granicus.if.org Git - postgresql/blob - contrib/pg_upgrade/pg_upgrade.c

   1 /*
   2  *      pg_upgrade.c
   3  *
   4  *      main source file
   5  *
   6  *      Copyright (c) 2010-2014, PostgreSQL Global Development Group
   7  *      contrib/pg_upgrade/pg_upgrade.c
   8  */
   9
  10 /*
  11  *      To simplify the upgrade process, we force certain system values to be
  12  *      identical between old and new clusters:
  13  *
  14  *      We control all assignments of pg_class.oid (and relfilenode) so toast
  15  *      oids are the same between old and new clusters.  This is important
  16  *      because toast oids are stored as toast pointers in user tables.
  17  *
  18  *      While pg_class.oid and pg_class.relfilenode are initially the same
  19  *      in a cluster, they can diverge due to CLUSTER, REINDEX, or VACUUM
  20  *      FULL.  In the new cluster, pg_class.oid and pg_class.relfilenode will
  21  *      be the same and will match the old pg_class.oid value.  Because of
  22  *      this, old/new pg_class.relfilenode values will not match if CLUSTER,
  23  *      REINDEX, or VACUUM FULL have been performed in the old cluster.
  24  *
  25  *      We control all assignments of pg_type.oid because these oids are stored
  26  *      in user composite type values.
  27  *
  28  *      We control all assignments of pg_enum.oid because these oids are stored
  29  *      in user tables as enum values.
  30  *
  31  *      We control all assignments of pg_authid.oid because these oids are stored
  32  *      in pg_largeobject_metadata.
  33  */
  34
  35
  36
  37 #include "postgres_fe.h"
  38
  39 #include "pg_upgrade.h"
  40
  41 #ifdef HAVE_LANGINFO_H
  42 #include <langinfo.h>
  43 #endif
  44
  45 static void prepare_new_cluster(void);
  46 static void prepare_new_databases(void);
  47 static void create_new_objects(void);
  48 static void copy_clog_xlog_xid(void);
  49 static void set_frozenxids(void);
  50 static void setup(char *argv0, bool *live_check);
  51 static void cleanup(void);
  52
  53 ClusterInfo old_cluster,
  54                         new_cluster;
  55 OSInfo          os_info;
  56
  57 char       *output_files[] = {
  58         SERVER_LOG_FILE,
  59 #ifdef WIN32
  60         /* unique file for pg_ctl start */
  61         SERVER_START_LOG_FILE,
  62 #endif
  63         UTILITY_LOG_FILE,
  64         INTERNAL_LOG_FILE,
  65         NULL
  66 };
  67
  68
  69 int
  70 main(int argc, char **argv)
  71 {
  72         char       *sequence_script_file_name = NULL;
  73         char       *analyze_script_file_name = NULL;
  74         char       *deletion_script_file_name = NULL;
  75         bool            live_check = false;
  76
  77         parseCommandLine(argc, argv);
  78
  79         adjust_data_dir(&old_cluster);
  80         adjust_data_dir(&new_cluster);
  81
  82         setup(argv[0], &live_check);
  83
  84         output_check_banner(live_check);
  85
  86         check_cluster_versions();
  87
  88         get_sock_dir(&old_cluster, live_check);
  89         get_sock_dir(&new_cluster, false);
  90
  91         check_cluster_compatibility(live_check);
  92
  93         check_and_dump_old_cluster(live_check, &sequence_script_file_name);
  94
  95
  96         /* -- NEW -- */
  97         start_postmaster(&new_cluster, true);
  98
  99         check_new_cluster();
 100         report_clusters_compatible();
 101
 102         pg_log(PG_REPORT, "\nPerforming Upgrade\n");
 103         pg_log(PG_REPORT, "------------------\n");
 104
 105         prepare_new_cluster();
 106
 107         stop_postmaster(false);
 108
 109         /*
 110          * Destructive Changes to New Cluster
 111          */
 112
 113         copy_clog_xlog_xid();
 114
 115         /* New now using xids of the old system */
 116
 117         /* -- NEW -- */
 118         start_postmaster(&new_cluster, true);
 119
 120         prepare_new_databases();
 121
 122         create_new_objects();
 123
 124         stop_postmaster(false);
 125
 126         /*
 127          * Most failures happen in create_new_objects(), which has completed at
 128          * this point.  We do this here because it is just before linking, which
 129          * will link the old and new cluster data files, preventing the old
 130          * cluster from being safely started once the new cluster is started.
 131          */
 132         if (user_opts.transfer_mode == TRANSFER_MODE_LINK)
 133                 disable_old_cluster();
 134
 135         transfer_all_new_tablespaces(&old_cluster.dbarr, &new_cluster.dbarr,
 136                                                                  old_cluster.pgdata, new_cluster.pgdata);
 137
 138         /*
 139          * Assuming OIDs are only used in system tables, there is no need to
 140          * restore the OID counter because we have not transferred any OIDs from
 141          * the old system, but we do it anyway just in case.  We do it late here
 142          * because there is no need to have the schema load use new oids.
 143          */
 144         prep_status("Setting next OID for new cluster");
 145         exec_prog(UTILITY_LOG_FILE, NULL, true,
 146                           "\"%s/pg_resetxlog\" -o %u \"%s\"",
 147                           new_cluster.bindir, old_cluster.controldata.chkpnt_nxtoid,
 148                           new_cluster.pgdata);
 149         check_ok();
 150
 151         prep_status("Sync data directory to disk");
 152         exec_prog(UTILITY_LOG_FILE, NULL, true,
 153                           "\"%s/initdb\" --sync-only \"%s\"", new_cluster.bindir,
 154                           new_cluster.pgdata);
 155         check_ok();
 156
 157         create_script_for_cluster_analyze(&analyze_script_file_name);
 158         create_script_for_old_cluster_deletion(&deletion_script_file_name);
 159
 160         issue_warnings(sequence_script_file_name);
 161
 162         pg_log(PG_REPORT, "\nUpgrade Complete\n");
 163         pg_log(PG_REPORT, "----------------\n");
 164
 165         output_completion_banner(analyze_script_file_name,
 166                                                          deletion_script_file_name);
 167
 168         pg_free(analyze_script_file_name);
 169         pg_free(deletion_script_file_name);
 170         pg_free(sequence_script_file_name);
 171
 172         cleanup();
 173
 174         return 0;
 175 }
 176
 177
 178 static void
 179 setup(char *argv0, bool *live_check)
 180 {
 181         char            exec_path[MAXPGPATH];   /* full path to my executable */
 182
 183         /*
 184          * make sure the user has a clean environment, otherwise, we may confuse
 185          * libpq when we connect to one (or both) of the servers.
 186          */
 187         check_pghost_envvar();
 188
 189         verify_directories();
 190
 191         /* no postmasters should be running, except for a live check */
 192         if (pid_lock_file_exists(old_cluster.pgdata))
 193         {
 194                 /*
 195                  * If we have a postmaster.pid file, try to start the server.  If it
 196                  * starts, the pid file was stale, so stop the server.  If it doesn't
 197                  * start, assume the server is running.  If the pid file is left over
 198                  * from a server crash, this also allows any committed transactions
 199                  * stored in the WAL to be replayed so they are not lost, because WAL
 200                  * files are not transfered from old to new servers.
 201                  */
 202                 if (start_postmaster(&old_cluster, false))
 203                         stop_postmaster(false);
 204                 else
 205                 {
 206                         if (!user_opts.check)
 207                                 pg_fatal("There seems to be a postmaster servicing the old cluster.\n"
 208                                                  "Please shutdown that postmaster and try again.\n");
 209                         else
 210                                 *live_check = true;
 211                 }
 212         }
 213
 214         /* same goes for the new postmaster */
 215         if (pid_lock_file_exists(new_cluster.pgdata))
 216         {
 217                 if (start_postmaster(&new_cluster, false))
 218                         stop_postmaster(false);
 219                 else
 220                         pg_fatal("There seems to be a postmaster servicing the new cluster.\n"
 221                                          "Please shutdown that postmaster and try again.\n");
 222         }
 223
 224         /* get path to pg_upgrade executable */
 225         if (find_my_exec(argv0, exec_path) < 0)
 226                 pg_fatal("Could not get path name to pg_upgrade: %s\n", getErrorText(errno));
 227
 228         /* Trim off program name and keep just path */
 229         *last_dir_separator(exec_path) = '\0';
 230         canonicalize_path(exec_path);
 231         os_info.exec_path = pg_strdup(exec_path);
 232 }
 233
 234
 235 static void
 236 prepare_new_cluster(void)
 237 {
 238         /*
 239          * It would make more sense to freeze after loading the schema, but that
 240          * would cause us to lose the frozenids restored by the load. We use
 241          * --analyze so autovacuum doesn't update statistics later
 242          */
 243         prep_status("Analyzing all rows in the new cluster");
 244         exec_prog(UTILITY_LOG_FILE, NULL, true,
 245                           "\"%s/vacuumdb\" %s --all --analyze %s",
 246                           new_cluster.bindir, cluster_conn_opts(&new_cluster),
 247                           log_opts.verbose ? "--verbose" : "");
 248         check_ok();
 249
 250         /*
 251          * We do freeze after analyze so pg_statistic is also frozen. template0 is
 252          * not frozen here, but data rows were frozen by initdb, and we set its
 253          * datfrozenxid and relfrozenxids later to match the new xid counter
 254          * later.
 255          */
 256         prep_status("Freezing all rows on the new cluster");
 257         exec_prog(UTILITY_LOG_FILE, NULL, true,
 258                           "\"%s/vacuumdb\" %s --all --freeze %s",
 259                           new_cluster.bindir, cluster_conn_opts(&new_cluster),
 260                           log_opts.verbose ? "--verbose" : "");
 261         check_ok();
 262
 263         get_pg_database_relfilenode(&new_cluster);
 264 }
 265
 266
 267 static void
 268 prepare_new_databases(void)
 269 {
 270         /*
 271          * We set autovacuum_freeze_max_age to its maximum value so autovacuum
 272          * does not launch here and delete clog files, before the frozen xids are
 273          * set.
 274          */
 275
 276         set_frozenxids();
 277
 278         prep_status("Restoring global objects in the new cluster");
 279
 280         /*
 281          * Install support functions in the global-object restore database to
 282          * preserve pg_authid.oid.  pg_dumpall uses 'template0' as its template
 283          * database so objects we add into 'template1' are not propogated.  They
 284          * are removed on pg_upgrade exit.
 285          */
 286         install_support_functions_in_new_db("template1");
 287
 288         /*
 289          * We have to create the databases first so we can install support
 290          * functions in all the other databases.  Ideally we could create the
 291          * support functions in template1 but pg_dumpall creates database using
 292          * the template0 template.
 293          */
 294         exec_prog(UTILITY_LOG_FILE, NULL, true,
 295                           "\"%s/psql\" " EXEC_PSQL_ARGS " %s -f \"%s\"",
 296                           new_cluster.bindir, cluster_conn_opts(&new_cluster),
 297                           GLOBALS_DUMP_FILE);
 298         check_ok();
 299
 300         /* we load this to get a current list of databases */
 301         get_db_and_rel_infos(&new_cluster);
 302 }
 303
 304
 305 static void
 306 create_new_objects(void)
 307 {
 308         int                     dbnum;
 309
 310         prep_status("Adding support functions to new cluster");
 311
 312         /*
 313          * Technically, we only need to install these support functions in new
 314          * databases that also exist in the old cluster, but for completeness we
 315          * process all new databases.
 316          */
 317         for (dbnum = 0; dbnum < new_cluster.dbarr.ndbs; dbnum++)
 318         {
 319                 DbInfo     *new_db = &new_cluster.dbarr.dbs[dbnum];
 320
 321                 /* skip db we already installed */
 322                 if (strcmp(new_db->db_name, "template1") != 0)
 323                         install_support_functions_in_new_db(new_db->db_name);
 324         }
 325         check_ok();
 326
 327         prep_status("Restoring database schemas in the new cluster\n");
 328
 329         for (dbnum = 0; dbnum < old_cluster.dbarr.ndbs; dbnum++)
 330         {
 331                 char            sql_file_name[MAXPGPATH],
 332                                         log_file_name[MAXPGPATH];
 333                 DbInfo     *old_db = &old_cluster.dbarr.dbs[dbnum];
 334
 335                 pg_log(PG_STATUS, "%s", old_db->db_name);
 336                 snprintf(sql_file_name, sizeof(sql_file_name), DB_DUMP_FILE_MASK, old_db->db_oid);
 337                 snprintf(log_file_name, sizeof(log_file_name), DB_DUMP_LOG_FILE_MASK, old_db->db_oid);
 338
 339                 /*
 340                  * pg_dump only produces its output at the end, so there is little
 341                  * parallelism if using the pipe.
 342                  */
 343                 parallel_exec_prog(log_file_name,
 344                                                    NULL,
 345                                                    "\"%s/pg_restore\" %s --exit-on-error --verbose --dbname \"%s\" \"%s\"",
 346                                                    new_cluster.bindir,
 347                                                    cluster_conn_opts(&new_cluster),
 348                                                    old_db->db_name,
 349                                                    sql_file_name);
 350         }
 351
 352         /* reap all children */
 353         while (reap_child(true) == true)
 354                 ;
 355
 356         end_progress_output();
 357         check_ok();
 358
 359         /* regenerate now that we have objects in the databases */
 360         get_db_and_rel_infos(&new_cluster);
 361
 362         uninstall_support_functions_from_new_cluster();
 363 }
 364
 365 /*
 366  * Delete the given subdirectory contents from the new cluster
 367  */
 368 static void
 369 remove_new_subdir(char *subdir, bool rmtopdir)
 370 {
 371         char            new_path[MAXPGPATH];
 372
 373         prep_status("Deleting files from new %s", subdir);
 374
 375         snprintf(new_path, sizeof(new_path), "%s/%s", new_cluster.pgdata, subdir);
 376         if (!rmtree(new_path, rmtopdir))
 377                 pg_fatal("could not delete directory \"%s\"\n", new_path);
 378
 379         check_ok();
 380 }
 381
 382 /*
 383  * Copy the files from the old cluster into it
 384  */
 385 static void
 386 copy_subdir_files(char *subdir)
 387 {
 388         char            old_path[MAXPGPATH];
 389         char            new_path[MAXPGPATH];
 390
 391         remove_new_subdir(subdir, true);
 392
 393         snprintf(old_path, sizeof(old_path), "%s/%s", old_cluster.pgdata, subdir);
 394         snprintf(new_path, sizeof(new_path), "%s/%s", new_cluster.pgdata, subdir);
 395
 396         prep_status("Copying old %s to new server", subdir);
 397
 398         exec_prog(UTILITY_LOG_FILE, NULL, true,
 399 #ifndef WIN32
 400                           "cp -Rf \"%s\" \"%s\"",
 401 #else
 402         /* flags: everything, no confirm, quiet, overwrite read-only */
 403                           "xcopy /e /y /q /r \"%s\" \"%s\\\"",
 404 #endif
 405                           old_path, new_path);
 406
 407         check_ok();
 408 }
 409
 410 static void
 411 copy_clog_xlog_xid(void)
 412 {
 413         /* copy old commit logs to new data dir */
 414         copy_subdir_files("pg_clog");
 415
 416         /* set the next transaction id of the new cluster */
 417         prep_status("Setting next transaction ID for new cluster");
 418         exec_prog(UTILITY_LOG_FILE, NULL, true,
 419                           "\"%s/pg_resetxlog\" -f -x %u \"%s\"",
 420                           new_cluster.bindir, old_cluster.controldata.chkpnt_nxtxid,
 421                           new_cluster.pgdata);
 422         check_ok();
 423
 424         /*
 425          * If the old server is before the MULTIXACT_FORMATCHANGE_CAT_VER change
 426          * (see pg_upgrade.h) and the new server is after, then we don't copy
 427          * pg_multixact files, but we need to reset pg_control so that the new
 428          * server doesn't attempt to read multis older than the cutoff value.
 429          */
 430         if (old_cluster.controldata.cat_ver >= MULTIXACT_FORMATCHANGE_CAT_VER &&
 431                 new_cluster.controldata.cat_ver >= MULTIXACT_FORMATCHANGE_CAT_VER)
 432         {
 433                 copy_subdir_files("pg_multixact/offsets");
 434                 copy_subdir_files("pg_multixact/members");
 435
 436                 prep_status("Setting next multixact ID and offset for new cluster");
 437
 438                 /*
 439                  * we preserve all files and contents, so we must preserve both "next"
 440                  * counters here and the oldest multi present on system.
 441                  */
 442                 exec_prog(UTILITY_LOG_FILE, NULL, true,
 443                                   "\"%s/pg_resetxlog\" -O %u -m %u,%u \"%s\"",
 444                                   new_cluster.bindir,
 445                                   old_cluster.controldata.chkpnt_nxtmxoff,
 446                                   old_cluster.controldata.chkpnt_nxtmulti,
 447                                   old_cluster.controldata.chkpnt_oldstMulti,
 448                                   new_cluster.pgdata);
 449                 check_ok();
 450         }
 451         else if (new_cluster.controldata.cat_ver >= MULTIXACT_FORMATCHANGE_CAT_VER)
 452         {
 453                 /*
 454                  * Remove files created by initdb that no longer match the
 455                  * new multi-xid value.
 456                  */
 457                 remove_new_subdir("pg_multixact/offsets", false);
 458                 remove_new_subdir("pg_multixact/members", false);
 459
 460                 prep_status("Setting oldest multixact ID on new cluster");
 461
 462                 /*
 463                  * We don't preserve files in this case, but it's important that the
 464                  * oldest multi is set to the latest value used by the old system, so
 465                  * that multixact.c returns the empty set for multis that might be
 466                  * present on disk.  We set next multi to the value following that; it
 467                  * might end up wrapped around (i.e. 0) if the old cluster had
 468                  * next=MaxMultiXactId, but multixact.c can cope with that just fine.
 469                  */
 470                 exec_prog(UTILITY_LOG_FILE, NULL, true,
 471                                   "\"%s/pg_resetxlog\" -m %u,%u \"%s\"",
 472                                   new_cluster.bindir,
 473                                   old_cluster.controldata.chkpnt_nxtmulti + 1,
 474                                   old_cluster.controldata.chkpnt_nxtmulti,
 475                                   new_cluster.pgdata);
 476                 check_ok();
 477         }
 478
 479         /* now reset the wal archives in the new cluster */
 480         prep_status("Resetting WAL archives");
 481         exec_prog(UTILITY_LOG_FILE, NULL, true,
 482                           "\"%s/pg_resetxlog\" -l %s \"%s\"", new_cluster.bindir,
 483                           old_cluster.controldata.nextxlogfile,
 484                           new_cluster.pgdata);
 485         check_ok();
 486 }
 487
 488
 489 /*
 490  *      set_frozenxids()
 491  *
 492  *      We have frozen all xids, so set relfrozenxid and datfrozenxid
 493  *      to be the old cluster's xid counter, which we just set in the new
 494  *      cluster.  User-table frozenxid values will be set by pg_dump
 495  *      --binary-upgrade, but objects not set by the pg_dump must have
 496  *      proper frozen counters.
 497  */
 498 static
 499 void
 500 set_frozenxids(void)
 501 {
 502         int                     dbnum;
 503         PGconn     *conn,
 504                            *conn_template1;
 505         PGresult   *dbres;
 506         int                     ntups;
 507         int                     i_datname;
 508         int                     i_datallowconn;
 509
 510         prep_status("Setting frozenxid counters in new cluster");
 511
 512         conn_template1 = connectToServer(&new_cluster, "template1");
 513
 514         /* set pg_database.datfrozenxid */
 515         PQclear(executeQueryOrDie(conn_template1,
 516                                                           "UPDATE pg_catalog.pg_database "
 517                                                           "SET  datfrozenxid = '%u'",
 518                                                           old_cluster.controldata.chkpnt_nxtxid));
 519
 520         /* get database names */
 521         dbres = executeQueryOrDie(conn_template1,
 522                                                           "SELECT       datname, datallowconn "
 523                                                           "FROM pg_catalog.pg_database");
 524
 525         i_datname = PQfnumber(dbres, "datname");
 526         i_datallowconn = PQfnumber(dbres, "datallowconn");
 527
 528         ntups = PQntuples(dbres);
 529         for (dbnum = 0; dbnum < ntups; dbnum++)
 530         {
 531                 char       *datname = PQgetvalue(dbres, dbnum, i_datname);
 532                 char       *datallowconn = PQgetvalue(dbres, dbnum, i_datallowconn);
 533
 534                 /*
 535                  * We must update databases where datallowconn = false, e.g.
 536                  * template0, because autovacuum increments their datfrozenxids and
 537                  * relfrozenxids even if autovacuum is turned off, and even though all
 538                  * the data rows are already frozen  To enable this, we temporarily
 539                  * change datallowconn.
 540                  */
 541                 if (strcmp(datallowconn, "f") == 0)
 542                         PQclear(executeQueryOrDie(conn_template1,
 543                                                                 "ALTER DATABASE %s ALLOW_CONNECTIONS = true",
 544                                                                           quote_identifier(datname)));
 545
 546                 conn = connectToServer(&new_cluster, datname);
 547
 548                 /* set pg_class.relfrozenxid */
 549                 PQclear(executeQueryOrDie(conn,
 550                                                                   "UPDATE       pg_catalog.pg_class "
 551                                                                   "SET  relfrozenxid = '%u' "
 552                 /* only heap, materialized view, and TOAST are vacuumed */
 553                                                                   "WHERE        relkind IN ('r', 'm', 't')",
 554                                                                   old_cluster.controldata.chkpnt_nxtxid));
 555                 PQfinish(conn);
 556
 557                 /* Reset datallowconn flag */
 558                 if (strcmp(datallowconn, "f") == 0)
 559                         PQclear(executeQueryOrDie(conn_template1,
 560                                                            "ALTER DATABASE %s ALLOW_CONNECTIONS = false",
 561                                                                           quote_identifier(datname)));
 562         }
 563
 564         PQclear(dbres);
 565
 566         PQfinish(conn_template1);
 567
 568         check_ok();
 569 }
 570
 571
 572 static void
 573 cleanup(void)
 574 {
 575         fclose(log_opts.internal);
 576
 577         /* Remove dump and log files? */
 578         if (!log_opts.retain)
 579         {
 580                 int                     dbnum;
 581                 char      **filename;
 582
 583                 for (filename = output_files; *filename != NULL; filename++)
 584                         unlink(*filename);
 585
 586                 /* remove dump files */
 587                 unlink(GLOBALS_DUMP_FILE);
 588
 589                 if (old_cluster.dbarr.dbs)
 590                         for (dbnum = 0; dbnum < old_cluster.dbarr.ndbs; dbnum++)
 591                         {
 592                                 char            sql_file_name[MAXPGPATH],
 593                                                         log_file_name[MAXPGPATH];
 594                                 DbInfo     *old_db = &old_cluster.dbarr.dbs[dbnum];
 595
 596                                 snprintf(sql_file_name, sizeof(sql_file_name), DB_DUMP_FILE_MASK, old_db->db_oid);
 597                                 unlink(sql_file_name);
 598
 599                                 snprintf(log_file_name, sizeof(log_file_name), DB_DUMP_LOG_FILE_MASK, old_db->db_oid);
 600                                 unlink(log_file_name);
 601                         }
 602         }
 603 }