granicus.if.org Git - postgresql/blob - contrib/pg_prewarm/autoprewarm.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * autoprewarm.c
   4  *              Periodically dump information about the blocks present in
   5  *              shared_buffers, and reload them on server restart.
   6  *
   7  *              Due to locking considerations, we can't actually begin prewarming
   8  *              until the server reaches a consistent state.  We need the catalogs
   9  *              to be consistent so that we can figure out which relation to lock,
  10  *              and we need to lock the relations so that we don't try to prewarm
  11  *              pages from a relation that is in the process of being dropped.
  12  *
  13  *              While prewarming, autoprewarm will use two workers.  There's a
  14  *              master worker that reads and sorts the list of blocks to be
  15  *              prewarmed and then launches a per-database worker for each
  16  *              relevant database in turn.  The former keeps running after the
  17  *              initial prewarm is complete to update the dump file periodically.
  18  *
  19  *      Copyright (c) 2016-2019, PostgreSQL Global Development Group
  20  *
  21  *      IDENTIFICATION
  22  *              contrib/pg_prewarm/autoprewarm.c
  23  *
  24  *-------------------------------------------------------------------------
  25  */
  26
  27 #include "postgres.h"
  28
  29 #include <unistd.h>
  30
  31 #include "access/relation.h"
  32 #include "access/xact.h"
  33 #include "catalog/pg_class.h"
  34 #include "catalog/pg_type.h"
  35 #include "miscadmin.h"
  36 #include "pgstat.h"
  37 #include "postmaster/bgworker.h"
  38 #include "storage/buf_internals.h"
  39 #include "storage/dsm.h"
  40 #include "storage/ipc.h"
  41 #include "storage/latch.h"
  42 #include "storage/lwlock.h"
  43 #include "storage/proc.h"
  44 #include "storage/procsignal.h"
  45 #include "storage/shmem.h"
  46 #include "storage/smgr.h"
  47 #include "tcop/tcopprot.h"
  48 #include "utils/acl.h"
  49 #include "utils/guc.h"
  50 #include "utils/memutils.h"
  51 #include "utils/rel.h"
  52 #include "utils/relfilenodemap.h"
  53 #include "utils/resowner.h"
  54
  55 #define AUTOPREWARM_FILE "autoprewarm.blocks"
  56
  57 /* Metadata for each block we dump. */
  58 typedef struct BlockInfoRecord
  59 {
  60         Oid                     database;
  61         Oid                     tablespace;
  62         Oid                     filenode;
  63         ForkNumber      forknum;
  64         BlockNumber blocknum;
  65 } BlockInfoRecord;
  66
  67 /* Shared state information for autoprewarm bgworker. */
  68 typedef struct AutoPrewarmSharedState
  69 {
  70         LWLock          lock;                   /* mutual exclusion */
  71         pid_t           bgworker_pid;   /* for main bgworker */
  72         pid_t           pid_using_dumpfile; /* for autoprewarm or block dump */
  73
  74         /* Following items are for communication with per-database worker */
  75         dsm_handle      block_info_handle;
  76         Oid                     database;
  77         int                     prewarm_start_idx;
  78         int                     prewarm_stop_idx;
  79         int                     prewarmed_blocks;
  80 } AutoPrewarmSharedState;
  81
  82 void            _PG_init(void);
  83 void            autoprewarm_main(Datum main_arg);
  84 void            autoprewarm_database_main(Datum main_arg);
  85
  86 PG_FUNCTION_INFO_V1(autoprewarm_start_worker);
  87 PG_FUNCTION_INFO_V1(autoprewarm_dump_now);
  88
  89 static void apw_load_buffers(void);
  90 static int      apw_dump_now(bool is_bgworker, bool dump_unlogged);
  91 static void apw_start_master_worker(void);
  92 static void apw_start_database_worker(void);
  93 static bool apw_init_shmem(void);
  94 static void apw_detach_shmem(int code, Datum arg);
  95 static int      apw_compare_blockinfo(const void *p, const void *q);
  96 static void apw_sigterm_handler(SIGNAL_ARGS);
  97 static void apw_sighup_handler(SIGNAL_ARGS);
  98
  99 /* Flags set by signal handlers */
 100 static volatile sig_atomic_t got_sigterm = false;
 101 static volatile sig_atomic_t got_sighup = false;
 102
 103 /* Pointer to shared-memory state. */
 104 static AutoPrewarmSharedState *apw_state = NULL;
 105
 106 /* GUC variables. */
 107 static bool autoprewarm = true; /* start worker? */
 108 static int      autoprewarm_interval;   /* dump interval */
 109
 110 /*
 111  * Module load callback.
 112  */
 113 void
 114 _PG_init(void)
 115 {
 116         DefineCustomIntVariable("pg_prewarm.autoprewarm_interval",
 117                                                         "Sets the interval between dumps of shared buffers",
 118                                                         "If set to zero, time-based dumping is disabled.",
 119                                                         &autoprewarm_interval,
 120                                                         300,
 121                                                         0, INT_MAX / 1000,
 122                                                         PGC_SIGHUP,
 123                                                         GUC_UNIT_S,
 124                                                         NULL,
 125                                                         NULL,
 126                                                         NULL);
 127
 128         if (!process_shared_preload_libraries_in_progress)
 129                 return;
 130
 131         /* can't define PGC_POSTMASTER variable after startup */
 132         DefineCustomBoolVariable("pg_prewarm.autoprewarm",
 133                                                          "Starts the autoprewarm worker.",
 134                                                          NULL,
 135                                                          &autoprewarm,
 136                                                          true,
 137                                                          PGC_POSTMASTER,
 138                                                          0,
 139                                                          NULL,
 140                                                          NULL,
 141                                                          NULL);
 142
 143         EmitWarningsOnPlaceholders("pg_prewarm");
 144
 145         RequestAddinShmemSpace(MAXALIGN(sizeof(AutoPrewarmSharedState)));
 146
 147         /* Register autoprewarm worker, if enabled. */
 148         if (autoprewarm)
 149                 apw_start_master_worker();
 150 }
 151
 152 /*
 153  * Main entry point for the master autoprewarm process.  Per-database workers
 154  * have a separate entry point.
 155  */
 156 void
 157 autoprewarm_main(Datum main_arg)
 158 {
 159         bool            first_time = true;
 160         TimestampTz last_dump_time = 0;
 161
 162         /* Establish signal handlers; once that's done, unblock signals. */
 163         pqsignal(SIGTERM, apw_sigterm_handler);
 164         pqsignal(SIGHUP, apw_sighup_handler);
 165         pqsignal(SIGUSR1, procsignal_sigusr1_handler);
 166         BackgroundWorkerUnblockSignals();
 167
 168         /* Create (if necessary) and attach to our shared memory area. */
 169         if (apw_init_shmem())
 170                 first_time = false;
 171
 172         /* Set on-detach hook so that our PID will be cleared on exit. */
 173         on_shmem_exit(apw_detach_shmem, 0);
 174
 175         /*
 176          * Store our PID in the shared memory area --- unless there's already
 177          * another worker running, in which case just exit.
 178          */
 179         LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
 180         if (apw_state->bgworker_pid != InvalidPid)
 181         {
 182                 LWLockRelease(&apw_state->lock);
 183                 ereport(LOG,
 184                                 (errmsg("autoprewarm worker is already running under PID %lu",
 185                                                 (unsigned long) apw_state->bgworker_pid)));
 186                 return;
 187         }
 188         apw_state->bgworker_pid = MyProcPid;
 189         LWLockRelease(&apw_state->lock);
 190
 191         /*
 192          * Preload buffers from the dump file only if we just created the shared
 193          * memory region.  Otherwise, it's either already been done or shouldn't
 194          * be done - e.g. because the old dump file has been overwritten since the
 195          * server was started.
 196          *
 197          * There's not much point in performing a dump immediately after we finish
 198          * preloading; so, if we do end up preloading, consider the last dump time
 199          * to be equal to the current time.
 200          */
 201         if (first_time)
 202         {
 203                 apw_load_buffers();
 204                 last_dump_time = GetCurrentTimestamp();
 205         }
 206
 207         /* Periodically dump buffers until terminated. */
 208         while (!got_sigterm)
 209         {
 210                 /* In case of a SIGHUP, just reload the configuration. */
 211                 if (got_sighup)
 212                 {
 213                         got_sighup = false;
 214                         ProcessConfigFile(PGC_SIGHUP);
 215                 }
 216
 217                 if (autoprewarm_interval <= 0)
 218                 {
 219                         /* We're only dumping at shutdown, so just wait forever. */
 220                         (void) WaitLatch(&MyProc->procLatch,
 221                                                          WL_LATCH_SET | WL_EXIT_ON_PM_DEATH,
 222                                                          -1L,
 223                                                          PG_WAIT_EXTENSION);
 224                 }
 225                 else
 226                 {
 227                         long            delay_in_ms = 0;
 228                         TimestampTz next_dump_time = 0;
 229                         long            secs = 0;
 230                         int                     usecs = 0;
 231
 232                         /* Compute the next dump time. */
 233                         next_dump_time =
 234                                 TimestampTzPlusMilliseconds(last_dump_time,
 235                                                                                         autoprewarm_interval * 1000);
 236                         TimestampDifference(GetCurrentTimestamp(), next_dump_time,
 237                                                                 &secs, &usecs);
 238                         delay_in_ms = secs + (usecs / 1000);
 239
 240                         /* Perform a dump if it's time. */
 241                         if (delay_in_ms <= 0)
 242                         {
 243                                 last_dump_time = GetCurrentTimestamp();
 244                                 apw_dump_now(true, false);
 245                                 continue;
 246                         }
 247
 248                         /* Sleep until the next dump time. */
 249                         (void) WaitLatch(&MyProc->procLatch,
 250                                                          WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH,
 251                                                          delay_in_ms,
 252                                                          PG_WAIT_EXTENSION);
 253                 }
 254
 255                 /* Reset the latch, loop. */
 256                 ResetLatch(&MyProc->procLatch);
 257         }
 258
 259         /*
 260          * Dump one last time.  We assume this is probably the result of a system
 261          * shutdown, although it's possible that we've merely been terminated.
 262          */
 263         apw_dump_now(true, true);
 264 }
 265
 266 /*
 267  * Read the dump file and launch per-database workers one at a time to
 268  * prewarm the buffers found there.
 269  */
 270 static void
 271 apw_load_buffers(void)
 272 {
 273         FILE       *file = NULL;
 274         int                     num_elements,
 275                                 i;
 276         BlockInfoRecord *blkinfo;
 277         dsm_segment *seg;
 278
 279         /*
 280          * Skip the prewarm if the dump file is in use; otherwise, prevent any
 281          * other process from writing it while we're using it.
 282          */
 283         LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
 284         if (apw_state->pid_using_dumpfile == InvalidPid)
 285                 apw_state->pid_using_dumpfile = MyProcPid;
 286         else
 287         {
 288                 LWLockRelease(&apw_state->lock);
 289                 ereport(LOG,
 290                                 (errmsg("skipping prewarm because block dump file is being written by PID %lu",
 291                                                 (unsigned long) apw_state->pid_using_dumpfile)));
 292                 return;
 293         }
 294         LWLockRelease(&apw_state->lock);
 295
 296         /*
 297          * Open the block dump file.  Exit quietly if it doesn't exist, but report
 298          * any other error.
 299          */
 300         file = AllocateFile(AUTOPREWARM_FILE, "r");
 301         if (!file)
 302         {
 303                 if (errno == ENOENT)
 304                 {
 305                         LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
 306                         apw_state->pid_using_dumpfile = InvalidPid;
 307                         LWLockRelease(&apw_state->lock);
 308                         return;                         /* No file to load. */
 309                 }
 310                 ereport(ERROR,
 311                                 (errcode_for_file_access(),
 312                                  errmsg("could not read file \"%s\": %m",
 313                                                 AUTOPREWARM_FILE)));
 314         }
 315
 316         /* First line of the file is a record count. */
 317         if (fscanf(file, "<<%d>>\n", &num_elements) != 1)
 318                 ereport(ERROR,
 319                                 (errcode_for_file_access(),
 320                                  errmsg("could not read from file \"%s\": %m",
 321                                                 AUTOPREWARM_FILE)));
 322
 323         /* Allocate a dynamic shared memory segment to store the record data. */
 324         seg = dsm_create(sizeof(BlockInfoRecord) * num_elements, 0);
 325         blkinfo = (BlockInfoRecord *) dsm_segment_address(seg);
 326
 327         /* Read records, one per line. */
 328         for (i = 0; i < num_elements; i++)
 329         {
 330                 unsigned        forknum;
 331
 332                 if (fscanf(file, "%u,%u,%u,%u,%u\n", &blkinfo[i].database,
 333                                    &blkinfo[i].tablespace, &blkinfo[i].filenode,
 334                                    &forknum, &blkinfo[i].blocknum) != 5)
 335                         ereport(ERROR,
 336                                         (errmsg("autoprewarm block dump file is corrupted at line %d",
 337                                                         i + 1)));
 338                 blkinfo[i].forknum = forknum;
 339         }
 340
 341         FreeFile(file);
 342
 343         /* Sort the blocks to be loaded. */
 344         pg_qsort(blkinfo, num_elements, sizeof(BlockInfoRecord),
 345                          apw_compare_blockinfo);
 346
 347         /* Populate shared memory state. */
 348         apw_state->block_info_handle = dsm_segment_handle(seg);
 349         apw_state->prewarm_start_idx = apw_state->prewarm_stop_idx = 0;
 350         apw_state->prewarmed_blocks = 0;
 351
 352         /* Get the info position of the first block of the next database. */
 353         while (apw_state->prewarm_start_idx < num_elements)
 354         {
 355                 int                     j = apw_state->prewarm_start_idx;
 356                 Oid                     current_db = blkinfo[j].database;
 357
 358                 /*
 359                  * Advance the prewarm_stop_idx to the first BlockRecordInfo that does
 360                  * not belong to this database.
 361                  */
 362                 j++;
 363                 while (j < num_elements)
 364                 {
 365                         if (current_db != blkinfo[j].database)
 366                         {
 367                                 /*
 368                                  * Combine BlockRecordInfos for global objects with those of
 369                                  * the database.
 370                                  */
 371                                 if (current_db != InvalidOid)
 372                                         break;
 373                                 current_db = blkinfo[j].database;
 374                         }
 375
 376                         j++;
 377                 }
 378
 379                 /*
 380                  * If we reach this point with current_db == InvalidOid, then only
 381                  * BlockRecordInfos belonging to global objects exist.  We can't
 382                  * prewarm without a database connection, so just bail out.
 383                  */
 384                 if (current_db == InvalidOid)
 385                         break;
 386
 387                 /* Configure stop point and database for next per-database worker. */
 388                 apw_state->prewarm_stop_idx = j;
 389                 apw_state->database = current_db;
 390                 Assert(apw_state->prewarm_start_idx < apw_state->prewarm_stop_idx);
 391
 392                 /* If we've run out of free buffers, don't launch another worker. */
 393                 if (!have_free_buffer())
 394                         break;
 395
 396                 /*
 397                  * Start a per-database worker to load blocks for this database; this
 398                  * function will return once the per-database worker exits.
 399                  */
 400                 apw_start_database_worker();
 401
 402                 /* Prepare for next database. */
 403                 apw_state->prewarm_start_idx = apw_state->prewarm_stop_idx;
 404         }
 405
 406         /* Clean up. */
 407         dsm_detach(seg);
 408         LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
 409         apw_state->block_info_handle = DSM_HANDLE_INVALID;
 410         apw_state->pid_using_dumpfile = InvalidPid;
 411         LWLockRelease(&apw_state->lock);
 412
 413         /* Report our success. */
 414         ereport(LOG,
 415                         (errmsg("autoprewarm successfully prewarmed %d of %d previously-loaded blocks",
 416                                         apw_state->prewarmed_blocks, num_elements)));
 417 }
 418
 419 /*
 420  * Prewarm all blocks for one database (and possibly also global objects, if
 421  * those got grouped with this database).
 422  */
 423 void
 424 autoprewarm_database_main(Datum main_arg)
 425 {
 426         int                     pos;
 427         BlockInfoRecord *block_info;
 428         Relation        rel = NULL;
 429         BlockNumber nblocks = 0;
 430         BlockInfoRecord *old_blk = NULL;
 431         dsm_segment *seg;
 432
 433         /* Establish signal handlers; once that's done, unblock signals. */
 434         pqsignal(SIGTERM, die);
 435         BackgroundWorkerUnblockSignals();
 436
 437         /* Connect to correct database and get block information. */
 438         apw_init_shmem();
 439         seg = dsm_attach(apw_state->block_info_handle);
 440         if (seg == NULL)
 441                 ereport(ERROR,
 442                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 443                                  errmsg("could not map dynamic shared memory segment")));
 444         BackgroundWorkerInitializeConnectionByOid(apw_state->database, InvalidOid, 0);
 445         block_info = (BlockInfoRecord *) dsm_segment_address(seg);
 446         pos = apw_state->prewarm_start_idx;
 447
 448         /*
 449          * Loop until we run out of blocks to prewarm or until we run out of free
 450          * buffers.
 451          */
 452         while (pos < apw_state->prewarm_stop_idx && have_free_buffer())
 453         {
 454                 BlockInfoRecord *blk = &block_info[pos++];
 455                 Buffer          buf;
 456
 457                 CHECK_FOR_INTERRUPTS();
 458
 459                 /*
 460                  * Quit if we've reached records for another database. If previous
 461                  * blocks are of some global objects, then continue pre-warming.
 462                  */
 463                 if (old_blk != NULL && old_blk->database != blk->database &&
 464                         old_blk->database != 0)
 465                         break;
 466
 467                 /*
 468                  * As soon as we encounter a block of a new relation, close the old
 469                  * relation. Note that rel will be NULL if try_relation_open failed
 470                  * previously; in that case, there is nothing to close.
 471                  */
 472                 if (old_blk != NULL && old_blk->filenode != blk->filenode &&
 473                         rel != NULL)
 474                 {
 475                         relation_close(rel, AccessShareLock);
 476                         rel = NULL;
 477                         CommitTransactionCommand();
 478                 }
 479
 480                 /*
 481                  * Try to open each new relation, but only once, when we first
 482                  * encounter it. If it's been dropped, skip the associated blocks.
 483                  */
 484                 if (old_blk == NULL || old_blk->filenode != blk->filenode)
 485                 {
 486                         Oid                     reloid;
 487
 488                         Assert(rel == NULL);
 489                         StartTransactionCommand();
 490                         reloid = RelidByRelfilenode(blk->tablespace, blk->filenode);
 491                         if (OidIsValid(reloid))
 492                                 rel = try_relation_open(reloid, AccessShareLock);
 493
 494                         if (!rel)
 495                                 CommitTransactionCommand();
 496                 }
 497                 if (!rel)
 498                 {
 499                         old_blk = blk;
 500                         continue;
 501                 }
 502
 503                 /* Once per fork, check for fork existence and size. */
 504                 if (old_blk == NULL ||
 505                         old_blk->filenode != blk->filenode ||
 506                         old_blk->forknum != blk->forknum)
 507                 {
 508                         RelationOpenSmgr(rel);
 509
 510                         /*
 511                          * smgrexists is not safe for illegal forknum, hence check whether
 512                          * the passed forknum is valid before using it in smgrexists.
 513                          */
 514                         if (blk->forknum > InvalidForkNumber &&
 515                                 blk->forknum <= MAX_FORKNUM &&
 516                                 smgrexists(rel->rd_smgr, blk->forknum))
 517                                 nblocks = RelationGetNumberOfBlocksInFork(rel, blk->forknum);
 518                         else
 519                                 nblocks = 0;
 520                 }
 521
 522                 /* Check whether blocknum is valid and within fork file size. */
 523                 if (blk->blocknum >= nblocks)
 524                 {
 525                         /* Move to next forknum. */
 526                         old_blk = blk;
 527                         continue;
 528                 }
 529
 530                 /* Prewarm buffer. */
 531                 buf = ReadBufferExtended(rel, blk->forknum, blk->blocknum, RBM_NORMAL,
 532                                                                  NULL);
 533                 if (BufferIsValid(buf))
 534                 {
 535                         apw_state->prewarmed_blocks++;
 536                         ReleaseBuffer(buf);
 537                 }
 538
 539                 old_blk = blk;
 540         }
 541
 542         dsm_detach(seg);
 543
 544         /* Release lock on previous relation. */
 545         if (rel)
 546         {
 547                 relation_close(rel, AccessShareLock);
 548                 CommitTransactionCommand();
 549         }
 550 }
 551
 552 /*
 553  * Dump information on blocks in shared buffers.  We use a text format here
 554  * so that it's easy to understand and even change the file contents if
 555  * necessary.
 556  * Returns the number of blocks dumped.
 557  */
 558 static int
 559 apw_dump_now(bool is_bgworker, bool dump_unlogged)
 560 {
 561         int                     num_blocks;
 562         int                     i;
 563         int                     ret;
 564         BlockInfoRecord *block_info_array;
 565         BufferDesc *bufHdr;
 566         FILE       *file;
 567         char            transient_dump_file_path[MAXPGPATH];
 568         pid_t           pid;
 569
 570         LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
 571         pid = apw_state->pid_using_dumpfile;
 572         if (apw_state->pid_using_dumpfile == InvalidPid)
 573                 apw_state->pid_using_dumpfile = MyProcPid;
 574         LWLockRelease(&apw_state->lock);
 575
 576         if (pid != InvalidPid)
 577         {
 578                 if (!is_bgworker)
 579                         ereport(ERROR,
 580                                         (errmsg("could not perform block dump because dump file is being used by PID %lu",
 581                                                         (unsigned long) apw_state->pid_using_dumpfile)));
 582
 583                 ereport(LOG,
 584                                 (errmsg("skipping block dump because it is already being performed by PID %lu",
 585                                                 (unsigned long) apw_state->pid_using_dumpfile)));
 586                 return 0;
 587         }
 588
 589         block_info_array =
 590                 (BlockInfoRecord *) palloc(sizeof(BlockInfoRecord) * NBuffers);
 591
 592         for (num_blocks = 0, i = 0; i < NBuffers; i++)
 593         {
 594                 uint32          buf_state;
 595
 596                 CHECK_FOR_INTERRUPTS();
 597
 598                 bufHdr = GetBufferDescriptor(i);
 599
 600                 /* Lock each buffer header before inspecting. */
 601                 buf_state = LockBufHdr(bufHdr);
 602
 603                 /*
 604                  * Unlogged tables will be automatically truncated after a crash or
 605                  * unclean shutdown. In such cases we need not prewarm them. Dump them
 606                  * only if requested by caller.
 607                  */
 608                 if (buf_state & BM_TAG_VALID &&
 609                         ((buf_state & BM_PERMANENT) || dump_unlogged))
 610                 {
 611                         block_info_array[num_blocks].database = bufHdr->tag.rnode.dbNode;
 612                         block_info_array[num_blocks].tablespace = bufHdr->tag.rnode.spcNode;
 613                         block_info_array[num_blocks].filenode = bufHdr->tag.rnode.relNode;
 614                         block_info_array[num_blocks].forknum = bufHdr->tag.forkNum;
 615                         block_info_array[num_blocks].blocknum = bufHdr->tag.blockNum;
 616                         ++num_blocks;
 617                 }
 618
 619                 UnlockBufHdr(bufHdr, buf_state);
 620         }
 621
 622         snprintf(transient_dump_file_path, MAXPGPATH, "%s.tmp", AUTOPREWARM_FILE);
 623         file = AllocateFile(transient_dump_file_path, "w");
 624         if (!file)
 625                 ereport(ERROR,
 626                                 (errcode_for_file_access(),
 627                                  errmsg("could not open file \"%s\": %m",
 628                                                 transient_dump_file_path)));
 629
 630         ret = fprintf(file, "<<%d>>\n", num_blocks);
 631         if (ret < 0)
 632         {
 633                 int                     save_errno = errno;
 634
 635                 FreeFile(file);
 636                 unlink(transient_dump_file_path);
 637                 errno = save_errno;
 638                 ereport(ERROR,
 639                                 (errcode_for_file_access(),
 640                                  errmsg("could not write to file \"%s\": %m",
 641                                                 transient_dump_file_path)));
 642         }
 643
 644         for (i = 0; i < num_blocks; i++)
 645         {
 646                 CHECK_FOR_INTERRUPTS();
 647
 648                 ret = fprintf(file, "%u,%u,%u,%u,%u\n",
 649                                           block_info_array[i].database,
 650                                           block_info_array[i].tablespace,
 651                                           block_info_array[i].filenode,
 652                                           (uint32) block_info_array[i].forknum,
 653                                           block_info_array[i].blocknum);
 654                 if (ret < 0)
 655                 {
 656                         int                     save_errno = errno;
 657
 658                         FreeFile(file);
 659                         unlink(transient_dump_file_path);
 660                         errno = save_errno;
 661                         ereport(ERROR,
 662                                         (errcode_for_file_access(),
 663                                          errmsg("could not write to file \"%s\": %m",
 664                                                         transient_dump_file_path)));
 665                 }
 666         }
 667
 668         pfree(block_info_array);
 669
 670         /*
 671          * Rename transient_dump_file_path to AUTOPREWARM_FILE to make things
 672          * permanent.
 673          */
 674         ret = FreeFile(file);
 675         if (ret != 0)
 676         {
 677                 int                     save_errno = errno;
 678
 679                 unlink(transient_dump_file_path);
 680                 errno = save_errno;
 681                 ereport(ERROR,
 682                                 (errcode_for_file_access(),
 683                                  errmsg("could not close file \"%s\": %m",
 684                                                 transient_dump_file_path)));
 685         }
 686
 687         (void) durable_rename(transient_dump_file_path, AUTOPREWARM_FILE, ERROR);
 688         apw_state->pid_using_dumpfile = InvalidPid;
 689
 690         ereport(DEBUG1,
 691                         (errmsg("wrote block details for %d blocks", num_blocks)));
 692         return num_blocks;
 693 }
 694
 695 /*
 696  * SQL-callable function to launch autoprewarm.
 697  */
 698 Datum
 699 autoprewarm_start_worker(PG_FUNCTION_ARGS)
 700 {
 701         pid_t           pid;
 702
 703         if (!autoprewarm)
 704                 ereport(ERROR,
 705                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 706                                  errmsg("autoprewarm is disabled")));
 707
 708         apw_init_shmem();
 709         LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
 710         pid = apw_state->bgworker_pid;
 711         LWLockRelease(&apw_state->lock);
 712
 713         if (pid != InvalidPid)
 714                 ereport(ERROR,
 715                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
 716                                  errmsg("autoprewarm worker is already running under PID %lu",
 717                                                 (unsigned long) pid)));
 718
 719         apw_start_master_worker();
 720
 721         PG_RETURN_VOID();
 722 }
 723
 724 /*
 725  * SQL-callable function to perform an immediate block dump.
 726  *
 727  * Note: this is declared to return int8, as insurance against some
 728  * very distant day when we might make NBuffers wider than int.
 729  */
 730 Datum
 731 autoprewarm_dump_now(PG_FUNCTION_ARGS)
 732 {
 733         int                     num_blocks;
 734
 735         apw_init_shmem();
 736
 737         PG_ENSURE_ERROR_CLEANUP(apw_detach_shmem, 0);
 738         {
 739                 num_blocks = apw_dump_now(false, true);
 740         }
 741         PG_END_ENSURE_ERROR_CLEANUP(apw_detach_shmem, 0);
 742
 743         PG_RETURN_INT64((int64) num_blocks);
 744 }
 745
 746 /*
 747  * Allocate and initialize autoprewarm related shared memory, if not already
 748  * done, and set up backend-local pointer to that state.  Returns true if an
 749  * existing shared memory segment was found.
 750  */
 751 static bool
 752 apw_init_shmem(void)
 753 {
 754         bool            found;
 755
 756         LWLockAcquire(AddinShmemInitLock, LW_EXCLUSIVE);
 757         apw_state = ShmemInitStruct("autoprewarm",
 758                                                                 sizeof(AutoPrewarmSharedState),
 759                                                                 &found);
 760         if (!found)
 761         {
 762                 /* First time through ... */
 763                 LWLockInitialize(&apw_state->lock, LWLockNewTrancheId());
 764                 apw_state->bgworker_pid = InvalidPid;
 765                 apw_state->pid_using_dumpfile = InvalidPid;
 766         }
 767         LWLockRelease(AddinShmemInitLock);
 768
 769         LWLockRegisterTranche(apw_state->lock.tranche, "autoprewarm");
 770
 771         return found;
 772 }
 773
 774 /*
 775  * Clear our PID from autoprewarm shared state.
 776  */
 777 static void
 778 apw_detach_shmem(int code, Datum arg)
 779 {
 780         LWLockAcquire(&apw_state->lock, LW_EXCLUSIVE);
 781         if (apw_state->pid_using_dumpfile == MyProcPid)
 782                 apw_state->pid_using_dumpfile = InvalidPid;
 783         if (apw_state->bgworker_pid == MyProcPid)
 784                 apw_state->bgworker_pid = InvalidPid;
 785         LWLockRelease(&apw_state->lock);
 786 }
 787
 788 /*
 789  * Start autoprewarm master worker process.
 790  */
 791 static void
 792 apw_start_master_worker(void)
 793 {
 794         BackgroundWorker worker;
 795         BackgroundWorkerHandle *handle;
 796         BgwHandleStatus status;
 797         pid_t           pid;
 798
 799         MemSet(&worker, 0, sizeof(BackgroundWorker));
 800         worker.bgw_flags = BGWORKER_SHMEM_ACCESS;
 801         worker.bgw_start_time = BgWorkerStart_ConsistentState;
 802         strcpy(worker.bgw_library_name, "pg_prewarm");
 803         strcpy(worker.bgw_function_name, "autoprewarm_main");
 804         strcpy(worker.bgw_name, "autoprewarm master");
 805         strcpy(worker.bgw_type, "autoprewarm master");
 806
 807         if (process_shared_preload_libraries_in_progress)
 808         {
 809                 RegisterBackgroundWorker(&worker);
 810                 return;
 811         }
 812
 813         /* must set notify PID to wait for startup */
 814         worker.bgw_notify_pid = MyProcPid;
 815
 816         if (!RegisterDynamicBackgroundWorker(&worker, &handle))
 817                 ereport(ERROR,
 818                                 (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
 819                                  errmsg("could not register background process"),
 820                                  errhint("You may need to increase max_worker_processes.")));
 821
 822         status = WaitForBackgroundWorkerStartup(handle, &pid);
 823         if (status != BGWH_STARTED)
 824                 ereport(ERROR,
 825                                 (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
 826                                  errmsg("could not start background process"),
 827                                  errhint("More details may be available in the server log.")));
 828 }
 829
 830 /*
 831  * Start autoprewarm per-database worker process.
 832  */
 833 static void
 834 apw_start_database_worker(void)
 835 {
 836         BackgroundWorker worker;
 837         BackgroundWorkerHandle *handle;
 838
 839         MemSet(&worker, 0, sizeof(BackgroundWorker));
 840         worker.bgw_flags =
 841                 BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION;
 842         worker.bgw_start_time = BgWorkerStart_ConsistentState;
 843         strcpy(worker.bgw_library_name, "pg_prewarm");
 844         strcpy(worker.bgw_function_name, "autoprewarm_database_main");
 845         strcpy(worker.bgw_name, "autoprewarm worker");
 846         strcpy(worker.bgw_type, "autoprewarm worker");
 847
 848         /* must set notify PID to wait for shutdown */
 849         worker.bgw_notify_pid = MyProcPid;
 850
 851         if (!RegisterDynamicBackgroundWorker(&worker, &handle))
 852                 ereport(ERROR,
 853                                 (errcode(ERRCODE_INSUFFICIENT_RESOURCES),
 854                                  errmsg("registering dynamic bgworker autoprewarm failed"),
 855                                  errhint("Consider increasing configuration parameter \"max_worker_processes\".")));
 856
 857         /*
 858          * Ignore return value; if it fails, postmaster has died, but we have
 859          * checks for that elsewhere.
 860          */
 861         WaitForBackgroundWorkerShutdown(handle);
 862 }
 863
 864 /* Compare member elements to check whether they are not equal. */
 865 #define cmp_member_elem(fld)    \
 866 do { \
 867         if (a->fld < b->fld)            \
 868                 return -1;                              \
 869         else if (a->fld > b->fld)       \
 870                 return 1;                               \
 871 } while(0)
 872
 873 /*
 874  * apw_compare_blockinfo
 875  *
 876  * We depend on all records for a particular database being consecutive
 877  * in the dump file; each per-database worker will preload blocks until
 878  * it sees a block for some other database.  Sorting by tablespace,
 879  * filenode, forknum, and blocknum isn't critical for correctness, but
 880  * helps us get a sequential I/O pattern.
 881  */
 882 static int
 883 apw_compare_blockinfo(const void *p, const void *q)
 884 {
 885         const BlockInfoRecord *a = (const BlockInfoRecord *) p;
 886         const BlockInfoRecord *b = (const BlockInfoRecord *) q;
 887
 888         cmp_member_elem(database);
 889         cmp_member_elem(tablespace);
 890         cmp_member_elem(filenode);
 891         cmp_member_elem(forknum);
 892         cmp_member_elem(blocknum);
 893
 894         return 0;
 895 }
 896
 897 /*
 898  * Signal handler for SIGTERM
 899  */
 900 static void
 901 apw_sigterm_handler(SIGNAL_ARGS)
 902 {
 903         int                     save_errno = errno;
 904
 905         got_sigterm = true;
 906
 907         if (MyProc)
 908                 SetLatch(&MyProc->procLatch);
 909
 910         errno = save_errno;
 911 }
 912
 913 /*
 914  * Signal handler for SIGHUP
 915  */
 916 static void
 917 apw_sighup_handler(SIGNAL_ARGS)
 918 {
 919         int                     save_errno = errno;
 920
 921         got_sighup = true;
 922
 923         if (MyProc)
 924                 SetLatch(&MyProc->procLatch);
 925
 926         errno = save_errno;
 927 }