pg_stat_statements.c

Go to the documentation of this file.

/*-------------------------------------------------------------------------
 *
 * pg_stat_statements.c
 *      Track statement execution times across a whole database cluster.
 *
 * Execution costs are totalled for each distinct source query, and kept in
 * a shared hashtable.  (We track only as many distinct queries as will fit
 * in the designated amount of shared memory.)
 *
 * As of Postgres 9.2, this module normalizes query entries.  Normalization
 * is a process whereby similar queries, typically differing only in their
 * constants (though the exact rules are somewhat more subtle than that) are
 * recognized as equivalent, and are tracked as a single entry.  This is
 * particularly useful for non-prepared queries.
 *
 * Normalization is implemented by fingerprinting queries, selectively
 * serializing those fields of each query tree's nodes that are judged to be
 * essential to the query.  This is referred to as a query jumble.  This is
 * distinct from a regular serialization in that various extraneous
 * information is ignored as irrelevant or not essential to the query, such
 * as the collations of Vars and, most notably, the values of constants.
 *
 * This jumble is acquired at the end of parse analysis of each query, and
 * a 64-bit hash of it is stored into the query's Query.queryId field.
 * The server then copies this value around, making it available in plan
 * tree(s) generated from the query.  The executor can then use this value
 * to blame query costs on the proper queryId.
 *
 * To facilitate presenting entries to users, we create "representative" query
 * strings in which constants are replaced with parameter symbols ($n), to
 * make it clearer what a normalized entry can represent.  To save on shared
 * memory, and to avoid having to truncate oversized query strings, we store
 * these strings in a temporary external query-texts file.  Offsets into this
 * file are kept in shared memory.
 *
 * Note about locking issues: to create or delete an entry in the shared
 * hashtable, one must hold pgss->lock exclusively.  Modifying any field
 * in an entry except the counters requires the same.  To look up an entry,
 * one must hold the lock shared.  To read or update the counters within
 * an entry, one must hold the lock shared or exclusive (so the entry doesn't
 * disappear!) and also take the entry's mutex spinlock.
 * The shared state variable pgss->extent (the next free spot in the external
 * query-text file) should be accessed only while holding either the
 * pgss->mutex spinlock, or exclusive lock on pgss->lock.  We use the mutex to
 * allow reserving file space while holding only shared lock on pgss->lock.
 * Rewriting the entire external query-text file, eg for garbage collection,
 * requires holding pgss->lock exclusively; this allows individual entries
 * in the file to be read or written while holding only shared lock.
 *
 *
 * Copyright (c) 2008-2018, PostgreSQL Global Development Group
 *
 * IDENTIFICATION
 *    contrib/pg_stat_statements/pg_stat_statements.c
 *
 *-------------------------------------------------------------------------
 */
#include "postgres.h"

#include <math.h>
#include <sys/stat.h>
#include <unistd.h>

#include "access/hash.h"
#include "catalog/pg_authid.h"
#include "executor/instrument.h"
#include "funcapi.h"
#include "mb/pg_wchar.h"
#include "miscadmin.h"
#include "parser/analyze.h"
#include "parser/parsetree.h"
#include "parser/scanner.h"
#include "parser/scansup.h"
#include "pgstat.h"
#include "storage/fd.h"
#include "storage/ipc.h"
#include "storage/spin.h"
#include "tcop/utility.h"
#include "utils/acl.h"
#include "utils/builtins.h"
#include "utils/memutils.h"

PG_MODULE_MAGIC;

/* Location of permanent stats file (valid when database is shut down) */
#define PGSS_DUMP_FILE  PGSTAT_STAT_PERMANENT_DIRECTORY "/pg_stat_statements.stat"

/*
 * Location of external query text file.  We don't keep it in the core
 * system's stats_temp_directory.  The core system can safely use that GUC
 * setting, because the statistics collector temp file paths are set only once
 * as part of changing the GUC, but pg_stat_statements has no way of avoiding
 * race conditions.  Besides, we only expect modest, infrequent I/O for query
 * strings, so placing the file on a faster filesystem is not compelling.
 */
#define PGSS_TEXT_FILE  PG_STAT_TMP_DIR "/pgss_query_texts.stat"

/* Magic number identifying the stats file format */
static const uint32 PGSS_FILE_HEADER = 0x20171004;

/* PostgreSQL major version number, changes in which invalidate all entries */
static const uint32 PGSS_PG_MAJOR_VERSION = PG_VERSION_NUM / 100;

/* XXX: Should USAGE_EXEC reflect execution time and/or buffer usage? */
#define USAGE_EXEC(duration)    (1.0)
#define USAGE_INIT              (1.0)   /* including initial planning */
#define ASSUMED_MEDIAN_INIT     (10.0)  /* initial assumed median usage */
#define ASSUMED_LENGTH_INIT     1024    /* initial assumed mean query length */
#define USAGE_DECREASE_FACTOR   (0.99)  /* decreased every entry_dealloc */
#define STICKY_DECREASE_FACTOR  (0.50)  /* factor for sticky entries */
#define USAGE_DEALLOC_PERCENT   5   /* free this % of entries at once */

#define JUMBLE_SIZE             1024    /* query serialization buffer size */

/*
 * Extension version number, for supporting older extension versions' objects
 */
typedef enum pgssVersion
{
    PGSS_V1_0 = 0,
    PGSS_V1_1,
    PGSS_V1_2,
    PGSS_V1_3
} pgssVersion;

/*
 * Hashtable key that defines the identity of a hashtable entry.  We separate
 * queries by user and by database even if they are otherwise identical.
 *
 * Right now, this structure contains no padding.  If you add any, make sure
 * to teach pgss_store() to zero the padding bytes.  Otherwise, things will
 * break, because pgss_hash is created using HASH_BLOBS, and thus tag_hash
 * is used to hash this.
 */
typedef struct pgssHashKey
{
    Oid         userid;         /* user OID */
    Oid         dbid;           /* database OID */
    uint64      queryid;        /* query identifier */
} pgssHashKey;

/*
 * The actual stats counters kept within pgssEntry.
 */
typedef struct Counters
{
    int64       calls;          /* # of times executed */
    double      total_time;     /* total execution time, in msec */
    double      min_time;       /* minimum execution time in msec */
    double      max_time;       /* maximum execution time in msec */
    double      mean_time;      /* mean execution time in msec */
    double      sum_var_time;   /* sum of variances in execution time in msec */
    int64       rows;           /* total # of retrieved or affected rows */
    int64       shared_blks_hit;    /* # of shared buffer hits */
    int64       shared_blks_read;   /* # of shared disk blocks read */
    int64       shared_blks_dirtied;    /* # of shared disk blocks dirtied */
    int64       shared_blks_written;    /* # of shared disk blocks written */
    int64       local_blks_hit; /* # of local buffer hits */
    int64       local_blks_read;    /* # of local disk blocks read */
    int64       local_blks_dirtied; /* # of local disk blocks dirtied */
    int64       local_blks_written; /* # of local disk blocks written */
    int64       temp_blks_read; /* # of temp blocks read */
    int64       temp_blks_written;  /* # of temp blocks written */
    double      blk_read_time;  /* time spent reading, in msec */
    double      blk_write_time; /* time spent writing, in msec */
    double      usage;          /* usage factor */
} Counters;

/*
 * Statistics per statement
 *
 * Note: in event of a failure in garbage collection of the query text file,
 * we reset query_offset to zero and query_len to -1.  This will be seen as
 * an invalid state by qtext_fetch().
 */
typedef struct pgssEntry
{
    pgssHashKey key;            /* hash key of entry - MUST BE FIRST */
    Counters    counters;       /* the statistics for this query */
    Size        query_offset;   /* query text offset in external file */
    int         query_len;      /* # of valid bytes in query string, or -1 */
    int         encoding;       /* query text encoding */
    slock_t     mutex;          /* protects the counters only */
} pgssEntry;

/*
 * Global shared state
 */
typedef struct pgssSharedState
{
    LWLock     *lock;           /* protects hashtable search/modification */
    double      cur_median_usage;   /* current median usage in hashtable */
    Size        mean_query_len; /* current mean entry text length */
    slock_t     mutex;          /* protects following fields only: */
    Size        extent;         /* current extent of query file */
    int         n_writers;      /* number of active writers to query file */
    int         gc_count;       /* query file garbage collection cycle count */
} pgssSharedState;

/*
 * Struct for tracking locations/lengths of constants during normalization
 */
typedef struct pgssLocationLen
{
    int         location;       /* start offset in query text */
    int         length;         /* length in bytes, or -1 to ignore */
} pgssLocationLen;

/*
 * Working state for computing a query jumble and producing a normalized
 * query string
 */
typedef struct pgssJumbleState
{
    /* Jumble of current query tree */
    unsigned char *jumble;

    /* Number of bytes used in jumble[] */
    Size        jumble_len;

    /* Array of locations of constants that should be removed */
    pgssLocationLen *clocations;

    /* Allocated length of clocations array */
    int         clocations_buf_size;

    /* Current number of valid entries in clocations array */
    int         clocations_count;

    /* highest Param id we've seen, in order to start normalization correctly */
    int         highest_extern_param_id;
} pgssJumbleState;

/*---- Local variables ----*/

/* Current nesting depth of ExecutorRun+ProcessUtility calls */
static int  nested_level = 0;

/* Saved hook values in case of unload */
static shmem_startup_hook_type prev_shmem_startup_hook = NULL;
static post_parse_analyze_hook_type prev_post_parse_analyze_hook = NULL;
static ExecutorStart_hook_type prev_ExecutorStart = NULL;
static ExecutorRun_hook_type prev_ExecutorRun = NULL;
static ExecutorFinish_hook_type prev_ExecutorFinish = NULL;
static ExecutorEnd_hook_type prev_ExecutorEnd = NULL;
static ProcessUtility_hook_type prev_ProcessUtility = NULL;

/* Links to shared memory state */
static pgssSharedState *pgss = NULL;
static HTAB *pgss_hash = NULL;

/*---- GUC variables ----*/

typedef enum
{
    PGSS_TRACK_NONE,            /* track no statements */
    PGSS_TRACK_TOP,             /* only top level statements */
    PGSS_TRACK_ALL              /* all statements, including nested ones */
}           PGSSTrackLevel;

static const struct config_enum_entry track_options[] =
{
    {"none", PGSS_TRACK_NONE, false},
    {"top", PGSS_TRACK_TOP, false},
    {"all", PGSS_TRACK_ALL, false},
    {NULL, 0, false}
};

static int  pgss_max;           /* max # statements to track */
static int  pgss_track;         /* tracking level */
static bool pgss_track_utility; /* whether to track utility commands */
static bool pgss_save;          /* whether to save stats across shutdown */


#define pgss_enabled() \
    (pgss_track == PGSS_TRACK_ALL || \
    (pgss_track == PGSS_TRACK_TOP && nested_level == 0))

#define record_gc_qtexts() \
    do { \
        volatile pgssSharedState *s = (volatile pgssSharedState *) pgss; \
        SpinLockAcquire(&s->mutex); \
        s->gc_count++; \
        SpinLockRelease(&s->mutex); \
    } while(0)

/*---- Function declarations ----*/

void        _PG_init(void);
void        _PG_fini(void);

PG_FUNCTION_INFO_V1(pg_stat_statements_reset);
PG_FUNCTION_INFO_V1(pg_stat_statements_1_2);
PG_FUNCTION_INFO_V1(pg_stat_statements_1_3);
PG_FUNCTION_INFO_V1(pg_stat_statements);

static void pgss_shmem_startup(void);
static void pgss_shmem_shutdown(int code, Datum arg);
static void pgss_post_parse_analyze(ParseState *pstate, Query *query);
static void pgss_ExecutorStart(QueryDesc *queryDesc, int eflags);
static void pgss_ExecutorRun(QueryDesc *queryDesc,
                 ScanDirection direction,
                 uint64 count, bool execute_once);
static void pgss_ExecutorFinish(QueryDesc *queryDesc);
static void pgss_ExecutorEnd(QueryDesc *queryDesc);
static void pgss_ProcessUtility(PlannedStmt *pstmt, const char *queryString,
                    ProcessUtilityContext context, ParamListInfo params,
                    QueryEnvironment *queryEnv,
                    DestReceiver *dest, char *completionTag);
static uint64 pgss_hash_string(const char *str, int len);
static void pgss_store(const char *query, uint64 queryId,
           int query_location, int query_len,
           double total_time, uint64 rows,
           const BufferUsage *bufusage,
           pgssJumbleState *jstate);
static void pg_stat_statements_internal(FunctionCallInfo fcinfo,
                            pgssVersion api_version,
                            bool showtext);
static Size pgss_memsize(void);
static pgssEntry *entry_alloc(pgssHashKey *key, Size query_offset, int query_len,
            int encoding, bool sticky);
static void entry_dealloc(void);
static bool qtext_store(const char *query, int query_len,
            Size *query_offset, int *gc_count);
static char *qtext_load_file(Size *buffer_size);
static char *qtext_fetch(Size query_offset, int query_len,
            char *buffer, Size buffer_size);
static bool need_gc_qtexts(void);
static void gc_qtexts(void);
static void entry_reset(void);
static void AppendJumble(pgssJumbleState *jstate,
             const unsigned char *item, Size size);
static void JumbleQuery(pgssJumbleState *jstate, Query *query);
static void JumbleRangeTable(pgssJumbleState *jstate, List *rtable);
static void JumbleExpr(pgssJumbleState *jstate, Node *node);
static void RecordConstLocation(pgssJumbleState *jstate, int location);
static char *generate_normalized_query(pgssJumbleState *jstate, const char *query,
                          int query_loc, int *query_len_p, int encoding);
static void fill_in_constant_lengths(pgssJumbleState *jstate, const char *query,
                         int query_loc);
static int  comp_location(const void *a, const void *b);


/*
 * Module load callback
 */
void
_PG_init(void)
{
    /*
     * In order to create our shared memory area, we have to be loaded via
     * shared_preload_libraries.  If not, fall out without hooking into any of
     * the main system.  (We don't throw error here because it seems useful to
     * allow the pg_stat_statements functions to be created even when the
     * module isn't active.  The functions must protect themselves against
     * being called then, however.)
     */
    if (!process_shared_preload_libraries_in_progress)
        return;

    /*
     * Define (or redefine) custom GUC variables.
     */
    DefineCustomIntVariable("pg_stat_statements.max",
                            "Sets the maximum number of statements tracked by pg_stat_statements.",
                            NULL,
                            &pgss_max,
                            5000,
                            100,
                            INT_MAX,
                            PGC_POSTMASTER,
                            0,
                            NULL,
                            NULL,
                            NULL);

    DefineCustomEnumVariable("pg_stat_statements.track",
                             "Selects which statements are tracked by pg_stat_statements.",
                             NULL,
                             &pgss_track,
                             PGSS_TRACK_TOP,
                             track_options,
                             PGC_SUSET,
                             0,
                             NULL,
                             NULL,
                             NULL);

    DefineCustomBoolVariable("pg_stat_statements.track_utility",
                             "Selects whether utility commands are tracked by pg_stat_statements.",
                             NULL,
                             &pgss_track_utility,
                             true,
                             PGC_SUSET,
                             0,
                             NULL,
                             NULL,
                             NULL);

    DefineCustomBoolVariable("pg_stat_statements.save",
                             "Save pg_stat_statements statistics across server shutdowns.",
                             NULL,
                             &pgss_save,
                             true,
                             PGC_SIGHUP,
                             0,
                             NULL,
                             NULL,
                             NULL);

    EmitWarningsOnPlaceholders("pg_stat_statements");

    /*
     * Request additional shared resources.  (These are no-ops if we're not in
     * the postmaster process.)  We'll allocate or attach to the shared
     * resources in pgss_shmem_startup().
     */
    RequestAddinShmemSpace(pgss_memsize());
    RequestNamedLWLockTranche("pg_stat_statements", 1);

    /*
     * Install hooks.
     */
    prev_shmem_startup_hook = shmem_startup_hook;
    shmem_startup_hook = pgss_shmem_startup;
    prev_post_parse_analyze_hook = post_parse_analyze_hook;
    post_parse_analyze_hook = pgss_post_parse_analyze;
    prev_ExecutorStart = ExecutorStart_hook;
    ExecutorStart_hook = pgss_ExecutorStart;
    prev_ExecutorRun = ExecutorRun_hook;
    ExecutorRun_hook = pgss_ExecutorRun;
    prev_ExecutorFinish = ExecutorFinish_hook;
    ExecutorFinish_hook = pgss_ExecutorFinish;
    prev_ExecutorEnd = ExecutorEnd_hook;
    ExecutorEnd_hook = pgss_ExecutorEnd;
    prev_ProcessUtility = ProcessUtility_hook;
    ProcessUtility_hook = pgss_ProcessUtility;
}

/*
 * Module unload callback
 */
void
_PG_fini(void)
{
    /* Uninstall hooks. */
    shmem_startup_hook = prev_shmem_startup_hook;
    post_parse_analyze_hook = prev_post_parse_analyze_hook;
    ExecutorStart_hook = prev_ExecutorStart;
    ExecutorRun_hook = prev_ExecutorRun;
    ExecutorFinish_hook = prev_ExecutorFinish;
    ExecutorEnd_hook = prev_ExecutorEnd;
    ProcessUtility_hook = prev_ProcessUtility;
}

/*
 * shmem_startup hook: allocate or attach to shared memory,
 * then load any pre-existing statistics from file.
 * Also create and load the query-texts file, which is expected to exist
 * (even if empty) while the module is enabled.
 */
static void
pgss_shmem_startup(void)
{
    bool        found;
    HASHCTL     info;
    FILE       *file = NULL;
    FILE       *qfile = NULL;
    uint32      header;
    int32       num;
    int32       pgver;
    int32       i;
    int         buffer_size;
    char       *buffer = NULL;

    if (prev_shmem_startup_hook)
        prev_shmem_startup_hook();

    /* reset in case this is a restart within the postmaster */
    pgss = NULL;
    pgss_hash = NULL;

    /*
     * Create or attach to the shared memory state, including hash table
     */
    LWLockAcquire(AddinShmemInitLock, LW_EXCLUSIVE);

    pgss = ShmemInitStruct("pg_stat_statements",
                           sizeof(pgssSharedState),
                           &found);

    if (!found)
    {
        /* First time through ... */
        pgss->lock = &(GetNamedLWLockTranche("pg_stat_statements"))->lock;
        pgss->cur_median_usage = ASSUMED_MEDIAN_INIT;
        pgss->mean_query_len = ASSUMED_LENGTH_INIT;
        SpinLockInit(&pgss->mutex);
        pgss->extent = 0;
        pgss->n_writers = 0;
        pgss->gc_count = 0;
    }

    memset(&info, 0, sizeof(info));
    info.keysize = sizeof(pgssHashKey);
    info.entrysize = sizeof(pgssEntry);
    pgss_hash = ShmemInitHash("pg_stat_statements hash",
                              pgss_max, pgss_max,
                              &info,
                              HASH_ELEM | HASH_BLOBS);

    LWLockRelease(AddinShmemInitLock);

    /*
     * If we're in the postmaster (or a standalone backend...), set up a shmem
     * exit hook to dump the statistics to disk.
     */
    if (!IsUnderPostmaster)
        on_shmem_exit(pgss_shmem_shutdown, (Datum) 0);

    /*
     * Done if some other process already completed our initialization.
     */
    if (found)
        return;

    /*
     * Note: we don't bother with locks here, because there should be no other
     * processes running when this code is reached.
     */

    /* Unlink query text file possibly left over from crash */
    unlink(PGSS_TEXT_FILE);

    /* Allocate new query text temp file */
    qfile = AllocateFile(PGSS_TEXT_FILE, PG_BINARY_W);
    if (qfile == NULL)
        goto write_error;

    /*
     * If we were told not to load old statistics, we're done.  (Note we do
     * not try to unlink any old dump file in this case.  This seems a bit
     * questionable but it's the historical behavior.)
     */
    if (!pgss_save)
    {
        FreeFile(qfile);
        return;
    }

    /*
     * Attempt to load old statistics from the dump file.
     */
    file = AllocateFile(PGSS_DUMP_FILE, PG_BINARY_R);
    if (file == NULL)
    {
        if (errno != ENOENT)
            goto read_error;
        /* No existing persisted stats file, so we're done */
        FreeFile(qfile);
        return;
    }

    buffer_size = 2048;
    buffer = (char *) palloc(buffer_size);

    if (fread(&header, sizeof(uint32), 1, file) != 1 ||
        fread(&pgver, sizeof(uint32), 1, file) != 1 ||
        fread(&num, sizeof(int32), 1, file) != 1)
        goto read_error;

    if (header != PGSS_FILE_HEADER ||
        pgver != PGSS_PG_MAJOR_VERSION)
        goto data_error;

    for (i = 0; i < num; i++)
    {
        pgssEntry   temp;
        pgssEntry  *entry;
        Size        query_offset;

        if (fread(&temp, sizeof(pgssEntry), 1, file) != 1)
            goto read_error;

        /* Encoding is the only field we can easily sanity-check */
        if (!PG_VALID_BE_ENCODING(temp.encoding))
            goto data_error;

        /* Resize buffer as needed */
        if (temp.query_len >= buffer_size)
        {
            buffer_size = Max(buffer_size * 2, temp.query_len + 1);
            buffer = repalloc(buffer, buffer_size);
        }

        if (fread(buffer, 1, temp.query_len + 1, file) != temp.query_len + 1)
            goto read_error;

        /* Should have a trailing null, but let's make sure */
        buffer[temp.query_len] = '\0';

        /* Skip loading "sticky" entries */
        if (temp.counters.calls == 0)
            continue;

        /* Store the query text */
        query_offset = pgss->extent;
        if (fwrite(buffer, 1, temp.query_len + 1, qfile) != temp.query_len + 1)
            goto write_error;
        pgss->extent += temp.query_len + 1;

        /* make the hashtable entry (discards old entries if too many) */
        entry = entry_alloc(&temp.key, query_offset, temp.query_len,
                            temp.encoding,
                            false);

        /* copy in the actual stats */
        entry->counters = temp.counters;
    }

    pfree(buffer);
    FreeFile(file);
    FreeFile(qfile);

    /*
     * Remove the persisted stats file so it's not included in
     * backups/replication slaves, etc.  A new file will be written on next
     * shutdown.
     *
     * Note: it's okay if the PGSS_TEXT_FILE is included in a basebackup,
     * because we remove that file on startup; it acts inversely to
     * PGSS_DUMP_FILE, in that it is only supposed to be around when the
     * server is running, whereas PGSS_DUMP_FILE is only supposed to be around
     * when the server is not running.  Leaving the file creates no danger of
     * a newly restored database having a spurious record of execution costs,
     * which is what we're really concerned about here.
     */
    unlink(PGSS_DUMP_FILE);

    return;

read_error:
    ereport(LOG,
            (errcode_for_file_access(),
             errmsg("could not read pg_stat_statement file \"%s\": %m",
                    PGSS_DUMP_FILE)));
    goto fail;
data_error:
    ereport(LOG,
            (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
             errmsg("ignoring invalid data in pg_stat_statement file \"%s\"",
                    PGSS_DUMP_FILE)));
    goto fail;
write_error:
    ereport(LOG,
            (errcode_for_file_access(),
             errmsg("could not write pg_stat_statement file \"%s\": %m",
                    PGSS_TEXT_FILE)));
fail:
    if (buffer)
        pfree(buffer);
    if (file)
        FreeFile(file);
    if (qfile)
        FreeFile(qfile);
    /* If possible, throw away the bogus file; ignore any error */
    unlink(PGSS_DUMP_FILE);

    /*
     * Don't unlink PGSS_TEXT_FILE here; it should always be around while the
     * server is running with pg_stat_statements enabled
     */
}

/*
 * shmem_shutdown hook: Dump statistics into file.
 *
 * Note: we don't bother with acquiring lock, because there should be no
 * other processes running when this is called.
 */
static void
pgss_shmem_shutdown(int code, Datum arg)
{
    FILE       *file;
    char       *qbuffer = NULL;
    Size        qbuffer_size = 0;
    HASH_SEQ_STATUS hash_seq;
    int32       num_entries;
    pgssEntry  *entry;

    /* Don't try to dump during a crash. */
    if (code)
        return;

    /* Safety check ... shouldn't get here unless shmem is set up. */
    if (!pgss || !pgss_hash)
        return;

    /* Don't dump if told not to. */
    if (!pgss_save)
        return;

    file = AllocateFile(PGSS_DUMP_FILE ".tmp", PG_BINARY_W);
    if (file == NULL)
        goto error;

    if (fwrite(&PGSS_FILE_HEADER, sizeof(uint32), 1, file) != 1)
        goto error;
    if (fwrite(&PGSS_PG_MAJOR_VERSION, sizeof(uint32), 1, file) != 1)
        goto error;
    num_entries = hash_get_num_entries(pgss_hash);
    if (fwrite(&num_entries, sizeof(int32), 1, file) != 1)
        goto error;

    qbuffer = qtext_load_file(&qbuffer_size);
    if (qbuffer == NULL)
        goto error;

    /*
     * When serializing to disk, we store query texts immediately after their
     * entry data.  Any orphaned query texts are thereby excluded.
     */
    hash_seq_init(&hash_seq, pgss_hash);
    while ((entry = hash_seq_search(&hash_seq)) != NULL)
    {
        int         len = entry->query_len;
        char       *qstr = qtext_fetch(entry->query_offset, len,
                                       qbuffer, qbuffer_size);

        if (qstr == NULL)
            continue;           /* Ignore any entries with bogus texts */

        if (fwrite(entry, sizeof(pgssEntry), 1, file) != 1 ||
            fwrite(qstr, 1, len + 1, file) != len + 1)
        {
            /* note: we assume hash_seq_term won't change errno */
            hash_seq_term(&hash_seq);
            goto error;
        }
    }

    free(qbuffer);
    qbuffer = NULL;

    if (FreeFile(file))
    {
        file = NULL;
        goto error;
    }

    /*
     * Rename file into place, so we atomically replace any old one.
     */
    (void) durable_rename(PGSS_DUMP_FILE ".tmp", PGSS_DUMP_FILE, LOG);

    /* Unlink query-texts file; it's not needed while shutdown */
    unlink(PGSS_TEXT_FILE);

    return;

error:
    ereport(LOG,
            (errcode_for_file_access(),
             errmsg("could not write pg_stat_statement file \"%s\": %m",
                    PGSS_DUMP_FILE ".tmp")));
    if (qbuffer)
        free(qbuffer);
    if (file)
        FreeFile(file);
    unlink(PGSS_DUMP_FILE ".tmp");
    unlink(PGSS_TEXT_FILE);
}

/*
 * Post-parse-analysis hook: mark query with a queryId
 */
static void
pgss_post_parse_analyze(ParseState *pstate, Query *query)
{
    pgssJumbleState jstate;

    if (prev_post_parse_analyze_hook)
        prev_post_parse_analyze_hook(pstate, query);

    /* Assert we didn't do this already */
    Assert(query->queryId == UINT64CONST(0));

    /* Safety check... */
    if (!pgss || !pgss_hash)
        return;

    /*
     * Utility statements get queryId zero.  We do this even in cases where
     * the statement contains an optimizable statement for which a queryId
     * could be derived (such as EXPLAIN or DECLARE CURSOR).  For such cases,
     * runtime control will first go through ProcessUtility and then the
     * executor, and we don't want the executor hooks to do anything, since we
     * are already measuring the statement's costs at the utility level.
     */
    if (query->utilityStmt)
    {
        query->queryId = UINT64CONST(0);
        return;
    }

    /* Set up workspace for query jumbling */
    jstate.jumble = (unsigned char *) palloc(JUMBLE_SIZE);
    jstate.jumble_len = 0;
    jstate.clocations_buf_size = 32;
    jstate.clocations = (pgssLocationLen *)
        palloc(jstate.clocations_buf_size * sizeof(pgssLocationLen));
    jstate.clocations_count = 0;
    jstate.highest_extern_param_id = 0;

    /* Compute query ID and mark the Query node with it */
    JumbleQuery(&jstate, query);
    query->queryId =
        DatumGetUInt64(hash_any_extended(jstate.jumble, jstate.jumble_len, 0));

    /*
     * If we are unlucky enough to get a hash of zero, use 1 instead, to
     * prevent confusion with the utility-statement case.
     */
    if (query->queryId == UINT64CONST(0))
        query->queryId = UINT64CONST(1);

    /*
     * If we were able to identify any ignorable constants, we immediately
     * create a hash table entry for the query, so that we can record the
     * normalized form of the query string.  If there were no such constants,
     * the normalized string would be the same as the query text anyway, so
     * there's no need for an early entry.
     */
    if (jstate.clocations_count > 0)
        pgss_store(pstate->p_sourcetext,
                   query->queryId,
                   query->stmt_location,
                   query->stmt_len,
                   0,
                   0,
                   NULL,
                   &jstate);
}

/*
 * ExecutorStart hook: start up tracking if needed
 */
static void
pgss_ExecutorStart(QueryDesc *queryDesc, int eflags)
{
    if (prev_ExecutorStart)
        prev_ExecutorStart(queryDesc, eflags);
    else
        standard_ExecutorStart(queryDesc, eflags);

    /*
     * If query has queryId zero, don't track it.  This prevents double
     * counting of optimizable statements that are directly contained in
     * utility statements.
     */
    if (pgss_enabled() && queryDesc->plannedstmt->queryId != UINT64CONST(0))
    {
        /*
         * Set up to track total elapsed time in ExecutorRun.  Make sure the
         * space is allocated in the per-query context so it will go away at
         * ExecutorEnd.
         */
        if (queryDesc->totaltime == NULL)
        {
            MemoryContext oldcxt;

            oldcxt = MemoryContextSwitchTo(queryDesc->estate->es_query_cxt);
            queryDesc->totaltime = InstrAlloc(1, INSTRUMENT_ALL);
            MemoryContextSwitchTo(oldcxt);
        }
    }
}

/*
 * ExecutorRun hook: all we need do is track nesting depth
 */
static void
pgss_ExecutorRun(QueryDesc *queryDesc, ScanDirection direction, uint64 count,
                 bool execute_once)
{
    nested_level++;
    PG_TRY();
    {
        if (prev_ExecutorRun)
            prev_ExecutorRun(queryDesc, direction, count, execute_once);
        else
            standard_ExecutorRun(queryDesc, direction, count, execute_once);
        nested_level--;
    }
    PG_CATCH();
    {
        nested_level--;
        PG_RE_THROW();
    }
    PG_END_TRY();
}

/*
 * ExecutorFinish hook: all we need do is track nesting depth
 */
static void
pgss_ExecutorFinish(QueryDesc *queryDesc)
{
    nested_level++;
    PG_TRY();
    {
        if (prev_ExecutorFinish)
            prev_ExecutorFinish(queryDesc);
        else
            standard_ExecutorFinish(queryDesc);
        nested_level--;
    }
    PG_CATCH();
    {
        nested_level--;
        PG_RE_THROW();
    }
    PG_END_TRY();
}

/*
 * ExecutorEnd hook: store results if needed
 */
static void
pgss_ExecutorEnd(QueryDesc *queryDesc)
{
    uint64      queryId = queryDesc->plannedstmt->queryId;

    if (queryId != UINT64CONST(0) && queryDesc->totaltime && pgss_enabled())
    {
        /*
         * Make sure stats accumulation is done.  (Note: it's okay if several
         * levels of hook all do this.)
         */
        InstrEndLoop(queryDesc->totaltime);

        pgss_store(queryDesc->sourceText,
                   queryId,
                   queryDesc->plannedstmt->stmt_location,
                   queryDesc->plannedstmt->stmt_len,
                   queryDesc->totaltime->total * 1000.0,    /* convert to msec */
                   queryDesc->estate->es_processed,
                   &queryDesc->totaltime->bufusage,
                   NULL);
    }

    if (prev_ExecutorEnd)
        prev_ExecutorEnd(queryDesc);
    else
        standard_ExecutorEnd(queryDesc);
}

/*
 * ProcessUtility hook
 */
static void
pgss_ProcessUtility(PlannedStmt *pstmt, const char *queryString,
                    ProcessUtilityContext context,
                    ParamListInfo params, QueryEnvironment *queryEnv,
                    DestReceiver *dest, char *completionTag)
{
    Node       *parsetree = pstmt->utilityStmt;

    /*
     * If it's an EXECUTE statement, we don't track it and don't increment the
     * nesting level.  This allows the cycles to be charged to the underlying
     * PREPARE instead (by the Executor hooks), which is much more useful.
     *
     * We also don't track execution of PREPARE.  If we did, we would get one
     * hash table entry for the PREPARE (with hash calculated from the query
     * string), and then a different one with the same query string (but hash
     * calculated from the query tree) would be used to accumulate costs of
     * ensuing EXECUTEs.  This would be confusing, and inconsistent with other
     * cases where planning time is not included at all.
     *
     * Likewise, we don't track execution of DEALLOCATE.
     */
    if (pgss_track_utility && pgss_enabled() &&
        !IsA(parsetree, ExecuteStmt) &&
        !IsA(parsetree, PrepareStmt) &&
        !IsA(parsetree, DeallocateStmt))
    {
        instr_time  start;
        instr_time  duration;
        uint64      rows;
        BufferUsage bufusage_start,
                    bufusage;

        bufusage_start = pgBufferUsage;
        INSTR_TIME_SET_CURRENT(start);

        nested_level++;
        PG_TRY();
        {
            if (prev_ProcessUtility)
                prev_ProcessUtility(pstmt, queryString,
                                    context, params, queryEnv,
                                    dest, completionTag);
            else
                standard_ProcessUtility(pstmt, queryString,
                                        context, params, queryEnv,
                                        dest, completionTag);
            nested_level--;
        }
        PG_CATCH();
        {
            nested_level--;
            PG_RE_THROW();
        }
        PG_END_TRY();

        INSTR_TIME_SET_CURRENT(duration);
        INSTR_TIME_SUBTRACT(duration, start);

        /* parse command tag to retrieve the number of affected rows. */
        if (completionTag &&
            strncmp(completionTag, "COPY ", 5) == 0)
            rows = pg_strtouint64(completionTag + 5, NULL, 10);
        else
            rows = 0;

        /* calc differences of buffer counters. */
        bufusage.shared_blks_hit =
            pgBufferUsage.shared_blks_hit - bufusage_start.shared_blks_hit;
        bufusage.shared_blks_read =
            pgBufferUsage.shared_blks_read - bufusage_start.shared_blks_read;
        bufusage.shared_blks_dirtied =
            pgBufferUsage.shared_blks_dirtied - bufusage_start.shared_blks_dirtied;
        bufusage.shared_blks_written =
            pgBufferUsage.shared_blks_written - bufusage_start.shared_blks_written;
        bufusage.local_blks_hit =
            pgBufferUsage.local_blks_hit - bufusage_start.local_blks_hit;
        bufusage.local_blks_read =
            pgBufferUsage.local_blks_read - bufusage_start.local_blks_read;
        bufusage.local_blks_dirtied =
            pgBufferUsage.local_blks_dirtied - bufusage_start.local_blks_dirtied;
        bufusage.local_blks_written =
            pgBufferUsage.local_blks_written - bufusage_start.local_blks_written;
        bufusage.temp_blks_read =
            pgBufferUsage.temp_blks_read - bufusage_start.temp_blks_read;
        bufusage.temp_blks_written =
            pgBufferUsage.temp_blks_written - bufusage_start.temp_blks_written;
        bufusage.blk_read_time = pgBufferUsage.blk_read_time;
        INSTR_TIME_SUBTRACT(bufusage.blk_read_time, bufusage_start.blk_read_time);
        bufusage.blk_write_time = pgBufferUsage.blk_write_time;
        INSTR_TIME_SUBTRACT(bufusage.blk_write_time, bufusage_start.blk_write_time);

        pgss_store(queryString,
                   0,           /* signal that it's a utility stmt */
                   pstmt->stmt_location,
                   pstmt->stmt_len,
                   INSTR_TIME_GET_MILLISEC(duration),
                   rows,
                   &bufusage,
                   NULL);
    }
    else
    {
        if (prev_ProcessUtility)
            prev_ProcessUtility(pstmt, queryString,
                                context, params, queryEnv,
                                dest, completionTag);
        else
            standard_ProcessUtility(pstmt, queryString,
                                    context, params, queryEnv,
                                    dest, completionTag);
    }
}

/*
 * Given an arbitrarily long query string, produce a hash for the purposes of
 * identifying the query, without normalizing constants.  Used when hashing
 * utility statements.
 */
static uint64
pgss_hash_string(const char *str, int len)
{
    return DatumGetUInt64(hash_any_extended((const unsigned char *) str,
                                            len, 0));
}

/*
 * Store some statistics for a statement.
 *
 * If queryId is 0 then this is a utility statement and we should compute
 * a suitable queryId internally.
 *
 * If jstate is not NULL then we're trying to create an entry for which
 * we have no statistics as yet; we just want to record the normalized
 * query string.  total_time, rows, bufusage are ignored in this case.
 */
static void
pgss_store(const char *query, uint64 queryId,
           int query_location, int query_len,
           double total_time, uint64 rows,
           const BufferUsage *bufusage,
           pgssJumbleState *jstate)
{
    pgssHashKey key;
    pgssEntry  *entry;
    char       *norm_query = NULL;
    int         encoding = GetDatabaseEncoding();

    Assert(query != NULL);

    /* Safety check... */
    if (!pgss || !pgss_hash)
        return;

    /*
     * Confine our attention to the relevant part of the string, if the query
     * is a portion of a multi-statement source string.
     *
     * First apply starting offset, unless it's -1 (unknown).
     */
    if (query_location >= 0)
    {
        Assert(query_location <= strlen(query));
        query += query_location;
        /* Length of 0 (or -1) means "rest of string" */
        if (query_len <= 0)
            query_len = strlen(query);
        else
            Assert(query_len <= strlen(query));
    }
    else
    {
        /* If query location is unknown, distrust query_len as well */
        query_location = 0;
        query_len = strlen(query);
    }

    /*
     * Discard leading and trailing whitespace, too.  Use scanner_isspace()
     * not libc's isspace(), because we want to match the lexer's behavior.
     */
    while (query_len > 0 && scanner_isspace(query[0]))
        query++, query_location++, query_len--;
    while (query_len > 0 && scanner_isspace(query[query_len - 1]))
        query_len--;

    /*
     * For utility statements, we just hash the query string to get an ID.
     */
    if (queryId == UINT64CONST(0))
        queryId = pgss_hash_string(query, query_len);

    /* Set up key for hashtable search */
    key.userid = GetUserId();
    key.dbid = MyDatabaseId;
    key.queryid = queryId;

    /* Lookup the hash table entry with shared lock. */
    LWLockAcquire(pgss->lock, LW_SHARED);

    entry = (pgssEntry *) hash_search(pgss_hash, &key, HASH_FIND, NULL);

    /* Create new entry, if not present */
    if (!entry)
    {
        Size        query_offset;
        int         gc_count;
        bool        stored;
        bool        do_gc;

        /*
         * Create a new, normalized query string if caller asked.  We don't
         * need to hold the lock while doing this work.  (Note: in any case,
         * it's possible that someone else creates a duplicate hashtable entry
         * in the interval where we don't hold the lock below.  That case is
         * handled by entry_alloc.)
         */
        if (jstate)
        {
            LWLockRelease(pgss->lock);
            norm_query = generate_normalized_query(jstate, query,
                                                   query_location,
                                                   &query_len,
                                                   encoding);
            LWLockAcquire(pgss->lock, LW_SHARED);
        }

        /* Append new query text to file with only shared lock held */
        stored = qtext_store(norm_query ? norm_query : query, query_len,
                             &query_offset, &gc_count);

        /*
         * Determine whether we need to garbage collect external query texts
         * while the shared lock is still held.  This micro-optimization
         * avoids taking the time to decide this while holding exclusive lock.
         */
        do_gc = need_gc_qtexts();

        /* Need exclusive lock to make a new hashtable entry - promote */
        LWLockRelease(pgss->lock);
        LWLockAcquire(pgss->lock, LW_EXCLUSIVE);

        /*
         * A garbage collection may have occurred while we weren't holding the
         * lock.  In the unlikely event that this happens, the query text we
         * stored above will have been garbage collected, so write it again.
         * This should be infrequent enough that doing it while holding
         * exclusive lock isn't a performance problem.
         */
        if (!stored || pgss->gc_count != gc_count)
            stored = qtext_store(norm_query ? norm_query : query, query_len,
                                 &query_offset, NULL);

        /* If we failed to write to the text file, give up */
        if (!stored)
            goto done;

        /* OK to create a new hashtable entry */
        entry = entry_alloc(&key, query_offset, query_len, encoding,
                            jstate != NULL);

        /* If needed, perform garbage collection while exclusive lock held */
        if (do_gc)
            gc_qtexts();
    }

    /* Increment the counts, except when jstate is not NULL */
    if (!jstate)
    {
        /*
         * Grab the spinlock while updating the counters (see comment about
         * locking rules at the head of the file)
         */
        volatile pgssEntry *e = (volatile pgssEntry *) entry;

        SpinLockAcquire(&e->mutex);

        /* "Unstick" entry if it was previously sticky */
        if (e->counters.calls == 0)
            e->counters.usage = USAGE_INIT;

        e->counters.calls += 1;
        e->counters.total_time += total_time;
        if (e->counters.calls == 1)
        {
            e->counters.min_time = total_time;
            e->counters.max_time = total_time;
            e->counters.mean_time = total_time;
        }
        else
        {
            /*
             * Welford's method for accurately computing variance. See
             * <http://www.johndcook.com/blog/standard_deviation/>
             */
            double      old_mean = e->counters.mean_time;

            e->counters.mean_time +=
                (total_time - old_mean) / e->counters.calls;
            e->counters.sum_var_time +=
                (total_time - old_mean) * (total_time - e->counters.mean_time);

            /* calculate min and max time */
            if (e->counters.min_time > total_time)
                e->counters.min_time = total_time;
            if (e->counters.max_time < total_time)
                e->counters.max_time = total_time;
        }
        e->counters.rows += rows;
        e->counters.shared_blks_hit += bufusage->shared_blks_hit;
        e->counters.shared_blks_read += bufusage->shared_blks_read;
        e->counters.shared_blks_dirtied += bufusage->shared_blks_dirtied;
        e->counters.shared_blks_written += bufusage->shared_blks_written;
        e->counters.local_blks_hit += bufusage->local_blks_hit;
        e->counters.local_blks_read += bufusage->local_blks_read;
        e->counters.local_blks_dirtied += bufusage->local_blks_dirtied;
        e->counters.local_blks_written += bufusage->local_blks_written;
        e->counters.temp_blks_read += bufusage->temp_blks_read;
        e->counters.temp_blks_written += bufusage->temp_blks_written;
        e->counters.blk_read_time += INSTR_TIME_GET_MILLISEC(bufusage->blk_read_time);
        e->counters.blk_write_time += INSTR_TIME_GET_MILLISEC(bufusage->blk_write_time);
        e->counters.usage += USAGE_EXEC(total_time);

        SpinLockRelease(&e->mutex);
    }

done:
    LWLockRelease(pgss->lock);

    /* We postpone this clean-up until we're out of the lock */
    if (norm_query)
        pfree(norm_query);
}

/*
 * Reset all statement statistics.
 */
Datum
pg_stat_statements_reset(PG_FUNCTION_ARGS)
{
    if (!pgss || !pgss_hash)
        ereport(ERROR,
                (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
                 errmsg("pg_stat_statements must be loaded via shared_preload_libraries")));
    entry_reset();
    PG_RETURN_VOID();
}

/* Number of output arguments (columns) for various API versions */
#define PG_STAT_STATEMENTS_COLS_V1_0    14
#define PG_STAT_STATEMENTS_COLS_V1_1    18
#define PG_STAT_STATEMENTS_COLS_V1_2    19
#define PG_STAT_STATEMENTS_COLS_V1_3    23
#define PG_STAT_STATEMENTS_COLS         23  /* maximum of above */

/*
 * Retrieve statement statistics.
 *
 * The SQL API of this function has changed multiple times, and will likely
 * do so again in future.  To support the case where a newer version of this
 * loadable module is being used with an old SQL declaration of the function,
 * we continue to support the older API versions.  For 1.2 and later, the
 * expected API version is identified by embedding it in the C name of the
 * function.  Unfortunately we weren't bright enough to do that for 1.1.
 */
Datum
pg_stat_statements_1_3(PG_FUNCTION_ARGS)
{
    bool        showtext = PG_GETARG_BOOL(0);

    pg_stat_statements_internal(fcinfo, PGSS_V1_3, showtext);

    return (Datum) 0;
}

Datum
pg_stat_statements_1_2(PG_FUNCTION_ARGS)
{
    bool        showtext = PG_GETARG_BOOL(0);

    pg_stat_statements_internal(fcinfo, PGSS_V1_2, showtext);

    return (Datum) 0;
}

/*
 * Legacy entry point for pg_stat_statements() API versions 1.0 and 1.1.
 * This can be removed someday, perhaps.
 */
Datum
pg_stat_statements(PG_FUNCTION_ARGS)
{
    /* If it's really API 1.1, we'll figure that out below */
    pg_stat_statements_internal(fcinfo, PGSS_V1_0, true);

    return (Datum) 0;
}

/* Common code for all versions of pg_stat_statements() */
static void
pg_stat_statements_internal(FunctionCallInfo fcinfo,
                            pgssVersion api_version,
                            bool showtext)
{
    ReturnSetInfo *rsinfo = (ReturnSetInfo *) fcinfo->resultinfo;
    TupleDesc   tupdesc;
    Tuplestorestate *tupstore;
    MemoryContext per_query_ctx;
    MemoryContext oldcontext;
    Oid         userid = GetUserId();
    bool        is_allowed_role = false;
    char       *qbuffer = NULL;
    Size        qbuffer_size = 0;
    Size        extent = 0;
    int         gc_count = 0;
    HASH_SEQ_STATUS hash_seq;
    pgssEntry  *entry;

    /* Superusers or members of pg_read_all_stats members are allowed */
    is_allowed_role = is_member_of_role(GetUserId(), DEFAULT_ROLE_READ_ALL_STATS);

    /* hash table must exist already */
    if (!pgss || !pgss_hash)
        ereport(ERROR,
                (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
                 errmsg("pg_stat_statements must be loaded via shared_preload_libraries")));

    /* check to see if caller supports us returning a tuplestore */
    if (rsinfo == NULL || !IsA(rsinfo, ReturnSetInfo))
        ereport(ERROR,
                (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
                 errmsg("set-valued function called in context that cannot accept a set")));
    if (!(rsinfo->allowedModes & SFRM_Materialize))
        ereport(ERROR,
                (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
                 errmsg("materialize mode required, but it is not " \
                        "allowed in this context")));

    /* Switch into long-lived context to construct returned data structures */
    per_query_ctx = rsinfo->econtext->ecxt_per_query_memory;
    oldcontext = MemoryContextSwitchTo(per_query_ctx);

    /* Build a tuple descriptor for our result type */
    if (get_call_result_type(fcinfo, NULL, &tupdesc) != TYPEFUNC_COMPOSITE)
        elog(ERROR, "return type must be a row type");

    /*
     * Check we have the expected number of output arguments.  Aside from
     * being a good safety check, we need a kluge here to detect API version
     * 1.1, which was wedged into the code in an ill-considered way.
     */
    switch (tupdesc->natts)
    {
        case PG_STAT_STATEMENTS_COLS_V1_0:
            if (api_version != PGSS_V1_0)
                elog(ERROR, "incorrect number of output arguments");
            break;
        case PG_STAT_STATEMENTS_COLS_V1_1:
            /* pg_stat_statements() should have told us 1.0 */
            if (api_version != PGSS_V1_0)
                elog(ERROR, "incorrect number of output arguments");
            api_version = PGSS_V1_1;
            break;
        case PG_STAT_STATEMENTS_COLS_V1_2:
            if (api_version != PGSS_V1_2)
                elog(ERROR, "incorrect number of output arguments");
            break;
        case PG_STAT_STATEMENTS_COLS_V1_3:
            if (api_version != PGSS_V1_3)
                elog(ERROR, "incorrect number of output arguments");
            break;
        default:
            elog(ERROR, "incorrect number of output arguments");
    }

    tupstore = tuplestore_begin_heap(true, false, work_mem);
    rsinfo->returnMode = SFRM_Materialize;
    rsinfo->setResult = tupstore;
    rsinfo->setDesc = tupdesc;

    MemoryContextSwitchTo(oldcontext);

    /*
     * We'd like to load the query text file (if needed) while not holding any
     * lock on pgss->lock.  In the worst case we'll have to do this again
     * after we have the lock, but it's unlikely enough to make this a win
     * despite occasional duplicated work.  We need to reload if anybody
     * writes to the file (either a retail qtext_store(), or a garbage
     * collection) between this point and where we've gotten shared lock.  If
     * a qtext_store is actually in progress when we look, we might as well
     * skip the speculative load entirely.
     */
    if (showtext)
    {
        int         n_writers;

        /* Take the mutex so we can examine variables */
        {
            volatile pgssSharedState *s = (volatile pgssSharedState *) pgss;

            SpinLockAcquire(&s->mutex);
            extent = s->extent;
            n_writers = s->n_writers;
            gc_count = s->gc_count;
            SpinLockRelease(&s->mutex);
        }

        /* No point in loading file now if there are active writers */
        if (n_writers == 0)
            qbuffer = qtext_load_file(&qbuffer_size);
    }

    /*
     * Get shared lock, load or reload the query text file if we must, and
     * iterate over the hashtable entries.
     *
     * With a large hash table, we might be holding the lock rather longer
     * than one could wish.  However, this only blocks creation of new hash
     * table entries, and the larger the hash table the less likely that is to
     * be needed.  So we can hope this is okay.  Perhaps someday we'll decide
     * we need to partition the hash table to limit the time spent holding any
     * one lock.
     */
    LWLockAcquire(pgss->lock, LW_SHARED);

    if (showtext)
    {
        /*
         * Here it is safe to examine extent and gc_count without taking the
         * mutex.  Note that although other processes might change
         * pgss->extent just after we look at it, the strings they then write
         * into the file cannot yet be referenced in the hashtable, so we
         * don't care whether we see them or not.
         *
         * If qtext_load_file fails, we just press on; we'll return NULL for
         * every query text.
         */
        if (qbuffer == NULL ||
            pgss->extent != extent ||
            pgss->gc_count != gc_count)
        {
            if (qbuffer)
                free(qbuffer);
            qbuffer = qtext_load_file(&qbuffer_size);
        }
    }

    hash_seq_init(&hash_seq, pgss_hash);
    while ((entry = hash_seq_search(&hash_seq)) != NULL)
    {
        Datum       values[PG_STAT_STATEMENTS_COLS];
        bool        nulls[PG_STAT_STATEMENTS_COLS];
        int         i = 0;
        Counters    tmp;
        double      stddev;
        int64       queryid = entry->key.queryid;

        memset(values, 0, sizeof(values));
        memset(nulls, 0, sizeof(nulls));

        values[i++] = ObjectIdGetDatum(entry->key.userid);
        values[i++] = ObjectIdGetDatum(entry->key.dbid);

        if (is_allowed_role || entry->key.userid == userid)
        {
            if (api_version >= PGSS_V1_2)
                values[i++] = Int64GetDatumFast(queryid);

            if (showtext)
            {
                char       *qstr = qtext_fetch(entry->query_offset,
                                               entry->query_len,
                                               qbuffer,
                                               qbuffer_size);

                if (qstr)
                {
                    char       *enc;

                    enc = pg_any_to_server(qstr,
                                           entry->query_len,
                                           entry->encoding);

                    values[i++] = CStringGetTextDatum(enc);

                    if (enc != qstr)
                        pfree(enc);
                }
                else
                {
                    /* Just return a null if we fail to find the text */
                    nulls[i++] = true;
                }
            }
            else
            {
                /* Query text not requested */
                nulls[i++] = true;
            }
        }
        else
        {
            /* Don't show queryid */
            if (api_version >= PGSS_V1_2)
                nulls[i++] = true;

            /*
             * Don't show query text, but hint as to the reason for not doing
             * so if it was requested
             */
            if (showtext)
                values[i++] = CStringGetTextDatum("<insufficient privilege>");
            else
                nulls[i++] = true;
        }

        /* copy counters to a local variable to keep locking time short */
        {
            volatile pgssEntry *e = (volatile pgssEntry *) entry;

            SpinLockAcquire(&e->mutex);
            tmp = e->counters;
            SpinLockRelease(&e->mutex);
        }

        /* Skip entry if unexecuted (ie, it's a pending "sticky" entry) */
        if (tmp.calls == 0)
            continue;

        values[i++] = Int64GetDatumFast(tmp.calls);
        values[i++] = Float8GetDatumFast(tmp.total_time);
        if (api_version >= PGSS_V1_3)
        {
            values[i++] = Float8GetDatumFast(tmp.min_time);
            values[i++] = Float8GetDatumFast(tmp.max_time);
            values[i++] = Float8GetDatumFast(tmp.mean_time);

            /*
             * Note we are calculating the population variance here, not the
             * sample variance, as we have data for the whole population, so
             * Bessel's correction is not used, and we don't divide by
             * tmp.calls - 1.
             */
            if (tmp.calls > 1)
                stddev = sqrt(tmp.sum_var_time / tmp.calls);
            else
                stddev = 0.0;
            values[i++] = Float8GetDatumFast(stddev);
        }
        values[i++] = Int64GetDatumFast(tmp.rows);
        values[i++] = Int64GetDatumFast(tmp.shared_blks_hit);
        values[i++] = Int64GetDatumFast(tmp.shared_blks_read);
        if (api_version >= PGSS_V1_1)
            values[i++] = Int64GetDatumFast(tmp.shared_blks_dirtied);
        values[i++] = Int64GetDatumFast(tmp.shared_blks_written);
        values[i++] = Int64GetDatumFast(tmp.local_blks_hit);
        values[i++] = Int64GetDatumFast(tmp.local_blks_read);
        if (api_version >= PGSS_V1_1)
            values[i++] = Int64GetDatumFast(tmp.local_blks_dirtied);
        values[i++] = Int64GetDatumFast(tmp.local_blks_written);
        values[i++] = Int64GetDatumFast(tmp.temp_blks_read);
        values[i++] = Int64GetDatumFast(tmp.temp_blks_written);
        if (api_version >= PGSS_V1_1)
        {
            values[i++] = Float8GetDatumFast(tmp.blk_read_time);
            values[i++] = Float8GetDatumFast(tmp.blk_write_time);
        }

        Assert(i == (api_version == PGSS_V1_0 ? PG_STAT_STATEMENTS_COLS_V1_0 :
                     api_version == PGSS_V1_1 ? PG_STAT_STATEMENTS_COLS_V1_1 :
                     api_version == PGSS_V1_2 ? PG_STAT_STATEMENTS_COLS_V1_2 :
                     api_version == PGSS_V1_3 ? PG_STAT_STATEMENTS_COLS_V1_3 :
                     -1 /* fail if you forget to update this assert */ ));

        tuplestore_putvalues(tupstore, tupdesc, values, nulls);
    }

    /* clean up and return the tuplestore */
    LWLockRelease(pgss->lock);

    if (qbuffer)
        free(qbuffer);

    tuplestore_donestoring(tupstore);
}

/*
 * Estimate shared memory space needed.
 */
static Size
pgss_memsize(void)
{
    Size        size;

    size = MAXALIGN(sizeof(pgssSharedState));
    size = add_size(size, hash_estimate_size(pgss_max, sizeof(pgssEntry)));

    return size;
}

/*
 * Allocate a new hashtable entry.
 * caller must hold an exclusive lock on pgss->lock
 *
 * "query" need not be null-terminated; we rely on query_len instead
 *
 * If "sticky" is true, make the new entry artificially sticky so that it will
 * probably still be there when the query finishes execution.  We do this by
 * giving it a median usage value rather than the normal value.  (Strictly
 * speaking, query strings are normalized on a best effort basis, though it
 * would be difficult to demonstrate this even under artificial conditions.)
 *
 * Note: despite needing exclusive lock, it's not an error for the target
 * entry to already exist.  This is because pgss_store releases and
 * reacquires lock after failing to find a match; so someone else could
 * have made the entry while we waited to get exclusive lock.
 */
static pgssEntry *
entry_alloc(pgssHashKey *key, Size query_offset, int query_len, int encoding,
            bool sticky)
{
    pgssEntry  *entry;
    bool        found;

    /* Make space if needed */
    while (hash_get_num_entries(pgss_hash) >= pgss_max)
        entry_dealloc();

    /* Find or create an entry with desired hash code */
    entry = (pgssEntry *) hash_search(pgss_hash, key, HASH_ENTER, &found);

    if (!found)
    {
        /* New entry, initialize it */

        /* reset the statistics */
        memset(&entry->counters, 0, sizeof(Counters));
        /* set the appropriate initial usage count */
        entry->counters.usage = sticky ? pgss->cur_median_usage : USAGE_INIT;
        /* re-initialize the mutex each time ... we assume no one using it */
        SpinLockInit(&entry->mutex);
        /* ... and don't forget the query text metadata */
        Assert(query_len >= 0);
        entry->query_offset = query_offset;
        entry->query_len = query_len;
        entry->encoding = encoding;
    }

    return entry;
}

/*
 * qsort comparator for sorting into increasing usage order
 */
static int
entry_cmp(const void *lhs, const void *rhs)
{
    double      l_usage = (*(pgssEntry *const *) lhs)->counters.usage;
    double      r_usage = (*(pgssEntry *const *) rhs)->counters.usage;

    if (l_usage < r_usage)
        return -1;
    else if (l_usage > r_usage)
        return +1;
    else
        return 0;
}

/*
 * Deallocate least-used entries.
 *
 * Caller must hold an exclusive lock on pgss->lock.
 */
static void
entry_dealloc(void)
{
    HASH_SEQ_STATUS hash_seq;
    pgssEntry **entries;
    pgssEntry  *entry;
    int         nvictims;
    int         i;
    Size        tottextlen;
    int         nvalidtexts;

    /*
     * Sort entries by usage and deallocate USAGE_DEALLOC_PERCENT of them.
     * While we're scanning the table, apply the decay factor to the usage
     * values, and update the mean query length.
     *
     * Note that the mean query length is almost immediately obsolete, since
     * we compute it before not after discarding the least-used entries.
     * Hopefully, that doesn't affect the mean too much; it doesn't seem worth
     * making two passes to get a more current result.  Likewise, the new
     * cur_median_usage includes the entries we're about to zap.
     */

    entries = palloc(hash_get_num_entries(pgss_hash) * sizeof(pgssEntry *));

    i = 0;
    tottextlen = 0;
    nvalidtexts = 0;

    hash_seq_init(&hash_seq, pgss_hash);
    while ((entry = hash_seq_search(&hash_seq)) != NULL)
    {
        entries[i++] = entry;
        /* "Sticky" entries get a different usage decay rate. */
        if (entry->counters.calls == 0)
            entry->counters.usage *= STICKY_DECREASE_FACTOR;
        else
            entry->counters.usage *= USAGE_DECREASE_FACTOR;
        /* In the mean length computation, ignore dropped texts. */
        if (entry->query_len >= 0)
        {
            tottextlen += entry->query_len + 1;
            nvalidtexts++;
        }
    }

    /* Sort into increasing order by usage */
    qsort(entries, i, sizeof(pgssEntry *), entry_cmp);

    /* Record the (approximate) median usage */
    if (i > 0)
        pgss->cur_median_usage = entries[i / 2]->counters.usage;
    /* Record the mean query length */
    if (nvalidtexts > 0)
        pgss->mean_query_len = tottextlen / nvalidtexts;
    else
        pgss->mean_query_len = ASSUMED_LENGTH_INIT;

    /* Now zap an appropriate fraction of lowest-usage entries */
    nvictims = Max(10, i * USAGE_DEALLOC_PERCENT / 100);
    nvictims = Min(nvictims, i);

    for (i = 0; i < nvictims; i++)
    {
        hash_search(pgss_hash, &entries[i]->key, HASH_REMOVE, NULL);
    }

    pfree(entries);
}

/*
 * Given a query string (not necessarily null-terminated), allocate a new
 * entry in the external query text file and store the string there.
 *
 * If successful, returns true, and stores the new entry's offset in the file
 * into *query_offset.  Also, if gc_count isn't NULL, *gc_count is set to the
 * number of garbage collections that have occurred so far.
 *
 * On failure, returns false.
 *
 * At least a shared lock on pgss->lock must be held by the caller, so as
 * to prevent a concurrent garbage collection.  Share-lock-holding callers
 * should pass a gc_count pointer to obtain the number of garbage collections,
 * so that they can recheck the count after obtaining exclusive lock to
 * detect whether a garbage collection occurred (and removed this entry).
 */
static bool
qtext_store(const char *query, int query_len,
            Size *query_offset, int *gc_count)
{
    Size        off;
    int         fd;

    /*
     * We use a spinlock to protect extent/n_writers/gc_count, so that
     * multiple processes may execute this function concurrently.
     */
    {
        volatile pgssSharedState *s = (volatile pgssSharedState *) pgss;

        SpinLockAcquire(&s->mutex);
        off = s->extent;
        s->extent += query_len + 1;
        s->n_writers++;
        if (gc_count)
            *gc_count = s->gc_count;
        SpinLockRelease(&s->mutex);
    }

    *query_offset = off;

    /* Now write the data into the successfully-reserved part of the file */
    fd = OpenTransientFile(PGSS_TEXT_FILE, O_RDWR | O_CREAT | PG_BINARY);
    if (fd < 0)
        goto error;

    if (lseek(fd, off, SEEK_SET) != off)
        goto error;

    if (write(fd, query, query_len) != query_len)
        goto error;
    if (write(fd, "\0", 1) != 1)
        goto error;

    CloseTransientFile(fd);

    /* Mark our write complete */
    {
        volatile pgssSharedState *s = (volatile pgssSharedState *) pgss;

        SpinLockAcquire(&s->mutex);
        s->n_writers--;
        SpinLockRelease(&s->mutex);
    }

    return true;

error:
    ereport(LOG,
            (errcode_for_file_access(),
             errmsg("could not write pg_stat_statement file \"%s\": %m",
                    PGSS_TEXT_FILE)));

    if (fd >= 0)
        CloseTransientFile(fd);

    /* Mark our write complete */
    {
        volatile pgssSharedState *s = (volatile pgssSharedState *) pgss;

        SpinLockAcquire(&s->mutex);
        s->n_writers--;
        SpinLockRelease(&s->mutex);
    }

    return false;
}

/*
 * Read the external query text file into a malloc'd buffer.
 *
 * Returns NULL (without throwing an error) if unable to read, eg
 * file not there or insufficient memory.
 *
 * On success, the buffer size is also returned into *buffer_size.
 *
 * This can be called without any lock on pgss->lock, but in that case
 * the caller is responsible for verifying that the result is sane.
 */
static char *
qtext_load_file(Size *buffer_size)
{
    char       *buf;
    int         fd;
    struct stat stat;

    fd = OpenTransientFile(PGSS_TEXT_FILE, O_RDONLY | PG_BINARY);
    if (fd < 0)
    {
        if (errno != ENOENT)
            ereport(LOG,
                    (errcode_for_file_access(),
                     errmsg("could not read pg_stat_statement file \"%s\": %m",
                            PGSS_TEXT_FILE)));
        return NULL;
    }

    /* Get file length */
    if (fstat(fd, &stat))
    {
        ereport(LOG,
                (errcode_for_file_access(),
                 errmsg("could not stat pg_stat_statement file \"%s\": %m",
                        PGSS_TEXT_FILE)));
        CloseTransientFile(fd);
        return NULL;
    }

    /* Allocate buffer; beware that off_t might be wider than size_t */
    if (stat.st_size <= MaxAllocHugeSize)
        buf = (char *) malloc(stat.st_size);
    else
        buf = NULL;
    if (buf == NULL)
    {
        ereport(LOG,
                (errcode(ERRCODE_OUT_OF_MEMORY),
                 errmsg("out of memory"),
                 errdetail("Could not allocate enough memory to read pg_stat_statement file \"%s\".",
                           PGSS_TEXT_FILE)));
        CloseTransientFile(fd);
        return NULL;
    }

    /*
     * OK, slurp in the file.  If we get a short read and errno doesn't get
     * set, the reason is probably that garbage collection truncated the file
     * since we did the fstat(), so we don't log a complaint --- but we don't
     * return the data, either, since it's most likely corrupt due to
     * concurrent writes from garbage collection.
     */
    errno = 0;
    if (read(fd, buf, stat.st_size) != stat.st_size)
    {
        if (errno)
            ereport(LOG,
                    (errcode_for_file_access(),
                     errmsg("could not read pg_stat_statement file \"%s\": %m",
                            PGSS_TEXT_FILE)));
        free(buf);
        CloseTransientFile(fd);
        return NULL;
    }

    CloseTransientFile(fd);

    *buffer_size = stat.st_size;
    return buf;
}

/*
 * Locate a query text in the file image previously read by qtext_load_file().
 *
 * We validate the given offset/length, and return NULL if bogus.  Otherwise,
 * the result points to a null-terminated string within the buffer.
 */
static char *
qtext_fetch(Size query_offset, int query_len,
            char *buffer, Size buffer_size)
{
    /* File read failed? */
    if (buffer == NULL)
        return NULL;
    /* Bogus offset/length? */
    if (query_len < 0 ||
        query_offset + query_len >= buffer_size)
        return NULL;
    /* As a further sanity check, make sure there's a trailing null */
    if (buffer[query_offset + query_len] != '\0')
        return NULL;
    /* Looks OK */
    return buffer + query_offset;
}

/*
 * Do we need to garbage-collect the external query text file?
 *
 * Caller should hold at least a shared lock on pgss->lock.
 */
static bool
need_gc_qtexts(void)
{
    Size        extent;

    /* Read shared extent pointer */
    {
        volatile pgssSharedState *s = (volatile pgssSharedState *) pgss;

        SpinLockAcquire(&s->mutex);
        extent = s->extent;
        SpinLockRelease(&s->mutex);
    }

    /* Don't proceed if file does not exceed 512 bytes per possible entry */
    if (extent < 512 * pgss_max)
        return false;

    /*
     * Don't proceed if file is less than about 50% bloat.  Nothing can or
     * should be done in the event of unusually large query texts accounting
     * for file's large size.  We go to the trouble of maintaining the mean
     * query length in order to prevent garbage collection from thrashing
     * uselessly.
     */
    if (extent < pgss->mean_query_len * pgss_max * 2)
        return false;

    return true;
}

/*
 * Garbage-collect orphaned query texts in external file.
 *
 * This won't be called often in the typical case, since it's likely that
 * there won't be too much churn, and besides, a similar compaction process
 * occurs when serializing to disk at shutdown or as part of resetting.
 * Despite this, it seems prudent to plan for the edge case where the file
 * becomes unreasonably large, with no other method of compaction likely to
 * occur in the foreseeable future.
 *
 * The caller must hold an exclusive lock on pgss->lock.
 *
 * At the first sign of trouble we unlink the query text file to get a clean
 * slate (although existing statistics are retained), rather than risk
 * thrashing by allowing the same problem case to recur indefinitely.
 */
static void
gc_qtexts(void)
{
    char       *qbuffer;
    Size        qbuffer_size;
    FILE       *qfile = NULL;
    HASH_SEQ_STATUS hash_seq;
    pgssEntry  *entry;
    Size        extent;
    int         nentries;

    /*
     * When called from pgss_store, some other session might have proceeded
     * with garbage collection in the no-lock-held interim of lock strength
     * escalation.  Check once more that this is actually necessary.
     */
    if (!need_gc_qtexts())
        return;

    /*
     * Load the old texts file.  If we fail (out of memory, for instance),
     * invalidate query texts.  Hopefully this is rare.  It might seem better
     * to leave things alone on an OOM failure, but the problem is that the
     * file is only going to get bigger; hoping for a future non-OOM result is
     * risky and can easily lead to complete denial of service.
     */
    qbuffer = qtext_load_file(&qbuffer_size);
    if (qbuffer == NULL)
        goto gc_fail;

    /*
     * We overwrite the query texts file in place, so as to reduce the risk of
     * an out-of-disk-space failure.  Since the file is guaranteed not to get
     * larger, this should always work on traditional filesystems; though we
     * could still lose on copy-on-write filesystems.
     */
    qfile = AllocateFile(PGSS_TEXT_FILE, PG_BINARY_W);
    if (qfile == NULL)
    {
        ereport(LOG,
                (errcode_for_file_access(),
                 errmsg("could not write pg_stat_statement file \"%s\": %m",
                        PGSS_TEXT_FILE)));
        goto gc_fail;
    }

    extent = 0;
    nentries = 0;

    hash_seq_init(&hash_seq, pgss_hash);
    while ((entry = hash_seq_search(&hash_seq)) != NULL)
    {
        int         query_len = entry->query_len;
        char       *qry = qtext_fetch(entry->query_offset,
                                      query_len,
                                      qbuffer,
                                      qbuffer_size);

        if (qry == NULL)
        {
            /* Trouble ... drop the text */
            entry->query_offset = 0;
            entry->query_len = -1;
            /* entry will not be counted in mean query length computation */
            continue;
        }

        if (fwrite(qry, 1, query_len + 1, qfile) != query_len + 1)
        {
            ereport(LOG,
                    (errcode_for_file_access(),
                     errmsg("could not write pg_stat_statement file \"%s\": %m",
                            PGSS_TEXT_FILE)));
            hash_seq_term(&hash_seq);
            goto gc_fail;
        }

        entry->query_offset = extent;
        extent += query_len + 1;
        nentries++;
    }

    /*
     * Truncate away any now-unused space.  If this fails for some odd reason,
     * we log it, but there's no need to fail.
     */
    if (ftruncate(fileno(qfile), extent) != 0)
        ereport(LOG,
                (errcode_for_file_access(),
                 errmsg("could not truncate pg_stat_statement file \"%s\": %m",
                        PGSS_TEXT_FILE)));

    if (FreeFile(qfile))
    {
        ereport(LOG,
                (errcode_for_file_access(),
                 errmsg("could not write pg_stat_statement file \"%s\": %m",
                        PGSS_TEXT_FILE)));
        qfile = NULL;
        goto gc_fail;
    }

    elog(DEBUG1, "pgss gc of queries file shrunk size from %zu to %zu",
         pgss->extent, extent);

    /* Reset the shared extent pointer */
    pgss->extent = extent;

    /*
     * Also update the mean query length, to be sure that need_gc_qtexts()
     * won't still think we have a problem.
     */
    if (nentries > 0)
        pgss->mean_query_len = extent / nentries;
    else
        pgss->mean_query_len = ASSUMED_LENGTH_INIT;

    free(qbuffer);

    /*
     * OK, count a garbage collection cycle.  (Note: even though we have
     * exclusive lock on pgss->lock, we must take pgss->mutex for this, since
     * other processes may examine gc_count while holding only the mutex.
     * Also, we have to advance the count *after* we've rewritten the file,
     * else other processes might not realize they read a stale file.)
     */
    record_gc_qtexts();

    return;

gc_fail:
    /* clean up resources */
    if (qfile)
        FreeFile(qfile);
    if (qbuffer)
        free(qbuffer);

    /*
     * Since the contents of the external file are now uncertain, mark all
     * hashtable entries as having invalid texts.
     */
    hash_seq_init(&hash_seq, pgss_hash);
    while ((entry = hash_seq_search(&hash_seq)) != NULL)
    {
        entry->query_offset = 0;
        entry->query_len = -1;
    }

    /*
     * Destroy the query text file and create a new, empty one
     */
    (void) unlink(PGSS_TEXT_FILE);
    qfile = AllocateFile(PGSS_TEXT_FILE, PG_BINARY_W);
    if (qfile == NULL)
        ereport(LOG,
                (errcode_for_file_access(),
                 errmsg("could not write new pg_stat_statement file \"%s\": %m",
                        PGSS_TEXT_FILE)));
    else
        FreeFile(qfile);

    /* Reset the shared extent pointer */
    pgss->extent = 0;

    /* Reset mean_query_len to match the new state */
    pgss->mean_query_len = ASSUMED_LENGTH_INIT;

    /*
     * Bump the GC count even though we failed.
     *
     * This is needed to make concurrent readers of file without any lock on
     * pgss->lock notice existence of new version of file.  Once readers
     * subsequently observe a change in GC count with pgss->lock held, that
     * forces a safe reopen of file.  Writers also require that we bump here,
     * of course.  (As required by locking protocol, readers and writers don't
     * trust earlier file contents until gc_count is found unchanged after
     * pgss->lock acquired in shared or exclusive mode respectively.)
     */
    record_gc_qtexts();
}

/*
 * Release all entries.
 */
static void
entry_reset(void)
{
    HASH_SEQ_STATUS hash_seq;
    pgssEntry  *entry;
    FILE       *qfile;

    LWLockAcquire(pgss->lock, LW_EXCLUSIVE);

    hash_seq_init(&hash_seq, pgss_hash);
    while ((entry = hash_seq_search(&hash_seq)) != NULL)
    {
        hash_search(pgss_hash, &entry->key, HASH_REMOVE, NULL);
    }

    /*
     * Write new empty query file, perhaps even creating a new one to recover
     * if the file was missing.
     */
    qfile = AllocateFile(PGSS_TEXT_FILE, PG_BINARY_W);
    if (qfile == NULL)
    {
        ereport(LOG,
                (errcode_for_file_access(),
                 errmsg("could not create pg_stat_statement file \"%s\": %m",
                        PGSS_TEXT_FILE)));
        goto done;
    }

    /* If ftruncate fails, log it, but it's not a fatal problem */
    if (ftruncate(fileno(qfile), 0) != 0)
        ereport(LOG,
                (errcode_for_file_access(),
                 errmsg("could not truncate pg_stat_statement file \"%s\": %m",
                        PGSS_TEXT_FILE)));

    FreeFile(qfile);

done:
    pgss->extent = 0;
    /* This counts as a query text garbage collection for our purposes */
    record_gc_qtexts();

    LWLockRelease(pgss->lock);
}

/*
 * AppendJumble: Append a value that is substantive in a given query to
 * the current jumble.
 */
static void
AppendJumble(pgssJumbleState *jstate, const unsigned char *item, Size size)
{
    unsigned char *jumble = jstate->jumble;
    Size        jumble_len = jstate->jumble_len;

    /*
     * Whenever the jumble buffer is full, we hash the current contents and
     * reset the buffer to contain just that hash value, thus relying on the
     * hash to summarize everything so far.
     */
    while (size > 0)
    {
        Size        part_size;

        if (jumble_len >= JUMBLE_SIZE)
        {
            uint64      start_hash;

            start_hash = DatumGetUInt64(hash_any_extended(jumble,
                                                          JUMBLE_SIZE, 0));
            memcpy(jumble, &start_hash, sizeof(start_hash));
            jumble_len = sizeof(start_hash);
        }
        part_size = Min(size, JUMBLE_SIZE - jumble_len);
        memcpy(jumble + jumble_len, item, part_size);
        jumble_len += part_size;
        item += part_size;
        size -= part_size;
    }
    jstate->jumble_len = jumble_len;
}

/*
 * Wrappers around AppendJumble to encapsulate details of serialization
 * of individual local variable elements.
 */
#define APP_JUMB(item) \
    AppendJumble(jstate, (const unsigned char *) &(item), sizeof(item))
#define APP_JUMB_STRING(str) \
    AppendJumble(jstate, (const unsigned char *) (str), strlen(str) + 1)

/*
 * JumbleQuery: Selectively serialize the query tree, appending significant
 * data to the "query jumble" while ignoring nonsignificant data.
 *
 * Rule of thumb for what to include is that we should ignore anything not
 * semantically significant (such as alias names) as well as anything that can
 * be deduced from child nodes (else we'd just be double-hashing that piece
 * of information).
 */
static void
JumbleQuery(pgssJumbleState *jstate, Query *query)
{
    Assert(IsA(query, Query));
    Assert(query->utilityStmt == NULL);

    APP_JUMB(query->commandType);
    /* resultRelation is usually predictable from commandType */
    JumbleExpr(jstate, (Node *) query->cteList);
    JumbleRangeTable(jstate, query->rtable);
    JumbleExpr(jstate, (Node *) query->jointree);
    JumbleExpr(jstate, (Node *) query->targetList);
    JumbleExpr(jstate, (Node *) query->onConflict);
    JumbleExpr(jstate, (Node *) query->returningList);
    JumbleExpr(jstate, (Node *) query->groupClause);
    JumbleExpr(jstate, (Node *) query->groupingSets);
    JumbleExpr(jstate, query->havingQual);
    JumbleExpr(jstate, (Node *) query->windowClause);
    JumbleExpr(jstate, (Node *) query->distinctClause);
    JumbleExpr(jstate, (Node *) query->sortClause);
    JumbleExpr(jstate, query->limitOffset);
    JumbleExpr(jstate, query->limitCount);
    /* we ignore rowMarks */
    JumbleExpr(jstate, query->setOperations);
}

/*
 * Jumble a range table
 */
static void
JumbleRangeTable(pgssJumbleState *jstate, List *rtable)
{
    ListCell   *lc;

    foreach(lc, rtable)
    {
        RangeTblEntry *rte = lfirst_node(RangeTblEntry, lc);

        APP_JUMB(rte->rtekind);
        switch (rte->rtekind)
        {
            case RTE_RELATION:
                APP_JUMB(rte->relid);
                JumbleExpr(jstate, (Node *) rte->tablesample);
                break;
            case RTE_SUBQUERY:
                JumbleQuery(jstate, rte->subquery);
                break;
            case RTE_JOIN:
                APP_JUMB(rte->jointype);
                break;
            case RTE_FUNCTION:
                JumbleExpr(jstate, (Node *) rte->functions);
                break;
            case RTE_TABLEFUNC:
                JumbleExpr(jstate, (Node *) rte->tablefunc);
                break;
            case RTE_VALUES:
                JumbleExpr(jstate, (Node *) rte->values_lists);
                break;
            case RTE_CTE:

                /*
                 * Depending on the CTE name here isn't ideal, but it's the
                 * only info we have to identify the referenced WITH item.
                 */
                APP_JUMB_STRING(rte->ctename);
                APP_JUMB(rte->ctelevelsup);
                break;
            case RTE_NAMEDTUPLESTORE:
                APP_JUMB_STRING(rte->enrname);
                break;
            default:
                elog(ERROR, "unrecognized RTE kind: %d", (int) rte->rtekind);
                break;
        }
    }
}

/*
 * Jumble an expression tree
 *
 * In general this function should handle all the same node types that
 * expression_tree_walker() does, and therefore it's coded to be as parallel
 * to that function as possible.  However, since we are only invoked on
 * queries immediately post-parse-analysis, we need not handle node types
 * that only appear in planning.
 *
 * Note: the reason we don't simply use expression_tree_walker() is that the
 * point of that function is to support tree walkers that don't care about
 * most tree node types, but here we care about all types.  We should complain
 * about any unrecognized node type.
 */
static void
JumbleExpr(pgssJumbleState *jstate, Node *node)
{
    ListCell   *temp;

    if (node == NULL)
        return;

    /* Guard against stack overflow due to overly complex expressions */
    check_stack_depth();

    /*
     * We always emit the node's NodeTag, then any additional fields that are
     * considered significant, and then we recurse to any child nodes.
     */
    APP_JUMB(node->type);

    switch (nodeTag(node))
    {
        case T_Var:
            {
                Var        *var = (Var *) node;

                APP_JUMB(var->varno);
                APP_JUMB(var->varattno);
                APP_JUMB(var->varlevelsup);
            }
            break;
        case T_Const:
            {
                Const      *c = (Const *) node;

                /* We jumble only the constant's type, not its value */
                APP_JUMB(c->consttype);
                /* Also, record its parse location for query normalization */
                RecordConstLocation(jstate, c->location);
            }
            break;
        case T_Param:
            {
                Param      *p = (Param *) node;

                APP_JUMB(p->paramkind);
                APP_JUMB(p->paramid);
                APP_JUMB(p->paramtype);
                /* Also, track the highest external Param id */
                if (p->paramkind == PARAM_EXTERN &&
                    p->paramid > jstate->highest_extern_param_id)
                    jstate->highest_extern_param_id = p->paramid;
            }
            break;
        case T_Aggref:
            {
                Aggref     *expr = (Aggref *) node;

                APP_JUMB(expr->aggfnoid);
                JumbleExpr(jstate, (Node *) expr->aggdirectargs);
                JumbleExpr(jstate, (Node *) expr->args);
                JumbleExpr(jstate, (Node *) expr->aggorder);
                JumbleExpr(jstate, (Node *) expr->aggdistinct);
                JumbleExpr(jstate, (Node *) expr->aggfilter);
            }
            break;
        case T_GroupingFunc:
            {
                GroupingFunc *grpnode = (GroupingFunc *) node;

                JumbleExpr(jstate, (Node *) grpnode->refs);
            }
            break;
        case T_WindowFunc:
            {
                WindowFunc *expr = (WindowFunc *) node;

                APP_JUMB(expr->winfnoid);
                APP_JUMB(expr->winref);
                JumbleExpr(jstate, (Node *) expr->args);
                JumbleExpr(jstate, (Node *) expr->aggfilter);
            }
            break;
        case T_ArrayRef:
            {
                ArrayRef   *aref = (ArrayRef *) node;

                JumbleExpr(jstate, (Node *) aref->refupperindexpr);
                JumbleExpr(jstate, (Node *) aref->reflowerindexpr);
                JumbleExpr(jstate, (Node *) aref->refexpr);
                JumbleExpr(jstate, (Node *) aref->refassgnexpr);
            }
            break;
        case T_FuncExpr:
            {
                FuncExpr   *expr = (FuncExpr *) node;

                APP_JUMB(expr->funcid);
                JumbleExpr(jstate, (Node *) expr->args);
            }
            break;
        case T_NamedArgExpr:
            {
                NamedArgExpr *nae = (NamedArgExpr *) node;

                APP_JUMB(nae->argnumber);
                JumbleExpr(jstate, (Node *) nae->arg);
            }
            break;
        case T_OpExpr:
        case T_DistinctExpr:    /* struct-equivalent to OpExpr */
        case T_NullIfExpr:      /* struct-equivalent to OpExpr */
            {
                OpExpr     *expr = (OpExpr *) node;

                APP_JUMB(expr->opno);
                JumbleExpr(jstate, (Node *) expr->args);
            }
            break;
        case T_ScalarArrayOpExpr:
            {
                ScalarArrayOpExpr *expr = (ScalarArrayOpExpr *) node;

                APP_JUMB(expr->opno);
                APP_JUMB(expr->useOr);
                JumbleExpr(jstate, (Node *) expr->args);
            }
            break;
        case T_BoolExpr:
            {
                BoolExpr   *expr = (BoolExpr *) node;

                APP_JUMB(expr->boolop);
                JumbleExpr(jstate, (Node *) expr->args);
            }
            break;
        case T_SubLink:
            {
                SubLink    *sublink = (SubLink *) node;

                APP_JUMB(sublink->subLinkType);
                APP_JUMB(sublink->subLinkId);
                JumbleExpr(jstate, (Node *) sublink->testexpr);
                JumbleQuery(jstate, castNode(Query, sublink->subselect));
            }
            break;
        case T_FieldSelect:
            {
                FieldSelect *fs = (FieldSelect *) node;

                APP_JUMB(fs->fieldnum);
                JumbleExpr(jstate, (Node *) fs->arg);
            }
            break;
        case T_FieldStore:
            {
                FieldStore *fstore = (FieldStore *) node;

                JumbleExpr(jstate, (Node *) fstore->arg);
                JumbleExpr(jstate, (Node *) fstore->newvals);
            }
            break;
        case T_RelabelType:
            {
                RelabelType *rt = (RelabelType *) node;

                APP_JUMB(rt->resulttype);
                JumbleExpr(jstate, (Node *) rt->arg);
            }
            break;
        case T_CoerceViaIO:
            {
                CoerceViaIO *cio = (CoerceViaIO *) node;

                APP_JUMB(cio->resulttype);
                JumbleExpr(jstate, (Node *) cio->arg);
            }
            break;
        case T_ArrayCoerceExpr:
            {
                ArrayCoerceExpr *acexpr = (ArrayCoerceExpr *) node;

                APP_JUMB(acexpr->resulttype);
                JumbleExpr(jstate, (Node *) acexpr->arg);
                JumbleExpr(jstate, (Node *) acexpr->elemexpr);
            }
            break;
        case T_ConvertRowtypeExpr:
            {
                ConvertRowtypeExpr *crexpr = (ConvertRowtypeExpr *) node;

                APP_JUMB(crexpr->resulttype);
                JumbleExpr(jstate, (Node *) crexpr->arg);
            }
            break;
        case T_CollateExpr:
            {
                CollateExpr *ce = (CollateExpr *) node;

                APP_JUMB(ce->collOid);
                JumbleExpr(jstate, (Node *) ce->arg);
            }
            break;
        case T_CaseExpr:
            {
                CaseExpr   *caseexpr = (CaseExpr *) node;

                JumbleExpr(jstate, (Node *) caseexpr->arg);
                foreach(temp, caseexpr->args)
                {
                    CaseWhen   *when = lfirst_node(CaseWhen, temp);

                    JumbleExpr(jstate, (Node *) when->expr);
                    JumbleExpr(jstate, (Node *) when->result);
                }
                JumbleExpr(jstate, (Node *) caseexpr->defresult);
            }
            break;
        case T_CaseTestExpr:
            {
                CaseTestExpr *ct = (CaseTestExpr *) node;

                APP_JUMB(ct->typeId);
            }
            break;
        case T_ArrayExpr:
            JumbleExpr(jstate, (Node *) ((ArrayExpr *) node)->elements);
            break;
        case T_RowExpr:
            JumbleExpr(jstate, (Node *) ((RowExpr *) node)->args);
            break;
        case T_RowCompareExpr:
            {
                RowCompareExpr *rcexpr = (RowCompareExpr *) node;

                APP_JUMB(rcexpr->rctype);
                JumbleExpr(jstate, (Node *) rcexpr->largs);
                JumbleExpr(jstate, (Node *) rcexpr->rargs);
            }
            break;
        case T_CoalesceExpr:
            JumbleExpr(jstate, (Node *) ((CoalesceExpr *) node)->args);
            break;
        case T_MinMaxExpr:
            {
                MinMaxExpr *mmexpr = (MinMaxExpr *) node;

                APP_JUMB(mmexpr->op);
                JumbleExpr(jstate, (Node *) mmexpr->args);
            }
            break;
        case T_SQLValueFunction:
            {
                SQLValueFunction *svf = (SQLValueFunction *) node;

                APP_JUMB(svf->op);
                /* type is fully determined by op */
                APP_JUMB(svf->typmod);
            }
            break;
        case T_XmlExpr:
            {
                XmlExpr    *xexpr = (XmlExpr *) node;

                APP_JUMB(xexpr->op);
                JumbleExpr(jstate, (Node *) xexpr->named_args);
                JumbleExpr(jstate, (Node *) xexpr->args);
            }
            break;
        case T_NullTest:
            {
                NullTest   *nt = (NullTest *) node;

                APP_JUMB(nt->nulltesttype);
                JumbleExpr(jstate, (Node *) nt->arg);
            }
            break;
        case T_BooleanTest:
            {
                BooleanTest *bt = (BooleanTest *) node;

                APP_JUMB(bt->booltesttype);
                JumbleExpr(jstate, (Node *) bt->arg);
            }
            break;
        case T_CoerceToDomain:
            {
                CoerceToDomain *cd = (CoerceToDomain *) node;

                APP_JUMB(cd->resulttype);
                JumbleExpr(jstate, (Node *) cd->arg);
            }
            break;
        case T_CoerceToDomainValue:
            {
                CoerceToDomainValue *cdv = (CoerceToDomainValue *) node;

                APP_JUMB(cdv->typeId);
            }
            break;
        case T_SetToDefault:
            {
                SetToDefault *sd = (SetToDefault *) node;

                APP_JUMB(sd->typeId);
            }
            break;
        case T_CurrentOfExpr:
            {
                CurrentOfExpr *ce = (CurrentOfExpr *) node;

                APP_JUMB(ce->cvarno);
                if (ce->cursor_name)
                    APP_JUMB_STRING(ce->cursor_name);
                APP_JUMB(ce->cursor_param);
            }
            break;
        case T_NextValueExpr:
            {
                NextValueExpr *nve = (NextValueExpr *) node;

                APP_JUMB(nve->seqid);
                APP_JUMB(nve->typeId);
            }
            break;
        case T_InferenceElem:
            {
                InferenceElem *ie = (InferenceElem *) node;

                APP_JUMB(ie->infercollid);
                APP_JUMB(ie->inferopclass);
                JumbleExpr(jstate, ie->expr);
            }
            break;
        case T_TargetEntry:
            {
                TargetEntry *tle = (TargetEntry *) node;

                APP_JUMB(tle->resno);
                APP_JUMB(tle->ressortgroupref);
                JumbleExpr(jstate, (Node *) tle->expr);
            }
            break;
        case T_RangeTblRef:
            {
                RangeTblRef *rtr = (RangeTblRef *) node;

                APP_JUMB(rtr->rtindex);
            }
            break;
        case T_JoinExpr:
            {
                JoinExpr   *join = (JoinExpr *) node;

                APP_JUMB(join->jointype);
                APP_JUMB(join->isNatural);
                APP_JUMB(join->rtindex);
                JumbleExpr(jstate, join->larg);
                JumbleExpr(jstate, join->rarg);
                JumbleExpr(jstate, join->quals);
            }
            break;
        case T_FromExpr:
            {
                FromExpr   *from = (FromExpr *) node;

                JumbleExpr(jstate, (Node *) from->fromlist);
                JumbleExpr(jstate, from->quals);
            }
            break;
        case T_OnConflictExpr:
            {
                OnConflictExpr *conf = (OnConflictExpr *) node;

                APP_JUMB(conf->action);
                JumbleExpr(jstate, (Node *) conf->arbiterElems);
                JumbleExpr(jstate, conf->arbiterWhere);
                JumbleExpr(jstate, (Node *) conf->onConflictSet);
                JumbleExpr(jstate, conf->onConflictWhere);
                APP_JUMB(conf->constraint);
                APP_JUMB(conf->exclRelIndex);
                JumbleExpr(jstate, (Node *) conf->exclRelTlist);
            }
            break;
        case T_List:
            foreach(temp, (List *) node)
            {
                JumbleExpr(jstate, (Node *) lfirst(temp));
            }
            break;
        case T_IntList:
            foreach(temp, (List *) node)
            {
                APP_JUMB(lfirst_int(temp));
            }
            break;
        case T_SortGroupClause:
            {
                SortGroupClause *sgc = (SortGroupClause *) node;

                APP_JUMB(sgc->tleSortGroupRef);
                APP_JUMB(sgc->eqop);
                APP_JUMB(sgc->sortop);
                APP_JUMB(sgc->nulls_first);
            }
            break;
        case T_GroupingSet:
            {
                GroupingSet *gsnode = (GroupingSet *) node;

                JumbleExpr(jstate, (Node *) gsnode->content);
            }
            break;
        case T_WindowClause:
            {
                WindowClause *wc = (WindowClause *) node;

                APP_JUMB(wc->winref);
                APP_JUMB(wc->frameOptions);
                JumbleExpr(jstate, (Node *) wc->partitionClause);
                JumbleExpr(jstate, (Node *) wc->orderClause);
                JumbleExpr(jstate, wc->startOffset);
                JumbleExpr(jstate, wc->endOffset);
            }
            break;
        case T_CommonTableExpr:
            {
                CommonTableExpr *cte = (CommonTableExpr *) node;

                /* we store the string name because RTE_CTE RTEs need it */
                APP_JUMB_STRING(cte->ctename);
                JumbleQuery(jstate, castNode(Query, cte->ctequery));
            }
            break;
        case T_SetOperationStmt:
            {
                SetOperationStmt *setop = (SetOperationStmt *) node;

                APP_JUMB(setop->op);
                APP_JUMB(setop->all);
                JumbleExpr(jstate, setop->larg);
                JumbleExpr(jstate, setop->rarg);
            }
            break;
        case T_RangeTblFunction:
            {
                RangeTblFunction *rtfunc = (RangeTblFunction *) node;

                JumbleExpr(jstate, rtfunc->funcexpr);
            }
            break;
        case T_TableFunc:
            {
                TableFunc  *tablefunc = (TableFunc *) node;

                JumbleExpr(jstate, tablefunc->docexpr);
                JumbleExpr(jstate, tablefunc->rowexpr);
                JumbleExpr(jstate, (Node *) tablefunc->colexprs);
            }
            break;
        case T_TableSampleClause:
            {
                TableSampleClause *tsc = (TableSampleClause *) node;

                APP_JUMB(tsc->tsmhandler);
                JumbleExpr(jstate, (Node *) tsc->args);
                JumbleExpr(jstate, (Node *) tsc->repeatable);
            }
            break;
        default:
            /* Only a warning, since we can stumble along anyway */
            elog(WARNING, "unrecognized node type: %d",
                 (int) nodeTag(node));
            break;
    }
}

/*
 * Record location of constant within query string of query tree
 * that is currently being walked.
 */
static void
RecordConstLocation(pgssJumbleState *jstate, int location)
{
    /* -1 indicates unknown or undefined location */
    if (location >= 0)
    {
        /* enlarge array if needed */
        if (jstate->clocations_count >= jstate->clocations_buf_size)
        {
            jstate->clocations_buf_size *= 2;
            jstate->clocations = (pgssLocationLen *)
                repalloc(jstate->clocations,
                         jstate->clocations_buf_size *
                         sizeof(pgssLocationLen));
        }
        jstate->clocations[jstate->clocations_count].location = location;
        /* initialize lengths to -1 to simplify fill_in_constant_lengths */
        jstate->clocations[jstate->clocations_count].length = -1;
        jstate->clocations_count++;
    }
}

/*
 * Generate a normalized version of the query string that will be used to
 * represent all similar queries.
 *
 * Note that the normalized representation may well vary depending on
 * just which "equivalent" query is used to create the hashtable entry.
 * We assume this is OK.
 *
 * If query_loc > 0, then "query" has been advanced by that much compared to
 * the original string start, so we need to translate the provided locations
 * to compensate.  (This lets us avoid re-scanning statements before the one
 * of interest, so it's worth doing.)
 *
 * *query_len_p contains the input string length, and is updated with
 * the result string length on exit.  The resulting string might be longer
 * or shorter depending on what happens with replacement of constants.
 *
 * Returns a palloc'd string.
 */
static char *
generate_normalized_query(pgssJumbleState *jstate, const char *query,
                          int query_loc, int *query_len_p, int encoding)
{
    char       *norm_query;
    int         query_len = *query_len_p;
    int         i,
                norm_query_buflen,  /* Space allowed for norm_query */
                len_to_wrt,     /* Length (in bytes) to write */
                quer_loc = 0,   /* Source query byte location */
                n_quer_loc = 0, /* Normalized query byte location */
                last_off = 0,   /* Offset from start for previous tok */
                last_tok_len = 0;   /* Length (in bytes) of that tok */

    /*
     * Get constants' lengths (core system only gives us locations).  Note
     * this also ensures the items are sorted by location.
     */
    fill_in_constant_lengths(jstate, query, query_loc);

    /*
     * Allow for $n symbols to be longer than the constants they replace.
     * Constants must take at least one byte in text form, while a $n symbol
     * certainly isn't more than 11 bytes, even if n reaches INT_MAX.  We
     * could refine that limit based on the max value of n for the current
     * query, but it hardly seems worth any extra effort to do so.
     */
    norm_query_buflen = query_len + jstate->clocations_count * 10;

    /* Allocate result buffer */
    norm_query = palloc(norm_query_buflen + 1);

    for (i = 0; i < jstate->clocations_count; i++)
    {
        int         off,        /* Offset from start for cur tok */
                    tok_len;    /* Length (in bytes) of that tok */

        off = jstate->clocations[i].location;
        /* Adjust recorded location if we're dealing with partial string */
        off -= query_loc;

        tok_len = jstate->clocations[i].length;

        if (tok_len < 0)
            continue;           /* ignore any duplicates */

        /* Copy next chunk (what precedes the next constant) */
        len_to_wrt = off - last_off;
        len_to_wrt -= last_tok_len;

        Assert(len_to_wrt >= 0);
        memcpy(norm_query + n_quer_loc, query + quer_loc, len_to_wrt);
        n_quer_loc += len_to_wrt;

        /* And insert a param symbol in place of the constant token */
        n_quer_loc += sprintf(norm_query + n_quer_loc, "$%d",
                              i + 1 + jstate->highest_extern_param_id);

        quer_loc = off + tok_len;
        last_off = off;
        last_tok_len = tok_len;
    }

    /*
     * We've copied up until the last ignorable constant.  Copy over the
     * remaining bytes of the original query string.
     */
    len_to_wrt = query_len - quer_loc;

    Assert(len_to_wrt >= 0);
    memcpy(norm_query + n_quer_loc, query + quer_loc, len_to_wrt);
    n_quer_loc += len_to_wrt;

    Assert(n_quer_loc <= norm_query_buflen);
    norm_query[n_quer_loc] = '\0';

    *query_len_p = n_quer_loc;
    return norm_query;
}

/*
 * Given a valid SQL string and an array of constant-location records,
 * fill in the textual lengths of those constants.
 *
 * The constants may use any allowed constant syntax, such as float literals,
 * bit-strings, single-quoted strings and dollar-quoted strings.  This is
 * accomplished by using the public API for the core scanner.
 *
 * It is the caller's job to ensure that the string is a valid SQL statement
 * with constants at the indicated locations.  Since in practice the string
 * has already been parsed, and the locations that the caller provides will
 * have originated from within the authoritative parser, this should not be
 * a problem.
 *
 * Duplicate constant pointers are possible, and will have their lengths
 * marked as '-1', so that they are later ignored.  (Actually, we assume the
 * lengths were initialized as -1 to start with, and don't change them here.)
 *
 * If query_loc > 0, then "query" has been advanced by that much compared to
 * the original string start, so we need to translate the provided locations
 * to compensate.  (This lets us avoid re-scanning statements before the one
 * of interest, so it's worth doing.)
 *
 * N.B. There is an assumption that a '-' character at a Const location begins
 * a negative numeric constant.  This precludes there ever being another
 * reason for a constant to start with a '-'.
 */
static void
fill_in_constant_lengths(pgssJumbleState *jstate, const char *query,
                         int query_loc)
{
    pgssLocationLen *locs;
    core_yyscan_t yyscanner;
    core_yy_extra_type yyextra;
    core_YYSTYPE yylval;
    YYLTYPE     yylloc;
    int         last_loc = -1;
    int         i;

    /*
     * Sort the records by location so that we can process them in order while
     * scanning the query text.
     */
    if (jstate->clocations_count > 1)
        qsort(jstate->clocations, jstate->clocations_count,
              sizeof(pgssLocationLen), comp_location);
    locs = jstate->clocations;

    /* initialize the flex scanner --- should match raw_parser() */
    yyscanner = scanner_init(query,
                             &yyextra,
                             ScanKeywords,
                             NumScanKeywords);

    /* we don't want to re-emit any escape string warnings */
    yyextra.escape_string_warning = false;

    /* Search for each constant, in sequence */
    for (i = 0; i < jstate->clocations_count; i++)
    {
        int         loc = locs[i].location;
        int         tok;

        /* Adjust recorded location if we're dealing with partial string */
        loc -= query_loc;

        Assert(loc >= 0);

        if (loc <= last_loc)
            continue;           /* Duplicate constant, ignore */

        /* Lex tokens until we find the desired constant */
        for (;;)
        {
            tok = core_yylex(&yylval, &yylloc, yyscanner);

            /* We should not hit end-of-string, but if we do, behave sanely */
            if (tok == 0)
                break;          /* out of inner for-loop */

            /*
             * We should find the token position exactly, but if we somehow
             * run past it, work with that.
             */
            if (yylloc >= loc)
            {
                if (query[loc] == '-')
                {
                    /*
                     * It's a negative value - this is the one and only case
                     * where we replace more than a single token.
                     *
                     * Do not compensate for the core system's special-case
                     * adjustment of location to that of the leading '-'
                     * operator in the event of a negative constant.  It is
                     * also useful for our purposes to start from the minus
                     * symbol.  In this way, queries like "select * from foo
                     * where bar = 1" and "select * from foo where bar = -2"
                     * will have identical normalized query strings.
                     */
                    tok = core_yylex(&yylval, &yylloc, yyscanner);
                    if (tok == 0)
                        break;  /* out of inner for-loop */
                }

                /*
                 * We now rely on the assumption that flex has placed a zero
                 * byte after the text of the current token in scanbuf.
                 */
                locs[i].length = strlen(yyextra.scanbuf + loc);
                break;          /* out of inner for-loop */
            }
        }

        /* If we hit end-of-string, give up, leaving remaining lengths -1 */
        if (tok == 0)
            break;

        last_loc = loc;
    }

    scanner_finish(yyscanner);
}

/*
 * comp_location: comparator for qsorting pgssLocationLen structs by location
 */
static int
comp_location(const void *a, const void *b)
{
    int         l = ((const pgssLocationLen *) a)->location;
    int         r = ((const pgssLocationLen *) b)->location;

    if (l < r)
        return -1;
    else if (l > r)
        return +1;
    else
        return 0;
}