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) 2001-2018, 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);
 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);
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);
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);
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)
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 = heap_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 = HeapTupleGetOid(tup);
1228
1229                 CHECK_FOR_INTERRUPTS();
1230
1231                 (void) hash_search(htab, (void *) &thisoid, HASH_ENTER, NULL);
1232         }
1233         heap_endscan(scan);
1234         UnregisterSnapshot(snapshot);
1235         heap_close(rel, AccessShareLock);
1236
1237         return htab;
1238 }
1239
1240
1241 /* ----------
1242  * pgstat_drop_database() -
1243  *
1244  *      Tell the collector that we just dropped a database.
1245  *      (If the message gets lost, we will still clean the dead DB eventually
1246  *      via future invocations of pgstat_vacuum_stat().)
1247  * ----------
1248  */
1249 void
1250 pgstat_drop_database(Oid databaseid)
1251 {
1252         PgStat_MsgDropdb msg;
1253
1254         if (pgStatSock == PGINVALID_SOCKET)
1255                 return;
1256
1257         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_DROPDB);
1258         msg.m_databaseid = databaseid;
1259         pgstat_send(&msg, sizeof(msg));
1260 }
1261
1262
1263 /* ----------
1264  * pgstat_drop_relation() -
1265  *
1266  *      Tell the collector that we just dropped a relation.
1267  *      (If the message gets lost, we will still clean the dead entry eventually
1268  *      via future invocations of pgstat_vacuum_stat().)
1269  *
1270  *      Currently not used for lack of any good place to call it; we rely
1271  *      entirely on pgstat_vacuum_stat() to clean out stats for dead rels.
1272  * ----------
1273  */
1274 #ifdef NOT_USED
1275 void
1276 pgstat_drop_relation(Oid relid)
1277 {
1278         PgStat_MsgTabpurge msg;
1279         int                     len;
1280
1281         if (pgStatSock == PGINVALID_SOCKET)
1282                 return;
1283
1284         msg.m_tableid[0] = relid;
1285         msg.m_nentries = 1;
1286
1287         len = offsetof(PgStat_MsgTabpurge, m_tableid[0]) + sizeof(Oid);
1288
1289         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_TABPURGE);
1290         msg.m_databaseid = MyDatabaseId;
1291         pgstat_send(&msg, len);
1292 }
1293 #endif                                                  /* NOT_USED */
1294
1295
1296 /* ----------
1297  * pgstat_reset_counters() -
1298  *
1299  *      Tell the statistics collector to reset counters for our database.
1300  *
1301  *      Permission checking for this function is managed through the normal
1302  *      GRANT system.
1303  * ----------
1304  */
1305 void
1306 pgstat_reset_counters(void)
1307 {
1308         PgStat_MsgResetcounter msg;
1309
1310         if (pgStatSock == PGINVALID_SOCKET)
1311                 return;
1312
1313         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_RESETCOUNTER);
1314         msg.m_databaseid = MyDatabaseId;
1315         pgstat_send(&msg, sizeof(msg));
1316 }
1317
1318 /* ----------
1319  * pgstat_reset_shared_counters() -
1320  *
1321  *      Tell the statistics collector to reset cluster-wide shared counters.
1322  *
1323  *      Permission checking for this function is managed through the normal
1324  *      GRANT system.
1325  * ----------
1326  */
1327 void
1328 pgstat_reset_shared_counters(const char *target)
1329 {
1330         PgStat_MsgResetsharedcounter msg;
1331
1332         if (pgStatSock == PGINVALID_SOCKET)
1333                 return;
1334
1335         if (strcmp(target, "archiver") == 0)
1336                 msg.m_resettarget = RESET_ARCHIVER;
1337         else if (strcmp(target, "bgwriter") == 0)
1338                 msg.m_resettarget = RESET_BGWRITER;
1339         else
1340                 ereport(ERROR,
1341                                 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
1342                                  errmsg("unrecognized reset target: \"%s\"", target),
1343                                  errhint("Target must be \"archiver\" or \"bgwriter\".")));
1344
1345         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_RESETSHAREDCOUNTER);
1346         pgstat_send(&msg, sizeof(msg));
1347 }
1348
1349 /* ----------
1350  * pgstat_reset_single_counter() -
1351  *
1352  *      Tell the statistics collector to reset a single counter.
1353  *
1354  *      Permission checking for this function is managed through the normal
1355  *      GRANT system.
1356  * ----------
1357  */
1358 void
1359 pgstat_reset_single_counter(Oid objoid, PgStat_Single_Reset_Type type)
1360 {
1361         PgStat_MsgResetsinglecounter msg;
1362
1363         if (pgStatSock == PGINVALID_SOCKET)
1364                 return;
1365
1366         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_RESETSINGLECOUNTER);
1367         msg.m_databaseid = MyDatabaseId;
1368         msg.m_resettype = type;
1369         msg.m_objectid = objoid;
1370
1371         pgstat_send(&msg, sizeof(msg));
1372 }
1373
1374 /* ----------
1375  * pgstat_report_autovac() -
1376  *
1377  *      Called from autovacuum.c to report startup of an autovacuum process.
1378  *      We are called before InitPostgres is done, so can't rely on MyDatabaseId;
1379  *      the db OID must be passed in, instead.
1380  * ----------
1381  */
1382 void
1383 pgstat_report_autovac(Oid dboid)
1384 {
1385         PgStat_MsgAutovacStart msg;
1386
1387         if (pgStatSock == PGINVALID_SOCKET)
1388                 return;
1389
1390         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_AUTOVAC_START);
1391         msg.m_databaseid = dboid;
1392         msg.m_start_time = GetCurrentTimestamp();
1393
1394         pgstat_send(&msg, sizeof(msg));
1395 }
1396
1397
1398 /* ---------
1399  * pgstat_report_vacuum() -
1400  *
1401  *      Tell the collector about the table we just vacuumed.
1402  * ---------
1403  */
1404 void
1405 pgstat_report_vacuum(Oid tableoid, bool shared,
1406                                          PgStat_Counter livetuples, PgStat_Counter deadtuples)
1407 {
1408         PgStat_MsgVacuum msg;
1409
1410         if (pgStatSock == PGINVALID_SOCKET || !pgstat_track_counts)
1411                 return;
1412
1413         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_VACUUM);
1414         msg.m_databaseid = shared ? InvalidOid : MyDatabaseId;
1415         msg.m_tableoid = tableoid;
1416         msg.m_autovacuum = IsAutoVacuumWorkerProcess();
1417         msg.m_vacuumtime = GetCurrentTimestamp();
1418         msg.m_live_tuples = livetuples;
1419         msg.m_dead_tuples = deadtuples;
1420         pgstat_send(&msg, sizeof(msg));
1421 }
1422
1423 /* --------
1424  * pgstat_report_analyze() -
1425  *
1426  *      Tell the collector about the table we just analyzed.
1427  *
1428  * Caller must provide new live- and dead-tuples estimates, as well as a
1429  * flag indicating whether to reset the changes_since_analyze counter.
1430  * --------
1431  */
1432 void
1433 pgstat_report_analyze(Relation rel,
1434                                           PgStat_Counter livetuples, PgStat_Counter deadtuples,
1435                                           bool resetcounter)
1436 {
1437         PgStat_MsgAnalyze msg;
1438
1439         if (pgStatSock == PGINVALID_SOCKET || !pgstat_track_counts)
1440                 return;
1441
1442         /*
1443          * Unlike VACUUM, ANALYZE might be running inside a transaction that has
1444          * already inserted and/or deleted rows in the target table. ANALYZE will
1445          * have counted such rows as live or dead respectively. Because we will
1446          * report our counts of such rows at transaction end, we should subtract
1447          * off these counts from what we send to the collector now, else they'll
1448          * be double-counted after commit.  (This approach also ensures that the
1449          * collector ends up with the right numbers if we abort instead of
1450          * committing.)
1451          */
1452         if (rel->pgstat_info != NULL)
1453         {
1454                 PgStat_TableXactStatus *trans;
1455
1456                 for (trans = rel->pgstat_info->trans; trans; trans = trans->upper)
1457                 {
1458                         livetuples -= trans->tuples_inserted - trans->tuples_deleted;
1459                         deadtuples -= trans->tuples_updated + trans->tuples_deleted;
1460                 }
1461                 /* count stuff inserted by already-aborted subxacts, too */
1462                 deadtuples -= rel->pgstat_info->t_counts.t_delta_dead_tuples;
1463                 /* Since ANALYZE's counts are estimates, we could have underflowed */
1464                 livetuples = Max(livetuples, 0);
1465                 deadtuples = Max(deadtuples, 0);
1466         }
1467
1468         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_ANALYZE);
1469         msg.m_databaseid = rel->rd_rel->relisshared ? InvalidOid : MyDatabaseId;
1470         msg.m_tableoid = RelationGetRelid(rel);
1471         msg.m_autovacuum = IsAutoVacuumWorkerProcess();
1472         msg.m_resetcounter = resetcounter;
1473         msg.m_analyzetime = GetCurrentTimestamp();
1474         msg.m_live_tuples = livetuples;
1475         msg.m_dead_tuples = deadtuples;
1476         pgstat_send(&msg, sizeof(msg));
1477 }
1478
1479 /* --------
1480  * pgstat_report_recovery_conflict() -
1481  *
1482  *      Tell the collector about a Hot Standby recovery conflict.
1483  * --------
1484  */
1485 void
1486 pgstat_report_recovery_conflict(int reason)
1487 {
1488         PgStat_MsgRecoveryConflict msg;
1489
1490         if (pgStatSock == PGINVALID_SOCKET || !pgstat_track_counts)
1491                 return;
1492
1493         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_RECOVERYCONFLICT);
1494         msg.m_databaseid = MyDatabaseId;
1495         msg.m_reason = reason;
1496         pgstat_send(&msg, sizeof(msg));
1497 }
1498
1499 /* --------
1500  * pgstat_report_deadlock() -
1501  *
1502  *      Tell the collector about a deadlock detected.
1503  * --------
1504  */
1505 void
1506 pgstat_report_deadlock(void)
1507 {
1508         PgStat_MsgDeadlock msg;
1509
1510         if (pgStatSock == PGINVALID_SOCKET || !pgstat_track_counts)
1511                 return;
1512
1513         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_DEADLOCK);
1514         msg.m_databaseid = MyDatabaseId;
1515         pgstat_send(&msg, sizeof(msg));
1516 }
1517
1518 /* --------
1519  * pgstat_report_tempfile() -
1520  *
1521  *      Tell the collector about a temporary file.
1522  * --------
1523  */
1524 void
1525 pgstat_report_tempfile(size_t filesize)
1526 {
1527         PgStat_MsgTempFile msg;
1528
1529         if (pgStatSock == PGINVALID_SOCKET || !pgstat_track_counts)
1530                 return;
1531
1532         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_TEMPFILE);
1533         msg.m_databaseid = MyDatabaseId;
1534         msg.m_filesize = filesize;
1535         pgstat_send(&msg, sizeof(msg));
1536 }
1537
1538
1539 /* ----------
1540  * pgstat_ping() -
1541  *
1542  *      Send some junk data to the collector to increase traffic.
1543  * ----------
1544  */
1545 void
1546 pgstat_ping(void)
1547 {
1548         PgStat_MsgDummy msg;
1549
1550         if (pgStatSock == PGINVALID_SOCKET)
1551                 return;
1552
1553         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_DUMMY);
1554         pgstat_send(&msg, sizeof(msg));
1555 }
1556
1557 /* ----------
1558  * pgstat_send_inquiry() -
1559  *
1560  *      Notify collector that we need fresh data.
1561  * ----------
1562  */
1563 static void
1564 pgstat_send_inquiry(TimestampTz clock_time, TimestampTz cutoff_time, Oid databaseid)
1565 {
1566         PgStat_MsgInquiry msg;
1567
1568         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_INQUIRY);
1569         msg.clock_time = clock_time;
1570         msg.cutoff_time = cutoff_time;
1571         msg.databaseid = databaseid;
1572         pgstat_send(&msg, sizeof(msg));
1573 }
1574
1575
1576 /*
1577  * Initialize function call usage data.
1578  * Called by the executor before invoking a function.
1579  */
1580 void
1581 pgstat_init_function_usage(FunctionCallInfoData *fcinfo,
1582                                                    PgStat_FunctionCallUsage *fcu)
1583 {
1584         PgStat_BackendFunctionEntry *htabent;
1585         bool            found;
1586
1587         if (pgstat_track_functions <= fcinfo->flinfo->fn_stats)
1588         {
1589                 /* stats not wanted */
1590                 fcu->fs = NULL;
1591                 return;
1592         }
1593
1594         if (!pgStatFunctions)
1595         {
1596                 /* First time through - initialize function stat table */
1597                 HASHCTL         hash_ctl;
1598
1599                 memset(&hash_ctl, 0, sizeof(hash_ctl));
1600                 hash_ctl.keysize = sizeof(Oid);
1601                 hash_ctl.entrysize = sizeof(PgStat_BackendFunctionEntry);
1602                 pgStatFunctions = hash_create("Function stat entries",
1603                                                                           PGSTAT_FUNCTION_HASH_SIZE,
1604                                                                           &hash_ctl,
1605                                                                           HASH_ELEM | HASH_BLOBS);
1606         }
1607
1608         /* Get the stats entry for this function, create if necessary */
1609         htabent = hash_search(pgStatFunctions, &fcinfo->flinfo->fn_oid,
1610                                                   HASH_ENTER, &found);
1611         if (!found)
1612                 MemSet(&htabent->f_counts, 0, sizeof(PgStat_FunctionCounts));
1613
1614         fcu->fs = &htabent->f_counts;
1615
1616         /* save stats for this function, later used to compensate for recursion */
1617         fcu->save_f_total_time = htabent->f_counts.f_total_time;
1618
1619         /* save current backend-wide total time */
1620         fcu->save_total = total_func_time;
1621
1622         /* get clock time as of function start */
1623         INSTR_TIME_SET_CURRENT(fcu->f_start);
1624 }
1625
1626 /*
1627  * find_funcstat_entry - find any existing PgStat_BackendFunctionEntry entry
1628  *              for specified function
1629  *
1630  * If no entry, return NULL, don't create a new one
1631  */
1632 PgStat_BackendFunctionEntry *
1633 find_funcstat_entry(Oid func_id)
1634 {
1635         if (pgStatFunctions == NULL)
1636                 return NULL;
1637
1638         return (PgStat_BackendFunctionEntry *) hash_search(pgStatFunctions,
1639                                                                                                            (void *) &func_id,
1640                                                                                                            HASH_FIND, NULL);
1641 }
1642
1643 /*
1644  * Calculate function call usage and update stat counters.
1645  * Called by the executor after invoking a function.
1646  *
1647  * In the case of a set-returning function that runs in value-per-call mode,
1648  * we will see multiple pgstat_init_function_usage/pgstat_end_function_usage
1649  * calls for what the user considers a single call of the function.  The
1650  * finalize flag should be TRUE on the last call.
1651  */
1652 void
1653 pgstat_end_function_usage(PgStat_FunctionCallUsage *fcu, bool finalize)
1654 {
1655         PgStat_FunctionCounts *fs = fcu->fs;
1656         instr_time      f_total;
1657         instr_time      f_others;
1658         instr_time      f_self;
1659
1660         /* stats not wanted? */
1661         if (fs == NULL)
1662                 return;
1663
1664         /* total elapsed time in this function call */
1665         INSTR_TIME_SET_CURRENT(f_total);
1666         INSTR_TIME_SUBTRACT(f_total, fcu->f_start);
1667
1668         /* self usage: elapsed minus anything already charged to other calls */
1669         f_others = total_func_time;
1670         INSTR_TIME_SUBTRACT(f_others, fcu->save_total);
1671         f_self = f_total;
1672         INSTR_TIME_SUBTRACT(f_self, f_others);
1673
1674         /* update backend-wide total time */
1675         INSTR_TIME_ADD(total_func_time, f_self);
1676
1677         /*
1678          * Compute the new f_total_time as the total elapsed time added to the
1679          * pre-call value of f_total_time.  This is necessary to avoid
1680          * double-counting any time taken by recursive calls of myself.  (We do
1681          * not need any similar kluge for self time, since that already excludes
1682          * any recursive calls.)
1683          */
1684         INSTR_TIME_ADD(f_total, fcu->save_f_total_time);
1685
1686         /* update counters in function stats table */
1687         if (finalize)
1688                 fs->f_numcalls++;
1689         fs->f_total_time = f_total;
1690         INSTR_TIME_ADD(fs->f_self_time, f_self);
1691
1692         /* indicate that we have something to send */
1693         have_function_stats = true;
1694 }
1695
1696
1697 /* ----------
1698  * pgstat_initstats() -
1699  *
1700  *      Initialize a relcache entry to count access statistics.
1701  *      Called whenever a relation is opened.
1702  *
1703  *      We assume that a relcache entry's pgstat_info field is zeroed by
1704  *      relcache.c when the relcache entry is made; thereafter it is long-lived
1705  *      data.  We can avoid repeated searches of the TabStatus arrays when the
1706  *      same relation is touched repeatedly within a transaction.
1707  * ----------
1708  */
1709 void
1710 pgstat_initstats(Relation rel)
1711 {
1712         Oid                     rel_id = rel->rd_id;
1713         char            relkind = rel->rd_rel->relkind;
1714
1715         /* We only count stats for things that have storage */
1716         if (!(relkind == RELKIND_RELATION ||
1717                   relkind == RELKIND_MATVIEW ||
1718                   relkind == RELKIND_INDEX ||
1719                   relkind == RELKIND_TOASTVALUE ||
1720                   relkind == RELKIND_SEQUENCE))
1721         {
1722                 rel->pgstat_info = NULL;
1723                 return;
1724         }
1725
1726         if (pgStatSock == PGINVALID_SOCKET || !pgstat_track_counts)
1727         {
1728                 /* We're not counting at all */
1729                 rel->pgstat_info = NULL;
1730                 return;
1731         }
1732
1733         /*
1734          * If we already set up this relation in the current transaction, nothing
1735          * to do.
1736          */
1737         if (rel->pgstat_info != NULL &&
1738                 rel->pgstat_info->t_id == rel_id)
1739                 return;
1740
1741         /* Else find or make the PgStat_TableStatus entry, and update link */
1742         rel->pgstat_info = get_tabstat_entry(rel_id, rel->rd_rel->relisshared);
1743 }
1744
1745 /*
1746  * get_tabstat_entry - find or create a PgStat_TableStatus entry for rel
1747  */
1748 static PgStat_TableStatus *
1749 get_tabstat_entry(Oid rel_id, bool isshared)
1750 {
1751         TabStatHashEntry *hash_entry;
1752         PgStat_TableStatus *entry;
1753         TabStatusArray *tsa;
1754         bool            found;
1755
1756         /*
1757          * Create hash table if we don't have it already.
1758          */
1759         if (pgStatTabHash == NULL)
1760         {
1761                 HASHCTL         ctl;
1762
1763                 memset(&ctl, 0, sizeof(ctl));
1764                 ctl.keysize = sizeof(Oid);
1765                 ctl.entrysize = sizeof(TabStatHashEntry);
1766
1767                 pgStatTabHash = hash_create("pgstat TabStatusArray lookup hash table",
1768                                                                         TABSTAT_QUANTUM,
1769                                                                         &ctl,
1770                                                                         HASH_ELEM | HASH_BLOBS);
1771         }
1772
1773         /*
1774          * Find an entry or create a new one.
1775          */
1776         hash_entry = hash_search(pgStatTabHash, &rel_id, HASH_ENTER, &found);
1777         if (!found)
1778         {
1779                 /* initialize new entry with null pointer */
1780                 hash_entry->tsa_entry = NULL;
1781         }
1782
1783         /*
1784          * If entry is already valid, we're done.
1785          */
1786         if (hash_entry->tsa_entry)
1787                 return hash_entry->tsa_entry;
1788
1789         /*
1790          * Locate the first pgStatTabList entry with free space, making a new list
1791          * entry if needed.  Note that we could get an OOM failure here, but if so
1792          * we have left the hashtable and the list in a consistent state.
1793          */
1794         if (pgStatTabList == NULL)
1795         {
1796                 /* Set up first pgStatTabList entry */
1797                 pgStatTabList = (TabStatusArray *)
1798                         MemoryContextAllocZero(TopMemoryContext,
1799                                                                    sizeof(TabStatusArray));
1800         }
1801
1802         tsa = pgStatTabList;
1803         while (tsa->tsa_used >= TABSTAT_QUANTUM)
1804         {
1805                 if (tsa->tsa_next == NULL)
1806                         tsa->tsa_next = (TabStatusArray *)
1807                                 MemoryContextAllocZero(TopMemoryContext,
1808                                                                            sizeof(TabStatusArray));
1809                 tsa = tsa->tsa_next;
1810         }
1811
1812         /*
1813          * Allocate a PgStat_TableStatus entry within this list entry.  We assume
1814          * the entry was already zeroed, either at creation or after last use.
1815          */
1816         entry = &tsa->tsa_entries[tsa->tsa_used++];
1817         entry->t_id = rel_id;
1818         entry->t_shared = isshared;
1819
1820         /*
1821          * Now we can fill the entry in pgStatTabHash.
1822          */
1823         hash_entry->tsa_entry = entry;
1824
1825         return entry;
1826 }
1827
1828 /*
1829  * find_tabstat_entry - find any existing PgStat_TableStatus entry for rel
1830  *
1831  * If no entry, return NULL, don't create a new one
1832  *
1833  * Note: if we got an error in the most recent execution of pgstat_report_stat,
1834  * it's possible that an entry exists but there's no hashtable entry for it.
1835  * That's okay, we'll treat this case as "doesn't exist".
1836  */
1837 PgStat_TableStatus *
1838 find_tabstat_entry(Oid rel_id)
1839 {
1840         TabStatHashEntry *hash_entry;
1841
1842         /* If hashtable doesn't exist, there are no entries at all */
1843         if (!pgStatTabHash)
1844                 return NULL;
1845
1846         hash_entry = hash_search(pgStatTabHash, &rel_id, HASH_FIND, NULL);
1847         if (!hash_entry)
1848                 return NULL;
1849
1850         /* Note that this step could also return NULL, but that's correct */
1851         return hash_entry->tsa_entry;
1852 }
1853
1854 /*
1855  * get_tabstat_stack_level - add a new (sub)transaction stack entry if needed
1856  */
1857 static PgStat_SubXactStatus *
1858 get_tabstat_stack_level(int nest_level)
1859 {
1860         PgStat_SubXactStatus *xact_state;
1861
1862         xact_state = pgStatXactStack;
1863         if (xact_state == NULL || xact_state->nest_level != nest_level)
1864         {
1865                 xact_state = (PgStat_SubXactStatus *)
1866                         MemoryContextAlloc(TopTransactionContext,
1867                                                            sizeof(PgStat_SubXactStatus));
1868                 xact_state->nest_level = nest_level;
1869                 xact_state->prev = pgStatXactStack;
1870                 xact_state->first = NULL;
1871                 pgStatXactStack = xact_state;
1872         }
1873         return xact_state;
1874 }
1875
1876 /*
1877  * add_tabstat_xact_level - add a new (sub)transaction state record
1878  */
1879 static void
1880 add_tabstat_xact_level(PgStat_TableStatus *pgstat_info, int nest_level)
1881 {
1882         PgStat_SubXactStatus *xact_state;
1883         PgStat_TableXactStatus *trans;
1884
1885         /*
1886          * If this is the first rel to be modified at the current nest level, we
1887          * first have to push a transaction stack entry.
1888          */
1889         xact_state = get_tabstat_stack_level(nest_level);
1890
1891         /* Now make a per-table stack entry */
1892         trans = (PgStat_TableXactStatus *)
1893                 MemoryContextAllocZero(TopTransactionContext,
1894                                                            sizeof(PgStat_TableXactStatus));
1895         trans->nest_level = nest_level;
1896         trans->upper = pgstat_info->trans;
1897         trans->parent = pgstat_info;
1898         trans->next = xact_state->first;
1899         xact_state->first = trans;
1900         pgstat_info->trans = trans;
1901 }
1902
1903 /*
1904  * pgstat_count_heap_insert - count a tuple insertion of n tuples
1905  */
1906 void
1907 pgstat_count_heap_insert(Relation rel, PgStat_Counter n)
1908 {
1909         PgStat_TableStatus *pgstat_info = rel->pgstat_info;
1910
1911         if (pgstat_info != NULL)
1912         {
1913                 /* We have to log the effect at the proper transactional level */
1914                 int                     nest_level = GetCurrentTransactionNestLevel();
1915
1916                 if (pgstat_info->trans == NULL ||
1917                         pgstat_info->trans->nest_level != nest_level)
1918                         add_tabstat_xact_level(pgstat_info, nest_level);
1919
1920                 pgstat_info->trans->tuples_inserted += n;
1921         }
1922 }
1923
1924 /*
1925  * pgstat_count_heap_update - count a tuple update
1926  */
1927 void
1928 pgstat_count_heap_update(Relation rel, bool hot)
1929 {
1930         PgStat_TableStatus *pgstat_info = rel->pgstat_info;
1931
1932         if (pgstat_info != NULL)
1933         {
1934                 /* We have to log the effect at the proper transactional level */
1935                 int                     nest_level = GetCurrentTransactionNestLevel();
1936
1937                 if (pgstat_info->trans == NULL ||
1938                         pgstat_info->trans->nest_level != nest_level)
1939                         add_tabstat_xact_level(pgstat_info, nest_level);
1940
1941                 pgstat_info->trans->tuples_updated++;
1942
1943                 /* t_tuples_hot_updated is nontransactional, so just advance it */
1944                 if (hot)
1945                         pgstat_info->t_counts.t_tuples_hot_updated++;
1946         }
1947 }
1948
1949 /*
1950  * pgstat_count_heap_delete - count a tuple deletion
1951  */
1952 void
1953 pgstat_count_heap_delete(Relation rel)
1954 {
1955         PgStat_TableStatus *pgstat_info = rel->pgstat_info;
1956
1957         if (pgstat_info != NULL)
1958         {
1959                 /* We have to log the effect at the proper transactional level */
1960                 int                     nest_level = GetCurrentTransactionNestLevel();
1961
1962                 if (pgstat_info->trans == NULL ||
1963                         pgstat_info->trans->nest_level != nest_level)
1964                         add_tabstat_xact_level(pgstat_info, nest_level);
1965
1966                 pgstat_info->trans->tuples_deleted++;
1967         }
1968 }
1969
1970 /*
1971  * pgstat_truncate_save_counters
1972  *
1973  * Whenever a table is truncated, we save its i/u/d counters so that they can
1974  * be cleared, and if the (sub)xact that executed the truncate later aborts,
1975  * the counters can be restored to the saved (pre-truncate) values.  Note we do
1976  * this on the first truncate in any particular subxact level only.
1977  */
1978 static void
1979 pgstat_truncate_save_counters(PgStat_TableXactStatus *trans)
1980 {
1981         if (!trans->truncated)
1982         {
1983                 trans->inserted_pre_trunc = trans->tuples_inserted;
1984                 trans->updated_pre_trunc = trans->tuples_updated;
1985                 trans->deleted_pre_trunc = trans->tuples_deleted;
1986                 trans->truncated = true;
1987         }
1988 }
1989
1990 /*
1991  * pgstat_truncate_restore_counters - restore counters when a truncate aborts
1992  */
1993 static void
1994 pgstat_truncate_restore_counters(PgStat_TableXactStatus *trans)
1995 {
1996         if (trans->truncated)
1997         {
1998                 trans->tuples_inserted = trans->inserted_pre_trunc;
1999                 trans->tuples_updated = trans->updated_pre_trunc;
2000                 trans->tuples_deleted = trans->deleted_pre_trunc;
2001         }
2002 }
2003
2004 /*
2005  * pgstat_count_truncate - update tuple counters due to truncate
2006  */
2007 void
2008 pgstat_count_truncate(Relation rel)
2009 {
2010         PgStat_TableStatus *pgstat_info = rel->pgstat_info;
2011
2012         if (pgstat_info != NULL)
2013         {
2014                 /* We have to log the effect at the proper transactional level */
2015                 int                     nest_level = GetCurrentTransactionNestLevel();
2016
2017                 if (pgstat_info->trans == NULL ||
2018                         pgstat_info->trans->nest_level != nest_level)
2019                         add_tabstat_xact_level(pgstat_info, nest_level);
2020
2021                 pgstat_truncate_save_counters(pgstat_info->trans);
2022                 pgstat_info->trans->tuples_inserted = 0;
2023                 pgstat_info->trans->tuples_updated = 0;
2024                 pgstat_info->trans->tuples_deleted = 0;
2025         }
2026 }
2027
2028 /*
2029  * pgstat_update_heap_dead_tuples - update dead-tuples count
2030  *
2031  * The semantics of this are that we are reporting the nontransactional
2032  * recovery of "delta" dead tuples; so t_delta_dead_tuples decreases
2033  * rather than increasing, and the change goes straight into the per-table
2034  * counter, not into transactional state.
2035  */
2036 void
2037 pgstat_update_heap_dead_tuples(Relation rel, int delta)
2038 {
2039         PgStat_TableStatus *pgstat_info = rel->pgstat_info;
2040
2041         if (pgstat_info != NULL)
2042                 pgstat_info->t_counts.t_delta_dead_tuples -= delta;
2043 }
2044
2045
2046 /* ----------
2047  * AtEOXact_PgStat
2048  *
2049  *      Called from access/transam/xact.c at top-level transaction commit/abort.
2050  * ----------
2051  */
2052 void
2053 AtEOXact_PgStat(bool isCommit)
2054 {
2055         PgStat_SubXactStatus *xact_state;
2056
2057         /*
2058          * Count transaction commit or abort.  (We use counters, not just bools,
2059          * in case the reporting message isn't sent right away.)
2060          */
2061         if (isCommit)
2062                 pgStatXactCommit++;
2063         else
2064                 pgStatXactRollback++;
2065
2066         /*
2067          * Transfer transactional insert/update counts into the base tabstat
2068          * entries.  We don't bother to free any of the transactional state, since
2069          * it's all in TopTransactionContext and will go away anyway.
2070          */
2071         xact_state = pgStatXactStack;
2072         if (xact_state != NULL)
2073         {
2074                 PgStat_TableXactStatus *trans;
2075
2076                 Assert(xact_state->nest_level == 1);
2077                 Assert(xact_state->prev == NULL);
2078                 for (trans = xact_state->first; trans != NULL; trans = trans->next)
2079                 {
2080                         PgStat_TableStatus *tabstat;
2081
2082                         Assert(trans->nest_level == 1);
2083                         Assert(trans->upper == NULL);
2084                         tabstat = trans->parent;
2085                         Assert(tabstat->trans == trans);
2086                         /* restore pre-truncate stats (if any) in case of aborted xact */
2087                         if (!isCommit)
2088                                 pgstat_truncate_restore_counters(trans);
2089                         /* count attempted actions regardless of commit/abort */
2090                         tabstat->t_counts.t_tuples_inserted += trans->tuples_inserted;
2091                         tabstat->t_counts.t_tuples_updated += trans->tuples_updated;
2092                         tabstat->t_counts.t_tuples_deleted += trans->tuples_deleted;
2093                         if (isCommit)
2094                         {
2095                                 tabstat->t_counts.t_truncated = trans->truncated;
2096                                 if (trans->truncated)
2097                                 {
2098                                         /* forget live/dead stats seen by backend thus far */
2099                                         tabstat->t_counts.t_delta_live_tuples = 0;
2100                                         tabstat->t_counts.t_delta_dead_tuples = 0;
2101                                 }
2102                                 /* insert adds a live tuple, delete removes one */
2103                                 tabstat->t_counts.t_delta_live_tuples +=
2104                                         trans->tuples_inserted - trans->tuples_deleted;
2105                                 /* update and delete each create a dead tuple */
2106                                 tabstat->t_counts.t_delta_dead_tuples +=
2107                                         trans->tuples_updated + trans->tuples_deleted;
2108                                 /* insert, update, delete each count as one change event */
2109                                 tabstat->t_counts.t_changed_tuples +=
2110                                         trans->tuples_inserted + trans->tuples_updated +
2111                                         trans->tuples_deleted;
2112                         }
2113                         else
2114                         {
2115                                 /* inserted tuples are dead, deleted tuples are unaffected */
2116                                 tabstat->t_counts.t_delta_dead_tuples +=
2117                                         trans->tuples_inserted + trans->tuples_updated;
2118                                 /* an aborted xact generates no changed_tuple events */
2119                         }
2120                         tabstat->trans = NULL;
2121                 }
2122         }
2123         pgStatXactStack = NULL;
2124
2125         /* Make sure any stats snapshot is thrown away */
2126         pgstat_clear_snapshot();
2127 }
2128
2129 /* ----------
2130  * AtEOSubXact_PgStat
2131  *
2132  *      Called from access/transam/xact.c at subtransaction commit/abort.
2133  * ----------
2134  */
2135 void
2136 AtEOSubXact_PgStat(bool isCommit, int nestDepth)
2137 {
2138         PgStat_SubXactStatus *xact_state;
2139
2140         /*
2141          * Transfer transactional insert/update counts into the next higher
2142          * subtransaction state.
2143          */
2144         xact_state = pgStatXactStack;
2145         if (xact_state != NULL &&
2146                 xact_state->nest_level >= nestDepth)
2147         {
2148                 PgStat_TableXactStatus *trans;
2149                 PgStat_TableXactStatus *next_trans;
2150
2151                 /* delink xact_state from stack immediately to simplify reuse case */
2152                 pgStatXactStack = xact_state->prev;
2153
2154                 for (trans = xact_state->first; trans != NULL; trans = next_trans)
2155                 {
2156                         PgStat_TableStatus *tabstat;
2157
2158                         next_trans = trans->next;
2159                         Assert(trans->nest_level == nestDepth);
2160                         tabstat = trans->parent;
2161                         Assert(tabstat->trans == trans);
2162                         if (isCommit)
2163                         {
2164                                 if (trans->upper && trans->upper->nest_level == nestDepth - 1)
2165                                 {
2166                                         if (trans->truncated)
2167                                         {
2168                                                 /* propagate the truncate status one level up */
2169                                                 pgstat_truncate_save_counters(trans->upper);
2170                                                 /* replace upper xact stats with ours */
2171                                                 trans->upper->tuples_inserted = trans->tuples_inserted;
2172                                                 trans->upper->tuples_updated = trans->tuples_updated;
2173                                                 trans->upper->tuples_deleted = trans->tuples_deleted;
2174                                         }
2175                                         else
2176                                         {
2177                                                 trans->upper->tuples_inserted += trans->tuples_inserted;
2178                                                 trans->upper->tuples_updated += trans->tuples_updated;
2179                                                 trans->upper->tuples_deleted += trans->tuples_deleted;
2180                                         }
2181                                         tabstat->trans = trans->upper;
2182                                         pfree(trans);
2183                                 }
2184                                 else
2185                                 {
2186                                         /*
2187                                          * When there isn't an immediate parent state, we can just
2188                                          * reuse the record instead of going through a
2189                                          * palloc/pfree pushup (this works since it's all in
2190                                          * TopTransactionContext anyway).  We have to re-link it
2191                                          * into the parent level, though, and that might mean
2192                                          * pushing a new entry into the pgStatXactStack.
2193                                          */
2194                                         PgStat_SubXactStatus *upper_xact_state;
2195
2196                                         upper_xact_state = get_tabstat_stack_level(nestDepth - 1);
2197                                         trans->next = upper_xact_state->first;
2198                                         upper_xact_state->first = trans;
2199                                         trans->nest_level = nestDepth - 1;
2200                                 }
2201                         }
2202                         else
2203                         {
2204                                 /*
2205                                  * On abort, update top-level tabstat counts, then forget the
2206                                  * subtransaction
2207                                  */
2208
2209                                 /* first restore values obliterated by truncate */
2210                                 pgstat_truncate_restore_counters(trans);
2211                                 /* count attempted actions regardless of commit/abort */
2212                                 tabstat->t_counts.t_tuples_inserted += trans->tuples_inserted;
2213                                 tabstat->t_counts.t_tuples_updated += trans->tuples_updated;
2214                                 tabstat->t_counts.t_tuples_deleted += trans->tuples_deleted;
2215                                 /* inserted tuples are dead, deleted tuples are unaffected */
2216                                 tabstat->t_counts.t_delta_dead_tuples +=
2217                                         trans->tuples_inserted + trans->tuples_updated;
2218                                 tabstat->trans = trans->upper;
2219                                 pfree(trans);
2220                         }
2221                 }
2222                 pfree(xact_state);
2223         }
2224 }
2225
2226
2227 /*
2228  * AtPrepare_PgStat
2229  *              Save the transactional stats state at 2PC transaction prepare.
2230  *
2231  * In this phase we just generate 2PC records for all the pending
2232  * transaction-dependent stats work.
2233  */
2234 void
2235 AtPrepare_PgStat(void)
2236 {
2237         PgStat_SubXactStatus *xact_state;
2238
2239         xact_state = pgStatXactStack;
2240         if (xact_state != NULL)
2241         {
2242                 PgStat_TableXactStatus *trans;
2243
2244                 Assert(xact_state->nest_level == 1);
2245                 Assert(xact_state->prev == NULL);
2246                 for (trans = xact_state->first; trans != NULL; trans = trans->next)
2247                 {
2248                         PgStat_TableStatus *tabstat;
2249                         TwoPhasePgStatRecord record;
2250
2251                         Assert(trans->nest_level == 1);
2252                         Assert(trans->upper == NULL);
2253                         tabstat = trans->parent;
2254                         Assert(tabstat->trans == trans);
2255
2256                         record.tuples_inserted = trans->tuples_inserted;
2257                         record.tuples_updated = trans->tuples_updated;
2258                         record.tuples_deleted = trans->tuples_deleted;
2259                         record.inserted_pre_trunc = trans->inserted_pre_trunc;
2260                         record.updated_pre_trunc = trans->updated_pre_trunc;
2261                         record.deleted_pre_trunc = trans->deleted_pre_trunc;
2262                         record.t_id = tabstat->t_id;
2263                         record.t_shared = tabstat->t_shared;
2264                         record.t_truncated = trans->truncated;
2265
2266                         RegisterTwoPhaseRecord(TWOPHASE_RM_PGSTAT_ID, 0,
2267                                                                    &record, sizeof(TwoPhasePgStatRecord));
2268                 }
2269         }
2270 }
2271
2272 /*
2273  * PostPrepare_PgStat
2274  *              Clean up after successful PREPARE.
2275  *
2276  * All we need do here is unlink the transaction stats state from the
2277  * nontransactional state.  The nontransactional action counts will be
2278  * reported to the stats collector immediately, while the effects on live
2279  * and dead tuple counts are preserved in the 2PC state file.
2280  *
2281  * Note: AtEOXact_PgStat is not called during PREPARE.
2282  */
2283 void
2284 PostPrepare_PgStat(void)
2285 {
2286         PgStat_SubXactStatus *xact_state;
2287
2288         /*
2289          * We don't bother to free any of the transactional state, since it's all
2290          * in TopTransactionContext and will go away anyway.
2291          */
2292         xact_state = pgStatXactStack;
2293         if (xact_state != NULL)
2294         {
2295                 PgStat_TableXactStatus *trans;
2296
2297                 for (trans = xact_state->first; trans != NULL; trans = trans->next)
2298                 {
2299                         PgStat_TableStatus *tabstat;
2300
2301                         tabstat = trans->parent;
2302                         tabstat->trans = NULL;
2303                 }
2304         }
2305         pgStatXactStack = NULL;
2306
2307         /* Make sure any stats snapshot is thrown away */
2308         pgstat_clear_snapshot();
2309 }
2310
2311 /*
2312  * 2PC processing routine for COMMIT PREPARED case.
2313  *
2314  * Load the saved counts into our local pgstats state.
2315  */
2316 void
2317 pgstat_twophase_postcommit(TransactionId xid, uint16 info,
2318                                                    void *recdata, uint32 len)
2319 {
2320         TwoPhasePgStatRecord *rec = (TwoPhasePgStatRecord *) recdata;
2321         PgStat_TableStatus *pgstat_info;
2322
2323         /* Find or create a tabstat entry for the rel */
2324         pgstat_info = get_tabstat_entry(rec->t_id, rec->t_shared);
2325
2326         /* Same math as in AtEOXact_PgStat, commit case */
2327         pgstat_info->t_counts.t_tuples_inserted += rec->tuples_inserted;
2328         pgstat_info->t_counts.t_tuples_updated += rec->tuples_updated;
2329         pgstat_info->t_counts.t_tuples_deleted += rec->tuples_deleted;
2330         pgstat_info->t_counts.t_truncated = rec->t_truncated;
2331         if (rec->t_truncated)
2332         {
2333                 /* forget live/dead stats seen by backend thus far */
2334                 pgstat_info->t_counts.t_delta_live_tuples = 0;
2335                 pgstat_info->t_counts.t_delta_dead_tuples = 0;
2336         }
2337         pgstat_info->t_counts.t_delta_live_tuples +=
2338                 rec->tuples_inserted - rec->tuples_deleted;
2339         pgstat_info->t_counts.t_delta_dead_tuples +=
2340                 rec->tuples_updated + rec->tuples_deleted;
2341         pgstat_info->t_counts.t_changed_tuples +=
2342                 rec->tuples_inserted + rec->tuples_updated +
2343                 rec->tuples_deleted;
2344 }
2345
2346 /*
2347  * 2PC processing routine for ROLLBACK PREPARED case.
2348  *
2349  * Load the saved counts into our local pgstats state, but treat them
2350  * as aborted.
2351  */
2352 void
2353 pgstat_twophase_postabort(TransactionId xid, uint16 info,
2354                                                   void *recdata, uint32 len)
2355 {
2356         TwoPhasePgStatRecord *rec = (TwoPhasePgStatRecord *) recdata;
2357         PgStat_TableStatus *pgstat_info;
2358
2359         /* Find or create a tabstat entry for the rel */
2360         pgstat_info = get_tabstat_entry(rec->t_id, rec->t_shared);
2361
2362         /* Same math as in AtEOXact_PgStat, abort case */
2363         if (rec->t_truncated)
2364         {
2365                 rec->tuples_inserted = rec->inserted_pre_trunc;
2366                 rec->tuples_updated = rec->updated_pre_trunc;
2367                 rec->tuples_deleted = rec->deleted_pre_trunc;
2368         }
2369         pgstat_info->t_counts.t_tuples_inserted += rec->tuples_inserted;
2370         pgstat_info->t_counts.t_tuples_updated += rec->tuples_updated;
2371         pgstat_info->t_counts.t_tuples_deleted += rec->tuples_deleted;
2372         pgstat_info->t_counts.t_delta_dead_tuples +=
2373                 rec->tuples_inserted + rec->tuples_updated;
2374 }
2375
2376
2377 /* ----------
2378  * pgstat_fetch_stat_dbentry() -
2379  *
2380  *      Support function for the SQL-callable pgstat* functions. Returns
2381  *      the collected statistics for one database or NULL. NULL doesn't mean
2382  *      that the database doesn't exist, it is just not yet known by the
2383  *      collector, so the caller is better off to report ZERO instead.
2384  * ----------
2385  */
2386 PgStat_StatDBEntry *
2387 pgstat_fetch_stat_dbentry(Oid dbid)
2388 {
2389         /*
2390          * If not done for this transaction, read the statistics collector stats
2391          * file into some hash tables.
2392          */
2393         backend_read_statsfile();
2394
2395         /*
2396          * Lookup the requested database; return NULL if not found
2397          */
2398         return (PgStat_StatDBEntry *) hash_search(pgStatDBHash,
2399                                                                                           (void *) &dbid,
2400                                                                                           HASH_FIND, NULL);
2401 }
2402
2403
2404 /* ----------
2405  * pgstat_fetch_stat_tabentry() -
2406  *
2407  *      Support function for the SQL-callable pgstat* functions. Returns
2408  *      the collected statistics for one table or NULL. NULL doesn't mean
2409  *      that the table doesn't exist, it is just not yet known by the
2410  *      collector, so the caller is better off to report ZERO instead.
2411  * ----------
2412  */
2413 PgStat_StatTabEntry *
2414 pgstat_fetch_stat_tabentry(Oid relid)
2415 {
2416         Oid                     dbid;
2417         PgStat_StatDBEntry *dbentry;
2418         PgStat_StatTabEntry *tabentry;
2419
2420         /*
2421          * If not done for this transaction, read the statistics collector stats
2422          * file into some hash tables.
2423          */
2424         backend_read_statsfile();
2425
2426         /*
2427          * Lookup our database, then look in its table hash table.
2428          */
2429         dbid = MyDatabaseId;
2430         dbentry = (PgStat_StatDBEntry *) hash_search(pgStatDBHash,
2431                                                                                                  (void *) &dbid,
2432                                                                                                  HASH_FIND, NULL);
2433         if (dbentry != NULL && dbentry->tables != NULL)
2434         {
2435                 tabentry = (PgStat_StatTabEntry *) hash_search(dbentry->tables,
2436                                                                                                            (void *) &relid,
2437                                                                                                            HASH_FIND, NULL);
2438                 if (tabentry)
2439                         return tabentry;
2440         }
2441
2442         /*
2443          * If we didn't find it, maybe it's a shared table.
2444          */
2445         dbid = InvalidOid;
2446         dbentry = (PgStat_StatDBEntry *) hash_search(pgStatDBHash,
2447                                                                                                  (void *) &dbid,
2448                                                                                                  HASH_FIND, NULL);
2449         if (dbentry != NULL && dbentry->tables != NULL)
2450         {
2451                 tabentry = (PgStat_StatTabEntry *) hash_search(dbentry->tables,
2452                                                                                                            (void *) &relid,
2453                                                                                                            HASH_FIND, NULL);
2454                 if (tabentry)
2455                         return tabentry;
2456         }
2457
2458         return NULL;
2459 }
2460
2461
2462 /* ----------
2463  * pgstat_fetch_stat_funcentry() -
2464  *
2465  *      Support function for the SQL-callable pgstat* functions. Returns
2466  *      the collected statistics for one function or NULL.
2467  * ----------
2468  */
2469 PgStat_StatFuncEntry *
2470 pgstat_fetch_stat_funcentry(Oid func_id)
2471 {
2472         PgStat_StatDBEntry *dbentry;
2473         PgStat_StatFuncEntry *funcentry = NULL;
2474
2475         /* load the stats file if needed */
2476         backend_read_statsfile();
2477
2478         /* Lookup our database, then find the requested function.  */
2479         dbentry = pgstat_fetch_stat_dbentry(MyDatabaseId);
2480         if (dbentry != NULL && dbentry->functions != NULL)
2481         {
2482                 funcentry = (PgStat_StatFuncEntry *) hash_search(dbentry->functions,
2483                                                                                                                  (void *) &func_id,
2484                                                                                                                  HASH_FIND, NULL);
2485         }
2486
2487         return funcentry;
2488 }
2489
2490
2491 /* ----------
2492  * pgstat_fetch_stat_beentry() -
2493  *
2494  *      Support function for the SQL-callable pgstat* functions. Returns
2495  *      our local copy of the current-activity entry for one backend.
2496  *
2497  *      NB: caller is responsible for a check if the user is permitted to see
2498  *      this info (especially the querystring).
2499  * ----------
2500  */
2501 PgBackendStatus *
2502 pgstat_fetch_stat_beentry(int beid)
2503 {
2504         pgstat_read_current_status();
2505
2506         if (beid < 1 || beid > localNumBackends)
2507                 return NULL;
2508
2509         return &localBackendStatusTable[beid - 1].backendStatus;
2510 }
2511
2512
2513 /* ----------
2514  * pgstat_fetch_stat_local_beentry() -
2515  *
2516  *      Like pgstat_fetch_stat_beentry() but with locally computed additions (like
2517  *      xid and xmin values of the backend)
2518  *
2519  *      NB: caller is responsible for a check if the user is permitted to see
2520  *      this info (especially the querystring).
2521  * ----------
2522  */
2523 LocalPgBackendStatus *
2524 pgstat_fetch_stat_local_beentry(int beid)
2525 {
2526         pgstat_read_current_status();
2527
2528         if (beid < 1 || beid > localNumBackends)
2529                 return NULL;
2530
2531         return &localBackendStatusTable[beid - 1];
2532 }
2533
2534
2535 /* ----------
2536  * pgstat_fetch_stat_numbackends() -
2537  *
2538  *      Support function for the SQL-callable pgstat* functions. Returns
2539  *      the maximum current backend id.
2540  * ----------
2541  */
2542 int
2543 pgstat_fetch_stat_numbackends(void)
2544 {
2545         pgstat_read_current_status();
2546
2547         return localNumBackends;
2548 }
2549
2550 /*
2551  * ---------
2552  * pgstat_fetch_stat_archiver() -
2553  *
2554  *      Support function for the SQL-callable pgstat* functions. Returns
2555  *      a pointer to the archiver statistics struct.
2556  * ---------
2557  */
2558 PgStat_ArchiverStats *
2559 pgstat_fetch_stat_archiver(void)
2560 {
2561         backend_read_statsfile();
2562
2563         return &archiverStats;
2564 }
2565
2566
2567 /*
2568  * ---------
2569  * pgstat_fetch_global() -
2570  *
2571  *      Support function for the SQL-callable pgstat* functions. Returns
2572  *      a pointer to the global statistics struct.
2573  * ---------
2574  */
2575 PgStat_GlobalStats *
2576 pgstat_fetch_global(void)
2577 {
2578         backend_read_statsfile();
2579
2580         return &globalStats;
2581 }
2582
2583
2584 /* ------------------------------------------------------------
2585  * Functions for management of the shared-memory PgBackendStatus array
2586  * ------------------------------------------------------------
2587  */
2588
2589 static PgBackendStatus *BackendStatusArray = NULL;
2590 static PgBackendStatus *MyBEEntry = NULL;
2591 static char *BackendAppnameBuffer = NULL;
2592 static char *BackendClientHostnameBuffer = NULL;
2593 static char *BackendActivityBuffer = NULL;
2594 static Size BackendActivityBufferSize = 0;
2595 #ifdef USE_SSL
2596 static PgBackendSSLStatus *BackendSslStatusBuffer = NULL;
2597 #endif
2598
2599
2600 /*
2601  * Report shared-memory space needed by CreateSharedBackendStatus.
2602  */
2603 Size
2604 BackendStatusShmemSize(void)
2605 {
2606         Size            size;
2607
2608         /* BackendStatusArray: */
2609         size = mul_size(sizeof(PgBackendStatus), NumBackendStatSlots);
2610         /* BackendAppnameBuffer: */
2611         size = add_size(size,
2612                                         mul_size(NAMEDATALEN, NumBackendStatSlots));
2613         /* BackendClientHostnameBuffer: */
2614         size = add_size(size,
2615                                         mul_size(NAMEDATALEN, NumBackendStatSlots));
2616         /* BackendActivityBuffer: */
2617         size = add_size(size,
2618                                         mul_size(pgstat_track_activity_query_size, NumBackendStatSlots));
2619 #ifdef USE_SSL
2620         /* BackendSslStatusBuffer: */
2621         size = add_size(size,
2622                                         mul_size(sizeof(PgBackendSSLStatus), NumBackendStatSlots));
2623 #endif
2624         return size;
2625 }
2626
2627 /*
2628  * Initialize the shared status array and several string buffers
2629  * during postmaster startup.
2630  */
2631 void
2632 CreateSharedBackendStatus(void)
2633 {
2634         Size            size;
2635         bool            found;
2636         int                     i;
2637         char       *buffer;
2638
2639         /* Create or attach to the shared array */
2640         size = mul_size(sizeof(PgBackendStatus), NumBackendStatSlots);
2641         BackendStatusArray = (PgBackendStatus *)
2642                 ShmemInitStruct("Backend Status Array", size, &found);
2643
2644         if (!found)
2645         {
2646                 /*
2647                  * We're the first - initialize.
2648                  */
2649                 MemSet(BackendStatusArray, 0, size);
2650         }
2651
2652         /* Create or attach to the shared appname buffer */
2653         size = mul_size(NAMEDATALEN, MaxBackends);
2654         BackendAppnameBuffer = (char *)
2655                 ShmemInitStruct("Backend Application Name Buffer", size, &found);
2656
2657         if (!found)
2658         {
2659                 MemSet(BackendAppnameBuffer, 0, size);
2660
2661                 /* Initialize st_appname pointers. */
2662                 buffer = BackendAppnameBuffer;
2663                 for (i = 0; i < NumBackendStatSlots; i++)
2664                 {
2665                         BackendStatusArray[i].st_appname = buffer;
2666                         buffer += NAMEDATALEN;
2667                 }
2668         }
2669
2670         /* Create or attach to the shared client hostname buffer */
2671         size = mul_size(NAMEDATALEN, MaxBackends);
2672         BackendClientHostnameBuffer = (char *)
2673                 ShmemInitStruct("Backend Client Host Name Buffer", size, &found);
2674
2675         if (!found)
2676         {
2677                 MemSet(BackendClientHostnameBuffer, 0, size);
2678
2679                 /* Initialize st_clienthostname pointers. */
2680                 buffer = BackendClientHostnameBuffer;
2681                 for (i = 0; i < NumBackendStatSlots; i++)
2682                 {
2683                         BackendStatusArray[i].st_clienthostname = buffer;
2684                         buffer += NAMEDATALEN;
2685                 }
2686         }
2687
2688         /* Create or attach to the shared activity buffer */
2689         BackendActivityBufferSize = mul_size(pgstat_track_activity_query_size,
2690                                                                                  NumBackendStatSlots);
2691         BackendActivityBuffer = (char *)
2692                 ShmemInitStruct("Backend Activity Buffer",
2693                                                 BackendActivityBufferSize,
2694                                                 &found);
2695
2696         if (!found)
2697         {
2698                 MemSet(BackendActivityBuffer, 0, size);
2699
2700                 /* Initialize st_activity pointers. */
2701                 buffer = BackendActivityBuffer;
2702                 for (i = 0; i < NumBackendStatSlots; i++)
2703                 {
2704                         BackendStatusArray[i].st_activity_raw = buffer;
2705                         buffer += pgstat_track_activity_query_size;
2706                 }
2707         }
2708
2709 #ifdef USE_SSL
2710         /* Create or attach to the shared SSL status buffer */
2711         size = mul_size(sizeof(PgBackendSSLStatus), NumBackendStatSlots);
2712         BackendSslStatusBuffer = (PgBackendSSLStatus *)
2713                 ShmemInitStruct("Backend SSL Status Buffer", size, &found);
2714
2715         if (!found)
2716         {
2717                 PgBackendSSLStatus *ptr;
2718
2719                 MemSet(BackendSslStatusBuffer, 0, size);
2720
2721                 /* Initialize st_sslstatus pointers. */
2722                 ptr = BackendSslStatusBuffer;
2723                 for (i = 0; i < NumBackendStatSlots; i++)
2724                 {
2725                         BackendStatusArray[i].st_sslstatus = ptr;
2726                         ptr++;
2727                 }
2728         }
2729 #endif
2730 }
2731
2732
2733 /* ----------
2734  * pgstat_initialize() -
2735  *
2736  *      Initialize pgstats state, and set up our on-proc-exit hook.
2737  *      Called from InitPostgres and AuxiliaryProcessMain. For auxiliary process,
2738  *      MyBackendId is invalid. Otherwise, MyBackendId must be set,
2739  *      but we must not have started any transaction yet (since the
2740  *      exit hook must run after the last transaction exit).
2741  *      NOTE: MyDatabaseId isn't set yet; so the shutdown hook has to be careful.
2742  * ----------
2743  */
2744 void
2745 pgstat_initialize(void)
2746 {
2747         /* Initialize MyBEEntry */
2748         if (MyBackendId != InvalidBackendId)
2749         {
2750                 Assert(MyBackendId >= 1 && MyBackendId <= MaxBackends);
2751                 MyBEEntry = &BackendStatusArray[MyBackendId - 1];
2752         }
2753         else
2754         {
2755                 /* Must be an auxiliary process */
2756                 Assert(MyAuxProcType != NotAnAuxProcess);
2757
2758                 /*
2759                  * Assign the MyBEEntry for an auxiliary process.  Since it doesn't
2760                  * have a BackendId, the slot is statically allocated based on the
2761                  * auxiliary process type (MyAuxProcType).  Backends use slots indexed
2762                  * in the range from 1 to MaxBackends (inclusive), so we use
2763                  * MaxBackends + AuxBackendType + 1 as the index of the slot for an
2764                  * auxiliary process.
2765                  */
2766                 MyBEEntry = &BackendStatusArray[MaxBackends + MyAuxProcType];
2767         }
2768
2769         /* Set up a process-exit hook to clean up */
2770         on_shmem_exit(pgstat_beshutdown_hook, 0);
2771 }
2772
2773 /* ----------
2774  * pgstat_bestart() -
2775  *
2776  *      Initialize this backend's entry in the PgBackendStatus array.
2777  *      Called from InitPostgres.
2778  *
2779  *      Apart from auxiliary processes, MyBackendId, MyDatabaseId,
2780  *      session userid, and application_name must be set for a
2781  *      backend (hence, this cannot be combined with pgstat_initialize).
2782  * ----------
2783  */
2784 void
2785 pgstat_bestart(void)
2786 {
2787         TimestampTz proc_start_timestamp;
2788         SockAddr        clientaddr;
2789         volatile PgBackendStatus *beentry;
2790
2791         /*
2792          * To minimize the time spent modifying the PgBackendStatus entry, fetch
2793          * all the needed data first.
2794          *
2795          * If we have a MyProcPort, use its session start time (for consistency,
2796          * and to save a kernel call).
2797          */
2798         if (MyProcPort)
2799                 proc_start_timestamp = MyProcPort->SessionStartTime;
2800         else
2801                 proc_start_timestamp = GetCurrentTimestamp();
2802
2803         /*
2804          * We may not have a MyProcPort (eg, if this is the autovacuum process).
2805          * If so, use all-zeroes client address, which is dealt with specially in
2806          * pg_stat_get_backend_client_addr and pg_stat_get_backend_client_port.
2807          */
2808         if (MyProcPort)
2809                 memcpy(&clientaddr, &MyProcPort->raddr, sizeof(clientaddr));
2810         else
2811                 MemSet(&clientaddr, 0, sizeof(clientaddr));
2812
2813         /*
2814          * Initialize my status entry, following the protocol of bumping
2815          * st_changecount before and after; and make sure it's even afterwards. We
2816          * use a volatile pointer here to ensure the compiler doesn't try to get
2817          * cute.
2818          */
2819         beentry = MyBEEntry;
2820
2821         /* pgstats state must be initialized from pgstat_initialize() */
2822         Assert(beentry != NULL);
2823
2824         if (MyBackendId != InvalidBackendId)
2825         {
2826                 if (IsAutoVacuumLauncherProcess())
2827                 {
2828                         /* Autovacuum Launcher */
2829                         beentry->st_backendType = B_AUTOVAC_LAUNCHER;
2830                 }
2831                 else if (IsAutoVacuumWorkerProcess())
2832                 {
2833                         /* Autovacuum Worker */
2834                         beentry->st_backendType = B_AUTOVAC_WORKER;
2835                 }
2836                 else if (am_walsender)
2837                 {
2838                         /* Wal sender */
2839                         beentry->st_backendType = B_WAL_SENDER;
2840                 }
2841                 else if (IsBackgroundWorker)
2842                 {
2843                         /* bgworker */
2844                         beentry->st_backendType = B_BG_WORKER;
2845                 }
2846                 else
2847                 {
2848                         /* client-backend */
2849                         beentry->st_backendType = B_BACKEND;
2850                 }
2851         }
2852         else
2853         {
2854                 /* Must be an auxiliary process */
2855                 Assert(MyAuxProcType != NotAnAuxProcess);
2856                 switch (MyAuxProcType)
2857                 {
2858                         case StartupProcess:
2859                                 beentry->st_backendType = B_STARTUP;
2860                                 break;
2861                         case BgWriterProcess:
2862                                 beentry->st_backendType = B_BG_WRITER;
2863                                 break;
2864                         case CheckpointerProcess:
2865                                 beentry->st_backendType = B_CHECKPOINTER;
2866                                 break;
2867                         case WalWriterProcess:
2868                                 beentry->st_backendType = B_WAL_WRITER;
2869                                 break;
2870                         case WalReceiverProcess:
2871                                 beentry->st_backendType = B_WAL_RECEIVER;
2872                                 break;
2873                         default:
2874                                 elog(FATAL, "unrecognized process type: %d",
2875                                          (int) MyAuxProcType);
2876                                 proc_exit(1);
2877                 }
2878         }
2879
2880         do
2881         {
2882                 pgstat_increment_changecount_before(beentry);
2883         } while ((beentry->st_changecount & 1) == 0);
2884
2885         beentry->st_procpid = MyProcPid;
2886         beentry->st_proc_start_timestamp = proc_start_timestamp;
2887         beentry->st_activity_start_timestamp = 0;
2888         beentry->st_state_start_timestamp = 0;
2889         beentry->st_xact_start_timestamp = 0;
2890         beentry->st_databaseid = MyDatabaseId;
2891
2892         /* We have userid for client-backends, wal-sender and bgworker processes */
2893         if (beentry->st_backendType == B_BACKEND
2894                 || beentry->st_backendType == B_WAL_SENDER
2895                 || beentry->st_backendType == B_BG_WORKER)
2896                 beentry->st_userid = GetSessionUserId();
2897         else
2898                 beentry->st_userid = InvalidOid;
2899
2900         beentry->st_clientaddr = clientaddr;
2901         if (MyProcPort && MyProcPort->remote_hostname)
2902                 strlcpy(beentry->st_clienthostname, MyProcPort->remote_hostname,
2903                                 NAMEDATALEN);
2904         else
2905                 beentry->st_clienthostname[0] = '\0';
2906 #ifdef USE_SSL
2907         if (MyProcPort && MyProcPort->ssl != NULL)
2908         {
2909                 beentry->st_ssl = true;
2910                 beentry->st_sslstatus->ssl_bits = be_tls_get_cipher_bits(MyProcPort);
2911                 beentry->st_sslstatus->ssl_compression = be_tls_get_compression(MyProcPort);
2912                 strlcpy(beentry->st_sslstatus->ssl_version, be_tls_get_version(MyProcPort), NAMEDATALEN);
2913                 strlcpy(beentry->st_sslstatus->ssl_cipher, be_tls_get_cipher(MyProcPort), NAMEDATALEN);
2914                 be_tls_get_peerdn_name(MyProcPort, beentry->st_sslstatus->ssl_clientdn, NAMEDATALEN);
2915         }
2916         else
2917         {
2918                 beentry->st_ssl = false;
2919         }
2920 #else
2921         beentry->st_ssl = false;
2922 #endif
2923         beentry->st_state = STATE_UNDEFINED;
2924         beentry->st_appname[0] = '\0';
2925         beentry->st_activity_raw[0] = '\0';
2926         /* Also make sure the last byte in each string area is always 0 */
2927         beentry->st_clienthostname[NAMEDATALEN - 1] = '\0';
2928         beentry->st_appname[NAMEDATALEN - 1] = '\0';
2929         beentry->st_activity_raw[pgstat_track_activity_query_size - 1] = '\0';
2930         beentry->st_progress_command = PROGRESS_COMMAND_INVALID;
2931         beentry->st_progress_command_target = InvalidOid;
2932
2933         /*
2934          * we don't zero st_progress_param here to save cycles; nobody should
2935          * examine it until st_progress_command has been set to something other
2936          * than PROGRESS_COMMAND_INVALID
2937          */
2938
2939         pgstat_increment_changecount_after(beentry);
2940
2941         /* Update app name to current GUC setting */
2942         if (application_name)
2943                 pgstat_report_appname(application_name);
2944 }
2945
2946 /*
2947  * Shut down a single backend's statistics reporting at process exit.
2948  *
2949  * Flush any remaining statistics counts out to the collector.
2950  * Without this, operations triggered during backend exit (such as
2951  * temp table deletions) won't be counted.
2952  *
2953  * Lastly, clear out our entry in the PgBackendStatus array.
2954  */
2955 static void
2956 pgstat_beshutdown_hook(int code, Datum arg)
2957 {
2958         volatile PgBackendStatus *beentry = MyBEEntry;
2959
2960         /*
2961          * If we got as far as discovering our own database ID, we can report what
2962          * we did to the collector.  Otherwise, we'd be sending an invalid
2963          * database ID, so forget it.  (This means that accesses to pg_database
2964          * during failed backend starts might never get counted.)
2965          */
2966         if (OidIsValid(MyDatabaseId))
2967                 pgstat_report_stat(true);
2968
2969         /*
2970          * Clear my status entry, following the protocol of bumping st_changecount
2971          * before and after.  We use a volatile pointer here to ensure the
2972          * compiler doesn't try to get cute.
2973          */
2974         pgstat_increment_changecount_before(beentry);
2975
2976         beentry->st_procpid = 0;        /* mark invalid */
2977
2978         pgstat_increment_changecount_after(beentry);
2979 }
2980
2981
2982 /* ----------
2983  * pgstat_report_activity() -
2984  *
2985  *      Called from tcop/postgres.c to report what the backend is actually doing
2986  *      (but note cmd_str can be NULL for certain cases).
2987  *
2988  * All updates of the status entry follow the protocol of bumping
2989  * st_changecount before and after.  We use a volatile pointer here to
2990  * ensure the compiler doesn't try to get cute.
2991  * ----------
2992  */
2993 void
2994 pgstat_report_activity(BackendState state, const char *cmd_str)
2995 {
2996         volatile PgBackendStatus *beentry = MyBEEntry;
2997         TimestampTz start_timestamp;
2998         TimestampTz current_timestamp;
2999         int                     len = 0;
3000
3001         TRACE_POSTGRESQL_STATEMENT_STATUS(cmd_str);
3002
3003         if (!beentry)
3004                 return;
3005
3006         if (!pgstat_track_activities)
3007         {
3008                 if (beentry->st_state != STATE_DISABLED)
3009                 {
3010                         volatile PGPROC *proc = MyProc;
3011
3012                         /*
3013                          * track_activities is disabled, but we last reported a
3014                          * non-disabled state.  As our final update, change the state and
3015                          * clear fields we will not be updating anymore.
3016                          */
3017                         pgstat_increment_changecount_before(beentry);
3018                         beentry->st_state = STATE_DISABLED;
3019                         beentry->st_state_start_timestamp = 0;
3020                         beentry->st_activity_raw[0] = '\0';
3021                         beentry->st_activity_start_timestamp = 0;
3022                         /* st_xact_start_timestamp and wait_event_info are also disabled */
3023                         beentry->st_xact_start_timestamp = 0;
3024                         proc->wait_event_info = 0;
3025                         pgstat_increment_changecount_after(beentry);
3026                 }
3027                 return;
3028         }
3029
3030         /*
3031          * To minimize the time spent modifying the entry, fetch all the needed
3032          * data first.
3033          */
3034         start_timestamp = GetCurrentStatementStartTimestamp();
3035         if (cmd_str != NULL)
3036         {
3037                 /*
3038                  * Compute length of to-be-stored string unaware of multi-byte
3039                  * characters. For speed reasons that'll get corrected on read, rather
3040                  * than computed every write.
3041                  */
3042                 len = Min(strlen(cmd_str), pgstat_track_activity_query_size - 1);
3043         }
3044         current_timestamp = GetCurrentTimestamp();
3045
3046         /*
3047          * Now update the status entry
3048          */
3049         pgstat_increment_changecount_before(beentry);
3050
3051         beentry->st_state = state;
3052         beentry->st_state_start_timestamp = current_timestamp;
3053
3054         if (cmd_str != NULL)
3055         {
3056                 memcpy((char *) beentry->st_activity_raw, cmd_str, len);
3057                 beentry->st_activity_raw[len] = '\0';
3058                 beentry->st_activity_start_timestamp = start_timestamp;
3059         }
3060
3061         pgstat_increment_changecount_after(beentry);
3062 }
3063
3064 /*-----------
3065  * pgstat_progress_start_command() -
3066  *
3067  * Set st_progress_command (and st_progress_command_target) in own backend
3068  * entry.  Also, zero-initialize st_progress_param array.
3069  *-----------
3070  */
3071 void
3072 pgstat_progress_start_command(ProgressCommandType cmdtype, Oid relid)
3073 {
3074         volatile PgBackendStatus *beentry = MyBEEntry;
3075
3076         if (!beentry || !pgstat_track_activities)
3077                 return;
3078
3079         pgstat_increment_changecount_before(beentry);
3080         beentry->st_progress_command = cmdtype;
3081         beentry->st_progress_command_target = relid;
3082         MemSet(&beentry->st_progress_param, 0, sizeof(beentry->st_progress_param));
3083         pgstat_increment_changecount_after(beentry);
3084 }
3085
3086 /*-----------
3087  * pgstat_progress_update_param() -
3088  *
3089  * Update index'th member in st_progress_param[] of own backend entry.
3090  *-----------
3091  */
3092 void
3093 pgstat_progress_update_param(int index, int64 val)
3094 {
3095         volatile PgBackendStatus *beentry = MyBEEntry;
3096
3097         Assert(index >= 0 && index < PGSTAT_NUM_PROGRESS_PARAM);
3098
3099         if (!beentry || !pgstat_track_activities)
3100                 return;
3101
3102         pgstat_increment_changecount_before(beentry);
3103         beentry->st_progress_param[index] = val;
3104         pgstat_increment_changecount_after(beentry);
3105 }
3106
3107 /*-----------
3108  * pgstat_progress_update_multi_param() -
3109  *
3110  * Update multiple members in st_progress_param[] of own backend entry.
3111  * This is atomic; readers won't see intermediate states.
3112  *-----------
3113  */
3114 void
3115 pgstat_progress_update_multi_param(int nparam, const int *index,
3116                                                                    const int64 *val)
3117 {
3118         volatile PgBackendStatus *beentry = MyBEEntry;
3119         int                     i;
3120
3121         if (!beentry || !pgstat_track_activities || nparam == 0)
3122                 return;
3123
3124         pgstat_increment_changecount_before(beentry);
3125
3126         for (i = 0; i < nparam; ++i)
3127         {
3128                 Assert(index[i] >= 0 && index[i] < PGSTAT_NUM_PROGRESS_PARAM);
3129
3130                 beentry->st_progress_param[index[i]] = val[i];
3131         }
3132
3133         pgstat_increment_changecount_after(beentry);
3134 }
3135
3136 /*-----------
3137  * pgstat_progress_end_command() -
3138  *
3139  * Reset st_progress_command (and st_progress_command_target) in own backend
3140  * entry.  This signals the end of the command.
3141  *-----------
3142  */
3143 void
3144 pgstat_progress_end_command(void)
3145 {
3146         volatile PgBackendStatus *beentry = MyBEEntry;
3147
3148         if (!beentry)
3149                 return;
3150         if (!pgstat_track_activities
3151                 && beentry->st_progress_command == PROGRESS_COMMAND_INVALID)
3152                 return;
3153
3154         pgstat_increment_changecount_before(beentry);
3155         beentry->st_progress_command = PROGRESS_COMMAND_INVALID;
3156         beentry->st_progress_command_target = InvalidOid;
3157         pgstat_increment_changecount_after(beentry);
3158 }
3159
3160 /* ----------
3161  * pgstat_report_appname() -
3162  *
3163  *      Called to update our application name.
3164  * ----------
3165  */
3166 void
3167 pgstat_report_appname(const char *appname)
3168 {
3169         volatile PgBackendStatus *beentry = MyBEEntry;
3170         int                     len;
3171
3172         if (!beentry)
3173                 return;
3174
3175         /* This should be unnecessary if GUC did its job, but be safe */
3176         len = pg_mbcliplen(appname, strlen(appname), NAMEDATALEN - 1);
3177
3178         /*
3179          * Update my status entry, following the protocol of bumping
3180          * st_changecount before and after.  We use a volatile pointer here to
3181          * ensure the compiler doesn't try to get cute.
3182          */
3183         pgstat_increment_changecount_before(beentry);
3184
3185         memcpy((char *) beentry->st_appname, appname, len);
3186         beentry->st_appname[len] = '\0';
3187
3188         pgstat_increment_changecount_after(beentry);
3189 }
3190
3191 /*
3192  * Report current transaction start timestamp as the specified value.
3193  * Zero means there is no active transaction.
3194  */
3195 void
3196 pgstat_report_xact_timestamp(TimestampTz tstamp)
3197 {
3198         volatile PgBackendStatus *beentry = MyBEEntry;
3199
3200         if (!pgstat_track_activities || !beentry)
3201                 return;
3202
3203         /*
3204          * Update my status entry, following the protocol of bumping
3205          * st_changecount before and after.  We use a volatile pointer here to
3206          * ensure the compiler doesn't try to get cute.
3207          */
3208         pgstat_increment_changecount_before(beentry);
3209         beentry->st_xact_start_timestamp = tstamp;
3210         pgstat_increment_changecount_after(beentry);
3211 }
3212
3213 /* ----------
3214  * pgstat_read_current_status() -
3215  *
3216  *      Copy the current contents of the PgBackendStatus array to local memory,
3217  *      if not already done in this transaction.
3218  * ----------
3219  */
3220 static void
3221 pgstat_read_current_status(void)
3222 {
3223         volatile PgBackendStatus *beentry;
3224         LocalPgBackendStatus *localtable;
3225         LocalPgBackendStatus *localentry;
3226         char       *localappname,
3227                            *localclienthostname,
3228                            *localactivity;
3229 #ifdef USE_SSL
3230         PgBackendSSLStatus *localsslstatus;
3231 #endif
3232         int                     i;
3233
3234         Assert(!pgStatRunningInCollector);
3235         if (localBackendStatusTable)
3236                 return;                                 /* already done */
3237
3238         pgstat_setup_memcxt();
3239
3240         localtable = (LocalPgBackendStatus *)
3241                 MemoryContextAlloc(pgStatLocalContext,
3242                                                    sizeof(LocalPgBackendStatus) * NumBackendStatSlots);
3243         localappname = (char *)
3244                 MemoryContextAlloc(pgStatLocalContext,
3245                                                    NAMEDATALEN * NumBackendStatSlots);
3246         localclienthostname = (char *)
3247                 MemoryContextAlloc(pgStatLocalContext,
3248                                                    NAMEDATALEN * NumBackendStatSlots);
3249         localactivity = (char *)
3250                 MemoryContextAlloc(pgStatLocalContext,
3251                                                    pgstat_track_activity_query_size * NumBackendStatSlots);
3252 #ifdef USE_SSL
3253         localsslstatus = (PgBackendSSLStatus *)
3254                 MemoryContextAlloc(pgStatLocalContext,
3255                                                    sizeof(PgBackendSSLStatus) * NumBackendStatSlots);
3256 #endif
3257
3258         localNumBackends = 0;
3259
3260         beentry = BackendStatusArray;
3261         localentry = localtable;
3262         for (i = 1; i <= NumBackendStatSlots; i++)
3263         {
3264                 /*
3265                  * Follow the protocol of retrying if st_changecount changes while we
3266                  * copy the entry, or if it's odd.  (The check for odd is needed to
3267                  * cover the case where we are able to completely copy the entry while
3268                  * the source backend is between increment steps.)      We use a volatile
3269                  * pointer here to ensure the compiler doesn't try to get cute.
3270                  */
3271                 for (;;)
3272                 {
3273                         int                     before_changecount;
3274                         int                     after_changecount;
3275
3276                         pgstat_save_changecount_before(beentry, before_changecount);
3277
3278                         localentry->backendStatus.st_procpid = beentry->st_procpid;
3279                         if (localentry->backendStatus.st_procpid > 0)
3280                         {
3281                                 memcpy(&localentry->backendStatus, (char *) beentry, sizeof(PgBackendStatus));
3282
3283                                 /*
3284                                  * strcpy is safe even if the string is modified concurrently,
3285                                  * because there's always a \0 at the end of the buffer.
3286                                  */
3287                                 strcpy(localappname, (char *) beentry->st_appname);
3288                                 localentry->backendStatus.st_appname = localappname;
3289                                 strcpy(localclienthostname, (char *) beentry->st_clienthostname);
3290                                 localentry->backendStatus.st_clienthostname = localclienthostname;
3291                                 strcpy(localactivity, (char *) beentry->st_activity_raw);
3292                                 localentry->backendStatus.st_activity_raw = localactivity;
3293                                 localentry->backendStatus.st_ssl = beentry->st_ssl;
3294 #ifdef USE_SSL
3295                                 if (beentry->st_ssl)
3296                                 {
3297                                         memcpy(localsslstatus, beentry->st_sslstatus, sizeof(PgBackendSSLStatus));
3298                                         localentry->backendStatus.st_sslstatus = localsslstatus;
3299                                 }
3300 #endif
3301                         }
3302
3303                         pgstat_save_changecount_after(beentry, after_changecount);
3304                         if (before_changecount == after_changecount &&
3305                                 (before_changecount & 1) == 0)
3306                                 break;
3307
3308                         /* Make sure we can break out of loop if stuck... */
3309                         CHECK_FOR_INTERRUPTS();
3310                 }
3311
3312                 beentry++;
3313                 /* Only valid entries get included into the local array */
3314                 if (localentry->backendStatus.st_procpid > 0)
3315                 {
3316                         BackendIdGetTransactionIds(i,
3317                                                                            &localentry->backend_xid,
3318                                                                            &localentry->backend_xmin);
3319
3320                         localentry++;
3321                         localappname += NAMEDATALEN;
3322                         localclienthostname += NAMEDATALEN;
3323                         localactivity += pgstat_track_activity_query_size;
3324 #ifdef USE_SSL
3325                         localsslstatus++;
3326 #endif
3327                         localNumBackends++;
3328                 }
3329         }
3330
3331         /* Set the pointer only after completion of a valid table */
3332         localBackendStatusTable = localtable;
3333 }
3334
3335 /* ----------
3336  * pgstat_get_wait_event_type() -
3337  *
3338  *      Return a string representing the current wait event type, backend is
3339  *      waiting on.
3340  */
3341 const char *
3342 pgstat_get_wait_event_type(uint32 wait_event_info)
3343 {
3344         uint32          classId;
3345         const char *event_type;
3346
3347         /* report process as not waiting. */
3348         if (wait_event_info == 0)
3349                 return NULL;
3350
3351         classId = wait_event_info & 0xFF000000;
3352
3353         switch (classId)
3354         {
3355                 case PG_WAIT_LWLOCK:
3356                         event_type = "LWLock";
3357                         break;
3358                 case PG_WAIT_LOCK:
3359                         event_type = "Lock";
3360                         break;
3361                 case PG_WAIT_BUFFER_PIN:
3362                         event_type = "BufferPin";
3363                         break;
3364                 case PG_WAIT_ACTIVITY:
3365                         event_type = "Activity";
3366                         break;
3367                 case PG_WAIT_CLIENT:
3368                         event_type = "Client";
3369                         break;
3370                 case PG_WAIT_EXTENSION:
3371                         event_type = "Extension";
3372                         break;
3373                 case PG_WAIT_IPC:
3374                         event_type = "IPC";
3375                         break;
3376                 case PG_WAIT_TIMEOUT:
3377                         event_type = "Timeout";
3378                         break;
3379                 case PG_WAIT_IO:
3380                         event_type = "IO";
3381                         break;
3382                 default:
3383                         event_type = "???";
3384                         break;
3385         }
3386
3387         return event_type;
3388 }
3389
3390 /* ----------
3391  * pgstat_get_wait_event() -
3392  *
3393  *      Return a string representing the current wait event, backend is
3394  *      waiting on.
3395  */
3396 const char *
3397 pgstat_get_wait_event(uint32 wait_event_info)
3398 {
3399         uint32          classId;
3400         uint16          eventId;
3401         const char *event_name;
3402
3403         /* report process as not waiting. */
3404         if (wait_event_info == 0)
3405                 return NULL;
3406
3407         classId = wait_event_info & 0xFF000000;
3408         eventId = wait_event_info & 0x0000FFFF;
3409
3410         switch (classId)
3411         {
3412                 case PG_WAIT_LWLOCK:
3413                         event_name = GetLWLockIdentifier(classId, eventId);
3414                         break;
3415                 case PG_WAIT_LOCK:
3416                         event_name = GetLockNameFromTagType(eventId);
3417                         break;
3418                 case PG_WAIT_BUFFER_PIN:
3419                         event_name = "BufferPin";
3420                         break;
3421                 case PG_WAIT_ACTIVITY:
3422                         {
3423                                 WaitEventActivity w = (WaitEventActivity) wait_event_info;
3424
3425                                 event_name = pgstat_get_wait_activity(w);
3426                                 break;
3427                         }
3428                 case PG_WAIT_CLIENT:
3429                         {
3430                                 WaitEventClient w = (WaitEventClient) wait_event_info;
3431
3432                                 event_name = pgstat_get_wait_client(w);
3433                                 break;
3434                         }
3435                 case PG_WAIT_EXTENSION:
3436                         event_name = "Extension";
3437                         break;
3438                 case PG_WAIT_IPC:
3439                         {
3440                                 WaitEventIPC w = (WaitEventIPC) wait_event_info;
3441
3442                                 event_name = pgstat_get_wait_ipc(w);
3443                                 break;
3444                         }
3445                 case PG_WAIT_TIMEOUT:
3446                         {
3447                                 WaitEventTimeout w = (WaitEventTimeout) wait_event_info;
3448
3449                                 event_name = pgstat_get_wait_timeout(w);
3450                                 break;
3451                         }
3452                 case PG_WAIT_IO:
3453                         {
3454                                 WaitEventIO w = (WaitEventIO) wait_event_info;
3455
3456                                 event_name = pgstat_get_wait_io(w);
3457                                 break;
3458                         }
3459                 default:
3460                         event_name = "unknown wait event";
3461                         break;
3462         }
3463
3464         return event_name;
3465 }
3466
3467 /* ----------
3468  * pgstat_get_wait_activity() -
3469  *
3470  * Convert WaitEventActivity to string.
3471  * ----------
3472  */
3473 static const char *
3474 pgstat_get_wait_activity(WaitEventActivity w)
3475 {
3476         const char *event_name = "unknown wait event";
3477
3478         switch (w)
3479         {
3480                 case WAIT_EVENT_ARCHIVER_MAIN:
3481                         event_name = "ArchiverMain";
3482                         break;
3483                 case WAIT_EVENT_AUTOVACUUM_MAIN:
3484                         event_name = "AutoVacuumMain";
3485                         break;
3486                 case WAIT_EVENT_BGWRITER_HIBERNATE:
3487                         event_name = "BgWriterHibernate";
3488                         break;
3489                 case WAIT_EVENT_BGWRITER_MAIN:
3490                         event_name = "BgWriterMain";
3491                         break;
3492                 case WAIT_EVENT_CHECKPOINTER_MAIN:
3493                         event_name = "CheckpointerMain";
3494                         break;
3495                 case WAIT_EVENT_LOGICAL_LAUNCHER_MAIN:
3496                         event_name = "LogicalLauncherMain";
3497                         break;
3498                 case WAIT_EVENT_LOGICAL_APPLY_MAIN:
3499                         event_name = "LogicalApplyMain";
3500                         break;
3501                 case WAIT_EVENT_PGSTAT_MAIN:
3502                         event_name = "PgStatMain";
3503                         break;
3504                 case WAIT_EVENT_RECOVERY_WAL_ALL:
3505                         event_name = "RecoveryWalAll";
3506                         break;
3507                 case WAIT_EVENT_RECOVERY_WAL_STREAM:
3508                         event_name = "RecoveryWalStream";
3509                         break;
3510                 case WAIT_EVENT_SYSLOGGER_MAIN:
3511                         event_name = "SysLoggerMain";
3512                         break;
3513                 case WAIT_EVENT_WAL_RECEIVER_MAIN:
3514                         event_name = "WalReceiverMain";
3515                         break;
3516                 case WAIT_EVENT_WAL_SENDER_MAIN:
3517                         event_name = "WalSenderMain";
3518                         break;
3519                 case WAIT_EVENT_WAL_WRITER_MAIN:
3520                         event_name = "WalWriterMain";
3521                         break;
3522                         /* no default case, so that compiler will warn */
3523         }
3524
3525         return event_name;
3526 }
3527
3528 /* ----------
3529  * pgstat_get_wait_client() -
3530  *
3531  * Convert WaitEventClient to string.
3532  * ----------
3533  */
3534 static const char *
3535 pgstat_get_wait_client(WaitEventClient w)
3536 {
3537         const char *event_name = "unknown wait event";
3538
3539         switch (w)
3540         {
3541                 case WAIT_EVENT_CLIENT_READ:
3542                         event_name = "ClientRead";
3543                         break;
3544                 case WAIT_EVENT_CLIENT_WRITE:
3545                         event_name = "ClientWrite";
3546                         break;
3547                 case WAIT_EVENT_LIBPQWALRECEIVER_CONNECT:
3548                         event_name = "LibPQWalReceiverConnect";
3549                         break;
3550                 case WAIT_EVENT_LIBPQWALRECEIVER_RECEIVE:
3551                         event_name = "LibPQWalReceiverReceive";
3552                         break;
3553                 case WAIT_EVENT_SSL_OPEN_SERVER:
3554                         event_name = "SSLOpenServer";
3555                         break;
3556                 case WAIT_EVENT_WAL_RECEIVER_WAIT_START:
3557                         event_name = "WalReceiverWaitStart";
3558                         break;
3559                 case WAIT_EVENT_WAL_SENDER_WAIT_WAL:
3560                         event_name = "WalSenderWaitForWAL";
3561                         break;
3562                 case WAIT_EVENT_WAL_SENDER_WRITE_DATA:
3563                         event_name = "WalSenderWriteData";
3564                         break;
3565                         /* no default case, so that compiler will warn */
3566         }
3567
3568         return event_name;
3569 }
3570
3571 /* ----------
3572  * pgstat_get_wait_ipc() -
3573  *
3574  * Convert WaitEventIPC to string.
3575  * ----------
3576  */
3577 static const char *
3578 pgstat_get_wait_ipc(WaitEventIPC w)
3579 {
3580         const char *event_name = "unknown wait event";
3581
3582         switch (w)
3583         {
3584                 case WAIT_EVENT_BGWORKER_SHUTDOWN:
3585                         event_name = "BgWorkerShutdown";
3586                         break;
3587                 case WAIT_EVENT_BGWORKER_STARTUP:
3588                         event_name = "BgWorkerStartup";
3589                         break;
3590                 case WAIT_EVENT_BTREE_PAGE:
3591                         event_name = "BtreePage";
3592                         break;
3593                 case WAIT_EVENT_EXECUTE_GATHER:
3594                         event_name = "ExecuteGather";
3595                         break;
3596                 case WAIT_EVENT_HASH_BATCH_ALLOCATING:
3597                         event_name = "Hash/Batch/Allocating";
3598                         break;
3599                 case WAIT_EVENT_HASH_BATCH_ELECTING:
3600                         event_name = "Hash/Batch/Electing";
3601                         break;
3602                 case WAIT_EVENT_HASH_BATCH_LOADING:
3603                         event_name = "Hash/Batch/Loading";
3604                         break;
3605                 case WAIT_EVENT_HASH_BUILD_ALLOCATING:
3606                         event_name = "Hash/Build/Allocating";
3607                         break;
3608                 case WAIT_EVENT_HASH_BUILD_ELECTING:
3609                         event_name = "Hash/Build/Electing";
3610                         break;
3611                 case WAIT_EVENT_HASH_BUILD_HASHING_INNER:
3612                         event_name = "Hash/Build/HashingInner";
3613                         break;
3614                 case WAIT_EVENT_HASH_BUILD_HASHING_OUTER:
3615                         event_name = "Hash/Build/HashingOuter";
3616                         break;
3617                 case WAIT_EVENT_HASH_GROW_BATCHES_ALLOCATING:
3618                         event_name = "Hash/GrowBatches/Allocating";
3619                         break;
3620                 case WAIT_EVENT_HASH_GROW_BATCHES_DECIDING:
3621                         event_name = "Hash/GrowBatches/Deciding";
3622                         break;
3623                 case WAIT_EVENT_HASH_GROW_BATCHES_ELECTING:
3624                         event_name = "Hash/GrowBatches/Electing";
3625                         break;
3626                 case WAIT_EVENT_HASH_GROW_BATCHES_FINISHING:
3627                         event_name = "Hash/GrowBatches/Finishing";
3628                         break;
3629                 case WAIT_EVENT_HASH_GROW_BATCHES_REPARTITIONING:
3630                         event_name = "Hash/GrowBatches/Repartitioning";
3631                         break;
3632                 case WAIT_EVENT_HASH_GROW_BUCKETS_ALLOCATING:
3633                         event_name = "Hash/GrowBuckets/Allocating";
3634                         break;
3635                 case WAIT_EVENT_HASH_GROW_BUCKETS_ELECTING:
3636                         event_name = "Hash/GrowBuckets/Electing";
3637                         break;
3638                 case WAIT_EVENT_HASH_GROW_BUCKETS_REINSERTING:
3639                         event_name = "Hash/GrowBuckets/Reinserting";
3640                         break;
3641                 case WAIT_EVENT_LOGICAL_SYNC_DATA:
3642                         event_name = "LogicalSyncData";
3643                         break;
3644                 case WAIT_EVENT_LOGICAL_SYNC_STATE_CHANGE:
3645                         event_name = "LogicalSyncStateChange";
3646                         break;
3647                 case WAIT_EVENT_MQ_INTERNAL:
3648                         event_name = "MessageQueueInternal";
3649                         break;
3650                 case WAIT_EVENT_MQ_PUT_MESSAGE:
3651                         event_name = "MessageQueuePutMessage";
3652                         break;
3653                 case WAIT_EVENT_MQ_RECEIVE:
3654                         event_name = "MessageQueueReceive";
3655                         break;
3656                 case WAIT_EVENT_MQ_SEND:
3657                         event_name = "MessageQueueSend";
3658                         break;
3659                 case WAIT_EVENT_PARALLEL_FINISH:
3660                         event_name = "ParallelFinish";
3661                         break;
3662                 case WAIT_EVENT_PARALLEL_BITMAP_SCAN:
3663                         event_name = "ParallelBitmapScan";
3664                         break;
3665                 case WAIT_EVENT_PARALLEL_CREATE_INDEX_SCAN:
3666                         event_name = "ParallelCreateIndexScan";
3667                         break;
3668                 case WAIT_EVENT_PROCARRAY_GROUP_UPDATE:
3669                         event_name = "ProcArrayGroupUpdate";
3670                         break;
3671                 case WAIT_EVENT_CLOG_GROUP_UPDATE:
3672                         event_name = "ClogGroupUpdate";
3673                         break;
3674                 case WAIT_EVENT_REPLICATION_ORIGIN_DROP:
3675                         event_name = "ReplicationOriginDrop";
3676                         break;
3677                 case WAIT_EVENT_REPLICATION_SLOT_DROP:
3678                         event_name = "ReplicationSlotDrop";
3679                         break;
3680                 case WAIT_EVENT_SAFE_SNAPSHOT:
3681                         event_name = "SafeSnapshot";
3682                         break;
3683                 case WAIT_EVENT_SYNC_REP:
3684                         event_name = "SyncRep";
3685                         break;
3686                         /* no default case, so that compiler will warn */
3687         }
3688
3689         return event_name;
3690 }
3691
3692 /* ----------
3693  * pgstat_get_wait_timeout() -
3694  *
3695  * Convert WaitEventTimeout to string.
3696  * ----------
3697  */
3698 static const char *
3699 pgstat_get_wait_timeout(WaitEventTimeout w)
3700 {
3701         const char *event_name = "unknown wait event";
3702
3703         switch (w)
3704         {
3705                 case WAIT_EVENT_BASE_BACKUP_THROTTLE:
3706                         event_name = "BaseBackupThrottle";
3707                         break;
3708                 case WAIT_EVENT_PG_SLEEP:
3709                         event_name = "PgSleep";
3710                         break;
3711                 case WAIT_EVENT_RECOVERY_APPLY_DELAY:
3712                         event_name = "RecoveryApplyDelay";
3713                         break;
3714                         /* no default case, so that compiler will warn */
3715         }
3716
3717         return event_name;
3718 }
3719
3720 /* ----------
3721  * pgstat_get_wait_io() -
3722  *
3723  * Convert WaitEventIO to string.
3724  * ----------
3725  */
3726 static const char *
3727 pgstat_get_wait_io(WaitEventIO w)
3728 {
3729         const char *event_name = "unknown wait event";
3730
3731         switch (w)
3732         {
3733                 case WAIT_EVENT_BUFFILE_READ:
3734                         event_name = "BufFileRead";
3735                         break;
3736                 case WAIT_EVENT_BUFFILE_WRITE:
3737                         event_name = "BufFileWrite";
3738                         break;
3739                 case WAIT_EVENT_CONTROL_FILE_READ:
3740                         event_name = "ControlFileRead";
3741                         break;
3742                 case WAIT_EVENT_CONTROL_FILE_SYNC:
3743                         event_name = "ControlFileSync";
3744                         break;
3745                 case WAIT_EVENT_CONTROL_FILE_SYNC_UPDATE:
3746                         event_name = "ControlFileSyncUpdate";
3747                         break;
3748                 case WAIT_EVENT_CONTROL_FILE_WRITE:
3749                         event_name = "ControlFileWrite";
3750                         break;
3751                 case WAIT_EVENT_CONTROL_FILE_WRITE_UPDATE:
3752                         event_name = "ControlFileWriteUpdate";
3753                         break;
3754                 case WAIT_EVENT_COPY_FILE_READ:
3755                         event_name = "CopyFileRead";
3756                         break;
3757                 case WAIT_EVENT_COPY_FILE_WRITE:
3758                         event_name = "CopyFileWrite";
3759                         break;
3760                 case WAIT_EVENT_DATA_FILE_EXTEND:
3761                         event_name = "DataFileExtend";
3762                         break;
3763                 case WAIT_EVENT_DATA_FILE_FLUSH:
3764                         event_name = "DataFileFlush";
3765                         break;
3766                 case WAIT_EVENT_DATA_FILE_IMMEDIATE_SYNC:
3767                         event_name = "DataFileImmediateSync";
3768                         break;
3769                 case WAIT_EVENT_DATA_FILE_PREFETCH:
3770                         event_name = "DataFilePrefetch";
3771                         break;
3772                 case WAIT_EVENT_DATA_FILE_READ:
3773                         event_name = "DataFileRead";
3774                         break;
3775                 case WAIT_EVENT_DATA_FILE_SYNC:
3776                         event_name = "DataFileSync";
3777                         break;
3778                 case WAIT_EVENT_DATA_FILE_TRUNCATE:
3779                         event_name = "DataFileTruncate";
3780                         break;
3781                 case WAIT_EVENT_DATA_FILE_WRITE:
3782                         event_name = "DataFileWrite";
3783                         break;
3784                 case WAIT_EVENT_DSM_FILL_ZERO_WRITE:
3785                         event_name = "DSMFillZeroWrite";
3786                         break;
3787                 case WAIT_EVENT_LOCK_FILE_ADDTODATADIR_READ:
3788                         event_name = "LockFileAddToDataDirRead";
3789                         break;
3790                 case WAIT_EVENT_LOCK_FILE_ADDTODATADIR_SYNC:
3791                         event_name = "LockFileAddToDataDirSync";
3792                         break;
3793                 case WAIT_EVENT_LOCK_FILE_ADDTODATADIR_WRITE:
3794                         event_name = "LockFileAddToDataDirWrite";
3795                         break;
3796                 case WAIT_EVENT_LOCK_FILE_CREATE_READ:
3797                         event_name = "LockFileCreateRead";
3798                         break;
3799                 case WAIT_EVENT_LOCK_FILE_CREATE_SYNC:
3800                         event_name = "LockFileCreateSync";
3801                         break;
3802                 case WAIT_EVENT_LOCK_FILE_CREATE_WRITE:
3803                         event_name = "LockFileCreateWRITE";
3804                         break;
3805                 case WAIT_EVENT_LOCK_FILE_RECHECKDATADIR_READ:
3806                         event_name = "LockFileReCheckDataDirRead";
3807                         break;
3808                 case WAIT_EVENT_LOGICAL_REWRITE_CHECKPOINT_SYNC:
3809                         event_name = "LogicalRewriteCheckpointSync";
3810                         break;
3811                 case WAIT_EVENT_LOGICAL_REWRITE_MAPPING_SYNC:
3812                         event_name = "LogicalRewriteMappingSync";
3813                         break;
3814                 case WAIT_EVENT_LOGICAL_REWRITE_MAPPING_WRITE:
3815                         event_name = "LogicalRewriteMappingWrite";
3816                         break;
3817                 case WAIT_EVENT_LOGICAL_REWRITE_SYNC:
3818                         event_name = "LogicalRewriteSync";
3819                         break;
3820                 case WAIT_EVENT_LOGICAL_REWRITE_TRUNCATE:
3821                         event_name = "LogicalRewriteTruncate";
3822                         break;
3823                 case WAIT_EVENT_LOGICAL_REWRITE_WRITE:
3824                         event_name = "LogicalRewriteWrite";
3825                         break;
3826                 case WAIT_EVENT_RELATION_MAP_READ:
3827                         event_name = "RelationMapRead";
3828                         break;
3829                 case WAIT_EVENT_RELATION_MAP_SYNC:
3830                         event_name = "RelationMapSync";
3831                         break;
3832                 case WAIT_EVENT_RELATION_MAP_WRITE:
3833                         event_name = "RelationMapWrite";
3834                         break;
3835                 case WAIT_EVENT_REORDER_BUFFER_READ:
3836                         event_name = "ReorderBufferRead";
3837                         break;
3838                 case WAIT_EVENT_REORDER_BUFFER_WRITE:
3839                         event_name = "ReorderBufferWrite";
3840                         break;
3841                 case WAIT_EVENT_REORDER_LOGICAL_MAPPING_READ:
3842                         event_name = "ReorderLogicalMappingRead";
3843                         break;
3844                 case WAIT_EVENT_REPLICATION_SLOT_READ:
3845                         event_name = "ReplicationSlotRead";
3846                         break;
3847                 case WAIT_EVENT_REPLICATION_SLOT_RESTORE_SYNC:
3848                         event_name = "ReplicationSlotRestoreSync";
3849                         break;
3850                 case WAIT_EVENT_REPLICATION_SLOT_SYNC:
3851                         event_name = "ReplicationSlotSync";
3852                         break;
3853                 case WAIT_EVENT_REPLICATION_SLOT_WRITE:
3854                         event_name = "ReplicationSlotWrite";
3855                         break;
3856                 case WAIT_EVENT_SLRU_FLUSH_SYNC:
3857                         event_name = "SLRUFlushSync";
3858                         break;
3859                 case WAIT_EVENT_SLRU_READ:
3860                         event_name = "SLRURead";
3861                         break;
3862                 case WAIT_EVENT_SLRU_SYNC:
3863                         event_name = "SLRUSync";
3864                         break;
3865                 case WAIT_EVENT_SLRU_WRITE:
3866                         event_name = "SLRUWrite";
3867                         break;
3868                 case WAIT_EVENT_SNAPBUILD_READ:
3869                         event_name = "SnapbuildRead";
3870                         break;
3871                 case WAIT_EVENT_SNAPBUILD_SYNC:
3872                         event_name = "SnapbuildSync";
3873                         break;
3874                 case WAIT_EVENT_SNAPBUILD_WRITE:
3875                         event_name = "SnapbuildWrite";
3876                         break;
3877                 case WAIT_EVENT_TIMELINE_HISTORY_FILE_SYNC:
3878                         event_name = "TimelineHistoryFileSync";
3879                         break;
3880                 case WAIT_EVENT_TIMELINE_HISTORY_FILE_WRITE:
3881                         event_name = "TimelineHistoryFileWrite";
3882                         break;
3883                 case WAIT_EVENT_TIMELINE_HISTORY_READ:
3884                         event_name = "TimelineHistoryRead";
3885                         break;
3886                 case WAIT_EVENT_TIMELINE_HISTORY_SYNC:
3887                         event_name = "TimelineHistorySync";
3888                         break;
3889                 case WAIT_EVENT_TIMELINE_HISTORY_WRITE:
3890                         event_name = "TimelineHistoryWrite";
3891                         break;
3892                 case WAIT_EVENT_TWOPHASE_FILE_READ:
3893                         event_name = "TwophaseFileRead";
3894                         break;
3895                 case WAIT_EVENT_TWOPHASE_FILE_SYNC:
3896                         event_name = "TwophaseFileSync";
3897                         break;
3898                 case WAIT_EVENT_TWOPHASE_FILE_WRITE:
3899                         event_name = "TwophaseFileWrite";
3900                         break;
3901                 case WAIT_EVENT_WALSENDER_TIMELINE_HISTORY_READ:
3902                         event_name = "WALSenderTimelineHistoryRead";
3903                         break;
3904                 case WAIT_EVENT_WAL_BOOTSTRAP_SYNC:
3905                         event_name = "WALBootstrapSync";
3906                         break;
3907                 case WAIT_EVENT_WAL_BOOTSTRAP_WRITE:
3908                         event_name = "WALBootstrapWrite";
3909                         break;
3910                 case WAIT_EVENT_WAL_COPY_READ:
3911                         event_name = "WALCopyRead";
3912                         break;
3913                 case WAIT_EVENT_WAL_COPY_SYNC:
3914                         event_name = "WALCopySync";
3915                         break;
3916                 case WAIT_EVENT_WAL_COPY_WRITE:
3917                         event_name = "WALCopyWrite";
3918                         break;
3919                 case WAIT_EVENT_WAL_INIT_SYNC:
3920                         event_name = "WALInitSync";
3921                         break;
3922                 case WAIT_EVENT_WAL_INIT_WRITE:
3923                         event_name = "WALInitWrite";
3924                         break;
3925                 case WAIT_EVENT_WAL_READ:
3926                         event_name = "WALRead";
3927                         break;
3928                 case WAIT_EVENT_WAL_SYNC_METHOD_ASSIGN:
3929                         event_name = "WALSyncMethodAssign";
3930                         break;
3931                 case WAIT_EVENT_WAL_WRITE:
3932                         event_name = "WALWrite";
3933                         break;
3934
3935                         /* no default case, so that compiler will warn */
3936         }
3937
3938         return event_name;
3939 }
3940
3941
3942 /* ----------
3943  * pgstat_get_backend_current_activity() -
3944  *
3945  *      Return a string representing the current activity of the backend with
3946  *      the specified PID.  This looks directly at the BackendStatusArray,
3947  *      and so will provide current information regardless of the age of our
3948  *      transaction's snapshot of the status array.
3949  *
3950  *      It is the caller's responsibility to invoke this only for backends whose
3951  *      state is expected to remain stable while the result is in use.  The
3952  *      only current use is in deadlock reporting, where we can expect that
3953  *      the target backend is blocked on a lock.  (There are corner cases
3954  *      where the target's wait could get aborted while we are looking at it,
3955  *      but the very worst consequence is to return a pointer to a string
3956  *      that's been changed, so we won't worry too much.)
3957  *
3958  *      Note: return strings for special cases match pg_stat_get_backend_activity.
3959  * ----------
3960  */
3961 const char *
3962 pgstat_get_backend_current_activity(int pid, bool checkUser)
3963 {
3964         PgBackendStatus *beentry;
3965         int                     i;
3966
3967         beentry = BackendStatusArray;
3968         for (i = 1; i <= MaxBackends; i++)
3969         {
3970                 /*
3971                  * Although we expect the target backend's entry to be stable, that
3972                  * doesn't imply that anyone else's is.  To avoid identifying the
3973                  * wrong backend, while we check for a match to the desired PID we
3974                  * must follow the protocol of retrying if st_changecount changes
3975                  * while we examine the entry, or if it's odd.  (This might be
3976                  * unnecessary, since fetching or storing an int is almost certainly
3977                  * atomic, but let's play it safe.)  We use a volatile pointer here to
3978                  * ensure the compiler doesn't try to get cute.
3979                  */
3980                 volatile PgBackendStatus *vbeentry = beentry;
3981                 bool            found;
3982
3983                 for (;;)
3984                 {
3985                         int                     before_changecount;
3986                         int                     after_changecount;
3987
3988                         pgstat_save_changecount_before(vbeentry, before_changecount);
3989
3990                         found = (vbeentry->st_procpid == pid);
3991
3992                         pgstat_save_changecount_after(vbeentry, after_changecount);
3993
3994                         if (before_changecount == after_changecount &&
3995                                 (before_changecount & 1) == 0)
3996                                 break;
3997
3998                         /* Make sure we can break out of loop if stuck... */
3999                         CHECK_FOR_INTERRUPTS();
4000                 }
4001
4002                 if (found)
4003                 {
4004                         /* Now it is safe to use the non-volatile pointer */
4005                         if (checkUser && !superuser() && beentry->st_userid != GetUserId())
4006                                 return "<insufficient privilege>";
4007                         else if (*(beentry->st_activity_raw) == '\0')
4008                                 return "<command string not enabled>";
4009                         else
4010                         {
4011                                 /* this'll leak a bit of memory, but that seems acceptable */
4012                                 return pgstat_clip_activity(beentry->st_activity_raw);
4013                         }
4014                 }
4015
4016                 beentry++;
4017         }
4018
4019         /* If we get here, caller is in error ... */
4020         return "<backend information not available>";
4021 }
4022
4023 /* ----------
4024  * pgstat_get_crashed_backend_activity() -
4025  *
4026  *      Return a string representing the current activity of the backend with
4027  *      the specified PID.  Like the function above, but reads shared memory with
4028  *      the expectation that it may be corrupt.  On success, copy the string
4029  *      into the "buffer" argument and return that pointer.  On failure,
4030  *      return NULL.
4031  *
4032  *      This function is only intended to be used by the postmaster to report the
4033  *      query that crashed a backend.  In particular, no attempt is made to
4034  *      follow the correct concurrency protocol when accessing the
4035  *      BackendStatusArray.  But that's OK, in the worst case we'll return a
4036  *      corrupted message.  We also must take care not to trip on ereport(ERROR).
4037  * ----------
4038  */
4039 const char *
4040 pgstat_get_crashed_backend_activity(int pid, char *buffer, int buflen)
4041 {
4042         volatile PgBackendStatus *beentry;
4043         int                     i;
4044
4045         beentry = BackendStatusArray;
4046
4047         /*
4048          * We probably shouldn't get here before shared memory has been set up,
4049          * but be safe.
4050          */
4051         if (beentry == NULL || BackendActivityBuffer == NULL)
4052                 return NULL;
4053
4054         for (i = 1; i <= MaxBackends; i++)
4055         {
4056                 if (beentry->st_procpid == pid)
4057                 {
4058                         /* Read pointer just once, so it can't change after validation */
4059                         const char *activity = beentry->st_activity_raw;
4060                         const char *activity_last;
4061
4062                         /*
4063                          * We mustn't access activity string before we verify that it
4064                          * falls within the BackendActivityBuffer. To make sure that the
4065                          * entire string including its ending is contained within the
4066                          * buffer, subtract one activity length from the buffer size.
4067                          */
4068                         activity_last = BackendActivityBuffer + BackendActivityBufferSize
4069                                 - pgstat_track_activity_query_size;
4070
4071                         if (activity < BackendActivityBuffer ||
4072                                 activity > activity_last)
4073                                 return NULL;
4074
4075                         /* If no string available, no point in a report */
4076                         if (activity[0] == '\0')
4077                                 return NULL;
4078
4079                         /*
4080                          * Copy only ASCII-safe characters so we don't run into encoding
4081                          * problems when reporting the message; and be sure not to run off
4082                          * the end of memory.  As only ASCII characters are reported, it
4083                          * doesn't seem necessary to perform multibyte aware clipping.
4084                          */
4085                         ascii_safe_strlcpy(buffer, activity,
4086                                                            Min(buflen, pgstat_track_activity_query_size));
4087
4088                         return buffer;
4089                 }
4090
4091                 beentry++;
4092         }
4093
4094         /* PID not found */
4095         return NULL;
4096 }
4097
4098 const char *
4099 pgstat_get_backend_desc(BackendType backendType)
4100 {
4101         const char *backendDesc = "unknown process type";
4102
4103         switch (backendType)
4104         {
4105                 case B_AUTOVAC_LAUNCHER:
4106                         backendDesc = "autovacuum launcher";
4107                         break;
4108                 case B_AUTOVAC_WORKER:
4109                         backendDesc = "autovacuum worker";
4110                         break;
4111                 case B_BACKEND:
4112                         backendDesc = "client backend";
4113                         break;
4114                 case B_BG_WORKER:
4115                         backendDesc = "background worker";
4116                         break;
4117                 case B_BG_WRITER:
4118                         backendDesc = "background writer";
4119                         break;
4120                 case B_CHECKPOINTER:
4121                         backendDesc = "checkpointer";
4122                         break;
4123                 case B_STARTUP:
4124                         backendDesc = "startup";
4125                         break;
4126                 case B_WAL_RECEIVER:
4127                         backendDesc = "walreceiver";
4128                         break;
4129                 case B_WAL_SENDER:
4130                         backendDesc = "walsender";
4131                         break;
4132                 case B_WAL_WRITER:
4133                         backendDesc = "walwriter";
4134                         break;
4135         }
4136
4137         return backendDesc;
4138 }
4139
4140 /* ------------------------------------------------------------
4141  * Local support functions follow
4142  * ------------------------------------------------------------
4143  */
4144
4145
4146 /* ----------
4147  * pgstat_setheader() -
4148  *
4149  *              Set common header fields in a statistics message
4150  * ----------
4151  */
4152 static void
4153 pgstat_setheader(PgStat_MsgHdr *hdr, StatMsgType mtype)
4154 {
4155         hdr->m_type = mtype;
4156 }
4157
4158
4159 /* ----------
4160  * pgstat_send() -
4161  *
4162  *              Send out one statistics message to the collector
4163  * ----------
4164  */
4165 static void
4166 pgstat_send(void *msg, int len)
4167 {
4168         int                     rc;
4169
4170         if (pgStatSock == PGINVALID_SOCKET)
4171                 return;
4172
4173         ((PgStat_MsgHdr *) msg)->m_size = len;
4174
4175         /* We'll retry after EINTR, but ignore all other failures */
4176         do
4177         {
4178                 rc = send(pgStatSock, msg, len, 0);
4179         } while (rc < 0 && errno == EINTR);
4180
4181 #ifdef USE_ASSERT_CHECKING
4182         /* In debug builds, log send failures ... */
4183         if (rc < 0)
4184                 elog(LOG, "could not send to statistics collector: %m");
4185 #endif
4186 }
4187
4188 /* ----------
4189  * pgstat_send_archiver() -
4190  *
4191  *      Tell the collector about the WAL file that we successfully
4192  *      archived or failed to archive.
4193  * ----------
4194  */
4195 void
4196 pgstat_send_archiver(const char *xlog, bool failed)
4197 {
4198         PgStat_MsgArchiver msg;
4199
4200         /*
4201          * Prepare and send the message
4202          */
4203         pgstat_setheader(&msg.m_hdr, PGSTAT_MTYPE_ARCHIVER);
4204         msg.m_failed = failed;
4205         StrNCpy(msg.m_xlog, xlog, sizeof(msg.m_xlog));
4206         msg.m_timestamp = GetCurrentTimestamp();
4207         pgstat_send(&msg, sizeof(msg));
4208 }
4209
4210 /* ----------
4211  * pgstat_send_bgwriter() -
4212  *
4213  *              Send bgwriter statistics to the collector
4214  * ----------
4215  */
4216 void
4217 pgstat_send_bgwriter(void)
4218 {
4219         /* We assume this initializes to zeroes */
4220         static const PgStat_MsgBgWriter all_zeroes;
4221
4222         /*
4223          * This function can be called even if nothing at all has happened. In
4224          * this case, avoid sending a completely empty message to the stats
4225          * collector.
4226          */
4227         if (memcmp(&BgWriterStats, &all_zeroes, sizeof(PgStat_MsgBgWriter)) == 0)
4228                 return;
4229
4230         /*
4231          * Prepare and send the message
4232          */
4233         pgstat_setheader(&BgWriterStats.m_hdr, PGSTAT_MTYPE_BGWRITER);
4234         pgstat_send(&BgWriterStats, sizeof(BgWriterStats));
4235
4236         /*
4237          * Clear out the statistics buffer, so it can be re-used.
4238          */
4239         MemSet(&BgWriterStats, 0, sizeof(BgWriterStats));
4240 }
4241
4242
4243 /* ----------
4244  * PgstatCollectorMain() -
4245  *
4246  *      Start up the statistics collector process.  This is the body of the
4247  *      postmaster child process.
4248  *
4249  *      The argc/argv parameters are valid only in EXEC_BACKEND case.
4250  * ----------
4251  */
4252 NON_EXEC_STATIC void
4253 PgstatCollectorMain(int argc, char *argv[])
4254 {
4255         int                     len;
4256         PgStat_Msg      msg;
4257         int                     wr;
4258
4259         /*
4260          * Ignore all signals usually bound to some action in the postmaster,
4261          * except SIGHUP and SIGQUIT.  Note we don't need a SIGUSR1 handler to
4262          * support latch operations, because we only use a local latch.
4263          */
4264         pqsignal(SIGHUP, pgstat_sighup_handler);
4265         pqsignal(SIGINT, SIG_IGN);
4266         pqsignal(SIGTERM, SIG_IGN);
4267         pqsignal(SIGQUIT, pgstat_exit);
4268         pqsignal(SIGALRM, SIG_IGN);
4269         pqsignal(SIGPIPE, SIG_IGN);
4270         pqsignal(SIGUSR1, SIG_IGN);
4271         pqsignal(SIGUSR2, SIG_IGN);
4272         pqsignal(SIGCHLD, SIG_DFL);
4273         pqsignal(SIGTTIN, SIG_DFL);
4274         pqsignal(SIGTTOU, SIG_DFL);
4275         pqsignal(SIGCONT, SIG_DFL);
4276         pqsignal(SIGWINCH, SIG_DFL);
4277         PG_SETMASK(&UnBlockSig);
4278
4279         /*
4280          * Identify myself via ps
4281          */
4282         init_ps_display("stats collector", "", "", "");
4283
4284         /*
4285          * Read in existing stats files or initialize the stats to zero.
4286          */
4287         pgStatRunningInCollector = true;
4288         pgStatDBHash = pgstat_read_statsfiles(InvalidOid, true, true);
4289
4290         /*
4291          * Loop to process messages until we get SIGQUIT or detect ungraceful
4292          * death of our parent postmaster.
4293          *
4294          * For performance reasons, we don't want to do ResetLatch/WaitLatch after
4295          * every message; instead, do that only after a recv() fails to obtain a
4296          * message.  (This effectively means that if backends are sending us stuff
4297          * like mad, we won't notice postmaster death until things slack off a
4298          * bit; which seems fine.)      To do that, we have an inner loop that
4299          * iterates as long as recv() succeeds.  We do recognize got_SIGHUP inside
4300          * the inner loop, which means that such interrupts will get serviced but
4301          * the latch won't get cleared until next time there is a break in the
4302          * action.
4303          */
4304         for (;;)
4305         {
4306                 /* Clear any already-pending wakeups */
4307                 ResetLatch(MyLatch);
4308
4309                 /*
4310                  * Quit if we get SIGQUIT from the postmaster.
4311                  */
4312                 if (need_exit)
4313                         break;
4314
4315                 /*
4316                  * Inner loop iterates as long as we keep getting messages, or until
4317                  * need_exit becomes set.
4318                  */
4319                 while (!need_exit)
4320                 {
4321                         /*
4322                          * Reload configuration if we got SIGHUP from the postmaster.
4323                          */
4324                         if (got_SIGHUP)
4325                         {
4326                                 got_SIGHUP = false;
4327                                 ProcessConfigFile(PGC_SIGHUP);
4328                         }
4329
4330                         /*
4331                          * Write the stats file(s) if a new request has arrived that is
4332                          * not satisfied by existing file(s).
4333                          */
4334                         if (pgstat_write_statsfile_needed())
4335                                 pgstat_write_statsfiles(false, false);
4336
4337                         /*
4338                          * Try to receive and process a message.  This will not block,
4339                          * since the socket is set to non-blocking mode.
4340                          *
4341                          * XXX On Windows, we have to force pgwin32_recv to cooperate,
4342                          * despite the previous use of pg_set_noblock() on the socket.
4343                          * This is extremely broken and should be fixed someday.
4344                          */
4345 #ifdef WIN32
4346                         pgwin32_noblock = 1;
4347 #endif
4348
4349                         len = recv(pgStatSock, (char *) &msg,
4350                                            sizeof(PgStat_Msg), 0);
4351
4352 #ifdef WIN32
4353                         pgwin32_noblock = 0;
4354 #endif
4355
4356                         if (len < 0)
4357                         {
4358                                 if (errno == EAGAIN || errno == EWOULDBLOCK || errno == EINTR)
4359                                         break;          /* out of inner loop */
4360                                 ereport(ERROR,
4361                                                 (errcode_for_socket_access(),
4362                                                  errmsg("could not read statistics message: %m")));
4363                         }
4364
4365                         /*
4366                          * We ignore messages that are smaller than our common header
4367                          */
4368                         if (len < sizeof(PgStat_MsgHdr))
4369                                 continue;
4370
4371                         /*
4372                          * The received length must match the length in the header
4373                          */
4374                         if (msg.msg_hdr.m_size != len)
4375                                 continue;
4376
4377                         /*
4378                          * O.K. - we accept this message.  Process it.
4379                          */
4380                         switch (msg.msg_hdr.m_type)
4381                         {
4382                                 case PGSTAT_MTYPE_DUMMY:
4383                                         break;
4384
4385                                 case PGSTAT_MTYPE_INQUIRY:
4386                                         pgstat_recv_inquiry((PgStat_MsgInquiry *) &msg, len);
4387                                         break;
4388
4389                                 case PGSTAT_MTYPE_TABSTAT:
4390                                         pgstat_recv_tabstat((PgStat_MsgTabstat *) &msg, len);
4391                                         break;
4392
4393                                 case PGSTAT_MTYPE_TABPURGE:
4394                                         pgstat_recv_tabpurge((PgStat_MsgTabpurge *) &msg, len);
4395                                         break;
4396
4397                                 case PGSTAT_MTYPE_DROPDB:
4398                                         pgstat_recv_dropdb((PgStat_MsgDropdb *) &msg, len);
4399                                         break;
4400
4401                                 case PGSTAT_MTYPE_RESETCOUNTER:
4402                                         pgstat_recv_resetcounter((PgStat_MsgResetcounter *) &msg,
4403                                                                                          len);
4404                                         break;
4405
4406                                 case PGSTAT_MTYPE_RESETSHAREDCOUNTER:
4407                                         pgstat_recv_resetsharedcounter(
4408                                                                                                    (PgStat_MsgResetsharedcounter *) &msg,
4409                                                                                                    len);
4410                                         break;
4411
4412                                 case PGSTAT_MTYPE_RESETSINGLECOUNTER:
4413                                         pgstat_recv_resetsinglecounter(
4414                                                                                                    (PgStat_MsgResetsinglecounter *) &msg,
4415                                                                                                    len);
4416                                         break;
4417
4418                                 case PGSTAT_MTYPE_AUTOVAC_START:
4419                                         pgstat_recv_autovac((PgStat_MsgAutovacStart *) &msg, len);
4420                                         break;
4421
4422                                 case PGSTAT_MTYPE_VACUUM:
4423                                         pgstat_recv_vacuum((PgStat_MsgVacuum *) &msg, len);
4424                                         break;
4425
4426                                 case PGSTAT_MTYPE_ANALYZE:
4427                                         pgstat_recv_analyze((PgStat_MsgAnalyze *) &msg, len);
4428                                         break;
4429
4430                                 case PGSTAT_MTYPE_ARCHIVER:
4431                                         pgstat_recv_archiver((PgStat_MsgArchiver *) &msg, len);
4432                                         break;
4433
4434                                 case PGSTAT_MTYPE_BGWRITER:
4435                                         pgstat_recv_bgwriter((PgStat_MsgBgWriter *) &msg, len);
4436                                         break;
4437
4438                                 case PGSTAT_MTYPE_FUNCSTAT:
4439                                         pgstat_recv_funcstat((PgStat_MsgFuncstat *) &msg, len);
4440                                         break;
4441
4442                                 case PGSTAT_MTYPE_FUNCPURGE:
4443                                         pgstat_recv_funcpurge((PgStat_MsgFuncpurge *) &msg, len);
4444                                         break;
4445
4446                                 case PGSTAT_MTYPE_RECOVERYCONFLICT:
4447                                         pgstat_recv_recoveryconflict((PgStat_MsgRecoveryConflict *) &msg, len);
4448                                         break;
4449
4450                                 case PGSTAT_MTYPE_DEADLOCK:
4451                                         pgstat_recv_deadlock((PgStat_MsgDeadlock *) &msg, len);
4452                                         break;
4453
4454                                 case PGSTAT_MTYPE_TEMPFILE:
4455                                         pgstat_recv_tempfile((PgStat_MsgTempFile *) &msg, len);
4456                                         break;
4457
4458                                 default:
4459                                         break;
4460                         }
4461                 }                                               /* end of inner message-processing loop */
4462
4463                 /* Sleep until there's something to do */
4464 #ifndef WIN32
4465                 wr = WaitLatchOrSocket(MyLatch,
4466                                                            WL_LATCH_SET | WL_POSTMASTER_DEATH | WL_SOCKET_READABLE,
4467                                                            pgStatSock, -1L,
4468                                                            WAIT_EVENT_PGSTAT_MAIN);
4469 #else
4470
4471                 /*
4472                  * Windows, at least in its Windows Server 2003 R2 incarnation,
4473                  * sometimes loses FD_READ events.  Waking up and retrying the recv()
4474                  * fixes that, so don't sleep indefinitely.  This is a crock of the
4475                  * first water, but until somebody wants to debug exactly what's
4476                  * happening there, this is the best we can do.  The two-second
4477                  * timeout matches our pre-9.2 behavior, and needs to be short enough
4478                  * to not provoke "using stale statistics" complaints from
4479                  * backend_read_statsfile.
4480                  */
4481                 wr = WaitLatchOrSocket(MyLatch,
4482                                                            WL_LATCH_SET | WL_POSTMASTER_DEATH | WL_SOCKET_READABLE | WL_TIMEOUT,
4483                                                            pgStatSock,
4484                                                            2 * 1000L /* msec */ ,
4485                                                            WAIT_EVENT_PGSTAT_MAIN);
4486 #endif
4487
4488                 /*
4489                  * Emergency bailout if postmaster has died.  This is to avoid the
4490                  * necessity for manual cleanup of all postmaster children.
4491                  */
4492                 if (wr & WL_POSTMASTER_DEATH)
4493                         break;
4494         }                                                       /* end of outer loop */
4495
4496         /*
4497          * Save the final stats to reuse at next startup.
4498          */
4499         pgstat_write_statsfiles(true, true);
4500
4501         exit(0);
4502 }
4503
4504
4505 /* SIGQUIT signal handler for collector process */
4506 static void
4507 pgstat_exit(SIGNAL_ARGS)
4508 {
4509         int                     save_errno = errno;
4510
4511         need_exit = true;
4512         SetLatch(MyLatch);
4513
4514         errno = save_errno;
4515 }
4516
4517 /* SIGHUP handler for collector process */
4518 static void
4519 pgstat_sighup_handler(SIGNAL_ARGS)
4520 {
4521         int                     save_errno = errno;
4522
4523         got_SIGHUP = true;
4524         SetLatch(MyLatch);
4525
4526         errno = save_errno;
4527 }
4528
4529 /*
4530  * Subroutine to clear stats in a database entry
4531  *
4532  * Tables and functions hashes are initialized to empty.
4533  */
4534 static void
4535 reset_dbentry_counters(PgStat_StatDBEntry *dbentry)
4536 {
4537         HASHCTL         hash_ctl;
4538
4539         dbentry->n_xact_commit = 0;
4540         dbentry->n_xact_rollback = 0;
4541         dbentry->n_blocks_fetched = 0;
4542         dbentry->n_blocks_hit = 0;
4543         dbentry->n_tuples_returned = 0;
4544         dbentry->n_tuples_fetched = 0;
4545         dbentry->n_tuples_inserted = 0;
4546         dbentry->n_tuples_updated = 0;
4547         dbentry->n_tuples_deleted = 0;
4548         dbentry->last_autovac_time = 0;
4549         dbentry->n_conflict_tablespace = 0;
4550         dbentry->n_conflict_lock = 0;
4551         dbentry->n_conflict_snapshot = 0;
4552         dbentry->n_conflict_bufferpin = 0;
4553         dbentry->n_conflict_startup_deadlock = 0;
4554         dbentry->n_temp_files = 0;
4555         dbentry->n_temp_bytes = 0;
4556         dbentry->n_deadlocks = 0;
4557         dbentry->n_block_read_time = 0;
4558         dbentry->n_block_write_time = 0;
4559
4560         dbentry->stat_reset_timestamp = GetCurrentTimestamp();
4561         dbentry->stats_timestamp = 0;
4562
4563         memset(&hash_ctl, 0, sizeof(hash_ctl));
4564         hash_ctl.keysize = sizeof(Oid);
4565         hash_ctl.entrysize = sizeof(PgStat_StatTabEntry);
4566         dbentry->tables = hash_create("Per-database table",
4567                                                                   PGSTAT_TAB_HASH_SIZE,
4568                                                                   &hash_ctl,
4569                                                                   HASH_ELEM | HASH_BLOBS);
4570
4571         hash_ctl.keysize = sizeof(Oid);
4572         hash_ctl.entrysize = sizeof(PgStat_StatFuncEntry);
4573         dbentry->functions = hash_create("Per-database function",
4574                                                                          PGSTAT_FUNCTION_HASH_SIZE,
4575                                                                          &hash_ctl,
4576                                                                          HASH_ELEM | HASH_BLOBS);
4577 }
4578
4579 /*
4580  * Lookup the hash table entry for the specified database. If no hash
4581  * table entry exists, initialize it, if the create parameter is true.
4582  * Else, return NULL.
4583  */
4584 static PgStat_StatDBEntry *
4585 pgstat_get_db_entry(Oid databaseid, bool create)
4586 {
4587         PgStat_StatDBEntry *result;
4588         bool            found;
4589         HASHACTION      action = (create ? HASH_ENTER : HASH_FIND);
4590
4591         /* Lookup or create the hash table entry for this database */
4592         result = (PgStat_StatDBEntry *) hash_search(pgStatDBHash,
4593                                                                                                 &databaseid,
4594                                                                                                 action, &found);
4595
4596         if (!create && !found)
4597                 return NULL;
4598
4599         /*
4600          * If not found, initialize the new one.  This creates empty hash tables
4601          * for tables and functions, too.
4602          */
4603         if (!found)
4604                 reset_dbentry_counters(result);
4605
4606         return result;
4607 }
4608
4609
4610 /*
4611  * Lookup the hash table entry for the specified table. If no hash
4612  * table entry exists, initialize it, if the create parameter is true.
4613  * Else, return NULL.
4614  */
4615 static PgStat_StatTabEntry *
4616 pgstat_get_tab_entry(PgStat_StatDBEntry *dbentry, Oid tableoid, bool create)
4617 {
4618         PgStat_StatTabEntry *result;
4619         bool            found;
4620         HASHACTION      action = (create ? HASH_ENTER : HASH_FIND);
4621
4622         /* Lookup or create the hash table entry for this table */
4623         result = (PgStat_StatTabEntry *) hash_search(dbentry->tables,
4624                                                                                                  &tableoid,
4625                                                                                                  action, &found);
4626
4627         if (!create && !found)
4628                 return NULL;
4629
4630         /* If not found, initialize the new one. */
4631         if (!found)
4632         {
4633                 result->numscans = 0;
4634                 result->tuples_returned = 0;
4635                 result->tuples_fetched = 0;
4636                 result->tuples_inserted = 0;
4637                 result->tuples_updated = 0;
4638                 result->tuples_deleted = 0;
4639                 result->tuples_hot_updated = 0;
4640                 result->n_live_tuples = 0;
4641                 result->n_dead_tuples = 0;
4642                 result->changes_since_analyze = 0;
4643                 result->blocks_fetched = 0;
4644                 result->blocks_hit = 0;
4645                 result->vacuum_timestamp = 0;
4646                 result->vacuum_count = 0;
4647                 result->autovac_vacuum_timestamp = 0;
4648                 result->autovac_vacuum_count = 0;
4649                 result->analyze_timestamp = 0;
4650                 result->analyze_count = 0;
4651                 result->autovac_analyze_timestamp = 0;
4652                 result->autovac_analyze_count = 0;
4653         }
4654
4655         return result;
4656 }
4657
4658
4659 /* ----------
4660  * pgstat_write_statsfiles() -
4661  *              Write the global statistics file, as well as requested DB files.
4662  *
4663  *      'permanent' specifies writing to the permanent files not temporary ones.
4664  *      When true (happens only when the collector is shutting down), also remove
4665  *      the temporary files so that backends starting up under a new postmaster
4666  *      can't read old data before the new collector is ready.
4667  *
4668  *      When 'allDbs' is false, only the requested databases (listed in
4669  *      pending_write_requests) will be written; otherwise, all databases
4670  *      will be written.
4671  * ----------
4672  */
4673 static void
4674 pgstat_write_statsfiles(bool permanent, bool allDbs)
4675 {
4676         HASH_SEQ_STATUS hstat;
4677         PgStat_StatDBEntry *dbentry;
4678         FILE       *fpout;
4679         int32           format_id;
4680         const char *tmpfile = permanent ? PGSTAT_STAT_PERMANENT_TMPFILE : pgstat_stat_tmpname;
4681         const char *statfile = permanent ? PGSTAT_STAT_PERMANENT_FILENAME : pgstat_stat_filename;
4682         int                     rc;
4683
4684         elog(DEBUG2, "writing stats file \"%s\"", statfile);
4685
4686         /*
4687          * Open the statistics temp file to write out the current values.
4688          */
4689         fpout = AllocateFile(tmpfile, PG_BINARY_W);
4690         if (fpout == NULL)
4691         {
4692                 ereport(LOG,
4693                                 (errcode_for_file_access(),
4694                                  errmsg("could not open temporary statistics file \"%s\": %m",
4695                                                 tmpfile)));
4696                 return;
4697         }
4698
4699         /*
4700          * Set the timestamp of the stats file.
4701          */
4702         globalStats.stats_timestamp = GetCurrentTimestamp();
4703
4704         /*
4705          * Write the file header --- currently just a format ID.
4706          */
4707         format_id = PGSTAT_FILE_FORMAT_ID;
4708         rc = fwrite(&format_id, sizeof(format_id), 1, fpout);
4709         (void) rc;                                      /* we'll check for error with ferror */
4710
4711         /*
4712          * Write global stats struct
4713          */
4714         rc = fwrite(&globalStats, sizeof(globalStats), 1, fpout);
4715         (void) rc;                                      /* we'll check for error with ferror */
4716
4717         /*
4718          * Write archiver stats struct
4719          */
4720         rc = fwrite(&archiverStats, sizeof(archiverStats), 1, fpout);
4721         (void) rc;                                      /* we'll check for error with ferror */
4722
4723         /*
4724          * Walk through the database table.
4725          */
4726         hash_seq_init(&hstat, pgStatDBHash);
4727         while ((dbentry = (PgStat_StatDBEntry *) hash_seq_search(&hstat)) != NULL)
4728         {
4729                 /*
4730                  * Write out the table and function stats for this DB into the
4731                  * appropriate per-DB stat file, if required.
4732                  */
4733                 if (allDbs || pgstat_db_requested(dbentry->databaseid))
4734                 {
4735                         /* Make DB's timestamp consistent with the global stats */
4736                         dbentry->stats_timestamp = globalStats.stats_timestamp;
4737
4738                         pgstat_write_db_statsfile(dbentry, permanent);
4739                 }
4740
4741                 /*
4742                  * Write out the DB entry. We don't write the tables or functions
4743                  * pointers, since they're of no use to any other process.
4744                  */
4745                 fputc('D', fpout);
4746                 rc = fwrite(dbentry, offsetof(PgStat_StatDBEntry, tables), 1, fpout);
4747                 (void) rc;                              /* we'll check for error with ferror */
4748         }
4749
4750         /*
4751          * No more output to be done. Close the temp file and replace the old
4752          * pgstat.stat with it.  The ferror() check replaces testing for error
4753          * after each individual fputc or fwrite above.
4754          */
4755         fputc('E', fpout);
4756
4757         if (ferror(fpout))
4758         {
4759                 ereport(LOG,
4760                                 (errcode_for_file_access(),
4761                                  errmsg("could not write temporary statistics file \"%s\": %m",
4762                                                 tmpfile)));
4763                 FreeFile(fpout);
4764                 unlink(tmpfile);
4765         }
4766         else if (FreeFile(fpout) < 0)
4767         {
4768                 ereport(LOG,
4769                                 (errcode_for_file_access(),
4770                                  errmsg("could not close temporary statistics file \"%s\": %m",
4771                                                 tmpfile)));
4772                 unlink(tmpfile);
4773         }
4774         else if (rename(tmpfile, statfile) < 0)
4775         {
4776                 ereport(LOG,
4777                                 (errcode_for_file_access(),
4778                                  errmsg("could not rename temporary statistics file \"%s\" to \"%s\": %m",
4779                                                 tmpfile, statfile)));
4780                 unlink(tmpfile);
4781         }
4782
4783         if (permanent)
4784                 unlink(pgstat_stat_filename);
4785
4786         /*
4787          * Now throw away the list of requests.  Note that requests sent after we
4788          * started the write are still waiting on the network socket.
4789          */
4790         list_free(pending_write_requests);
4791         pending_write_requests = NIL;
4792 }
4793
4794 /*
4795  * return the filename for a DB stat file; filename is the output buffer,
4796  * of length len.
4797  */
4798 static void
4799 get_dbstat_filename(bool permanent, bool tempname, Oid databaseid,
4800                                         char *filename, int len)
4801 {
4802         int                     printed;
4803
4804         /* NB -- pgstat_reset_remove_files knows about the pattern this uses */
4805         printed = snprintf(filename, len, "%s/db_%u.%s",
4806                                            permanent ? PGSTAT_STAT_PERMANENT_DIRECTORY :
4807                                            pgstat_stat_directory,
4808                                            databaseid,
4809                                            tempname ? "tmp" : "stat");
4810         if (printed > len)
4811                 elog(ERROR, "overlength pgstat path");
4812 }
4813
4814 /* ----------
4815  * pgstat_write_db_statsfile() -
4816  *              Write the stat file for a single database.
4817  *
4818  *      If writing to the permanent file (happens when the collector is
4819  *      shutting down only), remove the temporary file so that backends
4820  *      starting up under a new postmaster can't read the old data before
4821  *      the new collector is ready.
4822  * ----------
4823  */
4824 static void
4825 pgstat_write_db_statsfile(PgStat_StatDBEntry *dbentry, bool permanent)
4826 {
4827         HASH_SEQ_STATUS tstat;
4828         HASH_SEQ_STATUS fstat;
4829         PgStat_StatTabEntry *tabentry;
4830         PgStat_StatFuncEntry *funcentry;
4831         FILE       *fpout;
4832         int32           format_id;
4833         Oid                     dbid = dbentry->databaseid;
4834         int                     rc;
4835         char            tmpfile[MAXPGPATH];
4836         char            statfile[MAXPGPATH];
4837
4838         get_dbstat_filename(permanent, true, dbid, tmpfile, MAXPGPATH);
4839         get_dbstat_filename(permanent, false, dbid, statfile, MAXPGPATH);
4840
4841         elog(DEBUG2, "writing stats file \"%s\"", statfile);
4842
4843         /*
4844          * Open the statistics temp file to write out the current values.
4845          */
4846         fpout = AllocateFile(tmpfile, PG_BINARY_W);
4847         if (fpout == NULL)
4848         {
4849                 ereport(LOG,
4850                                 (errcode_for_file_access(),
4851                                  errmsg("could not open temporary statistics file \"%s\": %m",
4852                                                 tmpfile)));
4853                 return;
4854         }
4855
4856         /*
4857          * Write the file header --- currently just a format ID.
4858          */
4859         format_id = PGSTAT_FILE_FORMAT_ID;
4860         rc = fwrite(&format_id, sizeof(format_id), 1, fpout);
4861         (void) rc;                                      /* we'll check for error with ferror */
4862
4863         /*
4864          * Walk through the database's access stats per table.
4865          */
4866         hash_seq_init(&tstat, dbentry->tables);
4867         while ((tabentry = (PgStat_StatTabEntry *) hash_seq_search(&tstat)) != NULL)
4868         {
4869                 fputc('T', fpout);
4870                 rc = fwrite(tabentry, sizeof(PgStat_StatTabEntry), 1, fpout);
4871                 (void) rc;                              /* we'll check for error with ferror */
4872         }
4873
4874         /*
4875          * Walk through the database's function stats table.
4876          */
4877         hash_seq_init(&fstat, dbentry->functions);
4878         while ((funcentry = (PgStat_StatFuncEntry *) hash_seq_search(&fstat)) != NULL)
4879         {
4880                 fputc('F', fpout);
4881                 rc = fwrite(funcentry, sizeof(PgStat_StatFuncEntry), 1, fpout);
4882                 (void) rc;                              /* we'll check for error with ferror */
4883         }
4884
4885         /*
4886          * No more output to be done. Close the temp file and replace the old
4887          * pgstat.stat with it.  The ferror() check replaces testing for error
4888          * after each individual fputc or fwrite above.
4889          */
4890         fputc('E', fpout);
4891
4892         if (ferror(fpout))
4893         {
4894                 ereport(LOG,
4895                                 (errcode_for_file_access(),
4896                                  errmsg("could not write temporary statistics file \"%s\": %m",
4897                                                 tmpfile)));
4898                 FreeFile(fpout);
4899                 unlink(tmpfile);
4900         }
4901         else if (FreeFile(fpout) < 0)
4902         {
4903                 ereport(LOG,
4904                                 (errcode_for_file_access(),
4905                                  errmsg("could not close temporary statistics file \"%s\": %m",
4906                                                 tmpfile)));
4907                 unlink(tmpfile);
4908         }
4909         else if (rename(tmpfile, statfile) < 0)
4910         {
4911                 ereport(LOG,
4912                                 (errcode_for_file_access(),
4913                                  errmsg("could not rename temporary statistics file \"%s\" to \"%s\": %m",
4914                                                 tmpfile, statfile)));
4915                 unlink(tmpfile);
4916         }
4917
4918         if (permanent)
4919         {
4920                 get_dbstat_filename(false, false, dbid, statfile, MAXPGPATH);
4921
4922                 elog(DEBUG2, "removing temporary stats file \"%s\"", statfile);
4923                 unlink(statfile);
4924         }
4925 }
4926
4927 /* ----------
4928  * pgstat_read_statsfiles() -
4929  *
4930  *      Reads in some existing statistics collector files and returns the
4931  *      databases hash table that is the top level of the data.
4932  *
4933  *      If 'onlydb' is not InvalidOid, it means we only want data for that DB
4934  *      plus the shared catalogs ("DB 0").  We'll still populate the DB hash
4935  *      table for all databases, but we don't bother even creating table/function
4936  *      hash tables for other databases.
4937  *
4938  *      'permanent' specifies reading from the permanent files not temporary ones.
4939  *      When true (happens only when the collector is starting up), remove the
4940  *      files after reading; the in-memory status is now authoritative, and the
4941  *      files would be out of date in case somebody else reads them.
4942  *
4943  *      If a 'deep' read is requested, table/function stats are read, otherwise
4944  *      the table/function hash tables remain empty.
4945  * ----------
4946  */
4947 static HTAB *
4948 pgstat_read_statsfiles(Oid onlydb, bool permanent, bool deep)
4949 {
4950         PgStat_StatDBEntry *dbentry;
4951         PgStat_StatDBEntry dbbuf;
4952         HASHCTL         hash_ctl;
4953         HTAB       *dbhash;
4954         FILE       *fpin;
4955         int32           format_id;
4956         bool            found;
4957         const char *statfile = permanent ? PGSTAT_STAT_PERMANENT_FILENAME : pgstat_stat_filename;
4958
4959         /*
4960          * The tables will live in pgStatLocalContext.
4961          */
4962         pgstat_setup_memcxt();
4963
4964         /*
4965          * Create the DB hashtable
4966          */
4967         memset(&hash_ctl, 0, sizeof(hash_ctl));
4968         hash_ctl.keysize = sizeof(Oid);
4969         hash_ctl.entrysize = sizeof(PgStat_StatDBEntry);
4970         hash_ctl.hcxt = pgStatLocalContext;
4971         dbhash = hash_create("Databases hash", PGSTAT_DB_HASH_SIZE, &hash_ctl,
4972                                                  HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
4973
4974         /*
4975          * Clear out global and archiver statistics so they start from zero in
4976          * case we can't load an existing statsfile.
4977          */
4978         memset(&globalStats, 0, sizeof(globalStats));
4979         memset(&archiverStats, 0, sizeof(archiverStats));
4980
4981         /*
4982          * Set the current timestamp (will be kept only in case we can't load an
4983          * existing statsfile).
4984          */
4985         globalStats.stat_reset_timestamp = GetCurrentTimestamp();
4986         archiverStats.stat_reset_timestamp = globalStats.stat_reset_timestamp;
4987
4988         /*
4989          * Try to open the stats file. If it doesn't exist, the backends simply
4990          * return zero for anything and the collector simply starts from scratch
4991          * with empty counters.
4992          *
4993          * ENOENT is a possibility if the stats collector is not running or has
4994          * not yet written the stats file the first time.  Any other failure
4995          * condition is suspicious.
4996          */
4997         if ((fpin = AllocateFile(statfile, PG_BINARY_R)) == NULL)
4998         {
4999                 if (errno != ENOENT)
5000                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5001                                         (errcode_for_file_access(),
5002                                          errmsg("could not open statistics file \"%s\": %m",
5003                                                         statfile)));
5004                 return dbhash;
5005         }
5006
5007         /*
5008          * Verify it's of the expected format.
5009          */
5010         if (fread(&format_id, 1, sizeof(format_id), fpin) != sizeof(format_id) ||
5011                 format_id != PGSTAT_FILE_FORMAT_ID)
5012         {
5013                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5014                                 (errmsg("corrupted statistics file \"%s\"", statfile)));
5015                 goto done;
5016         }
5017
5018         /*
5019          * Read global stats struct
5020          */
5021         if (fread(&globalStats, 1, sizeof(globalStats), fpin) != sizeof(globalStats))
5022         {
5023                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5024                                 (errmsg("corrupted statistics file \"%s\"", statfile)));
5025                 memset(&globalStats, 0, sizeof(globalStats));
5026                 goto done;
5027         }
5028
5029         /*
5030          * In the collector, disregard the timestamp we read from the permanent
5031          * stats file; we should be willing to write a temp stats file immediately
5032          * upon the first request from any backend.  This only matters if the old
5033          * file's timestamp is less than PGSTAT_STAT_INTERVAL ago, but that's not
5034          * an unusual scenario.
5035          */
5036         if (pgStatRunningInCollector)
5037                 globalStats.stats_timestamp = 0;
5038
5039         /*
5040          * Read archiver stats struct
5041          */
5042         if (fread(&archiverStats, 1, sizeof(archiverStats), fpin) != sizeof(archiverStats))
5043         {
5044                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5045                                 (errmsg("corrupted statistics file \"%s\"", statfile)));
5046                 memset(&archiverStats, 0, sizeof(archiverStats));
5047                 goto done;
5048         }
5049
5050         /*
5051          * We found an existing collector stats file. Read it and put all the
5052          * hashtable entries into place.
5053          */
5054         for (;;)
5055         {
5056                 switch (fgetc(fpin))
5057                 {
5058                                 /*
5059                                  * 'D'  A PgStat_StatDBEntry struct describing a database
5060                                  * follows.
5061                                  */
5062                         case 'D':
5063                                 if (fread(&dbbuf, 1, offsetof(PgStat_StatDBEntry, tables),
5064                                                   fpin) != offsetof(PgStat_StatDBEntry, tables))
5065                                 {
5066                                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5067                                                         (errmsg("corrupted statistics file \"%s\"",
5068                                                                         statfile)));
5069                                         goto done;
5070                                 }
5071
5072                                 /*
5073                                  * Add to the DB hash
5074                                  */
5075                                 dbentry = (PgStat_StatDBEntry *) hash_search(dbhash,
5076                                                                                                                          (void *) &dbbuf.databaseid,
5077                                                                                                                          HASH_ENTER,
5078                                                                                                                          &found);
5079                                 if (found)
5080                                 {
5081                                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5082                                                         (errmsg("corrupted statistics file \"%s\"",
5083                                                                         statfile)));
5084                                         goto done;
5085                                 }
5086
5087                                 memcpy(dbentry, &dbbuf, sizeof(PgStat_StatDBEntry));
5088                                 dbentry->tables = NULL;
5089                                 dbentry->functions = NULL;
5090
5091                                 /*
5092                                  * In the collector, disregard the timestamp we read from the
5093                                  * permanent stats file; we should be willing to write a temp
5094                                  * stats file immediately upon the first request from any
5095                                  * backend.
5096                                  */
5097                                 if (pgStatRunningInCollector)
5098                                         dbentry->stats_timestamp = 0;
5099
5100                                 /*
5101                                  * Don't create tables/functions hashtables for uninteresting
5102                                  * databases.
5103                                  */
5104                                 if (onlydb != InvalidOid)
5105                                 {
5106                                         if (dbbuf.databaseid != onlydb &&
5107                                                 dbbuf.databaseid != InvalidOid)
5108                                                 break;
5109                                 }
5110
5111                                 memset(&hash_ctl, 0, sizeof(hash_ctl));
5112                                 hash_ctl.keysize = sizeof(Oid);
5113                                 hash_ctl.entrysize = sizeof(PgStat_StatTabEntry);
5114                                 hash_ctl.hcxt = pgStatLocalContext;
5115                                 dbentry->tables = hash_create("Per-database table",
5116                                                                                           PGSTAT_TAB_HASH_SIZE,
5117                                                                                           &hash_ctl,
5118                                                                                           HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
5119
5120                                 hash_ctl.keysize = sizeof(Oid);
5121                                 hash_ctl.entrysize = sizeof(PgStat_StatFuncEntry);
5122                                 hash_ctl.hcxt = pgStatLocalContext;
5123                                 dbentry->functions = hash_create("Per-database function",
5124                                                                                                  PGSTAT_FUNCTION_HASH_SIZE,
5125                                                                                                  &hash_ctl,
5126                                                                                                  HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
5127
5128                                 /*
5129                                  * If requested, read the data from the database-specific
5130                                  * file.  Otherwise we just leave the hashtables empty.
5131                                  */
5132                                 if (deep)
5133                                         pgstat_read_db_statsfile(dbentry->databaseid,
5134                                                                                          dbentry->tables,
5135                                                                                          dbentry->functions,
5136                                                                                          permanent);
5137
5138                                 break;
5139
5140                         case 'E':
5141                                 goto done;
5142
5143                         default:
5144                                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5145                                                 (errmsg("corrupted statistics file \"%s\"",
5146                                                                 statfile)));
5147                                 goto done;
5148                 }
5149         }
5150
5151 done:
5152         FreeFile(fpin);
5153
5154         /* If requested to read the permanent file, also get rid of it. */
5155         if (permanent)
5156         {
5157                 elog(DEBUG2, "removing permanent stats file \"%s\"", statfile);
5158                 unlink(statfile);
5159         }
5160
5161         return dbhash;
5162 }
5163
5164
5165 /* ----------
5166  * pgstat_read_db_statsfile() -
5167  *
5168  *      Reads in the existing statistics collector file for the given database,
5169  *      filling the passed-in tables and functions hash tables.
5170  *
5171  *      As in pgstat_read_statsfiles, if the permanent file is requested, it is
5172  *      removed after reading.
5173  *
5174  *      Note: this code has the ability to skip storing per-table or per-function
5175  *      data, if NULL is passed for the corresponding hashtable.  That's not used
5176  *      at the moment though.
5177  * ----------
5178  */
5179 static void
5180 pgstat_read_db_statsfile(Oid databaseid, HTAB *tabhash, HTAB *funchash,
5181                                                  bool permanent)
5182 {
5183         PgStat_StatTabEntry *tabentry;
5184         PgStat_StatTabEntry tabbuf;
5185         PgStat_StatFuncEntry funcbuf;
5186         PgStat_StatFuncEntry *funcentry;
5187         FILE       *fpin;
5188         int32           format_id;
5189         bool            found;
5190         char            statfile[MAXPGPATH];
5191
5192         get_dbstat_filename(permanent, false, databaseid, statfile, MAXPGPATH);
5193
5194         /*
5195          * Try to open the stats file. If it doesn't exist, the backends simply
5196          * return zero for anything and the collector simply starts from scratch
5197          * with empty counters.
5198          *
5199          * ENOENT is a possibility if the stats collector is not running or has
5200          * not yet written the stats file the first time.  Any other failure
5201          * condition is suspicious.
5202          */
5203         if ((fpin = AllocateFile(statfile, PG_BINARY_R)) == NULL)
5204         {
5205                 if (errno != ENOENT)
5206                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5207                                         (errcode_for_file_access(),
5208                                          errmsg("could not open statistics file \"%s\": %m",
5209                                                         statfile)));
5210                 return;
5211         }
5212
5213         /*
5214          * Verify it's of the expected format.
5215          */
5216         if (fread(&format_id, 1, sizeof(format_id), fpin) != sizeof(format_id) ||
5217                 format_id != PGSTAT_FILE_FORMAT_ID)
5218         {
5219                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5220                                 (errmsg("corrupted statistics file \"%s\"", statfile)));
5221                 goto done;
5222         }
5223
5224         /*
5225          * We found an existing collector stats file. Read it and put all the
5226          * hashtable entries into place.
5227          */
5228         for (;;)
5229         {
5230                 switch (fgetc(fpin))
5231                 {
5232                                 /*
5233                                  * 'T'  A PgStat_StatTabEntry follows.
5234                                  */
5235                         case 'T':
5236                                 if (fread(&tabbuf, 1, sizeof(PgStat_StatTabEntry),
5237                                                   fpin) != sizeof(PgStat_StatTabEntry))
5238                                 {
5239                                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5240                                                         (errmsg("corrupted statistics file \"%s\"",
5241                                                                         statfile)));
5242                                         goto done;
5243                                 }
5244
5245                                 /*
5246                                  * Skip if table data not wanted.
5247                                  */
5248                                 if (tabhash == NULL)
5249                                         break;
5250
5251                                 tabentry = (PgStat_StatTabEntry *) hash_search(tabhash,
5252                                                                                                                            (void *) &tabbuf.tableid,
5253                                                                                                                            HASH_ENTER, &found);
5254
5255                                 if (found)
5256                                 {
5257                                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5258                                                         (errmsg("corrupted statistics file \"%s\"",
5259                                                                         statfile)));
5260                                         goto done;
5261                                 }
5262
5263                                 memcpy(tabentry, &tabbuf, sizeof(tabbuf));
5264                                 break;
5265
5266                                 /*
5267                                  * 'F'  A PgStat_StatFuncEntry follows.
5268                                  */
5269                         case 'F':
5270                                 if (fread(&funcbuf, 1, sizeof(PgStat_StatFuncEntry),
5271                                                   fpin) != sizeof(PgStat_StatFuncEntry))
5272                                 {
5273                                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5274                                                         (errmsg("corrupted statistics file \"%s\"",
5275                                                                         statfile)));
5276                                         goto done;
5277                                 }
5278
5279                                 /*
5280                                  * Skip if function data not wanted.
5281                                  */
5282                                 if (funchash == NULL)
5283                                         break;
5284
5285                                 funcentry = (PgStat_StatFuncEntry *) hash_search(funchash,
5286                                                                                                                                  (void *) &funcbuf.functionid,
5287                                                                                                                                  HASH_ENTER, &found);
5288
5289                                 if (found)
5290                                 {
5291                                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5292                                                         (errmsg("corrupted statistics file \"%s\"",
5293                                                                         statfile)));
5294                                         goto done;
5295                                 }
5296
5297                                 memcpy(funcentry, &funcbuf, sizeof(funcbuf));
5298                                 break;
5299
5300                                 /*
5301                                  * 'E'  The EOF marker of a complete stats file.
5302                                  */
5303                         case 'E':
5304                                 goto done;
5305
5306                         default:
5307                                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5308                                                 (errmsg("corrupted statistics file \"%s\"",
5309                                                                 statfile)));
5310                                 goto done;
5311                 }
5312         }
5313
5314 done:
5315         FreeFile(fpin);
5316
5317         if (permanent)
5318         {
5319                 elog(DEBUG2, "removing permanent stats file \"%s\"", statfile);
5320                 unlink(statfile);
5321         }
5322 }
5323
5324 /* ----------
5325  * pgstat_read_db_statsfile_timestamp() -
5326  *
5327  *      Attempt to determine the timestamp of the last db statfile write.
5328  *      Returns true if successful; the timestamp is stored in *ts.
5329  *
5330  *      This needs to be careful about handling databases for which no stats file
5331  *      exists, such as databases without a stat entry or those not yet written:
5332  *
5333  *      - if there's a database entry in the global file, return the corresponding
5334  *      stats_timestamp value.
5335  *
5336  *      - if there's no db stat entry (e.g. for a new or inactive database),
5337  *      there's no stats_timestamp value, but also nothing to write so we return
5338  *      the timestamp of the global statfile.
5339  * ----------
5340  */
5341 static bool
5342 pgstat_read_db_statsfile_timestamp(Oid databaseid, bool permanent,
5343                                                                    TimestampTz *ts)
5344 {
5345         PgStat_StatDBEntry dbentry;
5346         PgStat_GlobalStats myGlobalStats;
5347         PgStat_ArchiverStats myArchiverStats;
5348         FILE       *fpin;
5349         int32           format_id;
5350         const char *statfile = permanent ? PGSTAT_STAT_PERMANENT_FILENAME : pgstat_stat_filename;
5351
5352         /*
5353          * Try to open the stats file.  As above, anything but ENOENT is worthy of
5354          * complaining about.
5355          */
5356         if ((fpin = AllocateFile(statfile, PG_BINARY_R)) == NULL)
5357         {
5358                 if (errno != ENOENT)
5359                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5360                                         (errcode_for_file_access(),
5361                                          errmsg("could not open statistics file \"%s\": %m",
5362                                                         statfile)));
5363                 return false;
5364         }
5365
5366         /*
5367          * Verify it's of the expected format.
5368          */
5369         if (fread(&format_id, 1, sizeof(format_id), fpin) != sizeof(format_id) ||
5370                 format_id != PGSTAT_FILE_FORMAT_ID)
5371         {
5372                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5373                                 (errmsg("corrupted statistics file \"%s\"", statfile)));
5374                 FreeFile(fpin);
5375                 return false;
5376         }
5377
5378         /*
5379          * Read global stats struct
5380          */
5381         if (fread(&myGlobalStats, 1, sizeof(myGlobalStats),
5382                           fpin) != sizeof(myGlobalStats))
5383         {
5384                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5385                                 (errmsg("corrupted statistics file \"%s\"", statfile)));
5386                 FreeFile(fpin);
5387                 return false;
5388         }
5389
5390         /*
5391          * Read archiver stats struct
5392          */
5393         if (fread(&myArchiverStats, 1, sizeof(myArchiverStats),
5394                           fpin) != sizeof(myArchiverStats))
5395         {
5396                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5397                                 (errmsg("corrupted statistics file \"%s\"", statfile)));
5398                 FreeFile(fpin);
5399                 return false;
5400         }
5401
5402         /* By default, we're going to return the timestamp of the global file. */
5403         *ts = myGlobalStats.stats_timestamp;
5404
5405         /*
5406          * We found an existing collector stats file.  Read it and look for a
5407          * record for the requested database.  If found, use its timestamp.
5408          */
5409         for (;;)
5410         {
5411                 switch (fgetc(fpin))
5412                 {
5413                                 /*
5414                                  * 'D'  A PgStat_StatDBEntry struct describing a database
5415                                  * follows.
5416                                  */
5417                         case 'D':
5418                                 if (fread(&dbentry, 1, offsetof(PgStat_StatDBEntry, tables),
5419                                                   fpin) != offsetof(PgStat_StatDBEntry, tables))
5420                                 {
5421                                         ereport(pgStatRunningInCollector ? LOG : WARNING,
5422                                                         (errmsg("corrupted statistics file \"%s\"",
5423                                                                         statfile)));
5424                                         goto done;
5425                                 }
5426
5427                                 /*
5428                                  * If this is the DB we're looking for, save its timestamp and
5429                                  * we're done.
5430                                  */
5431                                 if (dbentry.databaseid == databaseid)
5432                                 {
5433                                         *ts = dbentry.stats_timestamp;
5434                                         goto done;
5435                                 }
5436
5437                                 break;
5438
5439                         case 'E':
5440                                 goto done;
5441
5442                         default:
5443                                 ereport(pgStatRunningInCollector ? LOG : WARNING,
5444                                                 (errmsg("corrupted statistics file \"%s\"",
5445                                                                 statfile)));
5446                                 goto done;
5447                 }
5448         }
5449
5450 done:
5451         FreeFile(fpin);
5452         return true;
5453 }
5454
5455 /*
5456  * If not already done, read the statistics collector stats file into
5457  * some hash tables.  The results will be kept until pgstat_clear_snapshot()
5458  * is called (typically, at end of transaction).
5459  */
5460 static void
5461 backend_read_statsfile(void)
5462 {
5463         TimestampTz min_ts = 0;
5464         TimestampTz ref_ts = 0;
5465         Oid                     inquiry_db;
5466         int                     count;
5467
5468         /* already read it? */
5469         if (pgStatDBHash)
5470                 return;
5471         Assert(!pgStatRunningInCollector);
5472
5473         /*
5474          * In a normal backend, we check staleness of the data for our own DB, and
5475          * so we send MyDatabaseId in inquiry messages.  In the autovac launcher,
5476          * check staleness of the shared-catalog data, and send InvalidOid in
5477          * inquiry messages so as not to force writing unnecessary data.
5478          */
5479         if (IsAutoVacuumLauncherProcess())
5480                 inquiry_db = InvalidOid;
5481         else
5482                 inquiry_db = MyDatabaseId;
5483
5484         /*
5485          * Loop until fresh enough stats file is available or we ran out of time.
5486          * The stats inquiry message is sent repeatedly in case collector drops
5487          * it; but not every single time, as that just swamps the collector.
5488          */
5489         for (count = 0; count < PGSTAT_POLL_LOOP_COUNT; count++)
5490         {
5491                 bool            ok;
5492                 TimestampTz file_ts = 0;
5493                 TimestampTz cur_ts;
5494
5495                 CHECK_FOR_INTERRUPTS();
5496
5497                 ok = pgstat_read_db_statsfile_timestamp(inquiry_db, false, &file_ts);
5498
5499                 cur_ts = GetCurrentTimestamp();
5500                 /* Calculate min acceptable timestamp, if we didn't already */
5501                 if (count == 0 || cur_ts < ref_ts)
5502                 {
5503                         /*
5504                          * We set the minimum acceptable timestamp to PGSTAT_STAT_INTERVAL
5505                          * msec before now.  This indirectly ensures that the collector
5506                          * needn't write the file more often than PGSTAT_STAT_INTERVAL. In
5507                          * an autovacuum worker, however, we want a lower delay to avoid
5508                          * using stale data, so we use PGSTAT_RETRY_DELAY (since the
5509                          * number of workers is low, this shouldn't be a problem).
5510                          *
5511                          * We don't recompute min_ts after sleeping, except in the
5512                          * unlikely case that cur_ts went backwards.  So we might end up
5513                          * accepting a file a bit older than PGSTAT_STAT_INTERVAL.  In
5514                          * practice that shouldn't happen, though, as long as the sleep
5515                          * time is less than PGSTAT_STAT_INTERVAL; and we don't want to
5516                          * tell the collector that our cutoff time is less than what we'd
5517                          * actually accept.
5518                          */
5519                         ref_ts = cur_ts;
5520                         if (IsAutoVacuumWorkerProcess())
5521                                 min_ts = TimestampTzPlusMilliseconds(ref_ts,
5522                                                                                                          -PGSTAT_RETRY_DELAY);
5523                         else
5524                                 min_ts = TimestampTzPlusMilliseconds(ref_ts,
5525                                                                                                          -PGSTAT_STAT_INTERVAL);
5526                 }
5527
5528                 /*
5529                  * If the file timestamp is actually newer than cur_ts, we must have
5530                  * had a clock glitch (system time went backwards) or there is clock
5531                  * skew between our processor and the stats collector's processor.
5532                  * Accept the file, but send an inquiry message anyway to make
5533                  * pgstat_recv_inquiry do a sanity check on the collector's time.
5534                  */
5535                 if (ok && file_ts > cur_ts)
5536                 {
5537                         /*
5538                          * A small amount of clock skew between processors isn't terribly
5539                          * surprising, but a large difference is worth logging.  We
5540                          * arbitrarily define "large" as 1000 msec.
5541                          */
5542                         if (file_ts >= TimestampTzPlusMilliseconds(cur_ts, 1000))
5543                         {
5544                                 char       *filetime;
5545                                 char       *mytime;
5546
5547                                 /* Copy because timestamptz_to_str returns a static buffer */
5548                                 filetime = pstrdup(timestamptz_to_str(file_ts));
5549                                 mytime = pstrdup(timestamptz_to_str(cur_ts));
5550                                 elog(LOG, "stats collector's time %s is later than backend local time %s",
5551                                          filetime, mytime);
5552                                 pfree(filetime);
5553                                 pfree(mytime);
5554                         }
5555
5556                         pgstat_send_inquiry(cur_ts, min_ts, inquiry_db);
5557                         break;
5558                 }
5559
5560                 /* Normal acceptance case: file is not older than cutoff time */
5561                 if (ok && file_ts >= min_ts)
5562                         break;
5563
5564                 /* Not there or too old, so kick the collector and wait a bit */
5565                 if ((count % PGSTAT_INQ_LOOP_COUNT) == 0)
5566                         pgstat_send_inquiry(cur_ts, min_ts, inquiry_db);
5567
5568                 pg_usleep(PGSTAT_RETRY_DELAY * 1000L);
5569         }
5570
5571         if (count >= PGSTAT_POLL_LOOP_COUNT)
5572                 ereport(LOG,
5573                                 (errmsg("using stale statistics instead of current ones "
5574                                                 "because stats collector is not responding")));
5575
5576         /*
5577          * Autovacuum launcher wants stats about all databases, but a shallow read
5578          * is sufficient.  Regular backends want a deep read for just the tables
5579          * they can see (MyDatabaseId + shared catalogs).
5580          */
5581         if (IsAutoVacuumLauncherProcess())
5582                 pgStatDBHash = pgstat_read_statsfiles(InvalidOid, false, false);
5583         else
5584                 pgStatDBHash = pgstat_read_statsfiles(MyDatabaseId, false, true);
5585 }
5586
5587
5588 /* ----------
5589  * pgstat_setup_memcxt() -
5590  *
5591  *      Create pgStatLocalContext, if not already done.
5592  * ----------
5593  */
5594 static void
5595 pgstat_setup_memcxt(void)
5596 {
5597         if (!pgStatLocalContext)
5598                 pgStatLocalContext = AllocSetContextCreate(TopMemoryContext,
5599                                                                                                    "Statistics snapshot",
5600                                                                                                    ALLOCSET_SMALL_SIZES);
5601 }
5602
5603
5604 /* ----------
5605  * pgstat_clear_snapshot() -
5606  *
5607  *      Discard any data collected in the current transaction.  Any subsequent
5608  *      request will cause new snapshots to be read.
5609  *
5610  *      This is also invoked during transaction commit or abort to discard
5611  *      the no-longer-wanted snapshot.
5612  * ----------
5613  */
5614 void
5615 pgstat_clear_snapshot(void)
5616 {
5617         /* Release memory, if any was allocated */
5618         if (pgStatLocalContext)
5619                 MemoryContextDelete(pgStatLocalContext);
5620
5621         /* Reset variables */
5622         pgStatLocalContext = NULL;
5623         pgStatDBHash = NULL;
5624         localBackendStatusTable = NULL;
5625         localNumBackends = 0;
5626 }
5627
5628
5629 /* ----------
5630  * pgstat_recv_inquiry() -
5631  *
5632  *      Process stat inquiry requests.
5633  * ----------
5634  */
5635 static void
5636 pgstat_recv_inquiry(PgStat_MsgInquiry *msg, int len)
5637 {
5638         PgStat_StatDBEntry *dbentry;
5639
5640         elog(DEBUG2, "received inquiry for database %u", msg->databaseid);
5641
5642         /*
5643          * If there's already a write request for this DB, there's nothing to do.
5644          *
5645          * Note that if a request is found, we return early and skip the below
5646          * check for clock skew.  This is okay, since the only way for a DB
5647          * request to be present in the list is that we have been here since the
5648          * last write round.  It seems sufficient to check for clock skew once per
5649          * write round.
5650          */
5651         if (list_member_oid(pending_write_requests, msg->databaseid))
5652                 return;
5653
5654         /*
5655          * Check to see if we last wrote this database at a time >= the requested
5656          * cutoff time.  If so, this is a stale request that was generated before
5657          * we updated the DB file, and we don't need to do so again.
5658          *
5659          * If the requestor's local clock time is older than stats_timestamp, we
5660          * should suspect a clock glitch, ie system time going backwards; though
5661          * the more likely explanation is just delayed message receipt.  It is
5662          * worth expending a GetCurrentTimestamp call to be sure, since a large
5663          * retreat in the system clock reading could otherwise cause us to neglect
5664          * to update the stats file for a long time.
5665          */
5666         dbentry = pgstat_get_db_entry(msg->databaseid, false);
5667         if (dbentry == NULL)
5668         {
5669                 /*
5670                  * We have no data for this DB.  Enter a write request anyway so that
5671                  * the global stats will get updated.  This is needed to prevent
5672                  * backend_read_statsfile from waiting for data that we cannot supply,
5673                  * in the case of a new DB that nobody has yet reported any stats for.
5674                  * See the behavior of pgstat_read_db_statsfile_timestamp.
5675                  */
5676         }
5677         else if (msg->clock_time < dbentry->stats_timestamp)
5678         {
5679                 TimestampTz cur_ts = GetCurrentTimestamp();
5680
5681                 if (cur_ts < dbentry->stats_timestamp)
5682                 {
5683                         /*
5684                          * Sure enough, time went backwards.  Force a new stats file write
5685                          * to get back in sync; but first, log a complaint.
5686                          */
5687                         char       *writetime;
5688                         char       *mytime;
5689
5690                         /* Copy because timestamptz_to_str returns a static buffer */
5691                         writetime = pstrdup(timestamptz_to_str(dbentry->stats_timestamp));
5692                         mytime = pstrdup(timestamptz_to_str(cur_ts));
5693                         elog(LOG,
5694                                  "stats_timestamp %s is later than collector's time %s for database %u",
5695                                  writetime, mytime, dbentry->databaseid);
5696                         pfree(writetime);
5697                         pfree(mytime);
5698                 }
5699                 else
5700                 {
5701                         /*
5702                          * Nope, it's just an old request.  Assuming msg's clock_time is
5703                          * >= its cutoff_time, it must be stale, so we can ignore it.
5704                          */
5705                         return;
5706                 }
5707         }
5708         else if (msg->cutoff_time <= dbentry->stats_timestamp)
5709         {
5710                 /* Stale request, ignore it */
5711                 return;
5712         }
5713
5714         /*
5715          * We need to write this DB, so create a request.
5716          */
5717         pending_write_requests = lappend_oid(pending_write_requests,
5718                                                                                  msg->databaseid);
5719 }
5720
5721
5722 /* ----------
5723  * pgstat_recv_tabstat() -
5724  *
5725  *      Count what the backend has done.
5726  * ----------
5727  */
5728 static void
5729 pgstat_recv_tabstat(PgStat_MsgTabstat *msg, int len)
5730 {
5731         PgStat_StatDBEntry *dbentry;
5732         PgStat_StatTabEntry *tabentry;
5733         int                     i;
5734         bool            found;
5735
5736         dbentry = pgstat_get_db_entry(msg->m_databaseid, true);
5737
5738         /*
5739          * Update database-wide stats.
5740          */
5741         dbentry->n_xact_commit += (PgStat_Counter) (msg->m_xact_commit);
5742         dbentry->n_xact_rollback += (PgStat_Counter) (msg->m_xact_rollback);
5743         dbentry->n_block_read_time += msg->m_block_read_time;
5744         dbentry->n_block_write_time += msg->m_block_write_time;
5745
5746         /*
5747          * Process all table entries in the message.
5748          */
5749         for (i = 0; i < msg->m_nentries; i++)
5750         {
5751                 PgStat_TableEntry *tabmsg = &(msg->m_entry[i]);
5752
5753                 tabentry = (PgStat_StatTabEntry *) hash_search(dbentry->tables,
5754                                                                                                            (void *) &(tabmsg->t_id),
5755                                                                                                            HASH_ENTER, &found);
5756
5757                 if (!found)
5758                 {
5759                         /*
5760                          * If it's a new table entry, initialize counters to the values we
5761                          * just got.
5762                          */
5763                         tabentry->numscans = tabmsg->t_counts.t_numscans;
5764                         tabentry->tuples_returned = tabmsg->t_counts.t_tuples_returned;
5765                         tabentry->tuples_fetched = tabmsg->t_counts.t_tuples_fetched;
5766                         tabentry->tuples_inserted = tabmsg->t_counts.t_tuples_inserted;
5767                         tabentry->tuples_updated = tabmsg->t_counts.t_tuples_updated;
5768                         tabentry->tuples_deleted = tabmsg->t_counts.t_tuples_deleted;
5769                         tabentry->tuples_hot_updated = tabmsg->t_counts.t_tuples_hot_updated;
5770                         tabentry->n_live_tuples = tabmsg->t_counts.t_delta_live_tuples;
5771                         tabentry->n_dead_tuples = tabmsg->t_counts.t_delta_dead_tuples;
5772                         tabentry->changes_since_analyze = tabmsg->t_counts.t_changed_tuples;
5773                         tabentry->blocks_fetched = tabmsg->t_counts.t_blocks_fetched;
5774                         tabentry->blocks_hit = tabmsg->t_counts.t_blocks_hit;
5775
5776                         tabentry->vacuum_timestamp = 0;
5777                         tabentry->vacuum_count = 0;
5778                         tabentry->autovac_vacuum_timestamp = 0;
5779                         tabentry->autovac_vacuum_count = 0;
5780                         tabentry->analyze_timestamp = 0;
5781                         tabentry->analyze_count = 0;
5782                         tabentry->autovac_analyze_timestamp = 0;
5783                         tabentry->autovac_analyze_count = 0;
5784                 }
5785                 else
5786                 {
5787                         /*
5788                          * Otherwise add the values to the existing entry.
5789                          */
5790                         tabentry->numscans += tabmsg->t_counts.t_numscans;
5791                         tabentry->tuples_returned += tabmsg->t_counts.t_tuples_returned;
5792                         tabentry->tuples_fetched += tabmsg->t_counts.t_tuples_fetched;
5793                         tabentry->tuples_inserted += tabmsg->t_counts.t_tuples_inserted;
5794                         tabentry->tuples_updated += tabmsg->t_counts.t_tuples_updated;
5795                         tabentry->tuples_deleted += tabmsg->t_counts.t_tuples_deleted;
5796                         tabentry->tuples_hot_updated += tabmsg->t_counts.t_tuples_hot_updated;
5797                         /* If table was truncated, first reset the live/dead counters */
5798                         if (tabmsg->t_counts.t_truncated)
5799                         {
5800                                 tabentry->n_live_tuples = 0;
5801                                 tabentry->n_dead_tuples = 0;
5802                         }
5803                         tabentry->n_live_tuples += tabmsg->t_counts.t_delta_live_tuples;
5804                         tabentry->n_dead_tuples += tabmsg->t_counts.t_delta_dead_tuples;
5805                         tabentry->changes_since_analyze += tabmsg->t_counts.t_changed_tuples;
5806                         tabentry->blocks_fetched += tabmsg->t_counts.t_blocks_fetched;
5807                         tabentry->blocks_hit += tabmsg->t_counts.t_blocks_hit;
5808                 }
5809
5810                 /* Clamp n_live_tuples in case of negative delta_live_tuples */
5811                 tabentry->n_live_tuples = Max(tabentry->n_live_tuples, 0);
5812                 /* Likewise for n_dead_tuples */
5813                 tabentry->n_dead_tuples = Max(tabentry->n_dead_tuples, 0);
5814
5815                 /*
5816                  * Add per-table stats to the per-database entry, too.
5817                  */
5818                 dbentry->n_tuples_returned += tabmsg->t_counts.t_tuples_returned;
5819                 dbentry->n_tuples_fetched += tabmsg->t_counts.t_tuples_fetched;
5820                 dbentry->n_tuples_inserted += tabmsg->t_counts.t_tuples_inserted;
5821                 dbentry->n_tuples_updated += tabmsg->t_counts.t_tuples_updated;
5822                 dbentry->n_tuples_deleted += tabmsg->t_counts.t_tuples_deleted;
5823                 dbentry->n_blocks_fetched += tabmsg->t_counts.t_blocks_fetched;
5824                 dbentry->n_blocks_hit += tabmsg->t_counts.t_blocks_hit;
5825         }
5826 }
5827
5828
5829 /* ----------
5830  * pgstat_recv_tabpurge() -
5831  *
5832  *      Arrange for dead table removal.
5833  * ----------
5834  */
5835 static void
5836 pgstat_recv_tabpurge(PgStat_MsgTabpurge *msg, int len)
5837 {
5838         PgStat_StatDBEntry *dbentry;
5839         int                     i;
5840
5841         dbentry = pgstat_get_db_entry(msg->m_databaseid, false);
5842
5843         /*
5844          * No need to purge if we don't even know the database.
5845          */
5846         if (!dbentry || !dbentry->tables)
5847                 return;
5848
5849         /*
5850          * Process all table entries in the message.
5851          */
5852         for (i = 0; i < msg->m_nentries; i++)
5853         {
5854                 /* Remove from hashtable if present; we don't care if it's not. */
5855                 (void) hash_search(dbentry->tables,
5856                                                    (void *) &(msg->m_tableid[i]),
5857                                                    HASH_REMOVE, NULL);
5858         }
5859 }
5860
5861
5862 /* ----------
5863  * pgstat_recv_dropdb() -
5864  *
5865  *      Arrange for dead database removal
5866  * ----------
5867  */
5868 static void
5869 pgstat_recv_dropdb(PgStat_MsgDropdb *msg, int len)
5870 {
5871         Oid                     dbid = msg->m_databaseid;
5872         PgStat_StatDBEntry *dbentry;
5873
5874         /*
5875          * Lookup the database in the hashtable.
5876          */
5877         dbentry = pgstat_get_db_entry(dbid, false);
5878
5879         /*
5880          * If found, remove it (along with the db statfile).
5881          */
5882         if (dbentry)
5883         {
5884                 char            statfile[MAXPGPATH];
5885
5886                 get_dbstat_filename(false, false, dbid, statfile, MAXPGPATH);
5887
5888                 elog(DEBUG2, "removing stats file \"%s\"", statfile);
5889                 unlink(statfile);
5890
5891                 if (dbentry->tables != NULL)
5892                         hash_destroy(dbentry->tables);
5893                 if (dbentry->functions != NULL)
5894                         hash_destroy(dbentry->functions);
5895
5896                 if (hash_search(pgStatDBHash,
5897                                                 (void *) &dbid,
5898                                                 HASH_REMOVE, NULL) == NULL)
5899                         ereport(ERROR,
5900                                         (errmsg("database hash table corrupted during cleanup --- abort")));
5901         }
5902 }
5903
5904
5905 /* ----------
5906  * pgstat_recv_resetcounter() -
5907  *
5908  *      Reset the statistics for the specified database.
5909  * ----------
5910  */
5911 static void
5912 pgstat_recv_resetcounter(PgStat_MsgResetcounter *msg, int len)
5913 {
5914         PgStat_StatDBEntry *dbentry;
5915
5916         /*
5917          * Lookup the database in the hashtable.  Nothing to do if not there.
5918          */
5919         dbentry = pgstat_get_db_entry(msg->m_databaseid, false);
5920
5921         if (!dbentry)
5922                 return;
5923
5924         /*
5925          * We simply throw away all the database's table entries by recreating a
5926          * new hash table for them.
5927          */
5928         if (dbentry->tables != NULL)
5929                 hash_destroy(dbentry->tables);
5930         if (dbentry->functions != NULL)
5931                 hash_destroy(dbentry->functions);
5932
5933         dbentry->tables = NULL;
5934         dbentry->functions = NULL;
5935
5936         /*
5937          * Reset database-level stats, too.  This creates empty hash tables for
5938          * tables and functions.
5939          */
5940         reset_dbentry_counters(dbentry);
5941 }
5942
5943 /* ----------
5944  * pgstat_recv_resetshared() -
5945  *
5946  *      Reset some shared statistics of the cluster.
5947  * ----------
5948  */
5949 static void
5950 pgstat_recv_resetsharedcounter(PgStat_MsgResetsharedcounter *msg, int len)
5951 {
5952         if (msg->m_resettarget == RESET_BGWRITER)
5953         {
5954                 /* Reset the global background writer statistics for the cluster. */
5955                 memset(&globalStats, 0, sizeof(globalStats));
5956                 globalStats.stat_reset_timestamp = GetCurrentTimestamp();
5957         }
5958         else if (msg->m_resettarget == RESET_ARCHIVER)
5959         {
5960                 /* Reset the archiver statistics for the cluster. */
5961                 memset(&archiverStats, 0, sizeof(archiverStats));
5962                 archiverStats.stat_reset_timestamp = GetCurrentTimestamp();
5963         }
5964
5965         /*
5966          * Presumably the sender of this message validated the target, don't
5967          * complain here if it's not valid
5968          */
5969 }
5970
5971 /* ----------
5972  * pgstat_recv_resetsinglecounter() -
5973  *
5974  *      Reset a statistics for a single object
5975  * ----------
5976  */
5977 static void
5978 pgstat_recv_resetsinglecounter(PgStat_MsgResetsinglecounter *msg, int len)
5979 {
5980         PgStat_StatDBEntry *dbentry;
5981
5982         dbentry = pgstat_get_db_entry(msg->m_databaseid, false);
5983
5984         if (!dbentry)
5985                 return;
5986
5987         /* Set the reset timestamp for the whole database */
5988         dbentry->stat_reset_timestamp = GetCurrentTimestamp();
5989
5990         /* Remove object if it exists, ignore it if not */
5991         if (msg->m_resettype == RESET_TABLE)
5992                 (void) hash_search(dbentry->tables, (void *) &(msg->m_objectid),
5993                                                    HASH_REMOVE, NULL);
5994         else if (msg->m_resettype == RESET_FUNCTION)
5995                 (void) hash_search(dbentry->functions, (void *) &(msg->m_objectid),
5996                                                    HASH_REMOVE, NULL);
5997 }
5998
5999 /* ----------
6000  * pgstat_recv_autovac() -
6001  *
6002  *      Process an autovacuum signalling message.
6003  * ----------
6004  */
6005 static void
6006 pgstat_recv_autovac(PgStat_MsgAutovacStart *msg, int len)
6007 {
6008         PgStat_StatDBEntry *dbentry;
6009
6010         /*
6011          * Store the last autovacuum time in the database's hashtable entry.
6012          */
6013         dbentry = pgstat_get_db_entry(msg->m_databaseid, true);
6014
6015         dbentry->last_autovac_time = msg->m_start_time;
6016 }
6017
6018 /* ----------
6019  * pgstat_recv_vacuum() -
6020  *
6021  *      Process a VACUUM message.
6022  * ----------
6023  */
6024 static void
6025 pgstat_recv_vacuum(PgStat_MsgVacuum *msg, int len)
6026 {
6027         PgStat_StatDBEntry *dbentry;
6028         PgStat_StatTabEntry *tabentry;
6029
6030         /*
6031          * Store the data in the table's hashtable entry.
6032          */
6033         dbentry = pgstat_get_db_entry(msg->m_databaseid, true);
6034
6035         tabentry = pgstat_get_tab_entry(dbentry, msg->m_tableoid, true);
6036
6037         tabentry->n_live_tuples = msg->m_live_tuples;
6038         tabentry->n_dead_tuples = msg->m_dead_tuples;
6039
6040         if (msg->m_autovacuum)
6041         {
6042                 tabentry->autovac_vacuum_timestamp = msg->m_vacuumtime;
6043                 tabentry->autovac_vacuum_count++;
6044         }
6045         else
6046         {
6047                 tabentry->vacuum_timestamp = msg->m_vacuumtime;
6048                 tabentry->vacuum_count++;
6049         }
6050 }
6051
6052 /* ----------
6053  * pgstat_recv_analyze() -
6054  *
6055  *      Process an ANALYZE message.
6056  * ----------
6057  */
6058 static void
6059 pgstat_recv_analyze(PgStat_MsgAnalyze *msg, int len)
6060 {
6061         PgStat_StatDBEntry *dbentry;
6062         PgStat_StatTabEntry *tabentry;
6063
6064         /*
6065          * Store the data in the table's hashtable entry.
6066          */
6067         dbentry = pgstat_get_db_entry(msg->m_databaseid, true);
6068
6069         tabentry = pgstat_get_tab_entry(dbentry, msg->m_tableoid, true);
6070
6071         tabentry->n_live_tuples = msg->m_live_tuples;
6072         tabentry->n_dead_tuples = msg->m_dead_tuples;
6073
6074         /*
6075          * If commanded, reset changes_since_analyze to zero.  This forgets any
6076          * changes that were committed while the ANALYZE was in progress, but we
6077          * have no good way to estimate how many of those there were.
6078          */
6079         if (msg->m_resetcounter)
6080                 tabentry->changes_since_analyze = 0;
6081
6082         if (msg->m_autovacuum)
6083         {
6084                 tabentry->autovac_analyze_timestamp = msg->m_analyzetime;
6085                 tabentry->autovac_analyze_count++;
6086         }
6087         else
6088         {
6089                 tabentry->analyze_timestamp = msg->m_analyzetime;
6090                 tabentry->analyze_count++;
6091         }
6092 }
6093
6094
6095 /* ----------
6096  * pgstat_recv_archiver() -
6097  *
6098  *      Process a ARCHIVER message.
6099  * ----------
6100  */
6101 static void
6102 pgstat_recv_archiver(PgStat_MsgArchiver *msg, int len)
6103 {
6104         if (msg->m_failed)
6105         {
6106                 /* Failed archival attempt */
6107                 ++archiverStats.failed_count;
6108                 memcpy(archiverStats.last_failed_wal, msg->m_xlog,
6109                            sizeof(archiverStats.last_failed_wal));
6110                 archiverStats.last_failed_timestamp = msg->m_timestamp;
6111         }
6112         else
6113         {
6114                 /* Successful archival operation */
6115                 ++archiverStats.archived_count;
6116                 memcpy(archiverStats.last_archived_wal, msg->m_xlog,
6117                            sizeof(archiverStats.last_archived_wal));
6118                 archiverStats.last_archived_timestamp = msg->m_timestamp;
6119         }
6120 }
6121
6122 /* ----------
6123  * pgstat_recv_bgwriter() -
6124  *
6125  *      Process a BGWRITER message.
6126  * ----------
6127  */
6128 static void
6129 pgstat_recv_bgwriter(PgStat_MsgBgWriter *msg, int len)
6130 {
6131         globalStats.timed_checkpoints += msg->m_timed_checkpoints;
6132         globalStats.requested_checkpoints += msg->m_requested_checkpoints;
6133         globalStats.checkpoint_write_time += msg->m_checkpoint_write_time;
6134         globalStats.checkpoint_sync_time += msg->m_checkpoint_sync_time;
6135         globalStats.buf_written_checkpoints += msg->m_buf_written_checkpoints;
6136         globalStats.buf_written_clean += msg->m_buf_written_clean;
6137         globalStats.maxwritten_clean += msg->m_maxwritten_clean;
6138         globalStats.buf_written_backend += msg->m_buf_written_backend;
6139         globalStats.buf_fsync_backend += msg->m_buf_fsync_backend;
6140         globalStats.buf_alloc += msg->m_buf_alloc;
6141 }
6142
6143 /* ----------
6144  * pgstat_recv_recoveryconflict() -
6145  *
6146  *      Process a RECOVERYCONFLICT message.
6147  * ----------
6148  */
6149 static void
6150 pgstat_recv_recoveryconflict(PgStat_MsgRecoveryConflict *msg, int len)
6151 {
6152         PgStat_StatDBEntry *dbentry;
6153
6154         dbentry = pgstat_get_db_entry(msg->m_databaseid, true);
6155
6156         switch (msg->m_reason)
6157         {
6158                 case PROCSIG_RECOVERY_CONFLICT_DATABASE:
6159
6160                         /*
6161                          * Since we drop the information about the database as soon as it
6162                          * replicates, there is no point in counting these conflicts.
6163                          */
6164                         break;
6165                 case PROCSIG_RECOVERY_CONFLICT_TABLESPACE:
6166                         dbentry->n_conflict_tablespace++;
6167                         break;
6168                 case PROCSIG_RECOVERY_CONFLICT_LOCK:
6169                         dbentry->n_conflict_lock++;
6170                         break;
6171                 case PROCSIG_RECOVERY_CONFLICT_SNAPSHOT:
6172                         dbentry->n_conflict_snapshot++;
6173                         break;
6174                 case PROCSIG_RECOVERY_CONFLICT_BUFFERPIN:
6175                         dbentry->n_conflict_bufferpin++;
6176                         break;
6177                 case PROCSIG_RECOVERY_CONFLICT_STARTUP_DEADLOCK:
6178                         dbentry->n_conflict_startup_deadlock++;
6179                         break;
6180         }
6181 }
6182
6183 /* ----------
6184  * pgstat_recv_deadlock() -
6185  *
6186  *      Process a DEADLOCK message.
6187  * ----------
6188  */
6189 static void
6190 pgstat_recv_deadlock(PgStat_MsgDeadlock *msg, int len)
6191 {
6192         PgStat_StatDBEntry *dbentry;
6193
6194         dbentry = pgstat_get_db_entry(msg->m_databaseid, true);
6195
6196         dbentry->n_deadlocks++;
6197 }
6198
6199 /* ----------
6200  * pgstat_recv_tempfile() -
6201  *
6202  *      Process a TEMPFILE message.
6203  * ----------
6204  */
6205 static void
6206 pgstat_recv_tempfile(PgStat_MsgTempFile *msg, int len)
6207 {
6208         PgStat_StatDBEntry *dbentry;
6209
6210         dbentry = pgstat_get_db_entry(msg->m_databaseid, true);
6211
6212         dbentry->n_temp_bytes += msg->m_filesize;
6213         dbentry->n_temp_files += 1;
6214 }
6215
6216 /* ----------
6217  * pgstat_recv_funcstat() -
6218  *
6219  *      Count what the backend has done.
6220  * ----------
6221  */
6222 static void
6223 pgstat_recv_funcstat(PgStat_MsgFuncstat *msg, int len)
6224 {
6225         PgStat_FunctionEntry *funcmsg = &(msg->m_entry[0]);
6226         PgStat_StatDBEntry *dbentry;
6227         PgStat_StatFuncEntry *funcentry;
6228         int                     i;
6229         bool            found;
6230
6231         dbentry = pgstat_get_db_entry(msg->m_databaseid, true);
6232
6233         /*
6234          * Process all function entries in the message.
6235          */
6236         for (i = 0; i < msg->m_nentries; i++, funcmsg++)
6237         {
6238                 funcentry = (PgStat_StatFuncEntry *) hash_search(dbentry->functions,
6239                                                                                                                  (void *) &(funcmsg->f_id),
6240                                                                                                                  HASH_ENTER, &found);
6241
6242                 if (!found)
6243                 {
6244                         /*
6245                          * If it's a new function entry, initialize counters to the values
6246                          * we just got.
6247                          */
6248                         funcentry->f_numcalls = funcmsg->f_numcalls;
6249                         funcentry->f_total_time = funcmsg->f_total_time;
6250                         funcentry->f_self_time = funcmsg->f_self_time;
6251                 }
6252                 else
6253                 {
6254                         /*
6255                          * Otherwise add the values to the existing entry.
6256                          */
6257                         funcentry->f_numcalls += funcmsg->f_numcalls;
6258                         funcentry->f_total_time += funcmsg->f_total_time;
6259                         funcentry->f_self_time += funcmsg->f_self_time;
6260                 }
6261         }
6262 }
6263
6264 /* ----------
6265  * pgstat_recv_funcpurge() -
6266  *
6267  *      Arrange for dead function removal.
6268  * ----------
6269  */
6270 static void
6271 pgstat_recv_funcpurge(PgStat_MsgFuncpurge *msg, int len)
6272 {
6273         PgStat_StatDBEntry *dbentry;
6274         int                     i;
6275
6276         dbentry = pgstat_get_db_entry(msg->m_databaseid, false);
6277
6278         /*
6279          * No need to purge if we don't even know the database.
6280          */
6281         if (!dbentry || !dbentry->functions)
6282                 return;
6283
6284         /*
6285          * Process all function entries in the message.
6286          */
6287         for (i = 0; i < msg->m_nentries; i++)
6288         {
6289                 /* Remove from hashtable if present; we don't care if it's not. */
6290                 (void) hash_search(dbentry->functions,
6291                                                    (void *) &(msg->m_functionid[i]),
6292                                                    HASH_REMOVE, NULL);
6293         }
6294 }
6295
6296 /* ----------
6297  * pgstat_write_statsfile_needed() -
6298  *
6299  *      Do we need to write out any stats files?
6300  * ----------
6301  */
6302 static bool
6303 pgstat_write_statsfile_needed(void)
6304 {
6305         if (pending_write_requests != NIL)
6306                 return true;
6307
6308         /* Everything was written recently */
6309         return false;
6310 }
6311
6312 /* ----------
6313  * pgstat_db_requested() -
6314  *
6315  *      Checks whether stats for a particular DB need to be written to a file.
6316  * ----------
6317  */
6318 static bool
6319 pgstat_db_requested(Oid databaseid)
6320 {
6321         /*
6322          * If any requests are outstanding at all, we should write the stats for
6323          * shared catalogs (the "database" with OID 0).  This ensures that
6324          * backends will see up-to-date stats for shared catalogs, even though
6325          * they send inquiry messages mentioning only their own DB.
6326          */
6327         if (databaseid == InvalidOid && pending_write_requests != NIL)
6328                 return true;
6329
6330         /* Search to see if there's an open request to write this database. */
6331         if (list_member_oid(pending_write_requests, databaseid))
6332                 return true;
6333
6334         return false;
6335 }
6336
6337 /*
6338  * Convert a potentially unsafely truncated activity string (see
6339  * PgBackendStatus.st_activity_raw's documentation) into a correctly truncated
6340  * one.
6341  *
6342  * The returned string is allocated in the caller's memory context and may be
6343  * freed.
6344  */
6345 char *
6346 pgstat_clip_activity(const char *raw_activity)
6347 {
6348         char       *activity;
6349         int                     rawlen;
6350         int                     cliplen;
6351
6352         /*
6353          * Some callers, like pgstat_get_backend_current_activity(), do not
6354          * guarantee that the buffer isn't concurrently modified. We try to take
6355          * care that the buffer is always terminated by a NUL byte regardless, but
6356          * let's still be paranoid about the string's length. In those cases the
6357          * underlying buffer is guaranteed to be pgstat_track_activity_query_size
6358          * large.
6359          */
6360         activity = pnstrdup(raw_activity, pgstat_track_activity_query_size - 1);
6361
6362         /* now double-guaranteed to be NUL terminated */
6363         rawlen = strlen(activity);
6364
6365         /*
6366          * All supported server-encodings make it possible to determine the length
6367          * of a multi-byte character from its first byte (this is not the case for
6368          * client encodings, see GB18030). As st_activity is always stored using
6369          * server encoding, this allows us to perform multi-byte aware truncation,
6370          * even if the string earlier was truncated in the middle of a multi-byte
6371          * character.
6372          */
6373         cliplen = pg_mbcliplen(activity, rawlen,
6374                                                    pgstat_track_activity_query_size - 1);
6375
6376         activity[cliplen] = '\0';
6377
6378         return activity;
6379 }