error
730 * unconditionally; later we might allow read-only cases.
731 * Note that the error will cause us to come right back
732 * here with event == XACT_EVENT_ABORT, so we'll clean up
733 * the connection state at that point.
734 */
735 ereport(ERROR,
736 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
737 errmsg("cannot PREPARE a transaction that has operated on postgres_fdw foreign tables")));
738 break;
739 case XACT_EVENT_PARALLEL_COMMIT:
740 case XACT_EVENT_COMMIT:
741 case XACT_EVENT_PREPARE:
742 /* Pre-commit should have closed the open transaction */
743 elog(ERROR, "missed cleaning up connection during pre-commit");
744 break;
745 case XACT_EVENT_PARALLEL_ABORT:
746 case XACT_EVENT_ABORT:
747
748 /*
749 * Don't try to clean up the connection if we're already
750 * in error recursion trouble.
751 */
752 if (in_error_recursion_trouble())
753 entry->changing_xact_state = true;
754
755 /*
756 * If connection is already unsalvageable, don't touch it
757 * further.
758 */
759 if (entry->changing_xact_state)
760 break;
761
762 /*
763 * Mark this connection as in the process of changing
764 * transaction state.
765 */
766 entry->changing_xact_state = true;
767
768 /* Assume we might have lost track of prepared statements */
769 entry->have_error = true;
770
771 /*
772 * If a command has been submitted to the remote server by
773 * using an asynchronous execution function, the command
774 * might not have yet completed. Check to see if a
775 * command is still being processed by the remote server,
776 * and if so, request cancellation of the command.
777 */
778 if (PQtransactionStatus(entry->conn) == PQTRANS_ACTIVE &&
779 !pgfdw_cancel_query(entry->conn))
780 {
781 /* Unable to cancel running query. */
782 abort_cleanup_failure = true;
783 }
784 else if (!pgfdw_exec_cleanup_query(entry->conn,
785 "ABORT TRANSACTION",
786 false))
787 {
788 /* Unable to abort remote transaction. */
789 abort_cleanup_failure = true;
790 }
791 else if (entry->have_prep_stmt && entry->have_error &&
792 !pgfdw_exec_cleanup_query(entry->conn,
793 "DEALLOCATE ALL",
794 true))
795 {
796 /* Trouble clearing prepared statements. */
797 abort_cleanup_failure = true;
798 }
799 else
800 {
801 entry->have_prep_stmt = false;
802 entry->have_error = false;
803 }
804
805 /* Disarm changing_xact_state if it all worked. */
806 entry->changing_xact_state = abort_cleanup_failure;
807 break;
808 }
809 }
810
811 /* Reset state to show we're out of a transaction */
812 entry->xact_depth = 0;
813
814 /*
815 * If the connection isn't in a good idle state or it is marked as
816 * invalid, then discard it to recover. Next GetConnection will open a
817 * new connection.
818 */
819 if (PQstatus(entry->conn) != CONNECTION_OK ||
820 PQtransactionStatus(entry->conn) != PQTRANS_IDLE ||
821 entry->changing_xact_state ||
822 entry->invalidated)
823 {
824 elog(DEBUG3, "discarding connection %p", entry->conn);
825 disconnect_pg_server(entry);
826 }
827 }
828
829 /*
830 * Regardless of the event type, we can now mark ourselves as out of the
831 * transaction. (Note: if we are here during PRE_COMMIT or PRE_PREPARE,
832 * this saves a useless scan of the hashtable during COMMIT or PREPARE.)
833 */
834 xact_got_connection = false;
835
836 /* Also reset cursor numbering for next transaction */
837 cursor_number = 0;
838 }
839
840 /*
841 * pgfdw_subxact_callback --- cleanup at subtransaction end.
842 */
843 static void
844 pgfdw_subxact_callback(SubXactEvent event, SubTransactionId mySubid,
845 SubTransactionId parentSubid, void *arg)
846 {
847 HASH_SEQ_STATUS scan;
848 ConnCacheEntry *entry;
849 int curlevel;
850
851 /* Nothing to do at subxact start, nor after commit. */
852 if (!(event == SUBXACT_EVENT_PRE_COMMIT_SUB ||
853 event == SUBXACT_EVENT_ABORT_SUB))
854 return;
855
856 /* Quick exit if no connections were touched in this transaction. */
857 if (!xact_got_connection)
858 return;
859
860 /*
861 * Scan all connection cache entries to find open remote subtransactions
862 * of the current level, and close them.
863 */
864 curlevel = GetCurrentTransactionNestLevel();
865 hash_seq_init(&scan, ConnectionHash);
866 while ((entry = (ConnCacheEntry *) hash_seq_search(&scan)))
867 {
868 char sql[100];
869
870 /*
871 * We only care about connections with open remote subtransactions of
872 * the current level.
873 */
874 if (entry->conn == NULL || entry->xact_depth < curlevel)
875 continue;
876
877 if (entry->xact_depth > curlevel)
878 elog(ERROR, "missed cleaning up remote subtransaction at level %d",
879 entry->xact_depth);
880
881 if (event == SUBXACT_EVENT_PRE_COMMIT_SUB)
882 {
883 /*
884 * If abort cleanup previously failed for this connection, we
885 * can't issue any more commands against it.
886 */
887 pgfdw_reject_incomplete_xact_state_change(entry);
888
889 /* Commit all remote subtransactions during pre-commit */
890 snprintf(sql, sizeof(sql), "RELEASE SAVEPOINT s%d", curlevel);
891 entry->changing_xact_state = true;
892 do_sql_command(entry->conn, sql);
893 entry->changing_xact_state = false;
894 }
895 else if (in_error_recursion_trouble())
896 {
897 /*
898 * Don't try to clean up the connection if we're already in error
899 * recursion trouble.
900 */
901 entry->changing_xact_state = true;
902 }
903 else if (!entry->changing_xact_state)
904 {
905 bool abort_cleanup_failure = false;
906
907 /* Remember that abort cleanup is in progress. */
908 entry->changing_xact_state = true;
909
910 /* Assume we might have lost track of prepared statements */
911 entry->have_error = true;
912
913 /*
914 * If a command has been submitted to the remote server by using
915 * an asynchronous execution function, the command might not have
916 * yet completed. Check to see if a command is still being
917 * processed by the remote server, and if so, request cancellation
918 * of the command.
919 */
920 if (PQtransactionStatus(entry->conn) == PQTRANS_ACTIVE &&
921 !pgfdw_cancel_query(entry->conn))
922 abort_cleanup_failure = true;
923 else
924 {
925 /* Rollback all remote subtransactions during abort */
926 snprintf(sql, sizeof(sql),
927 "ROLLBACK TO SAVEPOINT s%d; RELEASE SAVEPOINT s%d",
928 curlevel, curlevel);
929 if (!pgfdw_exec_cleanup_query(entry->conn, sql, false))
930 abort_cleanup_failure = true;
931 }
932
933 /* Disarm changing_xact_state if it all worked. */
934 entry->changing_xact_state = abort_cleanup_failure;
935 }
936
937 /* OK, we're outta that level of subtransaction */
938 entry->xact_depth--;
939 }
940 }
941
942 /*
943 * Connection invalidation callback function
944 *
945 * After a change to a pg_foreign_server or pg_user_mapping catalog entry,
946 * close connections depending on that entry immediately if current transaction
947 * has not used those connections yet. Otherwise, mark those connections as
948 * invalid and then make pgfdw_xact_callback() close them at the end of current
949 * transaction, since they cannot be closed in the midst of the transaction
950 * using them. Closed connections will be remade at the next opportunity if
951 * necessary.
952 *
953 * Although most cache invalidation callbacks blow away all the related stuff
954 * regardless of the given hashvalue, connections are expensive enough that
955 * it's worth trying to avoid that.
956 *
957 * NB: We could avoid unnecessary disconnection more strictly by examining
958 * individual option values, but it seems too much effort for the gain.
959 */
960 static void
961 pgfdw_inval_callback(Datum arg, int cacheid, uint32 hashvalue)
962 {
963 HASH_SEQ_STATUS scan;
964 ConnCacheEntry *entry;
965
966 Assert(cacheid == FOREIGNSERVEROID || cacheid == USERMAPPINGOID);
967
968 /* ConnectionHash must exist already, if we're registered */
969 hash_seq_init(&scan, ConnectionHash);
970 while ((entry = (ConnCacheEntry *) hash_seq_search(&scan)))
971 {
972 /* Ignore invalid entries */
973 if (entry->conn == NULL)
974 continue;
975
976 /* hashvalue == 0 means a cache reset, must clear all state */
977 if (hashvalue == 0 ||
978 (cacheid == FOREIGNSERVEROID &&
979 entry->server_hashvalue == hashvalue) ||
980 (cacheid == USERMAPPINGOID &&
981 entry->mapping_hashvalue == hashvalue))
982 {
983 /*
984 * Close the connection immediately if it's not used yet in this
985 * transaction. Otherwise mark it as invalid so that
986 * pgfdw_xact_callback() can close it at the end of this
987 * transaction.
988 */
989 if (entry->xact_depth == 0)
990 {
991 elog(DEBUG3, "discarding connection %p", entry->conn);
992 disconnect_pg_server(entry);
993 }
994 else
995 entry->invalidated = true;
996 }
997 }
998 }
999
1000 /*
1001 * Raise an error if the given connection cache entry is marked as being
1002 * in the middle of an xact state change. This should be called at which no
1003 * such change is expected to be in progress; if one is found to be in
1004 * progress, it means that we aborted in the middle of a previous state change
1005 * and now don't know what the remote transaction state actually is.
1006 * Such connections can't safely be further used. Re-establishing the
1007 * connection would change the snapshot and roll back any writes already
1008 * performed, so that's not an option, either. Thus, we must abort.
1009 */
1010 static void
1011 pgfdw_reject_incomplete_xact_state_change(ConnCacheEntry *entry)
1012 {
1013 HeapTuple tup;
1014 Form_pg_user_mapping umform;
1015 ForeignServer *server;
1016
1017 /* nothing to do for inactive entries and entries of sane state */
1018 if (entry->conn == NULL || !entry->changing_xact_state)
1019 return;
1020
1021 /* make sure this entry is inactive */
1022 disconnect_pg_server(entry);
1023
1024 /* find server name to be shown in the message below */
1025 tup = SearchSysCache1(USERMAPPINGOID,
1026 ObjectIdGetDatum(entry->key));
1027 if (!HeapTupleIsValid(tup))
1028 elog(ERROR, "cache lookup failed for user mapping %u", entry->key);
1029 umform = (Form_pg_user_mapping) GETSTRUCT(tup);
1030 server = GetForeignServer(umform->umserver);
1031 ReleaseSysCache(tup);
1032
1033 ereport(ERROR,
1034 (errcode(ERRCODE_CONNECTION_EXCEPTION),
1035 errmsg("connection to server \"%s\" was lost",
1036 server->servername)));
1037 }
1038
1039 /*
1040 * Cancel the currently-in-progress query (whose query text we do not have)
1041 * and ignore the result. Returns true if we successfully cancel the query
1042 * and discard any pending result, and false if not.
1043 *
1044 * It's not a huge problem if we throw an ERROR here, but if we get into error
1045 * recursion trouble, we'll end up slamming the connection shut, which will
1046 * necessitate failing the entire toplevel transaction even if subtransactions
1047 * were used. Try to use WARNING where we can.
1048 */
1049 static bool
1050 pgfdw_cancel_query(PGconn *conn)
1051 {
1052 PGcancel *cancel;
1053 char errbuf[256];
1054 PGresult *result = NULL;
1055 TimestampTz endtime;
1056
1057 /*
1058 * If it takes too long to cancel the query and discard the result, assume
1059 * the connection is dead.
1060 */
1061 endtime = TimestampTzPlusMilliseconds(GetCurrentTimestamp(), 30000);
1062
1063 /*
1064 * Issue cancel request. Unfortunately, there's no good way to limit the
1065 * amount of time that we might block inside PQgetCancel().
1066 */
1067 if ((cancel = PQgetCancel(conn)))
1068 {
1069 if (!PQcancel(cancel, errbuf, sizeof(errbuf)))
1070 {
1071 ereport(WARNING,
1072 (errcode(ERRCODE_CONNECTION_FAILURE),
1073 errmsg("could not send cancel request: %s",
1074 errbuf)));
1075 PQfreeCancel(cancel);
1076 return false;
1077 }
1078 PQfreeCancel(cancel);
1079 }
1080
1081 /* Get and discard the result of the query. */
1082 if (pgfdw_get_cleanup_result(conn, endtime, &result))
1083 return false;
1084 PQclear(result);
1085
1086 return true;
1087 }
1088
1089 /*
1090 * Submit a query during (sub)abort cleanup and wait up to 30 seconds for the
1091 * result. If the query is executed without error, the return value is true.
1092 * If the query is executed successfully but returns an error, the return
1093 * value is true if and only if ignore_errors is set. If the query can't be
1094 * sent or times out, the return value is false.
1095 *
1096 * It's not a huge problem if we throw an ERROR here, but if we get into error
1097 * recursion trouble, we'll end up slamming the connection shut, which will
1098 * necessitate failing the entire toplevel transaction even if subtransactions
1099 * were used. Try to use WARNING where we can.
1100 */
1101 static bool
1102 pgfdw_exec_cleanup_query(PGconn *conn, const char *query, bool ignore_errors)
1103 {
1104 PGresult *result = NULL;
1105 TimestampTz endtime;
1106
1107 /*
1108 * If it takes too long to execute a cleanup query, assume the connection
1109 * is dead. It's fairly likely that this is why we aborted in the first
1110 * place (e.g. statement timeout, user cancel), so the timeout shouldn't
1111 * be too long.
1112 */
1113 endtime = TimestampTzPlusMilliseconds(GetCurrentTimestamp(), 30000);
1114
1115 /*
1116 * Submit a query. Since we don't use non-blocking mode, this also can
1117 * block. But its risk is relatively small, so we ignore that for now.
1118 */
1119 if (!PQsendQuery(conn, query))
1120 {
1121 pgfdw_report_error(WARNING, NULL, conn, false, query);
1122 return false;
1123 }
1124
1125 /* Get the result of the query. */
1126 if (pgfdw_get_cleanup_result(conn, endtime, &result))
1127 return false;
1128
1129 /* Issue a warning if not successful. */
1130 if (PQresultStatus(result) != PGRES_COMMAND_OK)
1131 {
1132 pgfdw_report_error(WARNING, result, conn, true, query);
1133 return ignore_errors;
1134 }
1135 PQclear(result);
1136
1137 return true;
1138 }
1139
1140 /*
1141 * Get, during abort cleanup, the result of a query that is in progress. This
1142 * might be a query that is being interrupted by transaction abort, or it might
1143 * be a query that was initiated as part of transaction abort to get the remote
1144 * side back to the appropriate state.
1145 *
1146 * endtime is the time at which we should give up and assume the remote
1147 * side is dead. Returns true if the timeout expired, otherwise false.
1148 * Sets *result except in case of a timeout.
1149 */
1150 static bool
1151 pgfdw_get_cleanup_result(PGconn *conn, TimestampTz endtime, PGresult **result)
1152 {
1153 volatile bool timed_out = false;
1154 PGresult *volatile last_res = NULL;
1155
1156 /* In what follows, do not leak any PGresults on an error. */
1157 PG_TRY();
1158 {
1159 for (;;)
1160 {
1161 PGresult *res;
1162
1163 while (PQisBusy(conn))
1164 {
1165 int wc;
1166 TimestampTz now = GetCurrentTimestamp();
1167 long cur_timeout;
1168
1169 /* If timeout has expired, give up, else get sleep time. */
1170 cur_timeout = TimestampDifferenceMilliseconds(now, endtime);
1171 if (cur_timeout <= 0)
1172 {
1173 timed_out = true;
1174 goto exit;
1175 }
1176
1177 /* Sleep until there's something to do */
1178 wc = WaitLatchOrSocket(MyLatch,
1179 WL_LATCH_SET | WL_SOCKET_READABLE |
1180 WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
1181 PQsocket(conn),
1182 cur_timeout, PG_WAIT_EXTENSION);
1183 ResetLatch(MyLatch);
1184
1185 CHECK_FOR_INTERRUPTS();
1186
1187 /* Data available in socket? */
1188 if (wc & WL_SOCKET_READABLE)
1189 {
1190 if (!PQconsumeInput(conn))
1191 {
1192 /* connection trouble; treat the same as a timeout */
1193 timed_out = true;
1194 goto exit;
1195 }
1196 }
1197 }
1198
1199 res = PQgetResult(conn);
1200 if (res == NULL)
1201 break; /* query is complete */
1202
1203 PQclear(last_res);
1204 last_res = res;
1205 }
1206 exit: ;
1207 }
1208 PG_CATCH();
1209 {
1210 PQclear(last_res);
1211 PG_RE_THROW();
1212 }
1213 PG_END_TRY();
1214
1215 if (timed_out)
1216 PQclear(last_res);
1217 else
1218 *result = last_res;
1219 return timed_out;
1220 }
1221