granicus.if.org Git - postgresql/blob - src/backend/postmaster/pgstat.c

   1 /* ----------
   2  * pgstat.c
   3  *
   4  *      All the statistics collector stuff hacked up in one big, ugly file.
   5  *
   6  *      TODO:   - Separate collector, postmaster and backend stuff
   7  *                        into different files.
   8  *
   9  *                      - Add some automatic call for pgstat vacuuming.
  10  *
  11  *                      - Add a pgstat config column to pg_database, so this
  12  *                        entire thing can be enabled/disabled on a per db basis.
  13  *
  14  *      Copyright (c) 2001-2019, PostgreSQL Global Development Group
  15  *
  16  *      src/backend/postmaster/pgstat.c
  17  * ----------
  18  */
  19 #include "postgres.h"
  20
  21 #include <unistd.h>
  22 #include <fcntl.h>
  23 #include <sys/param.h>
  24 #include <sys/time.h>
  25 #include <sys/socket.h>
  26 #include <netdb.h>
  27 #include <netinet/in.h>
  28 #include <arpa/inet.h>
  29 #include <signal.h>
  30 #include <time.h>
  31 #ifdef HAVE_SYS_SELECT_H
  32 #include <sys/select.h>
  33 #endif
  34
  35 #include "pgstat.h"
  36
  37 #include "access/heapam.h"
  38 #include "access/htup_details.h"
  39 #include "access/transam.h"
  40 #include "access/twophase_rmgr.h"
  41 #include "access/xact.h"
  42 #include "catalog/pg_database.h"
  43 #include "catalog/pg_proc.h"
  44 #include "common/ip.h"
  45 #include "libpq/libpq.h"
  46 #include "libpq/pqsignal.h"
  47 #include "mb/pg_wchar.h"
  48 #include "miscadmin.h"
  49 #include "pg_trace.h"
  50 #include "postmaster/autovacuum.h"
  51 #include "postmaster/fork_process.h"
  52 #include "postmaster/postmaster.h"
  53 #include "replication/walsender.h"
  54 #include "storage/backendid.h"
  55 #include "storage/dsm.h"
  56 #include "storage/fd.h"
  57 #include "storage/ipc.h"
  58 #include "storage/latch.h"
  59 #include "storage/lmgr.h"
  60 #include "storage/pg_shmem.h"
  61 #include "storage/procsignal.h"
  62 #include "storage/sinvaladt.h"
  63 #include "utils/ascii.h"
  64 #include "utils/guc.h"
  65 #include "utils/memutils.h"
  66 #include "utils/ps_status.h"
  67 #include "utils/rel.h"
  68 #include "utils/snapmgr.h"
  69 #include "utils/timestamp.h"
  70 #include "utils/tqual.h"
  71
  72
  73 /* ----------
  74  * Timer definitions.
  75  * ----------
  76  */
  77 #define PGSTAT_STAT_INTERVAL    500 /* Minimum time between stats file
  78                                                                          * updates; in milliseconds. */
  79
  80 #define PGSTAT_RETRY_DELAY              10      /* How long to wait between checks for a
  81                                                                          * new file; in milliseconds. */
  82
  83 #define PGSTAT_MAX_WAIT_TIME    10000   /* Maximum time to wait for a stats
  84                                                                                  * file update; in milliseconds. */
  85
  86 #define PGSTAT_INQ_INTERVAL             640 /* How often to ping the collector for a
  87                                                                          * new file; in milliseconds. */
  88
  89 #define PGSTAT_RESTART_INTERVAL 60      /* How often to attempt to restart a
  90                                                                          * failed statistics collector; in
  91                                                                          * seconds. */
  92
  93 #define PGSTAT_POLL_LOOP_COUNT  (PGSTAT_MAX_WAIT_TIME / PGSTAT_RETRY_DELAY)
  94 #define PGSTAT_INQ_LOOP_COUNT   (PGSTAT_INQ_INTERVAL / PGSTAT_RETRY_DELAY)
  95
  96 /* Minimum receive buffer size for the collector's socket. */
  97 #define PGSTAT_MIN_RCVBUF               (100 * 1024)
  98
  99
 100 /* ----------
 101  * The initial size hints for the hash tables used in the collector.
 102  * ----------
 103  */
 104 #define PGSTAT_DB_HASH_SIZE             16
 105 #define PGSTAT_TAB_HASH_SIZE    512
 106 #define PGSTAT_FUNCTION_HASH_SIZE       512
 107
 108
 109 /* ----------
 110  * Total number of backends including auxiliary
 111  *
 112  * We reserve a slot for each possible BackendId, plus one for each
 113  * possible auxiliary process type.  (This scheme assumes there is not
 114  * more than one of any auxiliary process type at a time.) MaxBackends
 115  * includes autovacuum workers and background workers as well.
 116  * ----------
 117  */
 118 #define NumBackendStatSlots (MaxBackends + NUM_AUXPROCTYPES)
 119
 120
 121 /* ----------
 122  * GUC parameters
 123  * ----------
 124  */
 125 bool            pgstat_track_activities = false;
 126 bool            pgstat_track_counts = false;
 127 int                     pgstat_track_functions = TRACK_FUNC_OFF;
 128 int                     pgstat_track_activity_query_size = 1024;
 129
 130 /* ----------
 131  * Built from GUC parameter
 132  * ----------
 133  */
 134 char       *pgstat_stat_directory = NULL;
 135 char       *pgstat_stat_filename = NULL;
 136 char       *pgstat_stat_tmpname = NULL;
 137
 138 /*
 139  * BgWriter global statistics counters (unused in other processes).
 140  * Stored directly in a stats message structure so it can be sent
 141  * without needing to copy things around.  We assume this inits to zeroes.
 142  */
 143 PgStat_MsgBgWriter BgWriterStats;
 144
 145 /* ----------
 146  * Local data
 147  * ----------
 148  */
 149 NON_EXEC_STATIC pgsocket pgStatSock = PGINVALID_SOCKET;
 150
 151 static struct sockaddr_storage pgStatAddr;
 152
 153 static time_t last_pgstat_start_time;
 154
 155 static bool pgStatRunningInCollector = false;
 156
 157 /*
 158  * Structures in which backends store per-table info that's waiting to be
 159  * sent to the collector.
 160  *
 161  * NOTE: once allocated, TabStatusArray structures are never moved or deleted
 162  * for the life of the backend.  Also, we zero out the t_id fields of the
 163  * contained PgStat_TableStatus structs whenever they are not actively in use.
 164  * This allows relcache pgstat_info pointers to be treated as long-lived data,
 165  * avoiding repeated searches in pgstat_initstats() when a relation is
 166  * repeatedly opened during a transaction.
 167  */
 168 #define TABSTAT_QUANTUM         100 /* we alloc this many at a time */
 169
 170 typedef struct TabStatusArray
 171 {
 172         struct TabStatusArray *tsa_next;        /* link to next array, if any */
 173         int                     tsa_used;               /* # entries currently used */
 174         PgStat_TableStatus tsa_entries[TABSTAT_QUANTUM];        /* per-table data */
 175 } TabStatusArray;
 176
 177 static TabStatusArray *pgStatTabList = NULL;
 178
 179 /*
 180  * pgStatTabHash entry: map from relation OID to PgStat_TableStatus pointer
 181  */
 182 typedef struct TabStatHashEntry
 183 {
 184         Oid                     t_id;
 185         PgStat_TableStatus *tsa_entry;
 186 } TabStatHashEntry;
 187
 188 /*
 189  * Hash table for O(1) t_id -> tsa_entry lookup
 190  */
 191 static HTAB *pgStatTabHash = NULL;
 192
 193 /*
 194  * Backends store per-function info that's waiting to be sent to the collector
 195  * in this hash table (indexed by function OID).
 196  */
 197 static HTAB *pgStatFunctions = NULL;
 198
 199 /*
 200  * Indicates if backend has some function stats that it hasn't yet
 201  * sent to the collector.
 202  */
 203 static bool have_function_stats = false;
 204
 205 /*
 206  * Tuple insertion/deletion counts for an open transaction can't be propagated
 207  * into PgStat_TableStatus counters until we know if it is going to commit
 208  * or abort.  Hence, we keep these counts in per-subxact structs that live
 209  * in TopTransactionContext.  This data structure is designed on the assumption
 210  * that subxacts won't usually modify very many tables.
 211  */
 212 typedef struct PgStat_SubXactStatus
 213 {
 214         int                     nest_level;             /* subtransaction nest level */
 215         struct PgStat_SubXactStatus *prev;      /* higher-level subxact if any */
 216         PgStat_TableXactStatus *first;  /* head of list for this subxact */
 217 } PgStat_SubXactStatus;
 218
 219 static PgStat_SubXactStatus *pgStatXactStack = NULL;
 220
 221 static int      pgStatXactCommit = 0;
 222 static int      pgStatXactRollback = 0;
 223 PgStat_Counter pgStatBlockReadTime = 0;
 224 PgStat_Counter pgStatBlockWriteTime = 0;
 225
 226 /* Record that's written to 2PC state file when pgstat state is persisted */
 227 typedef struct TwoPhasePgStatRecord
 228 {
 229         PgStat_Counter tuples_inserted; /* tuples inserted in xact */
 230         PgStat_Counter tuples_updated;  /* tuples updated in xact */
 231         PgStat_Counter tuples_deleted;  /* tuples deleted in xact */
 232         PgStat_Counter inserted_pre_trunc;      /* tuples inserted prior to truncate */
 233         PgStat_Counter updated_pre_trunc;       /* tuples updated prior to truncate */
 234         PgStat_Counter deleted_pre_trunc;       /* tuples deleted prior to truncate */
 235         Oid                     t_id;                   /* table's OID */
 236         bool            t_shared;               /* is it a shared catalog? */
 237         bool            t_truncated;    /* was the relation truncated? */
 238 } TwoPhasePgStatRecord;
 239
 240 /*
 241  * Info about current "snapshot" of stats file
 242  */
 243 static MemoryContext pgStatLocalContext = NULL;
 244 static HTAB *pgStatDBHash = NULL;
 245
 246 /* Status for backends including auxiliary */
 247 static LocalPgBackendStatus *localBackendStatusTable = NULL;
 248
 249 /* Total number of backends including auxiliary */
 250 static int      localNumBackends = 0;
 251
 252 /*
 253  * Cluster wide statistics, kept in the stats collector.
 254  * Contains statistics that are not collected per database
 255  * or per table.
 256  */
 257 static PgStat_ArchiverStats archiverStats;
 258 static PgStat_GlobalStats globalStats;
 259
 260 /*
 261  * List of OIDs of databases we need to write out.  If an entry is InvalidOid,
 262  * it means to write only the shared-catalog stats ("DB 0"); otherwise, we
 263  * will write both that DB's data and the shared stats.
 264  */
 265 static List *pending_write_requests = NIL;
 266
 267 /* Signal handler flags */
 268 static volatile bool need_exit = false;
 269 static volatile bool got_SIGHUP = false;
 270
 271 /*
 272  * Total time charged to functions so far in the current backend.
 273  * We use this to help separate "self" and "other" time charges.
 274  * (We assume this initializes to zero.)
 275  */
 276 static instr_time total_func_time;
 277
 278
 279 /* ----------
 280  * Local function forward declarations
 281  * ----------
 282  */
 283 #ifdef EXEC_BACKEND
 284 static pid_t pgstat_forkexec(void);
 285 #endif
 286
 287 NON_EXEC_STATIC void PgstatCollectorMain(int argc, char *argv[]) pg_attribute_noreturn();
 288 static void pgstat_exit(SIGNAL_ARGS);
 289 static void pgstat_beshutdown_hook(int code, Datum arg);
 290 static void pgstat_sighup_handler(SIGNAL_ARGS);
 291
 292 static PgStat_StatDBEntry *pgstat_get_db_entry(Oid databaseid, bool create);
 293 static PgStat_StatTabEntry *pgstat_get_tab_entry(PgStat_StatDBEntry *dbentry,
 294                                          Oid tableoid, bool create);
 295 static void pgstat_write_statsfiles(bool permanent, bool allDbs);
 296 static void pgstat_write_db_statsfile(PgStat_StatDBEntry *dbentry, bool permanent);
 297 static HTAB *pgstat_read_statsfiles(Oid onlydb, bool permanent, bool deep);
 298 static void pgstat_read_db_statsfile(Oid databaseid, HTAB *tabhash, HTAB *funchash, bool permanent);
 299 static void backend_read_statsfile(void);
 300 static void pgstat_read_current_status(void);
 301
 302 static bool pgstat_write_statsfile_needed(void);
 303 static bool pgstat_db_requested(Oid databaseid);
 304
 305 static void pgstat_send_tabstat(PgStat_MsgTabstat *tsmsg);
 306 static void pgstat_send_funcstats(void);
 307 static HTAB *pgstat_collect_oids(Oid catalogid, AttrNumber anum_oid);
 308
 309 static PgStat_TableStatus *get_tabstat_entry(Oid rel_id, bool isshared);
 310
 311 static void pgstat_setup_memcxt(void);
 312
 313 static const char *pgstat_get_wait_activity(WaitEventActivity w);
 314 static const char *pgstat_get_wait_client(WaitEventClient w);
 315 static const char *pgstat_get_wait_ipc(WaitEventIPC w);
 316 static const char *pgstat_get_wait_timeout(WaitEventTimeout w);
 317 static const char *pgstat_get_wait_io(WaitEventIO w);
 318
 319 static void pgstat_setheader(PgStat_MsgHdr *hdr, StatMsgType mtype);
 320 static void pgstat_send(void *msg, int len);
 321
 322 static void pgstat_recv_inquiry(PgStat_MsgInquiry *msg, int len);
 323 static void pgstat_recv_tabstat(PgStat_MsgTabstat *msg, int len);
 324 static void pgstat_recv_tabpurge(PgStat_MsgTabpurge *msg, int len);
 325 static void pgstat_recv_dropdb(PgStat_MsgDropdb *msg, int len);
 326 static void pgstat_recv_resetcounter(PgStat_MsgResetcounter *msg, int len);
 327 static void pgstat_recv_resetsharedcounter(PgStat_MsgResetsharedcounter *msg, int len);
 328 static void pgstat_recv_resetsinglecounter(PgStat_MsgResetsinglecounter *msg, int len);
 329 static void pgstat_recv_autovac(PgStat_MsgAutovacStart *msg, int len);
 330 static void pgstat_recv_vacuum(PgStat_MsgVacuum *msg, int len);
 331 static void pgstat_recv_analyze(PgStat_MsgAnalyze *msg, int len);
 332 static void pgstat_recv_archiver(PgStat_MsgArchiver *msg, int len);
 333 static void pgstat_recv_bgwriter(PgStat_MsgBgWriter *msg, int len);
 334 static void pgstat_recv_funcstat(PgStat_MsgFuncstat *msg, int len);
 335 static void pgstat_recv_funcpurge(PgStat_MsgFuncpurge *msg, int len);
 336 static void pgstat_recv_recoveryconflict(PgStat_MsgRecoveryConflict *msg, int len);
 337 static void pgstat_recv_deadlock(PgStat_MsgDeadlock *msg, int len);
 338 static void pgstat_recv_tempfile(PgStat_MsgTempFile *msg, int len);
 339
 340 /* ------------------------------------------------------------
 341  * Public functions called from postmaster follow
 342  * ------------------------------------------------------------
 343  */
 344
 345 /* ----------
 346  * pgstat_init() -
 347  *
 348  *      Called from postmaster at startup. Create the resources required
 349  *      by the statistics collector process.  If unable to do so, do not
 350  *      fail --- better to let the postmaster start with stats collection
 351  *      disabled.
 352  * ----------
 353  */
 354 void
 355 pgstat_init(void)
 356 {
 357         ACCEPT_TYPE_ARG3 alen;
 358         struct addrinfo *addrs = NULL,
 359                            *addr,
 360                                 hints;
 361         int                     ret;
 362         fd_set          rset;
 363         struct timeval tv;
 364         char            test_byte;
 365         int                     sel_res;
 366         int                     tries = 0;
 367
 368 #define TESTBYTEVAL ((char) 199)
 369
 370         /*
 371          * This static assertion verifies that we didn't mess up the calculations
 372          * involved in selecting maximum payload sizes for our UDP messages.
 373          * Because the only consequence of overrunning PGSTAT_MAX_MSG_SIZE would
 374          * be silent performance loss from fragmentation, it seems worth having a
 375          * compile-time cross-check that we didn't.
 376          */
 377         StaticAssertStmt(sizeof(PgStat_Msg) <= PGSTAT_MAX_MSG_SIZE,
 378                                          "maximum stats message size exceeds PGSTAT_MAX_MSG_SIZE");
 379
 380         /*
 381          * Create the UDP socket for sending and receiving statistic messages
 382          */
 383         hints.ai_flags = AI_PASSIVE;
 384         hints.ai_family = AF_UNSPEC;
 385         hints.ai_socktype = SOCK_DGRAM;
 386         hints.ai_protocol = 0;
 387         hints.ai_addrlen = 0;
 388         hints.ai_addr = NULL;
 389         hints.ai_canonname = NULL;
 390         hints.ai_next = NULL;
 391         ret = pg_getaddrinfo_all("localhost", NULL, &hints, &addrs);
 392         if (ret || !addrs)
 393         {
 394                 ereport(LOG,
 395                                 (errmsg("could not resolve \"localhost\": %s",
 396                                                 gai_strerror(ret))));
 397                 goto startup_failed;
 398         }
 399
 400         /*
 401          * On some platforms, pg_getaddrinfo_all() may return multiple addresses
 402          * only one of which will actually work (eg, both IPv6 and IPv4 addresses
 403          * when kernel will reject IPv6).  Worse, the failure may occur at the
 404          * bind() or perhaps even connect() stage.  So we must loop through the
 405          * results till we find a working combination. We will generate LOG
 406          * messages, but no error, for bogus combinations.
 407          */
 408         for (addr = addrs; addr; addr = addr->ai_next)
 409         {
 410 #ifdef HAVE_UNIX_SOCKETS
 411                 /* Ignore AF_UNIX sockets, if any are returned. */
 412                 if (addr->ai_family == AF_UNIX)
 413                         continue;
 414 #endif
 415
 416                 if (++tries > 1)
 417                         ereport(LOG,
 418                                         (errmsg("trying another address for the statistics collector")));
 419
 420                 /*
 421                  * Create the socket.
 422                  */
 423                 if ((pgStatSock = socket(addr->ai_family, SOCK_DGRAM, 0)) == PGINVALID_SOCKET)
 424                 {
 425                         ereport(LOG,
 426                                         (errcode_for_socket_access(),
 427                                          errmsg("could not create socket for statistics collector: %m")));
 428                         continue;
 429                 }
 430
 431                 /*
 432                  * Bind it to a kernel assigned port on localhost and get the assigned
 433                  * port via getsockname().
 434                  */
 435                 if (bind(pgStatSock, addr->ai_addr, addr->ai_addrlen) < 0)
 436                 {
 437                         ereport(LOG,
 438                                         (errcode_for_socket_access(),
 439                                          errmsg("could not bind socket for statistics collector: %m")));
 440                         closesocket(pgStatSock);
 441                         pgStatSock = PGINVALID_SOCKET;
 442                         continue;
 443                 }
 444
 445                 alen = sizeof(pgStatAddr);
 446                 if (getsockname(pgStatSock, (struct sockaddr *) &pgStatAddr, &alen) < 0)
 447                 {
 448                         ereport(LOG,
 449                                         (errcode_for_socket_access(),
 450                                          errmsg("could not get address of socket for statistics collector: %m")));
 451                         closesocket(pgStatSock);
 452                         pgStatSock = PGINVALID_SOCKET;
 453                         continue;
 454                 }
 455
 456                 /*
 457                  * Connect the socket to its own address.  This saves a few cycles by
 458                  * not having to respecify the target address on every send. This also
 459                  * provides a kernel-level check that only packets from this same
 460                  * address will be received.
 461                  */
 462                 if (connect(pgStatSock, (struct sockaddr *) &pgStatAddr, alen) < 0)
 463                 {
 464                         ereport(LOG,
 465                                         (errcode_for_socket_access(),
 466                                          errmsg("could not connect socket for statistics collector: %m")));
 467                         closesocket(pgStatSock);
 468                         pgStatSock = PGINVALID_SOCKET;
 469                         continue;
 470                 }
 471
 472                 /*
 473                  * Try to send and receive a one-byte test message on the socket. This
 474                  * is to catch situations where the socket can be created but will not
 475                  * actually pass data (for instance, because kernel packet filtering
 476                  * rules prevent it).
 477                  */
 478                 test_byte = TESTBYTEVAL;
 479
 480 retry1:
 481                 if (send(pgStatSock, &test_byte, 1, 0) != 1)
 482                 {
 483                         if (errno == EINTR)
 484                                 goto retry1;    /* if interrupted, just retry */
 485                         ereport(LOG,
 486                                         (errcode_for_socket_access(),
 487                                          errmsg("could not send test message on socket for statistics collector: %m")));
 488                         closesocket(pgStatSock);
 489                         pgStatSock = PGINVALID_SOCKET;
 490                         continue;
 491                 }
 492
 493                 /*
 494                  * There could possibly be a little delay before the message can be
 495                  * received.  We arbitrarily allow up to half a second before deciding
 496                  * it's broken.
 497                  */
 498                 for (;;)                                /* need a loop to handle EINTR */
 499                 {
 500                         FD_ZERO(&rset);
 501                         FD_SET(pgStatSock, &rset);
 502
 503                         tv.tv_sec = 0;
 504                         tv.tv_usec = 500000;
 505                         sel_res = select(pgStatSock + 1, &rset, NULL, NULL, &tv);
 506                         if (sel_res >= 0 || errno != EINTR)
 507                                 break;
 508                 }
 509                 if (sel_res < 0)
 510                 {
 511                         ereport(LOG,
 512                                         (errcode_for_socket_access(),
 513                                          errmsg("select() failed in statistics collector: %m")));
 514                         closesocket(pgStatSock);
 515                         pgStatSock = PGINVALID_SOCKET;
 516                         continue;
 517                 }
 518                 if (sel_res == 0 || !FD_ISSET(pgStatSock, &rset))
 519                 {
 520                         /*
 521                          * This is the case we actually think is likely, so take pains to
 522                          * give a specific message for it.
 523                          *
 524                          * errno will not be set meaningfully here, so don't use it.
 525                          */
 526                         ereport(LOG,
 527                                         (errcode(ERRCODE_CONNECTION_FAILURE),
 528                                          errmsg("test message did not get through on socket for statistics collector")));
 529                         closesocket(pgStatSock);
 530                         pgStatSock = PGINVALID_SOCKET;
 531                         continue;
 532                 }
 533
 534                 test_byte++;                    /* just make sure variable is changed */
 535
 536 retry2:
 537                 if (recv(pgStatSock, &test_byte, 1, 0) != 1)
 538                 {
 539                         if (errno == EINTR)
 540                                 goto retry2;    /* if interrupted, just retry */
 541                         ereport(LOG,
 542                                         (errcode_for_socket_access(),
 543                                          errmsg("could not receive test message on socket for statistics collector: %m")));
 544                         closesocket(pgStatSock);
 545                         pgStatSock = PGINVALID_SOCKET;
 546                         continue;
 547                 }
 548
 549                 if (test_byte != TESTBYTEVAL)   /* strictly paranoia ... */
 550                 {
 551                         ereport(LOG,
 552                                         (errcode(ERRCODE_INTERNAL_ERROR),
 553                                          errmsg("incorrect test message transmission on socket for statistics collector")));
 554                         closesocket(pgStatSock);
 555                         pgStatSock = PGINVALID_SOCKET;
 556                         continue;
 557                 }
 558
 559                 /* If we get here, we have a working socket */
 560                 break;
 561         }
 562
 563         /* Did we find a working address? */
 564         if (!addr || pgStatSock == PGINVALID_SOCKET)
 565                 goto startup_failed;
 566
 567         /*
 568          * Set the socket to non-blocking IO.  This ensures that if the collector
 569          * falls behind, statistics messages will be discarded; backends won't
 570          * block waiting to send messages to the collector.
 571          */
 572         if (!pg_set_noblock(pgStatSock))
 573         {
 574                 ereport(LOG,
 575                                 (errcode_for_socket_access(),
 576                                  errmsg("could not set statistics collector socket to nonblocking mode: %m")));
 577                 goto startup_failed;
 578         }
 579
 580         /*
 581          * Try to ensure that the socket's receive buffer is at least
 582          * PGSTAT_MIN_RCVBUF bytes, so that it won't easily overflow and lose
 583          * data.  Use of UDP protocol means that we are willing to lose data under
 584          * heavy load, but we don't want it to happen just because of ridiculously
 585          * small default buffer sizes (such as 8KB on older Windows versions).
 586          */
 587         {
 588                 int                     old_rcvbuf;
 589                 int                     new_rcvbuf;
 590                 ACCEPT_TYPE_ARG3 rcvbufsize = sizeof(old_rcvbuf);
 591
 592                 if (getsockopt(pgStatSock, SOL_SOCKET, SO_RCVBUF,
 593                                            (char *) &old_rcvbuf, &rcvbufsize) < 0)
 594                 {
 595                         elog(LOG, "getsockopt(SO_RCVBUF) failed: %m");
 596                         /* if we can't get existing size, always try to set it */
 597                         old_rcvbuf = 0;
 598                 }
 599
 600                 new_rcvbuf = PGSTAT_MIN_RCVBUF;
 601                 if (old_rcvbuf < new_rcvbuf)
 602                 {
 603                         if (setsockopt(pgStatSock, SOL_SOCKET, SO_RCVBUF,
 604                                                    (char *) &new_rcvbuf, sizeof(new_rcvbuf)) < 0)
 605                                 elog(LOG, "setsockopt(SO_RCVBUF) failed: %m");
 606                 }
 607         }
 608
 609         pg_freeaddrinfo_all(hints.ai_family, addrs);
 610
 611         return;
 612
 613 startup_failed:
 614         ereport(LOG,
 615                         (errmsg("disabling statistics collector for lack of working socket")));
 616
 617         if (addrs)
 618                 pg_freeaddrinfo_all(hints.ai_family, addrs);
 619
 620         if (pgStatSock != PGINVALID_SOCKET)
 621                 closesocket(pgStatSock);
 622         pgStatSock = PGINVALID_SOCKET;
 623
 624         /*
 625          * Adjust GUC variables to suppress useless activity, and for debugging
 626          * purposes (seeing track_counts off is a clue that we failed here). We
 627          * use PGC_S_OVERRIDE because there is no point in trying to turn it back
 628          * on from postgresql.conf without a restart.
 629          */
 630         SetConfigOption("track_counts", "off", PGC_INTERNAL, PGC_S_OVERRIDE);
 631 }
 632
 633 /*
 634  * subroutine for pgstat_reset_all
 635  */
 636 static void
 637 pgstat_reset_remove_files(const char *directory)
 638 {
 639         DIR                *dir;
 640         struct dirent *entry;
 641         char            fname[MAXPGPATH * 2];
 642
 643         dir = AllocateDir(directory);
 644         while ((entry = ReadDir(dir, directory)) != NULL)
 645         {
 646                 int                     nchars;
 647                 Oid                     tmp_oid;
 648
 649                 /*
 650                  * Skip directory entries that don't match the file names we write.
 651                  * See get_dbstat_filename for the database-specific pattern.
 652                  */
 653                 if (strncmp(entry->d_name, "global.", 7) == 0)
 654                         nchars = 7;
 655                 else
 656                 {
 657                         nchars = 0;
 658                         (void) sscanf(entry->d_name, "db_%u.%n",
 659                                                   &tmp_oid, &nchars);
 660                         if (nchars <= 0)
 661                                 continue;
 662                         /* %u allows leading whitespace, so reject that */
 663                         if (strchr("0123456789", entry->d_name[3]) == NULL)
 664                                 continue;
 665                 }
 666
 667                 if (strcmp(entry->d_name + nchars, "tmp") != 0 &&
 668                         strcmp(entry->d_name + nchars, "stat") != 0)
 669                         continue;
 670
 671                 snprintf(fname, sizeof(fname), "%s/%s", directory,
 672                                  entry->d_name);
 673                 unlink(fname);
 674         }
 675         FreeDir(dir);
 676 }
 677
 678 /*
 679  * pgstat_reset_all() -
 680  *
 681  * Remove the stats files.  This is currently used only if WAL
 682  * recovery is needed after a crash.
 683  */
 684 void
 685 pgstat_reset_all(void)
 686 {
 687         pgstat_reset_remove_files(pgstat_stat_directory);
 688         pgstat_reset_remove_files(PGSTAT_STAT_PERMANENT_DIRECTORY);
 689 }
 690
 691 #ifdef EXEC_BACKEND
 692
 693 /*
 694  * pgstat_forkexec() -
 695  *
 696  * Format up the arglist for, then fork and exec, statistics collector process
 697  */
 698 static pid_t
 699 pgstat_forkexec(void)
 700 {
 701         char       *av[10];
 702         int                     ac = 0;
 703
 704         av[ac++] = "postgres";
 705         av[ac++] = "--forkcol";
 706         av[ac++] = NULL;                        /* filled in by postmaster_forkexec */
 707
 708         av[ac] = NULL;
 709         Assert(ac < lengthof(av));
 710
 711         return postmaster_forkexec(ac, av);
 712 }
 713 #endif                                                  /* EXEC_BACKEND */
 714
 715
 716 /*
 717  * pgstat_start() -
 718  *
 719  *      Called from postmaster at startup or after an existing collector
 720  *      died.  Attempt to fire up a fresh statistics collector.
 721  *
 722  *      Returns PID of child process, or 0 if fail.
 723  *
 724  *      Note: if fail, we will be called again from the postmaster main loop.
 725  */
 726 int
 727 pgstat_start(void)
 728 {
 729         time_t          curtime;
 730         pid_t           pgStatPid;
 731
 732         /*
 733          * Check that the socket is there, else pgstat_init failed and we can do
 734          * nothing useful.
 735          */
 736         if (pgStatSock == PGINVALID_SOCKET)
 737                 return 0;
 738
 739         /*
 740          * Do nothing if too soon since last collector start.  This is a safety
 741          * valve to protect against continuous respawn attempts if the collector
 742          * is dying immediately at launch.  Note that since we will be re-called
 743          * from the postmaster main loop, we will get another chance later.
 744          */
 745         curtime = time(NULL);
 746         if ((unsigned int) (curtime - last_pgstat_start_time) <
 747                 (unsigned int) PGSTAT_RESTART_INTERVAL)
 748                 return 0;
 749         last_pgstat_start_time = curtime;
 750
 751         /*
 752          * Okay, fork off the collector.
 753          */
 754 #ifdef EXEC_BACKEND
 755         switch ((pgStatPid = pgstat_forkexec()))
 756 #else
 757         switch ((pgStatPid = fork_process()))
 758 #endif
 759         {
 760                 case -1:
 761                         ereport(LOG,
 762                                         (errmsg("could not fork statistics collector: %m")));
 763                         return 0;
 764
 765 #ifndef EXEC_BACKEND
 766                 case 0:
 767                         /* in postmaster child ... */
 768                         InitPostmasterChild();
 769
 770                         /* Close the postmaster's sockets */
 771                         ClosePostmasterPorts(false);
 772
 773                         /* Drop our connection to postmaster's shared memory, as well */
 774                         dsm_detach_all();
 775                         PGSharedMemoryDetach();
 776
 777                         PgstatCollectorMain(0, NULL);
 778                         break;
 779 #endif
 780
 781                 default:
 782                         return (int) pgStatPid;
 783         }
 784
 785         /* shouldn't get here */
 786         return 0;
 787 }
 788
 789 void
 790 allow_immediate_pgstat_restart(void)
 791 {
 792         last_pgstat_start_time = 0;
 793 }
 794
 795 /* ------------------------------------------------------------
 796  * Public functions used by backends follow
 797  *------------------------------------------------------------
 798  */
 799
 800
 801 /* ----------
 802  * pgstat_report_stat() -
 803  *
 804  *      Must be called by processes that performs DML: tcop/postgres.c, logical
 805  *      receiver processes, SPI worker, etc. to send the so far collected
 806  *      per-table and function usage statistics to the collector.  Note that this
 807  *      is called only when not within a transaction, so it is fair to use
 808  *      transaction stop time as an approximation of current time.
 809  * ----------
 810  */
 811 void
 812 pgstat_report_stat(bool force)
 813 {
 814         /* we assume this inits to all zeroes: */
 815         static const PgStat_TableCounts all_zeroes;
 816         static TimestampTz last_report = 0;
 817
 818         TimestampTz now;
 819         PgStat_MsgTabstat regular_msg;
 820         PgStat_MsgTabstat shared_msg;
 821         TabStatusArray *tsa;
 822         int                     i;
 823
 824         /* Don't expend a clock check if nothing to do */
 825         if ((pgStatTabList == NULL || pgStatTabList->tsa_used == 0) &&
 826                 pgStatXactCommit == 0 && pgStatXactRollback == 0 &&
 827                 !have_function_stats)
 828                 return;
 829
 830         /*
 831          * Don't send a message unless it's been at least PGSTAT_STAT_INTERVAL
 832          * msec since we last sent one, or the caller wants to force stats out.
 833          */
 834         now = GetCurrentTransactionStopTimestamp();
 835         if (!force &&
 836                 !TimestampDifferenceExceeds(last_report, now, PGSTAT_STAT_INTERVAL))
 837                 return;
 838         last_report = now;
 839
 840         /*
 841          * Destroy pgStatTabHash before we start invalidating PgStat_TableEntry
 842          * entries it points to.  (Should we fail partway through the loop below,
 843          * it's okay to have removed the hashtable already --- the only
 844          * consequence is we'd get multiple entries for the same table in the
 845          * pgStatTabList, and that's safe.)
 846          */
 847         if (pgStatTabHash)
 848                 hash_destroy(pgStatTabHash);
 849         pgStatTabHash = NULL;
 850
 851         /*
 852          * Scan through the TabStatusArray struct(s) to find tables that actually
 853          * have counts, and build messages to send.  We have to separate shared
 854          * relations from regular ones because the databaseid field in the message
 855          * header has to depend on that.
 856          */
 857         regular_msg.m_databaseid = MyDatabaseId;
 858         shared_msg.m_databaseid = InvalidOid;
 859         regular_msg.m_nentries = 0;
 860         shared_msg.m_nentries = 0;
 861
 862         for (tsa = pgStatTabList; tsa != NULL; tsa = tsa->tsa_next)
 863         {
 864                 for (i = 0; i < tsa->tsa_used; i++)
 865                 {
 866                         PgStat_TableStatus *entry = &tsa->tsa_entries[i];
 867                         PgStat_MsgTabstat *this_msg;
 868                         PgStat_TableEntry *this_ent;
 869
 870                         /* Shouldn't have any pending transaction-dependent counts */
 871                         Assert(entry->trans == NULL);
 872
 873                         /*
 874                          * Ignore entries that didn't accumulate any actual counts, such
 875                          * as indexes that were opened by the planner but not used.
 876                          */
 877                         if (memcmp(&entry->t_counts, &all_zeroes,
 878                                            sizeof(PgStat_TableCounts)) == 0)
 879                                 continue;
 880
 881                         /*
 882                          * OK, insert data into the appropriate message, and send if full.
 883                          */
 884                         this_msg = entry->t_shared ? &shared_msg : &regular_msg;
 885                         this_ent = &this_msg->m_entry[this_msg->m_nentries];
 886                         this_ent->t_id = entry->t_id;
 887                         memcpy(&this_ent->t_counts, &entry->t_counts,
 888                                    sizeof(PgStat_TableCounts));
 889                         if (++this_msg->m_nentries >= PGSTAT_NUM_TABENTRIES)
 890                         {
 891                                 pgstat_send_tabstat(this_msg);
 892                                 this_msg->m_nentries = 0;
 893                         }
 894                 }
 895                 /* zero out TableStatus structs after use */
 896                 MemSet(tsa->tsa_entries, 0,
 897                            tsa->tsa_used * sizeof(PgStat_TableStatus));
 898                 tsa->tsa_used = 0;
 899         }
 900
 901         /*
 902          * Send partial messages.  Make sure that any pending xact commit/abort
 903          * gets counted, even if there are no table stats to send.
 904          */
 905         if (regular_msg.m_nentries > 0 ||
 906                 pgStatXactCommit > 0 || pgStatXactRollback > 0)
 907                 pgstat_send_tabstat(&regular_msg);
 908         if (shared_msg.m_nentries > 0)
 909                 pgstat_send_tabstat(&shared_msg);
 910
 911         /* Now, send function statistics */
 912         pgstat_send_funcstats();
 913 }
 914
 915 /*
 916  * Subroutine for pgstat_report_stat: finish and send a tabstat message
 917  */
 918 static void
 919 pgstat_send_tabstat(PgStat_MsgTabstat *tsmsg)
 920 {
 921         int                     n;
 922         int                     len;
 923
 924         /* It's unlikely we'd get here with no socket, but maybe not impossible */
 925         if (pgStatSock == PGINVALID_SOCKET)
 926                 return;
 927
 928         /*
 929          * Report and reset accumulated xact commit/rollback and I/O timings
 930          * whenever we send a normal tabstat message
 931          */
 932         if (OidIsValid(tsmsg->m_databaseid))
 933         {
 934                 tsmsg->m_xact_commit = pgStatXactCommit;
 935                 tsmsg->m_xact_rollback = pgStatXactRollback;
 936                 tsmsg->m_block_read_time = pgStatBlockReadTime;
 937                 tsmsg->m_block_write_time = pgStatBlockWriteTime;
 938                 pgStatXactCommit = 0;
 939                 pgStatXactRollback = 0;
 940                 pgStatBlockReadTime = 0;
 941                 pgStatBlockWriteTime = 0;
 942         }
 943         else
 944         {
 945                 tsmsg->m_xact_commit = 0;
 946                 tsmsg->m_xact_rollback = 0;
 947                 tsmsg->m_block_read_time = 0;
 948                 tsmsg->m_block_write_time = 0;
 949         }
 950
 951         n = tsmsg->m_nentries;
 952         len = offsetof(PgStat_MsgTabstat, m_entry[0]) +
 953                 n * sizeof(PgStat_TableEntry);
 954
 955         pgstat_setheader(&tsmsg->m_hdr, PGSTAT_MTYPE_TABSTAT);
 956         pgstat_send(tsmsg, len);
 957 }
 958
 959 /*
 960  * Subroutine for pgstat_report_stat: populate and send a function stat message
 961  */
 962 static void
 963 pgstat_send_funcstats(void)
 964 {
 965         /* we assume this inits to all zeroes: */
 966         static const PgStat_FunctionCounts all_zeroes;
 967
 968         PgStat_MsgFuncstat msg;
 969         PgStat_BackendFunctionEntry *entry;
 970         HASH_SEQ_STATUS fstat;
 971
 972         if (pgStatFunctions == NULL)
 973                 return;
 974
 975         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_FUNCSTAT);
 976         msg.m_databaseid = MyDatabaseId;
 977         msg.m_nentries = 0;
 978
 979         hash_seq_init(&fstat, pgStatFunctions);
 980         while ((entry = (PgStat_BackendFunctionEntry *) hash_seq_search(&fstat)) != NULL)
 981         {
 982                 PgStat_FunctionEntry *m_ent;
 983
 984                 /* Skip it if no counts accumulated since last time */
 985                 if (memcmp(&entry->f_counts, &all_zeroes,
 986                                    sizeof(PgStat_FunctionCounts)) == 0)
 987                         continue;
 988
 989                 /* need to convert format of time accumulators */
 990                 m_ent = &msg.m_entry[msg.m_nentries];
 991                 m_ent->f_id = entry->f_id;
 992                 m_ent->f_numcalls = entry->f_counts.f_numcalls;
 993                 m_ent->f_total_time = INSTR_TIME_GET_MICROSEC(entry->f_counts.f_total_time);
 994                 m_ent->f_self_time = INSTR_TIME_GET_MICROSEC(entry->f_counts.f_self_time);
 995
 996                 if (++msg.m_nentries >= PGSTAT_NUM_FUNCENTRIES)
 997                 {
 998                         pgstat_send(&msg, offsetof(PgStat_MsgFuncstat, m_entry[0]) +
 999                                                 msg.m_nentries * sizeof(PgStat_FunctionEntry));
1000                         msg.m_nentries = 0;
1001                 }
1002
1003                 /* reset the entry's counts */
1004                 MemSet(&entry->f_counts, 0, sizeof(PgStat_FunctionCounts));
1005         }
1006
1007         if (msg.m_nentries > 0)
1008                 pgstat_send(&msg, offsetof(PgStat_MsgFuncstat, m_entry[0]) +
1009                                         msg.m_nentries * sizeof(PgStat_FunctionEntry));
1010
1011         have_function_stats = false;
1012 }
1013
1014
1015 /* ----------
1016  * pgstat_vacuum_stat() -
1017  *
1018  *      Will tell the collector about objects he can get rid of.
1019  * ----------
1020  */
1021 void
1022 pgstat_vacuum_stat(void)
1023 {
1024         HTAB       *htab;
1025         PgStat_MsgTabpurge msg;
1026         PgStat_MsgFuncpurge f_msg;
1027         HASH_SEQ_STATUS hstat;
1028         PgStat_StatDBEntry *dbentry;
1029         PgStat_StatTabEntry *tabentry;
1030         PgStat_StatFuncEntry *funcentry;
1031         int                     len;
1032
1033         if (pgStatSock == PGINVALID_SOCKET)
1034                 return;
1035
1036         /*
1037          * If not done for this transaction, read the statistics collector stats
1038          * file into some hash tables.
1039          */
1040         backend_read_statsfile();
1041
1042         /*
1043          * Read pg_database and make a list of OIDs of all existing databases
1044          */
1045         htab = pgstat_collect_oids(DatabaseRelationId, Anum_pg_database_oid);
1046
1047         /*
1048          * Search the database hash table for dead databases and tell the
1049          * collector to drop them.
1050          */
1051         hash_seq_init(&hstat, pgStatDBHash);
1052         while ((dbentry = (PgStat_StatDBEntry *) hash_seq_search(&hstat)) != NULL)
1053         {
1054                 Oid                     dbid = dbentry->databaseid;
1055
1056                 CHECK_FOR_INTERRUPTS();
1057
1058                 /* the DB entry for shared tables (with InvalidOid) is never dropped */
1059                 if (OidIsValid(dbid) &&
1060                         hash_search(htab, (void *) &dbid, HASH_FIND, NULL) == NULL)
1061                         pgstat_drop_database(dbid);
1062         }
1063
1064         /* Clean up */
1065         hash_destroy(htab);
1066
1067         /*
1068          * Lookup our own database entry; if not found, nothing more to do.
1069          */
1070         dbentry = (PgStat_StatDBEntry *) hash_search(pgStatDBHash,
1071                                                                                                  (void *) &MyDatabaseId,
1072                                                                                                  HASH_FIND, NULL);
1073         if (dbentry == NULL || dbentry->tables == NULL)
1074                 return;
1075
1076         /*
1077          * Similarly to above, make a list of all known relations in this DB.
1078          */
1079         htab = pgstat_collect_oids(RelationRelationId, Anum_pg_class_oid);
1080
1081         /*
1082          * Initialize our messages table counter to zero
1083          */
1084         msg.m_nentries = 0;
1085
1086         /*
1087          * Check for all tables listed in stats hashtable if they still exist.
1088          */
1089         hash_seq_init(&hstat, dbentry->tables);
1090         while ((tabentry = (PgStat_StatTabEntry *) hash_seq_search(&hstat)) != NULL)
1091         {
1092                 Oid                     tabid = tabentry->tableid;
1093
1094                 CHECK_FOR_INTERRUPTS();
1095
1096                 if (hash_search(htab, (void *) &tabid, HASH_FIND, NULL) != NULL)
1097                         continue;
1098
1099                 /*
1100                  * Not there, so add this table's Oid to the message
1101                  */
1102                 msg.m_tableid[msg.m_nentries++] = tabid;
1103
1104                 /*
1105                  * If the message is full, send it out and reinitialize to empty
1106                  */
1107                 if (msg.m_nentries >= PGSTAT_NUM_TABPURGE)
1108                 {
1109                         len = offsetof(PgStat_MsgTabpurge, m_tableid[0])
1110                                 + msg.m_nentries * sizeof(Oid);
1111
1112                         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_TABPURGE);
1113                         msg.m_databaseid = MyDatabaseId;
1114                         pgstat_send(&msg, len);
1115
1116                         msg.m_nentries = 0;
1117                 }
1118         }
1119
1120         /*
1121          * Send the rest
1122          */
1123         if (msg.m_nentries > 0)
1124         {
1125                 len = offsetof(PgStat_MsgTabpurge, m_tableid[0])
1126                         + msg.m_nentries * sizeof(Oid);
1127
1128                 pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_TABPURGE);
1129                 msg.m_databaseid = MyDatabaseId;
1130                 pgstat_send(&msg, len);
1131         }
1132
1133         /* Clean up */
1134         hash_destroy(htab);
1135
1136         /*
1137          * Now repeat the above steps for functions.  However, we needn't bother
1138          * in the common case where no function stats are being collected.
1139          */
1140         if (dbentry->functions != NULL &&
1141                 hash_get_num_entries(dbentry->functions) > 0)
1142         {
1143                 htab = pgstat_collect_oids(ProcedureRelationId, Anum_pg_proc_oid);
1144
1145                 pgstat_setheader(&f_msg.m_hdr, PGSTAT_MTYPE_FUNCPURGE);
1146                 f_msg.m_databaseid = MyDatabaseId;
1147                 f_msg.m_nentries = 0;
1148
1149                 hash_seq_init(&hstat, dbentry->functions);
1150                 while ((funcentry = (PgStat_StatFuncEntry *) hash_seq_search(&hstat)) != NULL)
1151                 {
1152                         Oid                     funcid = funcentry->functionid;
1153
1154                         CHECK_FOR_INTERRUPTS();
1155
1156                         if (hash_search(htab, (void *) &funcid, HASH_FIND, NULL) != NULL)
1157                                 continue;
1158
1159                         /*
1160                          * Not there, so add this function's Oid to the message
1161                          */
1162                         f_msg.m_functionid[f_msg.m_nentries++] = funcid;
1163
1164                         /*
1165                          * If the message is full, send it out and reinitialize to empty
1166                          */
1167                         if (f_msg.m_nentries >= PGSTAT_NUM_FUNCPURGE)
1168                         {
1169                                 len = offsetof(PgStat_MsgFuncpurge, m_functionid[0])
1170                                         + f_msg.m_nentries * sizeof(Oid);
1171
1172                                 pgstat_send(&f_msg, len);
1173
1174                                 f_msg.m_nentries = 0;
1175                         }
1176                 }
1177
1178                 /*
1179                  * Send the rest
1180                  */
1181                 if (f_msg.m_nentries > 0)
1182                 {
1183                         len = offsetof(PgStat_MsgFuncpurge, m_functionid[0])
1184                                 + f_msg.m_nentries * sizeof(Oid);
1185
1186                         pgstat_send(&f_msg, len);
1187                 }
1188
1189                 hash_destroy(htab);
1190         }
1191 }
1192
1193
1194 /* ----------
1195  * pgstat_collect_oids() -
1196  *
1197  *      Collect the OIDs of all objects listed in the specified system catalog
1198  *      into a temporary hash table.  Caller should hash_destroy the result
1199  *      when done with it.  (However, we make the table in CurrentMemoryContext
1200  *      so that it will be freed properly in event of an error.)
1201  * ----------
1202  */
1203 static HTAB *
1204 pgstat_collect_oids(Oid catalogid, AttrNumber anum_oid)
1205 {
1206         HTAB       *htab;
1207         HASHCTL         hash_ctl;
1208         Relation        rel;
1209         HeapScanDesc scan;
1210         HeapTuple       tup;
1211         Snapshot        snapshot;
1212
1213         memset(&hash_ctl, 0, sizeof(hash_ctl));
1214         hash_ctl.keysize = sizeof(Oid);
1215         hash_ctl.entrysize = sizeof(Oid);
1216         hash_ctl.hcxt = CurrentMemoryContext;
1217         htab = hash_create("Temporary table of OIDs",
1218                                            PGSTAT_TAB_HASH_SIZE,
1219                                            &hash_ctl,
1220                                            HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
1221
1222         rel = table_open(catalogid, AccessShareLock);
1223         snapshot = RegisterSnapshot(GetLatestSnapshot());
1224         scan = heap_beginscan(rel, snapshot, 0, NULL);
1225         while ((tup = heap_getnext(scan, ForwardScanDirection)) != NULL)
1226         {
1227                 Oid                     thisoid;
1228                 bool            isnull;
1229
1230                 thisoid = heap_getattr(tup, anum_oid, RelationGetDescr(rel), &isnull);
1231                 Assert(!isnull);
1232
1233                 CHECK_FOR_INTERRUPTS();
1234
1235                 (void) hash_search(htab, (void *) &thisoid, HASH_ENTER, NULL);
1236         }
1237         heap_endscan(scan);
1238         UnregisterSnapshot(snapshot);
1239         table_close(rel, AccessShareLock);
1240
1241         return htab;
1242 }
1243
1244
1245 /* ----------
1246  * pgstat_drop_database() -
1247  *
1248  *      Tell the collector that we just dropped a database.
1249  *      (If the message gets lost, we will still clean the dead DB eventually
1250  *      via future invocations of pgstat_vacuum_stat().)
1251  * ----------
1252  */
1253 void
1254 pgstat_drop_database(Oid databaseid)
1255 {
1256         PgStat_MsgDropdb msg;
1257
1258         if (pgStatSock == PGINVALID_SOCKET)
1259                 return;
1260
1261         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_DROPDB);
1262         msg.m_databaseid = databaseid;
1263         pgstat_send(&msg, sizeof(msg));
1264 }
1265
1266
1267 /* ----------
1268  * pgstat_drop_relation() -
1269  *
1270  *      Tell the collector that we just dropped a relation.
1271  *      (If the message gets lost, we will still clean the dead entry eventually
1272  *      via future invocations of pgstat_vacuum_stat().)
1273  *
1274  *      Currently not used for lack of any good place to call it; we rely
1275  *      entirely on pgstat_vacuum_stat() to clean out stats for dead rels.
1276  * ----------
1277  */
1278 #ifdef NOT_USED
1279 void
1280 pgstat_drop_relation(Oid relid)
1281 {
1282         PgStat_MsgTabpurge msg;
1283         int                     len;
1284
1285         if (pgStatSock == PGINVALID_SOCKET)
1286                 return;
1287
1288         msg.m_tableid[0] = relid;
1289         msg.m_nentries = 1;
1290
1291         len = offsetof(PgStat_MsgTabpurge, m_tableid[0]) + sizeof(Oid);
1292
1293         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_TABPURGE);
1294         msg.m_databaseid = MyDatabaseId;
1295         pgstat_send(&msg, len);
1296 }
1297 #endif                                                  /* NOT_USED */
1298
1299
1300 /* ----------
1301  * pgstat_reset_counters() -
1302  *
1303  *      Tell the statistics collector to reset counters for our database.
1304  *
1305  *      Permission checking for this function is managed through the normal
1306  *      GRANT system.
1307  * ----------
1308  */
1309 void
1310 pgstat_reset_counters(void)
1311 {
1312         PgStat_MsgResetcounter msg;
1313
1314         if (pgStatSock == PGINVALID_SOCKET)
1315                 return;
1316
1317         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_RESETCOUNTER);
1318         msg.m_databaseid = MyDatabaseId;
1319         pgstat_send(&msg, sizeof(msg));
1320 }
1321
1322 /* ----------
1323  * pgstat_reset_shared_counters() -
1324  *
1325  *      Tell the statistics collector to reset cluster-wide shared counters.
1326  *
1327  *      Permission checking for this function is managed through the normal
1328  *      GRANT system.
1329  * ----------
1330  */
1331 void
1332 pgstat_reset_shared_counters(const char *target)
1333 {
1334         PgStat_MsgResetsharedcounter msg;
1335
1336         if (pgStatSock == PGINVALID_SOCKET)
1337                 return;
1338
1339         if (strcmp(target, "archiver") == 0)
1340                 msg.m_resettarget = RESET_ARCHIVER;
1341         else if (strcmp(target, "bgwriter") == 0)
1342                 msg.m_resettarget = RESET_BGWRITER;
1343         else
1344                 ereport(ERROR,
1345                                 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1346                                  errmsg("unrecognized reset target: \"%s\"", target),
1347                                  errhint("Target must be \"archiver\" or \"bgwriter\".")));
1348
1349         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_RESETSHAREDCOUNTER);
1350         pgstat_send(&msg, sizeof(msg));
1351 }
1352
1353 /* ----------
1354  * pgstat_reset_single_counter() -
1355  *
1356  *      Tell the statistics collector to reset a single counter.
1357  *
1358  *      Permission checking for this function is managed through the normal
1359  *      GRANT system.
1360  * ----------
1361  */
1362 void
1363 pgstat_reset_single_counter(Oid objoid, PgStat_Single_Reset_Type type)
1364 {
1365         PgStat_MsgResetsinglecounter msg;
1366
1367         if (pgStatSock == PGINVALID_SOCKET)
1368                 return;
1369
1370         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_RESETSINGLECOUNTER);
1371         msg.m_databaseid = MyDatabaseId;
1372         msg.m_resettype = type;
1373         msg.m_objectid = objoid;
1374
1375         pgstat_send(&msg, sizeof(msg));
1376 }
1377
1378 /* ----------
1379  * pgstat_report_autovac() -
1380  *
1381  *      Called from autovacuum.c to report startup of an autovacuum process.
1382  *      We are called before InitPostgres is done, so can't rely on MyDatabaseId;
1383  *      the db OID must be passed in, instead.
1384  * ----------
1385  */
1386 void
1387 pgstat_report_autovac(Oid dboid)
1388 {
1389         PgStat_MsgAutovacStart msg;
1390
1391         if (pgStatSock == PGINVALID_SOCKET)
1392                 return;
1393
1394         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_AUTOVAC_START);
1395         msg.m_databaseid = dboid;
1396         msg.m_start_time = GetCurrentTimestamp();
1397
1398         pgstat_send(&msg, sizeof(msg));
1399 }
1400
1401
1402 /* ---------
1403  * pgstat_report_vacuum() -
1404  *
1405  *      Tell the collector about the table we just vacuumed.
1406  * ---------
1407  */
1408 void
1409 pgstat_report_vacuum(Oid tableoid, bool shared,
1410                                          PgStat_Counter livetuples, PgStat_Counter deadtuples)
1411 {
1412         PgStat_MsgVacuum msg;
1413
1414         if (pgStatSock == PGINVALID_SOCKET || !pgstat_track_counts)
1415                 return;
1416
1417         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_VACUUM);
1418         msg.m_databaseid = shared ? InvalidOid : MyDatabaseId;
1419         msg.m_tableoid = tableoid;
1420         msg.m_autovacuum = IsAutoVacuumWorkerProcess();
1421         msg.m_vacuumtime = GetCurrentTimestamp();
1422         msg.m_live_tuples = livetuples;
1423         msg.m_dead_tuples = deadtuples;
1424         pgstat_send(&msg, sizeof(msg));
1425 }
1426
1427 /* --------
1428  * pgstat_report_analyze() -
1429  *
1430  *      Tell the collector about the table we just analyzed.
1431  *
1432  * Caller must provide new live- and dead-tuples estimates, as well as a
1433  * flag indicating whether to reset the changes_since_analyze counter.
1434  * --------
1435  */
1436 void
1437 pgstat_report_analyze(Relation rel,
1438                                           PgStat_Counter livetuples, PgStat_Counter deadtuples,
1439                                           bool resetcounter)
1440 {
1441         PgStat_MsgAnalyze msg;
1442
1443         if (pgStatSock == PGINVALID_SOCKET || !pgstat_track_counts)
1444                 return;
1445
1446         /*
1447          * Unlike VACUUM, ANALYZE might be running inside a transaction that has
1448          * already inserted and/or deleted rows in the target table. ANALYZE will
1449          * have counted such rows as live or dead respectively. Because we will
1450          * report our counts of such rows at transaction end, we should subtract
1451          * off these counts from what we send to the collector now, else they'll
1452          * be double-counted after commit.  (This approach also ensures that the
1453          * collector ends up with the right numbers if we abort instead of
1454          * committing.)
1455          */
1456         if (rel->pgstat_info != NULL)
1457         {
1458                 PgStat_TableXactStatus *trans;
1459
1460                 for (trans = rel->pgstat_info->trans; trans; trans = trans->upper)
1461                 {
1462                         livetuples -= trans->tuples_inserted - trans->tuples_deleted;
1463                         deadtuples -= trans->tuples_updated + trans->tuples_deleted;
1464                 }
1465                 /* count stuff inserted by already-aborted subxacts, too */
1466                 deadtuples -= rel->pgstat_info->t_counts.t_delta_dead_tuples;
1467                 /* Since ANALYZE's counts are estimates, we could have underflowed */
1468                 livetuples = Max(livetuples, 0);
1469                 deadtuples = Max(deadtuples, 0);
1470         }
1471
1472         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_ANALYZE);
1473         msg.m_databaseid = rel->rd_rel->relisshared ? InvalidOid : MyDatabaseId;
1474         msg.m_tableoid = RelationGetRelid(rel);
1475         msg.m_autovacuum = IsAutoVacuumWorkerProcess();
1476         msg.m_resetcounter = resetcounter;
1477         msg.m_analyzetime = GetCurrentTimestamp();
1478         msg.m_live_tuples = livetuples;
1479         msg.m_dead_tuples = deadtuples;
1480         pgstat_send(&msg, sizeof(msg));
1481 }
1482
1483 /* --------
1484  * pgstat_report_recovery_conflict() -
1485  *
1486  *      Tell the collector about a Hot Standby recovery conflict.
1487  * --------
1488  */
1489 void
1490 pgstat_report_recovery_conflict(int reason)
1491 {
1492         PgStat_MsgRecoveryConflict msg;
1493
1494         if (pgStatSock == PGINVALID_SOCKET || !pgstat_track_counts)
1495                 return;
1496
1497         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_RECOVERYCONFLICT);
1498         msg.m_databaseid = MyDatabaseId;
1499         msg.m_reason = reason;
1500         pgstat_send(&msg, sizeof(msg));
1501 }
1502
1503 /* --------
1504  * pgstat_report_deadlock() -
1505  *
1506  *      Tell the collector about a deadlock detected.
1507  * --------
1508  */
1509 void
1510 pgstat_report_deadlock(void)
1511 {
1512         PgStat_MsgDeadlock msg;
1513
1514         if (pgStatSock == PGINVALID_SOCKET || !pgstat_track_counts)
1515                 return;
1516
1517         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_DEADLOCK);
1518         msg.m_databaseid = MyDatabaseId;
1519         pgstat_send(&msg, sizeof(msg));
1520 }
1521
1522 /* --------
1523  * pgstat_report_tempfile() -
1524  *
1525  *      Tell the collector about a temporary file.
1526  * --------
1527  */
1528 void
1529 pgstat_report_tempfile(size_t filesize)
1530 {
1531         PgStat_MsgTempFile msg;
1532
1533         if (pgStatSock == PGINVALID_SOCKET || !pgstat_track_counts)
1534                 return;
1535
1536         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_TEMPFILE);
1537         msg.m_databaseid = MyDatabaseId;
1538         msg.m_filesize = filesize;
1539         pgstat_send(&msg, sizeof(msg));
1540 }
1541
1542
1543 /* ----------
1544  * pgstat_ping() -
1545  *
1546  *      Send some junk data to the collector to increase traffic.
1547  * ----------
1548  */
1549 void
1550 pgstat_ping(void)
1551 {
1552         PgStat_MsgDummy msg;
1553
1554         if (pgStatSock == PGINVALID_SOCKET)
1555                 return;
1556
1557         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_DUMMY);
1558         pgstat_send(&msg, sizeof(msg));
1559 }
1560
1561 /* ----------
1562  * pgstat_send_inquiry() -
1563  *
1564  *      Notify collector that we need fresh data.
1565  * ----------
1566  */
1567 static void
1568 pgstat_send_inquiry(TimestampTz clock_time, TimestampTz cutoff_time, Oid databaseid)
1569 {
1570         PgStat_MsgInquiry msg;
1571
1572         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_INQUIRY);
1573         msg.clock_time = clock_time;
1574         msg.cutoff_time = cutoff_time;
1575         msg.databaseid = databaseid;
1576         pgstat_send(&msg, sizeof(msg));
1577 }
1578
1579
1580 /*
1581  * Initialize function call usage data.
1582  * Called by the executor before invoking a function.
1583  */
1584 void
1585 pgstat_init_function_usage(FunctionCallInfoData *fcinfo,
1586                                                    PgStat_FunctionCallUsage *fcu)
1587 {
1588         PgStat_BackendFunctionEntry *htabent;
1589         bool            found;
1590
1591         if (pgstat_track_functions <= fcinfo->flinfo->fn_stats)
1592         {
1593                 /* stats not wanted */
1594                 fcu->fs = NULL;
1595                 return;
1596         }
1597
1598         if (!pgStatFunctions)
1599         {
1600                 /* First time through - initialize function stat table */
1601                 HASHCTL         hash_ctl;
1602
1603                 memset(&hash_ctl, 0, sizeof(hash_ctl));
1604                 hash_ctl.keysize = sizeof(Oid);
1605                 hash_ctl.entrysize = sizeof(PgStat_BackendFunctionEntry);
1606                 pgStatFunctions = hash_create("Function stat entries",
1607                                                                           PGSTAT_FUNCTION_HASH_SIZE,
1608                                                                           &hash_ctl,
1609                                                                           HASH_ELEM | HASH_BLOBS);
1610         }
1611
1612         /* Get the stats entry for this function, create if necessary */
1613         htabent = hash_search(pgStatFunctions, &fcinfo->flinfo->fn_oid,
1614                                                   HASH_ENTER, &found);
1615         if (!found)
1616                 MemSet(&htabent->f_counts, 0, sizeof(PgStat_FunctionCounts));
1617
1618         fcu->fs = &htabent->f_counts;
1619
1620         /* save stats for this function, later used to compensate for recursion */
1621         fcu->save_f_total_time = htabent->f_counts.f_total_time;
1622
1623         /* save current backend-wide total time */
1624         fcu->save_total = total_func_time;
1625
1626         /* get clock time as of function start */
1627         INSTR_TIME_SET_CURRENT(fcu->f_start);
1628 }
1629
1630 /*
1631  * find_funcstat_entry - find any existing PgStat_BackendFunctionEntry entry
1632  *              for specified function
1633  *
1634  * If no entry, return NULL, don't create a new one
1635  */
1636 PgStat_BackendFunctionEntry *
1637 find_funcstat_entry(Oid func_id)
1638 {
1639         if (pgStatFunctions == NULL)
1640                 return NULL;
1641
1642         return (PgStat_BackendFunctionEntry *) hash_search(pgStatFunctions,
1643                                                                                                            (void *) &func_id,
1644                                                                                                            HASH_FIND, NULL);
1645 }
1646
1647 /*
1648  * Calculate function call usage and update stat counters.
1649  * Called by the executor after invoking a function.
1650  *
1651  * In the case of a set-returning function that runs in value-per-call mode,
1652  * we will see multiple pgstat_init_function_usage/pgstat_end_function_usage
1653  * calls for what the user considers a single call of the function.  The
1654  * finalize flag should be TRUE on the last call.
1655  */
1656 void
1657 pgstat_end_function_usage(PgStat_FunctionCallUsage *fcu, bool finalize)
1658 {
1659         PgStat_FunctionCounts *fs = fcu->fs;
1660         instr_time      f_total;
1661         instr_time      f_others;
1662         instr_time      f_self;
1663
1664         /* stats not wanted? */
1665         if (fs == NULL)
1666                 return;
1667
1668         /* total elapsed time in this function call */
1669         INSTR_TIME_SET_CURRENT(f_total);
1670         INSTR_TIME_SUBTRACT(f_total, fcu->f_start);
1671
1672         /* self usage: elapsed minus anything already charged to other calls */
1673         f_others = total_func_time;
1674         INSTR_TIME_SUBTRACT(f_others, fcu->save_total);
1675         f_self = f_total;
1676         INSTR_TIME_SUBTRACT(f_self, f_others);
1677
1678         /* update backend-wide total time */
1679         INSTR_TIME_ADD(total_func_time, f_self);
1680
1681         /*
1682          * Compute the new f_total_time as the total elapsed time added to the
1683          * pre-call value of f_total_time.  This is necessary to avoid
1684          * double-counting any time taken by recursive calls of myself.  (We do
1685          * not need any similar kluge for self time, since that already excludes
1686          * any recursive calls.)
1687          */
1688         INSTR_TIME_ADD(f_total, fcu->save_f_total_time);
1689
1690         /* update counters in function stats table */
1691         if (finalize)
1692                 fs->f_numcalls++;
1693         fs->f_total_time = f_total;
1694         INSTR_TIME_ADD(fs->f_self_time, f_self);
1695
1696         /* indicate that we have something to send */
1697         have_function_stats = true;
1698 }
1699
1700
1701 /* ----------
1702  * pgstat_initstats() -
1703  *
1704  *      Initialize a relcache entry to count access statistics.
1705  *      Called whenever a relation is opened.
1706  *
1707  *      We assume that a relcache entry's pgstat_info field is zeroed by
1708  *      relcache.c when the relcache entry is made; thereafter it is long-lived
1709  *      data.  We can avoid repeated searches of the TabStatus arrays when the
1710  *      same relation is touched repeatedly within a transaction.
1711  * ----------
1712  */
1713 void
1714 pgstat_initstats(Relation rel)
1715 {
1716         Oid                     rel_id = rel->rd_id;
1717         char            relkind = rel->rd_rel->relkind;
1718
1719         /* We only count stats for things that have storage */
1720         if (!(relkind == RELKIND_RELATION ||
1721                   relkind == RELKIND_MATVIEW ||
1722                   relkind == RELKIND_INDEX ||
1723                   relkind == RELKIND_TOASTVALUE ||
1724                   relkind == RELKIND_SEQUENCE))
1725         {
1726                 rel->pgstat_info = NULL;
1727                 return;
1728         }
1729
1730         if (pgStatSock == PGINVALID_SOCKET || !pgstat_track_counts)
1731         {
1732                 /* We're not counting at all */
1733                 rel->pgstat_info = NULL;
1734                 return;
1735         }
1736
1737         /*
1738          * If we already set up this relation in the current transaction, nothing
1739          * to do.
1740          */
1741         if (rel->pgstat_info != NULL &&
1742                 rel->pgstat_info->t_id == rel_id)
1743                 return;
1744
1745         /* Else find or make the PgStat_TableStatus entry, and update link */
1746         rel->pgstat_info = get_tabstat_entry(rel_id, rel->rd_rel->relisshared);
1747 }
1748
1749 /*
1750  * get_tabstat_entry - find or create a PgStat_TableStatus entry for rel
1751  */
1752 static PgStat_TableStatus *
1753 get_tabstat_entry(Oid rel_id, bool isshared)
1754 {
1755         TabStatHashEntry *hash_entry;
1756         PgStat_TableStatus *entry;
1757         TabStatusArray *tsa;
1758         bool            found;
1759
1760         /*
1761          * Create hash table if we don't have it already.
1762          */
1763         if (pgStatTabHash == NULL)
1764         {
1765                 HASHCTL         ctl;
1766
1767                 memset(&ctl, 0, sizeof(ctl));
1768                 ctl.keysize = sizeof(Oid);
1769                 ctl.entrysize = sizeof(TabStatHashEntry);
1770
1771                 pgStatTabHash = hash_create("pgstat TabStatusArray lookup hash table",
1772                                                                         TABSTAT_QUANTUM,
1773                                                                         &ctl,
1774                                                                         HASH_ELEM | HASH_BLOBS);
1775         }
1776
1777         /*
1778          * Find an entry or create a new one.
1779          */
1780         hash_entry = hash_search(pgStatTabHash, &rel_id, HASH_ENTER, &found);
1781         if (!found)
1782         {
1783                 /* initialize new entry with null pointer */
1784                 hash_entry->tsa_entry = NULL;
1785         }
1786
1787         /*
1788          * If entry is already valid, we're done.
1789          */
1790         if (hash_entry->tsa_entry)
1791                 return hash_entry->tsa_entry;
1792
1793         /*
1794          * Locate the first pgStatTabList entry with free space, making a new list
1795          * entry if needed.  Note that we could get an OOM failure here, but if so
1796          * we have left the hashtable and the list in a consistent state.
1797          */
1798         if (pgStatTabList == NULL)
1799         {
1800                 /* Set up first pgStatTabList entry */
1801                 pgStatTabList = (TabStatusArray *)
1802                         MemoryContextAllocZero(TopMemoryContext,
1803                                                                    sizeof(TabStatusArray));
1804         }
1805
1806         tsa = pgStatTabList;
1807         while (tsa->tsa_used >= TABSTAT_QUANTUM)
1808         {
1809                 if (tsa->tsa_next == NULL)
1810                         tsa->tsa_next = (TabStatusArray *)
1811                                 MemoryContextAllocZero(TopMemoryContext,
1812                                                                            sizeof(TabStatusArray));
1813                 tsa = tsa->tsa_next;
1814         }
1815
1816         /*
1817          * Allocate a PgStat_TableStatus entry within this list entry.  We assume
1818          * the entry was already zeroed, either at creation or after last use.
1819          */
1820         entry = &tsa->tsa_entries[tsa->tsa_used++];
1821         entry->t_id = rel_id;
1822         entry->t_shared = isshared;
1823
1824         /*
1825          * Now we can fill the entry in pgStatTabHash.
1826          */
1827         hash_entry->tsa_entry = entry;
1828
1829         return entry;
1830 }
1831
1832 /*
1833  * find_tabstat_entry - find any existing PgStat_TableStatus entry for rel
1834  *
1835  * If no entry, return NULL, don't create a new one
1836  *
1837  * Note: if we got an error in the most recent execution of pgstat_report_stat,
1838  * it's possible that an entry exists but there's no hashtable entry for it.
1839  * That's okay, we'll treat this case as "doesn't exist".
1840  */
1841 PgStat_TableStatus *
1842 find_tabstat_entry(Oid rel_id)
1843 {
1844         TabStatHashEntry *hash_entry;
1845
1846         /* If hashtable doesn't exist, there are no entries at all */
1847         if (!pgStatTabHash)
1848                 return NULL;
1849
1850         hash_entry = hash_search(pgStatTabHash, &rel_id, HASH_FIND, NULL);
1851         if (!hash_entry)
1852                 return NULL;
1853
1854         /* Note that this step could also return NULL, but that's correct */
1855         return hash_entry->tsa_entry;
1856 }
1857
1858 /*
1859  * get_tabstat_stack_level - add a new (sub)transaction stack entry if needed
1860  */
1861 static PgStat_SubXactStatus *
1862 get_tabstat_stack_level(int nest_level)
1863 {
1864         PgStat_SubXactStatus *xact_state;
1865
1866         xact_state = pgStatXactStack;
1867         if (xact_state == NULL || xact_state->nest_level != nest_level)
1868         {
1869                 xact_state = (PgStat_SubXactStatus *)
1870                         MemoryContextAlloc(TopTransactionContext,
1871                                                            sizeof(PgStat_SubXactStatus));
1872                 xact_state->nest_level = nest_level;
1873                 xact_state->prev = pgStatXactStack;
1874                 xact_state->first = NULL;
1875                 pgStatXactStack = xact_state;
1876         }
1877         return xact_state;
1878 }
1879
1880 /*
1881  * add_tabstat_xact_level - add a new (sub)transaction state record
1882  */
1883 static void
1884 add_tabstat_xact_level(PgStat_TableStatus *pgstat_info, int nest_level)
1885 {
1886         PgStat_SubXactStatus *xact_state;
1887         PgStat_TableXactStatus *trans;
1888
1889         /*
1890          * If this is the first rel to be modified at the current nest level, we
1891          * first have to push a transaction stack entry.
1892          */
1893         xact_state = get_tabstat_stack_level(nest_level);
1894
1895         /* Now make a per-table stack entry */
1896         trans = (PgStat_TableXactStatus *)
1897                 MemoryContextAllocZero(TopTransactionContext,
1898                                                            sizeof(PgStat_TableXactStatus));
1899         trans->nest_level = nest_level;
1900         trans->upper = pgstat_info->trans;
1901         trans->parent = pgstat_info;
1902         trans->next = xact_state->first;
1903         xact_state->first = trans;
1904         pgstat_info->trans = trans;
1905 }
1906
1907 /*
1908  * pgstat_count_heap_insert - count a tuple insertion of n tuples
1909  */
1910 void
1911 pgstat_count_heap_insert(Relation rel, PgStat_Counter n)
1912 {
1913         PgStat_TableStatus *pgstat_info = rel->pgstat_info;
1914
1915         if (pgstat_info != NULL)
1916         {
1917                 /* We have to log the effect at the proper transactional level */
1918                 int                     nest_level = GetCurrentTransactionNestLevel();
1919
1920                 if (pgstat_info->trans == NULL ||
1921                         pgstat_info->trans->nest_level != nest_level)
1922                         add_tabstat_xact_level(pgstat_info, nest_level);
1923
1924                 pgstat_info->trans->tuples_inserted += n;
1925         }
1926 }
1927
1928 /*
1929  * pgstat_count_heap_update - count a tuple update
1930  */
1931 void
1932 pgstat_count_heap_update(Relation rel, bool hot)
1933 {
1934         PgStat_TableStatus *pgstat_info = rel->pgstat_info;
1935
1936         if (pgstat_info != NULL)
1937         {
1938                 /* We have to log the effect at the proper transactional level */
1939                 int                     nest_level = GetCurrentTransactionNestLevel();
1940
1941                 if (pgstat_info->trans == NULL ||
1942                         pgstat_info->trans->nest_level != nest_level)
1943                         add_tabstat_xact_level(pgstat_info, nest_level);
1944
1945                 pgstat_info->trans->tuples_updated++;
1946
1947                 /* t_tuples_hot_updated is nontransactional, so just advance it */
1948                 if (hot)
1949                         pgstat_info->t_counts.t_tuples_hot_updated++;
1950         }
1951 }
1952
1953 /*
1954  * pgstat_count_heap_delete - count a tuple deletion
1955  */
1956 void
1957 pgstat_count_heap_delete(Relation rel)
1958 {
1959         PgStat_TableStatus *pgstat_info = rel->pgstat_info;
1960
1961         if (pgstat_info != NULL)
1962         {
1963                 /* We have to log the effect at the proper transactional level */
1964                 int                     nest_level = GetCurrentTransactionNestLevel();
1965
1966                 if (pgstat_info->trans == NULL ||
1967                         pgstat_info->trans->nest_level != nest_level)
1968                         add_tabstat_xact_level(pgstat_info, nest_level);
1969
1970                 pgstat_info->trans->tuples_deleted++;
1971         }
1972 }
1973
1974 /*
1975  * pgstat_truncate_save_counters
1976  *
1977  * Whenever a table is truncated, we save its i/u/d counters so that they can
1978  * be cleared, and if the (sub)xact that executed the truncate later aborts,
1979  * the counters can be restored to the saved (pre-truncate) values.  Note we do
1980  * this on the first truncate in any particular subxact level only.
1981  */
1982 static void
1983 pgstat_truncate_save_counters(PgStat_TableXactStatus *trans)
1984 {
1985         if (!trans->truncated)
1986         {
1987                 trans->inserted_pre_trunc = trans->tuples_inserted;
1988                 trans->updated_pre_trunc = trans->tuples_updated;
1989                 trans->deleted_pre_trunc = trans->tuples_deleted;
1990                 trans->truncated = true;
1991         }
1992 }
1993
1994 /*
1995  * pgstat_truncate_restore_counters - restore counters when a truncate aborts
1996  */
1997 static void
1998 pgstat_truncate_restore_counters(PgStat_TableXactStatus *trans)
1999 {
2000         if (trans->truncated)
2001         {
2002                 trans->tuples_inserted = trans->inserted_pre_trunc;
2003                 trans->tuples_updated = trans->updated_pre_trunc;
2004                 trans->tuples_deleted = trans->deleted_pre_trunc;
2005         }
2006 }
2007
2008 /*
2009  * pgstat_count_truncate - update tuple counters due to truncate
2010  */
2011 void
2012 pgstat_count_truncate(Relation rel)
2013 {
2014         PgStat_TableStatus *pgstat_info = rel->pgstat_info;
2015
2016         if (pgstat_info != NULL)
2017         {
2018                 /* We have to log the effect at the proper transactional level */
2019                 int                     nest_level = GetCurrentTransactionNestLevel();
2020
2021                 if (pgstat_info->trans == NULL ||
2022                         pgstat_info->trans->nest_level != nest_level)
2023                         add_tabstat_xact_level(pgstat_info, nest_level);
2024
2025                 pgstat_truncate_save_counters(pgstat_info->trans);
2026                 pgstat_info->trans->tuples_inserted = 0;
2027                 pgstat_info->trans->tuples_updated = 0;
2028                 pgstat_info->trans->tuples_deleted = 0;
2029         }
2030 }
2031
2032 /*
2033  * pgstat_update_heap_dead_tuples - update dead-tuples count
2034  *
2035  * The semantics of this are that we are reporting the nontransactional
2036  * recovery of "delta" dead tuples; so t_delta_dead_tuples decreases
2037  * rather than increasing, and the change goes straight into the per-table
2038  * counter, not into transactional state.
2039  */
2040 void
2041 pgstat_update_heap_dead_tuples(Relation rel, int delta)
2042 {
2043         PgStat_TableStatus *pgstat_info = rel->pgstat_info;
2044
2045         if (pgstat_info != NULL)
2046                 pgstat_info->t_counts.t_delta_dead_tuples -= delta;
2047 }
2048
2049
2050 /* ----------
2051  * AtEOXact_PgStat
2052  *
2053  *      Called from access/transam/xact.c at top-level transaction commit/abort.
2054  * ----------
2055  */
2056 void
2057 AtEOXact_PgStat(bool isCommit)
2058 {
2059         PgStat_SubXactStatus *xact_state;
2060
2061         /*
2062          * Count transaction commit or abort.  (We use counters, not just bools,
2063          * in case the reporting message isn't sent right away.)
2064          */
2065         if (isCommit)
2066                 pgStatXactCommit++;
2067         else
2068                 pgStatXactRollback++;
2069
2070         /*
2071          * Transfer transactional insert/update counts into the base tabstat
2072          * entries.  We don't bother to free any of the transactional state, since
2073          * it's all in TopTransactionContext and will go away anyway.
2074          */
2075         xact_state = pgStatXactStack;
2076         if (xact_state != NULL)
2077         {
2078                 PgStat_TableXactStatus *trans;
2079
2080                 Assert(xact_state->nest_level == 1);
2081                 Assert(xact_state->prev == NULL);
2082                 for (trans = xact_state->first; trans != NULL; trans = trans->next)
2083                 {
2084                         PgStat_TableStatus *tabstat;
2085
2086                         Assert(trans->nest_level == 1);
2087                         Assert(trans->upper == NULL);
2088                         tabstat = trans->parent;
2089                         Assert(tabstat->trans == trans);
2090                         /* restore pre-truncate stats (if any) in case of aborted xact */
2091                         if (!isCommit)
2092                                 pgstat_truncate_restore_counters(trans);
2093                         /* count attempted actions regardless of commit/abort */
2094                         tabstat->t_counts.t_tuples_inserted += trans->tuples_inserted;
2095                         tabstat->t_counts.t_tuples_updated += trans->tuples_updated;
2096                         tabstat->t_counts.t_tuples_deleted += trans->tuples_deleted;
2097                         if (isCommit)
2098                         {
2099                                 tabstat->t_counts.t_truncated = trans->truncated;
2100                                 if (trans->truncated)
2101                                 {
2102                                         /* forget live/dead stats seen by backend thus far */
2103                                         tabstat->t_counts.t_delta_live_tuples = 0;
2104                                         tabstat->t_counts.t_delta_dead_tuples = 0;
2105                                 }
2106                                 /* insert adds a live tuple, delete removes one */
2107                                 tabstat->t_counts.t_delta_live_tuples +=
2108                                         trans->tuples_inserted - trans->tuples_deleted;
2109                                 /* update and delete each create a dead tuple */
2110                                 tabstat->t_counts.t_delta_dead_tuples +=
2111                                         trans->tuples_updated + trans->tuples_deleted;
2112                                 /* insert, update, delete each count as one change event */
2113                                 tabstat->t_counts.t_changed_tuples +=
2114                                         trans->tuples_inserted + trans->tuples_updated +
2115                                         trans->tuples_deleted;
2116                         }
2117                         else
2118                         {
2119                                 /* inserted tuples are dead, deleted tuples are unaffected */
2120                                 tabstat->t_counts.t_delta_dead_tuples +=
2121                                         trans->tuples_inserted + trans->tuples_updated;
2122                                 /* an aborted xact generates no changed_tuple events */
2123                         }
2124                         tabstat->trans = NULL;
2125                 }
2126         }
2127         pgStatXactStack = NULL;
2128
2129         /* Make sure any stats snapshot is thrown away */
2130         pgstat_clear_snapshot();
2131 }
2132
2133 /* ----------
2134  * AtEOSubXact_PgStat
2135  *
2136  *      Called from access/transam/xact.c at subtransaction commit/abort.
2137  * ----------
2138  */
2139 void
2140 AtEOSubXact_PgStat(bool isCommit, int nestDepth)
2141 {
2142         PgStat_SubXactStatus *xact_state;
2143
2144         /*
2145          * Transfer transactional insert/update counts into the next higher
2146          * subtransaction state.
2147          */
2148         xact_state = pgStatXactStack;
2149         if (xact_state != NULL &&
2150                 xact_state->nest_level >= nestDepth)
2151         {
2152                 PgStat_TableXactStatus *trans;
2153                 PgStat_TableXactStatus *next_trans;
2154
2155                 /* delink xact_state from stack immediately to simplify reuse case */
2156                 pgStatXactStack = xact_state->prev;
2157
2158                 for (trans = xact_state->first; trans != NULL; trans = next_trans)
2159                 {
2160                         PgStat_TableStatus *tabstat;
2161
2162                         next_trans = trans->next;
2163                         Assert(trans->nest_level == nestDepth);
2164                         tabstat = trans->parent;
2165                         Assert(tabstat->trans == trans);
2166                         if (isCommit)
2167                         {
2168                                 if (trans->upper && trans->upper->nest_level == nestDepth - 1)
2169                                 {
2170                                         if (trans->truncated)
2171                                         {
2172                                                 /* propagate the truncate status one level up */
2173                                                 pgstat_truncate_save_counters(trans->upper);
2174                                                 /* replace upper xact stats with ours */
2175                                                 trans->upper->tuples_inserted = trans->tuples_inserted;
2176                                                 trans->upper->tuples_updated = trans->tuples_updated;
2177                                                 trans->upper->tuples_deleted = trans->tuples_deleted;
2178                                         }
2179                                         else
2180                                         {
2181                                                 trans->upper->tuples_inserted += trans->tuples_inserted;
2182                                                 trans->upper->tuples_updated += trans->tuples_updated;
2183                                                 trans->upper->tuples_deleted += trans->tuples_deleted;
2184                                         }
2185                                         tabstat->trans = trans->upper;
2186                                         pfree(trans);
2187                                 }
2188                                 else
2189                                 {
2190                                         /*
2191                                          * When there isn't an immediate parent state, we can just
2192                                          * reuse the record instead of going through a
2193                                          * palloc/pfree pushup (this works since it's all in
2194                                          * TopTransactionContext anyway).  We have to re-link it
2195                                          * into the parent level, though, and that might mean
2196                                          * pushing a new entry into the pgStatXactStack.
2197                                          */
2198                                         PgStat_SubXactStatus *upper_xact_state;
2199
2200                                         upper_xact_state = get_tabstat_stack_level(nestDepth - 1);
2201                                         trans->next = upper_xact_state->first;
2202                                         upper_xact_state->first = trans;
2203                                         trans->nest_level = nestDepth - 1;
2204                                 }
2205                         }
2206                         else
2207                         {
2208                                 /*
2209                                  * On abort, update top-level tabstat counts, then forget the
2210                                  * subtransaction
2211                                  */
2212
2213                                 /* first restore values obliterated by truncate */
2214                                 pgstat_truncate_restore_counters(trans);
2215                                 /* count attempted actions regardless of commit/abort */
2216                                 tabstat->t_counts.t_tuples_inserted += trans->tuples_inserted;
2217                                 tabstat->t_counts.t_tuples_updated += trans->tuples_updated;
2218                                 tabstat->t_counts.t_tuples_deleted += trans->tuples_deleted;
2219                                 /* inserted tuples are dead, deleted tuples are unaffected */
2220                                 tabstat->t_counts.t_delta_dead_tuples +=
2221                                         trans->tuples_inserted + trans->tuples_updated;
2222                                 tabstat->trans = trans->upper;
2223                                 pfree(trans);
2224                         }
2225                 }
2226                 pfree(xact_state);
2227         }
2228 }
2229
2230
2231 /*
2232  * AtPrepare_PgStat
2233  *              Save the transactional stats state at 2PC transaction prepare.
2234  *
2235  * In this phase we just generate 2PC records for all the pending
2236  * transaction-dependent stats work.
2237  */
2238 void
2239 AtPrepare_PgStat(void)
2240 {
2241         PgStat_SubXactStatus *xact_state;
2242
2243         xact_state = pgStatXactStack;
2244         if (xact_state != NULL)
2245         {
2246                 PgStat_TableXactStatus *trans;
2247
2248                 Assert(xact_state->nest_level == 1);
2249                 Assert(xact_state->prev == NULL);
2250                 for (trans = xact_state->first; trans != NULL; trans = trans->next)
2251                 {
2252                         PgStat_TableStatus *tabstat;
2253                         TwoPhasePgStatRecord record;
2254
2255                         Assert(trans->nest_level == 1);
2256                         Assert(trans->upper == NULL);
2257                         tabstat = trans->parent;
2258                         Assert(tabstat->trans == trans);
2259
2260                         record.tuples_inserted = trans->tuples_inserted;
2261                         record.tuples_updated = trans->tuples_updated;
2262                         record.tuples_deleted = trans->tuples_deleted;
2263                         record.inserted_pre_trunc = trans->inserted_pre_trunc;
2264                         record.updated_pre_trunc = trans->updated_pre_trunc;
2265                         record.deleted_pre_trunc = trans->deleted_pre_trunc;
2266                         record.t_id = tabstat->t_id;
2267                         record.t_shared = tabstat->t_shared;
2268                         record.t_truncated = trans->truncated;
2269
2270                         RegisterTwoPhaseRecord(TWOPHASE_RM_PGSTAT_ID, 0,
2271                                                                    &record, sizeof(TwoPhasePgStatRecord));
2272                 }
2273         }
2274 }
2275
2276 /*
2277  * PostPrepare_PgStat
2278  *              Clean up after successful PREPARE.
2279  *
2280  * All we need do here is unlink the transaction stats state from the
2281  * nontransactional state.  The nontransactional action counts will be
2282  * reported to the stats collector immediately, while the effects on live
2283  * and dead tuple counts are preserved in the 2PC state file.
2284  *
2285  * Note: AtEOXact_PgStat is not called during PREPARE.
2286  */
2287 void
2288 PostPrepare_PgStat(void)
2289 {
2290         PgStat_SubXactStatus *xact_state;
2291
2292         /*
2293          * We don't bother to free any of the transactional state, since it's all
2294          * in TopTransactionContext and will go away anyway.
2295          */
2296         xact_state = pgStatXactStack;
2297         if (xact_state != NULL)
2298         {
2299                 PgStat_TableXactStatus *trans;
2300
2301                 for (trans = xact_state->first; trans != NULL; trans = trans->next)
2302                 {
2303                         PgStat_TableStatus *tabstat;
2304
2305                         tabstat = trans->parent;
2306                         tabstat->trans = NULL;
2307                 }
2308         }
2309         pgStatXactStack = NULL;
2310
2311         /* Make sure any stats snapshot is thrown away */
2312         pgstat_clear_snapshot();
2313 }
2314
2315 /*
2316  * 2PC processing routine for COMMIT PREPARED case.
2317  *
2318  * Load the saved counts into our local pgstats state.
2319  */
2320 void
2321 pgstat_twophase_postcommit(TransactionId xid, uint16 info,
2322                                                    void *recdata, uint32 len)
2323 {
2324         TwoPhasePgStatRecord *rec = (TwoPhasePgStatRecord *) recdata;
2325         PgStat_TableStatus *pgstat_info;
2326
2327         /* Find or create a tabstat entry for the rel */
2328         pgstat_info = get_tabstat_entry(rec->t_id, rec->t_shared);
2329
2330         /* Same math as in AtEOXact_PgStat, commit case */
2331         pgstat_info->t_counts.t_tuples_inserted += rec->tuples_inserted;
2332         pgstat_info->t_counts.t_tuples_updated += rec->tuples_updated;
2333         pgstat_info->t_counts.t_tuples_deleted += rec->tuples_deleted;
2334         pgstat_info->t_counts.t_truncated = rec->t_truncated;
2335         if (rec->t_truncated)
2336         {
2337                 /* forget live/dead stats seen by backend thus far */
2338                 pgstat_info->t_counts.t_delta_live_tuples = 0;
2339                 pgstat_info->t_counts.t_delta_dead_tuples = 0;
2340         }
2341         pgstat_info->t_counts.t_delta_live_tuples +=
2342                 rec->tuples_inserted - rec->tuples_deleted;
2343         pgstat_info->t_counts.t_delta_dead_tuples +=
2344                 rec->tuples_updated + rec->tuples_deleted;
2345         pgstat_info->t_counts.t_changed_tuples +=
2346                 rec->tuples_inserted + rec->tuples_updated +
2347                 rec->tuples_deleted;
2348 }
2349
2350 /*
2351  * 2PC processing routine for ROLLBACK PREPARED case.
2352  *
2353  * Load the saved counts into our local pgstats state, but treat them
2354  * as aborted.
2355  */
2356 void
2357 pgstat_twophase_postabort(TransactionId xid, uint16 info,
2358                                                   void *recdata, uint32 len)
2359 {
2360         TwoPhasePgStatRecord *rec = (TwoPhasePgStatRecord *) recdata;
2361         PgStat_TableStatus *pgstat_info;
2362
2363         /* Find or create a tabstat entry for the rel */
2364         pgstat_info = get_tabstat_entry(rec->t_id, rec->t_shared);
2365
2366         /* Same math as in AtEOXact_PgStat, abort case */
2367         if (rec->t_truncated)
2368         {
2369                 rec->tuples_inserted = rec->inserted_pre_trunc;
2370                 rec->tuples_updated = rec->updated_pre_trunc;
2371                 rec->tuples_deleted = rec->deleted_pre_trunc;
2372         }
2373         pgstat_info->t_counts.t_tuples_inserted += rec->tuples_inserted;
2374         pgstat_info->t_counts.t_tuples_updated += rec->tuples_updated;
2375         pgstat_info->t_counts.t_tuples_deleted += rec->tuples_deleted;
2376         pgstat_info->t_counts.t_delta_dead_tuples +=
2377                 rec->tuples_inserted + rec->tuples_updated;
2378 }
2379
2380
2381 /* ----------
2382  * pgstat_fetch_stat_dbentry() -
2383  *
2384  *      Support function for the SQL-callable pgstat* functions. Returns
2385  *      the collected statistics for one database or NULL. NULL doesn't mean
2386  *      that the database doesn't exist, it is just not yet known by the
2387  *      collector, so the caller is better off to report ZERO instead.
2388  * ----------
2389  */
2390 PgStat_StatDBEntry *
2391 pgstat_fetch_stat_dbentry(Oid dbid)
2392 {
2393         /*
2394          * If not done for this transaction, read the statistics collector stats
2395          * file into some hash tables.
2396          */
2397         backend_read_statsfile();
2398
2399         /*
2400          * Lookup the requested database; return NULL if not found
2401          */
2402         return (PgStat_StatDBEntry *) hash_search(pgStatDBHash,
2403                                                                                           (void *) &dbid,
2404                                                                                           HASH_FIND, NULL);
2405 }
2406
2407
2408 /* ----------
2409  * pgstat_fetch_stat_tabentry() -
2410  *
2411  *      Support function for the SQL-callable pgstat* functions. Returns
2412  *      the collected statistics for one table or NULL. NULL doesn't mean
2413  *      that the table doesn't exist, it is just not yet known by the
2414  *      collector, so the caller is better off to report ZERO instead.
2415  * ----------
2416  */
2417 PgStat_StatTabEntry *
2418 pgstat_fetch_stat_tabentry(Oid relid)
2419 {
2420         Oid                     dbid;
2421         PgStat_StatDBEntry *dbentry;
2422         PgStat_StatTabEntry *tabentry;
2423
2424         /*
2425          * If not done for this transaction, read the statistics collector stats
2426          * file into some hash tables.
2427          */
2428         backend_read_statsfile();
2429
2430         /*
2431          * Lookup our database, then look in its table hash table.
2432          */
2433         dbid = MyDatabaseId;
2434         dbentry = (PgStat_StatDBEntry *) hash_search(pgStatDBHash,
2435                                                                                                  (void *) &dbid,
2436                                                                                                  HASH_FIND, NULL);
2437         if (dbentry != NULL && dbentry->tables != NULL)
2438         {
2439                 tabentry = (PgStat_StatTabEntry *) hash_search(dbentry->tables,
2440                                                                                                            (void *) &relid,
2441                                                                                                            HASH_FIND, NULL);
2442                 if (tabentry)
2443                         return tabentry;
2444         }
2445
2446         /*
2447          * If we didn't find it, maybe it's a shared table.
2448          */
2449         dbid = InvalidOid;
2450         dbentry = (PgStat_StatDBEntry *) hash_search(pgStatDBHash,
2451                                                                                                  (void *) &dbid,
2452                                                                                                  HASH_FIND, NULL);
2453         if (dbentry != NULL && dbentry->tables != NULL)
2454         {
2455                 tabentry = (PgStat_StatTabEntry *) hash_search(dbentry->tables,
2456                                                                                                            (void *) &relid,
2457                                                                                                            HASH_FIND, NULL);
2458                 if (tabentry)
2459                         return tabentry;
2460         }
2461
2462         return NULL;
2463 }
2464
2465
2466 /* ----------
2467  * pgstat_fetch_stat_funcentry() -
2468  *
2469  *      Support function for the SQL-callable pgstat* functions. Returns
2470  *      the collected statistics for one function or NULL.
2471  * ----------
2472  */
2473 PgStat_StatFuncEntry *
2474 pgstat_fetch_stat_funcentry(Oid func_id)
2475 {
2476         PgStat_StatDBEntry *dbentry;
2477         PgStat_StatFuncEntry *funcentry = NULL;
2478
2479         /* load the stats file if needed */
2480         backend_read_statsfile();
2481
2482         /* Lookup our database, then find the requested function.  */
2483         dbentry = pgstat_fetch_stat_dbentry(MyDatabaseId);
2484         if (dbentry != NULL && dbentry->functions != NULL)
2485         {
2486                 funcentry = (PgStat_StatFuncEntry *) hash_search(dbentry->functions,
2487                                                                                                                  (void *) &func_id,
2488                                                                                                                  HASH_FIND, NULL);
2489         }
2490
2491         return funcentry;
2492 }
2493
2494
2495 /* ----------
2496  * pgstat_fetch_stat_beentry() -
2497  *
2498  *      Support function for the SQL-callable pgstat* functions. Returns
2499  *      our local copy of the current-activity entry for one backend.
2500  *
2501  *      NB: caller is responsible for a check if the user is permitted to see
2502  *      this info (especially the querystring).
2503  * ----------
2504  */
2505 PgBackendStatus *
2506 pgstat_fetch_stat_beentry(int beid)
2507 {
2508         pgstat_read_current_status();
2509
2510         if (beid < 1 || beid > localNumBackends)
2511                 return NULL;
2512
2513         return &localBackendStatusTable[beid - 1].backendStatus;
2514 }
2515
2516
2517 /* ----------
2518  * pgstat_fetch_stat_local_beentry() -
2519  *
2520  *      Like pgstat_fetch_stat_beentry() but with locally computed additions (like
2521  *      xid and xmin values of the backend)
2522  *
2523  *      NB: caller is responsible for a check if the user is permitted to see
2524  *      this info (especially the querystring).
2525  * ----------
2526  */
2527 LocalPgBackendStatus *
2528 pgstat_fetch_stat_local_beentry(int beid)
2529 {
2530         pgstat_read_current_status();
2531
2532         if (beid < 1 || beid > localNumBackends)
2533                 return NULL;
2534
2535         return &localBackendStatusTable[beid - 1];
2536 }
2537
2538
2539 /* ----------
2540  * pgstat_fetch_stat_numbackends() -
2541  *
2542  *      Support function for the SQL-callable pgstat* functions. Returns
2543  *      the maximum current backend id.
2544  * ----------
2545  */
2546 int
2547 pgstat_fetch_stat_numbackends(void)
2548 {
2549         pgstat_read_current_status();
2550
2551         return localNumBackends;
2552 }
2553
2554 /*
2555  * ---------
2556  * pgstat_fetch_stat_archiver() -
2557  *
2558  *      Support function for the SQL-callable pgstat* functions. Returns
2559  *      a pointer to the archiver statistics struct.
2560  * ---------
2561  */
2562 PgStat_ArchiverStats *
2563 pgstat_fetch_stat_archiver(void)
2564 {
2565         backend_read_statsfile();
2566
2567         return &archiverStats;
2568 }
2569
2570
2571 /*
2572  * ---------
2573  * pgstat_fetch_global() -
2574  *
2575  *      Support function for the SQL-callable pgstat* functions. Returns
2576  *      a pointer to the global statistics struct.
2577  * ---------
2578  */
2579 PgStat_GlobalStats *
2580 pgstat_fetch_global(void)
2581 {
2582         backend_read_statsfile();
2583
2584         return &globalStats;
2585 }
2586
2587
2588 /* ------------------------------------------------------------
2589  * Functions for management of the shared-memory PgBackendStatus array
2590  * ------------------------------------------------------------
2591  */
2592
2593 static PgBackendStatus *BackendStatusArray = NULL;
2594 static PgBackendStatus *MyBEEntry = NULL;
2595 static char *BackendAppnameBuffer = NULL;
2596 static char *BackendClientHostnameBuffer = NULL;
2597 static char *BackendActivityBuffer = NULL;
2598 static Size BackendActivityBufferSize = 0;
2599 #ifdef USE_SSL
2600 static PgBackendSSLStatus *BackendSslStatusBuffer = NULL;
2601 #endif
2602
2603
2604 /*
2605  * Report shared-memory space needed by CreateSharedBackendStatus.
2606  */
2607 Size
2608 BackendStatusShmemSize(void)
2609 {
2610         Size            size;
2611
2612         /* BackendStatusArray: */
2613         size = mul_size(sizeof(PgBackendStatus), NumBackendStatSlots);
2614         /* BackendAppnameBuffer: */
2615         size = add_size(size,
2616                                         mul_size(NAMEDATALEN, NumBackendStatSlots));
2617         /* BackendClientHostnameBuffer: */
2618         size = add_size(size,
2619                                         mul_size(NAMEDATALEN, NumBackendStatSlots));
2620         /* BackendActivityBuffer: */
2621         size = add_size(size,
2622                                         mul_size(pgstat_track_activity_query_size, NumBackendStatSlots));
2623 #ifdef USE_SSL
2624         /* BackendSslStatusBuffer: */
2625         size = add_size(size,
2626                                         mul_size(sizeof(PgBackendSSLStatus), NumBackendStatSlots));
2627 #endif
2628         return size;
2629 }
2630
2631 /*
2632  * Initialize the shared status array and several string buffers
2633  * during postmaster startup.
2634  */
2635 void
2636 CreateSharedBackendStatus(void)
2637 {
2638         Size            size;
2639         bool            found;
2640         int                     i;
2641         char       *buffer;
2642
2643         /* Create or attach to the shared array */
2644         size = mul_size(sizeof(PgBackendStatus), NumBackendStatSlots);
2645         BackendStatusArray = (PgBackendStatus *)
2646                 ShmemInitStruct("Backend Status Array", size, &found);
2647
2648         if (!found)
2649         {
2650                 /*
2651                  * We're the first - initialize.
2652                  */
2653                 MemSet(BackendStatusArray, 0, size);
2654         }
2655
2656         /* Create or attach to the shared appname buffer */
2657         size = mul_size(NAMEDATALEN, NumBackendStatSlots);
2658         BackendAppnameBuffer = (char *)
2659                 ShmemInitStruct("Backend Application Name Buffer", size, &found);
2660
2661         if (!found)
2662         {
2663                 MemSet(BackendAppnameBuffer, 0, size);
2664
2665                 /* Initialize st_appname pointers. */
2666                 buffer = BackendAppnameBuffer;
2667                 for (i = 0; i < NumBackendStatSlots; i++)
2668                 {
2669                         BackendStatusArray[i].st_appname = buffer;
2670                         buffer += NAMEDATALEN;
2671                 }
2672         }
2673
2674         /* Create or attach to the shared client hostname buffer */
2675         size = mul_size(NAMEDATALEN, NumBackendStatSlots);
2676         BackendClientHostnameBuffer = (char *)
2677                 ShmemInitStruct("Backend Client Host Name Buffer", size, &found);
2678
2679         if (!found)
2680         {
2681                 MemSet(BackendClientHostnameBuffer, 0, size);
2682
2683                 /* Initialize st_clienthostname pointers. */
2684                 buffer = BackendClientHostnameBuffer;
2685                 for (i = 0; i < NumBackendStatSlots; i++)
2686                 {
2687                         BackendStatusArray[i].st_clienthostname = buffer;
2688                         buffer += NAMEDATALEN;
2689                 }
2690         }
2691
2692         /* Create or attach to the shared activity buffer */
2693         BackendActivityBufferSize = mul_size(pgstat_track_activity_query_size,
2694                                                                                  NumBackendStatSlots);
2695         BackendActivityBuffer = (char *)
2696                 ShmemInitStruct("Backend Activity Buffer",
2697                                                 BackendActivityBufferSize,
2698                                                 &found);
2699
2700         if (!found)
2701         {
2702                 MemSet(BackendActivityBuffer, 0, BackendActivityBufferSize);
2703
2704                 /* Initialize st_activity pointers. */
2705                 buffer = BackendActivityBuffer;
2706                 for (i = 0; i < NumBackendStatSlots; i++)
2707                 {
2708                         BackendStatusArray[i].st_activity_raw = buffer;
2709                         buffer += pgstat_track_activity_query_size;
2710                 }
2711         }
2712
2713 #ifdef USE_SSL
2714         /* Create or attach to the shared SSL status buffer */
2715         size = mul_size(sizeof(PgBackendSSLStatus), NumBackendStatSlots);
2716         BackendSslStatusBuffer = (PgBackendSSLStatus *)
2717                 ShmemInitStruct("Backend SSL Status Buffer", size, &found);
2718
2719         if (!found)
2720         {
2721                 PgBackendSSLStatus *ptr;
2722
2723                 MemSet(BackendSslStatusBuffer, 0, size);
2724
2725                 /* Initialize st_sslstatus pointers. */
2726                 ptr = BackendSslStatusBuffer;
2727                 for (i = 0; i < NumBackendStatSlots; i++)
2728                 {
2729                         BackendStatusArray[i].st_sslstatus = ptr;
2730                         ptr++;
2731                 }
2732         }
2733 #endif
2734 }
2735
2736
2737 /* ----------
2738  * pgstat_initialize() -
2739  *
2740  *      Initialize pgstats state, and set up our on-proc-exit hook.
2741  *      Called from InitPostgres and AuxiliaryProcessMain. For auxiliary process,
2742  *      MyBackendId is invalid. Otherwise, MyBackendId must be set,
2743  *      but we must not have started any transaction yet (since the
2744  *      exit hook must run after the last transaction exit).
2745  *      NOTE: MyDatabaseId isn't set yet; so the shutdown hook has to be careful.
2746  * ----------
2747  */
2748 void
2749 pgstat_initialize(void)
2750 {
2751         /* Initialize MyBEEntry */
2752         if (MyBackendId != InvalidBackendId)
2753         {
2754                 Assert(MyBackendId >= 1 && MyBackendId <= MaxBackends);
2755                 MyBEEntry = &BackendStatusArray[MyBackendId - 1];
2756         }
2757         else
2758         {
2759                 /* Must be an auxiliary process */
2760                 Assert(MyAuxProcType != NotAnAuxProcess);
2761
2762                 /*
2763                  * Assign the MyBEEntry for an auxiliary process.  Since it doesn't
2764                  * have a BackendId, the slot is statically allocated based on the
2765                  * auxiliary process type (MyAuxProcType).  Backends use slots indexed
2766                  * in the range from 1 to MaxBackends (inclusive), so we use
2767                  * MaxBackends + AuxBackendType + 1 as the index of the slot for an
2768                  * auxiliary process.
2769                  */
2770                 MyBEEntry = &BackendStatusArray[MaxBackends + MyAuxProcType];
2771         }
2772
2773         /* Set up a process-exit hook to clean up */
2774         on_shmem_exit(pgstat_beshutdown_hook, 0);
2775 }
2776
2777 /* ----------
2778  * pgstat_bestart() -
2779  *
2780  *      Initialize this backend's entry in the PgBackendStatus array.
2781  *      Called from InitPostgres.
2782  *
2783  *      Apart from auxiliary processes, MyBackendId, MyDatabaseId,
2784  *      session userid, and application_name must be set for a
2785  *      backend (hence, this cannot be combined with pgstat_initialize).
2786  * ----------
2787  */
2788 void
2789 pgstat_bestart(void)
2790 {
2791         SockAddr        clientaddr;
2792         volatile PgBackendStatus *beentry;
2793
2794         /*
2795          * To minimize the time spent modifying the PgBackendStatus entry, fetch
2796          * all the needed data first.
2797          */
2798
2799         /*
2800          * We may not have a MyProcPort (eg, if this is the autovacuum process).
2801          * If so, use all-zeroes client address, which is dealt with specially in
2802          * pg_stat_get_backend_client_addr and pg_stat_get_backend_client_port.
2803          */
2804         if (MyProcPort)
2805                 memcpy(&clientaddr, &MyProcPort->raddr, sizeof(clientaddr));
2806         else
2807                 MemSet(&clientaddr, 0, sizeof(clientaddr));
2808
2809         /*
2810          * Initialize my status entry, following the protocol of bumping
2811          * st_changecount before and after; and make sure it's even afterwards. We
2812          * use a volatile pointer here to ensure the compiler doesn't try to get
2813          * cute.
2814          */
2815         beentry = MyBEEntry;
2816
2817         /* pgstats state must be initialized from pgstat_initialize() */
2818         Assert(beentry != NULL);
2819
2820         if (MyBackendId != InvalidBackendId)
2821         {
2822                 if (IsAutoVacuumLauncherProcess())
2823                 {
2824                         /* Autovacuum Launcher */
2825                         beentry->st_backendType = B_AUTOVAC_LAUNCHER;
2826                 }
2827                 else if (IsAutoVacuumWorkerProcess())
2828                 {
2829                         /* Autovacuum Worker */
2830                         beentry->st_backendType = B_AUTOVAC_WORKER;
2831                 }
2832                 else if (am_walsender)
2833                 {
2834                         /* Wal sender */
2835                         beentry->st_backendType = B_WAL_SENDER;
2836                 }
2837                 else if (IsBackgroundWorker)
2838                 {
2839                         /* bgworker */
2840                         beentry->st_backendType = B_BG_WORKER;
2841                 }
2842                 else
2843                 {
2844                         /* client-backend */
2845                         beentry->st_backendType = B_BACKEND;
2846                 }
2847         }
2848         else
2849         {
2850                 /* Must be an auxiliary process */
2851                 Assert(MyAuxProcType != NotAnAuxProcess);
2852                 switch (MyAuxProcType)
2853                 {
2854                         case StartupProcess:
2855                                 beentry->st_backendType = B_STARTUP;
2856                                 break;
2857                         case BgWriterProcess:
2858                                 beentry->st_backendType = B_BG_WRITER;
2859                                 break;
2860                         case CheckpointerProcess:
2861                                 beentry->st_backendType = B_CHECKPOINTER;
2862                                 break;
2863                         case WalWriterProcess:
2864                                 beentry->st_backendType = B_WAL_WRITER;
2865                                 break;
2866                         case WalReceiverProcess:
2867                                 beentry->st_backendType = B_WAL_RECEIVER;
2868                                 break;
2869                         default:
2870                                 elog(FATAL, "unrecognized process type: %d",
2871                                          (int) MyAuxProcType);
2872                                 proc_exit(1);
2873                 }
2874         }
2875
2876         do
2877         {
2878                 pgstat_increment_changecount_before(beentry);
2879         } while ((beentry->st_changecount & 1) == 0);
2880
2881         beentry->st_procpid = MyProcPid;
2882         beentry->st_proc_start_timestamp = MyStartTimestamp;
2883         beentry->st_activity_start_timestamp = 0;
2884         beentry->st_state_start_timestamp = 0;
2885         beentry->st_xact_start_timestamp = 0;
2886         beentry->st_databaseid = MyDatabaseId;
2887
2888         /* We have userid for client-backends, wal-sender and bgworker processes */
2889         if (beentry->st_backendType == B_BACKEND
2890                 || beentry->st_backendType == B_WAL_SENDER
2891                 || beentry->st_backendType == B_BG_WORKER)
2892                 beentry->st_userid = GetSessionUserId();
2893         else
2894                 beentry->st_userid = InvalidOid;
2895
2896         beentry->st_clientaddr = clientaddr;
2897         if (MyProcPort && MyProcPort->remote_hostname)
2898                 strlcpy(beentry->st_clienthostname, MyProcPort->remote_hostname,
2899                                 NAMEDATALEN);
2900         else
2901                 beentry->st_clienthostname[0] = '\0';
2902 #ifdef USE_SSL
2903         if (MyProcPort && MyProcPort->ssl != NULL)
2904         {
2905                 beentry->st_ssl = true;
2906                 beentry->st_sslstatus->ssl_bits = be_tls_get_cipher_bits(MyProcPort);
2907                 beentry->st_sslstatus->ssl_compression = be_tls_get_compression(MyProcPort);
2908                 strlcpy(beentry->st_sslstatus->ssl_version, be_tls_get_version(MyProcPort), NAMEDATALEN);
2909                 strlcpy(beentry->st_sslstatus->ssl_cipher, be_tls_get_cipher(MyProcPort), NAMEDATALEN);
2910                 be_tls_get_peerdn_name(MyProcPort, beentry->st_sslstatus->ssl_clientdn, NAMEDATALEN);
2911         }
2912         else
2913         {
2914                 beentry->st_ssl = false;
2915         }
2916 #else
2917         beentry->st_ssl = false;
2918 #endif
2919         beentry->st_state = STATE_UNDEFINED;
2920         beentry->st_appname[0] = '\0';
2921         beentry->st_activity_raw[0] = '\0';
2922         /* Also make sure the last byte in each string area is always 0 */
2923         beentry->st_clienthostname[NAMEDATALEN - 1] = '\0';
2924         beentry->st_appname[NAMEDATALEN - 1] = '\0';
2925         beentry->st_activity_raw[pgstat_track_activity_query_size - 1] = '\0';
2926         beentry->st_progress_command = PROGRESS_COMMAND_INVALID;
2927         beentry->st_progress_command_target = InvalidOid;
2928
2929         /*
2930          * we don't zero st_progress_param here to save cycles; nobody should
2931          * examine it until st_progress_command has been set to something other
2932          * than PROGRESS_COMMAND_INVALID
2933          */
2934
2935         pgstat_increment_changecount_after(beentry);
2936
2937         /* Update app name to current GUC setting */
2938         if (application_name)
2939                 pgstat_report_appname(application_name);
2940 }
2941
2942 /*
2943  * Shut down a single backend's statistics reporting at process exit.
2944  *
2945  * Flush any remaining statistics counts out to the collector.
2946  * Without this, operations triggered during backend exit (such as
2947  * temp table deletions) won't be counted.
2948  *
2949  * Lastly, clear out our entry in the PgBackendStatus array.
2950  */
2951 static void
2952 pgstat_beshutdown_hook(int code, Datum arg)
2953 {
2954         volatile PgBackendStatus *beentry = MyBEEntry;
2955
2956         /*
2957          * If we got as far as discovering our own database ID, we can report what
2958          * we did to the collector.  Otherwise, we'd be sending an invalid
2959          * database ID, so forget it.  (This means that accesses to pg_database
2960          * during failed backend starts might never get counted.)
2961          */
2962         if (OidIsValid(MyDatabaseId))
2963                 pgstat_report_stat(true);
2964
2965         /*
2966          * Clear my status entry, following the protocol of bumping st_changecount
2967          * before and after.  We use a volatile pointer here to ensure the
2968          * compiler doesn't try to get cute.
2969          */
2970         pgstat_increment_changecount_before(beentry);
2971
2972         beentry->st_procpid = 0;        /* mark invalid */
2973
2974         pgstat_increment_changecount_after(beentry);
2975 }
2976
2977
2978 /* ----------
2979  * pgstat_report_activity() -
2980  *
2981  *      Called from tcop/postgres.c to report what the backend is actually doing
2982  *      (but note cmd_str can be NULL for certain cases).
2983  *
2984  * All updates of the status entry follow the protocol of bumping
2985  * st_changecount before and after.  We use a volatile pointer here to
2986  * ensure the compiler doesn't try to get cute.
2987  * ----------
2988  */
2989 void
2990 pgstat_report_activity(BackendState state, const char *cmd_str)
2991 {
2992         volatile PgBackendStatus *beentry = MyBEEntry;
2993         TimestampTz start_timestamp;
2994         TimestampTz current_timestamp;
2995         int                     len = 0;
2996
2997         TRACE_POSTGRESQL_STATEMENT_STATUS(cmd_str);
2998
2999         if (!beentry)
3000                 return;
3001
3002         if (!pgstat_track_activities)
3003         {
3004                 if (beentry->st_state != STATE_DISABLED)
3005                 {
3006                         volatile PGPROC *proc = MyProc;
3007
3008                         /*
3009                          * track_activities is disabled, but we last reported a
3010                          * non-disabled state.  As our final update, change the state and
3011                          * clear fields we will not be updating anymore.
3012                          */
3013                         pgstat_increment_changecount_before(beentry);
3014                         beentry->st_state = STATE_DISABLED;
3015                         beentry->st_state_start_timestamp = 0;
3016                         beentry->st_activity_raw[0] = '\0';
3017                         beentry->st_activity_start_timestamp = 0;
3018                         /* st_xact_start_timestamp and wait_event_info are also disabled */
3019                         beentry->st_xact_start_timestamp = 0;
3020                         proc->wait_event_info = 0;
3021                         pgstat_increment_changecount_after(beentry);
3022                 }
3023                 return;
3024         }
3025
3026         /*
3027          * To minimize the time spent modifying the entry, fetch all the needed
3028          * data first.
3029          */
3030         start_timestamp = GetCurrentStatementStartTimestamp();
3031         if (cmd_str != NULL)
3032         {
3033                 /*
3034                  * Compute length of to-be-stored string unaware of multi-byte
3035                  * characters. For speed reasons that'll get corrected on read, rather
3036                  * than computed every write.
3037                  */
3038                 len = Min(strlen(cmd_str), pgstat_track_activity_query_size - 1);
3039         }
3040         current_timestamp = GetCurrentTimestamp();
3041
3042         /*
3043          * Now update the status entry
3044          */
3045         pgstat_increment_changecount_before(beentry);
3046
3047         beentry->st_state = state;
3048         beentry->st_state_start_timestamp = current_timestamp;
3049
3050         if (cmd_str != NULL)
3051         {
3052                 memcpy((char *) beentry->st_activity_raw, cmd_str, len);
3053                 beentry->st_activity_raw[len] = '\0';
3054                 beentry->st_activity_start_timestamp = start_timestamp;
3055         }
3056
3057         pgstat_increment_changecount_after(beentry);
3058 }
3059
3060 /*-----------
3061  * pgstat_progress_start_command() -
3062  *
3063  * Set st_progress_command (and st_progress_command_target) in own backend
3064  * entry.  Also, zero-initialize st_progress_param array.
3065  *-----------
3066  */
3067 void
3068 pgstat_progress_start_command(ProgressCommandType cmdtype, Oid relid)
3069 {
3070         volatile PgBackendStatus *beentry = MyBEEntry;
3071
3072         if (!beentry || !pgstat_track_activities)
3073                 return;
3074
3075         pgstat_increment_changecount_before(beentry);
3076         beentry->st_progress_command = cmdtype;
3077         beentry->st_progress_command_target = relid;
3078         MemSet(&beentry->st_progress_param, 0, sizeof(beentry->st_progress_param));
3079         pgstat_increment_changecount_after(beentry);
3080 }
3081
3082 /*-----------
3083  * pgstat_progress_update_param() -
3084  *
3085  * Update index'th member in st_progress_param[] of own backend entry.
3086  *-----------
3087  */
3088 void
3089 pgstat_progress_update_param(int index, int64 val)
3090 {
3091         volatile PgBackendStatus *beentry = MyBEEntry;
3092
3093         Assert(index >= 0 && index < PGSTAT_NUM_PROGRESS_PARAM);
3094
3095         if (!beentry || !pgstat_track_activities)
3096                 return;
3097
3098         pgstat_increment_changecount_before(beentry);
3099         beentry->st_progress_param[index] = val;
3100         pgstat_increment_changecount_after(beentry);
3101 }
3102
3103 /*-----------
3104  * pgstat_progress_update_multi_param() -
3105  *
3106  * Update multiple members in st_progress_param[] of own backend entry.
3107  * This is atomic; readers won't see intermediate states.
3108  *-----------
3109  */
3110 void
3111 pgstat_progress_update_multi_param(int nparam, const int *index,
3112                                                                    const int64 *val)
3113 {
3114         volatile PgBackendStatus *beentry = MyBEEntry;
3115         int                     i;
3116
3117         if (!beentry || !pgstat_track_activities || nparam == 0)
3118                 return;
3119
3120         pgstat_increment_changecount_before(beentry);
3121
3122         for (i = 0; i < nparam; ++i)
3123         {
3124                 Assert(index[i] >= 0 && index[i] < PGSTAT_NUM_PROGRESS_PARAM);
3125
3126                 beentry->st_progress_param[index[i]] = val[i];
3127         }
3128
3129         pgstat_increment_changecount_after(beentry);
3130 }
3131
3132 /*-----------
3133  * pgstat_progress_end_command() -
3134  *
3135  * Reset st_progress_command (and st_progress_command_target) in own backend
3136  * entry.  This signals the end of the command.
3137  *-----------
3138  */
3139 void
3140 pgstat_progress_end_command(void)
3141 {
3142         volatile PgBackendStatus *beentry = MyBEEntry;
3143
3144         if (!beentry)
3145                 return;
3146         if (!pgstat_track_activities
3147                 && beentry->st_progress_command == PROGRESS_COMMAND_INVALID)
3148                 return;
3149
3150         pgstat_increment_changecount_before(beentry);
3151         beentry->st_progress_command = PROGRESS_COMMAND_INVALID;
3152         beentry->st_progress_command_target = InvalidOid;
3153         pgstat_increment_changecount_after(beentry);
3154 }
3155
3156 /* ----------
3157  * pgstat_report_appname() -
3158  *
3159  *      Called to update our application name.
3160  * ----------
3161  */
3162 void
3163 pgstat_report_appname(const char *appname)
3164 {
3165         volatile PgBackendStatus *beentry = MyBEEntry;
3166         int                     len;
3167
3168         if (!beentry)
3169                 return;
3170
3171         /* This should be unnecessary if GUC did its job, but be safe */
3172         len = pg_mbcliplen(appname, strlen(appname), NAMEDATALEN - 1);
3173
3174         /*
3175          * Update my status entry, following the protocol of bumping
3176          * st_changecount before and after.  We use a volatile pointer here to
3177          * ensure the compiler doesn't try to get cute.
3178          */
3179         pgstat_increment_changecount_before(beentry);
3180
3181         memcpy((char *) beentry->st_appname, appname, len);
3182         beentry->st_appname[len] = '\0';
3183
3184         pgstat_increment_changecount_after(beentry);
3185 }
3186
3187 /*
3188  * Report current transaction start timestamp as the specified value.
3189  * Zero means there is no active transaction.
3190  */
3191 void
3192 pgstat_report_xact_timestamp(TimestampTz tstamp)
3193 {
3194         volatile PgBackendStatus *beentry = MyBEEntry;
3195
3196         if (!pgstat_track_activities || !beentry)
3197                 return;
3198
3199         /*
3200          * Update my status entry, following the protocol of bumping
3201          * st_changecount before and after.  We use a volatile pointer here to
3202          * ensure the compiler doesn't try to get cute.
3203          */
3204         pgstat_increment_changecount_before(beentry);
3205         beentry->st_xact_start_timestamp = tstamp;
3206         pgstat_increment_changecount_after(beentry);
3207 }
3208
3209 /* ----------
3210  * pgstat_read_current_status() -
3211  *
3212  *      Copy the current contents of the PgBackendStatus array to local memory,
3213  *      if not already done in this transaction.
3214  * ----------
3215  */
3216 static void
3217 pgstat_read_current_status(void)
3218 {
3219         volatile PgBackendStatus *beentry;
3220         LocalPgBackendStatus *localtable;
3221         LocalPgBackendStatus *localentry;
3222         char       *localappname,
3223                            *localclienthostname,
3224                            *localactivity;
3225 #ifdef USE_SSL
3226         PgBackendSSLStatus *localsslstatus;
3227 #endif
3228         int                     i;
3229
3230         Assert(!pgStatRunningInCollector);
3231         if (localBackendStatusTable)
3232                 return;                                 /* already done */
3233
3234         pgstat_setup_memcxt();
3235
3236         localtable = (LocalPgBackendStatus *)
3237                 MemoryContextAlloc(pgStatLocalContext,
3238                                                    sizeof(LocalPgBackendStatus) * NumBackendStatSlots);
3239         localappname = (char *)
3240                 MemoryContextAlloc(pgStatLocalContext,
3241                                                    NAMEDATALEN * NumBackendStatSlots);
3242         localclienthostname = (char *)
3243                 MemoryContextAlloc(pgStatLocalContext,
3244                                                    NAMEDATALEN * NumBackendStatSlots);
3245         localactivity = (char *)
3246                 MemoryContextAlloc(pgStatLocalContext,
3247                                                    pgstat_track_activity_query_size * NumBackendStatSlots);
3248 #ifdef USE_SSL
3249         localsslstatus = (PgBackendSSLStatus *)
3250                 MemoryContextAlloc(pgStatLocalContext,
3251                                                    sizeof(PgBackendSSLStatus) * NumBackendStatSlots);
3252 #endif
3253
3254         localNumBackends = 0;
3255
3256         beentry = BackendStatusArray;
3257         localentry = localtable;
3258         for (i = 1; i <= NumBackendStatSlots; i++)
3259         {
3260                 /*
3261                  * Follow the protocol of retrying if st_changecount changes while we
3262                  * copy the entry, or if it's odd.  (The check for odd is needed to
3263                  * cover the case where we are able to completely copy the entry while
3264                  * the source backend is between increment steps.)      We use a volatile
3265                  * pointer here to ensure the compiler doesn't try to get cute.
3266                  */
3267                 for (;;)
3268                 {
3269                         int                     before_changecount;
3270                         int                     after_changecount;
3271
3272                         pgstat_save_changecount_before(beentry, before_changecount);
3273
3274                         localentry->backendStatus.st_procpid = beentry->st_procpid;
3275                         if (localentry->backendStatus.st_procpid > 0)
3276                         {
3277                                 memcpy(&localentry->backendStatus, (char *) beentry, sizeof(PgBackendStatus));
3278
3279                                 /*
3280                                  * strcpy is safe even if the string is modified concurrently,
3281                                  * because there's always a \0 at the end of the buffer.
3282                                  */
3283                                 strcpy(localappname, (char *) beentry->st_appname);
3284                                 localentry->backendStatus.st_appname = localappname;
3285                                 strcpy(localclienthostname, (char *) beentry->st_clienthostname);
3286                                 localentry->backendStatus.st_clienthostname = localclienthostname;
3287                                 strcpy(localactivity, (char *) beentry->st_activity_raw);
3288                                 localentry->backendStatus.st_activity_raw = localactivity;
3289                                 localentry->backendStatus.st_ssl = beentry->st_ssl;
3290 #ifdef USE_SSL
3291                                 if (beentry->st_ssl)
3292                                 {
3293                                         memcpy(localsslstatus, beentry->st_sslstatus, sizeof(PgBackendSSLStatus));
3294                                         localentry->backendStatus.st_sslstatus = localsslstatus;
3295                                 }
3296 #endif
3297                         }
3298
3299                         pgstat_save_changecount_after(beentry, after_changecount);
3300                         if (before_changecount == after_changecount &&
3301                                 (before_changecount & 1) == 0)
3302                                 break;
3303
3304                         /* Make sure we can break out of loop if stuck... */
3305                         CHECK_FOR_INTERRUPTS();
3306                 }
3307
3308                 beentry++;
3309                 /* Only valid entries get included into the local array */
3310                 if (localentry->backendStatus.st_procpid > 0)
3311                 {
3312                         BackendIdGetTransactionIds(i,
3313                                                                            &localentry->backend_xid,
3314                                                                            &localentry->backend_xmin);
3315
3316                         localentry++;
3317                         localappname += NAMEDATALEN;
3318                         localclienthostname += NAMEDATALEN;
3319                         localactivity += pgstat_track_activity_query_size;
3320 #ifdef USE_SSL
3321                         localsslstatus++;
3322 #endif
3323                         localNumBackends++;
3324                 }
3325         }
3326
3327         /* Set the pointer only after completion of a valid table */
3328         localBackendStatusTable = localtable;
3329 }
3330
3331 /* ----------
3332  * pgstat_get_wait_event_type() -
3333  *
3334  *      Return a string representing the current wait event type, backend is
3335  *      waiting on.
3336  */
3337 const char *
3338 pgstat_get_wait_event_type(uint32 wait_event_info)
3339 {
3340         uint32          classId;
3341         const char *event_type;
3342
3343         /* report process as not waiting. */
3344         if (wait_event_info == 0)
3345                 return NULL;
3346
3347         classId = wait_event_info & 0xFF000000;
3348
3349         switch (classId)
3350         {
3351                 case PG_WAIT_LWLOCK:
3352                         event_type = "LWLock";
3353                         break;
3354                 case PG_WAIT_LOCK:
3355                         event_type = "Lock";
3356                         break;
3357                 case PG_WAIT_BUFFER_PIN:
3358                         event_type = "BufferPin";
3359                         break;
3360                 case PG_WAIT_ACTIVITY:
3361                         event_type = "Activity";
3362                         break;
3363                 case PG_WAIT_CLIENT:
3364                         event_type = "Client";
3365                         break;
3366                 case PG_WAIT_EXTENSION:
3367                         event_type = "Extension";
3368                         break;
3369                 case PG_WAIT_IPC:
3370                         event_type = "IPC";
3371                         break;
3372                 case PG_WAIT_TIMEOUT:
3373                         event_type = "Timeout";
3374                         break;
3375                 case PG_WAIT_IO:
3376                         event_type = "IO";
3377                         break;
3378                 default:
3379                         event_type = "???";
3380                         break;
3381         }
3382
3383         return event_type;
3384 }
3385
3386 /* ----------
3387  * pgstat_get_wait_event() -
3388  *
3389  *      Return a string representing the current wait event, backend is
3390  *      waiting on.
3391  */
3392 const char *
3393 pgstat_get_wait_event(uint32 wait_event_info)
3394 {
3395         uint32          classId;
3396         uint16          eventId;
3397         const char *event_name;
3398
3399         /* report process as not waiting. */
3400         if (wait_event_info == 0)
3401                 return NULL;
3402
3403         classId = wait_event_info & 0xFF000000;
3404         eventId = wait_event_info & 0x0000FFFF;
3405
3406         switch (classId)
3407         {
3408                 case PG_WAIT_LWLOCK:
3409                         event_name = GetLWLockIdentifier(classId, eventId);
3410                         break;
3411                 case PG_WAIT_LOCK:
3412                         event_name = GetLockNameFromTagType(eventId);
3413                         break;
3414                 case PG_WAIT_BUFFER_PIN:
3415                         event_name = "BufferPin";
3416                         break;
3417                 case PG_WAIT_ACTIVITY:
3418                         {
3419                                 WaitEventActivity w = (WaitEventActivity) wait_event_info;
3420
3421                                 event_name = pgstat_get_wait_activity(w);
3422                                 break;
3423                         }
3424                 case PG_WAIT_CLIENT:
3425                         {
3426                                 WaitEventClient w = (WaitEventClient) wait_event_info;
3427
3428                                 event_name = pgstat_get_wait_client(w);
3429                                 break;
3430                         }
3431                 case PG_WAIT_EXTENSION:
3432                         event_name = "Extension";
3433                         break;
3434                 case PG_WAIT_IPC:
3435                         {
3436                                 WaitEventIPC w = (WaitEventIPC) wait_event_info;
3437
3438                                 event_name = pgstat_get_wait_ipc(w);
3439                                 break;
3440                         }
3441                 case PG_WAIT_TIMEOUT:
3442                         {
3443                                 WaitEventTimeout w = (WaitEventTimeout) wait_event_info;
3444
3445                                 event_name = pgstat_get_wait_timeout(w);
3446                                 break;
3447                         }
3448                 case PG_WAIT_IO:
3449                         {
3450                                 WaitEventIO w = (WaitEventIO) wait_event_info;
3451
3452                                 event_name = pgstat_get_wait_io(w);
3453                                 break;
3454                         }
3455                 default:
3456                         event_name = "unknown wait event";
3457                         break;
3458         }
3459
3460         return event_name;
3461 }
3462
3463 /* ----------
3464  * pgstat_get_wait_activity() -
3465  *
3466  * Convert WaitEventActivity to string.
3467  * ----------
3468  */
3469 static const char *
3470 pgstat_get_wait_activity(WaitEventActivity w)
3471 {
3472         const char *event_name = "unknown wait event";
3473
3474         switch (w)
3475         {
3476                 case WAIT_EVENT_ARCHIVER_MAIN:
3477                         event_name = "ArchiverMain";
3478                         break;
3479                 case WAIT_EVENT_AUTOVACUUM_MAIN:
3480                         event_name = "AutoVacuumMain";
3481                         break;
3482                 case WAIT_EVENT_BGWRITER_HIBERNATE:
3483                         event_name = "BgWriterHibernate";
3484                         break;
3485                 case WAIT_EVENT_BGWRITER_MAIN:
3486                         event_name = "BgWriterMain";
3487                         break;
3488                 case WAIT_EVENT_CHECKPOINTER_MAIN:
3489                         event_name = "CheckpointerMain";
3490                         break;
3491                 case WAIT_EVENT_LOGICAL_APPLY_MAIN:
3492                         event_name = "LogicalApplyMain";
3493                         break;
3494                 case WAIT_EVENT_LOGICAL_LAUNCHER_MAIN:
3495                         event_name = "LogicalLauncherMain";
3496                         break;
3497                 case WAIT_EVENT_PGSTAT_MAIN:
3498                         event_name = "PgStatMain";
3499                         break;
3500                 case WAIT_EVENT_RECOVERY_WAL_ALL:
3501                         event_name = "RecoveryWalAll";
3502                         break;
3503                 case WAIT_EVENT_RECOVERY_WAL_STREAM:
3504                         event_name = "RecoveryWalStream";
3505                         break;
3506                 case WAIT_EVENT_SYSLOGGER_MAIN:
3507                         event_name = "SysLoggerMain";
3508                         break;
3509                 case WAIT_EVENT_WAL_RECEIVER_MAIN:
3510                         event_name = "WalReceiverMain";
3511                         break;
3512                 case WAIT_EVENT_WAL_SENDER_MAIN:
3513                         event_name = "WalSenderMain";
3514                         break;
3515                 case WAIT_EVENT_WAL_WRITER_MAIN:
3516                         event_name = "WalWriterMain";
3517                         break;
3518                         /* no default case, so that compiler will warn */
3519         }
3520
3521         return event_name;
3522 }
3523
3524 /* ----------
3525  * pgstat_get_wait_client() -
3526  *
3527  * Convert WaitEventClient to string.
3528  * ----------
3529  */
3530 static const char *
3531 pgstat_get_wait_client(WaitEventClient w)
3532 {
3533         const char *event_name = "unknown wait event";
3534
3535         switch (w)
3536         {
3537                 case WAIT_EVENT_CLIENT_READ:
3538                         event_name = "ClientRead";
3539                         break;
3540                 case WAIT_EVENT_CLIENT_WRITE:
3541                         event_name = "ClientWrite";
3542                         break;
3543                 case WAIT_EVENT_LIBPQWALRECEIVER_CONNECT:
3544                         event_name = "LibPQWalReceiverConnect";
3545                         break;
3546                 case WAIT_EVENT_LIBPQWALRECEIVER_RECEIVE:
3547                         event_name = "LibPQWalReceiverReceive";
3548                         break;
3549                 case WAIT_EVENT_SSL_OPEN_SERVER:
3550                         event_name = "SSLOpenServer";
3551                         break;
3552                 case WAIT_EVENT_WAL_RECEIVER_WAIT_START:
3553                         event_name = "WalReceiverWaitStart";
3554                         break;
3555                 case WAIT_EVENT_WAL_SENDER_WAIT_WAL:
3556                         event_name = "WalSenderWaitForWAL";
3557                         break;
3558                 case WAIT_EVENT_WAL_SENDER_WRITE_DATA:
3559                         event_name = "WalSenderWriteData";
3560                         break;
3561                         /* no default case, so that compiler will warn */
3562         }
3563
3564         return event_name;
3565 }
3566
3567 /* ----------
3568  * pgstat_get_wait_ipc() -
3569  *
3570  * Convert WaitEventIPC to string.
3571  * ----------
3572  */
3573 static const char *
3574 pgstat_get_wait_ipc(WaitEventIPC w)
3575 {
3576         const char *event_name = "unknown wait event";
3577
3578         switch (w)
3579         {
3580                 case WAIT_EVENT_BGWORKER_SHUTDOWN:
3581                         event_name = "BgWorkerShutdown";
3582                         break;
3583                 case WAIT_EVENT_BGWORKER_STARTUP:
3584                         event_name = "BgWorkerStartup";
3585                         break;
3586                 case WAIT_EVENT_BTREE_PAGE:
3587                         event_name = "BtreePage";
3588                         break;
3589                 case WAIT_EVENT_CLOG_GROUP_UPDATE:
3590                         event_name = "ClogGroupUpdate";
3591                         break;
3592                 case WAIT_EVENT_EXECUTE_GATHER:
3593                         event_name = "ExecuteGather";
3594                         break;
3595                 case WAIT_EVENT_HASH_BATCH_ALLOCATING:
3596                         event_name = "Hash/Batch/Allocating";
3597                         break;
3598                 case WAIT_EVENT_HASH_BATCH_ELECTING:
3599                         event_name = "Hash/Batch/Electing";
3600                         break;
3601                 case WAIT_EVENT_HASH_BATCH_LOADING:
3602                         event_name = "Hash/Batch/Loading";
3603                         break;
3604                 case WAIT_EVENT_HASH_BUILD_ALLOCATING:
3605                         event_name = "Hash/Build/Allocating";
3606                         break;
3607                 case WAIT_EVENT_HASH_BUILD_ELECTING:
3608                         event_name = "Hash/Build/Electing";
3609                         break;
3610                 case WAIT_EVENT_HASH_BUILD_HASHING_INNER:
3611                         event_name = "Hash/Build/HashingInner";
3612                         break;
3613                 case WAIT_EVENT_HASH_BUILD_HASHING_OUTER:
3614                         event_name = "Hash/Build/HashingOuter";
3615                         break;
3616                 case WAIT_EVENT_HASH_GROW_BATCHES_ALLOCATING:
3617                         event_name = "Hash/GrowBatches/Allocating";
3618                         break;
3619                 case WAIT_EVENT_HASH_GROW_BATCHES_DECIDING:
3620                         event_name = "Hash/GrowBatches/Deciding";
3621                         break;
3622                 case WAIT_EVENT_HASH_GROW_BATCHES_ELECTING:
3623                         event_name = "Hash/GrowBatches/Electing";
3624                         break;
3625                 case WAIT_EVENT_HASH_GROW_BATCHES_FINISHING:
3626                         event_name = "Hash/GrowBatches/Finishing";
3627                         break;
3628                 case WAIT_EVENT_HASH_GROW_BATCHES_REPARTITIONING:
3629                         event_name = "Hash/GrowBatches/Repartitioning";
3630                         break;
3631                 case WAIT_EVENT_HASH_GROW_BUCKETS_ALLOCATING:
3632                         event_name = "Hash/GrowBuckets/Allocating";
3633                         break;
3634                 case WAIT_EVENT_HASH_GROW_BUCKETS_ELECTING:
3635                         event_name = "Hash/GrowBuckets/Electing";
3636                         break;
3637                 case WAIT_EVENT_HASH_GROW_BUCKETS_REINSERTING:
3638                         event_name = "Hash/GrowBuckets/Reinserting";
3639                         break;
3640                 case WAIT_EVENT_LOGICAL_SYNC_DATA:
3641                         event_name = "LogicalSyncData";
3642                         break;
3643                 case WAIT_EVENT_LOGICAL_SYNC_STATE_CHANGE:
3644                         event_name = "LogicalSyncStateChange";
3645                         break;
3646                 case WAIT_EVENT_MQ_INTERNAL:
3647                         event_name = "MessageQueueInternal";
3648                         break;
3649                 case WAIT_EVENT_MQ_PUT_MESSAGE:
3650                         event_name = "MessageQueuePutMessage";
3651                         break;
3652                 case WAIT_EVENT_MQ_RECEIVE:
3653                         event_name = "MessageQueueReceive";
3654                         break;
3655                 case WAIT_EVENT_MQ_SEND:
3656                         event_name = "MessageQueueSend";
3657                         break;
3658                 case WAIT_EVENT_PARALLEL_BITMAP_SCAN:
3659                         event_name = "ParallelBitmapScan";
3660                         break;
3661                 case WAIT_EVENT_PARALLEL_CREATE_INDEX_SCAN:
3662                         event_name = "ParallelCreateIndexScan";
3663                         break;
3664                 case WAIT_EVENT_PARALLEL_FINISH:
3665                         event_name = "ParallelFinish";
3666                         break;
3667                 case WAIT_EVENT_PROCARRAY_GROUP_UPDATE:
3668                         event_name = "ProcArrayGroupUpdate";
3669                         break;
3670                 case WAIT_EVENT_PROMOTE:
3671                         event_name = "Promote";
3672                         break;
3673                 case WAIT_EVENT_REPLICATION_ORIGIN_DROP:
3674                         event_name = "ReplicationOriginDrop";
3675                         break;
3676                 case WAIT_EVENT_REPLICATION_SLOT_DROP:
3677                         event_name = "ReplicationSlotDrop";
3678                         break;
3679                 case WAIT_EVENT_SAFE_SNAPSHOT:
3680                         event_name = "SafeSnapshot";
3681                         break;
3682                 case WAIT_EVENT_SYNC_REP:
3683                         event_name = "SyncRep";
3684                         break;
3685                         /* no default case, so that compiler will warn */
3686         }
3687
3688         return event_name;
3689 }
3690
3691 /* ----------
3692  * pgstat_get_wait_timeout() -
3693  *
3694  * Convert WaitEventTimeout to string.
3695  * ----------
3696  */
3697 static const char *
3698 pgstat_get_wait_timeout(WaitEventTimeout w)
3699 {
3700         const char *event_name = "unknown wait event";
3701
3702         switch (w)
3703         {
3704                 case WAIT_EVENT_BASE_BACKUP_THROTTLE:
3705                         event_name = "BaseBackupThrottle";
3706                         break;
3707                 case WAIT_EVENT_PG_SLEEP:
3708                         event_name = "PgSleep";
3709                         break;
3710                 case WAIT_EVENT_RECOVERY_APPLY_DELAY:
3711                         event_name = "RecoveryApplyDelay";
3712                         break;
3713                         /* no default case, so that compiler will warn */
3714         }
3715
3716         return event_name;
3717 }
3718
3719 /* ----------
3720  * pgstat_get_wait_io() -
3721  *
3722  * Convert WaitEventIO to string.
3723  * ----------
3724  */
3725 static const char *
3726 pgstat_get_wait_io(WaitEventIO w)
3727 {
3728         const char *event_name = "unknown wait event";
3729
3730         switch (w)
3731         {
3732                 case WAIT_EVENT_BUFFILE_READ:
3733                         event_name = "BufFileRead";
3734                         break;
3735                 case WAIT_EVENT_BUFFILE_WRITE:
3736                         event_name = "BufFileWrite";
3737                         break;
3738                 case WAIT_EVENT_CONTROL_FILE_READ:
3739                         event_name = "ControlFileRead";
3740                         break;
3741                 case WAIT_EVENT_CONTROL_FILE_SYNC:
3742                         event_name = "ControlFileSync";
3743                         break;
3744                 case WAIT_EVENT_CONTROL_FILE_SYNC_UPDATE:
3745                         event_name = "ControlFileSyncUpdate";
3746                         break;
3747                 case WAIT_EVENT_CONTROL_FILE_WRITE:
3748                         event_name = "ControlFileWrite";
3749                         break;
3750                 case WAIT_EVENT_CONTROL_FILE_WRITE_UPDATE:
3751                         event_name = "ControlFileWriteUpdate";
3752                         break;
3753                 case WAIT_EVENT_COPY_FILE_READ:
3754                         event_name = "CopyFileRead";
3755                         break;
3756                 case WAIT_EVENT_COPY_FILE_WRITE:
3757                         event_name = "CopyFileWrite";
3758                         break;
3759                 case WAIT_EVENT_DATA_FILE_EXTEND:
3760                         event_name = "DataFileExtend";
3761                         break;
3762                 case WAIT_EVENT_DATA_FILE_FLUSH:
3763                         event_name = "DataFileFlush";
3764                         break;
3765                 case WAIT_EVENT_DATA_FILE_IMMEDIATE_SYNC:
3766                         event_name = "DataFileImmediateSync";
3767                         break;
3768                 case WAIT_EVENT_DATA_FILE_PREFETCH:
3769                         event_name = "DataFilePrefetch";
3770                         break;
3771                 case WAIT_EVENT_DATA_FILE_READ:
3772                         event_name = "DataFileRead";
3773                         break;
3774                 case WAIT_EVENT_DATA_FILE_SYNC:
3775                         event_name = "DataFileSync";
3776                         break;
3777                 case WAIT_EVENT_DATA_FILE_TRUNCATE:
3778                         event_name = "DataFileTruncate";
3779                         break;
3780                 case WAIT_EVENT_DATA_FILE_WRITE:
3781                         event_name = "DataFileWrite";
3782                         break;
3783                 case WAIT_EVENT_DSM_FILL_ZERO_WRITE:
3784                         event_name = "DSMFillZeroWrite";
3785                         break;
3786                 case WAIT_EVENT_LOCK_FILE_ADDTODATADIR_READ:
3787                         event_name = "LockFileAddToDataDirRead";
3788                         break;
3789                 case WAIT_EVENT_LOCK_FILE_ADDTODATADIR_SYNC:
3790                         event_name = "LockFileAddToDataDirSync";
3791                         break;
3792                 case WAIT_EVENT_LOCK_FILE_ADDTODATADIR_WRITE:
3793                         event_name = "LockFileAddToDataDirWrite";
3794                         break;
3795                 case WAIT_EVENT_LOCK_FILE_CREATE_READ:
3796                         event_name = "LockFileCreateRead";
3797                         break;
3798                 case WAIT_EVENT_LOCK_FILE_CREATE_SYNC:
3799                         event_name = "LockFileCreateSync";
3800                         break;
3801                 case WAIT_EVENT_LOCK_FILE_CREATE_WRITE:
3802                         event_name = "LockFileCreateWrite";
3803                         break;
3804                 case WAIT_EVENT_LOCK_FILE_RECHECKDATADIR_READ:
3805                         event_name = "LockFileReCheckDataDirRead";
3806                         break;
3807                 case WAIT_EVENT_LOGICAL_REWRITE_CHECKPOINT_SYNC:
3808                         event_name = "LogicalRewriteCheckpointSync";
3809                         break;
3810                 case WAIT_EVENT_LOGICAL_REWRITE_MAPPING_SYNC:
3811                         event_name = "LogicalRewriteMappingSync";
3812                         break;
3813                 case WAIT_EVENT_LOGICAL_REWRITE_MAPPING_WRITE:
3814                         event_name = "LogicalRewriteMappingWrite";
3815                         break;
3816                 case WAIT_EVENT_LOGICAL_REWRITE_SYNC:
3817                         event_name = "LogicalRewriteSync";
3818                         break;
3819                 case WAIT_EVENT_LOGICAL_REWRITE_TRUNCATE:
3820                         event_name = "LogicalRewriteTruncate";
3821                         break;
3822                 case WAIT_EVENT_LOGICAL_REWRITE_WRITE:
3823                         event_name = "LogicalRewriteWrite";
3824                         break;
3825                 case WAIT_EVENT_RELATION_MAP_READ:
3826                         event_name = "RelationMapRead";
3827                         break;
3828                 case WAIT_EVENT_RELATION_MAP_SYNC:
3829                         event_name = "RelationMapSync";
3830                         break;
3831                 case WAIT_EVENT_RELATION_MAP_WRITE:
3832                         event_name = "RelationMapWrite";
3833                         break;
3834                 case WAIT_EVENT_REORDER_BUFFER_READ:
3835                         event_name = "ReorderBufferRead";
3836                         break;
3837                 case WAIT_EVENT_REORDER_BUFFER_WRITE:
3838                         event_name = "ReorderBufferWrite";
3839                         break;
3840                 case WAIT_EVENT_REORDER_LOGICAL_MAPPING_READ:
3841                         event_name = "ReorderLogicalMappingRead";
3842                         break;
3843                 case WAIT_EVENT_REPLICATION_SLOT_READ:
3844                         event_name = "ReplicationSlotRead";
3845                         break;
3846                 case WAIT_EVENT_REPLICATION_SLOT_RESTORE_SYNC:
3847                         event_name = "ReplicationSlotRestoreSync";
3848                         break;
3849                 case WAIT_EVENT_REPLICATION_SLOT_SYNC:
3850                         event_name = "ReplicationSlotSync";
3851                         break;
3852                 case WAIT_EVENT_REPLICATION_SLOT_WRITE:
3853                         event_name = "ReplicationSlotWrite";
3854                         break;
3855                 case WAIT_EVENT_SLRU_FLUSH_SYNC:
3856                         event_name = "SLRUFlushSync";
3857                         break;
3858                 case WAIT_EVENT_SLRU_READ:
3859                         event_name = "SLRURead";
3860                         break;
3861                 case WAIT_EVENT_SLRU_SYNC:
3862                         event_name = "SLRUSync";
3863                         break;
3864                 case WAIT_EVENT_SLRU_WRITE:
3865                         event_name = "SLRUWrite";
3866                         break;
3867                 case WAIT_EVENT_SNAPBUILD_READ:
3868                         event_name = "SnapbuildRead";
3869                         break;
3870                 case WAIT_EVENT_SNAPBUILD_SYNC:
3871                         event_name = "SnapbuildSync";
3872                         break;
3873                 case WAIT_EVENT_SNAPBUILD_WRITE:
3874                         event_name = "SnapbuildWrite";
3875                         break;
3876                 case WAIT_EVENT_TIMELINE_HISTORY_FILE_SYNC:
3877                         event_name = "TimelineHistoryFileSync";
3878                         break;
3879                 case WAIT_EVENT_TIMELINE_HISTORY_FILE_WRITE:
3880                         event_name = "TimelineHistoryFileWrite";
3881                         break;
3882                 case WAIT_EVENT_TIMELINE_HISTORY_READ:
3883                         event_name = "TimelineHistoryRead";
3884                         break;
3885                 case WAIT_EVENT_TIMELINE_HISTORY_SYNC:
3886                         event_name = "TimelineHistorySync";
3887                         break;
3888                 case WAIT_EVENT_TIMELINE_HISTORY_WRITE:
3889                         event_name = "TimelineHistoryWrite";
3890                         break;
3891                 case WAIT_EVENT_TWOPHASE_FILE_READ:
3892                         event_name = "TwophaseFileRead";
3893                         break;
3894                 case WAIT_EVENT_TWOPHASE_FILE_SYNC:
3895                         event_name = "TwophaseFileSync";
3896                         break;
3897                 case WAIT_EVENT_TWOPHASE_FILE_WRITE:
3898                         event_name = "TwophaseFileWrite";
3899                         break;
3900                 case WAIT_EVENT_WALSENDER_TIMELINE_HISTORY_READ:
3901                         event_name = "WALSenderTimelineHistoryRead";
3902                         break;
3903                 case WAIT_EVENT_WAL_BOOTSTRAP_SYNC:
3904                         event_name = "WALBootstrapSync";
3905                         break;
3906                 case WAIT_EVENT_WAL_BOOTSTRAP_WRITE:
3907                         event_name = "WALBootstrapWrite";
3908                         break;
3909                 case WAIT_EVENT_WAL_COPY_READ:
3910                         event_name = "WALCopyRead";
3911                         break;
3912                 case WAIT_EVENT_WAL_COPY_SYNC:
3913                         event_name = "WALCopySync";
3914                         break;
3915                 case WAIT_EVENT_WAL_COPY_WRITE:
3916                         event_name = "WALCopyWrite";
3917                         break;
3918                 case WAIT_EVENT_WAL_INIT_SYNC:
3919                         event_name = "WALInitSync";
3920                         break;
3921                 case WAIT_EVENT_WAL_INIT_WRITE:
3922                         event_name = "WALInitWrite";
3923                         break;
3924                 case WAIT_EVENT_WAL_READ:
3925                         event_name = "WALRead";
3926                         break;
3927                 case WAIT_EVENT_WAL_SYNC:
3928                         event_name = "WALSync";
3929                         break;
3930                 case WAIT_EVENT_WAL_SYNC_METHOD_ASSIGN:
3931                         event_name = "WALSyncMethodAssign";
3932                         break;
3933                 case WAIT_EVENT_WAL_WRITE:
3934                         event_name = "WALWrite";
3935                         break;
3936
3937                         /* no default case, so that compiler will warn */
3938         }
3939
3940         return event_name;
3941 }
3942
3943
3944 /* ----------
3945  * pgstat_get_backend_current_activity() -
3946  *
3947  *      Return a string representing the current activity of the backend with
3948  *      the specified PID.  This looks directly at the BackendStatusArray,
3949  *      and so will provide current information regardless of the age of our
3950  *      transaction's snapshot of the status array.
3951  *
3952  *      It is the caller's responsibility to invoke this only for backends whose
3953  *      state is expected to remain stable while the result is in use.  The
3954  *      only current use is in deadlock reporting, where we can expect that
3955  *      the target backend is blocked on a lock.  (There are corner cases
3956  *      where the target's wait could get aborted while we are looking at it,
3957  *      but the very worst consequence is to return a pointer to a string
3958  *      that's been changed, so we won't worry too much.)
3959  *
3960  *      Note: return strings for special cases match pg_stat_get_backend_activity.
3961  * ----------
3962  */
3963 const char *
3964 pgstat_get_backend_current_activity(int pid, bool checkUser)
3965 {
3966         PgBackendStatus *beentry;
3967         int                     i;
3968
3969         beentry = BackendStatusArray;
3970         for (i = 1; i <= MaxBackends; i++)
3971         {
3972                 /*
3973                  * Although we expect the target backend's entry to be stable, that
3974                  * doesn't imply that anyone else's is.  To avoid identifying the
3975                  * wrong backend, while we check for a match to the desired PID we
3976                  * must follow the protocol of retrying if st_changecount changes
3977                  * while we examine the entry, or if it's odd.  (This might be
3978                  * unnecessary, since fetching or storing an int is almost certainly
3979                  * atomic, but let's play it safe.)  We use a volatile pointer here to
3980                  * ensure the compiler doesn't try to get cute.
3981                  */
3982                 volatile PgBackendStatus *vbeentry = beentry;
3983                 bool            found;
3984
3985                 for (;;)
3986                 {
3987                         int                     before_changecount;
3988                         int                     after_changecount;
3989
3990                         pgstat_save_changecount_before(vbeentry, before_changecount);
3991
3992                         found = (vbeentry->st_procpid == pid);
3993
3994                         pgstat_save_changecount_after(vbeentry, after_changecount);
3995
3996                         if (before_changecount == after_changecount &&
3997                                 (before_changecount & 1) == 0)
3998                                 break;
3999
4000                         /* Make sure we can break out of loop if stuck... */
4001                         CHECK_FOR_INTERRUPTS();
4002                 }
4003
4004                 if (found)
4005                 {
4006                         /* Now it is safe to use the non-volatile pointer */
4007                         if (checkUser && !superuser() && beentry->st_userid != GetUserId())
4008                                 return "<insufficient privilege>";
4009                         else if (*(beentry->st_activity_raw) == '\0')
4010                                 return "<command string not enabled>";
4011                         else
4012                         {
4013                                 /* this'll leak a bit of memory, but that seems acceptable */
4014                                 return pgstat_clip_activity(beentry->st_activity_raw);
4015                         }
4016                 }
4017
4018                 beentry++;
4019         }
4020
4021         /* If we get here, caller is in error ... */
4022         return "<backend information not available>";
4023 }
4024
4025 /* ----------
4026  * pgstat_get_crashed_backend_activity() -
4027  *
4028  *      Return a string representing the current activity of the backend with
4029  *      the specified PID.  Like the function above, but reads shared memory with
4030  *      the expectation that it may be corrupt.  On success, copy the string
4031  *      into the "buffer" argument and return that pointer.  On failure,
4032  *      return NULL.
4033  *
4034  *      This function is only intended to be used by the postmaster to report the
4035  *      query that crashed a backend.  In particular, no attempt is made to
4036  *      follow the correct concurrency protocol when accessing the
4037  *      BackendStatusArray.  But that's OK, in the worst case we'll return a
4038  *      corrupted message.  We also must take care not to trip on ereport(ERROR).
4039  * ----------
4040  */
4041 const char *
4042 pgstat_get_crashed_backend_activity(int pid, char *buffer, int buflen)
4043 {
4044         volatile PgBackendStatus *beentry;
4045         int                     i;
4046
4047         beentry = BackendStatusArray;
4048
4049         /*
4050          * We probably shouldn't get here before shared memory has been set up,
4051          * but be safe.
4052          */
4053         if (beentry == NULL || BackendActivityBuffer == NULL)
4054                 return NULL;
4055
4056         for (i = 1; i <= MaxBackends; i++)
4057         {
4058                 if (beentry->st_procpid == pid)
4059                 {
4060                         /* Read pointer just once, so it can't change after validation */
4061                         const char *activity = beentry->st_activity_raw;
4062                         const char *activity_last;
4063
4064                         /*
4065                          * We mustn't access activity string before we verify that it
4066                          * falls within the BackendActivityBuffer. To make sure that the
4067                          * entire string including its ending is contained within the
4068                          * buffer, subtract one activity length from the buffer size.
4069                          */
4070                         activity_last = BackendActivityBuffer + BackendActivityBufferSize
4071                                 - pgstat_track_activity_query_size;
4072
4073                         if (activity < BackendActivityBuffer ||
4074                                 activity > activity_last)
4075                                 return NULL;
4076
4077                         /* If no string available, no point in a report */
4078                         if (activity[0] == '\0')
4079                                 return NULL;
4080
4081                         /*
4082                          * Copy only ASCII-safe characters so we don't run into encoding
4083                          * problems when reporting the message; and be sure not to run off
4084                          * the end of memory.  As only ASCII characters are reported, it
4085                          * doesn't seem necessary to perform multibyte aware clipping.
4086                          */
4087                         ascii_safe_strlcpy(buffer, activity,
4088                                                            Min(buflen, pgstat_track_activity_query_size));
4089
4090                         return buffer;
4091                 }
4092
4093                 beentry++;
4094         }
4095
4096         /* PID not found */
4097         return NULL;
4098 }
4099
4100 const char *
4101 pgstat_get_backend_desc(BackendType backendType)
4102 {
4103         const char *backendDesc = "unknown process type";
4104
4105         switch (backendType)
4106         {
4107                 case B_AUTOVAC_LAUNCHER:
4108                         backendDesc = "autovacuum launcher";
4109                         break;
4110                 case B_AUTOVAC_WORKER:
4111                         backendDesc = "autovacuum worker";
4112                         break;
4113                 case B_BACKEND:
4114                         backendDesc = "client backend";
4115                         break;
4116                 case B_BG_WORKER:
4117                         backendDesc = "background worker";
4118                         break;
4119                 case B_BG_WRITER:
4120                         backendDesc = "background writer";
4121                         break;
4122                 case B_CHECKPOINTER:
4123                         backendDesc = "checkpointer";
4124                         break;
4125                 case B_STARTUP:
4126                         backendDesc = "startup";
4127                         break;
4128                 case B_WAL_RECEIVER:
4129                         backendDesc = "walreceiver";
4130                         break;
4131                 case B_WAL_SENDER:
4132                         backendDesc = "walsender";
4133                         break;
4134                 case B_WAL_WRITER:
4135                         backendDesc = "walwriter";
4136                         break;
4137         }
4138
4139         return backendDesc;
4140 }
4141
4142 /* ------------------------------------------------------------
4143  * Local support functions follow
4144  * ------------------------------------------------------------
4145  */
4146
4147
4148 /* ----------
4149  * pgstat_setheader() -
4150  *
4151  *              Set common header fields in a statistics message
4152  * ----------
4153  */
4154 static void
4155 pgstat_setheader(PgStat_MsgHdr *hdr, StatMsgType mtype)
4156 {
4157         hdr->m_type = mtype;
4158 }
4159
4160
4161 /* ----------
4162  * pgstat_send() -
4163  *
4164  *              Send out one statistics message to the collector
4165  * ----------
4166  */
4167 static void
4168 pgstat_send(void *msg, int len)
4169 {
4170         int                     rc;
4171
4172         if (pgStatSock == PGINVALID_SOCKET)
4173                 return;
4174
4175         ((PgStat_MsgHdr *) msg)->m_size = len;
4176
4177         /* We'll retry after EINTR, but ignore all other failures */
4178         do
4179         {
4180                 rc = send(pgStatSock, msg, len, 0);
4181         } while (rc < 0 && errno == EINTR);
4182
4183 #ifdef USE_ASSERT_CHECKING
4184         /* In debug builds, log send failures ... */
4185         if (rc < 0)
4186                 elog(LOG, "could not send to statistics collector: %m");
4187 #endif
4188 }
4189
4190 /* ----------
4191  * pgstat_send_archiver() -
4192  *
4193  *      Tell the collector about the WAL file that we successfully
4194  *      archived or failed to archive.
4195  * ----------
4196  */
4197 void
4198 pgstat_send_archiver(const char *xlog, bool failed)
4199 {
4200         PgStat_MsgArchiver msg;
4201
4202         /*
4203          * Prepare and send the message
4204          */
4205         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_ARCHIVER);
4206         msg.m_failed = failed;
4207         StrNCpy(msg.m_xlog, xlog, sizeof(msg.m_xlog));
4208         msg.m_timestamp = GetCurrentTimestamp();
4209         pgstat_send(&msg, sizeof(msg));
4210 }
4211
4212 /* ----------
4213  * pgstat_send_bgwriter() -
4214  *
4215  *              Send bgwriter statistics to the collector
4216  * ----------
4217  */
4218 void
4219 pgstat_send_bgwriter(void)
4220 {
4221         /* We assume this initializes to zeroes */
4222         static const PgStat_MsgBgWriter all_zeroes;
4223
4224         /*
4225          * This function can be called even if nothing at all has happened. In
4226          * this case, avoid sending a completely empty message to the stats
4227          * collector.
4228          */
4229         if (memcmp(&BgWriterStats, &all_zeroes, sizeof(PgStat_MsgBgWriter)) == 0)
4230                 return;
4231
4232         /*
4233          * Prepare and send the message
4234          */
4235         pgstat_setheader(&BgWriterStats.m_hdr, PGSTAT_MTYPE_BGWRITER);
4236         pgstat_send(&BgWriterStats, sizeof(BgWriterStats));
4237
4238         /*
4239          * Clear out the statistics buffer, so it can be re-used.
4240          */
4241         MemSet(&BgWriterStats, 0, sizeof(BgWriterStats));
4242 }
4243
4244
4245 /* ----------
4246  * PgstatCollectorMain() -
4247  *
4248  *      Start up the statistics collector process.  This is the body of the
4249  *      postmaster child process.
4250  *
4251  *      The argc/argv parameters are valid only in EXEC_BACKEND case.
4252  * ----------
4253  */
4254 NON_EXEC_STATIC void
4255 PgstatCollectorMain(int argc, char *argv[])
4256 {
4257         int                     len;
4258         PgStat_Msg      msg;
4259         int                     wr;
4260
4261         /*
4262          * Ignore all signals usually bound to some action in the postmaster,
4263          * except SIGHUP and SIGQUIT.  Note we don't need a SIGUSR1 handler to
4264          * support latch operations, because we only use a local latch.
4265          */
4266         pqsignal(SIGHUP, pgstat_sighup_handler);
4267         pqsignal(SIGINT, SIG_IGN);
4268         pqsignal(SIGTERM, SIG_IGN);
4269         pqsignal(SIGQUIT, pgstat_exit);
4270         pqsignal(SIGALRM, SIG_IGN);
4271         pqsignal(SIGPIPE, SIG_IGN);
4272         pqsignal(SIGUSR1, SIG_IGN);
4273         pqsignal(SIGUSR2, SIG_IGN);
4274         /* Reset some signals that are accepted by postmaster but not here */
4275         pqsignal(SIGCHLD, SIG_DFL);
4276         PG_SETMASK(&UnBlockSig);
4277
4278         /*
4279          * Identify myself via ps
4280          */
4281         init_ps_display("stats collector", "", "", "");
4282
4283         /*
4284          * Read in existing stats files or initialize the stats to zero.
4285          */
4286         pgStatRunningInCollector = true;
4287         pgStatDBHash = pgstat_read_statsfiles(InvalidOid, true, true);
4288
4289         /*
4290          * Loop to process messages until we get SIGQUIT or detect ungraceful
4291          * death of our parent postmaster.
4292          *
4293          * For performance reasons, we don't want to do ResetLatch/WaitLatch after
4294          * every message; instead, do that only after a recv() fails to obtain a
4295          * message.  (This effectively means that if backends are sending us stuff
4296          * like mad, we won't notice postmaster death until things slack off a
4297          * bit; which seems fine.)      To do that, we have an inner loop that
4298          * iterates as long as recv() succeeds.  We do recognize got_SIGHUP inside
4299          * the inner loop, which means that such interrupts will get serviced but
4300          * the latch won't get cleared until next time there is a break in the
4301          * action.
4302          */
4303         for (;;)
4304         {
4305                 /* Clear any already-pending wakeups */
4306                 ResetLatch(MyLatch);
4307
4308                 /*
4309                  * Quit if we get SIGQUIT from the postmaster.
4310                  */
4311                 if (need_exit)
4312                         break;
4313
4314                 /*
4315                  * Inner loop iterates as long as we keep getting messages, or until
4316                  * need_exit becomes set.
4317                  */
4318                 while (!need_exit)
4319                 {
4320                         /*
4321                          * Reload configuration if we got SIGHUP from the postmaster.
4322                          */
4323                         if (got_SIGHUP)
4324                         {
4325                                 got_SIGHUP = false;
4326                                 ProcessConfigFile(PGC_SIGHUP);
4327                         }
4328
4329                         /*
4330                          * Write the stats file(s) if a new request has arrived that is
4331                          * not satisfied by existing file(s).
4332                          */
4333                         if (pgstat_write_statsfile_needed())
4334                                 pgstat_write_statsfiles(false, false);
4335
4336                         /*
4337                          * Try to receive and process a message.  This will not block,
4338                          * since the socket is set to non-blocking mode.
4339                          *
4340                          * XXX On Windows, we have to force pgwin32_recv to cooperate,
4341                          * despite the previous use of pg_set_noblock() on the socket.
4342                          * This is extremely broken and should be fixed someday.
4343                          */
4344 #ifdef WIN32
4345                         pgwin32_noblock = 1;
4346 #endif
4347
4348                         len = recv(pgStatSock, (char *) &msg,
4349                                            sizeof(PgStat_Msg), 0);
4350
4351 #ifdef WIN32
4352                         pgwin32_noblock = 0;
4353 #endif
4354
4355                         if (len < 0)
4356                         {
4357                                 if (errno == EAGAIN || errno == EWOULDBLOCK || errno == EINTR)
4358                                         break;          /* out of inner loop */
4359                                 ereport(ERROR,
4360                                                 (errcode_for_socket_access(),
4361                                                  errmsg("could not read statistics message: %m")));
4362                         }
4363
4364                         /*
4365                          * We ignore messages that are smaller than our common header
4366                          */
4367                         if (len < sizeof(PgStat_MsgHdr))
4368                                 continue;
4369
4370                         /*
4371                          * The received length must match the length in the header
4372                          */
4373                         if (msg.msg_hdr.m_size != len)
4374                                 continue;
4375
4376                         /*
4377                          * O.K. - we accept this message.  Process it.
4378                          */
4379                         switch (msg.msg_hdr.m_type)
4380                         {
4381                                 case PGSTAT_MTYPE_DUMMY:
4382                                         break;
4383
4384                                 case PGSTAT_MTYPE_INQUIRY:
4385                                         pgstat_recv_inquiry((PgStat_MsgInquiry *) &msg, len);
4386                                         break;
4387
4388                                 case PGSTAT_MTYPE_TABSTAT:
4389                                         pgstat_recv_tabstat((PgStat_MsgTabstat *) &msg, len);
4390                                         break;
4391
4392                                 case PGSTAT_MTYPE_TABPURGE:
4393                                         pgstat_recv_tabpurge((PgStat_MsgTabpurge *) &msg, len);
4394                                         break;
4395
4396                                 case PGSTAT_MTYPE_DROPDB:
4397                                         pgstat_recv_dropdb((PgStat_MsgDropdb *) &msg, len);
4398                                         break;
4399
4400                                 case PGSTAT_MTYPE_RESETCOUNTER:
4401                                         pgstat_recv_resetcounter((PgStat_MsgResetcounter *) &msg,
4402                                                                                          len);
4403                                         break;
4404
4405                                 case PGSTAT_MTYPE_RESETSHAREDCOUNTER:
4406                                         pgstat_recv_resetsharedcounter(
4407                                                                                                    (PgStat_MsgResetsharedcounter *) &msg,
4408                                                                                                    len);
4409                                         break;
4410
4411                                 case PGSTAT_MTYPE_RESETSINGLECOUNTER:
4412                                         pgstat_recv_resetsinglecounter(
4413                                                                                                    (PgStat_MsgResetsinglecounter *) &msg,
4414                                                                                                    len);
4415                                         break;
4416
4417                                 case PGSTAT_MTYPE_AUTOVAC_START:
4418                                         pgstat_recv_autovac((PgStat_MsgAutovacStart *) &msg, len);
4419                                         break;
4420
4421                                 case PGSTAT_MTYPE_VACUUM:
4422                                         pgstat_recv_vacuum((PgStat_MsgVacuum *) &msg, len);
4423                                         break;
4424
4425                                 case PGSTAT_MTYPE_ANALYZE:
4426                                         pgstat_recv_analyze((PgStat_MsgAnalyze *) &msg, len);
4427                                         break;
4428
4429                                 case PGSTAT_MTYPE_ARCHIVER:
4430                                         pgstat_recv_archiver((PgStat_MsgArchiver *) &msg, len);
4431                                         break;
4432
4433                                 case PGSTAT_MTYPE_BGWRITER:
4434                                         pgstat_recv_bgwriter((PgStat_MsgBgWriter *) &msg, len);
4435                                         break;
4436
4437                                 case PGSTAT_MTYPE_FUNCSTAT:
4438                                         pgstat_recv_funcstat((PgStat_MsgFuncstat *) &msg, len);
4439                                         break;
4440
4441                                 case PGSTAT_MTYPE_FUNCPURGE:
4442                                         pgstat_recv_funcpurge((PgStat_MsgFuncpurge *) &msg, len);
4443                                         break;
4444
4445                                 case PGSTAT_MTYPE_RECOVERYCONFLICT:
4446                                         pgstat_recv_recoveryconflict((PgStat_MsgRecoveryConflict *) &msg, len);
4447                                         break;
4448
4449                                 case PGSTAT_MTYPE_DEADLOCK:
4450                                         pgstat_recv_deadlock((PgStat_MsgDeadlock *) &msg, len);
4451                                         break;
4452
4453                                 case PGSTAT_MTYPE_TEMPFILE:
4454                                         pgstat_recv_tempfile((PgStat_MsgTempFile *) &msg, len);
4455                                         break;
4456
4457                                 default:
4458                                         break;
4459                         }
4460                 }                                               /* end of inner message-processing loop */
4461
4462                 /* Sleep until there's something to do */
4463 #ifndef WIN32
4464                 wr = WaitLatchOrSocket(MyLatch,
4465                                                            WL_LATCH_SET | WL_POSTMASTER_DEATH | WL_SOCKET_READABLE,
4466                                                            pgStatSock, -1L,
4467                                                            WAIT_EVENT_PGSTAT_MAIN);
4468 #else
4469
4470                 /*
4471                  * Windows, at least in its Windows Server 2003 R2 incarnation,
4472                  * sometimes loses FD_READ events.  Waking up and retrying the recv()
4473                  * fixes that, so don't sleep indefinitely.  This is a crock of the
4474                  * first water, but until somebody wants to debug exactly what's
4475                  * happening there, this is the best we can do.  The two-second
4476                  * timeout matches our pre-9.2 behavior, and needs to be short enough
4477                  * to not provoke "using stale statistics" complaints from
4478                  * backend_read_statsfile.
4479                  */
4480                 wr = WaitLatchOrSocket(MyLatch,
4481                                                            WL_LATCH_SET | WL_POSTMASTER_DEATH | WL_SOCKET_READABLE | WL_TIMEOUT,
4482                                                            pgStatSock,
4483                                                            2 * 1000L /* msec */ ,
4484                                                            WAIT_EVENT_PGSTAT_MAIN);
4485 #endif
4486
4487                 /*
4488                  * Emergency bailout if postmaster has died.  This is to avoid the
4489                  * necessity for manual cleanup of all postmaster children.
4490                  */
4491                 if (wr & WL_POSTMASTER_DEATH)
4492                         break;
4493         }                                                       /* end of outer loop */
4494
4495         /*
4496          * Save the final stats to reuse at next startup.
4497          */
4498         pgstat_write_statsfiles(true, true);
4499
4500         exit(0);
4501 }
4502
4503
4504 /* SIGQUIT signal handler for collector process */
4505 static void
4506 pgstat_exit(SIGNAL_ARGS)
4507 {
4508         int                     save_errno = errno;
4509
4510         need_exit = true;
4511         SetLatch(MyLatch);
4512
4513         errno = save_errno;
4514 }
4515
4516 /* SIGHUP handler for collector process */
4517 static void
4518 pgstat_sighup_handler(SIGNAL_ARGS)
4519 {
4520         int                     save_errno = errno;
4521
4522         got_SIGHUP = true;
4523         SetLatch(MyLatch);
4524
4525         errno = save_errno;
4526 }
4527
4528 /*
4529  * Subroutine to clear stats in a database entry
4530  *
4531  * Tables and functions hashes are initialized to empty.
4532  */
4533 static void
4534 reset_dbentry_counters(PgStat_StatDBEntry *dbentry)
4535 {
4536         HASHCTL         hash_ctl;
4537
4538         dbentry->n_xact_commit = 0;
4539         dbentry->n_xact_rollback = 0;
4540         dbentry->n_blocks_fetched = 0;
4541         dbentry->n_blocks_hit = 0;
4542         dbentry->n_tuples_returned = 0;
4543         dbentry->n_tuples_fetched = 0;
4544         dbentry->n_tuples_inserted = 0;
4545         dbentry->n_tuples_updated = 0;
4546         dbentry->n_tuples_deleted = 0;
4547         dbentry->last_autovac_time = 0;
4548         dbentry->n_conflict_tablespace = 0;
4549         dbentry->n_conflict_lock = 0;
4550         dbentry->n_conflict_snapshot = 0;
4551         dbentry->n_conflict_bufferpin = 0;
4552         dbentry->n_conflict_startup_deadlock = 0;
4553         dbentry->n_temp_files = 0;
4554         dbentry->n_temp_bytes = 0;
4555         dbentry->n_deadlocks = 0;
4556         dbentry->n_block_read_time = 0;
4557         dbentry->n_block_write_time = 0;
4558
4559         dbentry->stat_reset_timestamp = GetCurrentTimestamp();
4560         dbentry->stats_timestamp = 0;
4561
4562         memset(&hash_ctl, 0, sizeof(hash_ctl));
4563         hash_ctl.keysize = sizeof(Oid);
4564         hash_ctl.entrysize = sizeof(PgStat_StatTabEntry);
4565         dbentry->tables = hash_create("Per-database table",
4566                                                                   PGSTAT_TAB_HASH_SIZE,
4567                                                                   &hash_ctl,
4568                                                                   HASH_ELEM | HASH_BLOBS);
4569
4570         hash_ctl.keysize = sizeof(Oid);
4571         hash_ctl.entrysize = sizeof(PgStat_StatFuncEntry);
4572         dbentry->functions = hash_create("Per-database function",
4573                                                                          PGSTAT_FUNCTION_HASH_SIZE,
4574                                                                          &hash_ctl,
4575                                                                          HASH_ELEM | HASH_BLOBS);
4576 }
4577
4578 /*
4579  * Lookup the hash table entry for the specified database. If no hash
4580  * table entry exists, initialize it, if the create parameter is true.
4581  * Else, return NULL.
4582  */
4583 static PgStat_StatDBEntry *
4584 pgstat_get_db_entry(Oid databaseid, bool create)
4585 {
4586         PgStat_StatDBEntry *result;
4587         bool            found;
4588         HASHACTION      action = (create ? HASH_ENTER : HASH_FIND);
4589
4590         /* Lookup or create the hash table entry for this database */
4591         result = (PgStat_StatDBEntry *) hash_search(pgStatDBHash,
4592                                                                                                 &databaseid,
4593                                                                                                 action, &found);
4594
4595         if (!create && !found)
4596                 return NULL;
4597
4598         /*
4599          * If not found, initialize the new one.  This creates empty hash tables
4600          * for tables and functions, too.
4601          */
4602         if (!found)
4603                 reset_dbentry_counters(result);
4604
4605         return result;
4606 }
4607
4608
4609 /*
4610  * Lookup the hash table entry for the specified table. If no hash
4611  * table entry exists, initialize it, if the create parameter is true.
4612  * Else, return NULL.
4613  */
4614 static PgStat_StatTabEntry *
4615 pgstat_get_tab_entry(PgStat_StatDBEntry *dbentry, Oid tableoid, bool create)
4616 {
4617         PgStat_StatTabEntry *result;
4618         bool            found;
4619         HASHACTION      action = (create ? HASH_ENTER : HASH_FIND);
4620
4621         /* Lookup or create the hash table entry for this table */
4622         result = (PgStat_StatTabEntry *) hash_search(dbentry->tables,
4623                                                                                                  &tableoid,
4624                                                                                                  action, &found);
4625
4626         if (!create && !found)
4627                 return NULL;
4628
4629         /* If not found, initialize the new one. */
4630         if (!found)
4631         {
4632                 result->numscans = 0;
4633                 result->tuples_returned = 0;
4634                 result->tuples_fetched = 0;
4635                 result->tuples_inserted = 0;
4636                 result->tuples_updated = 0;
4637                 result->tuples_deleted = 0;
4638                 result->tuples_hot_updated = 0;
4639                 result->n_live_tuples = 0;
4640                 result->n_dead_tuples = 0;
4641                 result->changes_since_analyze = 0;
4642                 result->blocks_fetched = 0;
4643                 result->blocks_hit = 0;
4644                 result->vacuum_timestamp = 0;
4645                 result->vacuum_count = 0;
4646                 result->autovac_vacuum_timestamp = 0;
4647                 result->autovac_vacuum_count = 0;
4648                 result->analyze_timestamp = 0;
4649                 result->analyze_count = 0;
4650                 result->autovac_analyze_timestamp = 0;
4651                 result->autovac_analyze_count = 0;
4652         }
4653
4654         return result;
4655 }
4656
4657
4658 /* ----------
4659  * pgstat_write_statsfiles() -
4660  *              Write the global statistics file, as well as requested DB files.
4661  *
4662  *      'permanent' specifies writing to the permanent files not temporary ones.
4663  *      When true (happens only when the collector is shutting down), also remove
4664  *      the temporary files so that backends starting up under a new postmaster
4665  *      can't read old data before the new collector is ready.
4666  *
4667  *      When 'allDbs' is false, only the requested databases (listed in
4668  *      pending_write_requests) will be written; otherwise, all databases
4669  *      will be written.
4670  * ----------
4671  */
4672 static void
4673 pgstat_write_statsfiles(bool permanent, bool allDbs)
4674 {
4675         HASH_SEQ_STATUS hstat;
4676         PgStat_StatDBEntry *dbentry;
4677         FILE       *fpout;
4678         int32           format_id;
4679         const char *tmpfile = permanent ? PGSTAT_STAT_PERMANENT_TMPFILE : pgstat_stat_tmpname;
4680         const char *statfile = permanent ? PGSTAT_STAT_PERMANENT_FILENAME : pgstat_stat_filename;
4681         int                     rc;
4682
4683         elog(DEBUG2, "writing stats file \"%s\"", statfile);
4684
4685         /*
4686          * Open the statistics temp file to write out the current values.
4687          */
4688         fpout = AllocateFile(tmpfile, PG_BINARY_W);
4689         if (fpout == NULL)
4690         {
4691                 ereport(LOG,
4692                                 (errcode_for_file_access(),
4693                                  errmsg("could not open temporary statistics file \"%s\": %m",
4694                                                 tmpfile)));
4695                 return;
4696         }
4697
4698         /*
4699          * Set the timestamp of the stats file.
4700          */
4701         globalStats.stats_timestamp = GetCurrentTimestamp();
4702
4703         /*
4704          * Write the file header --- currently just a format ID.
4705          */
4706         format_id = PGSTAT_FILE_FORMAT_ID;
4707         rc = fwrite(&format_id, sizeof(format_id), 1, fpout);
4708         (void) rc;                                      /* we'll check for error with ferror */
4709
4710         /*
4711          * Write global stats struct
4712          */
4713         rc = fwrite(&globalStats, sizeof(globalStats), 1, fpout);
4714         (void) rc;                                      /* we'll check for error with ferror */
4715
4716         /*
4717          * Write archiver stats struct
4718          */
4719         rc = fwrite(&archiverStats, sizeof(archiverStats), 1, fpout);
4720         (void) rc;                                      /* we'll check for error with ferror */
4721
4722         /*
4723          * Walk through the database table.
4724          */
4725         hash_seq_init(&hstat, pgStatDBHash);
4726         while ((dbentry = (PgStat_StatDBEntry *) hash_seq_search(&hstat)) != NULL)
4727         {
4728                 /*
4729                  * Write out the table and function stats for this DB into the
4730                  * appropriate per-DB stat file, if required.
4731                  */
4732                 if (allDbs || pgstat_db_requested(dbentry->databaseid))
4733                 {
4734                         /* Make DB's timestamp consistent with the global stats */
4735                         dbentry->stats_timestamp = globalStats.stats_timestamp;
4736
4737                         pgstat_write_db_statsfile(dbentry, permanent);
4738                 }
4739
4740                 /*
4741                  * Write out the DB entry. We don't write the tables or functions
4742                  * pointers, since they're of no use to any other process.
4743                  */
4744                 fputc('D', fpout);
4745                 rc = fwrite(dbentry, offsetof(PgStat_StatDBEntry, tables), 1, fpout);
4746                 (void) rc;                              /* we'll check for error with ferror */
4747         }
4748
4749         /*
4750          * No more output to be done. Close the temp file and replace the old
4751          * pgstat.stat with it.  The ferror() check replaces testing for error
4752          * after each individual fputc or fwrite above.
4753          */
4754         fputc('E', fpout);
4755
4756         if (ferror(fpout))
4757         {
4758                 ereport(LOG,
4759                                 (errcode_for_file_access(),
4760                                  errmsg("could not write temporary statistics file \"%s\": %m",
4761                                                 tmpfile)));
4762                 FreeFile(fpout);
4763                 unlink(tmpfile);
4764         }
4765         else if (FreeFile(fpout) < 0)
4766         {
4767                 ereport(LOG,
4768                                 (errcode_for_file_access(),
4769                                  errmsg("could not close temporary statistics file \"%s\": %m",
4770                                                 tmpfile)));
4771                 unlink(tmpfile);
4772         }
4773         else if (rename(tmpfile, statfile) < 0)
4774         {
4775                 ereport(LOG,
4776                                 (errcode_for_file_access(),
4777                                  errmsg("could not rename temporary statistics file \"%s\" to \"%s\": %m",
4778                                                 tmpfile, statfile)));
4779                 unlink(tmpfile);
4780         }
4781
4782         if (permanent)
4783                 unlink(pgstat_stat_filename);
4784
4785         /*
4786          * Now throw away the list of requests.  Note that requests sent after we
4787          * started the write are still waiting on the network socket.
4788          */
4789         list_free(pending_write_requests);
4790         pending_write_requests = NIL;
4791 }
4792
4793 /*
4794  * return the filename for a DB stat file; filename is the output buffer,
4795  * of length len.
4796  */
4797 static void
4798 get_dbstat_filename(bool permanent, bool tempname, Oid databaseid,
4799                                         char *filename, int len)
4800 {
4801         int                     printed;
4802
4803         /* NB -- pgstat_reset_remove_files knows about the pattern this uses */
4804         printed = snprintf(filename, len, "%s/db_%u.%s",
4805                                            permanent ? PGSTAT_STAT_PERMANENT_DIRECTORY :
4806                                            pgstat_stat_directory,
4807                                            databaseid,
4808                                            tempname ? "tmp" : "stat");
4809         if (printed >= len)
4810                 elog(ERROR, "overlength pgstat path");
4811 }
4812
4813 /* ----------
4814  * pgstat_write_db_statsfile() -
4815  *              Write the stat file for a single database.
4816  *
4817  *      If writing to the permanent file (happens when the collector is
4818  *      shutting down only), remove the temporary file so that backends
4819  *      starting up under a new postmaster can't read the old data before
4820  *      the new collector is ready.
4821  * ----------
4822  */
4823 static void
4824 pgstat_write_db_statsfile(PgStat_StatDBEntry *dbentry, bool permanent)
4825 {
4826         HASH_SEQ_STATUS tstat;
4827         HASH_SEQ_STATUS fstat;
4828         PgStat_StatTabEntry *tabentry;
4829         PgStat_StatFuncEntry *funcentry;
4830         FILE       *fpout;
4831         int32           format_id;
4832         Oid                     dbid = dbentry->databaseid;
4833         int                     rc;
4834         char            tmpfile[MAXPGPATH];
4835         char            statfile[MAXPGPATH];
4836
4837         get_dbstat_filename(permanent, true, dbid, tmpfile, MAXPGPATH);
4838         get_dbstat_filename(permanent, false, dbid, statfile, MAXPGPATH);
4839
4840         elog(DEBUG2, "writing stats file \"%s\"", statfile);
4841
4842         /*
4843          * Open the statistics temp file to write out the current values.
4844          */
4845         fpout = AllocateFile(tmpfile, PG_BINARY_W);
4846         if (fpout == NULL)
4847         {
4848                 ereport(LOG,
4849                                 (errcode_for_file_access(),
4850                                  errmsg("could not open temporary statistics file \"%s\": %m",
4851                                                 tmpfile)));
4852                 return;
4853         }
4854
4855         /*
4856          * Write the file header --- currently just a format ID.
4857          */
4858         format_id = PGSTAT_FILE_FORMAT_ID;
4859         rc = fwrite(&format_id, sizeof(format_id), 1, fpout);
4860         (void) rc;                                      /* we'll check for error with ferror */
4861
4862         /*
4863          * Walk through the database's access stats per table.
4864          */
4865         hash_seq_init(&tstat, dbentry->tables);
4866         while ((tabentry = (PgStat_StatTabEntry *) hash_seq_search(&tstat)) != NULL)
4867         {
4868                 fputc('T', fpout);
4869                 rc = fwrite(tabentry, sizeof(PgStat_StatTabEntry), 1, fpout);
4870                 (void) rc;                              /* we'll check for error with ferror */
4871         }
4872
4873         /*
4874          * Walk through the database's function stats table.
4875          */
4876         hash_seq_init(&fstat, dbentry->functions);
4877         while ((funcentry = (PgStat_StatFuncEntry *) hash_seq_search(&fstat)) != NULL)
4878         {
4879                 fputc('F', fpout);
4880                 rc = fwrite(funcentry, sizeof(PgStat_StatFuncEntry), 1, fpout);
4881                 (void) rc;                              /* we'll check for error with ferror */
4882         }
4883
4884         /*
4885          * No more output to be done. Close the temp file and replace the old
4886          * pgstat.stat with it.  The ferror() check replaces testing for error
4887          * after each individual fputc or fwrite above.
4888          */
4889         fputc('E', fpout);
4890
4891         if (ferror(fpout))
4892         {
4893                 ereport(LOG,
4894                                 (errcode_for_file_access(),
4895                                  errmsg("could not write temporary statistics file \"%s\": %m",
4896                                                 tmpfile)));
4897                 FreeFile(fpout);
4898                 unlink(tmpfile);
4899         }
4900         else if (FreeFile(fpout) < 0)
4901         {
4902                 ereport(LOG,
4903                                 (errcode_for_file_access(),
4904                                  errmsg("could not close temporary statistics file \"%s\": %m",
4905                                                 tmpfile)));
4906                 unlink(tmpfile);
4907         }
4908         else if (rename(tmpfile, statfile) < 0)
4909         {
4910                 ereport(LOG,
4911                                 (errcode_for_file_access(),
4912                                  errmsg("could not rename temporary statistics file \"%s\" to \"%s\": %m",
4913                                                 tmpfile, statfile)));
4914                 unlink(tmpfile);
4915         }
4916
4917         if (permanent)
4918         {
4919                 get_dbstat_filename(false, false, dbid, statfile, MAXPGPATH);
4920
4921                 elog(DEBUG2, "removing temporary stats file \"%s\"", statfile);
4922                 unlink(statfile);
4923         }
4924 }
4925
4926 /* ----------
4927  * pgstat_read_statsfiles() -
4928  *
4929  *      Reads in some existing statistics collector files and returns the
4930  *      databases hash table that is the top level of the data.
4931  *
4932  *      If 'onlydb' is not InvalidOid, it means we only want data for that DB
4933  *      plus the shared catalogs ("DB 0").  We'll still populate the DB hash
4934  *      table for all databases, but we don't bother even creating table/function
4935  *      hash tables for other databases.
4936  *
4937  *      'permanent' specifies reading from the permanent files not temporary ones.
4938  *      When true (happens only when the collector is starting up), remove the
4939  *      files after reading; the in-memory status is now authoritative, and the
4940  *      files would be out of date in case somebody else reads them.
4941  *
4942  *      If a 'deep' read is requested, table/function stats are read, otherwise
4943  *      the table/function hash tables remain empty.
4944  * ----------
4945  */
4946 static HTAB *
4947 pgstat_read_statsfiles(Oid onlydb, bool permanent, bool deep)
4948 {
4949         PgStat_StatDBEntry *dbentry;
4950         PgStat_StatDBEntry dbbuf;
4951         HASHCTL         hash_ctl;
4952         HTAB       *dbhash;
4953         FILE       *fpin;
4954         int32           format_id;
4955         bool            found;
4956         const char *statfile = permanent ? PGSTAT_STAT_PERMANENT_FILENAME : pgstat_stat_filename;
4957
4958         /*
4959          * The tables will live in pgStatLocalContext.
4960          */
4961         pgstat_setup_memcxt();
4962
4963         /*
4964          * Create the DB hashtable
4965          */
4966         memset(&hash_ctl, 0, sizeof(hash_ctl));
4967         hash_ctl.keysize = sizeof(Oid);
4968         hash_ctl.entrysize = sizeof(PgStat_StatDBEntry);
4969         hash_ctl.hcxt = pgStatLocalContext;
4970         dbhash = hash_create("Databases hash", PGSTAT_DB_HASH_SIZE, &hash_ctl,
4971                                                  HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
4972
4973         /*
4974          * Clear out global and archiver statistics so they start from zero in
4975          * case we can't load an existing statsfile.
4976          */
4977         memset(&globalStats, 0, sizeof(globalStats));
4978         memset(&archiverStats, 0, sizeof(archiverStats));
4979
4980         /*
4981          * Set the current timestamp (will be kept only in case we can't load an
4982          * existing statsfile).
4983          */
4984         globalStats.stat_reset_timestamp = GetCurrentTimestamp();
4985         archiverStats.stat_reset_timestamp = globalStats.stat_reset_timestamp;
4986
4987         /*
4988          * Try to open the stats file. If it doesn't exist, the backends simply
4989          * return zero for anything and the collector simply starts from scratch
4990          * with empty counters.
4991          *
4992          * ENOENT is a possibility if the stats collector is not running or has
4993          * not yet written the stats file the first time.  Any other failure
4994          * condition is suspicious.
4995          */
4996         if ((fpin = AllocateFile(statfile, PG_BINARY_R)) == NULL)
4997         {
4998                 if (errno != ENOENT)
4999                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5000                                         (errcode_for_file_access(),
5001                                          errmsg("could not open statistics file \"%s\": %m",
5002                                                         statfile)));
5003                 return dbhash;
5004         }
5005
5006         /*
5007          * Verify it's of the expected format.
5008          */
5009         if (fread(&format_id, 1, sizeof(format_id), fpin) != sizeof(format_id) ||
5010                 format_id != PGSTAT_FILE_FORMAT_ID)
5011         {
5012                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5013                                 (errmsg("corrupted statistics file \"%s\"", statfile)));
5014                 goto done;
5015         }
5016
5017         /*
5018          * Read global stats struct
5019          */
5020         if (fread(&globalStats, 1, sizeof(globalStats), fpin) != sizeof(globalStats))
5021         {
5022                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5023                                 (errmsg("corrupted statistics file \"%s\"", statfile)));
5024                 memset(&globalStats, 0, sizeof(globalStats));
5025                 goto done;
5026         }
5027
5028         /*
5029          * In the collector, disregard the timestamp we read from the permanent
5030          * stats file; we should be willing to write a temp stats file immediately
5031          * upon the first request from any backend.  This only matters if the old
5032          * file's timestamp is less than PGSTAT_STAT_INTERVAL ago, but that's not
5033          * an unusual scenario.
5034          */
5035         if (pgStatRunningInCollector)
5036                 globalStats.stats_timestamp = 0;
5037
5038         /*
5039          * Read archiver stats struct
5040          */
5041         if (fread(&archiverStats, 1, sizeof(archiverStats), fpin) != sizeof(archiverStats))
5042         {
5043                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5044                                 (errmsg("corrupted statistics file \"%s\"", statfile)));
5045                 memset(&archiverStats, 0, sizeof(archiverStats));
5046                 goto done;
5047         }
5048
5049         /*
5050          * We found an existing collector stats file. Read it and put all the
5051          * hashtable entries into place.
5052          */
5053         for (;;)
5054         {
5055                 switch (fgetc(fpin))
5056                 {
5057                                 /*
5058                                  * 'D'  A PgStat_StatDBEntry struct describing a database
5059                                  * follows.
5060                                  */
5061                         case 'D':
5062                                 if (fread(&dbbuf, 1, offsetof(PgStat_StatDBEntry, tables),
5063                                                   fpin) != offsetof(PgStat_StatDBEntry, tables))
5064                                 {
5065                                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5066                                                         (errmsg("corrupted statistics file \"%s\"",
5067                                                                         statfile)));
5068                                         goto done;
5069                                 }
5070
5071                                 /*
5072                                  * Add to the DB hash
5073                                  */
5074                                 dbentry = (PgStat_StatDBEntry *) hash_search(dbhash,
5075                                                                                                                          (void *) &dbbuf.databaseid,
5076                                                                                                                          HASH_ENTER,
5077                                                                                                                          &found);
5078                                 if (found)
5079                                 {
5080                                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5081                                                         (errmsg("corrupted statistics file \"%s\"",
5082                                                                         statfile)));
5083                                         goto done;
5084                                 }
5085
5086                                 memcpy(dbentry, &dbbuf, sizeof(PgStat_StatDBEntry));
5087                                 dbentry->tables = NULL;
5088                                 dbentry->functions = NULL;
5089
5090                                 /*
5091                                  * In the collector, disregard the timestamp we read from the
5092                                  * permanent stats file; we should be willing to write a temp
5093                                  * stats file immediately upon the first request from any
5094                                  * backend.
5095                                  */
5096                                 if (pgStatRunningInCollector)
5097                                         dbentry->stats_timestamp = 0;
5098
5099                                 /*
5100                                  * Don't create tables/functions hashtables for uninteresting
5101                                  * databases.
5102                                  */
5103                                 if (onlydb != InvalidOid)
5104                                 {
5105                                         if (dbbuf.databaseid != onlydb &&
5106                                                 dbbuf.databaseid != InvalidOid)
5107                                                 break;
5108                                 }
5109
5110                                 memset(&hash_ctl, 0, sizeof(hash_ctl));
5111                                 hash_ctl.keysize = sizeof(Oid);
5112                                 hash_ctl.entrysize = sizeof(PgStat_StatTabEntry);
5113                                 hash_ctl.hcxt = pgStatLocalContext;
5114                                 dbentry->tables = hash_create("Per-database table",
5115                                                                                           PGSTAT_TAB_HASH_SIZE,
5116                                                                                           &hash_ctl,
5117                                                                                           HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
5118
5119                                 hash_ctl.keysize = sizeof(Oid);
5120                                 hash_ctl.entrysize = sizeof(PgStat_StatFuncEntry);
5121                                 hash_ctl.hcxt = pgStatLocalContext;
5122                                 dbentry->functions = hash_create("Per-database function",
5123                                                                                                  PGSTAT_FUNCTION_HASH_SIZE,
5124                                                                                                  &hash_ctl,
5125                                                                                                  HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
5126
5127                                 /*
5128                                  * If requested, read the data from the database-specific
5129                                  * file.  Otherwise we just leave the hashtables empty.
5130                                  */
5131                                 if (deep)
5132                                         pgstat_read_db_statsfile(dbentry->databaseid,
5133                                                                                          dbentry->tables,
5134                                                                                          dbentry->functions,
5135                                                                                          permanent);
5136
5137                                 break;
5138
5139                         case 'E':
5140                                 goto done;
5141
5142                         default:
5143                                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5144                                                 (errmsg("corrupted statistics file \"%s\"",
5145                                                                 statfile)));
5146                                 goto done;
5147                 }
5148         }
5149
5150 done:
5151         FreeFile(fpin);
5152
5153         /* If requested to read the permanent file, also get rid of it. */
5154         if (permanent)
5155         {
5156                 elog(DEBUG2, "removing permanent stats file \"%s\"", statfile);
5157                 unlink(statfile);
5158         }
5159
5160         return dbhash;
5161 }
5162
5163
5164 /* ----------
5165  * pgstat_read_db_statsfile() -
5166  *
5167  *      Reads in the existing statistics collector file for the given database,
5168  *      filling the passed-in tables and functions hash tables.
5169  *
5170  *      As in pgstat_read_statsfiles, if the permanent file is requested, it is
5171  *      removed after reading.
5172  *
5173  *      Note: this code has the ability to skip storing per-table or per-function
5174  *      data, if NULL is passed for the corresponding hashtable.  That's not used
5175  *      at the moment though.
5176  * ----------
5177  */
5178 static void
5179 pgstat_read_db_statsfile(Oid databaseid, HTAB *tabhash, HTAB *funchash,
5180                                                  bool permanent)
5181 {
5182         PgStat_StatTabEntry *tabentry;
5183         PgStat_StatTabEntry tabbuf;
5184         PgStat_StatFuncEntry funcbuf;
5185         PgStat_StatFuncEntry *funcentry;
5186         FILE       *fpin;
5187         int32           format_id;
5188         bool            found;
5189         char            statfile[MAXPGPATH];
5190
5191         get_dbstat_filename(permanent, false, databaseid, statfile, MAXPGPATH);
5192
5193         /*
5194          * Try to open the stats file. If it doesn't exist, the backends simply
5195          * return zero for anything and the collector simply starts from scratch
5196          * with empty counters.
5197          *
5198          * ENOENT is a possibility if the stats collector is not running or has
5199          * not yet written the stats file the first time.  Any other failure
5200          * condition is suspicious.
5201          */
5202         if ((fpin = AllocateFile(statfile, PG_BINARY_R)) == NULL)
5203         {
5204                 if (errno != ENOENT)
5205                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5206                                         (errcode_for_file_access(),
5207                                          errmsg("could not open statistics file \"%s\": %m",
5208                                                         statfile)));
5209                 return;
5210         }
5211
5212         /*
5213          * Verify it's of the expected format.
5214          */
5215         if (fread(&format_id, 1, sizeof(format_id), fpin) != sizeof(format_id) ||
5216                 format_id != PGSTAT_FILE_FORMAT_ID)
5217         {
5218                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5219                                 (errmsg("corrupted statistics file \"%s\"", statfile)));
5220                 goto done;
5221         }
5222
5223         /*
5224          * We found an existing collector stats file. Read it and put all the
5225          * hashtable entries into place.
5226          */
5227         for (;;)
5228         {
5229                 switch (fgetc(fpin))
5230                 {
5231                                 /*
5232                                  * 'T'  A PgStat_StatTabEntry follows.
5233                                  */
5234                         case 'T':
5235                                 if (fread(&tabbuf, 1, sizeof(PgStat_StatTabEntry),
5236                                                   fpin) != sizeof(PgStat_StatTabEntry))
5237                                 {
5238                                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5239                                                         (errmsg("corrupted statistics file \"%s\"",
5240                                                                         statfile)));
5241                                         goto done;
5242                                 }
5243
5244                                 /*
5245                                  * Skip if table data not wanted.
5246                                  */
5247                                 if (tabhash == NULL)
5248                                         break;
5249
5250                                 tabentry = (PgStat_StatTabEntry *) hash_search(tabhash,
5251                                                                                                                            (void *) &tabbuf.tableid,
5252                                                                                                                            HASH_ENTER, &found);
5253
5254                                 if (found)
5255                                 {
5256                                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5257                                                         (errmsg("corrupted statistics file \"%s\"",
5258                                                                         statfile)));
5259                                         goto done;
5260                                 }
5261
5262                                 memcpy(tabentry, &tabbuf, sizeof(tabbuf));
5263                                 break;
5264
5265                                 /*
5266                                  * 'F'  A PgStat_StatFuncEntry follows.
5267                                  */
5268                         case 'F':
5269                                 if (fread(&funcbuf, 1, sizeof(PgStat_StatFuncEntry),
5270                                                   fpin) != sizeof(PgStat_StatFuncEntry))
5271                                 {
5272                                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5273                                                         (errmsg("corrupted statistics file \"%s\"",
5274                                                                         statfile)));
5275                                         goto done;
5276                                 }
5277
5278                                 /*
5279                                  * Skip if function data not wanted.
5280                                  */
5281                                 if (funchash == NULL)
5282                                         break;
5283
5284                                 funcentry = (PgStat_StatFuncEntry *) hash_search(funchash,
5285                                                                                                                                  (void *) &funcbuf.functionid,
5286                                                                                                                                  HASH_ENTER, &found);
5287
5288                                 if (found)
5289                                 {
5290                                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5291                                                         (errmsg("corrupted statistics file \"%s\"",
5292                                                                         statfile)));
5293                                         goto done;
5294                                 }
5295
5296                                 memcpy(funcentry, &funcbuf, sizeof(funcbuf));
5297                                 break;
5298
5299                                 /*
5300                                  * 'E'  The EOF marker of a complete stats file.
5301                                  */
5302                         case 'E':
5303                                 goto done;
5304
5305                         default:
5306                                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5307                                                 (errmsg("corrupted statistics file \"%s\"",
5308                                                                 statfile)));
5309                                 goto done;
5310                 }
5311         }
5312
5313 done:
5314         FreeFile(fpin);
5315
5316         if (permanent)
5317         {
5318                 elog(DEBUG2, "removing permanent stats file \"%s\"", statfile);
5319                 unlink(statfile);
5320         }
5321 }
5322
5323 /* ----------
5324  * pgstat_read_db_statsfile_timestamp() -
5325  *
5326  *      Attempt to determine the timestamp of the last db statfile write.
5327  *      Returns true if successful; the timestamp is stored in *ts.
5328  *
5329  *      This needs to be careful about handling databases for which no stats file
5330  *      exists, such as databases without a stat entry or those not yet written:
5331  *
5332  *      - if there's a database entry in the global file, return the corresponding
5333  *      stats_timestamp value.
5334  *
5335  *      - if there's no db stat entry (e.g. for a new or inactive database),
5336  *      there's no stats_timestamp value, but also nothing to write so we return
5337  *      the timestamp of the global statfile.
5338  * ----------
5339  */
5340 static bool
5341 pgstat_read_db_statsfile_timestamp(Oid databaseid, bool permanent,
5342                                                                    TimestampTz *ts)
5343 {
5344         PgStat_StatDBEntry dbentry;
5345         PgStat_GlobalStats myGlobalStats;
5346         PgStat_ArchiverStats myArchiverStats;
5347         FILE       *fpin;
5348         int32           format_id;
5349         const char *statfile = permanent ? PGSTAT_STAT_PERMANENT_FILENAME : pgstat_stat_filename;
5350
5351         /*
5352          * Try to open the stats file.  As above, anything but ENOENT is worthy of
5353          * complaining about.
5354          */
5355         if ((fpin = AllocateFile(statfile, PG_BINARY_R)) == NULL)
5356         {
5357                 if (errno != ENOENT)
5358                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5359                                         (errcode_for_file_access(),
5360                                          errmsg("could not open statistics file \"%s\": %m",
5361                                                         statfile)));
5362                 return false;
5363         }
5364
5365         /*
5366          * Verify it's of the expected format.
5367          */
5368         if (fread(&format_id, 1, sizeof(format_id), fpin) != sizeof(format_id) ||
5369                 format_id != PGSTAT_FILE_FORMAT_ID)
5370         {
5371                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5372                                 (errmsg("corrupted statistics file \"%s\"", statfile)));
5373                 FreeFile(fpin);
5374                 return false;
5375         }
5376
5377         /*
5378          * Read global stats struct
5379          */
5380         if (fread(&myGlobalStats, 1, sizeof(myGlobalStats),
5381                           fpin) != sizeof(myGlobalStats))
5382         {
5383                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5384                                 (errmsg("corrupted statistics file \"%s\"", statfile)));
5385                 FreeFile(fpin);
5386                 return false;
5387         }
5388
5389         /*
5390          * Read archiver stats struct
5391          */
5392         if (fread(&myArchiverStats, 1, sizeof(myArchiverStats),
5393                           fpin) != sizeof(myArchiverStats))
5394         {
5395                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5396                                 (errmsg("corrupted statistics file \"%s\"", statfile)));
5397                 FreeFile(fpin);
5398                 return false;
5399         }
5400
5401         /* By default, we're going to return the timestamp of the global file. */
5402         *ts = myGlobalStats.stats_timestamp;
5403
5404         /*
5405          * We found an existing collector stats file.  Read it and look for a
5406          * record for the requested database.  If found, use its timestamp.
5407          */
5408         for (;;)
5409         {
5410                 switch (fgetc(fpin))
5411                 {
5412                                 /*
5413                                  * 'D'  A PgStat_StatDBEntry struct describing a database
5414                                  * follows.
5415                                  */
5416                         case 'D':
5417                                 if (fread(&dbentry, 1, offsetof(PgStat_StatDBEntry, tables),
5418                                                   fpin) != offsetof(PgStat_StatDBEntry, tables))
5419                                 {
5420                                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5421                                                         (errmsg("corrupted statistics file \"%s\"",
5422                                                                         statfile)));
5423                                         goto done;
5424                                 }
5425
5426                                 /*
5427                                  * If this is the DB we're looking for, save its timestamp and
5428                                  * we're done.
5429                                  */
5430                                 if (dbentry.databaseid == databaseid)
5431                                 {
5432                                         *ts = dbentry.stats_timestamp;
5433                                         goto done;
5434                                 }
5435
5436                                 break;
5437
5438                         case 'E':
5439                                 goto done;
5440
5441                         default:
5442                                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5443                                                 (errmsg("corrupted statistics file \"%s\"",
5444                                                                 statfile)));
5445                                 goto done;
5446                 }
5447         }
5448
5449 done:
5450         FreeFile(fpin);
5451         return true;
5452 }
5453
5454 /*
5455  * If not already done, read the statistics collector stats file into
5456  * some hash tables.  The results will be kept until pgstat_clear_snapshot()
5457  * is called (typically, at end of transaction).
5458  */
5459 static void
5460 backend_read_statsfile(void)
5461 {
5462         TimestampTz min_ts = 0;
5463         TimestampTz ref_ts = 0;
5464         Oid                     inquiry_db;
5465         int                     count;
5466
5467         /* already read it? */
5468         if (pgStatDBHash)
5469                 return;
5470         Assert(!pgStatRunningInCollector);
5471
5472         /*
5473          * In a normal backend, we check staleness of the data for our own DB, and
5474          * so we send MyDatabaseId in inquiry messages.  In the autovac launcher,
5475          * check staleness of the shared-catalog data, and send InvalidOid in
5476          * inquiry messages so as not to force writing unnecessary data.
5477          */
5478         if (IsAutoVacuumLauncherProcess())
5479                 inquiry_db = InvalidOid;
5480         else
5481                 inquiry_db = MyDatabaseId;
5482
5483         /*
5484          * Loop until fresh enough stats file is available or we ran out of time.
5485          * The stats inquiry message is sent repeatedly in case collector drops
5486          * it; but not every single time, as that just swamps the collector.
5487          */
5488         for (count = 0; count < PGSTAT_POLL_LOOP_COUNT; count++)
5489         {
5490                 bool            ok;
5491                 TimestampTz file_ts = 0;
5492                 TimestampTz cur_ts;
5493
5494                 CHECK_FOR_INTERRUPTS();
5495
5496                 ok = pgstat_read_db_statsfile_timestamp(inquiry_db, false, &file_ts);
5497
5498                 cur_ts = GetCurrentTimestamp();
5499                 /* Calculate min acceptable timestamp, if we didn't already */
5500                 if (count == 0 || cur_ts < ref_ts)
5501                 {
5502                         /*
5503                          * We set the minimum acceptable timestamp to PGSTAT_STAT_INTERVAL
5504                          * msec before now.  This indirectly ensures that the collector
5505                          * needn't write the file more often than PGSTAT_STAT_INTERVAL. In
5506                          * an autovacuum worker, however, we want a lower delay to avoid
5507                          * using stale data, so we use PGSTAT_RETRY_DELAY (since the
5508                          * number of workers is low, this shouldn't be a problem).
5509                          *
5510                          * We don't recompute min_ts after sleeping, except in the
5511                          * unlikely case that cur_ts went backwards.  So we might end up
5512                          * accepting a file a bit older than PGSTAT_STAT_INTERVAL.  In
5513                          * practice that shouldn't happen, though, as long as the sleep
5514                          * time is less than PGSTAT_STAT_INTERVAL; and we don't want to
5515                          * tell the collector that our cutoff time is less than what we'd
5516                          * actually accept.
5517                          */
5518                         ref_ts = cur_ts;
5519                         if (IsAutoVacuumWorkerProcess())
5520                                 min_ts = TimestampTzPlusMilliseconds(ref_ts,
5521                                                                                                          -PGSTAT_RETRY_DELAY);
5522                         else
5523                                 min_ts = TimestampTzPlusMilliseconds(ref_ts,
5524                                                                                                          -PGSTAT_STAT_INTERVAL);
5525                 }
5526
5527                 /*
5528                  * If the file timestamp is actually newer than cur_ts, we must have
5529                  * had a clock glitch (system time went backwards) or there is clock
5530                  * skew between our processor and the stats collector's processor.
5531                  * Accept the file, but send an inquiry message anyway to make
5532                  * pgstat_recv_inquiry do a sanity check on the collector's time.
5533                  */
5534                 if (ok && file_ts > cur_ts)
5535                 {
5536                         /*
5537                          * A small amount of clock skew between processors isn't terribly
5538                          * surprising, but a large difference is worth logging.  We
5539                          * arbitrarily define "large" as 1000 msec.
5540                          */
5541                         if (file_ts >= TimestampTzPlusMilliseconds(cur_ts, 1000))
5542                         {
5543                                 char       *filetime;
5544                                 char       *mytime;
5545
5546                                 /* Copy because timestamptz_to_str returns a static buffer */
5547                                 filetime = pstrdup(timestamptz_to_str(file_ts));
5548                                 mytime = pstrdup(timestamptz_to_str(cur_ts));
5549                                 elog(LOG, "stats collector's time %s is later than backend local time %s",
5550                                          filetime, mytime);
5551                                 pfree(filetime);
5552                                 pfree(mytime);
5553                         }
5554
5555                         pgstat_send_inquiry(cur_ts, min_ts, inquiry_db);
5556                         break;
5557                 }
5558
5559                 /* Normal acceptance case: file is not older than cutoff time */
5560                 if (ok && file_ts >= min_ts)
5561                         break;
5562
5563                 /* Not there or too old, so kick the collector and wait a bit */
5564                 if ((count % PGSTAT_INQ_LOOP_COUNT) == 0)
5565                         pgstat_send_inquiry(cur_ts, min_ts, inquiry_db);
5566
5567                 pg_usleep(PGSTAT_RETRY_DELAY * 1000L);
5568         }
5569
5570         if (count >= PGSTAT_POLL_LOOP_COUNT)
5571                 ereport(LOG,
5572                                 (errmsg("using stale statistics instead of current ones "
5573                                                 "because stats collector is not responding")));
5574
5575         /*
5576          * Autovacuum launcher wants stats about all databases, but a shallow read
5577          * is sufficient.  Regular backends want a deep read for just the tables
5578          * they can see (MyDatabaseId + shared catalogs).
5579          */
5580         if (IsAutoVacuumLauncherProcess())
5581                 pgStatDBHash = pgstat_read_statsfiles(InvalidOid, false, false);
5582         else
5583                 pgStatDBHash = pgstat_read_statsfiles(MyDatabaseId, false, true);
5584 }
5585
5586
5587 /* ----------
5588  * pgstat_setup_memcxt() -
5589  *
5590  *      Create pgStatLocalContext, if not already done.
5591  * ----------
5592  */
5593 static void
5594 pgstat_setup_memcxt(void)
5595 {
5596         if (!pgStatLocalContext)
5597                 pgStatLocalContext = AllocSetContextCreate(TopMemoryContext,
5598                                                                                                    "Statistics snapshot",
5599                                                                                                    ALLOCSET_SMALL_SIZES);
5600 }
5601
5602
5603 /* ----------
5604  * pgstat_clear_snapshot() -
5605  *
5606  *      Discard any data collected in the current transaction.  Any subsequent
5607  *      request will cause new snapshots to be read.
5608  *
5609  *      This is also invoked during transaction commit or abort to discard
5610  *      the no-longer-wanted snapshot.
5611  * ----------
5612  */
5613 void
5614 pgstat_clear_snapshot(void)
5615 {
5616         /* Release memory, if any was allocated */
5617         if (pgStatLocalContext)
5618                 MemoryContextDelete(pgStatLocalContext);
5619
5620         /* Reset variables */
5621         pgStatLocalContext = NULL;
5622         pgStatDBHash = NULL;
5623         localBackendStatusTable = NULL;
5624         localNumBackends = 0;
5625 }
5626
5627
5628 /* ----------
5629  * pgstat_recv_inquiry() -
5630  *
5631  *      Process stat inquiry requests.
5632  * ----------
5633  */
5634 static void
5635 pgstat_recv_inquiry(PgStat_MsgInquiry *msg, int len)
5636 {
5637         PgStat_StatDBEntry *dbentry;
5638
5639         elog(DEBUG2, "received inquiry for database %u", msg->databaseid);
5640
5641         /*
5642          * If there's already a write request for this DB, there's nothing to do.
5643          *
5644          * Note that if a request is found, we return early and skip the below
5645          * check for clock skew.  This is okay, since the only way for a DB
5646          * request to be present in the list is that we have been here since the
5647          * last write round.  It seems sufficient to check for clock skew once per
5648          * write round.
5649          */
5650         if (list_member_oid(pending_write_requests, msg->databaseid))
5651                 return;
5652
5653         /*
5654          * Check to see if we last wrote this database at a time >= the requested
5655          * cutoff time.  If so, this is a stale request that was generated before
5656          * we updated the DB file, and we don't need to do so again.
5657          *
5658          * If the requestor's local clock time is older than stats_timestamp, we
5659          * should suspect a clock glitch, ie system time going backwards; though
5660          * the more likely explanation is just delayed message receipt.  It is
5661          * worth expending a GetCurrentTimestamp call to be sure, since a large
5662          * retreat in the system clock reading could otherwise cause us to neglect
5663          * to update the stats file for a long time.
5664          */
5665         dbentry = pgstat_get_db_entry(msg->databaseid, false);
5666         if (dbentry == NULL)
5667         {
5668                 /*
5669                  * We have no data for this DB.  Enter a write request anyway so that
5670                  * the global stats will get updated.  This is needed to prevent
5671                  * backend_read_statsfile from waiting for data that we cannot supply,
5672                  * in the case of a new DB that nobody has yet reported any stats for.
5673                  * See the behavior of pgstat_read_db_statsfile_timestamp.
5674                  */
5675         }
5676         else if (msg->clock_time < dbentry->stats_timestamp)
5677         {
5678                 TimestampTz cur_ts = GetCurrentTimestamp();
5679
5680                 if (cur_ts < dbentry->stats_timestamp)
5681                 {
5682                         /*
5683                          * Sure enough, time went backwards.  Force a new stats file write
5684                          * to get back in sync; but first, log a complaint.
5685                          */
5686                         char       *writetime;
5687                         char       *mytime;
5688
5689                         /* Copy because timestamptz_to_str returns a static buffer */
5690                         writetime = pstrdup(timestamptz_to_str(dbentry->stats_timestamp));
5691                         mytime = pstrdup(timestamptz_to_str(cur_ts));
5692                         elog(LOG,
5693                                  "stats_timestamp %s is later than collector's time %s for database %u",
5694                                  writetime, mytime, dbentry->databaseid);
5695                         pfree(writetime);
5696                         pfree(mytime);
5697                 }
5698                 else
5699                 {
5700                         /*
5701                          * Nope, it's just an old request.  Assuming msg's clock_time is
5702                          * >= its cutoff_time, it must be stale, so we can ignore it.
5703                          */
5704                         return;
5705                 }
5706         }
5707         else if (msg->cutoff_time <= dbentry->stats_timestamp)
5708         {
5709                 /* Stale request, ignore it */
5710                 return;
5711         }
5712
5713         /*
5714          * We need to write this DB, so create a request.
5715          */
5716         pending_write_requests = lappend_oid(pending_write_requests,
5717                                                                                  msg->databaseid);
5718 }
5719
5720
5721 /* ----------
5722  * pgstat_recv_tabstat() -
5723  *
5724  *      Count what the backend has done.
5725  * ----------
5726  */
5727 static void
5728 pgstat_recv_tabstat(PgStat_MsgTabstat *msg, int len)
5729 {
5730         PgStat_StatDBEntry *dbentry;
5731         PgStat_StatTabEntry *tabentry;
5732         int                     i;
5733         bool            found;
5734
5735         dbentry = pgstat_get_db_entry(msg->m_databaseid, true);
5736
5737         /*
5738          * Update database-wide stats.
5739          */
5740         dbentry->n_xact_commit += (PgStat_Counter) (msg->m_xact_commit);
5741         dbentry->n_xact_rollback += (PgStat_Counter) (msg->m_xact_rollback);
5742         dbentry->n_block_read_time += msg->m_block_read_time;
5743         dbentry->n_block_write_time += msg->m_block_write_time;
5744
5745         /*
5746          * Process all table entries in the message.
5747          */
5748         for (i = 0; i < msg->m_nentries; i++)
5749         {
5750                 PgStat_TableEntry *tabmsg = &(msg->m_entry[i]);
5751
5752                 tabentry = (PgStat_StatTabEntry *) hash_search(dbentry->tables,
5753                                                                                                            (void *) &(tabmsg->t_id),
5754                                                                                                            HASH_ENTER, &found);
5755
5756                 if (!found)
5757                 {
5758                         /*
5759                          * If it's a new table entry, initialize counters to the values we
5760                          * just got.
5761                          */
5762                         tabentry->numscans = tabmsg->t_counts.t_numscans;
5763                         tabentry->tuples_returned = tabmsg->t_counts.t_tuples_returned;
5764                         tabentry->tuples_fetched = tabmsg->t_counts.t_tuples_fetched;
5765                         tabentry->tuples_inserted = tabmsg->t_counts.t_tuples_inserted;
5766                         tabentry->tuples_updated = tabmsg->t_counts.t_tuples_updated;
5767                         tabentry->tuples_deleted = tabmsg->t_counts.t_tuples_deleted;
5768                         tabentry->tuples_hot_updated = tabmsg->t_counts.t_tuples_hot_updated;
5769                         tabentry->n_live_tuples = tabmsg->t_counts.t_delta_live_tuples;
5770                         tabentry->n_dead_tuples = tabmsg->t_counts.t_delta_dead_tuples;
5771                         tabentry->changes_since_analyze = tabmsg->t_counts.t_changed_tuples;
5772                         tabentry->blocks_fetched = tabmsg->t_counts.t_blocks_fetched;
5773                         tabentry->blocks_hit = tabmsg->t_counts.t_blocks_hit;
5774
5775                         tabentry->vacuum_timestamp = 0;
5776                         tabentry->vacuum_count = 0;
5777                         tabentry->autovac_vacuum_timestamp = 0;
5778                         tabentry->autovac_vacuum_count = 0;
5779                         tabentry->analyze_timestamp = 0;
5780                         tabentry->analyze_count = 0;
5781                         tabentry->autovac_analyze_timestamp = 0;
5782                         tabentry->autovac_analyze_count = 0;
5783                 }
5784                 else
5785                 {
5786                         /*
5787                          * Otherwise add the values to the existing entry.
5788                          */
5789                         tabentry->numscans += tabmsg->t_counts.t_numscans;
5790                         tabentry->tuples_returned += tabmsg->t_counts.t_tuples_returned;
5791                         tabentry->tuples_fetched += tabmsg->t_counts.t_tuples_fetched;
5792                         tabentry->tuples_inserted += tabmsg->t_counts.t_tuples_inserted;
5793                         tabentry->tuples_updated += tabmsg->t_counts.t_tuples_updated;
5794                         tabentry->tuples_deleted += tabmsg->t_counts.t_tuples_deleted;
5795                         tabentry->tuples_hot_updated += tabmsg->t_counts.t_tuples_hot_updated;
5796                         /* If table was truncated, first reset the live/dead counters */
5797                         if (tabmsg->t_counts.t_truncated)
5798                         {
5799                                 tabentry->n_live_tuples = 0;
5800                                 tabentry->n_dead_tuples = 0;
5801                         }
5802                         tabentry->n_live_tuples += tabmsg->t_counts.t_delta_live_tuples;
5803                         tabentry->n_dead_tuples += tabmsg->t_counts.t_delta_dead_tuples;
5804                         tabentry->changes_since_analyze += tabmsg->t_counts.t_changed_tuples;
5805                         tabentry->blocks_fetched += tabmsg->t_counts.t_blocks_fetched;
5806                         tabentry->blocks_hit += tabmsg->t_counts.t_blocks_hit;
5807                 }
5808
5809                 /* Clamp n_live_tuples in case of negative delta_live_tuples */
5810                 tabentry->n_live_tuples = Max(tabentry->n_live_tuples, 0);
5811                 /* Likewise for n_dead_tuples */
5812                 tabentry->n_dead_tuples = Max(tabentry->n_dead_tuples, 0);
5813
5814                 /*
5815                  * Add per-table stats to the per-database entry, too.
5816                  */
5817                 dbentry->n_tuples_returned += tabmsg->t_counts.t_tuples_returned;
5818                 dbentry->n_tuples_fetched += tabmsg->t_counts.t_tuples_fetched;
5819                 dbentry->n_tuples_inserted += tabmsg->t_counts.t_tuples_inserted;
5820                 dbentry->n_tuples_updated += tabmsg->t_counts.t_tuples_updated;
5821                 dbentry->n_tuples_deleted += tabmsg->t_counts.t_tuples_deleted;
5822                 dbentry->n_blocks_fetched += tabmsg->t_counts.t_blocks_fetched;
5823                 dbentry->n_blocks_hit += tabmsg->t_counts.t_blocks_hit;
5824         }
5825 }
5826
5827
5828 /* ----------
5829  * pgstat_recv_tabpurge() -
5830  *
5831  *      Arrange for dead table removal.
5832  * ----------
5833  */
5834 static void
5835 pgstat_recv_tabpurge(PgStat_MsgTabpurge *msg, int len)
5836 {
5837         PgStat_StatDBEntry *dbentry;
5838         int                     i;
5839
5840         dbentry = pgstat_get_db_entry(msg->m_databaseid, false);
5841
5842         /*
5843          * No need to purge if we don't even know the database.
5844          */
5845         if (!dbentry || !dbentry->tables)
5846                 return;
5847
5848         /*
5849          * Process all table entries in the message.
5850          */
5851         for (i = 0; i < msg->m_nentries; i++)
5852         {
5853                 /* Remove from hashtable if present; we don't care if it's not. */
5854                 (void) hash_search(dbentry->tables,
5855                                                    (void *) &(msg->m_tableid[i]),
5856                                                    HASH_REMOVE, NULL);
5857         }
5858 }
5859
5860
5861 /* ----------
5862  * pgstat_recv_dropdb() -
5863  *
5864  *      Arrange for dead database removal
5865  * ----------
5866  */
5867 static void
5868 pgstat_recv_dropdb(PgStat_MsgDropdb *msg, int len)
5869 {
5870         Oid                     dbid = msg->m_databaseid;
5871         PgStat_StatDBEntry *dbentry;
5872
5873         /*
5874          * Lookup the database in the hashtable.
5875          */
5876         dbentry = pgstat_get_db_entry(dbid, false);
5877
5878         /*
5879          * If found, remove it (along with the db statfile).
5880          */
5881         if (dbentry)
5882         {
5883                 char            statfile[MAXPGPATH];
5884
5885                 get_dbstat_filename(false, false, dbid, statfile, MAXPGPATH);
5886
5887                 elog(DEBUG2, "removing stats file \"%s\"", statfile);
5888                 unlink(statfile);
5889
5890                 if (dbentry->tables != NULL)
5891                         hash_destroy(dbentry->tables);
5892                 if (dbentry->functions != NULL)
5893                         hash_destroy(dbentry->functions);
5894
5895                 if (hash_search(pgStatDBHash,
5896                                                 (void *) &dbid,
5897                                                 HASH_REMOVE, NULL) == NULL)
5898                         ereport(ERROR,
5899                                         (errmsg("database hash table corrupted during cleanup --- abort")));
5900         }
5901 }
5902
5903
5904 /* ----------
5905  * pgstat_recv_resetcounter() -
5906  *
5907  *      Reset the statistics for the specified database.
5908  * ----------
5909  */
5910 static void
5911 pgstat_recv_resetcounter(PgStat_MsgResetcounter *msg, int len)
5912 {
5913         PgStat_StatDBEntry *dbentry;
5914
5915         /*
5916          * Lookup the database in the hashtable.  Nothing to do if not there.
5917          */
5918         dbentry = pgstat_get_db_entry(msg->m_databaseid, false);
5919
5920         if (!dbentry)
5921                 return;
5922
5923         /*
5924          * We simply throw away all the database's table entries by recreating a
5925          * new hash table for them.
5926          */
5927         if (dbentry->tables != NULL)
5928                 hash_destroy(dbentry->tables);
5929         if (dbentry->functions != NULL)
5930                 hash_destroy(dbentry->functions);
5931
5932         dbentry->tables = NULL;
5933         dbentry->functions = NULL;
5934
5935         /*
5936          * Reset database-level stats, too.  This creates empty hash tables for
5937          * tables and functions.
5938          */
5939         reset_dbentry_counters(dbentry);
5940 }
5941
5942 /* ----------
5943  * pgstat_recv_resetshared() -
5944  *
5945  *      Reset some shared statistics of the cluster.
5946  * ----------
5947  */
5948 static void
5949 pgstat_recv_resetsharedcounter(PgStat_MsgResetsharedcounter *msg, int len)
5950 {
5951         if (msg->m_resettarget == RESET_BGWRITER)
5952         {
5953                 /* Reset the global background writer statistics for the cluster. */
5954                 memset(&globalStats, 0, sizeof(globalStats));
5955                 globalStats.stat_reset_timestamp = GetCurrentTimestamp();
5956         }
5957         else if (msg->m_resettarget == RESET_ARCHIVER)
5958         {
5959                 /* Reset the archiver statistics for the cluster. */
5960                 memset(&archiverStats, 0, sizeof(archiverStats));
5961                 archiverStats.stat_reset_timestamp = GetCurrentTimestamp();
5962         }
5963
5964         /*
5965          * Presumably the sender of this message validated the target, don't
5966          * complain here if it's not valid
5967          */
5968 }
5969
5970 /* ----------
5971  * pgstat_recv_resetsinglecounter() -
5972  *
5973  *      Reset a statistics for a single object
5974  * ----------
5975  */
5976 static void
5977 pgstat_recv_resetsinglecounter(PgStat_MsgResetsinglecounter *msg, int len)
5978 {
5979         PgStat_StatDBEntry *dbentry;
5980
5981         dbentry = pgstat_get_db_entry(msg->m_databaseid, false);
5982
5983         if (!dbentry)
5984                 return;
5985
5986         /* Set the reset timestamp for the whole database */
5987         dbentry->stat_reset_timestamp = GetCurrentTimestamp();
5988
5989         /* Remove object if it exists, ignore it if not */
5990         if (msg->m_resettype == RESET_TABLE)
5991                 (void) hash_search(dbentry->tables, (void *) &(msg->m_objectid),
5992                                                    HASH_REMOVE, NULL);
5993         else if (msg->m_resettype == RESET_FUNCTION)
5994                 (void) hash_search(dbentry->functions, (void *) &(msg->m_objectid),
5995                                                    HASH_REMOVE, NULL);
5996 }
5997
5998 /* ----------
5999  * pgstat_recv_autovac() -
6000  *
6001  *      Process an autovacuum signalling message.
6002  * ----------
6003  */
6004 static void
6005 pgstat_recv_autovac(PgStat_MsgAutovacStart *msg, int len)
6006 {
6007         PgStat_StatDBEntry *dbentry;
6008
6009         /*
6010          * Store the last autovacuum time in the database's hashtable entry.
6011          */
6012         dbentry = pgstat_get_db_entry(msg->m_databaseid, true);
6013
6014         dbentry->last_autovac_time = msg->m_start_time;
6015 }
6016
6017 /* ----------
6018  * pgstat_recv_vacuum() -
6019  *
6020  *      Process a VACUUM message.
6021  * ----------
6022  */
6023 static void
6024 pgstat_recv_vacuum(PgStat_MsgVacuum *msg, int len)
6025 {
6026         PgStat_StatDBEntry *dbentry;
6027         PgStat_StatTabEntry *tabentry;
6028
6029         /*
6030          * Store the data in the table's hashtable entry.
6031          */
6032         dbentry = pgstat_get_db_entry(msg->m_databaseid, true);
6033
6034         tabentry = pgstat_get_tab_entry(dbentry, msg->m_tableoid, true);
6035
6036         tabentry->n_live_tuples = msg->m_live_tuples;
6037         tabentry->n_dead_tuples = msg->m_dead_tuples;
6038
6039         if (msg->m_autovacuum)
6040         {
6041                 tabentry->autovac_vacuum_timestamp = msg->m_vacuumtime;
6042                 tabentry->autovac_vacuum_count++;
6043         }
6044         else
6045         {
6046                 tabentry->vacuum_timestamp = msg->m_vacuumtime;
6047                 tabentry->vacuum_count++;
6048         }
6049 }
6050
6051 /* ----------
6052  * pgstat_recv_analyze() -
6053  *
6054  *      Process an ANALYZE message.
6055  * ----------
6056  */
6057 static void
6058 pgstat_recv_analyze(PgStat_MsgAnalyze *msg, int len)
6059 {
6060         PgStat_StatDBEntry *dbentry;
6061         PgStat_StatTabEntry *tabentry;
6062
6063         /*
6064          * Store the data in the table's hashtable entry.
6065          */
6066         dbentry = pgstat_get_db_entry(msg->m_databaseid, true);
6067
6068         tabentry = pgstat_get_tab_entry(dbentry, msg->m_tableoid, true);
6069
6070         tabentry->n_live_tuples = msg->m_live_tuples;
6071         tabentry->n_dead_tuples = msg->m_dead_tuples;
6072
6073         /*
6074          * If commanded, reset changes_since_analyze to zero.  This forgets any
6075          * changes that were committed while the ANALYZE was in progress, but we
6076          * have no good way to estimate how many of those there were.
6077          */
6078         if (msg->m_resetcounter)
6079                 tabentry->changes_since_analyze = 0;
6080
6081         if (msg->m_autovacuum)
6082         {
6083                 tabentry->autovac_analyze_timestamp = msg->m_analyzetime;
6084                 tabentry->autovac_analyze_count++;
6085         }
6086         else
6087         {
6088                 tabentry->analyze_timestamp = msg->m_analyzetime;
6089                 tabentry->analyze_count++;
6090         }
6091 }
6092
6093
6094 /* ----------
6095  * pgstat_recv_archiver() -
6096  *
6097  *      Process a ARCHIVER message.
6098  * ----------
6099  */
6100 static void
6101 pgstat_recv_archiver(PgStat_MsgArchiver *msg, int len)
6102 {
6103         if (msg->m_failed)
6104         {
6105                 /* Failed archival attempt */
6106                 ++archiverStats.failed_count;
6107                 memcpy(archiverStats.last_failed_wal, msg->m_xlog,
6108                            sizeof(archiverStats.last_failed_wal));
6109                 archiverStats.last_failed_timestamp = msg->m_timestamp;
6110         }
6111         else
6112         {
6113                 /* Successful archival operation */
6114                 ++archiverStats.archived_count;
6115                 memcpy(archiverStats.last_archived_wal, msg->m_xlog,
6116                            sizeof(archiverStats.last_archived_wal));
6117                 archiverStats.last_archived_timestamp = msg->m_timestamp;
6118         }
6119 }
6120
6121 /* ----------
6122  * pgstat_recv_bgwriter() -
6123  *
6124  *      Process a BGWRITER message.
6125  * ----------
6126  */
6127 static void
6128 pgstat_recv_bgwriter(PgStat_MsgBgWriter *msg, int len)
6129 {
6130         globalStats.timed_checkpoints += msg->m_timed_checkpoints;
6131         globalStats.requested_checkpoints += msg->m_requested_checkpoints;
6132         globalStats.checkpoint_write_time += msg->m_checkpoint_write_time;
6133         globalStats.checkpoint_sync_time += msg->m_checkpoint_sync_time;
6134         globalStats.buf_written_checkpoints += msg->m_buf_written_checkpoints;
6135         globalStats.buf_written_clean += msg->m_buf_written_clean;
6136         globalStats.maxwritten_clean += msg->m_maxwritten_clean;
6137         globalStats.buf_written_backend += msg->m_buf_written_backend;
6138         globalStats.buf_fsync_backend += msg->m_buf_fsync_backend;
6139         globalStats.buf_alloc += msg->m_buf_alloc;
6140 }
6141
6142 /* ----------
6143  * pgstat_recv_recoveryconflict() -
6144  *
6145  *      Process a RECOVERYCONFLICT message.
6146  * ----------
6147  */
6148 static void
6149 pgstat_recv_recoveryconflict(PgStat_MsgRecoveryConflict *msg, int len)
6150 {
6151         PgStat_StatDBEntry *dbentry;
6152
6153         dbentry = pgstat_get_db_entry(msg->m_databaseid, true);
6154
6155         switch (msg->m_reason)
6156         {
6157                 case PROCSIG_RECOVERY_CONFLICT_DATABASE:
6158
6159                         /*
6160                          * Since we drop the information about the database as soon as it
6161                          * replicates, there is no point in counting these conflicts.
6162                          */
6163                         break;
6164                 case PROCSIG_RECOVERY_CONFLICT_TABLESPACE:
6165                         dbentry->n_conflict_tablespace++;
6166                         break;
6167                 case PROCSIG_RECOVERY_CONFLICT_LOCK:
6168                         dbentry->n_conflict_lock++;
6169                         break;
6170                 case PROCSIG_RECOVERY_CONFLICT_SNAPSHOT:
6171                         dbentry->n_conflict_snapshot++;
6172                         break;
6173                 case PROCSIG_RECOVERY_CONFLICT_BUFFERPIN:
6174                         dbentry->n_conflict_bufferpin++;
6175                         break;
6176                 case PROCSIG_RECOVERY_CONFLICT_STARTUP_DEADLOCK:
6177                         dbentry->n_conflict_startup_deadlock++;
6178                         break;
6179         }
6180 }
6181
6182 /* ----------
6183  * pgstat_recv_deadlock() -
6184  *
6185  *      Process a DEADLOCK message.
6186  * ----------
6187  */
6188 static void
6189 pgstat_recv_deadlock(PgStat_MsgDeadlock *msg, int len)
6190 {
6191         PgStat_StatDBEntry *dbentry;
6192
6193         dbentry = pgstat_get_db_entry(msg->m_databaseid, true);
6194
6195         dbentry->n_deadlocks++;
6196 }
6197
6198 /* ----------
6199  * pgstat_recv_tempfile() -
6200  *
6201  *      Process a TEMPFILE message.
6202  * ----------
6203  */
6204 static void
6205 pgstat_recv_tempfile(PgStat_MsgTempFile *msg, int len)
6206 {
6207         PgStat_StatDBEntry *dbentry;
6208
6209         dbentry = pgstat_get_db_entry(msg->m_databaseid, true);
6210
6211         dbentry->n_temp_bytes += msg->m_filesize;
6212         dbentry->n_temp_files += 1;
6213 }
6214
6215 /* ----------
6216  * pgstat_recv_funcstat() -
6217  *
6218  *      Count what the backend has done.
6219  * ----------
6220  */
6221 static void
6222 pgstat_recv_funcstat(PgStat_MsgFuncstat *msg, int len)
6223 {
6224         PgStat_FunctionEntry *funcmsg = &(msg->m_entry[0]);
6225         PgStat_StatDBEntry *dbentry;
6226         PgStat_StatFuncEntry *funcentry;
6227         int                     i;
6228         bool            found;
6229
6230         dbentry = pgstat_get_db_entry(msg->m_databaseid, true);
6231
6232         /*
6233          * Process all function entries in the message.
6234          */
6235         for (i = 0; i < msg->m_nentries; i++, funcmsg++)
6236         {
6237                 funcentry = (PgStat_StatFuncEntry *) hash_search(dbentry->functions,
6238                                                                                                                  (void *) &(funcmsg->f_id),
6239                                                                                                                  HASH_ENTER, &found);
6240
6241                 if (!found)
6242                 {
6243                         /*
6244                          * If it's a new function entry, initialize counters to the values
6245                          * we just got.
6246                          */
6247                         funcentry->f_numcalls = funcmsg->f_numcalls;
6248                         funcentry->f_total_time = funcmsg->f_total_time;
6249                         funcentry->f_self_time = funcmsg->f_self_time;
6250                 }
6251                 else
6252                 {
6253                         /*
6254                          * Otherwise add the values to the existing entry.
6255                          */
6256                         funcentry->f_numcalls += funcmsg->f_numcalls;
6257                         funcentry->f_total_time += funcmsg->f_total_time;
6258                         funcentry->f_self_time += funcmsg->f_self_time;
6259                 }
6260         }
6261 }
6262
6263 /* ----------
6264  * pgstat_recv_funcpurge() -
6265  *
6266  *      Arrange for dead function removal.
6267  * ----------
6268  */
6269 static void
6270 pgstat_recv_funcpurge(PgStat_MsgFuncpurge *msg, int len)
6271 {
6272         PgStat_StatDBEntry *dbentry;
6273         int                     i;
6274
6275         dbentry = pgstat_get_db_entry(msg->m_databaseid, false);
6276
6277         /*
6278          * No need to purge if we don't even know the database.
6279          */
6280         if (!dbentry || !dbentry->functions)
6281                 return;
6282
6283         /*
6284          * Process all function entries in the message.
6285          */
6286         for (i = 0; i < msg->m_nentries; i++)
6287         {
6288                 /* Remove from hashtable if present; we don't care if it's not. */
6289                 (void) hash_search(dbentry->functions,
6290                                                    (void *) &(msg->m_functionid[i]),
6291                                                    HASH_REMOVE, NULL);
6292         }
6293 }
6294
6295 /* ----------
6296  * pgstat_write_statsfile_needed() -
6297  *
6298  *      Do we need to write out any stats files?
6299  * ----------
6300  */
6301 static bool
6302 pgstat_write_statsfile_needed(void)
6303 {
6304         if (pending_write_requests != NIL)
6305                 return true;
6306
6307         /* Everything was written recently */
6308         return false;
6309 }
6310
6311 /* ----------
6312  * pgstat_db_requested() -
6313  *
6314  *      Checks whether stats for a particular DB need to be written to a file.
6315  * ----------
6316  */
6317 static bool
6318 pgstat_db_requested(Oid databaseid)
6319 {
6320         /*
6321          * If any requests are outstanding at all, we should write the stats for
6322          * shared catalogs (the "database" with OID 0).  This ensures that
6323          * backends will see up-to-date stats for shared catalogs, even though
6324          * they send inquiry messages mentioning only their own DB.
6325          */
6326         if (databaseid == InvalidOid && pending_write_requests != NIL)
6327                 return true;
6328
6329         /* Search to see if there's an open request to write this database. */
6330         if (list_member_oid(pending_write_requests, databaseid))
6331                 return true;
6332
6333         return false;
6334 }
6335
6336 /*
6337  * Convert a potentially unsafely truncated activity string (see
6338  * PgBackendStatus.st_activity_raw's documentation) into a correctly truncated
6339  * one.
6340  *
6341  * The returned string is allocated in the caller's memory context and may be
6342  * freed.
6343  */
6344 char *
6345 pgstat_clip_activity(const char *raw_activity)
6346 {
6347         char       *activity;
6348         int                     rawlen;
6349         int                     cliplen;
6350
6351         /*
6352          * Some callers, like pgstat_get_backend_current_activity(), do not
6353          * guarantee that the buffer isn't concurrently modified. We try to take
6354          * care that the buffer is always terminated by a NUL byte regardless, but
6355          * let's still be paranoid about the string's length. In those cases the
6356          * underlying buffer is guaranteed to be pgstat_track_activity_query_size
6357          * large.
6358          */
6359         activity = pnstrdup(raw_activity, pgstat_track_activity_query_size - 1);
6360
6361         /* now double-guaranteed to be NUL terminated */
6362         rawlen = strlen(activity);
6363
6364         /*
6365          * All supported server-encodings make it possible to determine the length
6366          * of a multi-byte character from its first byte (this is not the case for
6367          * client encodings, see GB18030). As st_activity is always stored using
6368          * server encoding, this allows us to perform multi-byte aware truncation,
6369          * even if the string earlier was truncated in the middle of a multi-byte
6370          * character.
6371          */
6372         cliplen = pg_mbcliplen(activity, rawlen,
6373                                                    pgstat_track_activity_query_size - 1);
6374
6375         activity[cliplen] = '\0';
6376
6377         return activity;
6378 }